AU2022358641A1

AU2022358641A1 - Spike echo state network model for fault prediction of aeroengine

Info

Publication number: AU2022358641A1
Application number: AU2022358641A
Authority: AU
Inventors: Hongxin Li; Moran LIU; Tao Sun; Ximing SUN
Original assignee: Dalian University of Technology
Current assignee: Dalian University of Technology
Priority date: 2022-08-30
Filing date: 2022-09-26
Publication date: 2024-03-14
Also published as: WO2024045246A1; CN115481658A

Abstract

The present invention relates to a spike echo state network model for fault prediction of an aero engine and belongs to the technical field of fault diagnosis of an aero-engine. 1)Various sensor signals of aircraft flight sorties are acquired, the appropriate sample characteristics of data are selected, the state data is preprocessed, and data segments are divided according to fault labels to form a sample set; 2) a spike echo state network comprising a spike input layer, a spike reservoir and a training output layer is constructed; 3) a spike echo state network model is trained by the sample set of aero engine data, and the prediction results of a training set and a test set are calculated by using the trained spike echo state network model; and 4) the prediction results are provided for aero-engine detection and early warning equipment. The present invention can automatically give the future development trend of the operating state according to the operation law of the aero-engine and accurately predict the changes of each performance parameter index of the aero-engine in a short term to assist the crew to judge whether an aero-engine will fail, so as to make the fault early warning more intelligent.

Description

SPIKE ECHO STATE NETWORK MODEL FOR FAULT PREDICTION OF AERO ENGINE Technical Field The present invention belongs to the technical field of fault diagnosis of an aero-engine and relates to a spike echo state network model for surge fault prediction of an aero-engine. Background As the core power source of civil or military aircraft, the aero-engine is a highly complex and precise pneumatic-thermal-mechanical system, of which the working state has direct or indirect influence on safety, reliability, economical efficiency and other performance of aircraft. Because the aero-engine works in a high-altitude adverse environment with high temperature, high pressure and strong vibration for a long time, the probability of failure is also increased, which will affect the working performance and even cause adverse flight accidents. In fact, the surge fault is the most representative and destructive type of aero-engine faults. When an aircraft flies in extreme conditions, the compressor of the aero-engine outputs lower pressure than the downstream of the system due to external influence, which causes high-pressure gas to flow backwards, resulting in violent vibration and hot-end overheating of the engine, which may quickly destroy the aero engine, thus causing serious consequences such as in-flight shutdown of the aircraft. Due to strong nonlinear and high-dimensional characteristics of the aero-engine, the fault early warning and prediction of the aero-engine often relies on the empirical criteria of the crew and the ground researchers in actual operation. However, it is often impossible to accurately realize fault monitoring based on such empirical estimation because of subjectivity, and as a result, aircraft operation faults cannot be judged or dealt with in time. Therefore, it is of great practical significance to predict the potential fault factors of an aero-engine scientifically and stop the surge fault in advance for the safe, stable and efficient operation of aircraft. The aero-engine fault diagnosis technology can be divided into a mechanism modeling method and a data-driven method, and the mechanism modeling method is the main research idea of early fault diagnosis of the aero-engine. The mechanism method is mainly to establish an accurate mathematical physical model based on the aero-thermodynamic characteristics of the aero-engine and apply the actual observed value of the aero-engine sensor to the model to calculate the corresponding parameter estimation and judgment results. Literature [1]={Urban L A . Gas Path Analysis Applied to Turbine Engine Condition Monitoring[J]. Journal of Aircraft, 1973, (7):400-406.} first proposes a fault influence coefficient matrix for aero-engine parameter estimation, which is still applied in various traditional fault diagnosis systems. Literature

[2]={Davison C R, Birk A M. Development of Fault Diagnosis and Failure Prediction Techniques for Small Gas Turbine Engines[C]// Asme Turbo Expo: Power for Land, Sea, & Air. 2001.} proposes a fault map method, in which each state of the aero-engine corresponds to a point or a region in the fault map, and the fault region is qualitatively classified according to the empirical criteria to determine the fault type. In addition, some scholars propose a variety of diagnosis methods based on the nonlinear Kalman filter algorithm for the Gaussian noise environment and nonlinear properties of aero-engine parameters. For example, literature [3]={Han D . A Study on Application of Fuzzy Adaptive Unscented Kalman Filter to Nonlinear Turbojet Engine Control[J]. International Journal of Aeronautical & Space Sciences, 2018.} proposes a fuzzy adaptive unscented Kalman filter algorithm. Literature [4]={Feng L, Gao T, Huang J, et al. A novel distributed extended Kalman filter for aircraft engine gas-path health estimation with sensor fusion uncertainty[J]. Aerospace Science and Technology, 2018, 84.} proposes a novel Kalman filter by using the multi-step recursive estimation strategy and a self-tuning buffer in the case of gas path measurement uncertainty. However, a real aero-engine is a highly complex nonlinear system, so the fault diagnosis method based on mechanism modeling has high requirements for the design accuracy of the model. The data-driven fault diagnosis methods do not need to obtain an accurate mathematical model, but only need sufficient judgment experience and historical data. So far, the data-driven fault diagnosis methods mainly include artificial neural network, support vector machine (SVM), extreme learning machine (ELM) and hidden Markov process. Literature

