Gas circuit fault identification method and identification model generation method

Abstract

The invention discloses a gas path fault recognition method and a recognition model generation method. Accurate judgment, positioning and timely recognition of various gas path faults of an aero-engine are achieved. According to the embodiment of the invention, the method achieves the recognition of a plurality of typical fault types, such as erosion, corrosion, scaling and foreign object damage, and is wider in recognition range. A module level positioning capability is possessed, a fault occurrence position is clear, and rapid maintenance is facilitated; the fault judgment model, the fault positioning model and the fault recognition model are subjected to simulation data training and optimization algorithm adjustment and optimization to adapt to complex working conditions, the diagnosis precision is improved, and the robustness is enhanced.

Description

Technical Field

[0001] This application relates to, but is not limited to, the design and manufacturing technology of aeroengines, and particularly to a gas path fault identification method and an identification model generation method. Background Art

[0002] Gas path faults in aeroengines are one of the main faults of the engine. Gas path faults will lead to a decrease in the efficiency of aeroengines and an increase in exhaust gas temperature.

[0003] In different operating environments, aeroengines will induce different forms of gas path faults. For example, in marine and humid environments, the compressor and turbine components are prone to corrosion; in desert or dusty environments, the compressor components are easily eroded and damaged by foreign objects. In addition, fuel nozzles are prone to blockage due to coking, and high lubricating oil temperature or blocked oil filters may cause fuel pump failures; while overheating or deformation of blade materials will also lead to a decrease in the compressor pressure ratio and flow rate. If these faults are not detected and handled in a timely manner, they will often accumulate and deteriorate, thus seriously affecting the performance and reliability of the engine.

[0004] Therefore, how to accurately and timely identify gas path faults and their occurrence locations is crucial for ensuring the normal operation of aeroengines. Summary of the Invention

[0005] This application provides a gas path fault identification method and an identification model generation method, which can accurately and timely identify gas path faults and their occurrence locations, thereby ensuring the normal operation of aeroengines.

[0006] An embodiment of the present invention provides a gas path fault identification method, including: Collecting operation monitoring data of an aeroengine; Performing feature extraction on the collected operation monitoring data to obtain an input feature vector applicable to fault judgment and identification; Inputting the obtained input feature vector into a pre-trained fault judgment model, and when the fault diagnosis model outputs that the current working condition belongs to an abnormal state, proceeding to the next step; Inputting the obtained input feature vector into a pre-trained fault location model to obtain the location where the fault occurs; Inputting the obtained input feature vector and fault location information into a pre-trained fault identification model to obtain the fault type.

[0007] In an exemplary example, the fault identification model is trained based on the structure of an autoencoder neural network.

[0008] In an exemplary example, the performing feature extraction on the collected operation monitoring data includes: Extracting an explicit feature vector from the collected operation monitoring data; Extract the implicit feature vectors of the collected operation monitoring data using a trained autoencoder or convolutional neural network model; Concatenate the explicit feature vectors and the implicit feature vectors to obtain the input feature vectors suitable for fault judgment and recognition; Among them, the implicit feature vectors include: extracting one-dimensional implicit features through a self-connected neural network; and performing time-frequency analysis on the collected operation monitoring data to obtain a time-frequency image, and using a convolutional neural network to extract image features from the time-frequency image to obtain two-dimensional implicit features.

[0009] In an exemplary example, the extracting one-dimensional implicit features through the self-connected neural network includes: Input the time series data of each channel into the self-connected neural network, which includes an input layer, a mapping layer, a feature layer, an inverse mapping layer, and an output layer. By minimizing the reconstruction error, extract a low-dimensional embedding vector as the one-dimensional implicit feature representing the time series dynamic change characteristics.

[0010] In an exemplary example, the using a convolutional neural network to extract image features from the time-frequency map to obtain two-dimensional implicit features includes: Perform time-frequency transformation on the collected operation monitoring data to obtain a multi-channel time-frequency image; then use the obtained time-frequency image as input, and use a convolutional neural network to extract deep pattern information in the image space to form the two-dimensional implicit feature representing the complex frequency evolution law.

[0011] The embodiment of the present application further provides a method for generating an identification model, including: Determine the information of the fault type and the fault occurrence location; Generate sample data of gas path faults for training the identification model, and extract the training sample feature vectors of the sample data; Use the training sample feature vectors and the information corresponding to whether there is a fault to train a preset first mathematical model to generate a fault judgment model; Use the training sample feature vectors and their corresponding fault locations to train a preset second mathematical model to generate a fault location model; Use the training sample feature vectors and their corresponding fault types to train a preset third mathematical model to generate a fault identification model.

[0012] In an exemplary example, the generating sample data of gas path faults for training the identification model includes: Using computational fluid dynamics (CFD) + Simulink technology, according to different types of faults, the operating states of the engine under faults are respectively simulated to generate the sample data of the gas path faults for training the recognition model corresponding to different gas path faults.

[0013] In an exemplary instance, the first mathematical model includes any one or more of the following classification algorithm models: random forest, support vector machine (SVM), extreme learning machine (ELM); When the first mathematical model includes SVM, before training, the SVM further includes: automatically optimizing its kernel function parameters, penalty factors, etc. through particle swarm optimization (PSO), grey wolf optimization algorithm (GWO), or improved grey wolf algorithm.

[0014] In an exemplary instance, the second mathematical model includes any one or more of the following classification algorithm models: random forest multi-class classification model, SVM multi-class extension, ELM multi-class structure, multi-class neural network model based on Softmax output layer; When the second mathematical model includes SVM, before training, the SVM further includes: automatically optimizing its kernel function parameters, penalty factors, etc. through particle swarm optimization (PSO), grey wolf optimization algorithm (GWO), or improved grey wolf algorithm.

