EV Battery Defects: AI Detection
For an electric vehicle, the battery is one of the most important components. For automakers, even a small defect in the battery can result in massive costs. As a result, it’s imperative that battery manufacturers closely inspect their products to ensure they are safe to use. This article discusses how deep learning can be used as a surface inspection technique for lithium ion battery cells.
Growing EV Market
The electric vehicle (EV) market is expected to boom in the coming years. Some analysts estimate that the EV market will be worth $957B by 2030. Each vehicle requires a battery, which often represents up to 30% of the cost of the vehicle.
Lithium Ion (Li-ion) is the leading type of battery used today for EVs. However, Li-ion batteries have experienced their fair share of product failures, ranging from recalled smartphones to electric vehicles. Even with an advanced battery management system (BMS), Li-ion batteries are still defect-prone. For these reasons, Li-ion batteries should be subject to a rigorous quality control process.
Existing Approach to Quality Control
Battery manufacturers understand the high cost of battery defects. Therefore it is no surprise that companies have made attempts to enforce rigorous quality controls on lithium-ion battery production. For example, Samsung has published its quality control process for its smartphone batteries.
- Durability Tests
- Visual Surface Inspection
- X-Ray Inspection
- Charge Cycle Tests
Out of these quality control methods, one stands out – visual inspection. Anytime humans are in the loop, quality control can become subjective instead of objective. Depending on how the human inspector is feeling that day, whether the person is tired, how good their vision is, etc., the inspection results can change.
Objective inspection is more reliable, as the quality control processes are standardized. That is a major benefit of automated optical inspection, which uses image processing or deep learning techniques to identify defects.
Maintaining Quality with Increased Demand for EV Batteries
If battery manufacturers are to meet this demand, they will need an inspection solution that can process several thousand samples per day. This can be difficult to achieve with conventional techniques, which are often plagued by low inspection repeatability and high false negative rates. Fortunately, there is a solution: Deep Learning.
Deep Learning for EV Battery Inspection
Deep learning enables battery manufacturers to automate the process of battery inspection, and this can be done with high accuracy.
Deep learning has been shown to be an effective technique for surface inspection. In particular, deep convolutional neural networks (CNNs) have demonstrated impressive results on a wide range of tasks, including object detection, facial recognition, and medical image analysis.
CNNs are particularly well-suited for battery surface inspection because they are able to learn the underlying patterns that indicate battery defects.
Training Data for EV Battery Inspection
One of the biggest bottlenecks with developing AI inspection for EV batteries is training data. In order to teach a deep learning algorithm to recognize defects, these algorithms must be trained on thousands of example images. Unfortunately, collecting real-world data is quite challenging since defects, by their nature, are rare.
Fortunately, Simerse has a solution. Simerse creates synthetic training data, which are computer-generated images that mimic real-world defects. Synthetic training data is proven to perform on par with real-world data. More importantly, Simerse is the only provider of high-performance synthetic training data for defect detection. If you are an automaker or battery manufacturer, request a meeting with us here.
In practical terms, Simerse will provide you with hundreds of thousands of training images, completely annotated for machine learning. This training data is ready for your AI engineer to use. If you don’t have an AI engineer, Simerse will lend you one of ours!
AI Model Development for Inspection
Now that you have access to training data, it’s time to start developing your AI model. There are a number of ML architectures to choose from, such as ResNet or YOLO. An AI engineer will be able to recommend an architecture best suited for your EV battery inspection challenge, whether that’s looking at surface defects or conducting X-Ray inspection.
Once you have decided on an ML architecture, it’s time to start training your AI. There are several options for AI training, the two most common being:
- On-Device GPU training
- Cloud training
There are advantages and disadvantages for each. If you already have a high-performance GPU, it is often cost effective to train on-device. You might have to wait around for training, but it is the cheapest route if you have an existing GPU.
If you prefer performance and value time, cloud training is the way to go. Almost all major cloud providers (AWS, Azure, GCP, IBM, etc.) have GPUs that can spun up on-demand. These cloud services take a lot of the pain out of model training and are generally easy to use.
After training, you will have an algorithm with the calibrated model weights. Congratulations! You have an AI algorithm!
Deploying an AI Algorithm
Now that you have an AI algorithm to classify defects in EV batteries, but how do you deploy that algorithm to your factory?
Well, the answer depends on your factory setup. Battery manufacturers and automakers alike generally have highly automated factories. That means that is likely you will have an existing machine vision camera that can be used for AI defect detection.
When considering factory deployment, it is important to think about where the AI processing will happen. Will it happen locally on the camera? Or will it occur in the cloud? We recommend consulting with your internal IT team to learn your factory’s cybersecurity practices.
Generally, IT teams prefer the processing to be done locally instead of in the cloud. The reason for this is that on-cloud processing can be hacked, while processing done on-site provides more safeguards. Of course, this is entirely dependent on the technical sophistication of your factory and the degree to which you have internet-connect devices already present in your facility.
Now that you have decided on a processing route, it is time to connect the camera to the AI algorithm. Many machine vision cameras have this functionality built-in, where the model weights can be inserted into the system. If not, a data scientist or IT person should be able to set up this connection.
You will also need a feedback mechanism for your human quality inspectors. In highly automated systems this may be text-based notification, or integration with an assembly system to take automatic action based on the AI inspection.
In less automated systems, the feedback system may be a light or mechanical indication that a defect is present. Again, it depends on your unique factory setup. Fortunately, this system can be tailored to a configuration that is most optimal for your workers.
Why Proof-of-Concepts are Great for AI
Everything about AI sounds great, but what if you’re a decision maker? Or, what if you’re making a recommendation to a decision maker? If you’re not ready for a full deployment, then a proof-of-concept is for you.
Proof-of-Concepts have several advantages over full deployments:
Proof-of-Concepts are cheap. Full deployments can be expensive, and any company will want to make sure the return on investment (ROI) hits the required threshold.
Proof-of-Concepts are simple. With limited scope, a proof-of-concept can prove the value of AI inspection without burdensome complexity.
Proof-of-Concepts reveal information. To make an ROI decision, managers need good information. That’s exactly what a proof-of-concept should yield: an early indication that a full deployment will be successful both from a technical perspective and a financial viewpoint.
How to Get Started
Contact Simerse. The best way to kickstart your EV battery AI project is to contact Simerse. Our experts will work with your team to rapidly develop and deploy a proof-of-concept. Ultimately, this project will prove the value and ROI of defect detection AI.