Feasibility of a Neural Network-Based Virtual Sensor for Vehicle Unsprung Mass Relative Velocity Estimation

Area of Science:

• Automotive Engineering
• Artificial Intelligence
• Sensor Technology

Background:

• Automated driving systems rely heavily on advanced sensing technologies.
• Virtual sensors offer a solution for obtaining difficult-to-measure or expensive data through data fusion.
• Accurate measurement of vehicle unsprung mass relative velocity is critical for suspension control.

Purpose of the Study:

• To develop and evaluate a virtual sensor for estimating vehicle unsprung mass relative velocity.
• To assess the effectiveness of Bidirectional Long-Short Term Memory (BiLSTM) neural networks for this regression task.

Main Methods:

• A neural network model, specifically BiLSTM, was trained using simulation data from an IPG Carmaker vehicle model.
• An extensive dataset covering 26 scenarios was utilized for training, validation, and testing.
• Bayesian Search optimized the neural network architecture, with Root Mean Square Error (RMSE) as the performance metric.

Main Results:

• The optimal neural network comprised 167 BiLSTM units, 256 fully connected hidden units, and 4 output units.
• Error histograms and spectral analysis confirmed the high accuracy of the predicted signals compared to reference data.
• The developed virtual sensor demonstrated strong applicability in estimating vehicle unsprung mass relative velocity.

Conclusions:

• Neural network-based virtual sensors are highly effective for estimating vehicle unsprung mass relative velocity.
• This technology holds significant promise for enhancing automated driving and vehicle suspension control systems.