CN-121980386-A - Intelligent diagnosis method and device for surge faults of aero-engine

CN121980386ACN 121980386 ACN121980386 ACN 121980386ACN-121980386-A

Abstract

The invention provides an intelligent diagnosis method and device for surge faults of an aero-engine, which belong to the technical field of fault diagnosis of the aero-engine and comprise the steps of obtaining historical working parameters, clustering, constructing a neural network of a digital twin network architecture, performing primary screening on the working parameters, performing normalization treatment, training a surge model, and finally realizing real-time surge monitoring; the invention directly utilizes historical frame-time flight data to train, the data naturally covers a real flight envelope, the invention adopts a casing vibration signal as a main monitoring object, the working environment of a vibration sensor is relatively mild, the reliability is higher, the accuracy and the robustness of surge diagnosis in a whole life cycle are obviously improved by the prediction model based on the digital twin network architecture, the model training convergence speed is accelerated, the demand of calculation resources is reduced, and the engineering deployment is facilitated.

Inventors

GAN TIAN
XIE WENJUN
LI ZHENGXIN
CHEN JICHENG
WANG QIONG
XIAO HONG
WANG DONGHUAN
Xiao Dasheng
DING BAOQING
ZHAO RUIFENG
LIANG HUA
ZHANG LIANG

Assignees

中国人民解放军空军工程大学

Dates

Publication Date: 20260505
Application Date: 20260126

Claims (10)

