Adaptive learning technology makes intelligent decisions regarding component, lead and solder joint integrity, and automatically improves and refines models for each region on a printed wiring assembly (PWA).
A key benefit of adaptive learning technology is that it eliminates the requirement that operators be trained in image processing techniques, complex rule-based algorithms and synthetic modelling. Instead, inspection programs are created by scanning example printed wiring assemblies (PWAs). The system itself is trained using multiple training boards and then uses complex statistical modelling to make future judgements. The ability of AIMS to learn and adapt as the process evolves is in striking contrast to traditional template or static pattern models that become ineffective every time the process changes. The AIMS system learns as a process drifts without degrading the inspection capabilities.
The functions of the AIMS automatic optical inspection (AOI) system from US company Photon Dynamics, are shown in Figure 1. The adaptive learning software automatically identifies and classifies component, lead and solder joint defects including critical defects such as insufficient solder and lifted leads that may escape other inspection systems. The system learns to discern acceptable or unacceptable process variations by inspecting live PWAs and automatically adapts to these variations. The operator does not have to reprogram the system every time there is a new variation.
Military background
The underlying algorithms that are the basis of the AIMS adaptive learning technology were initially developed for military smart vision systems under contract with the US Department of Defense. Military applications tend to be much more challenging than those found in the commercial world. The targets in missile and aircraft tracking systems move very fast in a noisy, totally uncontrollable environment. Even though the initial AIMS software ran on specialised hardware with custom analog-digital processors, the company found that its approach to solving vision problems and the algorithms it developed could have commercial applications when running on off-the-shelf computers.
No real world inspection problem is static. Seemingly static applications such as inspecting circuit boards are dynamic in that they are going to have many, many different forms of acceptable and unacceptable. As the process drifts, the appearance of what is acceptable and unacceptable will change over time. Some attributes drift over days, weeks or months sometimes simply because of the age of the equipment. Even seemingly innocent changes to the process such as the calibration of a pick-and-place machine or the cleaning of a paste stencil may make the boards appear to be different.
AIMS technology not only automatically updates its knowledge of the process, it also automatically updates the boundaries of what is acceptable and what is unacceptable and will distinguish between the two.
The ability of AIMS to learn and adapt as the process evolves is in striking contrast to traditional template or static pattern models that become ineffective every time the process changes. As a result, traditional approaches require highly skilled operators who routinely edit or completely scrap models and build new ones to accommodate changes. In effect, traditional approaches use the continual intervention of a specialised operator as the adaptation engine.
Algorithms turn raw images into refined knowledge-bases
AIMS technology is based on a wide array of algorithms including a family of filter algorithms that are sensitive to different geometries. The filters can recognise a change in a region of interest (ROI) structure and then accentuate that change. For example, the system is capable of detecting shifts in local grey-levels, disruptions to lines and edges, and changes in local topology. Each filter is applied only once to the ROI, which is in contrast with conventional filtering in which spatial filters are passed back and forth over the image, stepping a pixel at a time.
During basic training, the correct set of filters is automatically determined for each ROI, and the 'scores' for each filter obtained from each example form the ROI knowledge-base. The ROI knowledge-base is basically a set of typical filter scores for each ROI. At any time during operation, additional scores can be added to the knowledge-base.
There are two primary ideas behind AIMS. First, no matter how you represent the image that you are storing, if you use a single static example, it is going to be very difficult for it to capture all the forms of what is acceptable and all the forms of what is unacceptable.
Many times when you correlate a template image or use simple features from a fixed board, some variations that you see can deviate from that example a certain distance, and be considered good, while others that deviate that same distance or even less, are unacceptable. The more data points, the better the description or model. With a single data point, nothing is known about the typical values a feature variable might take over time. With multiple data points, a better description of the feature range is formed, and a statistical model may be constructed. Second, a rich feature set is critical for complex, diverse applications. No single feature can segment all of the above patterns as defective. Additionally, patterns can be created that these features cannot segment adequately while maintaining sensitivity. An additional problem that can arise is that the best features for one ROI may be the worst features for another.
