engiAI

1. Why engiAI and the ZONE-PAIR method?

engiAI Ethernet System – a visual inspection program for production lines. It takes a frame from an industrial or standard camera, divides it into zones, checks each zone using a trained neural network, and converts the result into an action: a signal to the camera's discrete output, a TCP message to a robot or PLC, or a record in a log. The ZONE-PAIR method is designed for products where each location has its "correct" component: printed circuit boards, assemblies, sets of parts in a tray, and packaging with specific layouts.

1.1. What does the method offer compared to classifying the entire frame?

A common approach is to use a single model with two classes, "normal/anomaly," applied to the entire image or to specific zones. However, this doesn't work on the circuit board: a gap under the BR1 bridge in zone 2 indicates a defect, while the same gap in another area of the board is considered normal. A single model cannot distinguish between these two cases. The other extreme is to use a separate model for each component, which would require dozens of training runs and numerous files in memory.

ZONE-PAIR addresses this by using a single model, with as many classes as there are zones. A "Zone" is a rectangle surrounding a single component. A "Class" is a folder containing images that the network learns from. Each zone has its own pair: "zone_001" (normal zone 1, for example, the C30 capacitor) and "zone_001_anomaly" (defect zone 1); "zone_002" and "zone_002_anomaly" represent the BR1 bridge. When checking zone 1, the program only looks at its pair; the answers for zone 2 are not used. Anomaly classes are optional: the "defect" changes the appearance of the zone, and the confidence in the "normal" state decreases automatically; however, with synthetic data (section 4.6), they can be created with a single click and noticeably improve reliability.

1.2. Speed: The AI only looks at the relevant sections

The neural network never receives the entire frame. Only the rectangle corresponding to each zone is extracted, and the zone number is selected from two classes. Zones from a single frame are processed in parallel (up to four threads), and the results from all zones are displayed together as a single batch, rather than one zone at a time.

• The analysis speed is determined by time. The "Analyses per second (AI)" slider on the Camera tab indicates exactly N analyses per second, not "every Nth frame." The limit is the camera's frame rate and inference time.
• GPU is used by default. Inference is performed using ONNX Runtime with CUDA; if there is no NVIDIA graphics card, the program automatically switches to the CPU and indicates this in the status (GPU (CUDA) or CPU).
• Clean buffers. Memory allocated for clippings is used once per stream and then reused; analysis time remains consistent across frames.
• Model for the task. For ZONE-PAIR, the program automatically sets the input size to 128 px and uses the YOLO11 "s" size network: the extracted zone is barely scaled, and the network has sufficient capacity to handle hundreds of pairs of classes.
• A live conveyor saves labor. The operator doesn't need to manually move the parts through the process as frequently as in a traditional system; the parts are picked and aligned once, and then the cutting is performed.
• The display does not interfere with analysis. The tile is redrawn no more than 25 times per second, and only based on the new frame from the camera. The image is reduced to the size of the tile before being displayed on the screen, and the log is written to a file by a background thread. The interface remains responsive even with 8-megapixel frames and hundreds of zones.

Displayed above the video, in the status bar, are the actual figures: fps (frames per second) and the milliseconds of the last analysis, as well as Eff, which indicates how many analyses were actually completed per second.

1.3. What accuracy to expect: Let's calculate using our camera

The program does not guarantee a "0.5 mm" shift regardless of the camera and lens. The sensitivity to displacement is determined by how many pixels correspond to one millimeter of the object, and by the boundary that the model has been trained on. This is calculated in two lines:

The number 5 px represents the lower limit of the synthetic shift in the recommended settings for anomalies (with a standard deviation of 0–2 px). The default settings are based on 11 px/mm: at this scale, 0.5 mm corresponds to 5–6 px, which is the target value that was agreed upon during development. Please enter your own values:

1.4. Quick Start: Settings pre-configured, zones and model ready in a few clicks

This method is designed to allow a technician with no prior experience in machine learning to obtain a working model in a single shift. Everything that could be pre-configured is already configured and stored in the program's database:

