Navigating the Challenges of In-line AI Vision Systems
Learn more from Eigen Innovations …
Pranav Bhatkal
A missed defect can halt production, drive up scrap, and eat into margins fast.
That’s why manufacturers are constantly seeking ways to improve production and product quality. This may involve increasing throughput and scrapping less material, or reducing machine downtime caused by the absence of an inline inspector. Often, their first step is investing in an automated inline machine vision system.
Inline machine-vision systems not only address issues related to efficiency and throughput but also provide consistency in product quality and traceability. A centralized system for inspection results and sensor data allows for trend analysis, root-cause identification, and continuous improvement.
But deploying and scaling these systems comes with its own set of challenges. Challenges can arise from factors that complicate installation and setup, or from data issues that destabilize AI models, increasing false-positive and false-negative rates. Either way, they create doubt and frustration, leading to a lack of trust in the system.
In this article, we break down common hurdles and offer practical strategies to overcome them.
Infographic Source: Eigen Innovations
Installation and Integration related challenges: Inconsistent Lighting and Environmental Conditions
Lighting and environmental factors can dramatically impact image quality in production environments. Natural light fluctuations, such as differences between sunny and cloudy days or varying light levels at different times of the day, can cause inconsistencies in image quality.
Reflective surfaces introduce additional variability, while in thermal applications like welding or injection molding, the plant’s temperature can also affect image clarity. Other factors like dust and vibrations can further degrade image quality.
While you can’t eliminate these conditions entirely, they can be managed. Using a consistent, controlled light source positioned at the right angles and distances can minimize the impact of external light sources. Additionally, protective camera enclosures and air-purge systems can shield cameras from dust, sparks, and vibrations. Image thresholding and filtering techniques can also help compensate for temperature-related shifts. Regular maintenance checks, including cleaning, are crucial to maintaining stable system performance over time.
Consistent image captures for High-Speed Requirements
Certain applications produce goods at such high rates that capturing images consistently becomes challenging. In these scenarios, polling read rates to PLCs can introduce latency in capturing the correct frame from the camera. This can lead to false alerts, where non-defective parts are flagged. Accurately accounting for this latency and synchronizing frame capture is crucial for high-speed operations.
One solution is to bypass the PLC and use I/O-triggered cameras to ensure sharp, well-timed image capture and reduce the risk of blurry images. In multi-camera setups, synchronizing frame capture is essential to correctly link images to the corresponding parts.
Challenges with limited space for cameras and other hardware
One big advantage of vision systems over manual inspection is the ability to provide visibility in areas that would be difficult for a manual inspector to view. The challenge lies in positioning cameras to ensure full part coverage in tight or awkward spaces.
This can be addressed by using custom mounts, wide-angle lenses, or fisheye lenses. Flexible mounting systems such as adjustable arms, sliding rails, and swivel joints allow for optimal placement, while wide-angle and fisheye lenses capture a broader field of view. Incorporating these lenses can reduce the need for multiple cameras while covering the same area.
Challenges with Network and connectivity
Most plants operate continuously across multiple shifts, generating a constant flow of data. For any reliable AI model, it’s crucial to tap into this data, and the first step is ensuring a strong, dependable connection to upload data. However, Wi-Fi connectivity is often unreliable, and if access to secure network gateways is required, strict security protocols must be followed.
Though it may be complex, aligning with IT and understanding network security protocols is critical to building a robust solution. A dedicated Ethernet connection or a segregated network for the vision system, combined with disciplined data traffic control, can help address these challenges effectively.
Data Related Challenges: Challenges with Data quality and imbalanced datasets
In a controlled environment, like a research lab, experimental data is typically clean and well-balanced. If not, it can be easily curated and pre-processed to fit our needs. However, this is not always the same for real-world data. Real-world production data often comes with noise, variability, and missing context, because it’s not generated in a lab.
For example, a nudge to the camera could obscure part of the object being inspected, or an operator’s hand or a wire might block the view. Another common issue is data imbalance.
Defects are rare in a well-run manufacturing plant, which can lead to an unstable model.
While camera mounts and enclosures offer durability, regular monitoring and clear
communication with the plant are essential for quick fixes. In some cases, simulated defects can be introduced during machine downtime to boost dataset diversity. Data augmentation techniques also help address imbalance and enhance model performance.
Bottlenecks in Labelling and Annotation
Most plants produce parts at a high rate. Some high-speed mass-producing applications can see around 20,000 parts per day, which is a dream for any company that works with data. However, most AI models deployed at these plants are usually supervised models, and one major bottleneck that these solutions see is that of labelling this data.
Manually reviewing and labeling large volumes of data demands significant time and attention, especially when defects are subtle or hard to detect, making manual review time-consuming and inefficient.
Implementing reliable data annotation tools and automated annotation processes can streamline the workflow, reduce manual labor, and free up valuable resources for other critical tasks.
Subjectivity and Drift in Data Over Time
Machines are typically operated and monitored by different shift workers, each with their own interpretation of whether a part is defective. While there is a standard to follow to decide if a particular part is defective or not, there are some borderline cases in which the opinions of the operators could differ.
When visually similar images are labeled inconsistently, the dataset suffers, and model
performance takes a hit. Furthermore, opinions could vary not just from person to person but also for one person over time. An operator who labeled a borderline case as defective three months ago might now mark a similar image as “Good.”
Over time, this noise can erode model confidence and degrade performance. The workaround for this is to establish clear, standardized definitions of defects and non-defects. Regular review sessions with operators at the plant to align on labeling practices can help minimize these inconsistencies and improve model performance.
While inline AI vision systems offer substantial benefits, they come with their own set of challenges in deployment and day-to-day use. But when these hurdles are addressed early and thoughtfully, the result is a system that’s not just stable, but trusted. That trust leads to smoother operations, faster decisions, and better-quality products.
