Nonlinear zero drift in force sensors is caused by factors such as non-uniform temperature heating of strain gauges during actual operation and temperature imbalance caused by self-heating of the acquisition and amplification module. To achieve effective compensation, this research puts forward a temperature prediction and compensation solution integrating VMD decomposition and the CNN-LSTM-Attention model. The workflow is structured as: first, applying VMD to decompose filtered data; then, using CNN to extract local features, LSTM to capture time-series dependencies, and the learnable attention mechanism to focus on temperature mutation points—measures that effectively elevate compensation precision.

deep learning; VMD; Temperature compensation； Learnable attention