[5]={Zhao Y P, Wang J J, Li X Y , et al. Extended least squares support vector machine with applications to fault diagnosis of aircraft engine[J]. ISA transactions, 2019, 97:189-201.} proposes an extended least square support vector machine which is proved to have the measuring ability of aero-engine fault diagnosis. Literature [6]={Yang X, Pang S, Shen W, et al. Aero Engine Fault Diagnosis Using an Optimized Extreme Learning Machine[J]. International Journal of Aerospace Engineering, 2016, (2016-1-26), 2016, 2016(pt.1):1-10.} proposes an extreme learning machine based on quantum-behaved particle swarm optimization, which is applied to diagnosis of gas turbofan engines. Compared with these machine learning methods, the artificial neural network is considered as one of the most potential diagnostic tools because of the advantages in nonlinear fitting. With the emergence of the recurrent neural network (RNN), the time series prediction analysis method based on historical data provides a new solution for fault diagnosis and prediction. Literature [7]={Mei Y, Wu Y, Li L . Fault diagnosis and remaining useful life estimation of aero engine using LSTM neural network[C]// IEEE International Conference on Aircraft Utility Systems. IEEE, 2016.} obtains good performance in aero-engine diagnosis and prediction through a long and short term memory (LSTM) neural network under the conditions of complex operation, hybrid faults and strong noise. However, the general recurrent neural network is not good in aero engine fault prediction because the short-term memory capacity is too small to solve the problem of long-term dependence. The present invention proposes a new aero-engine fault diagnosis and prediction model spike echo state network (Spike-ESN), which can accurately predict the surge fault data of an aero engine. The preset invention is funded by the China Postdoctoral Science Foundation funded project (2022TQ0179) and the projects (61890920, 61890921) funded by the National Natural Science Foundation of China. Summary To solve the problems in the prior art, the present invention provides a spike echo state network model for fault prediction of an aero-engine and belongs to the field of fault diagnosis of an aero-engine. To achieve the above purpose, the present invention adopts the following technical solution: The present invention relates to a spike echo state network model for fault prediction of an aero-engine. The method comprises the following steps: acquiring various sensor signals of aircraft flight sorties, and selecting the appropriate sample characteristics of data; preprocessing the state data by means of denoising, normalizing and resampling, and dividing data segments according to fault labels to form a sample set; constructing a spike echo state network comprising a spike input layer, a spike reservoir and a training output layer; training a spike echo state network model by the sample set of aero-engine data, and calculating the prediction results of a training set and a test set by using the trained spike echo state network model; and providing the prediction results for the detection and warning equipment of an aero-engine. Specifically: Step Sl: selecting the sample characteristics of data Firstly, a plurality of groups of sensors on an aero-engine are used to collect various operating state data of aircraft flight sorties; each type of operating state data represents a characteristic variable, with the data form of one dimension of data series; then, the fault time in the collected data series is determined by the traditional empirical criterion method to list fault labels in the series; secondly, the fluctuation of each parameter data near the fault label of each dimension of data is observed by the MATLAB data analysis method; when the change rate of the difference quotient of the data is greater than 5%, it can be considered that this dimension of data changes significantly with the occurrence of a fault, and the characteristic variable represented by this dimension of data is highly correlated with the aero-engine fault; and some dimension of collected data series is denoted as [s(1),s(2),..., s(t)]T, t is the sampling time of this dimension of data, and