[0015] In an exemplary instance, the third mathematical model includes any one or more of the following classification algorithm models: deep neural network (DNN), support vector machine or Softmax multi-classification model, optimized kernel extreme learning machine (KELM); When the third mathematical model includes KELM, it further includes: using improved IGWO to optimize the parameters of KELM.

[0016] The gas path fault recognition method provided by the embodiments of the present application realizes the accurate judgment, positioning, and timely recognition of multiple types of gas path faults of aeroengines. Through the embodiments of the present application, the recognition of multiple typical fault types is realized, such as erosion, corrosion, fouling, and foreign object damage, with a wider recognition range; it has module-level positioning ability, realizes the clear location of the fault occurrence, and is convenient for rapid maintenance; the fault judgment model, fault positioning model, and fault recognition model are trained with simulation data and optimized by optimization algorithms, adapt to complex working conditions, improve the diagnostic accuracy, and enhance the robustness.

[0017] Other features and advantages of the present invention will be described in the following specification, and, in part, will be obvious from the specification, or will be understood by implementing the present invention. The objectives and other advantages of the present invention can be achieved and obtained by the structures specifically pointed out in the specification, claims, and drawings. Brief Description of the Drawings

[0018] The accompanying drawings are used to provide a further understanding of the technical solution of the present application, and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the technical solution of the present application, and do not constitute a limitation to the technical solution of the present application.

[0019] Figure 1 It is a schematic structural diagram of a compressor erosion fault tree in the related art; Figure 2 It is a schematic flow diagram of a gas path fault identification method in an embodiment of the present application; Figure 3 It is a schematic flow diagram of a recognition model generation method in an embodiment of the present application; Figure 4 It is a schematic diagram of an embodiment of the generation process of sample data of gas path faults in an embodiment of the present application; Figure 5 It is a schematic flow diagram of an embodiment of obtaining fault data of a typical fault mode in an embodiment of the present application; Figure 6 It is a schematic diagram of the construction process of an aeroengine gas path fault diagnosis system in an embodiment of the present application; Figure 7 It is a schematic diagram of the generation process of a gas path fault diagnosis model in an embodiment of the present application; Figure 8 It is a schematic diagram of the neural network structure of a structure fault recognition model based on an autoencoder neural network in an embodiment of the present application; Figure 9 It is a schematic diagram of the composition structure of a gas path fault identification device in an embodiment of the present application; Figure 10 It is a schematic diagram of the composition structure of a recognition model generation device in an embodiment of the present application. Detailed implementation manners

[0020] To make the objectives, technical solutions and advantages of the present application clearer and more understandable, the embodiments of the present application will be described in detail below with reference to the accompanying drawings. It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments can be combined arbitrarily with each other.

[0021] In a typical configuration of the present application, a computing device includes one or more processors (CPUs), an input / output interface, a network interface, and a memory.

[0022] The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of, for example, read-only memory (ROM) or flash memory (flash RAM). The memory is an example of a computer-readable medium.

[0023] Computer-readable media includes both permanent and non-permanent, removable and non-removable media and can store information by any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape disk storage or other magnetic storage devices, or any other non-transmission medium that can store information accessible by a computing device. As defined herein, computer-readable media does not include transitory media such as modulated data signals and carrier waves.

[0024] The steps illustrated in the flowchart of the figures can be executed in a computer system such as a set of computer-executable instructions. And, although a logical order is illustrated in the flowchart, in some cases, the steps shown or described can be executed in a different order than herein.

[0025] Figure 1 It is a schematic structural diagram of a compressor erosion fault tree in the related art, as Figure 1 shown, including an air inlet duct (1), a compressor (2), a combustion chamber (3), a turbine (4), and a tail nozzle (5). In the gas path faults of aeroengines, erosion faults are the most common type, especially when operating in environments such as deserts and sandstorms. A large number of sand particles are sucked into the engine gas path and collide with the surface of the compressor blades at high speed, causing the surface of the blades to become rough and the profile to be damaged, thereby affecting their aerodynamic performance. As Figure 1 shown, erosion faults first manifest as anomalies in multiple performance indicators in the compressor module, such as increased temperature, decreased pressure, flow disorder, rotational speed fluctuation, pressure ratio drop, and efficiency drop, etc. As the performance of the compressor deteriorates, the anomalies will further propagate to the turbine, power turbine, and tail nozzle modules, and also manifest as phenomena such as increased temperature, pressure change, flow disturbance, rotational speed control instability, pressure ratio reduction, and efficiency drop in these modules. Erosion faults have obvious characteristics of chain effects and can cause a continuous decline in the overall performance of the aeroengine gas path system.

[0026] Similarly, corrosion failures are common in marine, humid, and salt spray environments. Moisture in the air combines with corrosive components to form a corrosive medium that acts on the surfaces of components such as compressors and turbines over a long period, leading to oxidation, corrosion, and reduced mechanical strength of metal materials. Blade damage caused by corrosion will result in a decrease in the pressure ratio and efficiency in the compressor module, as well as fluctuations in flow rate and pressure, and will gradually propagate through the system to the turbine, power turbine, and tail nozzle modules, affecting indicators such as temperature, pressure, flow rate, rotational speed, pressure ratio, and efficiency in these modules. Fouling failures mainly occur in the fuel system, especially phenomena such as carbon deposition blockage in the nozzles. Although not directly occurring in the four illustrated modules, the uneven combustion efficiency and flow rate changes caused by it will also feedback and affect the gas path performance indicators after the compressor, indirectly triggering fluctuations in parameters such as temperature, pressure ratio, and efficiency in each module of the gas path. In addition, foreign object damage is usually caused by bird strikes, ice, or foreign object ingestion, and most often affects the front section of the compressor. The damage will cause abnormal flow distribution, rotational speed instability, and a sudden drop in the pressure ratio inside the compressor. In severe cases, it will spread to the turbine, power turbine, and tail nozzle modules, resulting in drastic changes in the performance indicators (temperature, pressure, flow rate, rotational speed, pressure ratio, efficiency) of each module.

