WO2024045246A1 - 一种针对航空发动机故障预测的脉冲回声状态网络模型 - Google Patents

一种针对航空发动机故障预测的脉冲回声状态网络模型 Download PDF

Info

Publication number
WO2024045246A1
WO2024045246A1 PCT/CN2022/121214 CN2022121214W WO2024045246A1 WO 2024045246 A1 WO2024045246 A1 WO 2024045246A1 CN 2022121214 W CN2022121214 W CN 2022121214W WO 2024045246 A1 WO2024045246 A1 WO 2024045246A1
Authority
WO
WIPO (PCT)
Prior art keywords
pulse
data
fault
spike
state network
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2022/121214
Other languages
English (en)
French (fr)
Inventor
孙希明
孙涛
刘墨燃
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Dalian University of Technology
Original Assignee
Dalian University of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Dalian University of Technology filed Critical Dalian University of Technology
Priority to AU2022358641A priority Critical patent/AU2022358641B2/en
Publication of WO2024045246A1 publication Critical patent/WO2024045246A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation

Definitions

  • the invention belongs to the technical field of aero-engine fault diagnosis, and relates to a pulse echo state network model for aero-engine surge fault prediction.
  • the aeroengine As the core power source of civil or military aircraft, the aeroengine is a highly complex and precise aerodynamic-thermal-mechanical system. Its working status has a direct or indirect impact on the safety, reliability, economic efficiency and other performance issues of the aircraft. Because aeroengines have been working in harsh high-altitude environments with high temperatures, high pressures, and strong vibrations for a long time, their probability of failure increases, which affects their performance and even causes severe flight accidents. In fact, among various types of aeroengine faults, the most representative and destructive fault type is surge fault.
  • the fault diagnosis technology of aero-engines can be roughly divided into mechanism-based modeling methods and data-driven methods.
  • mechanism modeling methods are the main research ideas for early fault diagnosis of aero-engines.
  • This mechanism method mainly establishes an accurate mathematical physical model based on the aerodynamic and thermodynamic characteristics of the aeroengine, and applies the actual observation values of the aeroengine 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,1973,10(7):400-406. ⁇
  • the earliest proposed method for aeroengine parameter estimation problem The fault influence coefficient matrix method is still used in many traditional fault diagnosis systems.
  • Document [2] ⁇ Davison C R, Birk A M.
  • data-driven fault diagnosis methods mainly include artificial neural networks, support vector machines (SVM), extreme learning machines (ELM), and hidden Markov processes.
  • 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.
  • An extended least squares support vector machine is proposed and its measurement capability for aeroengine fault diagnosis is demonstrated.
  • Literature[6] ⁇ Yang 2016(pt.1):1-10. ⁇
  • An extreme learning machine based on quantum behavioral particle swarm optimization is proposed and applied to the diagnosis problem of gas turbine fan engine.
  • the artificial neural network method has more advantages than these machine learning methods in nonlinear fitting and other aspects, and is therefore considered to be one of the most potential diagnostic tools.
  • RNN recurrent neural networks
  • time series prediction analysis methods based on historical data provide new solutions to fault diagnosis and prediction problems.
  • 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. ⁇ Using long and short term Memory (LSTM) neural network achieves good aero-engine diagnosis and prediction performance under complex operations, mixed faults and strong noise conditions.
  • LSTM long and short term Memory
  • the present invention proposes a new aero-engine fault diagnosis and prediction model, Spike echo state network (Spike-ESN), which can accurately predict aero-engine surge fault data.
  • Spike-ESN Spike echo state network
  • the present invention provides a pulse echo state network model for aeroengine fault prediction, which belongs to the field of aeroengine fault diagnosis.
  • a pulse echo state network model for aircraft engine failure prediction includes the following steps: acquiring various sensor signals of each flight of the aircraft, selecting appropriate sample characteristics of the data; denoising, normalizing, and Resampling preprocessing, dividing data segments according to fault labels to form sample sets; constructing a pulse echo state network including a pulse input layer, a pulse reserve pool, and a training output layer; using the aeroengine data sample set to train the pulse echo state network model, and using the well-trained
  • the pulse echo state network model calculates the prediction results of the training set and the test set; the prediction results are provided to the aircraft engine detection and early warning equipment. details as follows:
  • Step S1 Data sample feature selection
  • each operating status data represents a characteristic variable, and its data form is a one-dimensional data sequence; then, traditional empirical judgment is used to According to the method, the time when the fault occurs in the collected data sequence is determined, and the fault tags in the sequence are listed; secondly, the MATLAB data analysis method is used to observe the fluctuation of each parameter data near the fault tag in each dimension of data. When the difference in the data If the change rate of the quotient is greater than 5%, it can be considered that the data in this dimension will change significantly with the occurrence of faults, and the characteristic variables represented by the data in this dimension are highly related to aero-engine faults.
  • Step S2 Sample preprocessing
  • step 1 perform denoising, normalization, and resampling preprocessing on the feature variable data selected in step 1; then, according to the fault labels listed in step 1, divide the data into multiple data containing the same fault segment, it is required that the number of sampling times before and after the fault occurs in each data segment is no less than 4000.
  • each data belonging to each fault will be The basic information of the segment mark, including the number of flights, sample feature names, running time, and fault labels, is stored in the database; finally, for a certain fault, select a data segment belonging to the fault type from the database as the pulse echo state network
  • the training sample set X train of the model select one of the remaining data segments as the test sample set X test ;
  • Step S3 Construct pulse echo state network model
  • the present invention improves the traditional Spike Echo State Network (Spike-ESN).
  • Spike-ESN Spike Echo State Network
  • the improvements are as follows: adding the pulse coding mechanism in brain-like computing to the input layer to form a pulse input layer; adding pulses to the reserve pool. Activation function to form a pulse reserve pool.
  • 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 used in the present invention consists of a pulse input layer, a pulse reserve pool and a training output layer.
  • the usage steps The training sample set X train obtained in 2 is used to build a model.
  • the specific construction process is as follows:
  • the input signal and output signal of the pulse echo state network at time t are recorded as u(t) and y(t) respectively; each dimensional data in the training sample set X train is used for modeling respectively, and one dimension is used here.
  • record the dimensional data sequence as [a(1),a(2),...,a(t)] T , t is the sampling time of the dimensional data; define the data a(t) to predict the data
  • the prediction step size is i
  • the impulse echo state network model when the prediction step size is i, the input signal u(t) is a(t), and the output signal y(t) is a (t+i);
  • the function of the pulse input layer is to convert the input signal u(t) into a pulse sequence, which can be expressed as formula (1);
  • f in ( ⁇ ) is the conversion function of the pulse input layer
  • t time is the number of pulse sampling times of the spike neuron for the same value.
  • k represents the number of event occurrences
  • ⁇ is the overall mean and variance of the Poisson distribution.
  • the actual meaning of the distribution can be expressed as the probability that the observed event actually occurs k times under the condition that the observed event occurs ⁇ times on average; input data The numerical value of determines the average value ⁇ of the generated pulse interval, and then generates a pulse sequence according to the Poisson distribution;
  • ⁇ (t) represents the average number of intervals at which u(t) generates pulses; substituting ⁇ (t) into equation (2), the number of intervals t time is randomly generated according to the Poisson distribution, that is Generate a pulse sequence with an interval number for each input data;
  • u i (t) is the element in the pulse sequence;
  • k represents the order of the pulse element "1”;
  • i represents the position of the pulse element "1" in the pulse sequence;
  • the pulse reserve pool is a sparse network composed of many neurons randomly connected.
  • the function of memorizing data can be achieved by adjusting the internal weights of the network; the number of pulse neurons in the pulse reserve pool is recorded as N res , and the internal state The signal is recorded as x(t), and the randomly fixed internal weight is W res .
  • the pulse reserve pool generation process is as shown in Equation (5);
  • ⁇ max(W) is the maximum eigenvalue of matrix W
  • W is a sparse matrix randomly generated according to uniform distribution in the interval [-1, 1], and its sparsity is ⁇ , which represents the internal non-zero value of matrix W.
  • the proportion of elements; the spectral radius ⁇ is an important control parameter that determines the generation of W res , representing the upper bound of the maximum eigenvalue of W res ; the state transition method of the pulse reserve pool is Equation (6) and Equation (7);
  • W out is the output weight
  • Step S4 train pulse echo state network model
  • the training model is to use the regression method to train the model generated in step 3, that is, to calculate W out , and the objective function is formula (9);
  • Equation (9)
  • 2 represents the L 2 regularization norm, and ⁇ is the regularization coefficient; using the ridge regression method to solve the objective function can be expressed as Equation (10);
  • the regularization coefficient is generally a positive number less than 0.01;
  • Step S5 Test and adjust pulse echo state network accuracy
  • Step S6 Use the trained pulse echo state network model to predict fault data
  • the sensor data is used as equipment input, and the detection and early warning equipment performs real-time calculations to predict the status data of the aircraft engine within 1-2 seconds in the short term in the future; using The preset empirical criteria determine the predicted data in the equipment. If the equipment detects abnormal data from the predicted data, it will issue a fault warning alarm and give preset operation suggestions based on the fault type for the crew to adopt;
  • the pulse echo state network contains a neurodynamic mechanism, has timing and spatial structure advantages in processing pulse signals, and is efficient and reliable in processing complex, sparse and noise-rich spatio-temporal information extraction. performance; the echo state network performs poorly in the aero-engine failure prediction problem due to problems such as small memory capacity.