-1)]- [s(t -1)- s(t -2)] the change rate of the difference quotient can be expressed as [s(t) - s(t s(t -1)- s(t - 2)

and finally, characteristic variables which are highly correlated with the occurrence of faults during aero-engine operation are selected for fault prediction; Step S2: preprocessing the samples In view of various fault types generated during aero-engine operation, multiple data sets are respectively established by using the characteristic variable data selected in Step 1, so as to construct a sample set for the spike echo state network model for various fault predictions, specifically: Firstly, the characteristic variable data selected in Step 1 is preprocessed by means of denoising, normalizing and resampling; then, multiple data segments containing the same type of fault are divided from the data according to the fault labels listed in step 1, the sampling times of each data segment before and after a fault occurs must be at least 4000, and multiple data segments belonging to different fault types can be obtained by repeating operations for each fault type; secondly, each data segment belonging to each fault type is labeled with basic information, including flight sorties, names of sample characteristics, operating time and fault labels, and stored in a database; and finally, aiming at a certain type of fault, a data segment belonging to the fault

type is selected from the database as a training sample set Xt,,i,, of the spike echo state network

model, and one of the remaining data segments is selected as a test sample set Xt;

Step S3: constructing the spike echo state network model The present invention improves the traditional spike echo state network (Spike-ESN) by adding a spike encoding mechanism of brain-like computing into the input layer to form a spike input layer; and adding a spike activation function into the reservoir to form a spike reservoir. The above improvements can better mine the hidden spatio-temporal information in the data and improve the long-term memory ability of the model; that is, the spike echo state network (Spike ESN) adopted by the present invention is composed of a spike input layer, a spike reservoir and a

training output layer. The training sample set X,,i,, obtained in step 2 is used to establish a model,

and the specific construction process is as follows: (1) Spike input layer The input signal and the output signal of the spike echo state network at the time t are

respectively denoted as u(t) and y(t); each dimension of data in the training sample set X,,i,, is

used for modeling; one dimension of data is used here for illustration, this dimension of data series

is denoted as [a(1), a(2),..., a(t)]T , and t is the sampling time of this dimension of data; when data a(t) is defined to predict data a(t+i), the prediction step is i , and then in the spike echo state network model, when the prediction step is i, the input signal u(t) is a(t), and the output signal y(t) is a(t+i); The spike input layer has the function of converting the input signal u(t) to a spike train, which can be expressed as formula (1); fin (U (t))= [ui (t), U2(t),..., Iut" (t)]T (1)

In formula (1), f,() is the transformation function of the spike input layer, and trime is the spike

sampling times of a spike neuron for the same value; after the spike sampling times of the Spike ESN network are specified, each input data will be transformed into a spike train of equal length, the format of the spike train has the length of spike sampling times, and each element is 1 or 0, representing activation or suppression; The Poisson distribution is used to generate a spike train for values, and is shown in formula (2); 2k P(x =k)=- e- (2) k! In formula (2), k represents the occurrence times of events; A is the population mean and variance of the Poisson distribution, and the practical significance of the distribution can be expressed as that when events are observed to occur A times on average, the actual probability of occurrence is k ; and the average value A of intervals to generate spike is determined by the value of the input data, and a spike train is generated according to the Poisson distribution.

max(u(t))- u(t) timemax(u(t))- min(u(t))

In formula (3), r(t) represents the average interval for u(t) to generate spike. r(t) is substituted t into formula (2), time intervals, i.e., [rc(t),K 2 (t),...,r, (t)], are randomly generated according to

the Poisson distribution, and spike trains are generated at intervals for each input data;

u()= W= ,15i time ke Z+(4) L0, otherwise

In formula (4), ui(t) is an element in a spike train; k represents the order of a spike element

"1"; and i represents the position of the spike element "1" in the spike train; (2) Spike reservoir The spike reservoir in the present invention is a sparse network composed of a plurality of neurons connected randomly, which can achieve the function of storing data by means of adjusting the internal weight of the network. Provided that the number of spike neurons in the spike reservoir is denoted as N, the internal state signal is denoted as x(t), and the random fixed internal weight is denoted as W,, the generation process of the spike reservoir is shown in formula (5);

W,., = p• (5) A max(W) In formula (5), A max(W) is the maximum eigenvalue of a matrix W ; W is a sparse matrix

randomly generated on the interval [1,1] according to uniform distribution, and the sparsity thereof is q, representing the proportion of non-zero elements in the matrix W ; the spectral radius p is

an important control parameter that determines the generation of W, and represents the upper

bound of the maximum eigenvalue of W,,; and the state transition modes of the spike reservoir

are formula (6) and formula (7); x(t)= tanh(Winfpik(u(t))+ W,,,x(t - 1)) (6)

f,pike(u(t))= exp(- " tsik)(7)

In formula (6), W,, is the input weight matrix of the reservoir, which is generated on the interval

