Aero-engine test state online identification method and device

Abstract

The invention provides an aero-engine test state on-line identification method and device, and the method comprises the steps: obtaining a minimum characteristic parameter set according to the technical indexes clearly specified in the general specification of an aero-engine flight test and the most relevant parameters when the corresponding state is reached; preprocessing flight parameter data corresponding to the minimum characteristic parameter set, and establishing an aero-engine test state online identification model; a sample data set of the aero-engine in the actual flight process is collected, data are input into the aero-engine test state online recognition model, and the aero-engine test state is determined based on output of the aero-engine test state online recognition model; the hidden risk in the real-time flight process can be reduced, and the safety coefficient is improved.

Description

Technical Field

[0001] This invention belongs to the field of flight test data processing and monitoring technology, and relates to an online identification method and device for the test status of aero-engines. Background Technology

[0002] An aero-engine is a device that provides the thrust required for aircraft flight. Based on thrust magnitude, the operating states of an aero-engine can be categorized into shutdown, idle, throttle, intermediate, partial afterburner, and full afterburner operating states. Accurate identification of these states during real-time monitoring of flight tests is a prerequisite and foundation for analyzing engine performance, verifying engine reliability, and reducing flight test risks, and thus has significant research importance and value.

[0003] Under the current monitoring model, online identification of aero-engine test status still relies mainly on manual data interpretation, lacking digital and automated support methods. With the explosive growth in the number of test flights and the significant increase in the amount of telemetry monitoring parameters, monitoring personnel face time-consuming and labor-intensive manual identification in scenarios where telemetry data has weak regularity and relies excessively on human experience, which is prone to misjudgment and omissions, seriously threatening test flight safety.

[0004] Therefore, there is an urgent need to study a method that can autonomously identify the test status of aero-engines, reduce hidden risks during real-time flight, and improve the safety factor. Summary of the Invention

[0005] To address the problems of time-consuming and labor-intensive manual identification of engine test conditions in existing technologies, which is prone to misjudgment and omission, seriously threatening flight test safety, this invention provides an online identification method and device for aero-engine test conditions. The invention designs a hybrid model based on a convolutional-conditional variational autoencoder and a multi-self-attention-long short-term memory network for aero-engine condition identification. The technical solution is as follows: Firstly, a method for online identification of the test status of an aero-engine is provided, comprising the following steps: The first step is to obtain the minimum set of characteristic parameters based on the technical indicators clearly specified in the general specifications for flight testing of aero engines and the parameters that are most relevant when the corresponding state is reached; The second step is to preprocess the flight parameter data corresponding to the minimum feature parameter set and establish an online identification model for the test status of aero-engines. The third step is to collect sample datasets of aero-engines during actual flight, input the data into the online identification model of aero-engine test status, and determine the test status of the aero-engine based on the output of the online identification model of aero-engine test status.

[0006] Optionally, the first step specifically includes: Based on the technical indicators clearly specified in the general specifications for aero-engine flight testing and the parameters most relevant when the corresponding state is reached, a preliminary set of original feature parameters required for online identification of aero-engine test states is formed. Principal component analysis is used to reduce the dimensionality of the feature parameter set and reduce its attributes to form the minimum feature parameter set.

[0007] Optionally, principal component analysis is used to reduce the dimensionality and attributes of the feature parameter set. The specific process includes: Prepare the original feature parameter set; Calculate the mean of the original feature parameter set; Calculate the covariance matrix C of the original feature parameter set; Calculate the eigenvalues ​​and corresponding eigenvectors of the covariance matrix C; Calculate the contribution of each principal component; Arrange the eigenvectors of the principal components in descending order of contribution to form the feature matrix U; The principal component is determined based on the defined cumulative contribution threshold; The sample feature matrix X is transformed using the feature matrix U to obtain the minimum set of feature parameters.

[0008] Optionally, the formula used to calculate the contribution of each principal element is: P k The contribution of each principal element.

[0009] Optionally, the formula used to determine the principal component sample set is: , This is the mean of the sample dataset.

