Orientation-Free Neural Network-Based Bias Estimation for Low-Cost Stationary Accelerometers

TL;DR

This work tackles the bias calibration problem for low-cost MEMS accelerometers by removing the need for known orientation or multi-position calibration. It introduces OFBENet, a lightweight orientation-free 1D CNN that maps raw stationary accelerometer data to a three-axis bias vector, trained with MSE and evaluated on gravity-aligned and rotated datasets totaling $13.39$ hours across six devices. Across five-fold cross-validation, OFBENet achieves the lowest RMSE and maximum bias errors, with improvements up to $>77\%$ on rotated data and $>52\%$ overall compared to model-based baselines, and statistical tests confirming significance. The method eliminates the need for leveling or complex orientation procedures, enabling rapid, field-ready bias calibration for low-cost inertial sensors in robotics, navigation, and industrial applications.

Abstract

Low-cost micro-electromechanical accelerometers are widely used in navigation, robotics, and consumer devices for motion sensing and position estimation. However, their performance is often degraded by bias errors. To eliminate deterministic bias terms a calibration procedure is applied under stationary conditions. It requires accelerom- eter leveling or complex orientation-dependent calibration procedures. To overcome those requirements, in this paper we present a model-free learning-based calibration method that estimates accelerometer bias under stationary conditions, without requiring knowledge of the sensor orientation and without the need to rotate the sensors. The proposed approach provides a fast, practical, and scalable solution suitable for rapid field deployment. Experimental validation on a 13.39-hour dataset collected from six accelerometers shows that the proposed method consistently achieves error levels more than 52% lower than traditional techniques. On a broader scale, this work contributes to the advancement of accurate calibration methods in orientation-free scenarios. As a consequence, it improves the reliability of low-cost inertial sensors in diverse scientific and industrial applications and eliminates the need for leveled calibration.

Figures (10)

• Figure 1: Error magnitude as a function of roll and pitch orientation angles. The red contour indicates the 50mg bias threshold, showing that for combined orientations within approximately ±2.87°, the orientation-induced error remains smaller than a typical accelerometer bias.
• Figure 2: Overview of the proposed calibration method. The pipeline begins with a stationary orientation-free accelerometer readings, which is fed into our simple, yet efficient, neural network OFBENet. The output of the network is the accelerometers deterministic bias vector.
• Figure 3: OFBENet architecture used for bias estimation. Each convolutional block includes a 1D convolutional layer, batch normalization, Leaky ReLU activation, and average pooling. The network concludes with global average pooling, a fully connected layer, dropout, and an output layer.
• Figure 4: Four SparkFun ADXL345 accelerometers used in the gravity-aligned dataset, placed on a leveled surface.
• Figure 5: Experimental setup used for recording the rotated accelerometer dataset. (a) Top view of the 3D-printed plate showing the fixed alignment of the IMUs. The sensor on the right is the Memsense MS-3025, and the two sensors on the left are SparkFun ADXL345 IMUs connected to the SparkFun development board. (b) Experimental setup mounted on a tripod used for orientation adjustment during data collection.