[1,1] according to uniform distribution; t, is the time series from the beginning to the end of

observation; tspike is the spike train interval after f.,(u(t)) spike digitization; fspike () is a spike

activation function; T is the global scaling factor of the spike activation function; and x(t)

represent the internal state signal; (3) Training output layer The internal state signals are integrated to obtain a collection state matrix

X(t) =[x(1), x(2),..., x(t)]e Nrext, [. , -] represents horizontal connection between states, and the

output vector j=j1),9(2),..., j(t)]E [ NT of Spike-ESN can be expressed as formula (8);

.9' = W,,(t) (8)

In formula (8), W, is an output weight;

Step S4: training the spike echo state network model Model training is to train the model generated in step 3 by means of regression, that is to

calculate W'U, , and the objective function is formula (9); min F(W.,) =||-W.,X(t)|2 + 0|W ,|| ( 2 9

) In formula (9), 2 represents a regularization norm of L 2 , and A is a regularization coefficient;

and the objective function is solved by means of ridge regression and can be expressed as formula (10);

T W., = yX T (XX + 1)- (10) In formula (10), the regularization coefficient is generally a positive number less than 0.01; Step S5: testing and adjusting the accuracy rate of the spike echo state network The test sample set Xtest is subjected to model prediction according to formula (1-8) in step

3, the output result of formula (8) is compared with the true result, the parameters of the spike echo state network model are adjusted, and steps 1-4 are repeated to test whether the model accuracy meets the requirement; Step S6: predicting fault data by using the trained spike echo state network model The spike echo state network for early warning of various fault types is loaded into the detection and warning equipment of an aero-engine, the sensor data is input to the equipment, and the detection and warning equipment carries out real-time calculation to predict the state data of the aero-engine in the short time of next 1-2 s. The predicted data in the equipment is judged according to the preset experience criteria. If abnormal data is detected from the predicted data, the equipment will give an early fault alert and give preset operation suggestions to the crew according to the fault type. The present invention has the following beneficial effects: Compared with the prior art, the spike echo state network contains a neurodynamic mechanism, which has advantages in timing and spatial structure in processing spike signals, and has efficient and reliable performance in processing complex, sparse and noisy spatio-temporal information extraction. The echo state network has poor performance in aero-engine fault prediction due to small memory capacity and other problems, and the spike echo state network improves the memory capacity and can achieve better results in aero-engine fault prediction. The present invention can automatically give the future development trend of the operating state according to the operation law of the aero-engine and accurately predict the changes of each performance parameter index of the aero-engine in a short term to assist the crew to judge whether the aero-engine will fail, so as to make the fault warning more intelligent.

Description of Drawings Fig. 1 is a flow chart of a spike echo state network model for fault prediction of an aero engine provided by the present invention used in a fault early warning flow of an aero-engine; Fig. 2 is an architecture diagram of a spike echo state network model for fault prediction of an aero-engine; Fig. 3 shows a data set of an aircraft of a certain sortie provided by the aero-engine research

institute; Fig. 3(a) is an original data diagram of parameters D., T6, Tl and a,, Fig. 3(b) is an

original data diagram of parameters a2, PLA, H and M , and Fig. 3(c) is an original data diagram

of parameters V, N 2 , N, and Signal;

Fig. 4 shows spike encoding results of partial aero-engine data; Fig. 4(a) shows an original

data diagram of a compressor opening angle a2 , and Fig. 4(b) shows a spike diagram of partial

data of a compressor opening angle a2 ;

Fig. 5 is a comparison diagram of prediction results of six parameters in the data set of an aircraft of a certain sortie by the Spike-ESN model with the prediction step of 1; Fig. 5(a) is a

comparison diagram of a predicted value and a true value of engine exhaust gas temperature T6

by the Spike-ESN model; Fig. 5(b) is an error diagram of a predicted value and a true value of

engine exhaust gas temperature T 6 by the Spike-ESN model; Fig. 5(c) is a comparison diagram of

a predicted value and a true value of engine exhaust gas temperature T by the Spike-ESN model;

and Fig. 5(d) is an error diagram of a predicted value and a true value of engine exhaust gas

temperature T by the Spike-ESN model;

Fig. 6 is a comparison diagram of prediction results of six parameters by Spike-ESN, ESN and ARMA models with the prediction steps of 1, 10 and 20; Fig. 6(a) is a comparison diagram of

prediction results of engine exhaust gas temperature T by three models; Fig. 6(b) is a comparison

diagram of prediction results of a compressor opening angle a, by three models; and Fig. 6(c) is

a comparison diagram of prediction results of a compressor opening angle a 2 by three models.

