Quality prediction of tractor rotary tillage based on BiConvLSTM with self-attention

Abstract

To accurately predict rototilling performance and rotary tillage quality based on multi-sensor measured data of tractor electro-hydraulic suspension system, an improved ConvLSTM-based model is proposed, and field tests of rototilling operation are carried out to verify the accuracy. The model is based on the SA-BiConvLSTM (Self-attention Bi-directional ConvLSTM, BiConvLSTM) network as the core, and the BiConvLSTM can take the time dependence of the load sequence’s contextual state information into account and extract the local features more deeply and accurately; the self-attention mechanism strengthens the inner correlation of long-range load features, it significantly reduces the number of model parameters, solving the problems of degraded prediction performance and low computing efficiency caused by the length of the load sequence. The experimental results showed that the accuracy and F1-score of the rotary tillage quality prediction based on the measured data reached 97.40% and 97.47%, respectively, and the model had a considerable improvement in the prediction efficiency with the highest increase of 14.09%, and all evaluation metrics were better than the experimental results of the control group. This model ensures the breadth, depth and correlation of the load feature, significantly simplifies the model complexity and directly improves the prediction effect and computing efficiency. This study explores an intelligent technological approach for predicting the quality of rotary tillage, providing a new technological reference for the research and application of precision agriculture.