• Zones per minute. In the ROI editor, drag a frame around the element and specify the zone size (e.g., 256 px). If there are many elements arranged in a grid, drag the frame across the entire layout – the program will automatically calculate the rows and columns of the grid.
• Synthetic data for normal and defective scenarios is pre-configured. The "Default," "Element Offset," and "Color and Texture" buttons switch between pre-defined profiles. The number of files, camera frames, and expected time are displayed before the "Capture" button is pressed.
• The training parameters are set automatically. After the initial scan, the program sets the Image Size to 128, Model Size to s, Epochs to 50, and Batch to 64, and disables augmentations during training that conflict with the anomaly classes.
• The camera is adjusted using a button. The "Auto-setting for AI (per frame)" function automatically selects the shutter speed, amplification, gamma, and white balance; for USB cameras, there are "Auto-selection for forms" and "Set parameters" options.
• Camera settings packs. The selected settings are saved under a specific name and can be switched between from a list. The camera immediately accepts and stores these settings in its memory.
• One ONNX file as output. The trained model is automatically exported and immediately assigned for inspection; no further steps are required.

The entire process: connect the camera → create a model → mark zones → define the alignment region → take "Zone Images" on multiple functional boards → identify "Zone Anomalies" → TRAIN AI MODEL → Inspect tab. The remaining steps involve assigning sound, outputs, and TCP connections to the zones. To get started, you will need a computer with Windows 10/11 (a NVIDIA graphics card is recommended for faster inspection), a camera, one functional board, and one internet connection for installing the training environment. Defective boards are not required.

2. Program Interface in One Minute

The main window is divided into two parts. On the left is a live video feed from the cameras and a status bar above it. On the right is a panel with tabs, where all the controls are located. The user can drag the boundary between the two sections; the program remembers the size, position, and layout of all windows in its database.

Buttons above the video

Tabs on the right

Status bar

The status of the active camera is displayed above the buttons: `LIVE (Android): 3264×2448 [HD USB Camera] – type, resolution, and name; `AI: WAITING or `AI: GPU (CUDA) – the model's assessment and the number of milliseconds taken for the last analysis; `Net: total … active … Mbit/s – network load from GigE cameras (green up to 40% of the link, orange up to 70%, and red thereafter). The `Exp Load bar shows what proportion of the frame interval is occupied by the exposure time: if it reaches the right edge, the camera will not maintain the specified frame rate.

Each camera's tile displays a header with the camera's name, status (LIVE | AI:ON), and a reconnect button. Right-clicking on a frame opens the zone menu, while right-clicking and dragging allows you to pan across the zoomed-in frame.

3. Cameras: connection and configuration

3.1. Supported cameras

There is no hard limit on the number of cameras: tiles can be arranged in a grid with any number; eight and nine cameras worked simultaneously on real setups. Each camera has its own model, its own zones, its own threshold, and its own analysis speed.

3.2. CAMERA WINDOW: Add, reconnect, delete

The checkboxes below apply to the selected camera: Camera enabled (otherwise, it will be disconnected from the device and disappear from the screen), Show video/zones (this only controls the display; analysis always runs), AI enabled for the selected camera. The Reconnect button performs a complete disconnection and reconnection, restarting the AI stream and reloading the models.

The right-hand side of the window displays all the camera parameters read from the hardware upon connection, grouped and with units of measurement: "Show parameters for," Re-read from camera, Copy to clipboard. This is only for viewing; you can change the parameters on the "Camera" tab.

3.3. Camera Tab for HIKROBOT and SICK

The tab controls the camera, with the camera tile highlighted by a blue border. When the program is connected, it does not write any data to the camera; instead, the camera serves as the source of truth, and the sliders adjust to match its values. Settings can only be saved intentionally by pressing the "Save settings" button.

"Hikrobot: Camera Parameters (GenICam)" block

Only the data captured by the camera is displayed. The following settings are available: Auto Exposure (Off / Once / Continuous), Auto Gain, Gamma and Gamma Mode, Sharpness, Hue, Saturation, Auto White Balance, White Balance by Channel, Digital Shift, Black Level Enabled, FPS Limit Enabled, Horizontal and Vertical Flip, GigE: Packet Size, GigE: Inter-Packet Delay, GigE: Bandwidth Limit, Test Image. The "Read Parameters from Camera" button updates the table.

"IMAGE COLOR" block

Bayer pattern (Auto / BGGR / RGGB / GRBG / GBRG), Saturation 0…200 %, Gamma 0.30…3.00, Contrast 50…200 %, reset button. Software processing is applied to all frame consumers: display, recording, inference, and training images, so the image is the same during training and inspection. For HIKROBOT, the pattern is taken from the camera pixel format, while for color SICK sensors, it is taken from the sensor and refined based on the first frame.

