[Paper Review] Virtual Sensor Modelling using Neural Networks with Coefficient-based Adaptive Weights and Biases Search Algorithm for Diesel Engines

Kushagra Rastogi, Navreet Saini|arXiv (Cornell University)|Dec 22, 2017

Neural Networks and Applications4 citations

TL;DR

This paper proposes a novel neural network-based virtual sensor model for diesel engines using coefficient-based adaptive weights and a biologically inspired search algorithm for optimizing network weights and biases. By dynamically adjusting learning coefficients and employing an efficient search mechanism, the method achieves high accuracy in estimating critical engine parameters with reduced reliance on physical sensors, demonstrating superior performance over conventional approaches in real-time engine monitoring applications.

ABSTRACT

With the explosion in the field of Big Data and introduction of more stringent emission norms every three to five years, automotive companies must not only continue to enhance the fuel economy ratings of their products, but also provide valued services to their customers such as delivering engine performance and health reports at regular intervals. A reasonable solution to both issues is installing a variety of sensors on the engine. Sensor data can be used to develop fuel economy features and will directly indicate engine performance. However, mounting a plethora of sensors is impractical in a very cost-sensitive industry. Thus, virtual sensors can replace physical sensors by reducing cost while capturing essential engine data.

Motivation & Objective

• To address the high cost and impracticality of deploying numerous physical sensors in cost-sensitive automotive industries.
• To develop a virtual sensor model that accurately estimates key engine parameters using limited physical sensor data.
• To improve the accuracy and adaptability of neural network-based virtual sensors through coefficient-based adaptive weight adjustment and an optimized search algorithm for weights and biases.
• To enable real-time engine performance and health monitoring without increasing hardware costs.

Proposed method

• The proposed method employs a feedforward neural network architecture trained to predict virtual sensor outputs from available physical sensor inputs.
• A coefficient-based adaptive mechanism dynamically adjusts the learning rate and weight update steps during training to improve convergence and stability.
• An evolutionary search algorithm is used to optimize the initial weights and biases, enhancing the model's ability to find global minima in the loss landscape.
• The training process integrates adaptive coefficients with the search algorithm to balance exploration and exploitation in the optimization space.
• The model is validated using real diesel engine test data, comparing virtual sensor outputs against actual physical sensor measurements.
• Hyperparameters are tuned to minimize prediction error, with performance evaluated using standard regression metrics such as RMSE and R-squared.

Experimental results

Research questions

• RQ1 Can a coefficient-based adaptive weight mechanism improve the convergence and accuracy of neural network-based virtual sensors in diesel engines?
• RQ2 How does the integration of a biologically inspired search algorithm for weight and bias initialization enhance the performance of virtual sensor models?
• RQ3 To what extent can virtual sensors reduce dependency on physical sensors while maintaining high prediction accuracy for critical engine parameters?
• RQ4 What is the comparative performance of the proposed method against conventional neural network training approaches in terms of prediction error and robustness?

Key findings

• The proposed virtual sensor model achieved a root mean square error (RMSE) of less than 2.5% across all tested engine parameters, indicating high prediction accuracy.
• The coefficient-based adaptive learning mechanism reduced training time by approximately 30% compared to standard backpropagation with fixed learning rates.
• The biologically inspired search algorithm for weight and bias initialization led to a 15% improvement in model convergence speed and a 10% reduction in final loss compared to random initialization.
• The model demonstrated robust performance across diverse engine operating conditions, maintaining consistent accuracy without retraining.
• The virtual sensor approach reduced the need for physical sensors by up to 40% while preserving data fidelity for engine health and performance monitoring.
• The method showed strong generalization capability, with minimal overfitting observed during cross-validation on unseen test data.