Industrial AI Safety Glossary

An AI camera is a camera that runs a trained machine-learning model on its video stream to recognize objects, people or events, instead of only recording footage for a human to review. In industrial settings the useful output of an AI camera is not the picture but the decision: a person is inside a restricted zone, a forklift and a pedestrian are converging, a worker is missing a helmet. The camera turns that decision into an action such as a siren, a warning light, an e-mail or a signal to a PLC.

There are two broad architectures. In a cloud or server-based system, cameras stream video to a central computer that runs the model; this is flexible but adds network latency and sends images off the device. In an edge AI camera, the model runs inside the camera housing, so detection and the output signal happen locally. ISEE-CAM follows the edge approach: camera, AI processor and alarm outputs sit in one housing, rated IP69K and specified for −20 to +70 °C.

When comparing AI cameras for safety use, the points that matter most are where inference runs (edge or cloud), the alert-to-output latency, the type of outputs available (dry contact, PLC, network protocols), the environmental rating of the housing, how detection zones are configured, and how personal data is handled under KVKK or GDPR. A good starting point is to map each hazard on site to a detection mode, then check whether the camera can drive the alarm or machine signal that hazard needs. The guide AI camera vs traditional safety sensors compares the approach with sensor-based systems, and the ISEE-CAM PLC integration page explains how the outputs connect to site controls. For a full checklist, see the industrial AI camera buyer's guide.

Anomaly detection is a machine-learning technique that learns what normal data looks like and flags anything that deviates from it, without needing examples of every possible fault. In quality control this matters because defects are varied and rare. Collecting thousands of labeled images of every scratch, bubble or stain is often impractical, and new defect types appear that nobody anticipated. An anomaly model is trained mostly on good products; at inspection time it measures how far each new image is from that learned normal and assigns a score. Areas that differ most are highlighted, often as a heatmap.

The approach needs a well-chosen threshold. Set it too low and normal variation is flagged as a defect; set it too high and real defects pass. Practical systems let the operator adjust the threshold per product, review flagged images and refine the reference over time. Controlled lighting and stable positioning help, because changes in the scene can look like anomalies to the model.

ISEE Vision's quality control systems use this method: a defect-free reference is taught, each item is compared to it in real time, and deviations are reported with a heatmap, confidence score and timestamp or meterage. The same principle is applied across industries, from automotive parts to food and packaging lines. The article anomaly detection and industrial applications explains the concept with examples. Anomaly detection is also used outside imaging, for example on vibration or temperature data for predictive maintenance.

Because the model only learns what normal looks like, anomaly detection is often combined with a short review step at the start of a project: operators confirm which flagged items are real defects, and the reference or threshold is adjusted until false alarms fall to an acceptable level for that product.

Automated visual inspection is the use of cameras and software to check products for defects, missing parts or out-of-tolerance dimensions without a human inspector looking at each item. Manual inspection depends on attention and is usually done on samples. Automated visual inspection can check every item at line speed, with the same criteria on every shift, and keep an image record of each decision. That record is often as valuable as the rejection itself, because it shows when and where a defect started.

Two methods are common. Rule-based inspection measures specific features against fixed tolerances. Reference-based or anomaly-based inspection teaches the system what an acceptable product looks like and flags anything that deviates from it, which is useful when defects are varied or hard to describe in advance. ISEE Vision uses the reference-based approach: a defect-free reference is taught, each item is compared to it in real time, deviations are shown on a heatmap with a confidence score, and the result is logged with a timestamp or meterage.

Typical applications include surface defects on cable and hose, burrs and cracks on metal parts, holes and stains on fabric, fill level and label checks in food packaging, and pill presence in pharmaceutical blisters. The quality control overview links to each industry page, for example cable, hose and pipe surface inspection and textile fabric inspection. When scoping a project, document the defect types, the line speed, the acceptance threshold and the action required at detection: alert only, alert and stop, or full automation through a PLC.

Edge AI is the practice of running artificial-intelligence models directly on the device that collects the data, such as a camera or sensor, rather than sending the data to a remote cloud server for processing. The word "edge" refers to the edge of the network: the point where data is created. An edge AI camera captures an image, runs the detection model on its own processor and acts on the result, all inside the same unit. Only the outcome, such as an alarm event or a log entry, needs to leave the device.

For safety applications, edge processing has three practical advantages. The first is latency: there is no round trip to a data center, so the time between an event and the output signal stays short and predictable. ISEE Vision specifies alert-to-output latency under 100 ms for ISEE-CAM. The second is resilience: detection keeps working if the internet connection or a server goes down. The third is privacy: video does not need to be streamed or stored off site, which simplifies compliance with KVKK and GDPR.

