An enormous number of production steps are required before a vehicle completes its first kilometers on the road. From the blinker glass, which costs in the cent range, to the drive unit, which has a significant influence on the overall price, every single component must be faultless and correctly installed. This prevents complaints and meets customer requirements. This applies to all vehicles, regardless of whether they are equipped with an electric drive or a combustion engine. However, electric vehicles require special quality inspection processes, especially for the production of their batteries. Machine vision systems based on artificial intelligence provide cost-effective solutions for this, as the following examples show.

Testing the cap weld

The essential core of an electric vehicle is its battery cell. If their quality is not perfect, their efficiency is reduced and uneven load distribution between the cells makes battery management more difficult. Both of these factors shorten the service life of the battery pack.

The problem with the production of battery cells is that faults are difficult to rectify once they have been connected to modules and packs. After the electrodes and separators have been packed together in the housing of a cylindrical or prismatic cell, the housing is sealed with a cap. To prevent damage to the sensitive electrical parts inside the housing, a low-temperature welding method is required. Such welds are usually made using a laser and must be extremely precise to ensure a secure seal around the cap. The welds must pass inspection before the cell is used in a battery module or as a single cell. Any leakage of electrolyte through a faulty weld will reduce the efficiency of the cell and could lead to short circuits in the battery. A correct assessment of the cap welds is therefore of great importance for the functionality and service life of the entire battery.

The role of image processing

Cognex's 'In-Sight D900' series of intelligent cameras have embedded deep learning inspection capabilities that allow for reliable differentiation of true coating defects from unacceptable defects and acceptable deviations.

© Cognex

From an image processing perspective, this task presents a challenge: Welds can vary significantly in appearance and exhibit a range of defects; on the other hand, there can be specific and possibly even large deviations that do not affect performance. With a conventional vision system, it is almost impossible to distinguish different appearances from functionally significant deviations, as their appearances overlap. Cognex's deep learning defect detection and classification tools offer a solution to this problem. They are trained using a wide range of welding variations and thus learn to accurately classify and differentiate between the various defect types despite variations in the objects and weld seams.

Inspection of the dosing opening

After welding and testing the top cap of a battery cell, it is filled with a liquid electrolyte that conducts the electrons in the battery. Once filling is complete, the required metering hole is welded shut. Due to the risk of heat damage to the anode, the cathode and the electrolyte, this step is also carried out by laser welding at a low temperature.

An electrical test of the filled and sealed cell could reveal any problems before the cell is installed in a module, but this method is not one hundred percent reliable. Since the appearance of the weld seams can also vary greatly in this case and deviations can lead to unacceptable functional losses, Cognex's deep learning-based tools are used to reliably distinguish functional defects from purely cosmetic ones.

Testing the battery surface

In the next process step, the battery cells are protected with a permanent coating. However, this can have defects - for example, bubbles and inclusions under the coating, scratches in the coating or inadequate application of the coating itself. If faulty cells are densely arranged in a battery module, several factors can lead to an electrical short circuit or overheating, for example if the cells are too close together, the charge of the individual cells is not correct, the heat generated by the cells becomes too high or there is insufficient contact with the heat-conducting material.

Checking whether these coatings have minor defects without impairing their function, or whether seemingly small scratches result in unsafe or unusable battery cells, can hardly be done reliably with conventional image processing algorithms. This can be remedied by more sophisticated image processing systems, such as intelligent cameras from Cognex's 'In-Sight D900' series, in which deep learning inspection functions are embedded in the image processing system. This system is trained on a set of images of both good and defective surfaces. The deep learning defect detection tool uses these training images to learn to recognize and accept surfaces within the acceptable range of deviations and to mark those with unacceptable defects. In doing so, it takes into account natural deviations in the image, such as light reflections.

Safe production processes

The author: Peter Stiefenhöfer is the founder of PS Marcom Services in Olching.

© Cognex

The above examples show a small selection of applications from the manufacture of electric vehicles in which machine vision in conjunction with deep learning can contribute to safe production processes. Other applications from this industrial sector can be found, for example, in the stacking of individual electrode sheets, in which the correct, µ-precise alignment to each other is checked.