Aeroengine Aerodynamic Instability Prediction Method and System Based on a Sampling Observer

Abstract

The present invention discloses a method and system for predicting aeroengine aerodynamic instability based on a sampling observer, including: using a differential equation with the single-point measurement data and its derivative of the aeroengine as state variables as a dynamic model for describing the normal mode and the instability precursor mode during the development of engine instability; constructing neural network identification models corresponding to different instability tests, fusing and expressing the neural network identification models under similar working conditions, and constructing a knowledge base of instability characteristics containing multiple decision groups; using the measured single-point pressure data to perform precursor mode detection in each decision group to obtain corresponding secondary decision results, and comprehensively processing the secondary decision results of different decision groups to obtain a primary decision result. The present invention uses single-sensor data to obtain the neural network expression of the aerodynamic instability model in a large-scale state space, which is beneficial to reducing the computational complexity of the decision-making process, thereby improving the real-time performance of instability prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of aero-engine safety control, and particularly to a method and system for predicting aero-engine aerodynamic instability based on a sampling observer. Background Technique

[0002] The statements in this part merely provide background technical information related to the present invention and do not necessarily constitute prior art.

[0003] As a core component of an aero-engine, the aerodynamic instability phenomenon of an axial-flow compressor (usually including rotating stall and surge) will greatly deteriorate the performance of the engine (thrust, economy), and even cause the engine to suddenly flame out, or the compressor blades vibrate violently so that the blades break and cause damage to the entire engine. In addition, the instability characteristics exhibited by the compressor at different speeds and different working environments also vary greatly. To ensure the stable operation of the engine under different working conditions, it is usually required that the compressor has a sufficient stall margin at the operating point. If the control system can effectively improve the stability of the compressor system before stall occurs, the requirement for the stall margin can be reduced. Therefore, how to obtain a longer warning time before instability begins is of great significance for ensuring the safe and efficient operation of the engine.

[0004] Since rotating stall usually occurs before surge, how to quickly and accurately detect the occurrence of stall precursors has become a key issue in the research of instability prediction. To address this issue, people usually use signal analysis methods in the frequency domain, time domain, or time-frequency domain to study the characteristics of stall initial disturbances, and monitor the occurrence of stall precursors by reasonably setting thresholds. The harmonic Fourier coefficient method, traveling wave energy method, random statistical method, correlation analysis method, and fast wavelet analysis are typical representatives. These detection methods perform stall warning by extracting partial characteristics such as the amplitude, phase, and energy of the compressor, but they cannot describe the comprehensive system dynamic characteristics of the compressor when instability occurs under various working conditions. As Professor Day of the University of Cambridge said, even after 75 years of research on compressor instability, for a new compressor, we still cannot accurately predict its stall characteristics.

[0005] In addition, fault diagnosis methods have been widely studied in the task of instability measurement. The early detection of stall precursors can be understood as the detection of minor faults in the field of fault diagnosis. This method usually uses approximate tools such as neural networks to learn the nonlinear dynamic functions of the engine in healthy and faulty states respectively, and compares the output of each neural network with the measured output of the engine, so as to detect and diagnose fault modes by setting appropriate thresholds. One of the key issues lies in how to ensure the accurate learning of the neural network for the instability characteristics of the engine. If the neural network cannot accurately describe the dynamic model of aerodynamic instability, it may mislead the decision-making process.

[0006] In the field of adaption, the accurate learning of a neural network (i.e., the convergence of the neural network weights to the true values) usually needs to satisfy the persistent excitation condition. By virtue of the persistent excitation characteristic of the RBF neural network, the prior art has proposed a stall precursor detection method based on deterministic learning. This method uses the RBF neural network to accurately identify the system dynamics from the multi-sensor data of a compressor, and then conducts early detection of instability precursors. However, in the research on instability precursor detection based on deterministic learning, it is required to arrange multiple sensors at the circumferential positions of the compressor. This will increase the computational burden and economic cost of the instability prediction system. More importantly, the arrangement of sensors requires drilling holes in the casing wall, which will damage the compressor structure and may affect the internal air flow. In addition, the non-linear dynamic characteristics of an axial compressor under different working conditions vary greatly. The dynamic modeling for different working conditions means a huge-scale model cluster, which will lead to a sharp increase in the computational complexity of the detection algorithm and is not conducive to ensuring the real-time performance of aerodynamic instability prediction. Summary of the Invention

[0007] To solve the above problems, the present invention proposes an aerodynamic instability prediction method and system for an aero-engine based on a sampling observer, which combines the sampling observer and the RBF neural network to accurately identify the dynamics of the instability mode from single-sensor data.

