Stop Wasting Time Transcribing Gauges: Automate Gauge Readings with AI in Days, Not Months

Introduction

Struggling to convert your legacy machinery's cryptic dials into actionable data? Stuck waiting for months to train custom AI models? We hear you. This blog is your fast track to decoding analog gauge in days, not months, starting with no data leveraging the power of AI models trained on synthetic data. Say goodbye to manual transcription and hello to real-time insights.

A working demo is on 🤗 check it out! For details continue on.

How hard can reading analog dials be? Kids can tell time, anyone with elementary education can be taught to read gauges, but GPT4 can't 🫨. Foundational models still can’t do learned tasks hence the need for custom models, but custom models are hard to train and may not perform - it is almost entirely because of data. Here we will show how these are not the case and how data solve them.

Problem

Unplanned outages, equipment inefficiency, and repair downtime cost billions in manufacturing & asset-intensive industries. Asset reliability management, ensuring critical equipment performs optimally, suffers from massive data gathering bottlenecks.  It's been realized that a significant amount of time is spent on gathering data than any other task. If you have the capital, there are existing solutions such as Boston Dynamics Spot and Lilz that automate the data collection process. As it turns out, the actual reading of the gauge isn’t an entirely plug-and-play solution. 

Synthetic Data

AI requires a lot of training data, but publicly available gauge datasets are scarce or closed-source. That's why we're open-sourcing our 42k+ image dataset, the largest of its kind! You can find it here. This dataset is for the two stages of the process described. The annotation of the data is perfect, the environment highly varied, and it costs half-day of compute to generate all thanks to synthetic data.

Conclusion

We've shown how a major digital transformation bottleneck can be tackled with open-source models and synthetic data. By combining synthetic data and open source models anyone can easily, quickly, and cost effectively build out solutions. It takes 2 days from a problem statement with no training data to a working model - Imagine all the other problems that can be solved. If you haven’t already, Try our live demo on Hugging Face and see for yourself!

Solution

Reading values in the wild is significantly more challenging. There's a lot of implementations out there, from good old computer vision, AI vision systems. Many implementations expect a direct flat view image of the face of the gauge with the user also entering in the minimum and maximum values of the gauge. This may work for highly controlled and static environments - will you do this manual work for the hundred and thousands of gauges?

Our process involves detection, perspective correction, and gauge reading, end-to-end process you would expect if you just took a picture on site. The first stage involves detection and perspective correction via homography - correcting the image such that you get a direct facing view of the gauge face. The second stage is to identify the needle tip, start and end marking values to determine the value from OCR. For both stages we used YOLO v8 pose and for OCR you can use EasyOCR, Tesseract, or some cloud service. The training takes about a day on lowest tier T4 Google Colab.

Subscribe to our newsletter