Extrapolates from known features
The AIMS system does not have to see every example in the world to decide if a PWA is acceptable or unacceptable. Instead the technology can make inferences and extrapolate from known features to new ones. This extrapolation does not require intervention by the operator. AIMS develops a flavour of what is good, based on numerous samples. The system generates features for each component site and toe fillet, for example. It may generate features about an entire region of interest, or it may adaptively break that block into smaller pieces and generate features inside of those. Either way, a string of descriptors is used to describe each ROI, for each of the training samples.
The system was designed for standard line operators and does not require sophisticated knowledge of image processing or data analysis. During system programming the graphical user interface (GUI) interacts with the database and presents examples to the operator that allow him to make decisions about what is acceptable and unacceptable. In effect, the operator serves as the system's mentor. The system will display the current image for a region of interest and then show the operator images from its knowledge-base of the parts that it thinks are the best match. The operator can then select the correct match from multiple choices or define a new class. Research has shown that operators are very good at multiple choice problems versus attaching an acceptable or unacceptable label to a single example. If you show operators the same defects over time, they will label them differently if they do not have a reference. However, if you show operators the closest example of what is acceptable, it is easy for them to make a consistent decision.
Supports both optical and X-ray
AIMS is applicable to both vision and X-ray. Many competing systems use image-based algorithms that rely on the physical state of the region of interest. This approach has problems. For example, if you look at a pre-reflow image using visible light optics, the paste is dull and diffuse and the leads are shiny and spectral. Once the paste reflows, it becomes very difficult to see the lead edge because it looks like chrome in sunlight and the light scatters everywhere. This means that in pre-reflow, you can leverage an algorithm that uses the leads to reliably find where a part is placed. If that same algorithm is used post-reflow to position a part, it will fail because the solder is irregular and the algorithm is depending on shiny spots on leads to place the part. AIMS works anywhere in the PWA line: paste, pre-reflow, post reflow, and post wave solder, and operates the same independent of line position.
The algorithms used are also independent of the inspection mode, eg, vision or X-ray. The AIMS creators realised that the software would have to be compatible with optical-only, X-ray only and X-ray/vision combination systems. Moreover, the software architecture and processing design are not locked to a type of inspection imagery or slot in the line. Everything is programmed to be general purpose. For example, even though the algorithms were not specifically designed to do paste, they have been proven to find bad paste examples. This is because, with training, the presence/absence algorithm will identify an acceptable paste and an unacceptable paste. The flexibility of this new, adaptable, knowledge-based software makes it compatible with a variety of sensors as well as X-ray and optical, including laser and IR. AIMS knowledge-bases are developed by loading PWA CAD data and by scanning production boards without template modifications, time-consuming software adjustments and algorithm revisions. The system learns as it works. When a questionable image is analysed, AIMS alerts the operator so he can identify a condition as acceptable or as a defect. This information is automatically incorporated into the knowledge-base and becomes part of the inspection criteria from that point forward. The learning ability of AIMS will eventually eliminate false alarms by refining the knowledge-base so that future alerts for that condition no longer occur.
AIMS technology is capable of detecting defects at all process levels. The system will learn to detect component level defects such as missing or wrong parts, wrong polarity, damaged, and critical skew (X, Y, O displacement). It also detects lead-level defects such as bridging, missing, bent, lifted, and off pad. Difficult to detect solder-joint level defects such as insufficient, fractured, excess, and poor wetting are also easily found. The modular design is scaleable in functionality, speed, and price. With the adaptive, self-organising capability of the software, the AIMS system can be placed anywhere in the line to address priority process problems.
| Tel: | +27 11 609 1244 |
| Email: | [email protected] |
| www: | www.zetech.co.za |
| Articles: | More information and articles about ZETECH ONE |
© Technews Publishing (Pty) Ltd | All Rights Reserved
printer friendly version