1. An intelligent diagnosis method for surge faults of an aero-engine, comprising the following steps: Step S1, acquiring and clustering historical working parameters, namely acquiring working parameters related to vibration of a historical single engine and a case through a flight parameter module, wherein the working parameters comprise environmental parameters and state parameters, and then clustering the environmental parameters and the state parameters according to a part structure and a working system of the aeroengine; S2, constructing a neural network of a digital twin network architecture, namely constructing a neural network architecture based on embedded aeroengine physical constraints, and forming a prediction model; Step S3, working parameter preliminary screening, namely, preliminary screening is carried out on the working parameters obtained in the step S1 according to the rotating speed; Step S4, normalizing, namely normalizing the working parameters after primary screening; Step 5, training a surge model, selecting P historical frames of a surge fault of a single engine, sequentially processing the P historical frames of working parameter data according to the steps S1, S3 and S4, inputting the data into a prediction model trained in the step 2, outputting a predicted value of a vibration parameter by the prediction model, and further determining a deviation value between the predicted value and a true value of the surge fault historical frames; And step S6, real-time surge monitoring, namely acquiring the working parameters of the engine in real time, processing the working parameters according to the steps S1, S3 and S4, inputting and guiding the prediction model in the step S2, outputting the predicted value of the vibration parameter in real time by the prediction model, and comparing the predicted value with the vibration true value of the target data to make surge judgment.
2. The intelligent diagnosis method for surge faults of an aeroengine as claimed in claim 1, wherein the step 1 comprises the following steps: S101, selecting working parameters, wherein the working parameters comprise environment parameters and state parameters, the environment parameters comprise flying height H, mach number Ma, inlet total temperature T 0 and engine inlet total pressure P 0 , the state parameters comprise engine low-pressure compressor conversion rotating speed N 1 , engine high-pressure compressor conversion rotating speed N 2 , fan inlet guide vane angle alpha 1 , high-pressure compressor inlet guide vane angle alpha 2 , high-pressure compressor inlet total temperature T t2.5 , high-pressure compressor outlet total pressure P t3 , accelerator angle PLA, turbine outlet total pressure P t5 , turbine outlet total temperature T t5 , tail nozzle throat diameter D 8 , forward and backward overload g 1 of an airplane, left and right yaw overload g 2 of the airplane and ascending and descending overload g 3 of the airplane; Step S102, acquiring historical data and clustering, namely acquiring the working parameters of the latest L historical single engines through a flight parameter module, constructing a time sequence, and carrying out parameter clustering according to （T 0 ）、（P 0 、α 1 ）、（α 2 、P t3 ）、（T t2.5 ）、（N 2 ）、（N 1 ）、（P t5 、T t5 ）、（PLA）、（D 8 ）、（g 1 ,g 2 ,g 3 ） according to the part structure and the working flow of the aero-engine, wherein the number of clusters is the length of the super-parameter sequence.
3. The intelligent diagnosis method for surge faults of an aeroengine as claimed in claim 1, wherein the step S2 comprises the following steps: Step S201, constructing a three-layer neural network architecture embedded with physical constraints of an aeroengine based on a Bi-LSTM network to form a digital twin network architecture, wherein the digital twin network architecture comprises a physical information input layer, a coupling layer and a parameter mapping layer; step S202, inputting a feature matrix; Step S203, determining characteristic dimensions; Step S204, setting output parameters.
4. The intelligent diagnosis method of surge faults of an aeroengine, as set forth in claim 3, characterized in that a physical information input layer in a digital twin network architecture is constructed by a layer of two-way long-short-term memory cyclic neural network, the physical information input layer outputs a feature matrix to a coupling layer, the physical information input layer structure in the digital twin network architecture is composed of 10 parallel Bi-LSTM sub-networks, each sub-network corresponds to a cluster parameter group, the number of clusters is 10, each sub-network input parameter is time sequence data for receiving the corresponding cluster parameter group, the number of Bi-LSTM units of each sub-network is determined by the number of parameters in the cluster parameter group, the number of parameters is set to be 2 times of the number of parameters, the minimum number of parameters is not less than 4, the dimension of a hidden state is set to be 64 dimension, the hidden state of the last time step of each sub-network is output, and the dimension is 64 dimension; The neural network structure adopted by the coupling layer is consistent with the physical information input layer, namely, a single-layer Bi-LSTM, the input parameters of the single-layer Bi-LSTM come from the output parameters of the physical information input layer, the input parameters are spliced into a 640-dimensional vector in sequence, the parameters are set to be 128 in the number of Bi-LSTM units, the dimension of the hidden state is 128, and the hidden state is output as the hidden state of the last time step; The parameter mapping layer is composed of a full-connection layer, the output of the coupling layer is mapped to a target vibration parameter, the first full-connection layer of the parameter mapping layer is 128-dimensional input, 64-dimensional output and the activation function is ReLU, the second full-connection layer is 64-dimensional input, 32-dimensional output, the activation function is ReLU, the output layer is 32-dimensional input, n-dimensional output is the number of vibration parameter components, the activation function is Linear, the output is a predicted value of a vibration parameter B, and the dimension is n.
5. The intelligent diagnosis method for surge faults of the aeroengine according to claim 3, which is characterized in that a physical information input layer in step S202 obtains parameter input by a flight parameter module, extracts characteristics, transmits the characteristics to a coupling layer, circulates according to a clustering sequence, and has a super parameter sequence length of iteration times, the number of network elements of a bidirectional long-short-term memory circulating neural network in the physical information input layer in step S203 is consistent with the number of clusters, the sequence length of the network elements is the sum of the number of cluster parameters, the characteristic dimension is fixed to be 1, the parameter mapping layer in step S204 outputs vibration parameters B, the physical structure represented by the vibration sensor of the machine case is represented by the parameter characteristics: N represents the number of vibration parameter components acquired by the sensor, the loss function of model training is represented by mean square error, M represents the total number of parameter samples input by single model training, y (x) is target vibration data, and f (x) is vibration parameter B output in the model training process.
6. The intelligent diagnosis method of the surge fault of the aeroengine as claimed in claim 1, wherein in the step S3, the working parameters obtained in the step S1 are subjected to primary screening, the primary screening is based on the working parameters corresponding to the low-pressure rotor in a rotating speed N 1 rotating speed range of 40% -80%, in the step S4, the working parameters are normalized, and the normalization method is to obtain the relative value R of each group of data by subtracting a minimum value X min in original data from a data original value X i at a certain moment and then dividing the minimum value X max in the original data by dividing the data original value X i at the moment.
7. The intelligent diagnosis method of surge faults of an aeroengine, as set forth in claim 1, is characterized in that in step S5, P historical frames of surge faults of a single engine are selected, P historical frame work parameter data are processed according to step S1, step S3 and step S4 in sequence, then the processed data are input into a prediction model trained in step S2, and the prediction model outputs predicted values of vibration parameters, and the method comprises the following steps: step S501, training setting, namely setting time step and interval of each parameter; And S502, training fault data and setting an early warning threshold value.
8. The intelligent diagnosis method of surge fault of aeroengine as claimed in claim 7, wherein in step S501, samples are divided according to a time sequence sliding window, the length of the sliding window is determined by the sampling frequency of engine parameters, the size is sampling frequency×H, H is an empirical value, 2 is set, the sampling frequency of engine parameters is 8Hz, training algorithm is RMSProp, loss function adopts mean square error MSE, Wherein X is normalized data of samples received by the model in a single training process, M is the total number of input parameter samples in the single training process, a L is output data in the model training process, the initial learning rate is set to be 0.001, the single training time is set to be 150, a vibration real-time prediction model is obtained, P historical frames of the current surge fault are selected in step S502, the value of P is 50-100, the environmental parameters of the P historical frames and the data of the state parameters are sequentially processed according to steps S1, S3 and S4, the model of step S2 is input, the model outputs the predicted value of the vibration parameters, and the deviation value between the predicted value and the true value of each surge fault historical frame is further determined Selecting the minimum value of the average deviation values in all the frames As a surge early warning threshold.
9. The intelligent diagnosis device for surge faults of the aeroengine comprises airborne equipment, a computing platform and a wireless network, and is characterized in that the airborne equipment and the computing platform are connected in a bidirectional mode through the wireless network, the airborne equipment comprises a flight parameter module and a vibration sensor arranged on an engine case, the computing platform is a high-performance server and can be deployed on a ground station or an air early warning machine, and the wireless network is a long-short wave wireless communication network, a satellite communication network, a data link or a laser communication network and realizes real-time mutual transmission of data through the data communication link.
10. The intelligent diagnosis device for surge faults of the aero-engine, as set forth in claim 9, is characterized in that the airborne equipment acquires working parameters related to vibration of a historical single engine and a casing through a flight parameter module, and clusters environmental parameters and state parameter characteristics according to engine components and systems according to the structure and working system of the components of the aero-engine, wherein the specific model of a vibration sensor is GGZ-B; A neural network of a digital twin network architecture is constructed on a computing platform, namely, a neural network architecture based on physical constraint of an embedded aeroengine is constructed according to clustering parameters to form a prediction model, a high-performance server is mainly configured to be two X86 series processors, the number of cores of each CPU is more than 20, the main frequency is 2.4GHz,256GB (32 GB multiplied by 8) DDR4 EEC memory is provided, and 8 blocks NVIDIA TESLA V and 100 GB GPU cards are provided; The wireless network is an aviation security service aviation satellite communication link of the international maritime satellite organization or a data link of an aviation VHF band.