[0027] In summary, gas path failures such as erosion, corrosion, fouling, and foreign object damage, although different in cause and location, may all affect the turbine, power turbine, and tail nozzle modules step by step through their impact on the compressor, such as along Figure 1 the path shown in the figure, and manifest as changes in the corresponding performance indicators at each level. Therefore, if these key indicators such as temperature, pressure, flow rate, rotational speed, pressure ratio, and efficiency can be monitored, it will help to achieve timely identification of gas path failures and rapid location of the fault positions.

[0028] To accurately and timely identify gas path failures and their occurrence locations and ensure the normal operation of aero-engines, the embodiments of this application provide a gas path failure identification method, as Figure 2 shown, which may include: Step 200: Collect operation monitoring data of the aero-engine.

[0029] In this step, multi-channel sensor data during the operation of the aero-engine is collected in real time, including but not limited to time series data such as temperature, pressure, flow rate, and rotational speed. The sensors are arranged to cover key components of the typical gas path, such as positions of the compressor, combustion chamber, turbine, and tail nozzle.

[0030] In one embodiment, the sensor data is collected in real time in a time-synchronized manner to form operation monitoring data that is a monitoring sample set containing multiple parameter channels.

[0031] Step 201: Extract features from the collected operation monitoring data to obtain input feature vectors suitable for fault judgment and identification.

[0032] In an exemplary instance, step 201 may include: adopting a feature extraction method consistent with that in the training of the recognition model: extracting explicit feature vectors from the collected operation monitoring data; using a trained autoencoder or convolutional neural network model to extract implicit feature vectors of the collected operation monitoring data; concatenating the explicit feature vectors and the implicit feature vectors to obtain input feature vectors suitable for fault judgment and recognition.

[0033] In an exemplary instance, extracting explicit feature vectors may include: within a sliding time window (such as 5 seconds or 10 seconds), performing various feature statistics and time-frequency feature extraction on each sensor parameter sequence.

[0034] Performing explicit feature extraction within the sliding window for each channel respectively, that is, extracting the following types of features from the time series data of this channel (it should be noted that this is only an example here and is not used to limit the protection scope of this application): Basic statistical features: maximum value, minimum value, mean value, variance, standard deviation, kurtosis, skewness; Linear trend features: slope and intercept of the fitted straight line; Seasonal and residual features: trend term, periodic term, residual term (applicable to periodic disturbance environments); Correlation features: autocorrelation coefficient, cross-correlation coefficient (used to identify co-variation between multiple channels); Spectrum features: main frequency, band energy, bandwidth, frequency centroid, etc.; Entropy features: Shannon entropy, fuzzy entropy, permutation entropy, used to characterize the dynamic complexity of the system; Fractal dimension features: box dimension, Hurst dimension, used to measure the roughness and time correlation of the signal.

[0035] All the above features are independently extracted for each channel, and each channel forms an explicit feature vector including multiple extracted features. Finally, after concatenation by channel, the final explicit feature vector with a dimension of the number of channels × the number of features per channel is obtained.

[0036] In the monitoring data of an aeroengine, a channel usually refers to a data stream of a sensor measuring a certain physical quantity. That is, each parameter corresponds to a data channel. Each channel is a set of time series data. For example, the 10-second record of the compressor inlet temperature constitutes the data of an independent channel. The monitoring data of multiple sensors form a data matrix, and each column of this data matrix represents the sampled point data of a channel, that is, time series data.

[0037] In an exemplary instance, extracting the implicit feature vector of the collected operation monitoring data may include: Extracting one-dimensional implicit features through a self-connected neural network; Performing time-frequency analysis on the collected operation monitoring data to obtain a time-frequency image, and using a convolutional neural network to extract image features from the time-frequency image to obtain two-dimensional implicit features.

[0038] In one embodiment, extracting one-dimensional implicit features through a self-connected neural network may include: Inputting the time series data of each channel into a self-connected neural network (such as an autoencoder neural network), which includes an input layer, a mapping layer, a feature layer, a demapping layer, and an output layer, and extracting a low-dimensional embedding vector by minimizing the reconstruction error as the one-dimensional implicit feature representing the time series dynamic change characteristics.

[0039] In one embodiment, using a convolutional neural network to extract image features from the time-frequency map to obtain two-dimensional implicit features may include: Performing time-frequency transformation (such as short-time Fourier transform, wavelet transform, etc.) on the collected operation monitoring data to obtain a multi-channel time-frequency image; then using the obtained time-frequency image as the input, and using a convolutional neural network (CNN) to extract the deep pattern information in the image space to form a two-dimensional implicit feature representing the complex frequency evolution law.

[0040] In an exemplary instance, splicing the explicit feature vector and the implicit feature vector to obtain an input feature vector suitable for fault judgment and recognition may include: Splicing or fusing the obtained final explicit feature vector, one-dimensional implicit feature, and two-dimensional implicit feature, and performing normalization or standardization processing to form an input feature vector suitable for fault judgment and recognition.

[0041] Step 202: Input the obtained input feature vector into a pre-trained fault judgment model. When the fault diagnosis model outputs that the current working condition belongs to an abnormal state, enter Step 203.

[0042] The fault diagnosis model is pre-trained. The input of the fault diagnosis model is the input feature vector, and the output is information indicating whether there is a fault, such as classification probability or label. If the output result is normal, end this process; if the output result is abnormal, continue to enter Step 203.

[0043] In one embodiment, the fault judgment model can judge whether the current working condition belongs to the normal state by outputting a probability or a classification result.

