A machine learning model for dynamic payload estimation of a wheel loader

Abstract

Wheel loader scales require high accuracy under challenging operating conditions. Many of the current industry solutions rely on proximity switches placed at specific positions along the boom path to determine when weight measurements should be taken. This is not ideal because proximity switches are difficult to install and induce more possible breakpoints. This thesis presents a more robust solution that is not dependent on proximity switches while also delivering superior accuracy. This thesis was done in collaboration with Tamtron, a manufacturer of industrial scales and weighing solutions.

Sensor data was collected from a Volvo L220 wheel loader using pressure sensors and inertial measurement unit (IMU) sensors. The sensors were connected to a controller area network (CAN) interface. The captured data includes boom cylinder pressure, boom angle, and boom acceleration. Data was collected during lifting operations with known payloads in the range of 0 kg – 15 000 kg. Two different models for payload estimation were developed: a neural network approach and a novel angle-interval sampling technique that uses LASSO regression. The angle-interval method collects sensor data from a specific boom angle range at 1.25° intervals, transforming the otherwise non-linear relationship between the sensor data and payload into an approximately linear one.

The angle-interval regression model significantly outperformed the neural network model, achieving a mean absolute error of 17.9 kg and a root mean square error of 24.6 kg. This accuracy exceeds the industry requirements, which require a maximum permissible error of 25 kg – 75 kg, depending on the payload. The model is computationally efficient for microcontroller implementation and can be easily calibrated across different wheel loader models. These things make the regression model a potential solution for Tamtron’s next-generation, manufacturer-agnostic wheel loader scales.