Product quality is the foundation of the enterprise. Recalling products and returning to the union is costly and permanently damaging to a company's reputation. Especially with the growing demand for customization, companies are under tremendous pressure to ensure that any product on the production line is intact. But with the rise of machine learning and machine analytics, businesses can now change their quality management. Relimetrics is using this technology to help enterprises analyze video data in real time at the edge, bringing them one step closer to achieving "zero defects."
Headquartered in Germany, Relimetrics is an HPE OEM and NVIDIA Metropolis platform partner. The company helps manufacturing companies revolutionize the way products are designed and created with fully digital quality audit (QA) software. The solution provided uses GPU acceleration, real-time video analysis and machine learning to detect the configuration and attributes of product parts, thereby improving the accuracy of defect detection.
Relimetrics is well aware that defects can make manufacturers sleep. In today's fast-paced production environment, companies are under increasing pressure to properly manage quality and respond to the growing demand for product customization from customers. Manufacturing and assembling products is becoming more complex, precision requirements are increasing, and the likelihood of anomalies is increasing.
Kemal Levi, founder and CEO of Relimetrics, illustrates this by way of example, "The car you drive is made up of more than 35,000 parts. If there is a problem with one part, it can lead to a vehicle recall, and the overall cost of completing the recall and the loss of the manufacturer's brand value can be as high as billions of dollars. ”
Use edge analysis to reduce defects and reduce inspection time
By enabling enterprises to run video analytics at the edge, Relimetrics helps them detect quality issues more accurately and resolve them quickly. This reduces rework, increases profitability, and increases customer satisfaction and retention, bringing businesses one step closer to zero defects.
Levi said, "We are taking QA automation to the next level, and reducing rework is no longer a dream, but a real reality." ”
Levi and his team showed in action how to automate and transform the way companies manufacture by using the intelligent edge to automate when performing QA digitization of Foxconn's production line.
The plant, located in Kutná Hora, Czech Republic, manufactures servers for HPE. These customizable products are complex and varied. For example, a server model may be equipped with 2 to 16 memory modules, each of which can be 16, 32, 64, or 128GB.
The memory configuration is just one of more than 20 different product manufacturing variables associated with it. Therefore, having workers inspect these servers on fast-moving conveyor belts inevitably misses some defects.
Levi explains, "It takes 2 to 5 minutes to manually detect each server. For a production line like Foxconn, this means increased costs and reduced productivity. ”
Improve detection accuracy
HPE and Foxconn decided to use machine vision at the edge to automate production line QA. They want to improve the accuracy of defect detection and reduce the number of rework servers that customers receive through more efficient quality traceability.
Levi said, "The key goal is to reduce warranty claims and improve the overall quality of service. A large part of these claims stems from poor product quality delivered to customers, so adopting technology that adapts to changes in production and is more accurate than manual inspection is critical for products like HPE brands. ”
"We are now able to automatically check the configuration and assembly quality of our products and eliminate human error in areas where solutions can be validated."
—Foxconn spokesperson
Enable analytics on the production floor
Relimetrics has teamed up with HPE Pointnext Services to implement video analytics and machine learning technologies on Foxconn's production line.
The solution consisted of installing 5 cameras to capture high-resolution images entering the conveyor belt product and transferring the images to an embedded system for on-site processing using machine learning algorithms. Relimetrics' machine vision system then compares the image captured by the camera with the reference image. Where, the reference image shows whether the server component was implemented correctly or incorrectly. If a problem is detected, it is immediately flagged so that Foxconn operators and line managers can immediately correct the problem. If the server is in good condition, you can proceed with the final wrapping and send it to the customer.
The system uses the HPE Edgeline Converged Edge System, an HPE EDGE OEM solution with NVIDIA® GPUs, to analyze images at the edge instead of the cloud. This setup saves time and prevents latency issues, as each camera installed on the conveyor belt transmits a large amount of data (about 3 GB per hour).
Levi noted, "By introducing NVIDIA GPUs into the HPE Edgeline EL4000 chassis and using NVIDIA TensorRT to optimize network performance, latency was reduced from 4165 milliseconds to 3 milliseconds, which equates to a nearly 1400-fold performance improvement." ”
Speed up algorithm training
Leveraging the combined capabilities of the CPU and GPU, the Relimetrics solution can train AI models online.
"In this way, manufacturers can digitalize QA without having to worry about retraining the model offline every time a new configuration appears in production," says Levi. ”
When the products involved, such as HPE servers, are highly diverse and customizable, training machine learning algorithms to detect problems or anomalies can be difficult and time-consuming. To save time, Relimetrics and HPE Pointnext Services built a machine vision system at the Foxconn plant that requires storing reference images of server components rather than the entire device. Based on this design, the MANUFACTUR EXECUTION system provides a bill of materials for each product on the conveyor belt so that the system can produce a complete reference image based on the reference image component.
Improve production performance
With the ability to process data in real time at the edge, Foxconn's production lines speed up the quality inspection process.
According to Heinle, the plant can save 96 seconds of inspection time per server, increase test coverage by 20%, and cut related costs. "Additional audit checks and machine learning solutions are error-free due to fatigue and are able to adapt to large production changes, resulting in a two-percentage point increase in initial pass rates and a significant improvement in overall production performance at the plant."
"By introducing NVIDIA GPUs into the HPE Edgeline EL4000 chassis and using NVIDIA TensorRT to optimize network performance, inference latency was reduced from 4165 milliseconds to 3 milliseconds. This equates to a nearly 1400-fold increase in performance. ”
—KEMAL LEVI, Founder and CEO of RELIMETRICS
Eliminate human error
As the QA process improved, Foxconn factories improved the accuracy of visual inspections, and as a result, customers received a 25 percent reduction in the number of HPE rework servers.
A spokesperson for Foxconn said: "We are now able to automatically check the configuration and assembly quality of our products and eliminate human error in areas where solutions can be verified. Levi also expressed his opinion on this, and now we estimate that the detection accuracy of automated QA is more than 999%, which truly achieves zero-defect manufacturing."
At the same time, a Foxconn spokesperson also believes that QA digitalization can also reduce some of the burden on employees, "with automation and machine vision systems, we can improve employee productivity and thus provide customers with highly competitive products and services." Machine vision systems support audit operators and reduce some of the workload for inspectors, allowing them to focus more on important aspects of their work. ”