3.4. Camera Tab for USB/Webcam

For USB cameras, the tab displays the "USB Camera: Resolution, FPS, and Properties" block. Properties of the driver that the camera does not provide are not displayed; the field boundaries are set by the program because DirectShow does not report ranges.

3.5. Saved Camera Settings (Sets)

The "Saved Camera Settings" block is displayed for USB, SICK, and HIKROBOT cameras. By selecting the parameters for the product, save them under a name: for another board or different lighting, simply select a set from the list.

• Save as… – Reads the current values from the camera and prompts for a set name. If the name is already in use, it offers to overwrite it.
• Selecting a set from the list immediately writes the parameters to the camera and saves them: for USB, to the database; for SICK and HIKROBOT, to a parameter file and the camera's memory (UserSet1 / EEPROM). Then, the panel reads the values from the hardware and displays how many parameters have been accepted and how many have been deviated from.
• Overwrite overwrites the selected set with the current values, Rename and Delete only affect the database entry; the camera settings remain unchanged.

The kit includes: for USB – resolution, FPS, driver properties, and auto-modes; for SICK – gain, exposure, FPS, black level, pixel clock, gain boost, white balance R/G/B, hardware gamma, and software color; for HIKROBOT – gain, exposure, FPS, black level, and all GenICam nodes. The frame size of industrial cameras is not included in the kit; changing it restarts the capture. The kits are the same for all cameras of the same type, so one kit is placed on each camera on the stand.

3.6. GigE Network and Reconnection

• The HIKROBOT GigE packet size is automatically set to optimal when opened; it can also be manually adjusted in the GenICam block. Jumbo frames (8164–9000) reduce the load; when using fragmented frames, set 1500 and increase the inter-packet delay.
• The `Net` line calculates the load from each camera based on its own data (useful load × frequency) and compares it to the actual incoming network traffic. A significant discrepancy indicates dropped frames or extraneous traffic. Virtual adapters and NDIS filters (GigEVision, uEye, Npcap) are excluded from the calculation.
• If the camera is visible but the stream isn't working, click "Reconnect" (in the CAMERA window or on the tile). The most common causes are network bandwidth limitations or USB bandwidth issues.
• The SICK picoCam loses its settings when the descriptor is closed; if there is no configuration file in the camera itself, the program applies the configuration file and prompts the user to click "Save Settings" to save them to the camera.

4. ZONE-PAIR Model: From Grid to Trained Model

4.1. How the method works: One model, one class per zone

The ZONE-PAIR model is a standard YOLO11 classification network, with a number of classes equal to the number of zones, plus optional anomaly classes. A class consists of a folder within the `train` and `val` directories. The program automatically creates all folders; no manual creation is required. Here is an example of a model's folder structure after training:

The files named `real_…jpg` are actual images captured by the camera (and are included in both the training and validation sets). The files named `syn_…jpg` represent synthetic data. The files named `edit_…jpg` are images taken from the zone editor. In the example provided, after capturing images of several boards: `zone_001` contains 502 files in the training set and 111 in the validation set; `zone_001_anomaly` contains 680 files in the training set and 144 in the validation set; `zone_002` contains 813 files in the training set and 181 in the validation set; `zone_002_anomaly` contains 340 files in the training set and 60 in the validation set.

The program automatically recognizes the model as a "payment" based on the presence of the zone_NNN class. The class number is the zone number within the grid (1…N). Therefore, the main rule of the method is:

4.2. Create the model

4.3. Mark zones on the board

Red zones: Exclude a zone from synthetic data with a single click

On the "AI Model" tab, click on the already created area directly on the video with the left mouse button: the area will turn red and will no longer be part of the zone-based and anomaly detection. A second click will restore it. The file counters in the synthetic data windows are recalculated immediately, and the "off N" label above the frame indicates how many areas are turned off. The list of red areas is stored per camera and model; on the "Inspect" tab, the same area remains normal (with its own yellow areas – section 5.5).

4.4. Fix reference: alignment region

The inspection zones are defined based on the position of a single circuit board. To allow for the insertion of another board with a shift and rotation, the software remembers a reference point: a live frame and an alignment region – a section of the board with features that are used to locate the product in each new frame. Once the area is found, the software transfers the inspection zones to the new board. This is how the image acquisition, synthetic data generation, and inspection processes work.

