A novel neural network with variable-activation-function for analytical redundancy of variable cycle engine sensors

Analytical redundancy based on data-driven technology is an effective method of reconstructing faulty sensors in the aeroengine control system. Compared to traditional aeroengines, variable cycle engine (VCE) has more control parameters and more complex analytical relationships among parameters. This fact requires higher nonlinear expression ability of data-driven analytical redundancy models which have more input variables and deeper layers, resulting in slower calculation speeds. To enhance the nonlinear expressive ability of algorithms and simultaneously reduce calculation time, variable-activation-function neural network (VANN) is proposed to construct the analytical redundancy for VCE sensors. The activation function of VANN is variable in response to the time-varying analytical relationships. Moreover, the weights and biases of VANN are input functions, and the biases can be utilized to rapidly eliminate the error between the actual sensor output and the estimated value at the terminal node. Therefore, the nonlinear expression ability of VANN is enhanced while reducing the network parameters. Compared with other neural networks, VANN has the lowest Relative Root Mean Square Error for estimating various sensor signals. Besides, for an aeroengine control system, the calculation time of VANN is 0.73 ms, far less than each interval of the control step, which is usually around 20 ms.