Detailed Description To make the technical problem to be solved, the technical solution and the advantages of the present invention more clear, specific embodiments of the present invention will be described below in detail in combination with the accompanying drawings and the analysis of experimental data. The embodiment is a spike echo state network model for fault prediction of an aero-engine.

Fig. 1 shows an aero-engine fault early warning process of the model, comprising the following steps: Step SI: selecting the sample characteristics of data In order to describe the implementation process of the method in more detail, the present invention introduces the implementation process of the method in detail based on the aero-engine data provided by Dalian University of Technology and an aero-engine research institute. Experimental data is divided into seven groups, which are generated by engines in different flight sorties, and data sets contain normal data and fault data. The data contains 12 parameters, as shown in Table 1. The sampling interval of the data is 0.062 s, and each data set has about 100,000 data points. Table 1 Parameters of Data Set Parameter Parameter Dimension Name Meaning Dimension Name Meaning

1 D8 Ambient pressure 7 H Flight altitude

Engine exhaust gas 8Flight Mach temperature number

3 3 T1 Engine combustion 9 V Flight speed chamber temperature

Compressor opening High pressure rotor angle speed

Compressor opening Low pressure rotor 5 a2 11 N1 angle speed Throttle lever 6 PLA 12 Signal Failure or not stroke

To select appropriate data characteristic samples, the present invention describes in detail the

changes of each parameter in the case of surge. In the case of surge, the ambient pressure D drops

sharply from 1 to about 0.2, the engine exhaust gas temperature T has a slight fluctuation drop,

and the engine combustion chamber temperature T, slightly increases, proving that at this time,

the engine cannot effectively exhaust heated gas, which leads to increase of the engine combustion chamber temperature and may also lead to compressor overtemperature; the high pressure rotor

speed N 2 and the low pressure rotor speed N, are kept in a high state and decrease, the compressor opening angle al and the compressor opening angle a 2 change obviously at this time, the aero engine is currently in a state of rotating stall, the compressor opening angle significantly decreases to reduce the intake volume, and the surge is actively eliminated in this way. The throttle lever stroke PLA, flight altitude H, flight Mach number M and flight speed V do not change, which proves that these operating parameters and state parameters do not change and the surge is caused by the sharp drop of ambient pressure rather than operation. With the implementation of proactive measures to eliminate surge, the parameters gradually return to the normal state in the fluctuation, and the system returns to normal after the surge occurs again several times. Except for manual control parameters and state parameters such as altitude and speed which do not change obviously when the surge occurs, other aero-engine parameters can effectively represent the state change when the surge occurs, wherein the ambient pressure D, is the cause of surge, the engine exhaust gas temperature T, engine combustion chamber temperature TI, high pressure rotor speed N 2 and low pressure rotor speed N, are affected by rotating stall before the surge, and the compressor opening angle a, and compressor opening angle a 2 are parameters that actively change when the surge is eliminated. Therefore, the above parameters affected by surge can be used as prediction data. Step S2: preprocessing the samples The aero-engine state data samples selected in step 1 are processed by means of denoising, normalizing and resampling. A sliding window average method is adopted for denoising, the maximum and minimum normalization method is adopted for normalization, and the resampling frequency is 10 times the original frequency. According to the fault labels, multiple data segments containing the same type of fault are divided from the data, and the sampling times of each data segment before and after a fault occurs must be at least 4000. The experimental data contains the surge fault caused by the drop of ambient pressure. By dividing the 4000 data points before and after each fault label, multiple data segments with the length of 8000 can be obtained, and the basic information of each data label is stored in the database. Step S3: constructing the spike echo state network model

Based on the spike echo state model proposed in step 3 in Summary, the dimension N of

the reservoir of the spike echo state network is set to 100, the spectral radius p is set to 0.9, the

sparsity r is set to 0.1, and the regularization coefficient / for regression calculation is set to 10

8. To keep the echo state within the nonlinear range of a tanh activation function, the scaling factor is used to reduce the echo state as a whole, and the scaling factor is set to 0.8.