[0044] Step 203: Input the obtained input feature vector into a pre-trained fault location model to obtain the location where the fault occurs.

[0045] The fault location model is pre-trained. The input of the fault diagnosis model is the input feature vector, and the output is the module-level label indicating the location where the fault occurs, such as the turbine.

[0046] Through step 203, the fault is located to a certain component module, such as the compressor, combustion chamber, turbine, or tail nozzle, etc., and the identification of the fault location at the module level is initially completed.

[0047] Step 204: Input the obtained input feature vector and fault location information into the pre-trained fault identification model to obtain the fault type.

[0048] The fault identification model is pre-trained. The input of the fault diagnosis model is the input feature vector and fault location information, and the output is the fault type, such as erosion fault, corrosion fault, scaling fault, foreign object damage, etc. After the type discrimination of the fault is completed, the identification result can be further encoded and saved as an important reference for subsequent maintenance or control systems.

[0049] Through the gas path fault identification method in the embodiments of the present application, accurate judgment, location, and timely identification of various gas path faults of aero-engines are realized. The gas path fault identification method in the embodiments of the present application realizes the identification of multiple typical fault types, such as erosion, corrosion, scaling, and foreign object damage, with a wider identification range; has the module-level location ability, realizes the clear location of the fault occurrence, and is convenient for rapid maintenance; the fault judgment model, fault location model, and fault identification model are trained with simulation data and optimized by optimization algorithms, adapt to complex working conditions, improve the diagnostic accuracy, and enhance the robustness.

[0050] To implement the embodiments of the present application, the embodiments of the present application further provide a method for generating an identification model, where the identification model includes a fault judgment model, a fault location model, and a fault identification model. The method for generating an identification model provided by the embodiments of the present application is used to train a model for diagnosing gas path faults of aero-engines. The embodiments of the present application combine the simulation modeling and the training process of multiple intelligent algorithms to construct an identification model for multiple gas path faults with the capabilities of judgment, location, and identification, such as Figure 3 as shown, combined with Figure 7 it may include: Step 300: Determine the information of the fault type and the location where the fault occurs.

[0051] In this step, first, determine the typical gas path fault types in the aero-engine, including but not limited to: erosion fault, corrosion fault, scaling fault, foreign object damage, etc. At the same time, determine the key component positions that may be affected by the above faults, that is, in which modules the faults occur, such as gas path components such as compressors and turbines.

[0052] In one embodiment, the information on the fault type and the fault occurrence area may include, for example: Fault type: such as erosion, corrosion, scaling, foreign object damage; Fault occurrence location: modules such as compressors, turbines, combustion chambers, etc.; Degree of damage: mild, moderate, severe (such as a 3%, 5%, 10% decrease in blade efficiency, etc.); Fault mode parameters: such as material corrosion rate, blade roughness, pore blockage degree, foreign object impact shape, etc.

[0053] Step 301: Generate sample data for training the gas path fault recognition model, and extract the training sample feature vectors of the sample data.

[0054] In an exemplary instance, if there is enough real fault data in reality, step 301 may include: directly using these real fault data to form the sample data for training the gas path fault recognition model.

[0055] If there is very little real fault data in reality, especially in the case where some extreme or major faults are not easy to collect, one cannot rely solely on real data, but needs to rely on simulation to synthesize the sample data for training the gas path fault recognition model. In an exemplary instance, step 301 may include: Adopt computational fluid dynamics (CFD) + Simulink technology, and according to different types of faults, respectively simulate the operating state of the engine under faults to generate the sample data for training the gas path fault recognition model corresponding to different gas path faults.

[0056] In one embodiment, generating the sample data for training the gas path fault recognition model corresponding to different gas path faults may include: First, as Figure 4 Shown in the first row, according to the fault type, occurrence location, and degree of damage in the information on the determined fault type and fault occurrence location, configure the fault scenarios, such as erosion fault, corrosion fault, scaling fault, foreign object damage fault, and construct the CFD flow field simulation models corresponding to different fault scenarios to simulate the flow characteristics of the gas under different fault states. In the embodiments of the present application, the fault scenario is a specific fault combination situation set by humans, including factors such as the type of fault, fault location, degree of damage, environmental conditions, etc. Each fault combination can also be understood as a set of simulation parameter configurations used to simulate the reaction of the engine under a certain fault state.

[0057] Next, as Figure 4As shown in the second row, based on a CAD tool or an existing component model, a three-dimensional geometric model of the fault area is constructed. For example, a three-dimensional model of a compressor / turbine component after a fault is constructed, and key parts in the geometric model (such as blade surfaces, edge notches, carbon deposition areas, etc.) are parametrically modeled during the model construction process to express the damage effect.

[0058] Then, as Figure 4 As shown in the second row, the entire fault impact area of the three-dimensional geometric model (such as the blade damage surface, corrosion boundary, etc.) is meshed. In one embodiment, tools such as ANSYS Meshing can be used to generate high-quality structured or hybrid meshes to improve the stability and accuracy of numerical calculations and ensure the simulation accuracy. ANSYS Meshing is a commonly used mesh generation tool in CFD.