[0010] Optionally, the second step specifically includes: The flight parameter data corresponding to the minimum feature parameter set are preprocessed by outlier removal and normalization correction methods, and known engine test states are labeled by sample labels to form the sample dataset required for model training. A conditional variational autoencoder containing a decoder, encoder, and conditional input module is constructed, and a convolutional module is introduced into the conditional variational autoencoder to obtain a convolutional-conditional variational autoencoder. The local feature patterns of the aero-engine sample dataset are extracted by optimizing the loss function. Using the sample dataset as the input to the encoder in the convolutional-conditional variational autoencoder, the output of the encoder is the local feature pattern of the aero-engine sample dataset. A long short-term memory network model is constructed, and a multi-self-attention mechanism is introduced to assign weights to the hidden states output by the long short-term memory network model, forming a multi-self-attention mechanism-long short-term memory network model. The output of the convolutional-conditional variational autoencoder is used as the input of the multi-self-attention mechanism-long short-term memory network model. The output of the multi-self-attention mechanism-long short-term memory network model is the long-term dependency relationship and different spatial and temporal information in the aero-engine sample dataset. The output of the multi-self-attention mechanism-long short-term memory network model is used as the input of the decoder in the convolutional-conditional variational autoencoder. The output of the decoder is the final prediction result of the aero-engine test state. Thus, the online recognition model of aero-engine test state is trained.

[0011] Optionally, the formula used when preprocessing the flight parameter data corresponding to the minimum feature parameter set using outlier removal and normalization correction is as follows: X max X represents the maximum value of the sample elements. min It represents the minimum value of the sample elements.

[0012] Optionally, the formula used to guide the extraction of local feature patterns from the aero-engine sample dataset by optimizing the loss function is as follows:

[0013] In the formula, Represents the reconstruction loss, which is the difference between the model output and the input; KL represents the Kullback-Leible divergence, which is the difference between the variational distribution in the latent space and the prior distribution.

[0014] Optionally, a multiple self-attention mechanism is introduced, and the formula used to assign weights to the hidden states output by the Long Short-Term Memory network model is as follows: w and b are the weight matrix and bias of the final linear layer, respectively.

[0015] In a second aspect, an online identification device for the test status of an aero-engine is provided, for performing any of the methods described in the first aspect, the device comprising: The processing module is used to obtain a minimum set of characteristic parameters based on the technical specifications clearly defined in the engine's general specifications and the main parameters directly related to achieving the corresponding state. A module is established to preprocess the flight parameter data corresponding to the minimum feature parameter set and to build an online identification model for aero-engine test conditions. The determination module is used to collect sample datasets of aero-engines during actual flight, input the data into the online identification model of aero-engine test status, and determine the test status of aero-engines based on the output of the online identification model of aero-engine test status.

[0016] The beneficial effects of this invention are as follows: 1. Principal component analysis was used to construct the minimum set of feature parameters required for aero-engine state identification, which reduced the model size while ensuring the model identification accuracy; 2. A pre-trained convolutional-conditional variational autoencoder method is used to learn the latent representation of aero-engine time-series feature data, forming stable local features in each time window, and accurately extracting the hidden feature patterns in aero-engine time-series feature data; 3. Introduce multiple self-attention mechanisms to optimize the long short-term memory neural network, learn the long-term dependencies and change patterns of time-series feature data, capture the relationships between multiple variables and relevant information in different subspaces, highlight the importance of different features, and improve the performance of the aero-engine operating status recognition model; 4. The intelligent recognition method based on convolutional variational autoencoder and multi-attention-long short-term memory network can effectively improve the accuracy and efficiency of identifying the working status of aero-engines during real-time monitoring of flight tests, reduce over-reliance on expert experience, and reduce the problems of misjudgment and time-consuming manual identification. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the overall structure of the aircraft engine condition recognition model; Figure 2 This is a schematic diagram of a convolutional-conditional variational autoencoder structure. Figure 3 This is a schematic diagram of a long short-term memory neural network structure; Figure 4 This is a schematic diagram of a long short-term memory network model based on multiple self-attention. Figure 5 This is a schematic diagram of the online status identification results for an aircraft engine. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The features and illustrative embodiments of various aspects of the present invention will now be described in detail. Numerous specific details are set forth in the following detailed description to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention may be practiced without requiring some of these specific details. The following description of embodiments is merely intended to provide a better understanding of the invention by illustrating examples of the invention. The invention is by no means limited to any specific setups and methods set forth below, but covers any improvements, substitutions, and modifications to structures, methods, and devices without departing from the spirit of the invention. Well-known structures and techniques are not shown in the drawings and the following description to avoid unnecessarily obscuring the invention.