At the beginning of training, the echo state network needs to be initialized, and the first 200 data points of the data set are used for initialization, which will not be included in network evaluation. During spike digitization, the sequence length is set to 100, the observation time is set to 10000, and the overall scaling factor r of the spike activation function is set to 5000. The model construction and the subsequent model training of the spike echo state network model are realized by MATLAB. Step S4: training the spike echo state network The data points from 13000 to 20000 of the data set are used as a training set, 200 data points between 11400 and 11600 when the fault occurs are randomly and smoothly inserted into the training set to form a training set containing all operating modes of an aero-engine, and the test set contains the real data of the occurring fault between 11200 and 12500. For the data sets, the method of Poisson distribution to generate spike is adopted for encoding. After spike encoding, the larger the data is, the higher the frequency of the converted spike train is. The state collection matrix X is initialized as an empty matrix. The training data is successively input to the network, and the echo state x(t) is updated according to formula (6). Then all the updates x(t) are integrated into the state collection matrix X . Finally, the output

weight matrix W, is calculated according to formula (10), the model training is completed, and

the model can be used for prediction verification. In formula (10) of the embodiment, the value of the regularization coefficient A is 10-8. Step S5: predicting fault data by the trained spike echo state network model The error, the root mean squared error (RMSE) and the mean absolute percentage error (MAPE) are selected as the basis for testing the prediction results. The data of the test set is input to the trained spike echo state network, and the prediction effect of the model can be known by comparing the prediction results with the real results of the test set according to the calculation result of formula (8). The fault of the test set occurs within the data range of 0 to 200. When the step is 1, the model prediction error is in the magnitude of 10-4, which can accurately reflect the development trend of the aero-engine state data in a short term. When the prediction step is increased to 10, the model prediction error increases and is in the magnitude of 10-2, but the development trend of the aero engine state data still can be reflected. The spike echo state network (ESN) model achieves good effect on the whole, and can accurately reflect the future data of an aero-engine. To determine the advantages of the spike echo state network in extraction of temporal information, a comparative experiment is conducted between the echo state network model and an autoregressive moving average model. The model parameters of the echo state network model are set the same as the parameters of the spike echo state network model. The dimension Nr of the reservoir is set to 100, the spectral radius p is set to 0.9, the sparsity r is set to 0.1, the regularization coefficient A for regression calculation is set to 10-8, and the scaling factor is set to 0.8. The autoregressive moving average model (ARMA) is set to ARMA(4,4), i.e., both AR and MA parts have four parameters, and then the first 200 data of the test sequence is used to predict the following data. The effect of the autoregressive moving average model increases significantly with the increase of parameters. When the ARMA model is set to ARMA(4,4), the model will achieve the best effect. Further increase of the parameters will lead to an increase in the calculated amount, and the prediction effect will be improved little. The prediction errors of three models for each parameter with the steps of 1, 10 and 20 are sorted into Table 2, with three significant digits reserved. Table 2 Prediction Errors of Models for Each Parameter with Different Steps

RMSE (x 10- 3 ) MAPE (x 10-3

) Data set Model Step-I Step-10 Step-20 Step-i Step-10 Step-20 Spike-ESN 0.482 10.2 23.5 0.0895 2.00 4.82

T6 ESN 0.509 11.4 27.8 0.0981 3.00 7.14 ARMA 1.44 10.2 21.9 0.289 2.37 4.46 Spike-ESN 0.577 7.83 14.3 0.211 3.94 7.10 T ESN 0.589 8.02 16.8 0.217 3.71 7.48 ARMA 1.63 9.79 16.9 0.664 4.39 7.81 Spike-ESN 2.68 39.0 95.9 0.590 10.4 24.3 N2 ESN 2.53 40.4 97.7 0.713 13.2 33.9 ARMA 9.80 70.3 102 1.84 15.5 30.6 Spike-ESN 1.56 22.4 46.0 0.417 7.10 16.1

N, ESN 1.75 27.9 52.0 0.526 9.77 19.3 ARMA 5.08 30.7 48.8 1.45 10.2 17.1 Spike-ESN 0.527 11.7 23.7 0.0810 2.01 4.25

al ESN 0.616 12.5 33.5 0.0917 2.21 5.30 ARMA 1.55 10.5 15.7 0.280 1.98 3.11 Spike-ESN 0.905 18.1 36.2 0.199 4.07 9.00

