Fluorescent penetrant inspection (FPI) is a popular way to inspect manufacturing products, thanks to its relative accuracy, simplicity, and non-destructiveness. However, the process isn’t perfect, with its reliance on human vision introducing the potential for error and inefficiency. If you could automate your inspection process, those concerns would fade.
Quality control automation has become increasingly viable with the rise of machine learning (ML). ML is excellent at spotting patterns and, more importantly, outliers from those parameters, making them a promising solution for quality inspections. With the right approach, machine learning could unlock FPI’s full potential.
Automated FPI comes with many potential benefits. One of the most significant is machine learning’s speed. Human vision is not uniform and can take 100 to 500 milliseconds to perceive visual input. Computers can process the same information at several times the speed, helping accelerate the FPI process which would otherwise form a bottleneck in the production pipeline.
Machine learning is also more reliable than humans when it comes to repetitive classification tasks like FPI. Because ML works off hard data and fast, in-depth analysis of pre-learned standards, it’s often more accurate and delivers the same accuracy every time. Humans, by contrast, can easily make mistakes, especially as they get tired.
It’s also important to note that employees often don’t enjoy repetitive, non-engaging tasks like quality control. Consequently, by automating it through ML, you can free that part of your workforce to focus on workflows they find more enjoyable or engaging. That could boost morale and job satisfaction, helping reduce turnover and improve productivity.
While automating your inspection process has substantial potential, it carries some risks, too. Most notably, machine learning models can only fulfill their accuracy and reliability promises with an effective training process. That can be challenging if data scientists don’t have sufficient, reliable data or understand the end goal well enough.
Similarly, relevant experience, or a lack thereof, also poses a challenge. Almost half of all ML-deploying companies cite issues with programming and framework support as a barrier to machine learning. Another 37% report trouble monitoring model performance, and, because this technology is still relatively new, many organizations lack the necessary expertise to develop and monitor it effectively. Costs are another prominent concern with machine learning. Training these models can be expensive and disruptive upfront, even though an effective end product will save money over time.
Despite these lingering challenges, several recent studies have proven it’s possible to automate your inspection process effectively with ML. Some algorithms have achieved 95% accuracy and a 0.07 false positive rate, far surpassing manual processes. Here’s how you can design and implement machine learning models to automate your FPI.
The first step in effective ML application is determining what data points your model should recognize. In an FPI context, this means finding specific visual markers for what constitutes a defect. Common surface defects to look for in FPI include cracks and fractures, though in some cases, bumps or warping may be a more prominent concern. Your inspection process likely scans for multiple defect types, so compile a list of what your workers look for and what those look like. Be sure to include acceptable parameters for each, too, to prevent your ML model from flagging small imperfections as significant problems.
Designing an ML model also means deciding the type of data to analyze, but that’s fairly straightforward in this context. Your machine vision system will look at images of the products under ultraviolet light. Be sure to train it on the same image file type as its cameras will use in practice.
Next, you must collect the right data to train your algorithm with. These ML models learn to detect defects by learning what a passable part looks like and what an unacceptable defect is. That means providing plenty of examples, in this case, images of quality-passing cases and defects.
Machine learning works best when it has a lot of data to learn from. Every time quality inspectors notice a defect, take a picture of it to provide more data for your ML model. Ideally, you should take photos from multiple angles and in different lighting conditions for each instance.
An important but easily overlooked part of this process is cleaning your data. Make sure everything is accurate, high-quality, organized, and in the same format. This cleaning process will produce more accurate results.
It’s also important to find the right type of model to automate your inspection process. There are many different machine learning algorithms available today, and each offers unique advantages and disadvantages.
Two of the most popular for FPI automation are random forests and residual neural networks (ResNets). A random forest combines multiple decision trees, which follow flow charts for data classification, giving them high accuracy and flexibility to learn new information.
ResNets, by contrast, are a type of deep learning using multiple complex layers of processing but skipping layers as necessary to speed up results. ResNets are often more accurate, even with limited data, than random forests in FPI. However, random forests may be easier to program and deploy quickly. Determine how sensitive your FPI process is and consider your budgetary and expertise constraints to choose between them.
Once you’ve selected a base model and have enough data, it’s time to train the algorithm. This involves programming the model on what constitutes a failed quality check and what’s acceptable. Then, you’ll feed it more and more data to test it, correcting it when it misclassifies an image to make it increasingly accurate over time.
Training can be lengthy and involved, but many automated tools available today can help quicken the process. You may also need to reach out to a third-party machine learning developer to help build and train your FPI model.
After your model has reached an acceptable, consistent level of accuracy in training, you can deploy it. However, given machine learning’s costs, it’s best to apply at a small scale initially. Start by automating just one FPI line and monitor it closely. Refer to your most experienced quality inspectors to judge how well the algorithm performs and where it can improve.
ML projects can often cost upwards of $100,000 depending on their complexity and scale, so it’s best to scale slowly and watch them carefully. A slower approach with more expert input will improve model accuracy before you increase its costs too dramatically. This, in turn, will provide a better return on your investment, helping you make the most of FPI automation.
It will take time and money to automate your inspection process, but it comes with many benefits. Research shows that machine learning can make FPI inspections more accurate, efficient, and reliable than manual alternatives. You can then ensure a high level of manufacturing quality while maximizing cost-effectiveness.