[0021] It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other, and the various embodiments can be referenced and cited in each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0023] The specific implementation process of this invention is as follows: (1) Selection of characteristic flight parameter data The technical specifications clearly defined in the general specifications for flight testing of aero engines, and the parameters most relevant to achieving the corresponding states, are used as benchmarks. For example, the required speed N = Nmax is specified under the maximum engine operating condition; under the engine idle operating condition, the speed is usually 0.4Nmax to 0.6Nmax, etc.

[0024] If the parameter with the highest relevance is missing, a parameter with a similar physical characterization is selected for replacement. For example, if the turbine inlet gas temperature T4 is missing, the turbine outlet gas temperature T6, which has a similar physical characterization, can be selected as the characteristic flight parameter data.

[0025] If there is deep coupling between characteristic flight parameter data, parameters that are strongly correlated with engine operating conditions are selected. For example, if there is a correlation between high-pressure speed and low-pressure speed, considering the strong sensitivity of high-pressure to high-pressure speed, high-pressure speed is selected as the characteristic flight parameter data.

[0026] (2) Feature parameter attribute reduction Principal Component Analysis (PCA) was used for feature dimensionality reduction and data compression of flight parameter data. The principle is to preserve as many hidden features as possible from the original data, using linear transformations to reduce the dimensionality of multiple dependent variables to at least a few independent composite variables. The specific process is as follows: ① Prepare the original feature parameter set; ② Calculate the mean of the original feature parameter set; ③ Calculate the covariance matrix C of the original feature parameter set; ④ Calculate the eigenvalues ​​and corresponding eigenvectors of the covariance matrix C; ⑤ Calculate the contribution of each principal component; ⑥ Arrange the eigenvectors of the principal components in descending order of contribution to form the feature matrix U; ⑦ Determine the principal components based on the defined cumulative contribution threshold; ⑧ Transform the sample feature matrix X using the feature matrix U; ⑨ Determine the minimum feature parameter set.

[0027] The specific calculation process is as follows: Suppose we have an n-dimensional original feature parameter set:

[0028] Calculate the mean of the original feature parameter set. :

[0029] Calculate the covariance matrix C of the original feature parameter set:

[0030] Through orthogonal decomposition, the eigenvalues ​​and corresponding eigenvectors of the covariance matrix C are obtained:

[0031] In the formula: It is a diagonal matrix consisting of n eigenvalues ​​of C ordered in descending order of dimensionality. Composition. Feature matrix From eigenvalues Corresponding feature vector The characteristic matrix formed by this.

[0032] Calculate the contribution of each element:

[0033] Arrange the eigenvectors of the principal components in descending order of contribution to form the feature matrix U; use To represent the cumulative contribution rate, the contribution rates of the first d principal components are:

[0034] Set the threshold to P, such that ≥P, determining the principal components, yields the principal component model:

[0035] The original feature parameter set X is reconstructed to form the minimal feature parameter set:

[0036] This allows most of the feature information of the PCA subspace formed by the first d feature vectors to be represented, achieving the goal of attribute reduction.

[0037] (3) Flight parameter data preprocessing Outlier removal: Due to the instability of telemetry link transmission, there are random jump points in the data that deviate significantly from normal operating conditions; secondly, when the airborne data acquisition system is not powered on, the engine status data is generally filled with 43690, which is a large difference compared with the normal operating condition.

[0038] Normalization Correction: Because flight parameter data characterizes different engine operating conditions, directly using it would lead to parameters with large data values ​​dominating model training, while parameters with small data values ​​would be ignored. Therefore, to eliminate the influence of the dimensions of each parameter, a linear normalization method is used to normalize all parameters to between 0 and 1.

[0039] Among them, X max X represents the maximum value of the sample elements. min It represents the minimum value of the sample elements.

[0040] (4) Construct a convolutional-conditional variational autoencoder This paper optimizes the variational autoencoder by combining the concepts of variational autoencoder and conditional generation. Sample class labels are used as conditional inputs to control the state prediction and generation process, forming a conditional variational autoencoder based on semi-supervised learning. The main structure of the conditional variational autoencoder still consists of an encoder and a decoder. During model training, sample class labels are used as inputs to the decoder to iteratively optimize the loss function, generating samples that conform to the original data distribution under specific states.