[0008] In some embodiments, the following technical solutions are adopted:

[0009] An aerodynamic instability prediction method for an aero-engine based on a sampling observer, comprising:

[0010] Obtaining single-sensor pressure data during the instability development process of the aero-engine under different working conditions;

[0011] Taking the differential equation with the single-sensor data and its derivative as state variables as the dynamic model describing the normal mode and the instability precursor mode during the instability development process of the engine;

[0012] Using a sampling high-gain observer to estimate the derivative information from the single-sensor data, and constructing a neural network identification model within the state trajectory distribution region formed by the single-sensor pressure data and its derivative information, where the neural network is used to identify the dynamic models of the normal mode and the instability precursor mode in each instability test;

[0013] Fusing and expressing the neural network identification models of the normal mode or the instability precursor mode in all instability tests under similar working conditions, constructing a knowledge base containing the instability characteristics under multiple working conditions, and dividing the fusion models in the knowledge base into different decision groups according to different working conditions;

[0014] Construct a dynamic estimator using the neural network fusion model in the knowledge base, and use the single-measurement point pressure data measured by the aero-engine as the input of the dynamic estimator. Each decision group independently performs precursor mode detection to obtain the corresponding secondary decision results, and comprehensively processes the secondary decision results of different decision groups to obtain the primary decision result.

[0015] As a further solution, the dynamic models of the normal mode and the instability precursor mode during the instability development process of the engine are specifically:

[0016]

[0017] Among them, A, B, and C are coefficient matrices respectively, x is the state variable, φ is the output variable, f(·) is a nonlinear function, and the dynamic function corresponding to the normal mode is denoted as f N (·), and the dynamic function corresponding to the instability precursor mode is denoted as f S (·).

[0018] As a further solution, estimate the state of the dynamic model using a sampling observer, specifically:

[0019]

[0020] Where ψ(k) ∈ R 2 , is the observation result of the state of the dynamic model. A do , B do , C do and D do are coefficient matrices.

[0021] As a further solution, construct a neural network identification model corresponding to different instability tests, specifically:

[0022]

[0023] Among them, χ is the identifier state, is the derivative of the single-measurement point data, α is the identifier gain, is the estimated value of the neural network weight, is the neural network basis, T is the sampling period, used to learn the nonlinear dynamics of the normal or instability precursor mode.

[0024] As a further solution, fuse and express the neural network identification models of the normal mode or the instability precursor mode under similar working conditions, specifically:

[0025]

[0026]

[0027]

[0028]

[0029] Among them, respectively represent the radial basis function vectors corresponding to multiple instability tests under similar working conditions, and respectively represent the neural network identification results of multiple instability tests under similar working conditions, and respectively represent the radial basis function sub-vectors near the trajectory in multiple instability tests under similar working conditions.

[0030] As a further solution, within each decision group, an instability prediction index reflecting the degree of dynamic matching is used for precursor mode detection, and the instability prediction index is:

[0031] In the dynamic estimator corresponding to each decision group, based on the single-point pressure data measured by the aero-engine, the estimator state is obtained, and the difference between the estimator state and the derivative of the single-point pressure data is taken to obtain the estimator residual; the average L1 norm of the residuals on the sliding window is used as the instability prediction index.

[0032] As a further solution, the dynamic estimator is specifically:

[0033]

[0034] Among them, T is the sampling period, is the estimator state, is the neural network fusion model in the knowledge base, with the subscript p used to distinguish different working conditions and the superscript j used to distinguish the normal mode or the instability precursor mode, is the neural network basis, is the observed result of the dynamic model state corresponding to the measured sequence data φ r (k), is the derivative of the measured single-point pressure data.

[0035] As a further solution, the instability prediction index is specifically:

[0036]

[0037]

[0038] Among them, T e is the length of the sliding window; is the estimator state, is the derivative of the measured single-point pressure data.

