An extra set of eyes never hurts. Electronics and packaging industries have long relied on machine vision to guide, inspect, measure, and identify objects whizzing by. Now, pharmaceutical manufacturers are catching on to the benefits of having an extra set or two of electronic “eyes” watching over packaging and “back end” fill-and-finish operations.
Machine vision system deployment has been anything but a slam dunk for pharmaceutical applications. Issues such as programming, computing, integration and control interfaces prevented the first generation of machine vision technologies from providing the accuracy and reliability that pharmaceutical manufacturing demands.
However, since that time, costs for the systems have fallen dramatically, by a factor of five. At the same time, inspection speed, field vision and overall accessibility have improved ten-fold. Pharmaceutical and medical device manufacturing demand for the systems grew by 10-20% last year to $86.7 million, or 7% of the total machine vision market.
Many machine vision vendors entered pharmaceutical markets through the back door, via medical device manufacturing. Such was the case with Retina Systems, Seymour, Conn., whose technology evolved from laser position measurement in the late 1970s.
Laser systems are the “legacy technology” of machine vision, providing simple dimensional measurement, according to Retina Systems marketing director Robert Tibbing. As imaging and control functions have become more affordable through advances in computing and semiconductor technology, applications have multiplied exponentially.
Critical components of any machine vision platform are image acquisition, optics, lighting, material handling, computing, and machine control. Hundreds of vendors sell equipment that can perform some of these functions, but only a handful can fashion components into integrated systems for manufacturing.
In North America, most pharmaceutical manufacturers and their suppliers use vision systems only when big pharma customers request it, according to Luis Cruz, director of international marketing at IC-Vision (Saint-Laurent, PQ). “The drivers over here are external, not internal, whereas in Europe the opposite is true,” he observes. Cruz attributes U.S. drug-makers’ reticence to a general avoidance of new technologies and a lower commitment to operational quality compared with overseas manufacturers.
Greater Than the Sum of its Parts
Modern machine vision is more an amalgamation of interoperational components, tied in to a specific application, than an off-the-shelf device. A typical system consists of:
- Cameras, usually color CCD devices similar to those in camcorders
- Frame grabbers, which digitize analog images and interface with bus structures
- Software and fast data transfer
Imaging functions may be further broken down into image acquisition, image processing, and feature extraction.
Machine vision performs four tasks very well: guidance, inspection, gauging, and identification, suiting them for packaging. However, with the right programming, vision systems can also pick out fill errors, detect smudges on ampoules, inspect lyophilization cakes for contaminants, and accept or reject solid materials based on average particle size. Some biotech companies have even used cameras to monitor froth and turbidity during fermentations.
But process applications lie at the fringes of Zuech’s definition because of the required human intervention. Besides, traditional measurements (temperature, pH, extent of reaction) are relatively straightforward and require no new thinking. With drug-makers just getting comfortable with conventional real-time process analytics, machine vision for process monitoring is probably a long way off, at least in North America.
For example, Japanese drug makers have embraced machine vision inspection of solid dosage forms, where their peers in the U.S. have not. Vision systems excel at for inspecting pills and tablets for defects, engravings, and overall integrity. About a half-dozen companies sell inspection equipment for this type of application to Japanese markets, but Nello Zuech says vendors can barely give their equipment away to U.S. pharmaceutical manufacturers.
Seidenader Markt (Schwaben, Germany) offers a barrier-enclosed inspection system with a throughput of 2,500 tablets per minute. The company also develops machine vision systems for ampoules, prefilled syringes, and infusion bottles plus ampoule cleaning/drying machines and capsule polishers. The InspectoRx tablet/capsule inspection system from American SensoRx (Glen Rock, N.J.) accepts or rejects solid doses based on integrity, color, shape, and presence of foreign object contamination. Other solid dose inspection vendors include Driam USA (Spartanburg, S.C.), AC Compacting (North Brunswick, N.J.), and Japan’s Kanebo. A few blister-pack equipment vendors, most notably Bosch and Uhlmann, integrate inspection systems on their machinery.
Machine vision can be deployed out of the box, built into a process, or custom-installed. Off-the-shelf apps like identification code recognition perform one task very well and do not require customization. “They’re about as close to a standard application as you can get,” notes Justin Testa, senior vice president for Identification Products at Cognex (Natick, Mass.).
Although high-end implementations tend to be custom-built, with unique lighting, integration, and object presentation, vision companies have been striving to develop more general-purpose products. “One goal is to make these deployments simpler and cheaper so users can be up and running quickly,” says Testa. Another is to provide products with the capability of vision systems and the ease of use of barcode readers.