The "reference with zones" button opens the saved reference with a grid and zone numbers (section 4.8). The found component is checked for reasonableness: the area should be between 0.5 and 2 times the area of the reference, it should be a convex quadrilateral, and the support points should occupy at least a quarter of the area along each axis – otherwise, the result is discarded, and the message "component not found" is displayed in the console.

4.5. Zone-based image: normal and its synthetic data

The "ZONE-PAIR" button: creating a zone-based image involves cutting the live image into zones, placing each zone's excerpt into its corresponding class zone_NNN, and adding synthetic data representing normal conditions. Synthetic data representing normal conditions is not a defect; it reflects variations in the image, such as a slightly different placement of the board, slightly different lighting, and sensor noise.

4.6. Synthetic data defect: element shifts with the base

Two clicks:

No further configuration is required: you don't need to draw defects or search for faulty boards. Next, TRAIN AI MODEL.

Before and after: what the program achieves

Folders representing anomalies as seen by the user

Compare with the "normal" folders from section 4.5: the only difference is the position of the detail relative to the base.

Detail marking: why is the "thick" line and how is it configured?

The parts have markings – labels and 1–2 px thick lines within the frame. These markings vary from batch to batch: different fonts, labels on capacitors are rotated, and some symbols are erased. If a synthetic defect, such as a "scratch," is drawn with the same thin line, the system will not be able to distinguish it from the marking line and sensor noise, and will begin to reject boards from a different batch.

The synthetic data ZONE-PAIR addresses this from both sides. First, the main defect – the displacement of the element – moves the marking along with the body, so the marking does not become a distinguishing feature: it looks the same in both normal and anomalous states. Second, the default line/scratch is thick (⅛ of the zone, no less than 3 px) and is clearly not a marking stroke. The thickness is now explicitly defined: the Line thickness, px (0 = auto) field applies to both the line and the cross.

Hand-drawn defects (rectangles, circles, lines, "missing element," blurred spots, and others) are grouped together in a collapsible section Synthetic defects on top of the cutout. By default, it is collapsed and disabled: the main defect on the board is misalignment. Enable it for specific tasks: flooded area, extra drop, worn-off marking.

4.7. Zone Anomalies: Settings Window

The `ZONE-PAIR` button: "Anomaly Zones" creates classes `zone_NNN_anomaly` from normal cuts. All zones are selected, except for the red (disabled) zones (created by clicking on the AI Model frame in section 4.3). Anomaly classes are not needed everywhere: add them to zones where defects actually occur, and the model will check the remaining zones for recognition.

4.8. Reference with zones: solder pads and zone editor

The "reference with zones" button opens a saved reference image with a grid. Synthetic data is created here without a camera: the zoom function, the right button for panorama, and the left button for selecting zones (Ctrl to add, drag with the mouse to define the area, Ctrl+A to select all), and double-clicking opens the zone editor.

• Solder pads for selected areas… – "Solder pads: short circuit between tracks". The program automatically identifies tracks within the area (Ozu threshold, light-colored strips ≥ 2 px), placing a solder pad in the gap between adjacent tracks, using the color of the tracks, with a dark border and a highlight. Settings: between vertical or horizontal tracks, pads for images 1…3, size 8…18% of the area, images per area (5), where to place – in an anomaly or in normal, surface lighting, in val – a new image with different lighting or an exact copy.
• Zone Editor: Tools: Brush, Pipette, Color Spot, Copy Area, Paste, Stamp, Area for pads; palette: "solder / pad", "board", "shadow / mask", white, black; "Copy" field and Create Image button. This allows you to draw a defect that was not present on the line: a filled pad, an extra element, or a missing label.

The editor's files are named edit_… and are categorized into either the "zone" class or its "anomaly" class. On the "Inspect" tab, the same window opens, but it is only for viewing purposes.

4.9. Model training

The "Training Settings" section on the AI Model tab is automatically populated after the initial scan based on the specified zones: Image Size 128, Model Size s, Epochs 50, Batch 64. These values correspond to the 108–137 px zones. For the 256 px zones, as in the example, set Image Size to 192 or 256 – this will minimize compression (the Epochs 100 and Image Size 192 settings are shown in the main window). Verify the values and click "TRAIN AI MODEL".