  • the pulse echo state network improves the memory capacity and can achieve better results in the aero-engine failure prediction problem; the present invention can It automatically gives the future development trend of the operating status according to the operating rules of the aero-engine, and accurately predicts the short-term changes in various performance parameters of the aero-engine, which can be used to assist the crew in determining whether the aero-engine will malfunction, making fault warning more intelligent.
  • Figure 1 is a flow chart of the aero-engine failure early warning process using the pulse echo state network model model for aero-engine failure prediction provided by the present invention
  • Figure 2 is the architecture diagram of the pulse echo state network model for aeroengine failure prediction
  • Figure 3 is a certain aircraft data set provided by the Aero Engine Research Institute;
  • Figure 3(a) is the original data chart of parameters D 8 , T 6 , T 1 , ⁇ 1
  • Figure 3(b) is the parameters ⁇ 2 , PLA,
  • Figure 3(c) is the original data diagram of parameters V, N 2 , N 1 , and Signal;
  • Figure 4 is the pulse encoding result of part of the data of the aeroengine;
  • Figure 4(a) is a partial original data diagram of the compressor opening angle ⁇ 2
  • Figure 4(b) is a partial data pulsed diagram of the compressor opening angle ⁇ 2 ;
  • Figure 5 is a comparison chart of the prediction results of the Spike-ESN model for 6 parameters in a certain aircraft sortie data set when the step size is 1;
  • Figure 5(a) is the predicted value of the engine exhaust temperature T 6 by the Spike-ESN model and the actual value.
  • Value comparison chart Figure 5(b) is the error chart between the Spike-ESN model's predicted value and the real value of the engine exhaust temperature T 6
  • Figure 5(c) is the Spike-ESN model's predicted value and the real value of the engine exhaust temperature T 1 Value comparison chart
  • Figure 5(d) is the error chart between the predicted value and the true value of the engine exhaust temperature T 1 by the Spike-ESN model;
  • Figure 6 is a comparison of the prediction results of the three models of Spike-ESN, ESN, and ARMA when the prediction steps of the six parameters are 1, 10, and 20;
  • Figure 6(a) is a comparison of the prediction results of the three models for the engine exhaust temperature T 1
  • Figure 6 (b) is a comparison chart of the prediction results of the three models for the compressor opening angle ⁇ 1
  • Figure 6 (c) is a comparison chart of the prediction results of the three models for the compressor opening angle ⁇ 2 .
  • This implementation is a pulse echo state network model for aero-engine failure prediction.
  • Figure 1 shows that the modified model is used in the aero-engine failure early warning process, which includes the following steps.
  • Step S1 Data sample feature selection
  • the present invention introduces the execution process of the method in detail based on the aeroengine data provided by Dalian University of Technology and an aerospace engine research institute.
  • the data set contains normal and fault data.
  • the data contains 12 parameters, as shown in Table 1.
  • the sampling interval of the data is 0.062s, and each data set has about 100,000 data points.
  • dimension parameter name meaning dimension parameter name meaning 1 D 8 Environmental pressure 7 H flight altitude 2 T 6 Engine exhaust temperature 8 M Flight Mach number 3 T 1 engine combustion chamber temperature 9 V flight speed 4 ⁇ 1 Compressor opening angle 10 N 2 High pressure rotor speed 5 ⁇ 2 Compressor opening angle 11 n 1 Low pressure rotor speed 6 PLA Throttle lever push 12 Signal Whether a malfunction occurred
  • the present invention introduces in detail the changes in various parameters of a certain surge.
  • surge occurs, the ambient pressure D 8 drops sharply from 1 to about 0.2, the engine exhaust temperature T 6 fluctuates slightly, and the engine combustion chamber temperature T 1 rises slightly, proving that the engine cannot effectively discharge the heated gas at this time. , causing the temperature of the engine combustion chamber to rise, and may also cause the compressor to overheat.
  • Both the high-pressure rotor speed N 2 and the low-pressure rotor speed N 1 remain at a very high state and decrease.
  • the compressor opening angle ⁇ 1 There is a significant change in the compressor opening angle ⁇ 2.
  • the aeroengine is in a rotating stall state.
  • the compressor opening angle is significantly reduced to reduce the air intake volume.
  • surge is actively eliminated.
  • the throttle stick thrust PLA, flight altitude H, flight Mach number M and flight speed V have not changed suddenly, proving that these operating parameters and state parameters have not changed, proving that this surge is not caused by operation, but due to the environment.
  • causesd by a sharp drop in pressure As active measures to eliminate surge are carried out, each parameter gradually returns to normal state amid fluctuations. After surge occurs again several times, the system returns to normal.
  • the ambient pressure D 8 is the inducement for surge
  • the exhaust gas temperature T 6 , the engine combustion chamber temperature T 1 , the high-pressure rotor speed N 2 and the low-pressure rotor speed N 1 are all affected by the rotational stall before the surge occurs.
  • the compressor opening angle ⁇ 1 and the compressor opening angle ⁇ 2 are eliminated. Parameters that actively change during surge. Therefore, the above parameters affected by surge can be used as prediction data.
  • Step S2 Sample preprocessing
  • the denoising method uses the sliding window average method
  • the normalization uses the maximum and minimum value normalization method
  • the resampling frequency takes 10 times the original frequency.
  • the fault label multiple data segments containing the same fault are divided from the data, and the number of sampling times before and after the fault occurs in each data segment is required to be no less than 4,000.
  • the fault included in this experimental data is a surge fault caused by a drop in environmental pressure. By segmenting 4000 data points before and after each fault label, multiple data segments with a length of 8000 can be obtained.
  • the basic information of each segment of data label is stored in database.
  • Step S3 Construct pulse echo state network model
  • the reserve pool dimension N res of the pulse echo state network is set to 100, the spectral radius ⁇ is set to 0.9, the sparsity ⁇ is set to 0.1, and the regularization coefficient ⁇ of the regression calculation is set is 10 -8 .
  • a scaling factor is used to reduce the overall echo state, and the scaling factor is set to 0.8.
  • the echo state network needs to be initialized.
  • the first 200 data points of the data set are used for initialization and are not counted in the network evaluation.
  • the sequence length is set to 100
  • the observation time is 10000
  • the overall scaling factor ⁇ of the pulse activation function is set to 5000.
  • Step S4 Pulse echo state network training process
  • the 13000th to 20000th data points of the data set are used as the training set, and the 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 that contains all operating states of the aeroengine, and
  • the test set is selected between 11200 and 12500 to contain real data where faults occur.
  • the data set is encoded using a Poisson distribution to generate pulses. After pulse encoding, the larger the data, the higher the frequency of the pulse sequence converted into.
  • the initialization state acquisition matrix X is an empty matrix.
  • the training data is input into the network in turn, and the echo state x(t) is updated according to equation (6). Then integrate all updates x(t) into the state acquisition matrix X. Finally, the output weight matrix W out is calculated according to formula (10), the model training is completed, and the model can be used for prediction verification.
  • the value of the regularization coefficient ⁇ is 10 -8 .
  • Step S5 Use the trained pulse echo state network model to predict fault data
  • RMSE Root Mean Squared Error
  • MAAE Mean Absolute Percentage Error
  • Test set failures occur in the range of 0 to 200 data.
  • step size is 1, the model prediction errors are all on the order of 10 -4 , which can accurately reflect the short-term development trend of aeroengine status data.
  • prediction step size increases to 10, although the model prediction error increases, and the prediction errors are all in the order of 10 -2 , it can still reflect the development trend of aeroengine status data.
  • the pulse echo state network model (Echo state network, ESN) has achieved good results on the whole, and can accurately reflect the data that will occur in the future of aeroengines.
  • the echo state Comparative experiments were also conducted on the network model and the autoregressive moving average model.
  • Set the model parameters of the echo state network model to be the same as the pulse echo state network model parameters.
  • the reserve pool dimension N res is set to 100
  • the spectral radius ⁇ is set to 0.9
  • the sparsity ⁇ is set to 0.1
  • the regularization coefficient ⁇ of the regression calculation is set to 10 -8
  • the scaling factor is set to 0.8.
  • ARMA Autoregressive moving average model
  • ARMA(4,4) that is, both the AR and MA parts have 4 parameters
  • the effect of the autoregressive moving average model increases significantly with the increase of parameters.
  • the ARMA model is set to ARMA(4,4), the model effect reaches the best.
  • Continuing to increase the parameters will lead to an increase in the amount of calculation and an improvement in the prediction effect. rare.
  • the prediction errors of the three models for each parameter with step sizes of 1, 10, and 20 are organized in Table 2, retaining three significant figures.
  • each parameter in the above Table 2 is: engine exhaust temperature T 6 , engine combustion chamber temperature T 1 , high-pressure rotor speed N 2 , low-pressure rotor speed N 1 , compressor opening angle ⁇ 1 , compressor opening angle ⁇ 2 .
  • Spike-ESN of the present invention can achieve better results than ESN, especially in the long-term prediction with a step size of 20, the accuracy of each parameter is improved. Two thousandths. This may play a key role in aeroengine parameter prediction and fault diagnosis. For example, in some small-scale numerical fluctuations, Spike-ESN can be used to detect problems, while ESN may ignore them. Compared with ESN and ARMA, a more stable and better-performing model is obtained. The reason is that the Spike-ESN model is improved on the basis of ESN and absorbs its advantages of small calculation amount, convenient training, and suitability for time series. At the same time, a pulsating mechanism is added to make the new model more sensitive to time information in time series. The experimental results show that the prediction effect has been improved, and it is more capable of long-term prediction than ESN.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Molecular Biology (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Combined Controls Of Internal Combustion Engines (AREA)
  • Measurement Of Mechanical Vibrations Or Ultrasonic Waves (AREA)

