Optimization of Signal Detection Using Machine Learning

Increasing the Accuracy of Device Recognition via RFID Chips with the Help of Machine Learning


Tools and equipment from companies are often distributed among vehicles and employees or stored at the headquarters. It is a challenge for companies to keep track of where each piece of equipment is currently located. Manually recording lending and return transactions is time-consuming and error-prone. Therefore, Wattro GmbH has developed an automatic scan terminal for devices using RFID chips ("radio-frequency identification chips"). These chips, attached to the devices, store data that can be read wirelessly.

In an initial version, the automated detection provided a suboptimal accuracy of only 90%. For example, it was not yet reliably distinguishable whether the device was moving towards the terminal or at the back of the warehouse.

Therefore, INWT investigated whether the accuracy could be improved using machine learning methods.


For the proof of concept, data was generated for various scenarios, such as:

  • An item is moved towards the terminal.
  • An item is moved towards the terminal while other items are nearby.
  • An item is moved towards the terminal, and at the same time, another item, which is not currently of interest, is being moved in the warehouse behind the terminal.

The terminal sends out a signal that is returned by the passive RFID chips attached to the tool or device. This generates data such as the ID of the item, the strength of the returned signal, and the phase shift of the signal. Additionally, an ultrasonic sensor is used to measure whether an item is directly in front of the terminal; however, this sensor cannot determine the device ID or detect devices that are further away.

All sensor data was available in the form of measurement series.


The essential part of the modeling involved calculating features. This was done in close coordination with the client to meaningfully represent the relevant physical processes and make them tangible for the machine learning model.

The features included, for example:

  • Aggregation of the measurement series at the item level, such as by calculating the average speed
  • Deriving information about the item's speed from the phase shift
  • Difference between the speed measured by ultrasound and the speed measured by RFID, which indicates which ultrasound measurement belongs to which chip. If both are (nearly) the same, they likely belong to the same device, indicating that it must be close to the terminal since no machting ultrasound measurement would be available otherwise.

To predict whether an item is moving towards the terminal or not, an XGBoost model with a binary outcome was trained. The model provided as output the probability that a certain object would move towards the terminal.

To verify plausibility, the relevance of the features was calculated using Explainable AI. The results confirmed the insights Wattro had already gained during the previous development, for example:

  • The speed derived from the RFID data was relevant to detect whether an item was moving.
  • The difference between ultrasound and RFID speed was particularly important to distinguish whether an item was actually moving towards the terminal or at the back of the warehouse.
  • The signal strength, on the other hand, had little relevance because it exhibited high inaccuracy due to factors such as the positioning of the chip on the device.


With the help of the machine learning model, the accuracy of detection could be increased from 90% to 99%. This was a necessary prerequisite for Wattro's business model. The project showed with manageable effort that the machine learning approach is very well suited for increasing the accuracy of recognising device handover. After the project was completed, the code was handed over from INWT to Wattro, enabling Wattro to use the model in live operations, train it on additional data, and further develop it.