An Assessment of a Sensor Network Using Bayesian Analysis Demonstrated on an Inlet Manifold

Modern control strategies for internal combustion engines use increasingly complex networks of sensors and actuators to measure different physical parameters. Often indirect measurements and estimation of variables, based off sensor data, are used in the closed loop control of the engine and its subsystems. Thus, sensor fusion techniques and virtual instrumentation have become more significant to the control strategy. With the large volumes of data produced by the increasing number of sensors, the analysis of sensor networks has become more important. Understanding the value of the information they contain and how well it is extracted through uncertainty quantification will also become essential to the development of control architecture. This paper proposes a methodology to quantify how valuable a sensor is relative to the architecture. By modelling the sensor network as a Bayesian network, Bayesian analysis and control metrics were used to assess the value of the sensor. This was demonstrated on charge mass flow estimation in the inlet manifold. Four control architectures modelled using a Bayesian network were compared: balanced sensors, redundant sensors, synergistic sensors and unbalanced sensors. The assessment metrics included uncertainty propagation, area of one sigma ellipses and the relative gain in information entropy of the estimated variable. The unique uncertainty characteristics of each case were identified using these assessment metrics, allowing for direct comparison between the architectures. Multivariate analysis by Gaussian modelling of the covariance matrix of the model was also performed. These results were used to quantifiably assess how each sensor and variable affects the charge mass flow estimation.