As Conductive Anodic Filament (CAF) testing on Printed Circuit Boards (PCBs) approaches 50 years of industry attention, the chemistry of formation in the lab environment is well documented, significant resources have been expended in passing the lab test, and new efforts are in progress to make the test even more difficult to pass with voltages up to 1000v. But what about the operating environment for this shorting mechanism— to what extent is CAF a “field fail” risk today? Is it possible to prove that text book CAF failures are actually impossible in today’s typical office system environment? With future PCB technology driving progressively smaller via-via spacings, is there a point where tomorrow’s reliable product will never pass today’s CAF test? Either way, considering CAF testing as a subset of the broader Insulation Resistance (IR) test, what are the shorting failure mechanisms the industry should be concerned about and what is the best test for them?

This paper surveys 40 years of experience in a reliability and failure analysis lab with IR testing, product reliability qualifications, and field return forensics. Shorting based field failures are described and ranked based on impact, risk, and changes over time with industry wide improvements in electrical test and laminate materials. Two test methods 65C/85%rh and 50C/80%rh are compared in terms of time to fail and resulting failure mechanisms. Also presented is a data-generated temperature and humidity (T&H) model from a CAF prone material using a unique test vehicle, along with data for the important question of CAF formation humidity threshold. The IR/CAF fire triangle of moisture + path + ions is discussed-- how it perfectly explains all types of PCB insulation resistance failures. The CAF triangle is further validated by test results from purposely populating an office system with pre-soaked CAF prone boards, and by data for this question: is the T&H part of the CAF test reversible? The criticality of T&H control in accurate CAF testing is demonstrated with chamber measurements and test data. Finally, the importance of finding the right IR/CAF test for future HDI/tight grid technology is addressed.

As the manufacturing sector is under constant pressure to satisfy customers’ demands in a competitive market by applying complex processes to meet manufacturing cost and schedule goals, the need to identify quality variables within processes is occurring at a faster rate. Locating the source of process variations becomes more challenging for engineers. Each day, the manufacturing sector generates tremendous amounts of data that provide valuable information. This data is crucial to supporting strategic business operations decision-making. Traditional ways of data interpretation are labor intensive and time consuming. Failure to accurately and precisely translate data will lead to subjective “opinion” or “speculation-based” decision-making.

In this paper, we will review general opportunities for the application of machine learning (ML) algorithms and methods to the test data troubleshooting process. A method is developed for analyzing data and identifying patterns that are consistent with poorly performing units. This method uses a “quasi-supervised” learning technique to identify drivers of variance within a dataset, visualize the trends among the primary drivers of variance, and establish some screening limits based on those trends. The method employs Principal Components Analysis (PCA) to review patterns, trends, and uses some knowledge of better or worse performing groups. The output is a set of screening limits that characterize parts likely to have similar performance. The method provides clear knowledge, visualization, and understanding of the trends that are driving failures or poor performers.In addition, it does not require the rigorous data capture that a true supervised learning method. This method can be used on any dataset with observations in the rows and attributes/variables in the columns if there is some knowledge of an identifiable batch that is better or worse than the others. A performance characterization on a batch of units was successfully performed to identify the anomalies within a dataset.

To ensure the highest possible quality standards in automotive electronics production, an extensive implementation of testing and inspection systems throughout production is mandatory. In SMT production optical inspection systems are the standard technology for evaluation of quality in SMT soldering processes.

To ensure the highest possible level of quality, these systems are enhanced by human verification experts that review results from the automated process and thereby ensure a high level of quality while minimizing production losses through false calls.

In this contribution we introduce an Artificial Intelligence platform designed for application in SMT inspection processes, enhancing and eventually outperforming the human verification operator. The design of the platform is chosen to be process agnostic and can be applied in any quality inspection process that relies on visual information in pictures.

For the design of the platform, we have collected more than 1 billion solder joint pictures and labels from the shop floor as the foundation for the development work. To ensure proper training results, we have worked with a team of soldering experts to ensure correct labeling on a significant portion of these pictures.  Based on this data we designed deep learning algorithms that are capable of properly clustering the error images into the error classes predefined by internationally accepted standards and reliably identifying the large class of false calls. 

To make the algorithms usable for production and specifically to enable non-AI-experts to work with the algorithms, we embedded them into a tool suite based on apps that are easy to use for soldering experts. In the applications, datasets can be handled, new decision models can be trained, neural network quality can be evaluated and eventually the decision models can be deployed to the production line. On the shop floor the decision model can support the operator with suggestions, or it can also completely take over the task of verification in certain scenarios. 