ESN 0.891 18.7 38.5 0.179 3.90 9.44 a2 ARMA 2.94 19.0 36.6 0.648 4.43 10.3 wherein the parameters in Table 2 are: engine exhaust gas temperature T6, engine combustion

chamber temperature TI, high pressure rotor speed N 2 , low pressure rotor speed N, compressor

opening angle a, and compressor opening angle a 2

It can be seen from the table that the Spike-ESN of the present invention can achieve better effect than ESN in the prediction of each parameter, and especially in the long-term prediction with a step of 20, the accuracy of each parameter is improved by 2%o. This may play a key role in aero-engine parameter prediction and fault diagnosis. For example, the Spike-ESN can be used to detect problems in a small range of numerical fluctuations, which may be ignored by ESN. Compared with ESN and ARMA, a more stable and effective model is obtained. The reason is that the Spike-ESN model is improved on the basis of ESN, which absorbs the advantages of small calculated amount, convenient training and applicability to time series, and the spike mechanism is added to make the new model more sensitive to temporal information in time series. In terms of experimental results, the prediction effect is improved, and the Spike-ESN is more capable of long term prediction than ESN. Although the embodiments of the present invention have been shown and described above, it will be appreciated that the above embodiments are only used for describing the technical solution of the present invention and shall not be understood as limitations to the present invention. Those ordinary skilled in the art can make amendments and replacements to the above embodiments within the scope of the present invention without departing from the principle and purpose of the present invention.

Claims

CLAIMS:

1. A spike echo state network model for fault prediction of an aero-engine, comprising the following steps: step Sl: selecting the sample characteristics of data firstly, a plurality of groups of sensors on an aero-engine are used to collect various operating state data of aircraft flight sorties; each type of operating state data represents a characteristic variable, with the data form of one dimension of data series; then, the fault time in the collected data series is determined by the traditional empirical criterion method to list fault labels in the series; and finally, the fluctuation of each parameter data near the fault label of each dimension of data is observed by the MATLAB data analysis method, and characteristic variables which are highly correlated with the occurrence of faults during aero-engine operation are selected for fault prediction; step S2: preprocessing the samples in view of various fault types generated during aero-engine operation, multiple data sets are respectively established by using the characteristic variable data selected in Step 1, so as to construct a sample set for the spike echo state network model for various fault predictions, specifically: firstly, the characteristic variable data selected in Step 1 is preprocessed; then, multiple data segments containing the same type of fault are divided from the data according to the fault labels listed in step 1, the sampling times of each data segment before and after a fault occurs must be at least 4000, and multiple data segments belonging to different fault types can be obtained by repeating operations for each fault type; secondly, each data segment belonging to each fault type is labeled with basic information, and stored in a database; and finally, aiming at a certain type of fault, a data segment belonging to the fault type is selected from the database as a training sample

set Xtrain of the spike echo state network model, and one of the remaining data segments is selected

as a test sample set X ;

step S3: constructing the spike echo state network model the spike echo state network (Spike-ESN) is improved by adding a spike encoding mechanism into the input layer to form a spike input layer; and adding a spike activation function into the reservoir to form a spike reservoir; that is, the spike echo state network (Spike-ESN) adopted is composed of a spike input layer, a spike reservoir and a training output layer, and the training

sample set Xtran obtained in step 2 is used to establish a model; step S4: training the spike echo state network model the model generated in step 3 is trained by means of regression, that is to calculate W , ,and the objective function is formula (9); min F (WO,,) =||j - W,.,X (t)| 1 | WOt,1|2(9) WouE ] 1xNres in formula (9), 2 represents a regularization norm of L2 , and A is a regularization coefficient; and the objective function is solved by means of ridge regression and can be expressed as formula (10);

Wou, = yX T (XX T +1)- 1 (10) in formula (10), the regularization coefficient A is a positive number less than 0.01;

step S5: testing and adjusting the accuracy rate of the spike echo state network

the test sample set Xtest is subjected to model prediction according to formula (1-8) in step 3,

the output result of formula (8) is compared with the true result, the parameters of the spike echo state network model are adjusted, and steps 1-4 are repeated to test whether the model accuracy meets the requirement; step S6: predicting fault data by using the trained spike echo state network model the spike echo state network for early warning of various fault types is loaded into the detection and warning equipment of an aero-engine, the sensor data is input to the equipment, and the detection and warning equipment carries out real-time calculation to predict the state data of the aero-engine in the short time of next 1-2 s; and the predicted data in the equipment is judged according to the preset experience criteria, and if abnormal data is detected from the predicted data, the equipment will give an early fault alert and give preset operation suggestions to the crew according to the fault type.