[0039] As a further solution, the multiple secondary decision results are comprehensively processed to obtain the primary decision result, specifically:

[0040] If the number of secondary decision results of the instability precursor mode detected exceeds the set number, the primary decision result outputs an alarm signal for aerodynamic instability.

[0041] In some other embodiments, the following technical solution is adopted:

[0042] An aeroengine aerodynamic instability prediction system based on a sampling observer, comprising:

[0043] A data acquisition module, configured to acquire single-point pressure data during the instability development process of the aeroengine under different working conditions;

[0044] A model construction module, configured to use a differential equation with single-point data and its derivative as state variables as a dynamic model describing the normal mode and the instability precursor mode during the instability development process of the engine;

[0045] A model identification module, configured to estimate derivative information from single-point data by sampling a high-gain observer, and construct a neural network identification model within the state trajectory distribution region formed by the single-point pressure data and its derivative information, where the neural network is used to identify the dynamic models of the normal mode and the instability precursor mode in each instability test;

[0046] A knowledge base construction module, configured to fuse and express the neural network identification models of the normal mode or the instability precursor mode in all instability tests under similar working conditions, construct a knowledge base containing instability characteristics under multiple working conditions, and divide the fusion models in the knowledge base into different decision groups according to different working conditions;

[0047] An aerodynamic instability prediction module, configured to construct a dynamic estimator by using the neural network fusion model in the knowledge base, and use the actually measured single-point pressure data of the aeroengine as the input of the dynamic estimator. Each decision group independently performs precursor mode detection to obtain corresponding secondary decision results, and comprehensively processes the secondary decision results of different decision groups to obtain the primary decision result.

[0048] In some other embodiments, the following technical solution is adopted:

[0049] A terminal device, comprising a processor and a memory, where the processor is configured to implement instructions; the memory is configured to store multiple instructions, and the instructions are suitable for being loaded and executed by the processor to perform the above-mentioned aeroengine aerodynamic instability prediction method based on a sampling observer.

[0050] Compared with the prior art, the beneficial effects of the present invention are:

[0051] (1) The present invention designs a learning and prediction algorithm for aero-engine aerodynamic instability from the perspective of the sampling system, using only the measurement data of a single sensor. The present invention can weaken the damage to the hardware structure of the aero-engine caused by drilling and installing sensors and the influence on the internal air flow, and reduce the cost consumption of the sensors. At the same time, this algorithm design based on the sampling system is more suitable for the digital execution of engineering tasks.

[0052] (2) Compared with the instability precursor detection method based on signal analysis, the present invention can obtain and utilize the state and dynamic information in the aero-engine instability process more comprehensively and fully. The state information of the dynamic mode is recovered from the single-point measurement data by using a sampling observer, and then along the estimated state trajectory, the dynamic information of the instability mode is obtained by using an RBF neural network, and the dynamic modeling result along the state trajectory is stored in the form of a constant neural network. This neural network model containing both state information and dynamic information can more profoundly reflect the instability characteristics of the aero-engine.

[0053] (3) Compared with the fault detection method based on neural networks, the present invention combines a sampling observer and an RBF neural network to accurately identify the dynamics of the instability mode from single-point measurement data. Along the estimated trajectory of the sampling observer, the RBF neural network satisfies the condition of persistent excitation. The satisfaction of this condition can ensure that the neuron weights near the trajectory converge exponentially to the ideal true value, and the neuron weights far from the trajectory remain zero, so as to accurately identify the nonlinear dynamic function along the instability mode trajectory.

[0054] (4) Compared with the existing instability prediction algorithms based on deterministic learning, the present invention combines a sampling observer and a dynamic mode fusion algorithm to fuse and reconstruct the local accurate identification results of different instability modes under similar working conditions. This fusion process can enhance the expression and generalization ability of the RBF neural network, effectively reduce the scale of the knowledge base, and thus ensure the real-time performance of subsequent instability prediction.

[0055] (5) Compared with the existing instability prediction algorithms for a single fixed working condition, the present invention proposes a hierarchical decision-making mechanism for the instability prediction task under multiple working conditions. This mechanism detects the precursor mode by dividing different decision-making groups, avoiding misjudgment caused by the confusion of decision-making indicators under different working conditions, and improving the accuracy of the engine operating state prediction.