[0059] After that, as Figure 4 As shown in the third row, physical solution settings and numerical calculation processing are performed on the meshed CFD fluid simulation model. For example, the performance parameters of the compressor / turbine after the fault are calculated to obtain the aerodynamic performance data under the fault condition, such as the distribution of field variables such as air flow velocity, pressure, and temperature. In one embodiment, it may include: after completing the meshing of the three-dimensional geometric model of the fault, corresponding boundary conditions and physical model parameters (including turbulence models, energy equations, etc.) are set based on the meshed model, and the CFD solver is used to perform a numerical simulation of the flow process under the fault scenario to solve the physical variables such as the flow velocity, pressure distribution, and temperature field of the gas inside the model to obtain the typical aerodynamic response characteristics under this fault condition, that is, the aerodynamic performance data under the fault state. The setting of boundary conditions may include: the inlet boundary usually adopts total pressure, total temperature, or velocity inlet conditions, the outlet boundary is set as a static pressure outlet, and the wall surface is set as a no-slip boundary, and properties such as heat flux or surface roughness can be further specified to reflect the fault surface characteristics. In the selection of the physical model, a suitable turbulence model (such as the standard k-ε model or k-ω SST model) can be selected according to the simulation accuracy requirements and the type of fault, the energy equation is enabled to simulate the temperature field distribution, and it is determined whether to turn on the compressibility solution setting according to the Mach number of the air flow. After completing the configuration of the boundary conditions and the physical model, a CFD software (such as ANSYS Fluent, CFX, or OpenFOAM) can be used to perform numerical solutions on the above network area. Appropriate time step, maximum number of iterations, and convergence criteria (such as the residual is limited within 1e-5) are set, and the residual curve, mass conservation, and energy conservation are monitored in real time during the solution process to ensure the accuracy and stability of the simulation results. When the simulation converges, a flow field solution containing multi-dimensional distributions such as velocity field, pressure field, and temperature field can be obtained.

[0060] Finally, post-process the simulation output results, and extract the key performance parameters of the target components (such as compressors or turbines) as sample data for training the gas path fault recognition model, including indicators such as total pressure ratio, flow efficiency, total temperature change, and loss factor. In this way, based on the data of multiple typical operating points, performance characteristic curves under fault conditions can be constructed, such as efficiency-flow curves, pressure ratio-speed curves, etc. As Figure 4 shown in the fourth and fifth lines, these performance characteristic curves can be used later to replace the standard characteristic curves of the compressor or turbine module in the engine Simulink model, forming a system model with actual fault response behavior. So that through subsequent system simulations, time series data such as temperature, pressure, flow rate, and speed of each sensor point during engine operation under this fault scenario can be simulated at the whole engine level, and finally used to construct a training sample set for the neural network fault diagnosis model, such as importing an aero-engine performance simulation model to obtain gas path fault simulation data.

[0061] The specific implementation of extracting and generating the training sample feature vectors of the sample data in step 301 is the same as that in step 201, which will not be elaborated here.

[0062] Figure 5 It is a schematic flow chart of an embodiment for obtaining fault data of a typical fault mode in the embodiments of the present application. As Figure 5 shown, it generally includes a three-layer structure: the first layer is to build models, build a CFD flow field simulation model, and establish an engine Simulink model; the second layer is to simulate gas path faults and inject different fault configurations into the simulation model; the third layer is to output different types of fault simulation data, including multiple types of fault data such as erosion, corrosion, fouling, foreign object damage, lubricating oil pump failure, blade overheating / bending, fuel nozzle blockage, etc. Among them, building a CFD flow field simulation model is used to construct the performance characteristics of components such as compressors / turbines in the fault state; establishing an engine Simulink model is used to import the performance parameters in the fault state into the system model; simulating gas path faults is used to load fault curves in the engine system-level model and simulate the operating state of the whole engine; the various types of fault data in the bottom layer represent that each simulation fault scenario → generates a set of multi-channel operating data for training.

[0063] Step 302: Use the training sample feature vectors and the information corresponding to whether there is a fault to train a preset first mathematical model to generate a fault judgment model.

[0064] The training sample feature vectors are derived from the data output by the engine system simulation under various fault scenarios (such as temperature, pressure, flow rate, speed, etc.). After feature extraction, each set of training sample feature vectors is assigned corresponding label information, that is, the normal or fault state (i.e., binary classification label).

[0065] In an exemplary instance, step 302 may include: Using the extracted training sample feature vectors as the input information of a preset first mathematical model, and using the information indicating whether a fault exists corresponding to the training sample feature vectors as the output information of the preset first mathematical model, training the first mathematical model to determine the model parameters of the first mathematical model, so as to generate a fault judgment model according to the trained model parameters.

[0066] In an exemplary instance, the first mathematical model may include, but is not limited to, any one or more of the following classification algorithm models: Random Forest, Support Vector Machine (SVM), Extreme Learning Machine (ELM), etc. The selection of specific algorithms can be configured and optimized according to the characteristics of training data, the complexity of classification tasks, and the required model generalization ability.

[0067] In an embodiment, for the case where the first mathematical model includes a random forest model, a classification result can be output by constructing multiple decision trees and adopting a majority voting method. During the training process, parameters such as the number of trees (e.g., 100 - 500 trees), maximum depth, and minimum number of samples can be adjusted to control the model complexity and generalization ability. In the embodiments of the present application, the random forest algorithm improves the generalization ability of the model by regulating the number of decision trees.

[0068] In an embodiment, for the case where the first mathematical model includes SVM, the samples can be projected into a high-dimensional space, and boundary classification can be performed by selecting different kernel functions (such as linear kernel, polynomial kernel, Gaussian kernel, etc.). Grid search tuning can be performed on the penalty coefficient C, kernel function parameter γ, etc. In the embodiments of the present application, the support vector machine can select different kernel functions according to the data characteristics and finely tune the kernel function parameters.

[0069] In an embodiment, for the case where the first mathematical model includes ELM, a single-hidden layer feedforward neural network can be constructed. By randomly initializing the input weights and biases, in this way, the training can be completed only by calculating the output weights once, and the training speed is fast. In an embodiment, the number of neurons in the hidden layer needs to be set during training, and an activation function (such as sigmoid, relu, sine, etc.) is selected according to the data characteristics to obtain the best classification effect. In the embodiments of the present application, the extreme learning machine optimizes the diagnosis effect by determining the number of neurons in the hidden layer and selecting a suitable function from multiple activation functions.