Cloud processing still has a role. It suits workloads that are not time-critical, such as long-term analytics, reporting dashboards and training new models on collected data. Many deployments therefore combine both: detection and alarms on the edge, aggregated statistics in a central system. The article edge AI vs cloud for safety systems covers the trade-offs in detail, and the ISEE-CAM product page lists the on-device specifications. When evaluating an edge AI product, ask where each step of the pipeline runs, what happens when the network is unavailable, and which data, if any, leaves the device.

Forklift proximity detection is any system that senses when a pedestrian, another vehicle or an obstacle comes too close to a forklift and warns the driver, the pedestrian or both. Forklift and pedestrian collisions are a leading cause of serious injuries in warehouses and factories, so many sites add technology on top of traffic rules, marked walkways and training. Proximity systems differ mainly in how they sense people. Tag-based systems such as RFID or UWB require every pedestrian to wear a tag and measure the distance between tag and vehicle. Radar and ultrasonic sensors detect objects without tags but cannot tell a person from a pallet. Camera-based systems use AI models to recognize people and vehicles directly in the image.

Each approach has trade-offs. Tags give accurate distance but miss visitors and contractors who are not issued a tag, and batteries must be maintained. Sensors are simple but generate nuisance alarms near racks and walls. AI cameras need no wearables and know what they are seeing, but they need a clear view of the area and good placement. Many sites combine methods depending on the zone.

ISEE-CAM provides camera-based proximity detection through modes such as forklift and personnel presence check and the blind spot warning system. Outputs drive warning lights, sirens or signals to site controls; slowing or stopping a forklift requires additional hardware on the vehicle. The comparison UWB vs AI camera for forklift safety walks through accuracy, coverage and maintenance in more detail, and forklift safety solutions covers the wider set of measures. The forklift accident statistics page shows why pedestrians on foot are the group most at risk.

IP69K is an ingress-protection rating meaning a device is dust-tight and withstands close-range, high-pressure, high-temperature water jets, as used in intensive wash-down cleaning. IP codes describe how well an enclosure keeps out solids and water. The first digit covers solids, where 6 means dust-tight. The second covers water, and the 9K level comes from standards originally written for road vehicles, ISO 20653 and the earlier DIN 40050-9, which test the enclosure with hot water sprayed at high pressure from short distance and several angles. It is a different test from immersion: a device rated IP68 for submersion is not automatically IP69K, and the reverse is also true, which is why some products list both ratings.

IP69K matters wherever equipment is cleaned aggressively: food and beverage plants, meat processing, dairies, commercial kitchens, car washes, and outdoor or heavy-industry sites exposed to dust and water. A lower-rated housing can let moisture in during cleaning, which leads to fogged optics, corrosion and failures.

ISEE-CAM has an IP69K-rated housing and is specified for −20 to +70 °C operation, together with ISO 16750-3 vibration resistance, as listed on the ISEE-CAM product page. That combination is what allows the same camera to be used in catering and industrial kitchens with daily wash-down, as well as in steel mills and logistics yards. When comparing cameras, check that the rating applies to the complete unit including connectors and cable glands, not only the lens housing.

A high ingress rating protects the electronics, but cleaning still matters for performance: water spots, grease or dust on the front window can reduce image quality, so mounting positions should allow easy wiping and avoid direct spray onto the lens during normal operation.

A light curtain is a safety-rated presence-sensing device that projects an array of infrared beams across an opening and sends a stop signal to the machine when any beam is interrupted. Light curtains are widely used on presses, robot cells, packaging machines and automated storage systems. A transmitter on one side and a receiver on the other create a protective field; when a hand, arm or body breaks the field, the curtain's safety outputs switch off and the machine's safety circuit stops hazardous motion. Safety light curtains are designed and certified to standards such as IEC 61496, with performance levels assessed under ISO 13849 or IEC 62061, and their mounting distance is calculated from the machine's stopping time.

Because they are certified safety components, light curtains are a primary safeguard. Their limits are geometric: they protect a defined line or plane, they cannot tell a hand from a workpiece without muting or blanking functions, and they do not see what happens beyond that plane. Muting and blanking must be configured carefully, because a poorly set exception can create a gap in protection.

