Artificial Intelligence is gaining use in a variety of manufacturing applications. The ability to learn patterns and prescribe actions in repetitive scenarios is an invaluable tool for businesses large and small. In industrial settings, a variety of business problems are particularly well-suited for AI-based solutions. When traditional systems for visual inspection and automation are complemented by AI, stakeholders realize large gains in productivity and reduce operational costs. Quality assurance is often cited as one of the most common use-cases for AI in the industrial world.
Visual inspection systems which are typical in most plant-floor quality inspections can be enhanced with AI-based analysis to reduce the cost of quality and compliance. In the ensuing sections, we examine the various elements of AI-driven quality assurance approaches.
Quality Assurance
Quality assurance best practices focus on the prevention of mistakes and defects during the manufacturing process. A typical production line involves a streamlined process with a series of operations – manufacturing, machining, pick and place, inspection, packaging, shipping, etc. – to ensure that the product being produced adheres to exacting standards demanded by customers or regulations. For example, in computer numerical control (CNC) operations, several machines are used with an external controller, such as conveyors or robotic arms, to move components from machine to machine and align them under the guidance of industrial cameras before cutting operations commence. The objects are then conveyed to the next operation for flaw inspection. While human inspection is not uncommon, visual inspection systems are employed to reduce human errors and save costs.
A visual inspection (machine vision) system constitutes a camera and rules-based image processing componentry. These machine vision systems analyze photos of parts and components for defects, specification adherence, and other criteria. Since defects found by customers can result in expensive warranty or repair costs, inspection systems should be sophisticated to identify failures on a variety of dimensions. In addition, these systems need to be integrated into other production processes so corrective actions can be taken without delay. Since production environments are dynamic and often complex, vision systems that can learn and adapt can lead to higher quality products and lower operating costs for stakeholders.
Enhancing Quality Inspection with AI
Many businesses still rely on human inspectors or inflexible rules-based machine vision for their quality needs. Industry statistics indicate that human inspection has an error rate of 20-30%1. This high rate is due to subjective assessment and overkill (products marked defective when they are acceptable). Advances in artificial intelligence (AI), machine learning (ML), and deep learning (DL) are bringing new capabilities that help overcome many challenges in visual inspection. At the core of the AI-based quality inspection process is machine learning. By learning from defect histories, computers arrive at highly sophisticated algorithms for quality inspection. AI platforms create dynamic and real-time defect books that systematically help human inspectors with clear defect definitions. Reducing ambiguity and applying a uniform ruleset greatly increases accuracy and consistency.
Figure 1: Data sources aid in key business outcomes.
Another area where AI brings superior capabilities is in dynamically changing quality measures to meet new requirements. AI-based inspection methods collect and systematically organize quality data and can also identify even small changes in operating performance. This data can be used to alter operations and retrain the algorithms for even more precise performance. Abrupt changes can also be noticed well before humans can observe and intervene with corrective actions.
An emerging area for AI impact is in generating unique insights from integrating disparate data from supply chains, operational KPIs, and returns/warranty data with defect data. This type of integration allows stakeholders to assess the true root causes of quality issues and make changes to improve quality.
AI-Based Visual Inspection System
Cameras are a crucial component of any visual inspection system. Camera imagery is processed by a computer and the data is analyzed against a defect model and a pass/fail determination is made. The result and camera data are archived in a data storage area. The defect model is itself a result of a machine learning process that uses the historical data archive to determine patterns. The algorithm continuously learns from archived images that represent both good and bad quality parameters to build a comprehensive model of the quality measure.
Figure 2: An example AI-based quality inspection system (Source: QualityMag)
In designing AI-based visual inspection systems, several factors need to be considered.
• Inspection goals - Estimating the cost and application benefits is essential. Ease of use, configuration, calibration, and maintenance are all important when determining the cost of the inspection system. The cost of the visual inspection system needs to be then compared to the cost of errors when manual or human-only inspection is employed. Ensuring that the system has the right payback and a low total cost of ownership will provide the right framework for improving the manufacturing process and product yield. Cost targets can also help stakeholders make build vs buy decisions on the system needed to meet the inspection goals.
• Inspection characteristics (inspection time, features/defects, material handling) - Quick inspection times will require fast and reliable image processing required for defect determination. A full characterization of the defect profiles and creating an image database will provide a good estimation for the requirements and sophistication of the hardware and learning algorithms. A key factor is also the amount of lighting and type of optics required to ensure speedy and accurate determination of defects. The system can be set up to be in-line or off-line with the manufacturing process. This will also create requirements on the overall integration, image acquisition, learning model, and processing hardware of the system.