[0070] The fault judgment model generated after being trained in step 302 can be used as the first-layer classifier of the system to determine whether there is a fault in the working condition corresponding to the currently monitored engine operation data. If the output of the fault judgment model is in the normal state, the system will not perform further processing; if the output of the fault judgment model is in the fault state, then the subsequent fault location and fault identification processes will be triggered, thereby realizing hierarchical intelligent diagnosis.

[0071] In an exemplary instance, before the support vector machine is trained, it may further include: automatically optimizing its kernel function parameters and penalty factors through the particle swarm optimization algorithm (PSO), gray wolf optimization algorithm (GWO), or improved gray wolf algorithm, so as to improve the classification accuracy and generalization ability of the fault judgment model.

[0072] In an exemplary instance, in order to enhance the reliability and robustness of the fault judgment model in a complex environment, it may further include: injecting white noise of different intensities into the training dataset to simulate sensor noise and abnormal disturbances, and dynamically adjusting the hyperparameter configuration of the optimization algorithm on this basis to test the sensitivity and stability of the fault judgment model to noise interference.

[0073] Step 303: Use the training sample feature vectors and their corresponding fault locations to train a preset second mathematical model to generate a fault location model.

[0074] After the training of the fault judgment model is completed, in order to further accurately locate the module where the fault occurs, a fault location model needs to be trained in the embodiments of the present application. The fault location model is used to further output the specific module location where the fault occurs on the premise that a fault has been determined, such as the compressor, combustion chamber, turbine, or tail nozzle.

[0075] In an exemplary instance, step 303 may include: Taking the extracted training sample feature vectors as the input information of a preset second mathematical model, and taking the fault locations corresponding to the training sample feature vectors as the output information of the preset second mathematical model, training the second mathematical model to determine the model parameters of the second mathematical model, so as to generate a fault location model according to the trained model parameters.

[0076] In one embodiment, the location where the fault occurs can be represented by a module-level label. The module-level label can assign a module-level label to each training sample according to the location of the fault set in the simulation scenario, such as: compressor fault, combustion chamber fault, etc.

[0077] In an exemplary instance, the second mathematical model may include, but is not limited to, any one or more of the following classification algorithm models: random forest multi-class classification model, SVM multi-class extension (such as one-versus-one or one-versus-all), ELM multi-class structure, multi-class neural network model based on Softmax output layer, etc. The selection of specific algorithms can be configured and optimized according to the characteristics of training data, the complexity of classification tasks, and the required model generalization ability.

[0078] The fault location model obtained through training in step 303 can perform module-level determination on real-time monitoring data during the operation of the system, providing support for subsequent fault identification.

[0079] In an exemplary instance, when constructing a multi-class classifier using a support vector machine, it may further include: searching for parameters through PSO, GWO, or improved grey wolf algorithm to further improve the effectiveness of the classification boundary.

[0080] In an exemplary instance, to enhance the adaptability of the fault location model to operating condition disturbances, it may further include: adding random noises of different magnitudes to the training samples and testing the module recognition accuracy of the fault location model during the training process to evaluate its fault location ability in complex environments.

[0081] Step 304: Use the training sample feature vectors and their corresponding fault types to train a pre-set third mathematical model to generate a fault identification model.

[0082] As the last link in the diagnostic chain, the goal of the fault identification model is to identify specific fault types, such as erosion faults, corrosion faults, scaling faults, or foreign object damage, etc., after the fault location is completed.

[0083] In an exemplary instance, step 304 may include: Take the extracted training sample feature vectors as the input information of the pre-set third mathematical model, and take the fault types corresponding to the training sample feature vectors as the output information of the pre-set third mathematical model, and train the third mathematical model to determine the model parameters of the third mathematical model, so as to generate a fault identification model according to the trained model parameters.

[0084] In an embodiment, the fault types corresponding to the training sample feature vectors can be represented by assigning specific fault type labels to each training sample using corresponding fault simulation types, and the fault type labels may include, for example, erosion, corrosion, etc.

[0085] In an embodiment, the fault location results can be further fused as the input information of the third mathematical model.

[0086] The fault identification model can adopt an intelligent algorithm with greater discriminative capabilities. In an exemplary instance, the third mathematical model can include but is not limited to any one or more of the following classification algorithm models: deep neural network (DNN) such as multi-layer perceptron (MLP), support vector machine or Softmax multi-classification model, optimized kernel extreme learning machine (KELM), etc.

[0087] In one embodiment, if Figure 8 As shown in the figure, the fault identification model can adopt a structure based on an autoencoder neural network. The autoencoder neural network consists of an input layer, a mapping layer, a principal component compression layer, a demapping layer, and an output layer, forming a nonlinear compression and decoding process for the input feature vector. The mapping layer compresses the input feature vector dimension, and principal component analysis (PC analysis) is performed in the intermediate layer. Key fault features are then recovered through decoding. Finally, the output layer maps the fault samples to corresponding fault type labels, such as erosion, corrosion, scaling, foreign object damage, nozzle blockage, blade overheating, and lubricating oil pump failure. This is used to classify and identify typical gas path faults such as erosion, corrosion, scaling, foreign object damage, and overheating deformation. The autoencoder neural network effectively improves the ability to distinguish multiple types of gas path faults. The autoencoder neural network is suitable for both supervised classification in the training phase and efficient fault identification in the inference phase. The embodiment of the present application performs data mining and analysis on aircraft engine operation and maintenance data and engine operation simulation data, extracts features, and uses neural network methods to diagnose common aircraft engine faults such as erosion, corrosion, scaling, and foreign object damage, thereby obtaining the probability of occurrence of the fault and the possible components, laying the foundation for timely diagnosis of aircraft engine faults.

