MLP-based Error Compensation Method for Single-Direction Survey Lines in Land Vehicle Strapdown Gravimetry

In land vehicle-borne dynamic strapdown gravimetry, horizontal accelerometer biases project onto the navigation frame through the time-varying heading angle, producing systematic errors in gravity disturbance estimation. Due to the inherent heading instability of ground vehicles, these bias-induced errors exhibit low-frequency, continuous, and heading-correlated characteristics along the survey line. The conventional forward-backward fusion method exploits the mirror symmetry of repeated lines to cancel such errors, but at the cost of halving the effective survey coverage and precluding single-pass operation.

To overcome this limitation, this study proposes a MLP-based (multilayer perceptron) compensation approach that directly learns the mapping from vehicle motion states to the systematic gravity estimation error. The input features include the forward-only gravity disturbance (east and north), heading representation (sine and cosine of yaw), speed, and yaw rate. The supervision target is defined as the residual between the forward-only solution and the forward-backward fused reference, which inherently encodes the heading-dependent bias effect. A compact two-hidden-layer MLP (32 neurons each, ReLU activation) is trained with mean squared error loss and early stopping.

Experiments on a vehicle-borne gravimetry dataset (4782 samples, 70%/30% sequential split) show that the proposed method reduces the east-component RMSE from 1.188 mGal to 0.188 mGal (84.1% improvement) and the north-component RMSE from 0.478 mGal to 0.134 mGal (72.1% improvement). The compensated results closely approximate the fused reference, confirming that the MLP effectively learns the slowly varying and heading-correlated error characteristics.