[0056] Other features and advantages of the present invention will be partially given in the following description, partially become obvious from the following description, or be understood through the practice of this aspect. Brief Description of the Drawings

[0057] Figure 1 It is a schematic structural diagram of the aero-engine in the embodiment of the present invention;

[0058] Figure 2 It is the flow chart of the aero - engine aerodynamic instability prediction method based on a sampling observer in the embodiment of the present invention;

[0059] Figure 3 It is the state trajectory of the instability mode reconstructed by the sampling observer in the embodiment of the present invention;

[0060] Figure 4 It is the convergence result of the weights of some neurons in the RBF neural network in the embodiment of the present invention;

[0061] Figure 5 It is the RBF neural network identification result of different instability modes in a single instability test in the embodiment of the present invention;

[0062] Figure 6 It is the fusion result of the normal mode under similar working conditions in the embodiment of the present invention;

[0063] Figure 7 It is the fusion result of the instability precursor mode under similar working conditions in the embodiment of the present invention;

[0064] Figure 8 It is the identification index of the estimator under multiple working conditions in the embodiment of the present invention. Detailed implementation manners

[0065] It should be noted that the following detailed description is exemplary and is intended to provide further illustration of the present application. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meaning as commonly understood by those of ordinary skill in the technical field to which this application belongs.

[0066] It should be noted that the terms used herein are only for describing the specific implementation manners and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they specify the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0067] Embodiment 1

[0068] In one or more embodiments, an aero - engine aerodynamic instability prediction method based on a sampling observer is disclosed. Combining Figure 2 , the specific process is as follows:

[0069] S101: Obtain the single - point pressure data of the aero - engine during the instability development process under different working conditions;

[0070] In this embodiment, the overall structure of the aero - engine is as Figure 1As shown. A dynamic pressure sensor is installed on the casing wall in front of the compressor rotor. By gradually closing the throttle plug cone, the flow rate of the compressor is reduced to induce surging, thereby obtaining single-point measurement data during the instability development process. In this embodiment, different working conditions are simulated by adjusting the ambient temperature, humidity, etc., which can be roughly divided into 4 types. 6 instability tests are carried out for each working condition, and a total of 4×6 = 24 single-variable time series are used as the training set.

[0071] First, the baseline drift in the original data is eliminated by the following method:

[0072]

[0073] where, φ ori is the original single-point measurement data, and φ pre is the single-point measurement data after eliminating the baseline drift. In this embodiment, T s is set to 100. In addition, considering that the frequency range of the surge signal is 5 - 30 Hz and the frequency range of rotating stall is 20 - 130 Hz, in order to filter out the influence of high-frequency noise, a Butterworth filter is used for low-pass filtering with a cut-off frequency of 140 Hz.

[0074] The filtered single-point measurement data is used as the test data (denoted as φ(k)), and the learning and prediction research on aerodynamic instability is carried out.

[0075] S102: Use the differential equation with the single-point measurement data and its derivative as state variables as the dynamic model to describe the normal mode and instability precursor mode during the instability development process of the engine;

[0076] Specifically, the dynamic model is as follows:

[0077]

[0078] where, x ∈ R 2 and φ ∈ R are the state and output variables respectively, and the coefficient matrix is C = [1 0], and the non-linear function f(·): R 2 →R represents the internal dynamics (unknown non-linear function) of the engine operating mode. In this embodiment, 5 normal modes and 1 instability precursor mode are intercepted from the single-point measurement data of each instability experiment. The dynamics corresponding to the normal mode are denoted as f N (·), N ∈ {1, 2, 3, 4, 5}; the dynamics corresponding to the instability precursor mode are denoted as f S (·).

[0079] The single-point measurement sampling data is denoted as φ(k): = φ(kT), where T > 0 is the sampling period and k represents time.

[0080] In this embodiment, T = 0.001 s is selected, and the single measurement point data φ(k) can be described by the following discrete model

[0081]

[0082] where

[0083] S103: Estimate the derivative information of the single measurement point data by using a sampled high-gain observer. Within the state trajectory distribution region formed by the single measurement point pressure data and its derivative information, construct a neural network identification model, and the neural network is used to identify the dynamic models of the normal mode and the instability precursor mode for each instability test;

[0084] For the single measurement point data φ(k) of each instability test in the training set, use a sampled high-gain observer to estimate the derivative information:

[0085]

[0086] where ψ(k) ∈ R 2 , represents the state estimation result of the engine dynamic model.

