Edge AI camera with OCR for hot-water meter readings, for industrial plant engineering teams in Thailand

How the brief came about

This system uses a camera to read values from a hot water meter with a numeric dial and sends them into a system, without anyone having to go and note them down. The customer's meter shows its value only on the dial. To know the water usage, someone has to read it at the installation point and then note it down or key it into the system by hand, so data arrives late and may be written down or keyed in incorrectly.

The customer is a factory engineering company in Thailand. The customer's IoT (devices that send data over a network) engineering team has its own servers and dashboards. What it needs is a device that turns the numbers on the dial into data and sends them into that system.

Requirements and constraints

The meter reading set must work with the existing meter and send values into the system the customer already has. The system requirements are as follows.

• Fit to the existing hot water meter without replacing it, so the camera housing case must be designed to fit the size of the meter body.
• Read the digits on the device, then send the value via API or MQTT to the customer's server.
• The equipment set has no built-in database, and the dashboard and alerts sit on the customer's system side.
• Configuration of 1 camera set, with the camera supporting a reading cycle of every 1–2 minutes.
• 2 test criteria: the camera reads the digits on the dial correctly, and the value sent via API matches the dial.

Why fit a camera to read the existing dial instead of replacing the meter?

We chose to fit a camera viewing the existing dial because the brief was to bring the value the meter already displays into the system. This approach avoids removing and replacing the meter, and our work ends at sending the value via API or MQTT to the customer's server.

We have the device read the digits itself on site (Edge AI), so images do not need to be sent to a server for processing. The only data sent out is the digits that were read. The system works as follows.

1. 1

   Camera in a housing case views the dial

   The camera sits in a housing case designed to fit the size of the meter body and photographs the number window on the dial at the set interval.

2. 2

   Read the digits on the device

   The Embedded Controller board runs an OCR model and reads the digits from the image on its own.

3. 3

   Send values via API or MQTT

   The digits read are sent as data to the customer's server. The equipment set has no built-in database.

4. 4

   The customer uses the values in its own system

   The customer's team takes the values to show on a dashboard, set alerts or combine with other data in its existing system.

System components

The system consists of a meter reading hardware set, AI camera configuration and value output. It is designed for 1 camera set, a reading cycle of every 1–2 minutes and 2 test criteria. The components are as follows.

• Camera for reading the meter (Camera for check meter)
• Embedded Controller board with power adapter
• Camera housing case
• AI camera configuration to read the digits on the dial and send values via API or MQTT

There are 2 versions of the work specification: the first in 2024 and a revised version in 2025, which adds hardware housing case design work.

What we built and tested

We tested Python OCR code with photographs of a real meter dial on a small computer board, and it read the digits. This result comes from our internal testing with photographs. We also measured the actual meter body to design the camera housing case.

The system's 2 test criteria are for measurement on the meter at the installation point, so accuracy and speed figures must be measured at that stage. This page reports only results that have been tested.

What we learned

This work offers the following principles for reading meters with a camera.

• Acceptance criteria must be compared with the numbers on the real dial at both the source and the destination. This system separates its test criteria into 2: the value read by the camera, and the value received at the destination via API. A confidence value (confidence is the score a model gives its own result) is not accuracy until the result has been compared with the real dial.
• Camera position and the housing case are part of the scope of work from the start. The housing case sets the camera's distance and angle. In this work, the actual meter body was measured, and the revised work specification added housing case design work.
• Divide responsibilities clearly between the device and the destination system. In this work, the camera set only reads and sends values, while the dashboard and alerts sit on the customer side. The API or MQTT channel is a handover point that both sides can check.

How can you apply this approach to the meters in your factory?

If your factory still has people walking round to read meters with numeric dials, and has a system that receives data via API or MQTT, you can use the same approach as this system: fit a camera viewing the existing dial, read the digits on the device and send only the values into the system. What needs adapting is the housing case and the testing with your own meters.

• Prepare device power and a network at the installation point, and check your organisation's network policy.
• Plan testing that compares the digits read by the system with the numbers on the real dial before putting the values into use.