• Image-processing hardware – The right choice of image processing hardware will optimize power consumption, size, cost, and the overall ecosystem of tools and training. Graphics processing units (GPU), field programmable gate arrays (FPGA), and vision processing units (VPU) are the main choices available for system builders. All these systems offer a comprehensive ecosystem of software, technical resources, and support to aid in implementation:
o GPUs: GPUs have thousands of processor cores and target computationally demanding tasks and training algorithms. Due to a massively parallel architecture, GPUs are suitable for accelerating deep learning applications in visual inspection systems. NVIDIA has been the market leader for GPUs and has invested heavily in the development of tools for enabling deep learning and inference processes.
o FPGAs: FPGAs are widely deployed in machine vision cameras and frame grabbers. FPGAs offer the middle ground between the flexibility and programmability of software running on a general-purpose CPU, and the speed and efficiency of a custom-designed application-specific integrated circuit (ASIC). Intel FPGAs come with an extensive ecosystem of support and tools.
o VPUs: A VPU is a System-on-Chip (SoC) designed for the acquisition and interpretation of visual information. VPUs target mobile applications and are optimized for small size and power efficiency.
• Learning models - The learning model is critical to ensuring that the visual inspection system handles the complexity of the inspection task, required delivery time, and cost.
o A deep learning model using pre-set services (e.g: Google Cloud ML Engine, Amazon ML, etc.) is appropriate when requirements for defect detection features are in line with templates provided by a given service. These services can save both time and budget as there is no need to develop models from scratch. Deploying them can be as simple as uploading data and setting model options to perform the relevant tasks. These models are not customizable and may not be suitable for all inspection requirements.
o Using pre-trained models - A pre-trained model is created from a separate deep learning model that accomplishes tasks required for visual inspection. A pre-trained model may not comply with all requirements, but it offers significant time and cost savings. Using previously trained models on large dataset models allow customizations to suit specific needs.
o Deep Learning model development from scratch - This method is ideal for complex and secure visual inspection systems. The approach is time and effort-intensive, but the result is a model that suits the specific requirements of the inspection system. When developing custom visual inspection models, data scientists use one or several computer vision algorithms. These include image classification, object detection, and instance segmentation. The choice of a deep learning algorithm can be made by incorporating business goals, size of objects/defects, lighting conditions, number of products to inspect, types of defects, resolution of images, etc. and arrive at a highly specialized system.
• Build or Buy - Stakeholders have the choice to assemble various components of the system themselves by procuring individual components or to buy fully integrated systems. Time-to-market, in-house expertise, and the risk of failure drive stakeholder choice. Building a system has the advantage of offering the greatest flexibility of customization and can be very cost-effective for high-volume applications. However, building a system can involve high upfront development costs and pose project risks. Also, future-proofing the system and integration with other aspects of the manufacturing process will require additional effort and constant upkeep.
In contrast, buying pre-configured and integrated systems offer a high level of integration and limit customization. Fully integrated systems that come with smart cameras are small, compact, all-in-one vision systems that incorporate lens, image sensors, system storage, and processors into a single device. These are increasingly popular as they take away the hassle of assembling all the components. Fast time to market and low risk are additional benefits accrued by buying a pre-built system.
Figure 3: ADLINK’s NEON Machine Vision Camera comes with both NVIDIA Jetson and Intel Movidius processor options.
Summary
The above are the chief considerations that go into deciding on an AI-based visual inspection system. While the choices can be overwhelming, there are market solutions that can cater to any need. Implementing a visual inspection system requires a thorough analysis of the requirements, evaluation, and prototyping to ensure that the final solution achieves business objectives.
Comprehensive solution providers like Arrow offer solutions for build and buy options and can also support with design services to aid in the deployment of vision inspection systems. Arrow’s experts have experience defining solutions across the entire AI stack and assisting companies in adopting AI/ML technology for optical inspection. Further, Arrow’s proven experience developing distributed architecture helps in implementing edge-to-cloud solutions without compromising on time to market. These capabilities will help customers strike the right balance of cloud/edge infrastructure for their visual inspection systems.
References
• Ten basic steps to successful machine-vision-system design
https://www.vision-systems.com/boards-software/article/16736434/ten-basic-steps-to-successful-machinevisionsystem-design
• Finding the optimal hardware for deep learning inference in machine vision
https://www.vision-systems.com/boards-software/article/14039580/finding-the-optimal-hardware-for-deep-learning-inference-in-machine-vision"
