A PARAMETRIC INVESTIGATION OF THE DIAGNOSTIC ABILITY OF PROBABILISTIC NEURAL NETWORKS ON TURBOFAN ENGINES

Transcription

1 A PARAMETRIC INVESTIGATION OF THE DIAGNOSTIC ABILITY OF PROBABILISTIC NEURAL NETWORKS ON TURBOFAN ENGINES C. Romessis Research Assistant A. Stamatis Research associate K. Mathioudakis Assistant Professor Laboratory of Thermal Turbomachines National Technical University of Athens 1

2 A PARAMETRIC INVESTIGATION OF THE DIAGNOSTIC ABILITY OF PROBABILISTIC NEURAL NETWORKS ON TURBOFAN ENGINES Formulation of the diagnostic problem Features of Probabilistic Neural Networks Study of factors affecting diagnostic performance of PNN oeffect of Training Set oeffect of Measurements noise oeffect of the severity of a fault oeffect of the operating conditions oeffect of the selected diagnostic parameters Summary -Conclusions 2

3 Formulation Of The Diagnostic Problem (I) OBJECTIVE: PNN to diagnose component faults of High-by-Pass ratio, partially mixed, TURBOFAN engine Layout of a turbofan engine and station numbering for positions of interest. 3

4 Formulation Of The Diagnostic Problem (II) Information Provided Quantities defining operating Condition (4) Ambient Conditions Flight Speed Set Point Variable (Fuel flow rate) Quantities for Deducing Engine Condition (7) Shaft speeds Pressure, Temperatures Information Required Condition of individual engine components Kind, Location, Severity (magnitude) of faults 4

5 PNN Generation (I) Architecture Input Layer: Each node represents a measured quantity (Inputs :ΔΥ=[(Y-Yref)/ Yref ] 100%) Hidden Layer: Each node represents a training pattern (a-priori given information) Output Layer: Each node represents a fault parameter (probability values derived by NN) 5

6 Structure of a Probabilistic Neural Network Input Layer XNLP XNHP P13 P3 T3 T5 T Hidden Layer m-4 m-3 m-2 m-1 m SW12 SE12 SW2 SE2 SW26 SE26 SW41 SE41 SW49 SE49 A8IMP Output Layer 6

7 PNN Generation (II) A priori information Training Patterns Generated by Engine Performance model Reference information ( healthy -Y ref ) Faulty Engine Information ( Y ) 7

8 Ambient Pressure Turbofan Engine Modeling Quantities defining the operating Conditions (4): Fuel consumption Engine Inlet Conditions (pressure, temperature) Fault Parameters (11): Flow factors of positions of interest: SW i = W i p i T i W i p i T i ref Efficiency factors of positions of interest: SE i = (η η i i ) ref Exhaust area (numbered station 8) Measured quantities (7): Shafts speed (low and high pressure) Pressures and temperatures at stations along the engine 8

9 Creating Fault Patterns (Containing noise) + Noise σ u u Procedure for generating noisefree patterns by EPM and adding noise on them u* EPM Faulty operation; some f i 0 Healthy operation; all f i =0 Noise σ Y EPM Y + Y o * Y * Y=(Y * -Y o * )/Y o * 100 9

10 Generating the Training Testing patterns Patterns (generated by EPM) Deltas of measurements for a specific set of deltas of the fault parameters Considered faults 11 Single faults: Delta of 1 fault parameter deviates within a range of +/- 3% 5 Combined faults: Deltas of 2 fault parameters, of the same module,deviate. * Delta of a value Υ: ΔΥ=[(Y-Yref)/ Yref ] 100% 10

11 Noise Situations Considered (reproduce realistically the data collected from an actual engine) 5 noise levels considered noise level noise in u noise in y 1 0 σ y /2 2 0 σ y 3 0 3σ y/ σ y 5 σ u σ y σu: noise due to Operating Condition uncertainty σy: noise due to sensor inaccuracies 11

12 Considered Probabilistic Neural Networks ID No. of input nodes Operating point Training patterns Noise level included No. of training patterns No. of output nodes 1 7 cruise No noise st,2nd, 2 7 cruise 3rd, 5th cruise 5th level st,2nd, 4 7 take-off 3rd, 5th take-off 5th level cruise 5th level cruise & take-off 5th level

13 Effect of noise level on the training patterns score (%) Fault ID Noise-free training all noise levels 1 noise level (σu,σy) 13

14 Performance of a network tested on patterns partially used for training and on non-training patterns score (%) fault ID set partially trained set not trained 14

15 Effect of noise level in the patterns presented to the network. Network trained using data from all noise situations score (%) fault ID Noise-free (0,σy/2) (0,σy) (0,3σy/2) (σu,σy) 15

16 Effect of magnitude of fault on the diagnostic performance of PNN score (%) fault ID f <1% 1%< f <2% 2%< f 16

17 Effect of noise magnitude on test patterns 120 Fault 3: SW2 Fault 8: SE Score (%) (0,0) (0,σy/2) (0,σy) Noise (0,3σy/2) (0,2σy) fault:3, f <1% fault:3, 1%< f <2% fault:3, 2%< f fault:8, f <1% fault:8, 1%< f <2% fault:8, 2%< f 17

18 Diagnostic performance of three networks, each one trained and tested at a different operating point score (%) fault ID cruise data take-off data mixed data 18

19 Diagnostic performance of two networks with different output configuration score (%) Fault ID individual parameters output component condition output 19

20 Conclusions The influence of noise is important. Training with noisy data improves the performance. When a network is used, the performance is better for less noise in the data. Filtering procedure of measurements before they are fed to such a network may thus be beneficial. Possibility of correct detection of a fault is better for more severe faults, namely when the magnitude of deviation is bigger. It is safer to make diagnosis based on observations at higher power settings (e.g. take-off versus cruise conditions). 20