Abstract

一种针对航空发动机故障预测的脉冲回声状态网络模型，属于航空发动机故障诊断技术领域。1)获取飞机飞行各架次的各种传感器信号，选择数据的适当样本特征；对各状态数据进行预处理，根据故障标签划分数据段组成样本集；2)构建包含脉冲输入层、脉冲储备池、训练输出层的脉冲回声状态网络；3)利用航空发动机数据样本集训练脉冲回声状态网络模型，利用训练好的脉冲回声状态网络模型计算训练集和测试集的预测结果；4)将预测结果提供给航空发动机检测预警设备使用。本申请能够自动根据航空发动机运行规律给出运行状态的未来发展趋势，准确预测出航空发动机短期内各性能参数指标变化情况，用于辅助机组判断航空发动机是否会发生故障，使故障预警更加智能化。

Description

一种针对航空发动机故障预测的脉冲回声状态网络模型 技术领域
本发明属于航空发动机故障诊断技术领域，涉及到一种针对航空发动机喘振故障预测的脉冲回声状态网络模型。
背景技术
作为民用或军用飞机的核心动力来源，航空发动机是一种高度复杂而精密的气动-热力-机械系统，其工作状态对飞机的安全可靠性和经济高效性等性能问题有着直接或间接的影响。由于航空发动机长期工作于高温、高压、强震动的高空恶劣环境中，因此其发生故障的概率也随之增高，从而影响其工作性能，甚至会造成恶劣的飞行事故。实际上，在航空发动机的各类故障中，最具代表性和破坏性的故障类型就是喘振故障。当飞机在极端条件下飞行时，航空发动机的压气机受到外界影响而导致输出压力小于系统下游，引起高压气体向前倒流，导致发动机产生剧烈的震动和热端的超温，可能迅速破坏航空发动机，从而造成飞机的空中停车等严重后果。由于航空发动机具有较强的非线性和高维特性，所以在实际运行中，针对航空发动机的故障预警与预测往往是依靠机组及地面科研人员的经验判据进行估计。然而这样的经验估计也往往因为其主观性常常无法准确地实现故障监测，从而导致飞机运行故障无法得到及时判断和处理。因此，科学地预测航空发动机潜在故障因素，将喘振故障终止在提前阶段，对飞机安全、稳定、高效运行有着重要现实意义。
航空发动机的故障诊断技术大体可以分为基于机理建模方法和基于数据驱动方法，其中机理建模方法是航空发动机早期故障诊断的主要研究思路。该机理方法主要是根据航空发动机的气动热力学特性建立精确的数学物理模型，并将航空发动机传感器的实际观测值作用于模型，来计算出相应的参数估计和判断结果。文献[1]＝{Urban L A.Gas Path Analysis Applied to Turbine Engine Condition Monitoring[J].Journal of Aircraft,1973,10(7):400-406.}针对航空发动机参数估计问题最早提出了一种故障影响系数矩阵法，并仍然应用于多种传统故障诊断系统中。文献[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.}提出了一种故障图方法，将航空发动机的每一个状态对应到故障图中一个点或一个区域，通过经验判据来定性划分故障区域以确定故障类型。另外，也有学者针对航空发动机参数的高斯噪声环境和非线性性质，提出了多种基于非线性卡尔曼滤波算法的诊断方法。例如，文献[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.}提出了模糊自适应无迹卡尔曼滤波算法；文献[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.}针对存在气路测量不确定度的情况下，采用多步递归估计策略和自整定缓冲器，提出了一种新型卡尔曼滤波器。然而，真实的航空发动机是一种高度复杂的非线性系统，因此基于机理建模的故障诊断方法对模型的设计精度具有较高要求。基于数据驱动的故障诊断方法无需获得精确的数学模型，只需要有充足的判断经验和历史数据即可。迄今为止，基于数据驱动的故障诊断方法主要包括人工神经网络、支持向量机(SVM)、极限学习机(ELM)以及隐马尔可夫过程等方法。文献[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.}提出了一种扩展型最小二乘支持向量机，并证明了其具备航空发动机故障诊断的测量能力。文献[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.}提出了一种基于量子行为粒子群优化的极限学习机，并把它应用于燃气涡轮风扇发动机的诊断问题。而人工神经网络方法比这些机器学习方法在非线性拟合等方面更具有优势，因而被认为是最具潜力的诊断工具之一。随着循环神经网络(RNN)的出现，采用基于历史数据的时间序列预测分析方法，为故障诊断和预测问题提供了新的解决方案。文献[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.}利用长短期记忆(LSTM)神经网络在复杂操作、混合故障和强噪声情况下获得良好的航空发动机诊断和预测性能。但由于一般循环神经网络短时记忆容量太小，无法解决长时依赖问题，在航空发动机故障预测中表现欠佳。
本发明提出了一种新的航空发动机故障诊断预测模型，脉冲回声状态网络(Spike echo state network，Spike-ESN)，能够精准预测航空发动机喘振故障数据。本发明由中国博士后科学基金项目(2022TQ0179)和国家自然科学基金项目(61890920、61890921)资助。
发明内容
为了解决现有技术存在的问题，本发明提供一种针对航空发动机故障预测的脉冲回声状态网络模型，属于航空发动机故障诊断领域。
为了达到上述目的，本发明采用的技术方案为：
一种针对航空发动机故障预测的脉冲回声状态网络模型，该方法包括如下步骤：获取飞机飞行各架次的各种传感器信号，选择数据的适当样本特征；对各状态数据进行去噪、归一化、重采样预处理，根据故障标签划分数据段组成样本集；构建包含脉冲输入层、脉冲储备池、训练输出层的脉冲回声状态网络；利用航空发动机数据样本集训练脉冲回声状态网络模 型，利用训练好的脉冲回声状态网络模型计算训练集和测试集的预测结果；将预测结果提供给航空发动机检测预警设备使用。具体如下：
步骤S1：数据样本特征选择
首先，使用航空发动机上的多组传感器，采集飞机飞行多架次的各种运行状态数据；每一种运行状态数据代表一个特征变量，其数据形式为一维数据序列；然后，利用传统的经验判据方法确定采集到的数据序列中发生故障的时间，将序列中的故障标签列出；其次，使用MATLAB数据分析方法，观测每一维数据在故障标签附近各参数数据的波动，当数据的差商变化率大于5％，可以认为该维数据随着故障出现而发生明显变化，而且该维数据所代表的特征变量与航空发动机故障相关程度高，将某一维采集数据序列记为[s(1),s(2),...,s(t)] T，t为该维数据的采样时间，那么差商变化率可以表示为
Figure PCTCN2022121214-appb-000001
最后，选取航空发动机运行过程中与故障发生相关程度高的特征变量用于故障预测；
步骤S2：样本预处理
针对航空发动机运行中产生的多种故障类型，使用步骤1中挑选出的特征变量数据分别建立多个数据集，以构建用于各种故障预测的脉冲回声状态网络模型样本集，具体如下：
首先，对步骤1中挑选出的特征变量数据进行去噪、归一化、重采样预处理；然后，根据步骤1中列出的故障标签，从数据中划分出包含同一种故障的多个数据段，要求每个数据段中故障发生前后采样次数均不低于4000次，对每种故障重复操作可以得到属于不同故障类型的多个数据段；其次，将属于每一种故障的每个数据段标记基本信息，包括飞行架次数、样本特征名称、运行时间、故障标签，存储至数据库；最后，针对某一种故障，从数据库中，选择属于该故障类型的一个数据段作为脉冲回声状态网络模型的训练样本集X train，选择剩余数据段的其中一个作为测试样本集X test；