5. Inspection

All the work is done on the "Inspect" tab. The model has already been trained; now it's necessary to check the regions, set the threshold, select the speed source, and, if the board arrives in any orientation, enable the "live conveyor" feature.

5.1. Assign the model and check the zones

5.2. Zone recognition threshold

For each zone, the network provides probabilities for all classes, but the decision is made only within each pair. The probability of the "normal" class, q, is calculated as p(zone_N) / (p(zone_N) + p(zone_N_anomaly)) and is then multiplied by a "confidence" factor, which represents how likely the pair is to be considered a valid output by the network (full confidence is achieved at 10% or higher). The resulting score, ranging from 0 to 100%, is compared to a threshold: a score above the threshold indicates a "normal" state, while a score below the threshold indicates an anomaly.

The default threshold is 50%, and it is stored in `board.json` alongside the model and applied immediately without requiring a restart. Below 50%, the threshold shifts towards normal more strongly: 25% → a cutoff of 6%, 10% → 0.4%, 1% → almost all green; 99% only allows through confident areas. This scale is needed so that the operator can distinguish between 5% and 99%, rather than seeing only areas in red.

5.3. Live pipeline: payment in any position

Halo Live Conveyor: Finding the PCB in Any Position includes searching for the PCB based on its alignment region before each analysis: the frame is aligned with the reference image, the cutting zones are extracted from the aligned image, and the alignment region and zones (green – normal, red – anomaly) are drawn on the video. The field Period, ms specifies how often to search for the PCB (default: 1000, 2000 in the image). The indicator next to it indicates: "normal · N points", "ANOMALY · N points", or "PCB not found" or "No reference – confirm". The verdict is based on a single analysis, without accumulation.

The reference and alignment region are fixed using the same button Fix reference, as when taking a picture (section 4.4); reference with zones opens it for viewing. The boundaries of the alignment region and the blue reference points, which align the board, are drawn on the video. Areas that extend beyond the frame after alignment are marked with the SKIP label.

5.4. Analyze based on which signal

The source of the speed is always the same: when a new one is included, the program removes the previous one.

5.5. Yellow Zones: Turn off zone by clicking

In some cases, certain zones may need to be temporarily removed from control during the process. This might be because a particular element is not yet installed in this batch, a zone is blocked by a fixture, or the system is undergoing debugging. To do this, you don't need to modify the grid or model; instead, the zones are simply disabled directly in the video feed, specifically within the "Inspect" tab.

The actions of the yellow zones are determined by the main checkbox: "Yellow (muted by clicking) zones are silent" in the "Audio" tab of the OUTPUT window (enabled by default):

5.6. What you can see on the screen and in the console

• The result for each zone is displayed; the border is colored according to the zone's status (red for active, green for normal). When there are more than 50 zones, only the normal status is displayed in green, while anomalies and uncertain zones are indicated with labels. Hundreds of labels on each frame made it difficult to read the image and increased the rendering load. Yellow zones are disabled by clicking (section 5.5).
• The "Trigger Analysis" panel at the bottom of the tab displays each trigger event as a row: `[hh:mm:ss.mmm] Door Camera + TRIGGER 4ms | AI 12ms #57`, along with the result for each zone. These rows are not written to the database.
• Right-clicking on a region in the video: "Classify this region as '...' → Rename the model's response for this region (this does not modify the model itself, but the rejector and log will see the new name); Also, "Create a rejector for this region" and the rule counter.
• Check the "Save inspection images" box. The folder: images with marked areas are saved as files with the extension inspect_….jpg. The settings for this feature vary for each camera.
• Every 10 minutes, a frame from the live conveyor and a text list of zones are written to a subfolder named `diag`. The last 20 images are also stored there, for reviewing disputed cases without expanding the folder.

6. Triggers: External sensor on the camera's input

Main mode for the conveyor: the camera captures a frame based on an electrical signal from the sensor's discrete input. This is the most precise moment of capture: the product is always in the same position in the frame, and the analysis is performed only once per product. The TRIGGER WINDOW collects the "sensor → camera" connection and draws a wiring diagram for the selected connector.

Identifying and labeling the connectors

7. Outputs: rejector, camera discrete outputs, and TCP

The rejector converts the inspection result into an action. The rule is created for a specific area (or for all areas of the model, class, or marked areas) and is triggered when the model detects the specified class in that area. The actions are three: sending a TCP packet to the robot or PLC, switching the camera's digital output, and issuing an audible alert to the operator. Each trigger is recorded in the log.

