AI for Quality Control in Manufacturing

Defect detection models trained on your product images using architectures matched to your inspection task. ResNet-50 and EfficientNet-B4 are strong baseline choices for image classification tasks where the defect occupies a known region of the frame. YOLOv8 is preferred for object detection tasks where multiple defect instances can appear at arbitrary positions within the image and bounding box localisation is required alongside classification. The model architecture selection is driven by your defect type, throughput requirement, and available inference hardware -- not by a single-architecture default.

Image datasets are built using annotation tools such as CVAT (Computer Vision Annotation Tool) or Label Studio, with your quality engineers defining the labelling standard: what constitutes a reject, what is a borderline case requiring secondary review, and what is acceptable product variation that the model should pass. A consistent labelling standard applied before annotation starts is the most important factor in model performance -- ambiguous labelling standards produce ambiguous models. Models are trained to generalise across lighting condition variation, product colour and size variants, and orientation differences on the line. Confidence-scored detection with configurable threshold setting lets your quality team balance sensitivity (catching more defects) against specificity (reducing false rejections) based on the cost of a false positive versus the cost of a defect escaping to the customer. The Type I / Type II error tradeoff -- false rejection rate versus defect escape rate -- is calibrated against your quality cost model during threshold tuning rather than set arbitrarily.

Camera and lighting system design for your specific inspection task. GigE Vision cameras from Basler or Allied Vision are the standard for industrial line inspection: GigE Vision is the machine vision protocol standard that covers camera control, image acquisition, and data transfer over standard Ethernet, allowing deterministic frame triggering at production line speed without the latency variability of USB cameras. Basler ace and pulse series and Allied Vision Alvium series are commonly used for inline inspection at line speeds from 100 to 600 units per minute, depending on lens, resolution, and shutter speed selection. Lighting configuration -- diffuse dome for surface defect detection, telecentric backlight for dimensional measurement, coaxial illumination for reflective surfaces -- is specified during hardware scoping based on the defect types and product geometry.

Inference runs on an NVIDIA Jetson Orin edge module for installations requiring on-premise processing without a server room, or on a rack-mounted GPU inference server for higher throughput or multi-camera installations. The ONNX runtime and TensorRT optimisation pipeline converts the trained model to a TensorRT engine optimised for the target GPU, reducing inference latency to 10-30ms per image on Jetson Orin for typical detection models. OpenCV handles image preprocessing: adaptive thresholding, morphological operations (erosion, dilation) to separate connected defect regions, and normalisation before inference. Integration with production line PLCs via digital I/O or OPC UA protocol outputs a rejection signal to trigger the reject gate, air blower, or diverter mechanism within the inter-unit gap, removing the defective unit from the line without stopping throughput.

Classification models that distinguish defect types, not just defect presence. A binary pass/fail model tells you a defect was found; a classification model tells you it was a surface scratch, not a dimensional failure, and that matters because the corrective actions are different: a scratch may trace to a handling issue at a specific production stage, while a dimensional failure traces to tooling wear or a machine calibration drift. YOLOv8 classification heads or separate EfficientNet-B4 classifiers are trained per defect category using bounding-box annotated images from CVAT or Label Studio, with each defect type as a separate class.

Severity grading within defect types adds a second dimension to the classification: a minor surface mark that is cosmetic-only gets a different disposition than a major surface defect that affects sealing or structural integrity. Severity grades are defined by your quality engineering team and encoded in the labelling standard before annotation. The model learns to map visual features to severity grades based on that standard. Every rejection is captured with: defect class label, severity grade, confidence score, bounding box coordinates on the inspection image, a cropped image of the detected defect region, timestamp, production line, and shift. This image archive is stored in S3 (or equivalent object storage) with structured metadata -- defect class, severity, product SKU, line, date -- indexed for retrieval by query. The indexed archive is the dataset for model retraining: as production conditions change and new defect types emerge, the archive provides the foundation for iterative model improvement without rebuilding the dataset from scratch.