步骤S3：构建脉冲回声状态网络模型
本发明对传统的脉冲回声状态网络(Spike-ESN)进行了改进，改进之处为：在输入层中加入了类脑计算中的脉冲编码机制，形成脉冲输入层；在储备池中加入了脉冲激活函数，形成脉冲储备池。上述改进能够更好地挖掘数据中的隐藏时空信息，提高模型时间长时记忆能力；即本发明采用的脉冲回声状态网络Spike-ESN由脉冲输入层，脉冲储备池和训练输出层组成，使用步骤2中获得的训练样本集X train建立模型，具体构建过程如下：
(1)脉冲输入层
所述脉冲回声状态网络在t时刻的输入信号和输出信号分别记为u(t)和y(t)；将训练样本集X train中的各维数据分别用于建模，这里使用其中一维数据进行说明，将该维数据序列记为 [a(1),a(2),...,a(t)] T，t为该维数据的采样时间；定义数据a(t)预测数据a(t+i)时预测步长为i，那么在脉冲回声状态网络模型中，当预测步长为i时，输入信号u(t)为a(t)，输出信号y(t)为a(t+i)；
脉冲输入层的作用是将输入信号u(t)转化为脉冲序列，可以表示为式(1)；
Figure PCTCN2022121214-appb-000002
式(1)中，f in(·)为脉冲输入层的转化函数，t time为脉冲神经元对于同一数值的脉冲采样次数，在规定了Spike-ESN网络的脉冲采样次数后，每个输入数据都将变成等长的脉冲序列，脉冲序列的格式为以脉冲采样次数为长度，每个元素为1或0，代表激活或抑制；使用泊松分布对数值产生脉冲序列泊松分布如式(2)所示；
Figure PCTCN2022121214-appb-000003
式(2)中，k表示事件发生次数；λ为泊松分布的总体均值和方差，分布的实际意义可以表示为，观察事件平均发生λ次的条件下，实际发生k次的概率；输入数据的数值大小决定了生成脉冲间隔的平均值λ，再根据泊松分布生成脉冲序列；
Figure PCTCN2022121214-appb-000004
式(3)中，κ(t)表示u(t)产生脉冲的平均间隔数；将κ(t)代入式(2)中，根据泊松分布随机生成得到t time个间隔数，即
Figure PCTCN2022121214-appb-000005
对每个输入数据以间隔数生成脉冲序列；
Figure PCTCN2022121214-appb-000006
式(4)中，u i(t)为脉冲化序列中的元素；k表示脉冲元素“1”的次序；i表示脉冲元素“1”在脉冲序列中位置；
(2)脉冲储备池
本发明中脉冲储备池是一个由很多神经元随机连接组成的稀疏网络，通过调整网络内部权值的方法可以达到记忆数据的功能；将脉冲储备池的脉冲神经元数量记为N res，内部状态信号记为x(t)，随机固定内部权重为W res，则脉冲储备池生成过程如式(5)所示；
Figure PCTCN2022121214-appb-000007
式(5)中，λmax(W)是矩阵W的最大特征值；W是在区间[-1，1]上根据均匀分布随机生成的稀疏矩阵，其稀疏度为η，代表矩阵W内部非零元素的比例；谱半径ρ是决定W res生成的 重要控制参数，代表W res最大特征值的上界；脉冲储备池的状态转移方式为式(6)和式(7)；
x(t)＝tanh(W inf spike(u(t))+W resx(t-1))(6)
Figure PCTCN2022121214-appb-000008
式(6)中，W in为储备池的输入权重矩阵，依均匀分布在[-1，1]区间上产生；t n为从开始观察到结束观察的时间数列；t spike为f in(u(t))脉冲数据化后的脉冲序列间隔；f spike(·)是脉冲激活函数；τ为脉冲激活函数的整体缩放因子；x(t)表示内部状态信号；
(3)训练输出层
将内部状态信号进行整合，获得采集状态矩阵
Figure PCTCN2022121214-appb-000009
[·,·]表示状态间水平连接，Spike-ESN的输出向量
Figure PCTCN2022121214-appb-000010
可以表示为式(8)；
Figure PCTCN2022121214-appb-000011
式(8)中，W out是输出权重；
步骤S4，训练脉冲回声状态网络模型
训练模型就是采用回归方式训练步骤3中生成的模型，即计算W out，目标函数为式(9)；
Figure PCTCN2022121214-appb-000012
式(9)中，||·|| 2表示L 2正则化范数，λ为正则化系数；使用岭回归方法求解目标函数，可以表示为式(10)；
Figure PCTCN2022121214-appb-000013
式(10)中，正则化系数一般取小于0.01的正数；
步骤S5：测试并调整脉冲回声状态网络准确率
将测试样本集X test按步骤3中的式(1-8)进行模型预测，式(8)输出结果与真实结果进行对比，调整脉冲回声状态网络模型参数，并重复步骤1-4，测试模型准确度是否达到要求；
步骤S6：利用训练好的脉冲回声状态网络模型预测故障数据
将预警各种故障类型的脉冲回声状态网络装载到航空发动机的检测预警设备中，传感器数据作为设备输入，检测预警设备进行实时计算，预测航空发动机在未来短期时间1-2s内的状态数据；利用预设经验判据对设备中的预测数据进行判定，若设备从预测数据中检测到异常数据，则会发出故障预警警报，根据故障类型给出预设好的操作建议供机组人员采用；
本发明的有益效果：
相较于现有技术而言，脉冲回声状态网络包含神经动力学机制，对脉冲信号的处理有着时序和空间结构上的优势，在处理复杂、稀疏和富含噪声的时空信息提取方面有着高效可靠的性能；回声状态网络因记忆容量小等问题，在航空发动机故障预测问题中表现欠佳，脉冲回声状态网络提高了记忆容量，能够在航空发动机故障预测问题中取得更好的效果；本发明能够自动根据航空发动机运行规律给出运行状态的未来发展趋势，准确预测出航空发动机短期内各性能参数指标变化情况，用于辅助机组判断航空发动机是否会发生故障，使故障预警更加智能化。
附图说明
图1是本发明提供的针对航空发动机故障预测的脉冲回声状态网络模型模型用于航空发动机故障预警流程流程图；
图2是针对航空发动机故障预测的脉冲回声状态网络模型架构图；
图3是航空发动机研究所提供的某架次飞机数据集；图3(a)为参数D 8、T 6、T 1、α 1的原始数据图，图3(b)为参数α 2、PLA、H、M的原始数据图，图3(c)为参数V、N 2、N 1、Signal的原始数据图；
图4是航空发动机部分数据的脉冲编码结果；图4(a)为压气机开口角度α 2的部分原始数据图，图4(b)为压气机开口角度α 2的部分数据脉冲化图；
图5是Spike-ESN模型对飞机某架次数据集中6个参数的预测步长为1时的预测结果对比图；图5(a)为Spike-ESN模型对发动机排气温度T 6预测值与真实值对比图，图5(b)为Spike-ESN模型对发动机排气温度T 6预测值与真实值误差图，图5(c)为Spike-ESN模型对发动机排气温度T 1预测值与真实值对比图；图5(d)为Spike-ESN模型对发动机排气温度T 1预测值与真实值误差图；
图6是Spike-ESN、ESN、ARMA三个模型对6个参数预测步长为1、10、20时预测结果对比；图6(a)为三个模型对发动机排气温度T 1预测结果对比图，图6(b)为三个模型对压气机开口角度α 1预测结果对比图，图6(c)为三个模型对压气机开口角度α 2预测结果对比图。
具体实施方式