Cognex, which sells sensors, components, and vision systems to several industries, offers In-Sight technology, which reads and verifies RSS (Reduced Space Symbology) codes on pharmaceutical packages for product identification and traceability. RSS codes, whose compactness allows them to go where no barcode has gone before, can uniquely identify unit doses, even individual tablets.
|Shown above is a printed logo inspection station sustaining a flow of 3600 parts per minute. The flow of caps has been separated on two adjacent conveyors (see below) to ease mechanical handling of the parts, but inspection is handled by a single vision processor. Photos courtesy of IC-Vision|
IC-Vision offers two systems based on a self-learning “neuro-based” image analyzing algorithm. The flagship Eye-Q system inspects caps, closures, bottles, tubes, and print quality, detecting malformations, flash (extraneous plastic at joints), missing liner, contamination, outer or inner missing cap, line defects (missing, half-moon, deep cut, folded, half-seated, off-center), defects in tamper-evident bands and printing defects. Since Eye-Q’s processing algorithm memorizes the flawless product rather than likely flaws, it recognizes all product flaws that yield visual differences with the reference group, resulting in more accurate defect recognition and reduced changeover and production downtime.
IC-Vision's Inspector system, which also uses a self-learning algorithm, memorizes up to 13 million features on a part in relation to other features. Inspector extracts these characteristics related to size, color, position, and other criteria and uses them to calculate simple decision-making criteria.
Although they’re smart enough to catch tiny perforations in blisterpaks or unnoticeable skew in labels or cartons, IC-Vision’s imaging systems cannot verify the integrity of
printed materials beyond their physical orientation.
“We can detect if the label has been cut improperly, if the print is the right color, and if all the print is present, but we can’t detect if the printer switched an âI’ with an âl,” Cruz explains. “We don’t do OCR and actually we don’t do absolute measurement, even though we detect differences much smaller than one millimeter.”
To Cruz, the most difficult aspects of creating a functioning vision system are mechanical handling and presentation of products to the camera. Issues such as what type of guide rail to use, product stability as it passes by the camera, lighting, and how it’s all put together spell the difference between success and failure. “Vision systems are only as good as their installation,” Cruz says. “Components may be great by themselves, but you’ll get mediocre results if the people installing them don’t know enough about the product and operation.”
The Eyes Have It
The greatest difficulty in implementing machine vision is defining appropriate inspection criteria. Particularly in a regulated industry like pharmaceuticals, where everything is validated, QA groups must be in alignment with what the vision system determines is acceptable or unacceptable product.
“When switching over from manual inspection to machine vision everyone must agree, and then understand, what’s satisfactory from the vision standpoint,” says Dan Freed, Business Development Manager at Xyntek (Yardley, Pa.). “Figuring out the inspection criteria is the starting point of any machine vision project,” Freed remarks. “Then comes the art of determining if you can get what you need from an image standpoint.”
Machine vision implementations are almost never as straightforward as they appear. “Some applications appear to be simple, but when you get onsite you discover that they’re difficult or impossible,” Freed observes. He notes a recent biotech project involving hundreds of roller bottles, which Xyntek was asked to monitor for cell culture health. “We didn’t think we’d be able to image color and turbidity, which can be indicators of dead cells or contaminated cultures. But we came up with an incredible lighting scheme that gave us what we needed without resorting to software tricks.”
Optical character verification (OCV), a classic machine vision application, involves matching printed symbols on a label (e.g. date, lot code) or manufacturing packaging with a standard image. OCV differs from optical character recognition (OCR) in that the machine is not reading the print but instead verifying that everything is there, in the right place, as it would for a part.
“An OCR system has no idea what it’s reading, whereas with OCV, you already know what you’re looking at,” says Dan Freed of Xyntek, which offers a turnkey, configurable OCV system it calls MVX-OCV. Like other machine vision implementations, MVX-OCV needs to be trained using, in this case, a standard alphanumeric set or library. Xyntek’s other packaging products include MVX-Blister (for tablet color, integrity, presence), MVX-PQI (labels/printing), MVX-TFI (tablets on slat filling machines), MVX-2D, and MVX-CPI.
Xyntek also serves drug discovery and laboratory automation markets through its MVX-PAS for grading biological samples by color, area, density, and height. Another biotech-oriented product, MVX-CCA, inspects and grades cell cultures based on visual properties like color, cell clumping, and light transmittance. MVX-CCA, which was the basis of Xyntek’s cell culture product, consists of a high-resolution color CCD camera, lighting, machine vision processor, and application software.
Although most machine vision vendors specialize in a narrow application area--for example, packaging or print imaging--Xyntek tackles most applications that come its way, even process monitoring. According to Freed, sensor or component vendors have an easier time specializing, whereas value-added engineering firms recognize that even within inspection classifications every application is different.
“What one manufacturer considers a quality defect may not be an issue for anyone else,” he observes. "Vision systems are only as good as their installation... Components may be great by themselves, but you’ll get mediocre results if the people installing them don’t know enough about the product and operation."