Description

Intelligent diagnosis method and device for surge faults of aero-engine Technical Field The invention belongs to the technical field of aeroengine fault diagnosis, and particularly relates to an intelligent diagnosis method for identifying surge faults in the flight process of an aeroengine. Background In aeroengines, surge is a common gas path fault, and once the engine is surmounted, air accident symptoms are induced, if mishandling is not carried out, the engine is stopped in the air, and if heavy, parts are broken, so that the engine is damaged and fails, and a flight accident is caused. The current surge monitoring method applied to the aeroengine is mainly an onboard pressure sensor, and whether the engine is in surge or not is determined according to a threshold criterion of surge identification by monitoring the difference between the total pressure and the static pressure of an outlet of a compressor. However, the electronic components of the airborne sensor are complex, the working environment is bad, and the fault can not be avoided when working for a long time. Secondly, the threshold criterion of surge identification is to pass a large number of complex complete machine surge high altitude bench tests and flight tests, so that the cost is too high, the full flight envelope cannot be covered, and the variation of the surge pressure difference signal identification threshold value can be caused by the decline of the engine performance along with the increase of the service life of the engine, so that false alarms are induced. When the engine is in surge, the engine casing can vibrate to a greater extent, vibration signals caused by surge have strong differences, corresponding vibration signals can be recorded into the engine parameter management system through the vibration sensor, vibration parameters can be transmitted back to the ground communication interface in real time through the airborne communication link, and real-time monitoring and diagnosis analysis of air abnormal vibration on the ground are realized. At present, the development of artificial intelligence makes the advantages of an AI fault diagnosis and prediction method more obvious, establishes an aeroengine surge fault monitoring model based on casing vibration by combining big data and a machine learning technology, automatically identifies complex abnormal state signs and trends, and carries out effective and rapid evaluation. The method of the Chinese patent CN119043731A is a mode of using a differential pressure sensor and an airborne experience mechanism model at an airborne end, and specifically comprises the steps of calculating a theoretical differential pressure in real time according to design parameters of an engine and current rotating speed, total pressure, total temperature and the like, collecting the real-time differential pressure, and comparing the differential values of two groups of data after filtering and correction. Because the pressure sensor and the surge have strong linear correlation, and the vibration sensor and the surge are nonlinear correlation, the pressure sensor is more accurate, but the model parameters of the method are calibrated once by a design phase test, and the mechanism model cannot be updated in real time after the aging of the engine, the replacement of parts and the performance degradation, so that the accuracy is reduced. In addition, the pressure sensor is positioned in the surge gas path flow passage for a long time, the working environment is bad, once the pressure sensor fails, the whole monitoring system is paralyzed, and the risk of misjudgment or missed detection can exist under the complex working condition. The methods of Chinese patent CN114564996A and Chinese patent CN119043731A are similar, and a pressure sensor and an empirical calculation model are adopted to monitor surging on line, and Chinese patent CN114564996A introduces an improved complete integrated empirical mode decomposition (ICEEMDAN) algorithm with self-adaptive noise on the model, and combines Hilbert transformation and time spectrum analysis to extract surge precursor characteristics (SPI) and realize early warning of surging. The method solves the problems of dependence on high-frequency sampling and a plurality of sensors in the traditional method, has the advantages of low cost, early warning time advance (0.05-0.3 seconds), high detection success rate and the like, and is suitable for real-time monitoring in an airborne environment. However, while empirical mode decomposition algorithms improve modal aliasing and end point effects, certain demands are placed on computational effort, which can place a certain stress on the on-board computing resources. The method of the Chinese patent CN118641206A is also a method adopting a pressure sensor and an empirical mechanism model, the precision and model correction can be adjusted along with the change of the state of an engine, but the method is only used