为使本发明要解决的技术问题、技术方案和优点更加清楚，下面将结合附图及实验数据的分析来详细叙述本发明的具体实施方式。
本实施中为一种针对航空发动机故障预测的脉冲回声状态网络模型，附图1表示改模型用于航空发动机故障预警流程，包括以下步骤。
步骤S1：数据样本特征选择
为了更详细说明本方法的实施流程细节，本发明以在大连理工大学联合某航空发动机研究所提供的航空发动机数据为基础，详细介绍方法的执行过程。实验数据共7组，分别为不同飞行架次中发动机产生数据，数据集中均包含正常和故障数据。数据中包含12个参数，如表1所示。数据的采样间隔为0.062s，每个数据集有10万个数据点左右。
表1 数据集参数表
维数 参数名 含义 维数 参数名 含义
1 D 8 环境压力 7 H 飞行高度
2 T 6 发动机排气温度 8 M 飞行马赫数
3 T 1 发动机燃烧室温度 9 V 飞行速度
4 α 1 压气机开口角度 10 N 2 高压转子转速
5 α 2 压气机开口角度 11 N 1 低压转子转速
6 PLA 油门杆推程 12 Signal 是否发生故障
为了选择合适数据特征样本，本发明详细介绍某一次发生喘振各参数变化情况。发生喘振时，环境压力D 8从1急剧下降到了0.2左右，发动机排气温度T 6发生了轻微波动下降，而发动机燃烧室温度T 1略微升高，证明此时发动机不能有效将加热气体排出，导致了发动机燃烧室温度升高，也可能导致压气机超温，高压转子转速N 2和低压转子转速N 1都保持在一个很高的状态并有所降低，此时压气机开口角度α 1和压气机开口角度α 2发生明显变化，当前航空发动机处于旋转失速状态，压气机开口角度明显减小用于减少进气量，通过这种方式主动消除喘振。而油门杆推程PLA，飞行高度H，飞行马赫数M和飞行速度V都未发生突变，证明这些操作参数和状态参数未发生改变，证明本次喘振不是由操作引起的，而是由于环境压力急剧下降引起的。随着主动消除喘振措施的进行，各参数在波动中逐渐恢复到正常状态，在喘振再次发生几次后，系统恢复正常。
除人为控制参数和高度速度等状态参数在喘振发生时未发生明显变化以外，其他航空发动机参数均能有效表示喘振发生时的状态变化，其中环境压力D 8是喘振发生的诱因，发动机排气温度T 6，发动机燃烧室温度T 1，高压转子转速N 2和低压转子转速N 1均受到喘振发生前旋转失速的影响，压气机开口角度α 1和压气机开口角度α 2是消除喘振时主动发生变化的参数。因此，以上受到喘振影响的参数均可作为预测数据。
步骤S2：样本预处理
对步骤1中选择航空发动机状态数据样本进行去噪、归一化、重采样预处理。去噪方法 采用滑动窗口平均法，归一化采用最大最小值归一化方法，重采样频率采取10倍原频率。根据故障标签，从数据中划分出包含同一种故障的多个数据段，要求每个数据段中故障发生前后采样次数均不低于4000次。本实验数据中包含故障是因环境压力下降引起的喘振故障，在各个故障标签前后4000个数据点进行分段，可以得到长度为8000的多个数据段，将每段数据标记基本信息存储至数据库。
步骤S3：构建脉冲回声状态网络模型
基于发明内容步骤3提出的脉冲回声状态模型，将脉冲回声状态网络的储备池维数N res设置为100，谱半径ρ设置为0.9，稀疏度η设置为0.1，回归计算的正则化系数λ设置为10 -8，为了使回声状态能够处在tanh()激活函数的非线性范围之内，因此使用缩放因子对回声状态整体进行缩减，缩放因子设置为0.8。
训练开始时，回声状态网络需要进行初始化，使用数据集的前200个数据点用于初始化，不计入网络评价中。脉冲数据化中设置序列长度为100，观察时间为10000，脉冲激活函数的整体缩放因子τ设置为5000。
利用MATLAB，实现脉冲回声状态网络模型模型构建及后续的模型训练。
步骤S4：脉冲回声状态网络训练过程
将数据集第13000到20000的数据点作为训练集，将故障发生的11400到11600之间的200个数据点随机平滑地插入到训练集中，形成一个包含了航空发动机所有运行状态的训练集，而测试集选择11200到12500之间包含有故障发生的真实数据。
针对数据集采用泊松分布生成脉冲方式进行编码，脉冲编码后，数据越大转换成的脉冲序列频率越高。
初始化状态采集矩阵X为空矩阵。依次将训练数据输入到网络中，根据式(6)更新回声状态x(t)。再将所有更新x(t)整合到状态采集矩阵X中。最后根据公式(10)计算输出权重矩阵W out，完成了模型训练，即可利用模型进行预测验证。本实施例公式(10)中，正则化系数λ取值为10 -8。
步骤S5：利用训练好的脉冲回声状态网络模型预测故障数据
选取误差(Error)、均方根误差(Root Mean SquaredError，RMSE)和平均绝对百分比误差(Mean Absolute Percentage Error，MAPE)作为检验预测结果的依据。
将测试集数据输入到训练好的脉冲回声状态网络中，根据式(8)计算结果，将预测结果与测试集真实结果进行对比，可以得知模型预测效果。
测试集故障发生在0到200数据范围内。当步长为1时，模型预测误差均在10 -4数量级，能够准确反映出航空发动机状态数据在短期内的发展趋势。当预测步长增加到10时，虽然模 型预测误差有所增加，预测误差均在10 -2数量级，但是仍能反映出航空发动机状态数据的发展趋势。
脉冲回声状态网络模型(Echo state network,ESN)在整体上取得了不错的效果，对航空发动机未来将要发生数据的反映较为准确，为了确定脉冲回声状态网络在时间信息提取上的优势，对回声状态网络模型和自回归滑动平均模型同样进行了对比实验。将回声状态网络模型的模型参数设置与脉冲回声状态网络模型参数相同。其中，储备池维数N res设置为100，谱半径ρ设置为0.9，稀疏度η设置为0.1，回归计算的正则化系数λ设置为10 -8，缩放因子设置为0.8。将自回归滑动平均模型(Autoregressive moving average model,ARMA)设置为ARMA(4,4)，即AR和MA部分都有4个参数，再利用测试序列的前200个数据进行接下来数据的预测。自回归滑动平均模型的效果随着参数的增加有显著的升高，当ARMA模型设置为ARMA(4,4)时，模型效果达到最佳，继续增加参数会导致计算量增加，且预测效果提升很少。将三种模型对各参数的步长为1、10、20的预测误差整理为表2，保留三位有效数字。
表2 模型对各参数在不同步长时的预测误差
Figure PCTCN2022121214-appb-000014
Figure PCTCN2022121214-appb-000015
其中，上述表2中各参数含义为：发动机排气温度T 6，发动机燃烧室温度T 1，高压转子转速N 2、低压转子转速N 1、压气机开口角度α 1、压气机开口角度α 2。
可以从表中看出，在各参数的预测中，本发明的Spike-ESN均能取得比ESN更好的效果，尤其是在步长为20的长期预测中，各参数的准确度均提高了千分之二。这对于航空发动机参数预测和故障诊断中可能起到了关键作用，例如一些小范围的数值波动，利用Spike-ESN就可以检查出其中的问题，而ESN可能会予以忽略。与ESN和ARMA进行对比，得到了更加稳定且效果更好的模型，其原因在于，Spike-ESN模型在ESN基础上进行改进，吸收了其计算量小，训练方便，适用于时间序列等优点，同时增加了脉冲化机制，使新模型对时间序列中的时间信息更加敏感。表现在实验结果上为，预测效果得到了提升，且比ESN更加具备长期预测的能力。
尽管上面已经示出和描述了本发明的实施例，可以理解的是，上述实施例仅用以说明本发明的技术方案，不能理解为对本发明的限制，本领域的普通技术人员在不脱离本发明原理和宗旨情况下，在本发明的范围内可以对以上述实施例进行修改和替换。

