
Summary: AI vision and smart factories were among the hottest topics at the show. At the programming station, though, the questions are more concrete: how to feed devices reliably, position them accurately, program the right version and keep the data. The value of an automated programming system isn't shaving a few seconds off each programming cycle. It is linking these steps into one manageable process.
At this year's NEPCON ASIA, AI and automation were among the key themes, with a dedicated forum on integrating smart factories and AI-driven automation. VISION China (Shenzhen), held alongside the show, featured 3D cameras, intelligent vision equipment, AI vision software and vision inspection solutions. Inspection equipment vendors also introduced process optimization solutions that combine measurement data with AI.
Beyond the buzz, one question deserves a closer look: what do these technologies actually have to do with an automated programming system on the production line?
The answer is simpler than it sounds. The vision an automated programming system needs is, first of all, about seeing accurately. Is the device there? Is it oriented correctly? How far is it offset? The intelligence it needs is, first of all, about control. Which program? Which lot? Which device failed, and why? Getting these two things right is worth more than any AI label.
Machine vision growth offers a useful window into manufacturing automation. According to the 2026 China Machine Vision Report from the China Machine Vision Union (CMVU), China's machine vision market is expected to grow from RMB 50.30 billion to RMB 78.40 billion between 2026 and 2028, a CAGR of 24.8%. In 2025, manufacturing accounted for 78.0% of machine vision sales. Consumer electronics was the largest application, with a 30% share, and semiconductors (including PCBs) accounted for 12.2% of manufacturing applications.
The same report highlights a shift: machine vision is moving from single-point inspection toward system-level solutions. Electronics production lines are moving the same way. No matter how good a single machine is, its value drops if it can't connect with the processes before and after it.
On electronics production lines, automation upgrades center on five areas:
Material handling: Feeding, pick-and-place, material changeover and buffering, reducing manual handling and wait times.
Machine vision: Positioning guidance, orientation recognition and code reading, so equipment can "see" the material.
Automated inspection: Appearance, marking and dimensional checks that catch defects at the station.
Data collection: Equipment status, job results and exception records, moving from "reading the screen" to "keeping records."
MES management: Work order dispatch, result reporting and lot association, bringing machine-level data into factory systems.
The machine vision seminars held alongside this year's show also focused on two fundamental applications: vision inspection and positioning guidance. Automating the programming station involves the same five areas.
An offline automated programming system usually sits between incoming material from the warehouse and SMT placement. Upstream, it receives blank devices in the semiconductor vendor's original packaging. Downstream, it delivers programmed devices ready for the placement machine. Inside the system, a complete cycle looks roughly like this:
| Stage | What Happens | What to Watch |
|---|---|---|
| Input | Devices are picked from trays, tape or tubes | Feeding method matches incoming packaging; material changes without stopping |
| Recognition and positioning | Device position and orientation are confirmed, and the nozzle is guided to place it in the socket | Positioning accuracy, orientation recognition, placement stability |
| Programming and verification | Program and Verify are completed in the socket | Correct program version; failures detected and diverted |
| Inspection and marking | Appearance inspection, laser marking or code reading as needed | Programmed devices distinguishable; markings readable |
| Output and data logging | Good devices taped or returned to trays, rejects diverted, results recorded | Output format meets the next step's needs; data traceable |
Each stage has details that are easy to underestimate:
Input: It looks like simple material handling, but it determines whether the system can run continuously. How quickly the system detects, alarms on and handles empty trays, material shortages and jams matters more than its speed with full trays.
Recognition and positioning: With fine-pitch, small-form-factor packages, a device that is offset or rotated may cause poor contact at best. At worst, it can damage the device, the socket or even the mechanism.
Programming and verification: Program writes the data, and Verify checks the result using the configured method. Together, they confirm that the data was written correctly. This is not a product quality test, and it cannot replace the semiconductor vendor's electrical testing.
Inspection and marking: What the vision system can detect depends on the camera, lighting, lens, algorithm and pass/fail criteria. Adding a vision system does not mean every defect will be caught.
Output: Good devices, failed devices and devices pending re-inspection each need a defined destination. These rules should be built into the equipment logic, not left to manual sorting.
Not every project needs all five stages. Some need only automated loading and unloading plus programming. Others also need vision, marking and data upload. The right configuration depends on the devices, the volume and the customer's requirements.
What really needs careful design is the handoff between stages. Once one step is complete, how is the status passed to the next? Could a device that failed programming end up at the output for good devices? Before marking, how does the system confirm that the device actually passed programming? These handoff rules determine whether a system is just several stations bolted together or a truly closed-loop process.
The UPH on a datasheet is usually a maximum rating for a specific device and configuration. On the production line, actual output depends on five factors.
Programming time depends on device capacity, data size, algorithm and verification method. When comparing speeds, first clarify the test conditions. Which device? What capacity? Is Verify included? Which program file? Was it a single-device test or continuous volume production? A speed figure without its test conditions can't be mapped directly to the production line.