[0087] Taking the bilinear transformation as an example in this embodiment, the coefficient matrix is designed as H o =[h 1 ,h 2 T , A o =A - H o C, ∈ is a very small positive constant. Select H o such that the characteristic roots of s 2 +h 1 s + h 2 = 0 have negative real parts. In this embodiment, H o =[4, 3] T ,

[0088] The trajectory reconstruction results of the normal and precursor modes in a certain instability test by the observer are as Figure 3 shown.

[0089] Within the distribution region of the reconstructed trajectory, construct an RBF neural network identifier:

[0090]

[0091] where χ is the identifier state, is the derivative of the single measurement point data, α = 0.5 is the identifier gain, ​​​​​​​​L = 1456 is the estimated value of the neural network weights, which is the neural network basis, used to learn the non - linear dynamics of normal or instability precursor patterns.

[0092] In this embodiment, the Gaussian function is selected as the neural network basis, that is:

[0093]

[0094] where η = 0.2, and the neuron center ξ l ∈R 2 is uniformly distributed in the plane of [- 3,2]×[- 6,5].

[0095] The neural network weight adjustment law is:

[0096]

[0097] where γ = 10000, σ = 0.001.

[0098] According to the certainty learning theory, the RBF sub - vector composed of neurons near the trajectory satisfies the persistent excitation condition, and the weight sub - vector of the corresponding neurons exponentially converges to the ideal true value; for neurons far from the trajectory, the weight sub - vector remains 0 (as shown in Figure 4 ). Select the estimated weight after the transient to calculate the constant weight:

[0099]

[0100] where [k a ,k b represents the time period after the transient. The calculated constant RBF neural network is used as the modeling result of the instability mode dynamics f(·) of the aero - engine.

[0101] In this embodiment, according to the 24 univariate time series in the training set, a total of 5×24 = 120 modeling results of normal patterns and 24 modeling results of precursor patterns are obtained.

[0102] Figure 5 Shows the modeling results of 5 normal - mode dynamics f N (·) and 1 precursor - mode dynamics f in a certain instability test S (·)

[0103] In this embodiment, the processing result of the sampling high-gain observer for single-measurement point data is used as the reconstruction trajectory of the engine operation mode. Local accurate learning of the dynamic function of the instability mode is realized within the reconstruction trajectory region. Within the distribution region of the reconstruction trajectory, the sub-vectors composed of the RBF neural network satisfy the persistent excitation condition, which means that the estimated weights corresponding to the neurons in the region near the trajectory can exponentially converge to the ideal true value, and the estimated weights corresponding to the neurons far from the trajectory remain zero, thus ensuring the accurate identification of the dynamic function of the instability mode by the RBF neural network along the trajectory.

[0104] S104: Fuse and express the neural network identification models under similar working conditions to construct a knowledge base containing the pneumatic instability characteristics under various working conditions, and divide the fusion models in the knowledge base into different decision groups according to different working conditions;

[0105] In this embodiment, for the modeling results of Q modes under similar working conditions, the neural network parameter reconstruction is carried out using the following formula

[0106]

[0107] Among them,

[0108]

[0109]

[0110] Among them, respectively represent the radial basis function vectors corresponding to multiple instability tests under similar working conditions, and respectively represent the neural network identification results of multiple instability tests under similar working conditions, and respectively represent the radial basis function sub-vectors near the trajectory in multiple instability tests under similar working conditions.

[0111] Using the above fusion model, a dynamic knowledge base is constructed This knowledge base contains the dynamic fusion models of the normal mode j = N and the precursor mode j = S under P different working conditions. In this embodiment, mode fusion is carried out for 5×6 = 30 normal modes and 1×6 = 6 precursor modes under similar working conditions respectively. Figure 6 and Figure 7 respectively show the fusion models characterizing the normal and unstable operating states of the engine under a certain working condition. Finally, a knowledge base characterizing the engine operating states under four working conditions can be obtained, with a total of (1 + 1)×4 = 8 neural network fusion models. These fusion models are divided into four decision groups, and each decision group contains the normal mode fusion model and the precursor mode fusion model under the same working condition.

