Your message is highly valuable for us. One of our experts will follow up with you within 1-2 business days to discuss your request or to inquire for additional information if needed.
Many plants with continuous production still rely on manual data collection from old gauges. Installed years ago, they perfectly fulfill their main functions. Replacing such working equipment with the next-gen APIs can result in large expenses. Hence, manufacturers seek economically viable solutions to automate data retrieval and sending it to the cloud.
Our customer, known for its innovative digital initiatives, asked us to help with the development of a Computer Vision and IoT solution for their metal manufacturing client. It could automate taking readings of old gauge equipment, thus optimizing manufacturing performance and reducing human error. It is rather complicated to retrofit and integrate old legacy gauge equipment (weighers, thermocouples, barometers, sensors, etc.) into a modern IT ecosystem according to the process digitization requirements of Industry 4.0 / OPC Unified Architecture and IoT standards. We had to figure out how to quickly recognize data on gauge pointers or 7-segment display with a pocket-size device.
Infopulse’s R&D team trained a convolutional neural network (CNN) to instantly recognize digits on the customer’s gauges despite image defects from 7-segment LCD display such as glares, white spots, reflections of objects or people, etc. To vary digits, our designers processed 10 GB of gauge images received from a manufacturer, their ad hoc image updates and studied other open source examples, ending up with 20 GB of images for CNN building.
The team chose Theano & Keras open source tech stack as their Machine Learning Toolkit and performed one-time feeding of all images to a neural network on a powerful Amazon AWS Virtual Machine. Then, we started the actual training of a network including counting, discerning small and large numbers, transforming the pointer position on a circular gauge into specific digital value. Upon reaching the needed result, the compiled CNN was deployed to Raspberry Pi 3 device with the HD camera.
Technologies: Raspberry Pi 3, ARM32, Theano, Keras, Python, AWS
Technology domains: Computer Vision, IoT, Embedded, Data Science
- AI image recognition is achieved in less than 2 seconds with a small handheld device.
- Received data is automatically stored in a database via Wi-Fi.
- MVP release was developed in 2 weeks whereas production-ready system was released in 3 months by the Infopulse team of two engineers and one architect.
- We helped the customer save development costs by utilizing Theano & Keras free and open source software (FOSS).
- Achieving high fault tolerance due to the reliability of chosen hardware and developed bug-free software.
- On top of that, after reaching one customer, the developed solution made interested other huge clients, such as automotive and beverage producers.