Claims (4)

  1. 一种针对航空发动机故障预测的脉冲回声状态网络模型，其特征在于，包括如下步骤：
    步骤S1：数据样本特征选择
    首先，使用航空发动机上的多组传感器，采集飞机飞行多架次的各种运行状态数据；每一种运行状态数据代表一个特征变量，其数据形式为一维数据序列；然后，利用传统的经验判据方法确定采集到的数据序列中发生故障的时间，将序列中的故障标签列出；最后，使用MATLAB数据分析方法，观测每一维数据在故障标签附近各参数数据的波动，选取航空发动机运行过程中与故障发生相关程度高的特征变量用于故障预测；
    步骤S2：样本预处理
    针对航空发动机运行中产生的多种故障类型，使用步骤1中挑选出的特征变量数据分别建立多个数据集，以构建用于各种故障预测的脉冲回声状态网络模型样本集，具体如下：
    首先，对步骤1中挑选出的特征变量数据进行预处理；然后，根据步骤1中列出的故障标签，从数据中划分出包含同一种故障的多个数据段，要求每个数据段中故障发生前后采样次数均不低于4000次，对每种故障重复操作可以得到属于不同故障类型的多个数据段；其次，将属于每一种故障的每个数据段标记基本信息，并存储至数据库；最后，针对某一种故障，从数据库中，选择属于该故障类型的一个数据段作为脉冲回声状态网络模型的训练样本集X train，选择剩余数据段的其中一个作为测试样本集X test；
    步骤S3：构建脉冲回声状态网络模型
    对脉冲回声状态网络Spike-ESN进行了改进，改进之处为：在输入层中加入脉冲编码机制，形成脉冲输入层；在储备池中加入脉冲激活函数，形成脉冲储备池；即采用的脉冲回声状态网络Spike-ESN由脉冲输入层、脉冲储备池和训练输出层组成，使用步骤2中获得的训练样本集X train建立模型；
    步骤S4，训练脉冲回声状态网络模型
    采用回归方式训练步骤3中生成的模型，即计算W out，目标函数为式(9)；
    Figure PCTCN2022121214-appb-100001
    式(9)中，||·|| 2表示L 2正则化范数，λ为正则化系数；使用岭回归方法求解目标函数，可以表示为式(10)；
    Figure PCTCN2022121214-appb-100002
    式(10)中，正则化系数λ取小于0.01的正数；
    步骤S5：测试并调整脉冲回声状态网络准确率
    将测试样本集X test按步骤3中的式(1-8)进行模型预测，式(8)输出结果与真实结果进行对比，调整脉冲回声状态网络模型参数，并重复步骤1-4，测试模型准确度是否达到要求；
    步骤S6：利用训练好的脉冲回声状态网络模型预测故障数据
    将预警各种故障类型的脉冲回声状态网络装载到航空发动机的检测预警设备中，传感器数据作为设备输入，检测预警设备进行实时计算，预测航空发动机在未来短期时间1-2s内的状态数据；利用预设经验判据对设备中的预测数据进行判定，若设备从预测数据中检测到异常数据，则会发出故障预警警报，根据故障类型给出预设好的操作建议供机组人员采用。
  2. 根据权利要求1所述的一种针对航空发动机故障预测的脉冲回声状态网络模型，其特征在于，所述的步骤S3模型具体如下：
    (1)脉冲输入层
    所述脉冲回声状态网络在t时刻的输入信号和输出信号分别记为u(t)和y(t)；将训练样本集X train中的各维数据分别用于建模，采用其中一维数据进行说明，将该维数据序列记为[a(1),a(2),...,a(t)] T，t为该维数据的采样时间；定义数据a(t)预测数据a(t+i)时预测步长为i，则在脉冲回声状态网络模型中，当预测步长为i时，输入信号u(t)为a(t)，输出信号y(t)为a(t+i)；
    脉冲输入层的作用是将输入信号u(t)转化为脉冲序列，可以表示为式(1)；
    Figure PCTCN2022121214-appb-100003
    式(1)中，f in(·)为脉冲输入层的转化函数，t time为脉冲神经元对于同一数值的脉冲采样次数，在规定Spike-ESN网络的脉冲采样次数后，每个输入数据都将变成等长的脉冲序列，脉冲序列的格式为以脉冲采样次数为长度，每个元素为1或0，代表激活或抑制；数值产生脉冲序列泊松分布如公式(2)所示；
    Figure PCTCN2022121214-appb-100004
    式(2)中，k表示事件发生次数；λ为泊松分布的总体均值和方差，分布的实际意义可以表示为，观察事件平均发生λ次的条件下，实际发生k次的概率；输入数据的数值大小决定了生成脉冲间隔的平均值λ，再根据泊松分布生成脉冲序列；
    Figure PCTCN2022121214-appb-100005
    式(3)中，κ(t)表示u(t)产生脉冲的平均间隔数；将κ(t)代入式(2)中，根据泊松分布随机生 成得到t time个间隔数，即
    Figure PCTCN2022121214-appb-100006
    对每个输入数据以间隔数生成脉冲序列；
    Figure PCTCN2022121214-appb-100007
    式(4)中，u i(t)为脉冲化序列中的元素；k表示脉冲元素“1”的次序；i表示脉冲元素“1”在脉冲序列中位置；
    (2)脉冲储备池
    脉冲储备池是一个由很多神经元随机连接组成的稀疏网络，通过调整网络内部权值的方法可以达到记忆数据的功能；将脉冲储备池的脉冲神经元数量记为N res，内部状态信号记为x(t)，随机固定内部权重为W res，则脉冲储备池生成过程如式(5)所示；
    Figure PCTCN2022121214-appb-100008
    式(5)中，λmax(W)是矩阵W的最大特征值；W是在区间[-1，1]上根据均匀分布随机生成的稀疏矩阵，其稀疏度为η，代表矩阵W内部非零元素的比例；谱半径ρ是决定W res生成的重要控制参数，代表W res最大特征值的上界；脉冲储备池的状态转移方式为式(6)和式(7)；
    x(t)＝tanh(W inf spike(u(t))+W resx(t-1))  (6)
    Figure PCTCN2022121214-appb-100009
    式(6)中，W in为储备池的输入权重矩阵，依均匀分布在[-1，1]区间上产生；t n为从开始观察到结束观察的时间数列；t spike为f in(u(t))脉冲数据化后的脉冲序列间隔；f spike(·)是脉冲激活函数；τ为脉冲激活函数的整体缩放因子；x(t)表示内部状态信号；
    (3)训练输出层
    将内部状态信号进行整合，获得采集状态矩阵
    Figure PCTCN2022121214-appb-100010
    [·,·]表示状态间水平连接，Spike-ESN的输出向量
    Figure PCTCN2022121214-appb-100011
    可以表示为式(8)；