Every move from the input tray to the socket, and from the socket to the output, takes time. If pick-and-place and positioning can't keep up, overall output is limited, however fast the programming is.
Running multiple sockets in parallel lets the pick-and-place mechanism serve some sockets while others are programming. But the number of parallel sites doesn't multiply output directly. Output also depends on which actions can overlap, how programmer resources are allocated, and whether shared mechanisms cause waiting.
The bottleneck shifts with the device. For high-capacity devices with long programming times, the bottleneck is usually the number of sockets, and the pick-and-place mechanism has to wait. For low-capacity devices with short programming times, sockets free up quickly, and the bottleneck moves to pick-and-place speed. On the same machine, switching to a different device can produce a completely different UPH.
Changing to a different part number may require adjusting the trays, sockets, pick-and-place parameters, algorithm and program file. A high single-product UPH doesn't make a system suitable for high-mix production if changeovers are cumbersome. During selection, ask:
How easy is it to change sockets?
Is the program configuration for each part number clearly organized, and are parameter changes access-controlled?
Is revalidation required after a changeover?
Can production jobs be linked to program files and called up automatically?
Material shortages, pick-and-place failures, contact problems, programming failures and equipment alarms are hard to avoid completely in volume production. What matters is not avoiding exceptions altogether, but whether the system can:
Detect material shortages and pick-and-place failures
Distinguish programming failures from equipment faults
Divert rejects
Keep exception records
Resume safely after manual intervention
The first three factors determine how fast the system runs. The last two determine how long it keeps running, and that is exactly what UPH doesn't show.
The international standard ISO 22400-2 defines key performance indicators for manufacturing operations management. Among them, overall equipment effectiveness (OEE) is the product of three factors: availability, performance and quality. Availability is actual run time as a share of planned production time. Performance compares actual output speed with design speed. Quality is the share of good units in total output.
Applied to automated programming:
Availability is eroded by changeovers, material changes and breakdowns.
Performance is eroded by pick-and-place waiting, minor stops and running below design cycle time.
Quality is eroded by programming failures, verification failures and devices that need re-programming.
So when evaluating automated programming equipment, don't just ask about UPH. Run a continuous test with the target devices, actual program files and the expected changeover frequency, and record the three types of loss separately. Only then will the numbers match your production scheduling.
Start with the division of labor. Machine vision handles "seeing and judging." The automated programming system handles "executing the programming process." The two can work together in the same flow, but their functional boundaries differ.
Next, distinguish between two kinds of vision. Rule-based machine vision relies on algorithms such as template matching, edge extraction and dimensional measurement. It suits well-defined tasks with stable features, such as positioning, orientation detection and code reading. Deep learning–based AI vision relies on models trained with sample images. It is better at identifying defects with wide visual variation, where rules are hard to define exhaustively. The two don't replace each other, and many applications combine them.
At the programming station, vision handles four main types of tasks:
Vision positioning: Guides the nozzle to place the device accurately in the socket. Results depend on the package, surface features, lighting, camera resolution and mechanical accuracy, and must be evaluated with actual devices.
Device identification: Detects orientation and, where needed, reads surface markings or barcodes to prevent wrong-part errors. Surface markings are not electrical identification, however. They cannot confirm a device's internal part number, capacity or programming compatibility. Those must still be confirmed through the part number, the algorithm and official support documentation.
Process checks: Confirms that the device is seated correctly in the socket, with no offset or debris, and that the marking is in the right position and clearly legible.
Quality control: Combines inspection and programming results to make a pass/fail decision, diverts rejects and archives the data.
One boundary is firm: vision inspection is not functional device testing, and it cannot replace Verify after programming. Whether stored data is complete and whether electrical performance is acceptable can only be confirmed through verification and electrical testing.
The first three types of tasks are mostly well-defined, and rule-based vision is often sufficient for them. AI vision is more likely to add value in appearance defect judgment, where variation is high and rules are hard to write. Whether AI is needed should depend on the inspection target, sample quality, false reject rate, escape rate, processing speed and model maintenance cost, not on whether the AI label is present.
Whichever type of vision is used, its inspection capability must be validated with representative samples. Adding a vision system does not mean every defect will be caught.
Looking ahead, the more important question may not be whether the vision system uses AI. It is whether vision data, programming results and equipment status can be brought together. When images of failed devices can be analyzed alongside socket IDs, insertion counts and verification results, it becomes possible to spot early that a socket is wearing out or that an incoming lot has problems. This is intelligence at the data level, and it requires the data to be fully recorded first.
Completing the production steps only covers process execution. Factories with lot management and quality traceability requirements must also be able to answer: Which program file was used for a given lot of devices? On which machine were they programmed, and what were the results?
In the hierarchy model of ISA-95 (IEC 62264), the standard for enterprise-control system integration, MES sits at Level 3, manufacturing operations management. It coordinates people, equipment and materials in production, and exchanges information between business systems (Level 4) and shop-floor control (Levels 1–2). Automated programming equipment sits on the shop-floor side. Its job is to deliver execution results to Level 3 in the agreed format.