[0112] In this embodiment, the dynamic identification results of instability modes under similar working conditions are reconstructed to achieve the dynamic expression of instability modes within a larger state space range, significantly reducing the scale and redundancy of the knowledge base.

[0113] S105: Obtain the single-point pressure data measured by the aero-engine, use the measured single-point pressure data to perform precursor mode detection in each decision group to obtain the corresponding secondary decision results, and comprehensively process the secondary decision results of different decision groups to obtain the primary decision results.

[0114] In this embodiment, within each decision group, precursor mode detection is performed using an instability prediction index that reflects the degree of dynamic matching. The instability prediction index is:

[0115] Dynamic estimators are respectively constructed using the normal mode fusion model and the precursor mode fusion model of each decision group. Based on the single-point pressure data measured by the aero-engine, the estimator state is obtained, and the difference between the estimator state and the derivative of the measured single-point pressure data is calculated to obtain the estimator residual; the average L1 norm of the residuals on the sliding window is used as the instability prediction index.

[0116] Specifically, for the measured flow data φ r (k) (with different working conditions from the training sequence), a dynamic estimator is constructed using the knowledge base Ψ as follows to monitor the operating state of the aero-engine:

[0117]

[0118] where b = 0.5, is the estimator state, the subscript p is used to distinguish different working conditions, and the superscript j is used to distinguish the normal mode or the instability precursor mode. is the state observation result of the dynamic model corresponding to the measured sequence data φ r (k).

[0119] In this embodiment, is provided by the following sampled high-gain observer:

[0120]

[0121] Define the estimator residual Design the following instability prediction index:

[0122]

[0123] where T e = 500 is the length of the sliding window.

[0124] In this embodiment, the dynamic estimator embeds a fusion model of instability modes under different working conditions. Based on the dynamic differences between the fusion model and the measured mode, an estimator residual is generated. Thus, the estimator residual is used as an external manifestation of the internal dynamic differences, and real-time detection / identification of the instability precursor mode is performed from the perspective of the stability analysis of the residual system.

[0125] Based on the obtained instability prediction index, this embodiment formulates the following decision-making mechanism:

[0126] S1051: For the prediction index of each estimator, it is divided into different decision groups according to the working condition category. Each decision group independently performs precursor mode detection, and the detection result is used as the secondary decision result of the instability prediction system (see Figure 8 ). Specifically, when the engine is operating normally, the dynamic difference between the actual operating mode and the normal mode in the knowledge base is relatively small. Then, within each decision group, the prediction index corresponding to the normal estimator is less than the prediction index corresponding to the precursor mode

[0127] Once there is a precursor to unstable operation of the engine, within the decision group that matches the working condition, the prediction index corresponding to the normal estimator will gradually increase, while the prediction index corresponding to the precursor estimator will gradually decrease until it is less than

[0128] At this time, it indicates that the operating state of the engine may match the instability precursor mode in this decision group, and this decision group generates a switch signal as the secondary decision result.

[0129] S1052: The primary decision-making mechanism simultaneously monitors all secondary decision groups and gives the final instability prediction result.

[0130] Considering that there may be significant differences in the dynamic characteristics of engine aerodynamic instability under different working conditions, even when the engine is operating normally, within the decision group with unmatched working conditions, it is possible that the prediction index corresponding to the precursor estimator is less than the prediction index corresponding to the normal estimator This will lead to false alarms of aerodynamic instability. Therefore, it is necessary to comprehensively process the secondary decision results of different decision groups in combination with actual engineering applications.

[0131] In this embodiment, the primary decision-making mechanism monitors that decision groups 2 and 3 simultaneously generate switch signals at 3.349 s. At this time, an alarm signal is given, realizing the early prediction of aerodynamic instability, as shown in Figure 8 shown.

[0132] The hierarchical decision-making mechanism in this embodiment divides dynamic estimators with significant operating condition differences into different decision groups, enhancing the robustness of the decision-making process, thereby effectively weakening the instability misjudgment caused by the confusion of estimator residuals when the operating conditions change.