7.1. OUTPUT REFERENCE: Contacts and testing with a multimeter

The "EXIT" button on the video opens a help window that does not configure anything: it shows which contacts are used for the output, how to test it without power, and how to connect a load.

• Output — "dry contact" of the photo-relay: it does not provide its own voltage, only closes the circuit. An external power supply is required.
• Testing with a multimeter in Ω mode: OUTPUT OFF — open circuit (OL), OUTPUT ON — hundreds of Ohms. The buzzer is silent in both cases, see the numbers. If nothing changes, reverse the probes; the photo-relay is polarized.
• Load: +24V → lamp → pin "+", pin "-" → GND. For HIKROBOT, the limit is 25mA – any signal exceeding 27mA will cause problems; reduce the power supply or use an intermediate relay. For SICK, the power consumption is high, typically 20-30mA per lamp, which can be connected directly.

7.2. OUTPUT Window: Zone Rule

The "OUTPUT" button on the right panel (or right-clicking on the area → "Create a rejector for this area") opens the "OUTPUT – TCP packets and camera discrete outputs" window. On the left, there is a tree structure "Cameras → AI models → classes → zones" with checkboxes on the zones; on the right, there are tabs for "Zone rules", "Trigger log", and "Database".

One zone can contain multiple TCP rules (different receivers), no more than one discrete output rule per class, and no more than one audio stream per class. The rule is stored in the database and operates from the inference stream: rules are stored in memory, a search for the zone's address is performed per frame, and the camera's sockets and SDK are handled by a separate dispatcher thread to prevent the analysis from waiting for the network.

7.3. TCP Packet to Robot or PLC

The program functions as a TCP client: it connects to the specified IP address and port of the receiver, maintains the connection, and reuses it. If the connection is lost, it attempts to reconnect and resend the data once. The data is transmitted in UTF-8 encoding and includes a carriage return and line feed character at the end.

Example of a structured template: zoneId={zoneId};zone={zone};class={class};conf={conf};camera={camera};model={model};time={time};date={date};data={data}. Numbers are always in a consistent format to prevent parsing errors in PLCs.

• `Check connection` only opens the socket and reports either "connection established" or "no response after N ms".
• Send Test actually sends a string; the result is displayed next to the button, not within the window.
• Confirmation. Without the "repeat" checkbox, the rule sends one message upon class change and repeats it until the recipient responds with the line ok <message>. With the checkbox, the message is repeated at regular intervals while the class remains active.
• "Last command wins": If the model has generated multiple results while the operator is busy, the last one will be executed. The queue is limited to 512 tasks.

7.4. Discrete Camera Output

The buttons Enable Output, Disable Output, and Pulse are checked using a tester or lamp; What the camera can do queries all methods and prints a report. Manual tests are written in the same log.

7.5. Audio Alert: Custom sound for each zone

The third type of rule involves audio alerts from the computer running the software. The operator doesn't need to look at the screen; instead, they hear an alert when a zone enters a defective state, and different sounds in different zones indicate the specific location. This audio alert is available for any cameras, including USB and RTSP cameras, which do not have discrete outputs.

7.6. Trigger Log and Database

The "Operation Log" tab displays the most recent 300 entries in rows of the format "HH:MM:SS.mmm ✔/✖ Camera · Model · Class · Zone → Action [message]", and it is updated in real-time. The "Refresh" and "Clear Log" buttons.

Tab DatabaseThe program stores all data in a SQLite file rejectors.db Next to the executable file, there is no separate service running, and the port is not active. No installation is required. Tables: "settings" (settings), "cameras" (camera profiles) camera_models (models on cameras), zones (zone names and geometry), rejectors (rules) reject_log (journal) camera_param_files (copies of camera settings), camera_presets (camera settings configurations). Save a copy of the database... It creates a comprehensive snapshot of the VACUUM INTO process directly on the active database. Load the database from a copy... It restores the database, but first saves the current version. You can view the database in DBeaver without closing the program; however, you should avoid modifying it from outside the program.