The solution presented here is in practical use on large scale and therefore the contribution offers a theoretical approach to the topic, an implementation example with a platform solution and a view on the business impact of the solution.

The key challenge for industry in adopting AI into their manufacturing processes surrounds the accessibility of their data. Many manufacturing industries, especially electronics manufacturing, suffer from a lack of standard data protocols often due to the use of legacy equipment. The absence of standards makes it difficult to integrate assets on the shop floor which is a primary step before introducing AI into the process for creating intelligent manufacturing environments. The AI for Electronics Manufacturing project aims to provide a reference architecture which solves these connectivity issues for the PCB assembly process, which may be adopted by other industries with adjustments. The objectives of the architecture are to enable operational data transfer from production assets in a modularised manner in order to coincide with reconfigurable manufacturing systems. Additionally, the project explores the application of novel AI techniques to propose a flexible condition monitoring framework as a solution to the challenges that industry face when adopting these types of systems. Such challenges include the lack of suitable training data available to industry and the specialisation of these systems which make it difficult to transfer the system over to other applications or even equipment of similar function. The framework was evaluated with the results presented in this paper.

Neural network based deep learning models increasingly demonstrates high accuracy in object detection and image classification in digital image processing. The manufacturing industry is adapting this advanced technology to assist in automated quality assurance. Successful in implementing prototypes and small-scale deployment to employ AI models for quality inspection has been achieved. AI-assisted quality inspection significantly improves inspection accuracy, operation throughput and efficiency. “A Framework for Large-Scale AI-Assisted Quality Inspection Implementation in Manufacturing Using Edge Computing” [1] was previously presented, in which details are discussed highlighting challenges in large-scale deployment of AI models for quality inspection operation and focused on IT architectural decisions to fulfill the OT requirement, including user experience in the quality inspection ecosystem.

This paper focuses on AI model benchmarking at the edge, with respect to the architecture presented in [1]. It discusses the technical challenges to meet specific inference performance requirement at the edge. Benchmarking study of various AI models on a set of edge hardware including Nvidia Jetson TX2 and IBM Power servers are performed and recommendations on AI model and edge hardware selection is presented.

Keyword: Quality Inspection, AI Models Benchmarking, Edge Computing

Photonic soldering utilizes high intensity flashes of visible light to achieve wide area heating with exceptional uniformity. Solder paste is heated to its liquidus temperature using radiative energy transfer, and light is converted to heat through optical absorption. This process can be made selective by exploiting the high absorptivity of solder pastes relative to most other printed circuit board (PCB) materials, or with the aid of shadow masks. The optical flash can be modulated digitally, with high temporal resolution, which enables highly customizable processing flows ranging from traditional to highly innovative.

Photonic soldering is compatible with standard high temperature lead free solder alloys (e.g., SAC305) in combination with temperature-sensitive substrates (e.g., PET). The nonequilibrium nature of the heating process enables thermal isolation of active regions from temperature sensitive regions. The resulting flexibility in material selection gives designers significant freedom and new options in outlining device architectures.

Previous presentations of this technology focused on the quality of junctions formed through this process. This paper focuses on the unique features of the photonic soldering process, as they relate to production line design and operation. The main advantages of the process are rapid change of process conditions with limited hysteresis combined with short dwell time and high throughput of the system. Together, these unique advantages enable a fresh approach to tool setup and timing, which better meets the needs of next generation electronics. This paper highlights the advantages of the new technology and discusses the application space for the photonic soldering technology.

These innovations enable product designers to combine components, substrates and solder alloys that are not feasible with reflow ovens while allowing very high volume – and high throughput – manufacturing processes in a digital format.

IPC-HERMES-9852 as the smart replacement for the long-used IPC-SMEMA-9851 provides machine-to-machine (M2M) communication that ensures consistency of each PCB and its individual data while traveling down an SMT Line in production.  Thus, Hermes enables machines to consistently transfer a PCB together with its Digital Twin. This Digital Twin alone already provides valuable support for basic reporting functionality, such as monitoring and traceability reporting. But this data together with the M2M communication can do much more: It can be used for further automating certain workflows of a flexible production, bringing a cost-effective solution for automated mixed production.

In this presentation, we look at some advanced workflow examples in an automated flexible production, which can be easily automated using M2M communication provided by IPC-HERMES-9852, including:

•Automated machine program selection based on PCB related data such as barcode, product name, work order ID, etc.

•Control of the oven error loop to prevent PCBs from entering the oven while the buffer after is full and cannot take anymore PCBs from the oven

•Coordination of the interaction between AOI and Flipping Unit to allow inspection of top and bottom side of a PCB