【正文】 agnosis sensor fault, by paring the output of work and the corresponding sensor output to detect sensor the difference between a sensor measurement and its estimation exceeds the threshold while the differences of other sensors with their corresponding estimation ( e. g. relatively low) , then a sensor fault is declared to happen. Once a faulty sensor measurement is detected, it will be disconnected from the input layer of work. However, the neural work will continue to function by using the most recent corresponding output of the NN as input instead of the faulty sensor measurement, because the most recent output is a good estimation of the faulty sensor measurement when there are enough information on input variables. And AANN has the ability of fault tolerance for the fact that the disturbance from input nodes can be distributed to the work and has little impact on controller will be switched to the estimated value to continue the system operation. Under this scheme, the system can remain operable even with multiple sensors faults as long as normal sensors are not less than bottleneck nodes. The ability to bine detection, isolation and acmodation in one step is the key advantage of AANN based sensor validation scheme. This ability is based on the dimensionality reduction property of AANN. There will be performance degeneration or installation and manufacture tolerance which are the sources of uncertainties and will cause estimation error in the optimal estimation of AANN. These uncertainties may be taken for sensor fault or vice verse. If the degeneration is taken for sensor fault,fault will be wrongly warned, causing incorrect fault acmodation. And if sensor fault is taken for degeneration, it will cause incorrect work pensation. Fault control gain together with soft fault detection logic is developed to distinguish optimal estimation error from sensor faults in this paper. Axial directional fault signature is used to identify the cause of optimal estimation error. If the residual is caused by optimal estimation error, then the weights and biases of AANN will be pensated online. If the residual is caused by sensor fault, then corresponding estimation is used to replace the failed sensor, providing analytical redundancy. In the fault acmodation logic the fault control gains are used to provide a smooth transition from the failing sensor to its corresponding estimation. 4 Example of Digital Simulation Lets take a turbo shaft engine for example . shows the closed loop control system of a engine system consisting of the engine, controller and AANN based sensor fault diagnosis. The primary variables of interest are, ng, np, Tt45, Ps3, Ml and WfB. np is the given speed of power turbine, and Ml and npg are inputs. The control feedback variables are ng, np, Tt45, when the input variables of AANN are correlative, the valid feature of the variables can be extracted from the bottleneck layer. The covariance matrix of 6 variables, ng, np, Tt45, Ps3, WfB,Ml can reflex the correlation: Close loop control system of a turbo shaft engine R= ng Tt45 Ps3 WfB np Ml ng Tt45 Ps3 WfB np Ml When | Rij | in the covariance matrix R, the ith and jth variables are considered uncorrelated. As discussed above, 2 nodes are selected in bottleneck layer. Using cross validation theory, AANN is trained with mapping nodes of 8 16 respectively. When a sensor fault is declared, it is cut off from the input of AANN. The input of work is then replaced by the last estimation of the sensor,and the work can still works well. ERS is adopted to detect sensor faults for the realtime requirement. Considering the robustness to uncertainty, AANN is trained with normal data and pensated online. The procedure of sensor fault diagnosis and reconstruction of engine control system based on AANN is (1) Collect data from test or simulation and use normal data to train AANN offline. Relative variables are adopted in this paper. (2)Adopt