[0041] The loss function is:

[0042] In the formula, the first term on the right-hand side represents the reconstruction loss, which is the difference between the model output and the input; KL represents the Kullback-Leible divergence, which is the difference between the variational distribution in the latent space and the prior distribution.

[0043] To accurately extract hidden local feature patterns from the time-series feature data of aero-engines, a convolutional module is introduced into the conditional variational autoencoder to replace the fully connected layer in the encoder. The goal is to extract local features of the data with higher training and inference efficiency while preserving the spatial features of the data.

[0044] (5) Pre-trained convolutional-conditional variational autoencoder If feature extraction is performed using the entire time series data as input, it is impossible to effectively capture the key features of the engine at each stage and under each operating condition. Therefore, a window-based method is adopted to divide the time series data into multiple subsequences for pre-training of the model.

[0045] Define a w t Given a time t and a length of Time window: , where t corresponds to the step size of the observation x. Using Each data prediction output can generate Each window is used to train a convolutional-conditional variational autoencoder model, and each window has n-dimensional variables. Thus, X... train This can be represented by the input window sequence W:

[0046] Therefore, a preprocessed engine state dataset is obtained, denoted as... .

[0047] To learn the long-term dependencies among multiple related variables, a separate convolutional channel is assigned to the time series data of each variable. Each channel extracts the temporal features (such as local state and fluctuation trends) of its corresponding variable through independent convolution operations. After feature extraction is completed for each channel, a fusion operation is used to integrate the key features of all channels, ultimately forming a set that comprehensively reflects the core features of all variables. This is achieved through a multidimensional window. The input is fed into the encoder, where it undergoes local feature extraction through a set of filters in the convolutional layer, and the final encoded output is obtained through an activation function. The encoder transforms this into a set of latent feature data represented by c. The convolution process is as follows:

[0048] In the formula, X represents the input window. ; Let be the weight matrix and bias of the k-th convolutional kernel; * represents the activation function; * represents a 2D convolution operation. This is the latent representation of the feature data k. For N-dimensional data, the feature outputs of each dimension are stacked together for output, expressed as:

[0049] The output of the final convolutional layer splits into two branches: the mean and the variance, generating latent vectors corresponding to a Gaussian distribution. .

[0050] The decoder's structure corresponds to the encoder's. It uses deconvolution to reconstruct the window size of the latent feature data C, making it as similar as possible to the original data. The calculation formula is:

[0051] In the formula c i It is the potential representation of each dimension of data; This is a weight flipping operation; b is a single bias for each channel. The model is trained using the training data, optimizing the model parameters while minimizing the loss function until the network converges.

[0052] (6) Construct a long short-term memory network model based on multiple self-attention mechanisms Long Short-Term Memory (LSTM) neural networks control the flow and memory of information through a "gating mechanism," thereby avoiding the "vanishing gradient" or "exploding gradient" problems when processing long sequences. They effectively capture long-term dependencies in time-series data and extract hidden temporal dynamic features. The core structure includes an input gate, a forget gate, an output gate, and a memory cell. The input gate, composed of a sigmoid layer and a tanh layer, is responsible for storing new information into the cell state. The forget gate uses a sigmoid function to output weights between 0 and 1, determining which historical information in the cell is invalid and needs to be discarded. The output gate uses the sigmoid layer to determine which information in the cell state is output, and then processes it through the tanh layer to generate the output for the current time step. The memory cell is the core of long-term memory, recording the state information of the time series at the current moment, i.e., the mathematical expression of the current moment.

[0053] The update formulas for the three gates and memory cells are shown below:

[0054]

[0055]

[0056]

[0057]

[0058]

[0059] In the formula, sigmoid and tanh are activation functions; i t f represents the input gate. t Represents the Gate of Oblivion, o tRepresents the output gate. Candidate values ​​representing memory cells, c t "Represents memory cells, W represents weight parameters, and b represents bias parameters." "Represents the dot product operation.

[0060] Using the output of the encoder in the convolutional-conditional variational autoencoder as the input to the long short-term memory network, the pre-training results of a single step can be fused with the input data of the next time step to participate in the training process of the next time step. This process is recursively repeated to ultimately achieve deep mining of dynamic information in long time series. The input sample set is constructed using the sliding window method, containing T time steps... The time series sample is set as the first time window. This becomes the second time window, and so on, with each time window inputting to deeply mine the gradually evolving temporal characteristics in the engine data.