[0088] After the fault identification model is trained through step 304, the output results of the first two models, namely the fault judgment model and the fault location model, are combined to achieve module-type cascade identification and output structured diagnostic results, such as: turbine module + corrosion fault, providing accurate support for the engine health management system.

[0089] In an exemplary embodiment, it is recommended to use the kernel extreme learning machine (KELM) model for training, and to adjust the kernel parameters and network structure parameters in the KELM in combination with an intelligent optimization algorithm.

[0090] In one embodiment, the improved grey wolf optimization algorithm (IGWO) is used to optimize the parameters of the KELM. White noise of varying intensities is also introduced into the training data to simulate sensor errors and environmental noise in the measured signals, thereby evaluating the recognition model's anti-interference ability and generalization level.

[0091] It should be noted that if Figure 7As shown in the figure, during the model training phase, the fault diagnosis system utilizes a variety of classification algorithm structures, including support vector machines, extreme learning machines, and random forests. Model parameters can be further fine-tuned using intelligent algorithms such as particle swarm optimization and gray wolf optimization. This optimization and training process is completed in one go before model deployment. During system operation, there's no need to perform the optimization again. Simply input the feature vectors of the collected operational monitoring data into the trained model to sequentially complete fault diagnosis, location, and type identification, achieving efficient online intelligent fault diagnosis.

[0092] The identification model generation method provided in the embodiment of the present application has good adaptability, scalability and fault differentiation capabilities. In some embodiments, by introducing simulation modeling and CFD aerodynamic characteristics extraction of multiple typical fault scenarios, combined with Simulink system-level response simulation, it is possible to generate high-quality, structured training sample sets in the absence of real fault data, effectively covering multiple complex working conditions such as erosion, corrosion, scaling, and foreign object damage. In the training stage, the identification model generation method provided in the embodiment of the present application integrates explicit features and implicit deep feature extraction technology to improve the model's perception of changes in gas path performance; and by introducing multiple classification structures such as support vector machines, extreme learning machines, and deep neural networks, as well as intelligent algorithms such as particle swarm optimization and gray wolf optimization to tune model parameters, the accuracy and robustness of the diagnostic model are significantly enhanced. Therefore, the identification model generation method provided in the embodiment of the present application not only improves the recognition accuracy of multiple types of faults, but also has good anti-interference ability and cross-working condition adaptability, and is suitable for intelligent diagnosis and predictive maintenance systems during the operation of aircraft engines.

[0093] Through the recognition model generation method provided in the embodiments of this application, after model training is completed, a fault judgment model, a fault location model, and a fault identification model are obtained. In this way, in the actual application stage, the collected engine operation monitoring data is input and the trained fault judgment model, fault location model, and fault identification model are called in sequence. This will complete the entire process from operating condition abnormality identification, module-level fault location, to fault type identification, and ultimately output a structured diagnostic result. This diagnostic result can be used in subsequent maintenance scheduling, flight support, or health management systems, providing support for improving aircraft engine operational reliability and maintenance decision-making efficiency.

[0094] Figure 6 FIG. 1 is a schematic diagram of the construction process of the aircraft engine gas path fault diagnosis system in the embodiment of the present application. Figure 6As shown in the figure, from top to bottom, it includes: the modeling stage of the first layer, including: building a CFD flow field simulation model for generating local aerodynamic characteristics (such as a decrease in compressor efficiency) during component failures, and constructing an engine Simulink model for the whole machine system response simulation to obtain operation data (such as multi-channel sensor analog outputs). The data preparation stage of the second layer, where the data includes: engine operation and maintenance data collected from actual operation or experiments, and engine simulation data output from the fault simulation scenarios in the upper-layer modeling. The feature extraction stage of the third layer, including: performing fault-related data mining on the collected or simulated data, and using statistical features, spectra, neural networks, etc. to extract explicit / implicit feature vectors. The model training stage of the fourth layer, including: building an identification model for aeroengine gas path faults based on the extracted features, including the training of a fault judgment module, a fault location module, and a fault identification module. The fifth layer shows the recognizable typical fault types, including: erosion, corrosion, fouling, foreign object damage, lubricating oil pump failure, blade overheating / bending, nozzle blockage, etc.

[0095] The present application also provides a computer-readable storage medium storing computer-executable instructions for executing the gas path fault identification method described in any one of the above, and / or the identification model generation method described in any one of the above.

[0096] The present application further provides a computer device including a memory and a processor, where the memory stores the following instructions executable by the processor: steps for executing the gas path fault identification method described in any one of the above, and / or steps for executing the identification model generation method described in any one of the above.

[0097] Figure 9 It is a schematic structural diagram of the gas path fault identification device in the embodiment of the present application, as Figure 9 shown, at least including: a collection module, a first extraction module, a fault judgment module, a fault location module, and a fault identification module; where The collection module is used to collect aeroengine operation monitoring data; The first extraction module is used to extract features from the collected operation monitoring data to obtain input feature vectors suitable for fault judgment and identification; The fault judgment model is used to determine that the current working condition belongs to an abnormal state according to the input feature vector; The fault location model is used to obtain the location where the fault occurs according to the input feature vector to obtain fault location information; The fault identification model is used to obtain the fault type according to the input feature vector and the fault location information.

[0098] Through the gas path fault identification device in the embodiments of the present application, accurate judgment, positioning, and timely identification of various types of gas path faults in aeroengines are realized. The gas path fault identification method in the embodiments of the present application realizes the identification of multiple typical fault types, such as erosion, corrosion, fouling, and foreign object damage, with a wider identification range; it has the module-level positioning ability, realizes the clear location of the fault occurrence, and is convenient for rapid maintenance; the fault judgment model, fault positioning model, and fault identification model are trained with simulation data and optimized with optimization algorithms, adapt to complex working conditions, improve the diagnostic accuracy, and enhance the robustness.