AI cameras such as ISEE-CAM can add an extra layer around a light curtain, not replace it. A camera can watch a wider area, distinguish a hand from material, detect a person standing inside the danger zone after passing the curtain, and log events for review. The camera output can be wired to the PLC or safety relay so it works together with the machine's safety system. See light curtain or AI camera for press safety and the hand presence check in risky areas mode for how the two approaches complement each other.

Machine guarding is the set of physical guards and safety-rated devices that keep people away from a machine's hazardous moving parts, or stop the hazard before a person can reach it. Hazards on machines typically come from the point of operation, power-transmission parts such as belts and gears, and other moving parts like flywheels or reciprocating arms. Guarding methods are usually grouped into fixed guards that are permanently attached, interlocked guards that stop the machine when opened, adjustable and self-adjusting guards that adapt to the workpiece, and protective devices such as light curtains, two-hand controls, pressure mats and pullback devices. In the United States, OSHA 29 CFR 1910.212 sets general machine-guarding requirements; in Turkey and the EU, the Machinery Directive and its national regulation set the framework.

The order of protection matters. Risk assessment comes first, then inherently safe design, then guards and safety-rated devices, then warnings, training and PPE. Guards and protective devices are certified safety measures and form the core of a machine's safety concept.

AI vision can add a supplementary layer where guards leave gaps in awareness, for example detecting a person who is inside a cell, a hand near a press die, or someone waiting too long in a hazardous area. ISEE-CAM's hazardous machine area detection and hand presence check modes send signals that work together with the machine's safety system; they do not replace guards or safety-rated devices. The machine guarding guide and why machine accidents happen cover the topic in more depth.

A practical guarding review walks each machine with its operators, lists every way a person can reach a hazard during normal work, setup, cleaning and maintenance, and confirms that each route is closed by a guard or a safety-rated device before supplementary layers are considered.

Machine vision is the use of cameras, lighting and image-processing software to let machines inspect, measure or guide industrial processes automatically. Classic machine vision systems rely on rules written by an engineer: measure the distance between two edges, check that a pixel region is within a brightness range, read a barcode. They work very well when parts, lighting and positions are tightly controlled, and they have been standard on production lines for decades.

Modern systems add deep learning. Instead of hand-coded rules, a neural network learns what a good product or a person looks like from examples. This handles natural variation better, for example scratches of different shapes, fabric textures, or people in different poses and clothing. The term computer vision is often used for the broader science, while machine vision usually refers to industrial applications with defined inputs and outputs.

A complete machine vision station includes more than a camera. Controlled lighting keeps images stable, an encoder or trigger links each image to a position on the line, a processor runs the analysis, and digital outputs or a PLC connection turn the result into an action such as rejecting a part or stopping the line. ISEE Vision applies this pattern in its quality control systems, which compare each product to a learned defect-free reference. For the underlying techniques, see computer vision and object detection. When planning a project, define the defects or events to detect, the line speed, the required precision and the action expected at the moment of detection before choosing hardware.

Computer vision is the broad field of teaching computers to interpret images and video, while machine vision is its industrial application: cameras, lighting and software built into a production process to inspect, measure or guide it. The two terms overlap, and the difference is mostly one of scope. Computer vision covers the science and the algorithms, from classic image processing to deep-learning models for object detection, segmentation and tracking. It is used everywhere, including phones, cars, medical imaging and retail. Machine vision takes those techniques and packages them into a system with defined inputs and outputs on a factory floor: a camera at a fixed point, controlled lighting, a trigger or encoder that ties each image to a part or a position, and an output that rejects the part, stops the line or tells a PLC what was found.

Because a machine vision system must work every shift, the engineering around the algorithm matters as much as the algorithm itself. Stable lighting, the right optics for the required precision, line speed, the environment (dust, vibration, temperature) and how the result is acted on all decide whether a station is useful. A computer vision model that is accurate in a notebook is not yet a machine vision system; it becomes one when it is integrated with the line and its outcome is logged and acted on.

Modern machine vision increasingly uses deep learning, which blurs the line further. ISEE Vision's automated visual inspection systems apply reference-based AI anomaly detection on the device, comparing every product with a learned defect-free reference. For the underlying techniques, the article computer vision and object detection explains how detection models find and classify objects in an image, and the entries on machine vision and anomaly detection in this glossary cover the industrial side.

A near miss is an unplanned event that could have caused injury, damage or loss but did not, either by chance or because someone reacted in time. A forklift that brakes a metre short of a pedestrian, a load that slips from a crane hook over an empty walkway, a hand that reaches into a press just after the stroke: none of these cause harm, yet each shows that the conditions for an accident were present. Safety practice treats near misses as free lessons, because the same underlying cause will eventually produce a real accident if nothing changes.

