A 2025 MDPI case study paired a Cognex smart camera with an Allen-Bradley PLC over implicit cyclic Ethernet/IP messaging, cutting false positives by re-validating every pass/fail inside the PLC scan rather than trusting the camera verdict alone [S1]. The same architecture, camera, trigger, and PLC handshake, scaled across 3D-printed parts, automotive connectors, and fiber-optic tubes without rewriting the inspection logic [S1].
Vision-PLC cells now ship in three structural variants: hardwired discrete I/O for binary pass/fail, industrial Ethernet (EtherNet/IP, PROFINET, Modbus TCP) for measurement arrays, and OPC UA for SCADA and MES hand-off [S2][S3]. Choice is driven by data volume, deterministic timing, and whether the line must stop a defective part or merely log it.
Protocol selection: discrete I/O vs EtherNet/IP vs PROFINET vs OPC UA
Discrete I/O is the simplest cell layout: one trigger input, one result output, one ready, one busy, deterministic to the PLC scan, immune to network faults, but limited to a single bit of pass/fail information [S2]. Vision systems that support EtherNet/IP appear as nodes on the network and exchange data through implicit or explicit messaging, carrying measurement arrays, error codes, and positional data alongside the basic result [S2]. PROFINET supports both RT and IRT classes, the latter isochronous and aimed at high-speed inspection where deterministic timing is critical, and remains the default on Siemens architectures [S2].
OPC UA has emerged as the SCADA-facing protocol of choice because of its platform-independent architecture and built-in security model, and many modern vision platforms now expose a native OPC UA server [S2]. Modbus TCP is a pragmatic retrofit choice when integrating cameras into legacy PLC control systems that cannot host EtherNet/IP or PROFINET scanners [S2]. For plants already running on a safety PLC for e-stop and zone logic, vision results still flow through the standard PLC scan, with safety-relevant reject commands handled as hardwired outputs to keep the safety chain independent of the vision network.
Hardware stack: sensor, interface, lighting, and processing
Industrial cameras in inspection cells differ from consumer units in four ways: global shutter (captures the whole frame at once, required for moving parts), industrial interfaces (GigE Vision, USB3 Vision, CoaXPress for deterministic data transfer), ruggedized construction for vibration, dust, and thermal extremes, and precise hardware triggering synchronized to line events [S4]. Lighting is the dominant variable: ring lights for flat surfaces, bar lights for surface-feature emphasis, dome lights for diffuse reflection suppression, backlights for edge and hole detection, and strobed/pulsed LED for high-speed lines [S4].
AI inference processing sits in one of three tiers: industrial PC with discrete NVIDIA GPU (standard, flexible), edge AI device such as NVIDIA Jetson (space-constrained cells), or fully integrated smart cameras with on-board GPU and IP65 enclosure (fastest deployment) [S4]. Vision cells that mix a line scan camera with discrete PLCs typically need a frame buffer or streaming PC because the sensor produces data continuously rather than on trigger, while area-scan smart cameras fit trigger-event models more cleanly [S4]. A representative integrated unit ships in an IP65 housing with photometric-stereo lighting, on-board NVIDIA GPU inference, and a trainable model workflow reported by the vendor at under one hour of training time [S4].
PLC-side design: scan timing, result handling, and reject logic
PLC programs must account for the asynchronous nature of vision processing: a photoelectric sensor returns in microseconds, a vision inspection returns tens to hundreds of milliseconds later, and the PLC scan must hold the part state until the result lands [S2]. A common failure mode is the PLC advancing the conveyor before the camera has emitted a result, causing a reject command to fire against the wrong index or against an empty station.
The MDPI 2025 case study addressed this with a lightweight Ladder Diagram interface and an implicit cyclic Ethernet/IP assembly that the PLC polls each scan, re-validating the camera's verdict against a tag map and only then driving the reject output [S1]. For an industrial camera cell tied to a high-speed molding line or similar indexed conveyor, the practical rule is to gate downstream motion on a Result_Valid bit plus a Result_Code register, not on a single pass/fail wire.
Acceptance criteria and when not to repair
A vision-PLC cell passes commissioning when the camera-to-PLC round trip holds below the indexed station pitch, the reject actuator fires within one station of the bad part, and false-pass rate is below the agreed quality level (commonly referenced as zero-defect target for automotive and medical cells) [S1][S3]. A cell should be escalated to the vision vendor or replaced, not repaired in-line, when the camera unit-to-unit image variance drifts past the AI model's training distribution, or when Ethernet error counters on the PLC's industrial Ethernet port rise above the supplier's documented threshold [S4].
Two systemic failure modes recur: lighting drift, where LED aging or contamination shifts pixel statistics and silently degrades AI accuracy, and protocol mismatch, where a team specifies EtherNet/IP on a PROFINET-only PLC scanner or vice versa [S2][S3]. Both are caught by routine image-log review and by validating the PLC's input assembly tag list against the camera's exported EDS/GSD file at FAT, not at line start-up.
Decision matrix: when to choose which integration
For binary pass/fail at speeds below one inspection per second and a PLC already on the line, discrete I/O is the lowest-risk path: four wires, no EDS file, deterministic, and any electrician can troubleshoot it [S2]. For measurement arrays, defect codes, or part traceability, EtherNet/IP on Allen-Bradley/Rockwell or PROFINET on Siemens are the default, with IRT-class PROFINET preferred where deterministic sub-millisecond timing is required [S2].
OPC UA is the right choice when the inspection result must travel from the cell to SCADA, MES, or a cloud analytics layer without bespoke middleware, and is increasingly bundled by vision vendors as a native server [S2]. Modbus TCP fits only when integrating into a legacy PLC that cannot host an industrial Ethernet scanner, accepting the loss of implicit messaging in exchange for universal support [S2]. Across all options, the cell should log every image plus the PLC tag snapshot at the moment of reject, because that paired record is what makes a post-event root cause analysis possible instead of a guessing game.
Trackable signal: the next revision wave to watch is hybrid AI-PLC integration, where the camera's deep-learning verdict is cross-checked against PLC-side rule logic before any reject output, an approach the MDPI 2025 case study explicitly flags as the direction for scalable Industry 4.0 deployment [S1]. Related reading on adjacent process-instrumentation topics: edge versus cloud analytics for vibration data and why thermal cameras click and freeze on NUC.