Real-time defect rate data fed into your Statistical Process Control system as each inspection event is generated. SPC integration sends structured inspection results -- defect count, defect type, production line, timestamp, and SKU -- to your SPC platform via API (InfinityQS, Minitab Workspace, or custom SPC systems) or a direct data stream to the SPC database. This replaces the end-of-shift manual tally sheet as the data source for SPC charts, giving your quality engineers live control chart data rather than a picture of what happened eight hours ago.

Defect rate P-charts and NP-charts by defect type, production line, shift, and SKU update in real time as the inspection generates data. When a data point crosses a control limit (typically set at three sigma from the mean), an alert is triggered to the quality engineer and line supervisor before the next inspection cycle, not at the next scheduled quality review. Process capability metrics (Cpk, Ppk) calculated from the inspection measurement data identify which lines and SKUs are most at risk of producing out-of-spec product. SPC trend detection using Western Electric rules (eight consecutive points on one side of the centreline, six consecutive points trending in one direction) flags systematic process drift before a control limit is crossed. The SPC integration also feeds the CMMS (Computerised Maintenance Management System): a sustained increase in a defect type associated with tooling wear or machine degradation triggers a preventive maintenance work order in the CMMS automatically, rather than waiting for a quality hold to prompt the investigation.

Rejection management workflows triggered automatically when the inspection system classifies a unit as a reject. The rejection event creates a record with defect class, severity grade, confidence score, and inspection image, and simultaneously sends a signal to the line PLC to divert or eject the unit. Downstream from the line, the rejected unit is tracked through your WMS (Warehouse Management System) or ERP using the serial or batch number captured at inspection: rework routing sends correctable defects to the rework station with the defect description and corrective procedure attached; scrap classification sends non-correctable defects directly to scrap with a disposal record. Rework completion is recorded against the original rejection so the rework conversion rate (how many rejections are recoverable) is tracked over time.

Rejection reason codes are automatically populated from the defect classification model output and mapped to your ERP's quality notification reason codes -- the quality notification in SAP QM or the nonconformance record in your QMS is created from the inspection data without manual entry. Shift-end rejection summary reports are generated automatically: total units inspected, total rejects by defect type, rejection rate by line and SKU, top three defect causes by frequency. Quality hold workflows are triggered when a batch's cumulative rejection rate exceeds a defined threshold -- the batch is flagged for hold in the WMS, the quality manager is notified, and a hold disposition workflow starts that requires a quality decision before the batch moves to the next production stage. The false positive rate is tracked separately: units flagged by the system as rejects but confirmed as acceptable on secondary review feed a false positive metric that is monitored alongside true reject rate, because an excessive false positive rate drives line downtime and rework cost that partially offsets the defect escape savings.

Quality analytics dashboards built for two audiences with different data needs. Production-floor dashboards display real-time inspection status: units inspected in the current shift, current defect rate versus target, top defect type in the current run, and a live feed of the most recent rejection images -- information a line supervisor can act on immediately. Management dashboards aggregate across lines, shifts, SKUs, and time periods: first-pass yield trends, defect rate by line and shift, top defect causes ranked by frequency and cost impact, and comparison across production lines to identify which equipment or process generates the most quality risk.

Correlation analysis between production parameters and defect rates is built into the analytics layer: machine speed, die temperature, raw material lot number, and operator shift are stored alongside each inspection event so the data team can query whether defect rate increased when lot 4472 was introduced, or whether Line 3 consistently produces more surface scratches on the night shift. This correlation data is what drives root cause analysis from hypothesis to evidence. Pareto charts of defect causes, ranked by frequency and by downstream quality cost (escape cost, rework cost, customer return cost), prioritise which corrective actions to invest in first. Defect image archive in S3 with metadata indexing (defect class, severity, line, SKU, date, production parameters) supports model retraining: when a new defect type appears in production, the quality team uses Label Studio to annotate examples from the archive and the model is retrained on the augmented dataset without rebuilding the annotation infrastructure. The analytics layer is where inspection data becomes quality improvement intelligence -- the difference between knowing your defect rate and knowing what is causing it and how to reduce it.