8. Other Modes and Tasks

Standard models: normal / anomaly and any classes

The same ROI editor also works for standard classification: pills, caps, labels, and parts in the tray. The classes are defined in the "New AI Model" dialogue (default: normal and anomaly, with Latin characters), and the zones are assigned to the items. The "Take a picture" button with numbers 1…50, the "Create images" series, either from the camera (frame/shot) or from the encoder, and the "Synthetic data for AI" panel: precise rotations of 90°/180°, a full 360° rotation with a step and range, a mirror, "classes by degrees" (each angle in its own class – the rejector sends the angle via TCP), and brightness and contrast ranges. Training is performed with full Ultralytics augmentations. The confidence threshold for these models is set in the same place as for ZONE-PAIR – on the Inspect tab.

Multi-model inspection (slots 1–20)

With a single camera, you can manage up to 20 models: Add ONNX to the slot, then ROI → Add a class to the selected model – each model in the slot has its own zone labeling, and the results from all models are displayed on a single frame.

When multi-model inspection is actually used:

The threshold, analysis speed, yellow zones, and rejector rules apply to all camera models. The "REMOVE MODEL" button clears and resets the slots, so it asks for confirmation; "Remove selected" removes one model from the slot, but its files and zones remain in the database.

Common models

The "GENERAL MODELS" button saves a set of "camera → model" entries under the name: "Create from current models", "Apply to all cameras", "Update from current", "Rename", "Delete set". Switching the stand from eight cameras to another product is a single selection. Deleting the set only removes the grouping: folders, ONNX, zones, and rules remain. The zones belong to the pair "camera + model": one model on eight cameras results in eight independent annotations.

Encoder and rejector on the conveyor

For lines with an encoder on the COM port (115200, ESP32): connect CONNECT ENCODER, auto-connect, position, image capture step, and correction, Set zero + CAPTURE. Capture for training and analysis can be done "via the encoder" - the frames are evenly spaced along the product's path. The old rejector, with a time delay and anti-spam function, sends the REJECT command to the encoder controller.

Video and image recording

Settings tab: START VIDEO RECORDING from the active camera to a file named record_date_time.avi (MJPG, actual camera FPS). Training images: real_….jpg, inspection images: inspect_….jpg with marked areas.

9. Settings, console, requirements, and license

Settings

• SAVE TO FILE / LOAD FROM FILE – all settings (camera, models, rules) in JSON format, which is convenient before reinstalling or transferring to a second computer. The source of truth is the database, and the JSON file serves as a backup in case the database becomes unavailable. All changes to the settings are automatically saved.
• RESET TO DEFAULT – This action is irreversible: it removes the configuration and window layout.
• Interface language: 14 languages; the main window switches immediately, and the other windows switch upon the next opening.
• Position, size, and monitor of all windows, and the position of dividers – are stored in the database; when the monitor is turned off, the window returns to the visible area.

Console

The "Camera" and "Trigger" checkboxes are disabled by default, and the corresponding messages are displayed (in trigger mode, camera messages go to the "Trigger" section); all data is written to the logs\engiai-*.log files before any filtering is applied. The console is designed to operate for weeks without intervention: up to 800 lines are displayed on the screen, with 1000 old lines being removed; new lines are collected in batches of 4 every second, rather than one at a time; downward scrolling is enabled until you manually scroll upwards; the console has no sound. Changes to zone classes are written no more than 10 lines per second (the rest is stored in a counter), the log file is 20 MB in size, and files older than 14 days are deleted upon startup; the log of events in the database stores the last 20,000 entries. Unprocessed errors are located in crash.log, and the reasons for forced termination are located in runtime.log next to the exe file. If something is not working correctly, check this section first.

Requirements

License and Trial Mode

Upon startup, the program connects to the engi.live license server. Only hashes of the hardware identifiers and version numbers are sent, without serial numbers or user names. If the server is unavailable, but a license has been cached, the program operates using the cached license (for offline environments). Without a valid license, a 20-minute trial mode is activated: a "TEST MM:SS" countdown appears in the window title bar and in the AI status bar. The "BUY" button on the Settings tab opens the purchase page. When a valid license is active, the countdown is not displayed.

10. Quick Checklist and Fault Finding

Zero-based checklist for the board

If something goes wrong:

engiAI Ethernet System. The names of buttons and fields are as they appear in the program on your language (screenshots are from the Russian interface). The screenshots were taken on the "model_4" model (two 256x256 zones: capacitor C30 and bridge BR1), a 3264x2448 USB camera, and the model was trained in grayscale. Version 3, September 2026: example with two zones, creating model folders, alignment region, synthetic data "element with base" with "before / after" images, color palette and line thickness, red zones, console.