[0133] It should be noted that the knowledge base of the instability characteristics in this embodiment is established through offline learning. By collecting more instability test data, the dynamic knowledge base can be updated, increasing the expression ability of the knowledge base for the instability characteristics of aeroengines. The instability prediction process is applicable to online prediction tasks. This hierarchical decision-making mechanism based on the fusion model and estimator residuals can avoid constructing a large number of dynamic estimators and does not require complex dynamic identification of the measured mode, thus effectively ensuring the real-time nature of the prediction process.

[0134] Embodiment 2

[0135] An aeroengine aerodynamic instability prediction system based on a sampling observer, comprising:

[0136] A data acquisition module for acquiring single-point pressure data during the instability development process of an aeroengine under different operating conditions;

[0137] A model construction module for using a differential equation with single-point data and its derivative as state variables as the dynamic model describing the normal mode and instability precursor mode during the instability development process of the engine;

[0138] A model identification module for using a high-gain observer to estimate derivative information from single-point data, and constructing a neural network identification model within the state trajectory distribution area formed by the single-point pressure data and its derivative information, where the neural network is used to identify the dynamic models of the normal mode and instability precursor mode in each instability test;

[0139] A knowledge base construction module for fusing and expressing the neural network identification models of the normal mode or instability precursor mode in all instability tests under similar operating conditions, constructing a knowledge base containing instability characteristics under various operating conditions, and dividing the fusion models in the knowledge base into different decision groups according to different operating conditions;

[0140] An aerodynamic instability prediction module for constructing a dynamic estimator using the neural network fusion model in the knowledge base, taking the actually measured single-point pressure data of the aeroengine as the input of the dynamic estimator, independently detecting the precursor mode for each decision group to obtain corresponding secondary decision results, and comprehensively processing the secondary decision results of different decision groups to obtain a primary decision result.

[0141] It should be noted that the specific implementation methods of the above modules are the same as those in Embodiment 1 and will not be elaborated here.

[0142] Embodiment 3

[0143] In one or more embodiments, a terminal device is disclosed, including a server. The server includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the method for predicting aeroengine aerodynamic instability based on a sampling observer in the first embodiment. For the sake of brevity, it will not be elaborated here.

[0144] It should be understood that in this embodiment, the processor may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0145] The memory may include a read-only memory and a random access memory, and provide instructions and data to the processor. A part of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

[0146] In the implementation process, the steps of the above method may be completed by the integrated logic circuit in the hardware of the processor or the instructions in the form of software.

[0147] Although the specific implementation manners of the present invention are described above in conjunction with the drawings, it is not a limitation on the protection scope of the present invention. Those skilled in the art should understand that on the basis of the technical solution of the present invention, various modifications or deformations that can be made without creative labor by those skilled in the art are still within the protection scope of the present invention.

Claims

1. A method for predicting aeroengine aerodynamic instability based on a sampling observer, characterized in that, it includes: Obtain the single-point pressure data during the instability development process of the aeroengine under different working conditions; Use the differential equation with the single-point data and its derivative as state variables as the dynamic model describing the normal mode and instability precursor mode during the instability development process of the engine; Use a sampling high-gain observer to estimate the derivative information from the single-point data. In the state trajectory distribution area formed by the single-point pressure data and its derivative information, construct a neural network identification model, and the neural network is used to identify the dynamic models of the normal mode and instability precursor mode in each instability test; Fuse and express the neural network identification models of the normal mode or instability precursor mode in all instability tests under similar working conditions, construct a knowledge base containing instability characteristics under various working conditions, and divide the fusion models in the knowledge base into different decision groups according to different working conditions; Use the neural network fusion model in the knowledge base to construct a dynamic estimator, and use the actually measured single-point pressure data of the aeroengine as the input of the dynamic estimator. Each decision group independently conducts precursor mode detection to obtain corresponding secondary decision results, and comprehensively processes the secondary decision results of different decision groups to obtain a primary decision result; Among them, constructing the neural network identification models corresponding to different instability tests specifically is: where χ is the identifier state, is the derivative of the single measurement point data, α is the identifier gain, is the estimated value of the neural network weights, is the neural network basis, T is the sampling period, used to learn the non-linear dynamics of normal or pre-failure modes; Fuse and express the neural network identification models under similar working conditions, specifically: Among them, respectively represent the radial basis function vectors corresponding to multiple instability tests under similar working conditions, and respectively represent the neural network identification results of multiple instability tests under similar working conditions, and respectively represent the radial basis function sub-vectors near the trajectory in multiple instability tests under similar working conditions.