    Figure PCTCN2022121214-appb-100012
    式(8)中，W out是输出权重。
  3. 根据权利要求1所述的一种针对航空发动机故障预测的脉冲回声状态网络模型，其特征在于，所述的步骤S1中，当观测的数据的差商变化率大于5％，可以认为该维数据随着故障出现而发生明显变化，而且该维数据所代表的特征变量与航空发动机故障相关程度高。
  4. 根据权利要求1所述的一种针对航空发动机故障预测的脉冲回声状态网络模型，其特征在 于，所述的步骤S2中基本信息包括飞行架次数、样本特征名称、运行时间、故障标签。
PCT/CN2022/121214 2022-08-30 2022-09-26 一种针对航空发动机故障预测的脉冲回声状态网络模型 Ceased WO2024045246A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
AU2022358641A AU2022358641B2 (en) 2022-08-30 2022-09-26 Spike echo state network model for fault prediction of aeroengine

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202211052012.XA CN115481658B (zh) 2022-08-30 2022-08-30 一种针对航空发动机故障预测的脉冲回声状态网络模型
CN202211052012.X 2022-08-30

Publications (1)

Publication Number Publication Date
WO2024045246A1 true WO2024045246A1 (zh) 2024-03-07

Family

ID=84422619

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/121214 Ceased WO2024045246A1 (zh) 2022-08-30 2022-09-26 一种针对航空发动机故障预测的脉冲回声状态网络模型

Country Status (3)

Country Link
CN (1) CN115481658B (zh)
AU (1) AU2022358641B2 (zh)
WO (1) WO2024045246A1 (zh)

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117950906A (zh) * 2024-03-27 2024-04-30 西南石油大学 一种基于表格图神经网络的服务器故障原因推断方法
CN118332295A (zh) * 2024-06-14 2024-07-12 太仓点石航空动力有限公司 一种航空发动机压气机叶片涂层检测方法及系统
CN118607419A (zh) * 2024-08-08 2024-09-06 四川大学 一种基于迭代容积Kalman滤波的飞行气动热估计方法
CN119532646A (zh) * 2024-11-15 2025-02-28 常州大学 基于apso-hmm的管道泄漏孔径识别方法
CN119669763A (zh) * 2024-12-13 2025-03-21 开普云信息科技股份有限公司 特定人员的聚集场所模型的训练方法、装置、介质及设备
CN119669826A (zh) * 2024-11-29 2025-03-21 南京航空航天大学 一种航天器复杂装备的故障检测与隔离方法、设备及介质
CN119917711A (zh) * 2024-11-29 2025-05-02 贵州电网有限责任公司 一种基于增量回声状态网络的负荷预测方法及系统
CN120105213A (zh) * 2025-05-07 2025-06-06 山东大学 一种机床轴承故障诊断方法、系统、存储介质及设备
CN120180936A (zh) * 2025-05-21 2025-06-20 大连理工大学 一种基于tcsnn的航空发动机气路部件衰退预测系统
CN120509442A (zh) * 2025-06-03 2025-08-19 南京中汇电气科技有限公司 一种基于风光功率联合预测模型的算法及系统
CN120808310A (zh) * 2025-07-23 2025-10-17 济南大学 基于esn-cv协同处理的实时视觉处理方法及系统
CN121115705A (zh) * 2025-11-17 2025-12-12 西安卓越蔚来氢能源科技股份有限公司 基于反应釜实时工况反馈的动态定量投料控制系统及方法

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116449685B (zh) * 2023-04-23 2024-03-29 上海交通大学 一种快速跟踪精准控制的舰船燃机控制方法及装置
CN116956197B (zh) * 2023-09-14 2024-01-19 山东理工昊明新能源有限公司 基于深度学习的能源设施故障预测方法、装置及电子设备
CN118070153B (zh) * 2024-04-22 2024-07-02 中国航发四川燃气涡轮研究院 一种航空发动机整机试验性能诊断方法
CN120317735B (zh) * 2025-03-25 2026-02-03 山东永正产业技术研究院有限公司 一种基于物联采集的胶凝材料生产能耗评估方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130090831A1 (en) * 2010-06-14 2013-04-11 Inha-Industry Partnership Institute System for fault detection and diagnosis of aircraft engine and method thereof
CN111175054A (zh) * 2020-01-08 2020-05-19 沈阳航空航天大学 一种基于数据驱动的航空发动机故障诊断方法
CN112149233A (zh) * 2020-09-30 2020-12-29 南京航空航天大学 基于回声状态网络的航空发动机动态推力估计方法
CN112749764A (zh) * 2021-01-28 2021-05-04 厦门大学 一种基于qar数据的航空发动机运行状态分类方法
CN113821888A (zh) * 2021-09-23 2021-12-21 西安热工研究院有限公司 基于周期性冲击特征提取和回声状态网络的振动数据故障诊断方法
CN114741945A (zh) * 2022-02-17 2022-07-12 成都飞机工业（集团）有限责任公司 一种航空发动机在线故障的诊断方法