2. The spike echo state network model for fault prediction of an aero-engine according to claim 1, wherein the model in step S3 is specifically as follows: (1) spike input layer the input signal and the output signal of the spike echo state network at the time t are

respectively denoted as u(t) and y(t); each dimension of data in the training sample set Xtrain is

used for modeling, one dimension of data is used for illustration, this dimension of data series is denoted as [a(1),a(2),...,a(t)], and t is the sampling time of this dimension of data; and when data a(t) is defined to predict data a(t+i), the prediction step is i , and then in the spike echo state network model, when the prediction step is i , the input signal u(t) is a(t), and the output signal y(t) is a(t+i); the spike input layer has the function of converting the input signal u(t) to a spike train, which can be expressed as formula (1); fin (U (t))= [ui (t), U2(t),..., Iut" (t)]T (1) in formula (1), f,(-) is the transformation function of the spike input layer, and trime is the spike sampling times of a spike neuron for the same value; after the spike sampling times of the Spike ESN network are specified, each input data will be transformed into a spike train of equal length, the format of the spike train has the length of spike sampling times, and each element is 1 or 0, representing activation or suppression; and the Poisson distribution of the numerically generated spike train is shown in formula (2);

P(x =k)=- e- (2) k

! in formula (2), k represents the occurrence times of events; A is the population mean and variance of the Poisson distribution, and the practical significance of the distribution can be expressed as that when events are observed to occur A times on average, the actual probability of occurrence is k ; and the average value A of intervals to generate spike is determined by the value of the input data, and a spike train is generated according to the Poisson distribution; max(u(t))- u(t) jr(t) = t,,, ax O -Ut (3) t max(u(t))- min(u(t)) in formula (3), r(t) represents the average interval for u(t) to generate spike; r(t) is substituted

into formula (2), time intervals, i.e., [x,(t),x 2 (t),...,t, (t)], are randomly generated according to

1, i= ui (t) = I ,15 i ! t time Ik,= eZ'(4) 0, otherwise

informula(4), ui(t) is an element in a spike train; k represents the order of a spike element"1";

and i represents the position of the spike element "1" in the spike train; (2) spike reservoir the spike reservoir is a sparse network composed of a plurality of neurons connected randomly, which can achieve the function of storing data by means of adjusting the internal weight of the network. and provided that the number of spike neurons in the spike reservoir is denoted as

N,,, the internal state signal is denoted as x(t), and the random fixed internal weight is denoted

as W,,, the generation process of the spike reservoir is shown in formula (5);

W W,., = p• (5) A max(W) in formula (5), A max(W) is the maximum eigenvalue of a matrix W ; W is a sparse matrix

an important control parameter that determines the generation of W,, and represents the upper

bound of the maximum eigenvalue of W; and the state transition modes of the spike reservoir

are formula (6) and formula (7); x(t)= tanh(Wfk(u(t))+ ,x(t - 1)) (6)

fspik (u(t))- ~exp(-f -sik)(7)

in formula (6), W, is the input weight matrix of the reservoir, which is generated on the interval

[1,1] according to uniform distribution; t,, is the time series from the beginning to the end of

observation; tspie is the spike train interval after f1 ,(u(t)) spike digitization; fspike() is a spike

X(t) =[x(1), x(2),..., x(t)]ENxt - , ]represents horizontal connection between states, and the

output vector j=[(1), 9(2),...,9(t)]E of Spike-ESN can be expressed as formula (8);

.9=W,,X (t) (8)

in formula (8), Wt is an output weight.

3. The spike echo state network model for fault prediction of an aero-engine according to claim 1, wherein in step S1, when the change rate of the difference quotient of the observed data is greater than 5%, it can be considered that this dimension of data changes significantly with the occurrence of a fault, and the characteristic variable represented by this dimension of data is highly correlated with the aero-engine fault.

4. The spike echo state network model for fault prediction of an aero-engine according to claim 1, wherein the basic information in step S2 includes flight sorties, names of sample characteristics, operating time and fault labels.

Aero-engine Data driving

Before Buckling Rational Surge buckling inception buckling

Compre or pre ure Pre ure Pre ure unbalance pul ation

High-pa filter Sen or mea urement

High-frequency attenuation

Time Fault prediction

Fig. 1

() ()

Fig. 2 Value

Time

Fig. 3(a) Value

Time

Fig. 3(b) 1/4

Value

Time

Fig c Value

Time erie

Fig. 4(a) Spike train

Time erie

Fig. 4(b)

2/4 a

Value a True value Predicted value

Time b Fig. 5(a)

b Error

c

Time

Fig. 5(b) c Value

True value Predicted value d

Time d Fig. 5(c) Error

Time

Fig. 5(d)

3/4

Prediction tep

Fig. 6(a)

Prediction tep

Fig. 6(b)

Prediction tep

Fig. 6(c)

4/4