Updated: Sep 6
Defect detection and rejection have always been an industry classic. With the recent advancement in artificial intelligence (AI) we have seen newer and more reliable approaches in defect detection technology, especially in the consumer electronics sector. Nevertheless, we have not seen this technology being adopted in the food manufacturing industry successfully. Adopting such a system to a food manufacturing plant can be a difficult challenge especially when dealing with high-speed conveyors in a dynamic environment (e.g., changes in lighting, etc.). In addition, unlike electronics products, food products are usually clumped together on the conveyor line, which makes it even more challenging.
This led us to our partnership with one of our customers, which was interested in increasing sufficient quality checks. Prior to our partnership, our customer conducted manual defect detection and removal which depended on human vision and lab quality checks on their high-speed candy line. Due to the fast-paced and high-volume nature of the production process, the existing manual inspection and removal process was not an ideal solution. First, the allocation of labour (i.e., human inspectors) could be utilized somewhere else to optimize the overall production. Second, if the defect rate is particularly high, human inspectors might not be able to keep up with the defects, which can result in potential quality issues to their customers. Below is an example of the production conveyor. In addition to being accurate, a defect detection and rejection system must also be able to keep up with the speed of production.
To solve our customer’s problem, we installed our vision and robotic AI platforms. This comprises an industrial camera, a computing platform with a Graphic Processing Unit (GPU) and WiFi connectivity, and a collaborative robot (CoBot) arm.
The vision system enabled real-time monitoring and analysis of the candy line. Concretely, we successfully developed AI models that can perform accurate and efficient defect inspections on this high-speed candy line. We were also able to run the same platform to detect defects on more than 20 different product families. Furthermore, our AI models are not sensitive to changes in lighting at the production line and does not require any custom lighting solutions, making them very practical to use. Below is an example on how our AI system successfully detects defects on various candy products.
To facilitate automated defect removal from the main production line, we installed a six degrees of freedom lightweight CoBot attached with a food-grade hose and a protective jacket. In addition, together with our integrator network, we successfully installed our robot without interrupting the line. This is important to us and the customer because line interruption can be very costly. Once the defect is detected, our system instantly triggers the robot and moves it to the correct location to capture the defect. The video below shows how our robot captures a defect in slow-motion.
In addition to vision and robotic AI platforms, we also provide the customer with a dashboard. This dashboard allows the customer to see the historical data for traceability (e.g., which time period or which products produce the most defects). As a result, the customer can try to understand what problem might have happened during a specified period, and come up with a solution to improve their production.
As a result of this AI-driven solution, the customer observed a remarkable increase in the inspection rate, which led to a notable reduction in defects found in the packaged products. This provided a solution to increase quality checks on the high-speed candy line and enabled the client to allocate the human inspector for other tasks. Furthermore, from the defect dashboard, the client can gain more insights about their production to improve their production process. This demonstrates the value of integrating our technologies to optimize quality control processes, particularly in high-speed and high-volume manufacturing environments.
On average, our defect detection model achieved overall accuracy of 99.1% ,
In a test that we conducted with the customer, our system achieved relative performance of 162% compared to a human inspector .
The customer can now reallocate labour usually needed for manual inspections to other tasks, reducing direct product cost,
Digitization of defects data which can be accessed through our dashboard,
Real-time insights on defects, which can guide the customer in taking corrective actions to reduce the defects.
Dashboard allows the customer to see the historical data for traceability
: We collected a dataset of defective products and computed overall accuracy based on true positive, false positive, true negative, and false negative rate (last measured on 2023/09/02). Since the ratio of defective and non-defective data in our dataset does not reflect the ratio that we observe in the real-world, when computing overall accuracy, we adjusted the weight of non-defective data to reflect the real data distribution where the ratio of defective and non-defective images is around 1:20.
: Assuming a constant defect rate, our system was able to remove defects at 188 defects per hour, while human inspector is at 116 defects per hour during the same production run.