2. A method for predicting aeroengine aerodynamic instability based on a sampling observer according to claim 1, characterized in that, the dynamic models of the normal mode and instability precursor mode during the instability development process of the engine are specifically: Among them, A, B, and C are coefficient matrices respectively, x is the state variable, φ is the output variable, and f(·) is a non-linear function, which can be the function f corresponding to the normal mode N (·), or the dynamics corresponding to the pre-failure mode is denoted as f S (·).

3. A method for predicting aeroengine aerodynamic instability based on a sampling observer according to claim 1, characterized in that, Within each decision group, use an instability prediction index reflecting the degree of dynamic matching for precursor mode detection, and the instability prediction index is: Use the normal mode fusion model and instability precursor fusion model in each decision group to construct dynamic estimators respectively. Based on the actually measured single-point pressure data of the aeroengine, obtain the estimator state, subtract the derivative of the actually measured single-point pressure data from the estimator state to obtain the estimator residual; use the average L1 norm of the residuals on the sliding window as the instability prediction index.

4. A method for predicting aeroengine aerodynamic instability based on a sampling observer according to claim 3, characterized in that, the dynamic estimator is specifically: where T is the sampling period, is the estimator state, is the neural network fusion model in the knowledge base, the subscript p is used to distinguish different working conditions, and the superscript j is used to distinguish the normal mode or the precursor mode of instability, is the neural network basis, is the observed result of the kinetic model state corresponding to the measured sequence data φ r (k), is the derivative of the measured single-point pressure data.

5. A method for predicting aeroengine aerodynamic instability based on a sampling observer according to claim 3, characterized in that, the instability prediction index is specifically: Among them, T e is the length of the sliding window; is the estimator state, is the derivative of the measured single-point pressure data.

6. A method for predicting aeroengine aerodynamic instability based on a sampling observer according to claim 1, characterized in that, Comprehensively process the secondary decision results of different decision groups to obtain a primary decision result, specifically: If the number of secondary decision results detecting the instability precursor mode exceeds the set number, the primary decision result outputs an alarm signal for aerodynamic instability.

7. A system for predicting aeroengine aerodynamic instability based on a sampling observer, It is characterized in that including a data acquisition module for acquiring single-point pressure data during the instability development process of an aero-engine under different working conditions a model construction module for using a differential equation with single-point data and its derivative as state variables as a dynamic model describing the normal mode and instability precursor mode during the instability development process of the engine a model identification module for estimating derivative information from single-point data by sampling a high-gain observer, and constructing a neural network identification model within the state trajectory distribution region formed by the single-point pressure data and its derivative information, where the neural network is used to identify the dynamic models of the normal mode and instability precursor mode in each instability test a knowledge base construction module for fusing and expressing the neural network identification models of the normal mode or instability precursor mode in all instability tests under similar working conditions, constructing a knowledge base containing instability characteristics under various working conditions, and dividing the fusion models in the knowledge base into different decision groups according to different working conditions an aerodynamic instability prediction module for using the neural network fusion model in the knowledge base to construct a dynamic estimator, taking the single-point pressure data measured by the aero-engine as the input of the dynamic estimator, independently detecting the precursor mode for each decision group to obtain corresponding secondary decision results, and comprehensively processing the secondary decision results of different decision groups to obtain a primary decision result wherein, constructing the neural network identification models corresponding to different instability tests specifically includes where χ is the identifier state, is the derivative of single measurement point data, α is the identifier gain, is the estimated value of the neural network weight, is the neural network basis, T is the sampling period, used to learn the non-linear dynamics of normal or precursor-to-instability patterns; fusing and expressing the neural network identification models under similar working conditions, specifically including Among them, respectively represent the radial basis function vectors corresponding to multiple instability tests under similar working conditions, and respectively represent the neural network identification results of multiple instability tests under similar working conditions, and respectively represent the radial basis function sub-vectors near the trajectory in multiple instability tests under similar working conditions.

8. A terminal device, which includes a processor and a memory, the processor is used to implement instructions; the memory is used to store multiple instructions It is characterized in that the instructions are suitable for being loaded and executed by the processor to perform the aero-engine aerodynamic instability prediction method based on a sampling observer according to any one of claims 1-6

Images