Board-level products have serial numbers, so the traceability unit is the individual board. Offline-programmed devices have no serial numbers and look identical to blank parts, so the traceability unit is usually a tray, a reel or a lot. To link the chain together, build associations around the following data:
| Data Category | Key Information |
|---|---|
| Production job | Work order number, product model, production lot, planned quantity |
| Device information | Device part number, package, semiconductor vendor lot number, incoming package ID |
| Programming information | Program file and version, algorithm version, programming parameters |
| Equipment information | Equipment ID, socket ID, production time, equipment status |
| Results | Program/Verify results, pass count, fail count, re-programming count |
| Output | Output reel or tray ID; laser-marked code where applicable |
| Exceptions | Alarm codes, failure causes, handling records |
When finer-grained traceability is needed, laser marking can write a lot or version code onto the device surface, so each individual device can be linked to its programming record. Not all of these fields need to be captured by the programming equipment alone. Some can come from the MES, barcode system or host PC. What matters is defining data sources, transfer methods and ownership in advance.
Before implementation, confirm at least these five points:
Job dispatch: Can the MES send work orders, device part numbers and program information to the equipment?
Program management: How are program files and algorithm versions managed, and how are wrong versions prevented?
Result reporting: Can Program and Verify results and quantities be uploaded as required?
Exception handling: How are programming failures, equipment alarms and communication interruptions handled? How is data buffered during a network outage and uploaded afterward?
Data association: Can results be linked to the work order, lot, output package or individual device marking?
If the equipment uploads only production quantities and can't link them to work orders and program files, the data is collected but its traceability value remains limited.
The IPC-1782 standard series addresses traceability in electronics manufacturing and the supply chain, and can serve as a reference when discussing traceability levels. How far traceability needs to go depends on product risk and customer requirements. Not every application needs traceability down to the individual device.
The HILOMAX AT3 Series is designed for automated volume programming. According to current product documentation, the main models are:
| Model | Documented Rating | Structure and Configuration |
|---|---|---|
| AT3-150C | Up to 2,000 UPH | Up to 32 sockets, 2 nozzles |
| AT3-350C | Up to 3,000 UPH | Up to 64 sockets, 8 programmers; 3D vision, laser marking and MES-related configurations |
| AT3-350M4 | Up to 2,500 UPH | Up to 128 sockets, 4 nozzles, linear motor design |
Three notes on the table:
The UPH figures are the maximum ratings stated in each model's documentation. Test conditions are not uniformly specified, so the models should not be ranked against each other, and the figures do not represent actual output on any given line. As discussed in Section 3, actual performance must be confirmed through continuous testing with the target devices.
The vision configuration in the AT3-350C documentation is 3D vision. This article does not attribute to the AT3 Series any AI vision inspection features shown by other exhibitors at the show.
All specifications and features are subject to the official specifications of the relevant models and project validation results.
From a process integration perspective, use these questions to guide selection:
For devices with long programming times and high parallelism needs, focus on socket count and programmer configuration.
For devices with short programming times where pick-and-place dominates, focus on nozzle count, pick-and-place structure and motion design.
For projects that need to distinguish programmed devices or require lot traceability, focus on laser marking and data interfaces.
For fine-pitch packages or devices with demanding positioning requirements, focus on vision positioning capability and validate it with actual devices.
For lines with frequent changeovers, focus on the steps and time required for material changes, socket changes and program switching.
Equipment specifications are only the starting point. What an automated programming system really has to deliver is feeding, positioning, programming, inspection, marking and data all running together under the target process.
Vision handles recognition, positioning and inspection. The automated programming system executes programming and related processes. The MES manages jobs, process data and traceability. Each owns its part, and only when they work together does programming truly become part of the electronics production line.
For an automated programming system, the foundation of "intelligence" is three basic things: stable material handling, reliable programming results and complete data records. Once that foundation is in place, vision can do its job as a gatekeeper and data can actually be analyzed. Skipping these steps to chase an AI label usually produces little more than a demo at a trade show booth.
Looking ahead, production lines will increasingly judge automated programming systems on these capabilities, not just rated UPH:
Compatibility with a wide range of devices and packages
Stability of material handling and automated pick-and-place
Process coordination between programming, verification and auxiliary inspection
Multi-site parallel operation and exception recovery
Production data collection and MES integration
The value of an automated programming system isn't just how much faster each programming cycle runs. It is turning material handling, device programming, result inspection and production data into one manageable process.
Planning an automated programming solution? Contact HILOMAX. Share your target devices and packages, incoming and outgoing packing formats, target cycle time, changeover frequency and traceability requirements, and we'll work with you to confirm the configuration and validation plan for your actual process.
Next, the series finale: from device programming to production automation, and how to choose the right programming solution for each scenario.
