A weigh-in-motion (WIM) system continuously and automatically detects an object's weight during transmission. The WIM system is used widely in logistics and industry due to increasing labor and time costs. However, the accuracy and stability of WIM system measurements could be affected by shock and vibration under high speed and heavy load. A novel six degrees-of-freedom (DOF), mass-spring damping-based Kalman filter with time scale (KFTS) algorithm was proposed to filter noise due to the multiple-input noise and its frequency that is highly coupled with the basic sensor signal. Additionally, an attention-based long short-term memory (LSTM) model was built to predict the object's mass by using multiple time-series sensor signals. The results showed that the model has superior performance compared to support vector machine (SVM), fully connected network (FCN) and extreme gradient boosting (XGBoost) models. Experiments showed this improved deep learning model can provide remarkable accuracy under different loads, speed and working situations, which can be applied to the high-precision logistics industry.
Keywords: Kalman filter; deep learning; time-series analysis; weigh-in-motion.