NVIDIA Jetson Orin NX PCB Inspection System for Missing and Misplaced Components

High-Quality Image Capture and Calibration

Image quality determines whether machine-learning performance can be trusted. A capable AI model cannot fully compensate for blurred, overexposed or inconsistent images.

The process of our imaging engineering begins with the calculation of the field of vision. The chosen camera sensor's pixel density should be sufficient for defect spotting. For instance, determining the absence of a big connector will require lower resolution than verifying if chip capacitors are in their place.

The inspection system may involve either one high-definition camera, several cameras for big panels or a camera moving high over the object if its size is larger than the maximum allowed field of vision.

The capture pipeline includes:

• Fixed camera-to-board distance and rigid mechanical mounting
• Lens selection to minimize distortion and preserve edge sharpness
• Exposure, gain and white-balance control locked for production conditions
• Diffuse and directional lighting trials for solder and silkscreen contrast
• Triggered image capture after the PCB reaches a stable position
• Calibration routines for pixel-to-millimeter conversion and board alignment
• Image-quality checks for blur, excessive glare and incorrect board placement

Prior to performing defect analysis, the program utilizes fiducial markings, PCB borders, or other reference attributes to align the photo taken with the reference board coordinate grid. This step is extremely important as any small movement of the fixture may introduce errors of misclassification.

Defect Dataset Development for Missing and Misplaced Components

The development of an efficient inspection model relies on the availability of representative data, not just ideal reference images. Adequate Infosoft will offer support in the form of dataset development, collection of images, guidance on annotation as well as model training and validation.

To conduct missing component inspections, the various possible places of placement are linked to the respective component class or designator. The system operates using AI detection and reference position logic.

The process of inspection begins with the detection of component or area of interest and latter compares the output from detection against the expected position and tolerances.

For misplaced components, the system measures the position, rotation and bounding area of the detected component relative to the board reference. A component may be marked for review if its centre point, angle or overlap exceeds configurable limits.

The dataset should include:

• Approved boards across normal production variation
• Defective boards containing missing, shifted, rotated or reversed components
• Multiple PCB batches, printing conditions and component suppliers
• Lighting, focus and background conditions expected at the station
• Difficult cases such as reflective solder, labels, flux residue and partial occlusion
• Clear defect definitions and consistent annotation standards

In situations where real defects are rare, the teams can deploy controlled rework samples to produce safe and representative instances. Synthetic augmentation techniques like brightness variations, small rotations, blurring, and scaling can increase robustness, but the real images should still be at the core of any training setup.

The model is assessed using precision, recall, false-positive rate, false-negative rate and class-specific performance. In quality inspection, the business cost of each error matters.

A false negative may allow a defective board to proceed; too many false positives can slow production and cause operators to ignore alerts. Thresholds should therefore be tuned with manufacturing and quality stakeholders, not selected only from a generic benchmark score.

Edge AI Inference Optimization on Jetson Orin NX

The application utilized for inspecting products can employ various methods such as the approaches of object detection, segmentation or anomaly detection based on the kind of product as well as defect characteristics.

The method of object detection is suitable where there is a need to identify specific types of components and verify its proper positions. The method of segmentation can be required when there is a need to study soldering areas or boundaries.

The use of the anomaly detection method can help identify visual discrepancies on the boards. Nonetheless, it requires careful verification as it can check out harmless defects during manufacturing processes.

The engineering workflow includes model selection, training, export, deployment and runtime profiling. Models can be optimized using NVIDIA TensorRT to reduce inference latency and improve edge throughput.

FP16 inference commonly provides a practical balance between speed and accuracy; INT8 may be evaluated for higher performance after calibration confirms that inspection accuracy remains acceptable.

The Jetson application processes images through a controlled sequence:

1. Scan PCB or work-order identifier.
2. Load the relevant inspection recipe and approved reference configuration.
3. Trigger image capture and run image-quality checks.
4. Align the PCB image using fiducials or known board features.
5. Run AI inference and post-processing.
6. Compare observed components with expected positions and tolerances.
7. Classify the result as pass, fail or operator-review required.
8. Store the result, defect overlays, confidence values and source image.
9. Display a clear outcome to the operator and send line integration signals.

This structure makes the system explainable. Rather than returning only "failed," the operator can see the affected board area, expected component location, detected condition, confidence value and image evidence.

Operator Dashboard and Manufacturing Traceability

The operator dashboard is a crucial component of the whole solution. An accurate technical model has no operational value if the operator can't interpret its results or if the quality manager can't trace back defects to a batch.

Adequate Infosoft provides the capability of creating a web-based dashboard to be used on a touchscreen station, supervisor PC or tablet. It can show the live camera status, current PCB recipe, pass/fail counts, last defects, sample images, and device health.

For a failed inspection, an operator can zoom into the defect overlay and select actions such as:

• Confirm defect and route the board to rework
• Mark the alert as an accepted variation with reason code
• Request supervisor review
• Capture a note or photograph for quality records
• Re-run inspection after rework
• Place the station in maintenance or recipe-change mode

The backend can record serial number, board revision, operator ID, production line, timestamp, inspection duration, model version, confidence threshold and final disposition.

Traceability assists in performing root cause analysis. For instance, in case of an unexpected increase in alerts about shift components, we can presume that a pick and place feeder or issue with the stencil, fixture and board handling might be present.

Depending on the factory setup, the system can be interfaced with MES, ERP or quality management systems or PLC based line equipment using REST APIs, MQTT, OPC UA, Modbus TCP or file based interfaces.

Field Validation and Continuous Improvement

Deployment begins with controlled trials, followed by validation on real production lots. Engineers test camera stability, fixture repeatability, barcode handling, lighting performance, model latency, dashboard workflow, network recovery and data retention.

When discussing the validation plan, it is important to include aspects such as goodboard testing, purposely defected testing, repeated scans, different operators, and environmental changes.

All inaccuracies and failures of the process are examined together with the quality team and pictures from confirmed edge cases can be added to the controlled dataset for possible future improvements.

Whenever some model changes occur, it is important to version it, check it against specific validation data for the purpose of protecting inspection procedure from accuracy regressions.