[0099] Figure 10 It is a schematic structural diagram of the identification model generation device in the embodiments of the present application, as Figure 10 shown, and at least includes: a determination module, a generation module, a second extraction module, a first training module, a second training module, and a third training module; where The determination module is used to determine the information of the fault type and the fault occurrence location; The generation module is used to generate sample data of gas path faults for training the identification model; The second extraction module is used to extract the training sample feature vectors of the generated sample data; The first training module is used to train a preset first mathematical model with the training sample feature vectors and the information corresponding to whether there is a fault, and generate a fault judgment model; The second training module is used to train a preset second mathematical model with the training sample feature vectors and their corresponding fault locations, and generate a fault positioning model; The third training module is used to train a preset third mathematical model with the training sample feature vectors and their corresponding fault types, and generate a fault identification model.

[0100] In an exemplary instance, the third mathematical model can adopt a structure based on an autoencoder neural network, and the trained fault identification model is based on the structure of an autoencoder neural network.

[0101] The identification model generation device provided by the embodiments of the present application has good adaptability, scalability, and fault discrimination ability.

[0102] Although the disclosed embodiments of the present application are as above, the content described is only the embodiments adopted for facilitating the understanding of the present application, and is not used to limit the present application. Any person skilled in the art within the scope of the present application can make any modifications and changes in the form and details of the implementation without departing from the spirit and scope disclosed by the present application. However, the patent protection scope of the present application shall still be subject to the scope defined by the appended claims.

Claims

1. A gas path fault identification method, characterized in that, Including: Collecting operation monitoring data of an aero-engine; Performing feature extraction on the collected operation monitoring data to obtain an input feature vector applicable to fault judgment and recognition; Inputting the obtained input feature vector into a pre-trained fault judgment model. When the fault diagnosis model outputs that the current working condition belongs to an abnormal state, proceed to the next step; Inputting the obtained input feature vector into a pre-trained fault location model to obtain the location where the fault occurs; Inputting the obtained input feature vector and fault location information into a pre-trained fault recognition model to obtain the fault type.

2. The gas path fault identification method according to claim 1, wherein, The fault recognition model is trained based on the structure of an autoencoder neural network.

3. The gas path fault identification method according to claim 1, wherein, The performing feature extraction on the collected operation monitoring data includes: Extracting an explicit feature vector from the collected operation monitoring data; Using a trained autoencoder or convolutional neural network model to extract an implicit feature vector of the collected operation monitoring data; Concatenating the explicit feature vector and the implicit feature vector to obtain the input feature vector applicable to fault judgment and recognition; Wherein, the implicit feature vector includes: extracting one-dimensional implicit features through a self-connected neural network; and performing time-frequency analysis on the collected operation monitoring data to obtain a time-frequency image, and using a convolutional neural network to perform image feature extraction on the time-frequency image to obtain two-dimensional implicit features.

4. The gas path fault identification method according to claim 3, wherein, The extracting one-dimensional implicit features through the self-connected neural network includes: Inputting the time series data of each channel into the self-connected neural network, which includes an input layer, a mapping layer, a feature layer, a reflection layer, and an output layer. By minimizing the reconstruction error, a low-dimensional embedding vector is extracted therefrom as the one-dimensional implicit feature representing the time series dynamic change characteristics.

5. The gas path fault identification method according to claim 3, wherein, The using a convolutional neural network to perform image feature extraction on the time-frequency diagram to obtain two-dimensional implicit features includes: Performing time-frequency transformation on the collected operation monitoring data to obtain a multi-channel time-frequency image; then using the obtained time-frequency image as an input, and using a convolutional neural network to extract deep pattern information in the image space to form the two-dimensional implicit features representing the complex frequency evolution law.

6. A method for generating an identification model, characterized in that, Including: Determining information on the fault type and the location where the fault occurs; Generating sample data of a gas path fault for training an identification model, and extracting a training sample feature vector of the sample data; Using the training sample feature vector and the information corresponding to whether there is a fault to train a preset first mathematical model to generate a fault judgment model; Using the training sample feature vector and its corresponding fault location to train a preset second mathematical model to generate a fault location model; Using the training sample feature vector and its corresponding fault type to train a preset third mathematical model to generate a fault recognition model.

7. The method for generating an identification model according to claim 6, wherein, The generating sample data of a gas path fault for training an identification model includes: Adopting computational fluid dynamics CFD+Simulink technology, and respectively simulating the operation state of the engine under faults according to different types of faults to generate the sample data of the gas path fault for training the identification model corresponding to different gas path faults.

8. The method for generating an identification model according to claim 6, wherein, The first mathematical model includes any one or more of the following classification algorithm models: random forest, support vector machine (SVM), extreme learning machine (ELM); When the first mathematical model includes SVM, before training, the SVM further includes: automatically optimizing its kernel function parameters, penalty factors, etc. through particle swarm optimization (PSO), grey wolf optimization (GWO) or improved grey wolf algorithm.

9. The method for generating an identification model according to claim 6, wherein, The second mathematical model includes any one or more of the following classification algorithm models: random forest multi-class classification model, SVM multi-class extension, ELM multi-class structure, multi-class neural network model based on Softmax output layer; When the second mathematical model includes SVM, before training, the SVM further includes: automatically optimizing its kernel function parameters, penalty factors, etc. through particle swarm optimization (PSO), grey wolf optimization (GWO) or improved grey wolf algorithm.

10. The method for generating an identification model according to claim 6, wherein, The third mathematical model includes any one or more of the following classification algorithm models: deep neural network (DNN), support vector machine or Softmax multi-classification model, optimized kernel extreme learning machine (KELM); When the third mathematical model includes KELM, it further includes: using improved IGWO to optimize the parameters of KELM.

Images