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102262210B (zh) * 2011-04-20 2013-03-20 哈尔滨工业大学 基于随机子空间与多储备池集成分类的模拟电路故障诊断方法
CN102262198B (zh) * 2011-04-20 2013-02-27 哈尔滨工业大学 基于回声状态网络同步优化的模拟电路故障诊断方法
WO2019144386A1 (zh) * 2018-01-26 2019-08-01 大连理工大学 一种航空发动机过渡态关键性能参数预测方法
GB2577065A (en) * 2018-09-11 2020-03-18 Satavia Ltd System and method for aircraft health and schedule maintenance
CN113902105A (zh) * 2021-11-22 2022-01-07 郑州大学 一种基于回声状态网络的时间序列预测方法及设备

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130090831A1 (en) * 2010-06-14 2013-04-11 Inha-Industry Partnership Institute System for fault detection and diagnosis of aircraft engine and method thereof
CN111175054A (zh) * 2020-01-08 2020-05-19 沈阳航空航天大学 一种基于数据驱动的航空发动机故障诊断方法
CN112149233A (zh) * 2020-09-30 2020-12-29 南京航空航天大学 基于回声状态网络的航空发动机动态推力估计方法
CN112749764A (zh) * 2021-01-28 2021-05-04 厦门大学 一种基于qar数据的航空发动机运行状态分类方法
CN113821888A (zh) * 2021-09-23 2021-12-21 西安热工研究院有限公司 基于周期性冲击特征提取和回声状态网络的振动数据故障诊断方法
CN114741945A (zh) * 2022-02-17 2022-07-12 成都飞机工业（集团）有限责任公司 一种航空发动机在线故障的诊断方法

Cited By (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117950906A (zh) * 2024-03-27 2024-04-30 西南石油大学 一种基于表格图神经网络的服务器故障原因推断方法
CN117950906B (zh) * 2024-03-27 2024-06-04 西南石油大学 一种基于表格图神经网络的服务器故障原因推断方法
CN118332295A (zh) * 2024-06-14 2024-07-12 太仓点石航空动力有限公司 一种航空发动机压气机叶片涂层检测方法及系统
CN118607419A (zh) * 2024-08-08 2024-09-06 四川大学 一种基于迭代容积Kalman滤波的飞行气动热估计方法
CN119532646A (zh) * 2024-11-15 2025-02-28 常州大学 基于apso-hmm的管道泄漏孔径识别方法
CN119669826A (zh) * 2024-11-29 2025-03-21 南京航空航天大学 一种航天器复杂装备的故障检测与隔离方法、设备及介质
CN119917711A (zh) * 2024-11-29 2025-05-02 贵州电网有限责任公司 一种基于增量回声状态网络的负荷预测方法及系统
CN119669763A (zh) * 2024-12-13 2025-03-21 开普云信息科技股份有限公司 特定人员的聚集场所模型的训练方法、装置、介质及设备
CN120105213A (zh) * 2025-05-07 2025-06-06 山东大学 一种机床轴承故障诊断方法、系统、存储介质及设备
CN120180936A (zh) * 2025-05-21 2025-06-20 大连理工大学 一种基于tcsnn的航空发动机气路部件衰退预测系统
CN120180936B (zh) * 2025-05-21 2025-08-12 大连理工大学 一种基于tcsnn的航空发动机气路部件衰退预测系统
CN120509442A (zh) * 2025-06-03 2025-08-19 南京中汇电气科技有限公司 一种基于风光功率联合预测模型的算法及系统
CN120808310A (zh) * 2025-07-23 2025-10-17 济南大学 基于esn-cv协同处理的实时视觉处理方法及系统
CN121115705A (zh) * 2025-11-17 2025-12-12 西安卓越蔚来氢能源科技股份有限公司 基于反应釜实时工况反馈的动态定量投料控制系统及方法
CN121115705B (zh) * 2025-11-17 2026-02-06 西安卓越蔚来氢能源科技股份有限公司 基于反应釜实时工况反馈的动态定量投料控制系统及方法

Also Published As

Publication number Publication date
AU2022358641A1 (en) 2024-03-14
AU2022358641B2 (en) 2024-10-10
CN115481658B (zh) 2024-10-15
CN115481658A (zh) 2022-12-16

Similar Documents

Publication Publication Date Title
WO2024045246A1 (zh) 一种针对航空发动机故障预测的脉冲回声状态网络模型
CN110807257A (zh) 航空发动机剩余寿命预测方法
CN112580267B (zh) 基于多分支特征融合网络的航空发动机喘振预测方法
WO2023097705A1 (zh) 一种基于多源数据融合的压气机旋转失速预测方法
CN108256173B (zh) 一种航空发动机动态过程的气路故障诊断方法及系统
CN109766583A (zh) 基于无标签、不均衡、初值不确定数据的航空发动机寿命预测方法
WO2020000248A1 (zh) 一种基于空间重构的航空发动机过渡态加速过程关键性能参数预测方法
CN114818811B (zh) 一种基于孪生网络度量学习的航空发动机滚动轴承故障诊断方法
Xu et al. New RUL prediction method for rotating machinery via data feature distribution and spatial attention residual network
CN114330517A (zh) 一种基于神经网络的航空发动机传感器故障自诊断方法
CN108375474A (zh) 一种航空发动机过渡态关键性能参数预测方法
CN112560252B (zh) 一种航空发动机剩余寿命预测方法
CN115508102A (zh) 基于数据驱动的航空发动机全系统故障监测与识别方法
Liu et al. Brain-inspired spike echo state network dynamics for aero-engine intelligent fault prediction
CN115688609A (zh) 一种用于航空发动机的智能推力预测及实时预警方法
Zhong et al. Aero-engine exhaust gas temperature prognostic model based on gated recurrent unit network
CN112116101A (zh) 一种基于群体约减核极限学习机的航空发动机故障诊断方法
CN116521406A (zh) 一种基于残差门gru-vae模型的航空发动机未超限飞参数据异常检测方法
CN117421665A (zh) 一种考虑运行工况的两阶段航空发动机气路故障诊断方法
CN120805086B (zh) 融合对抗自编码与稀疏注意力的发动机飞行推力估计方法
Pan et al. A model of fuel consumption estimation and abnormality detection based on airplane flight data analysis
CN120336823A (zh) 一种基于变分神经异常检测架构的涡扇发动机故障检测方法
CN112416913A (zh) 一种基于gwo-bp算法的飞机燃油系统状态缺失值补充方法
CN116911024A (zh) 一种基于ngo优化lstm的航空发动机压气机建模与检测方法
CN110657031B (zh) 飞机发动机喘振识别方法

Legal Events

Date Code Title Description
WWE Wipo information: entry into national phase

Ref document number: 2022358641

Country of ref document: AU

ENP Entry into the national phase

Ref document number: 2022358641

Country of ref document: AU

Date of ref document: 20220926

Kind code of ref document: A

121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22957065

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 22957065

Country of ref document: EP

Kind code of ref document: A1