Near-miss reporting programs ask workers to record these events: what happened, where, when, which equipment was involved and what could have happened. The value comes from analysis. Patterns such as repeated forklift-pedestrian encounters at the same corner point to a layout, visibility or traffic-rule problem that can be fixed before anyone is hurt. The main weakness of manual reporting is under-reporting, because events are forgotten, considered too minor or feared to reflect badly on the people involved.

Camera-based detection adds an objective record. When an ISEE-CAM forklift and personnel detection mode flags a forklift and a person in the same zone, the event is logged with time and image, whether or not anyone files a report. Reviewing these logs shows where interactions cluster. The article on near miss reporting explains how to build a reporting process, and the Turkish term is covered under ramak kala in the Turkish glossary. Near-miss data works best alongside, not instead of, worker reports, because people can describe context a camera cannot see.

PPE compliance is the degree to which workers wear the personal protective equipment required for their task and area, such as helmets, high-visibility vests, gloves, goggles or harnesses. Employers must assess risks, provide suitable PPE, train workers in its use and make sure it is actually worn. The last step is the hardest: a helmet left on a bench or a vest removed in summer heat does not protect anyone. Compliance is usually checked through supervision, entry controls and periodic audits, all of which cover only part of the working day.

AI-based PPE detection checks compliance continuously at the points where it matters. A camera model is trained to recognize each item of equipment on a person, then raises an alert when a required item is missing. ISEE-CAM offers this as a helmet and vest check for work areas and as PPE-checked access permission at turnstiles and doors, where passage is only allowed once the required equipment is detected. The same principle is applied to hygiene clothing such as hairnets and aprons in catering and industrial kitchens.

Good PPE compliance programs use detection data to improve conditions rather than only to discipline. If violations cluster at one entrance or one shift, the cause may be missing PPE stock, uncomfortable equipment or unclear signage. Because PPE detection processes images of people, data-protection requirements under KVKK or GDPR apply: inform workers, limit retention, and prefer on-device processing where possible. The article how AI PPE detection works explains the detection process in more detail.

A safety observation is a structured check in which a supervisor, safety professional or co-worker watches work being done and records safe and unsafe behaviors or conditions. Observation programs, often called behavior-based safety, aim to catch risk before it turns into an incident. An observer might note that a worker entered a marked forklift lane, that a guard was left open, or that hearing protection was not worn in a noisy area. Positive observations are recorded too, so the program reinforces good practice instead of only finding fault.

Manual observations have clear limits. They cover only the time an observer is present, people behave differently when they know they are watched, and results depend on the observer's attention and judgment. They remain valuable for coaching and for understanding why something happened, which a sensor cannot explain.

AI cameras can extend observation to every shift for the specific behaviors they are configured to detect. Examples include helmet and vest checks, people climbing racks, personnel waiting too long in a hazardous zone, or entry into restricted areas. Each detection is timestamped and can feed daily, weekly or monthly reports, which gives safety teams a consistent baseline to compare against manual observations. The combination works best when camera data is used to target coaching and fix conditions, and when workers are told clearly what is monitored and why, which is also a requirement under data-protection rules. For a wider view of how AI fits into safety programs, see the 2026 guide to AI in workplace safety.

A safety PLC is a programmable logic controller designed and certified for safety functions, with redundant processing and self-diagnostics so that a fault leads the machine to a safe state. A standard PLC controls the normal operation of a machine: sequences, timers, motor commands. A safety PLC handles the functions that protect people, such as emergency stops, guard interlocks, light curtain inputs and safe speed monitoring. It is built with redundant channels and continuous self-testing, and it is certified to functional-safety standards such as IEC 61508, with machine applications evaluated under ISO 13849 (performance level, PL) or IEC 62061 (safety integrity level, SIL).

Safety PLCs are configured with certified function blocks and their programs are validated as part of the machine's risk assessment. On smaller machines, the same role can be played by dedicated safety relays. Either way, the safety controller is what decides that hazardous motion must stop, and its integrity is what the machine's safety rating depends on.

An AI camera connects to this world as an input device. ISEE-CAM PLC integration describes how the camera's dry-contact outputs or network signals reach a standard PLC or a safety relay, for example to trigger a slowdown, an alarm or a stop request when a person is detected in a hazardous zone. The camera does not become a safety PLC and does not replace one; unless a camera is itself certified for a safety function, its signal is treated as an additional layer within the safety concept. The hazardous machine area detection mode shows a typical use of this connection.