[0061] Self-attention mechanisms dynamically allocate weights by calculating the relevance of each element in the input sequence to other elements, thus highlighting important information. A linear transformation of the original data yields three vectors for each element: a query vector Q, a key vector K, and a value vector V. Attention scores are calculated based on the similarity between Q and each K element, and then normalized using a softmax function to obtain the weight of each element relative to other elements. Higher weights indicate a stronger connection. The V vectors of each element are then summed according to their attention weights to obtain the element's "context-aware feature," a new feature that integrates relevant information from all elements in the sequence.

[0062] The output of the attention matrix is ​​represented as:

[0063] In the formula This represents the number of hidden units in the neural network. In a self-attention mechanism, Q, K, and V all come from the output of the previous layer.

[0064] Multiple self-attention mechanisms are an extension of the self-attention mechanism. By executing multiple self-attention heads in parallel, they capture correlation information from different dimensions in the data, further enhancing the model's ability to model complex dependencies. The scaling dot product attention for each representation is calculated after generating Q, K, and V vectors through a linear transformation of each head, as shown in the formula:

[0065] Finally, each head is concatenated and converted to the same dimension as the input sequence:

[0066] In the formula Representative parameter matrix.

[0067] The multiple self-attention module is applied to the hidden state H output by the long short-term memory network module, and the parameters of Q, K, and V after linear transformation are calculated for each head. They are all different; they are obtained through model training, using W. i The attention layer takes the entire hidden state H as input and modifies it with the parameter W. i Multiply. The formula for calculating the self-attention value of each head is:

[0068] After p operations, the result u of each head is... i Connecting them together yields a feature representation v:

[0069] Finally, a linear layer is used to transform the concatenated value v into the output y:

[0070] In the formula, w and b are the weight matrix and bias of the final linear layer, respectively. The long short-term memory network model based on multiple self-attention learns e... i To predict the next embedding vector :

[0071] By minimizing the amount of raw data e p and forecast data The model is trained using the error between the two.

[0072] For example, in one embodiment, the implementation process is as follows: (1) According to the general specifications for flight testing of aero engines, 18 parameters were initially selected as the characteristic parameter set for judging the operating conditions of aero engines, including throttle position (PLA), low-pressure speed (N1), high-pressure speed (N2), turbine afterburner gas temperature (T6), turbine afterburner outlet pressure (P6), lubricating oil pressure (MPa), fuel flow rate (W), engine exhaust temperature (T9), and nozzle area (A9). (2) Collect flight parameter data of five aero-engine flights at altitudes of 0-10km. According to the flight mission plan, the flight will perform afterburner flight and in-flight start-up. The aero-engine test conditions of these five flights are labeled as 0-5: engine stop, idle, throttle, intermediate, partial afterburner and full afterburner.

[0073] (3) Data preprocessing, including outlier removal and normalization correction, was performed on the feature parameter set data, resulting in 187,566 sample data points. These included 19,326 parking condition data points, 71,268 slow vehicle condition data points, 81,401 throttling condition data points, 5,416 intermediate condition data points, 6,124 partial acceleration condition data points, and 4,031 full acceleration condition data points. 70% of the samples under each condition were selected as the training set, and the remaining 30% were selected as the test set.

[0074] (4) Using the training set as input, the 16 feature parameters are fused and reduced to obtain 16 eigenvalue matrices. and the corresponding feature vectors Define the cumulative contribution rate of principal components. 95%, to obtain the corresponding The value is 5. The cumulative contribution rates of the first 6 principal components are 60.2%, 71.5%, 78.9%, 86.4%, 94.6%, and 97.5%, respectively. Therefore, the first 6 principal component parameters are finally determined as the minimum feature parameter set.

[0075] (5) The recognition performance of the improved model (CVAE-LSTMA) was compared with that of the traditional autoencoder-long short-term memory network (VAE-LSTM). The test results are shown in Table 1. The results show that the improved model achieved a recognition rate of over 90% for all six engine operating conditions during the identification process. In particular, the improved model showed a significant improvement in classification performance in the intermediate state, partial afterburner state, and full afterburner state. However, due to the wide distribution range and strong oscillation and fluctuation of flight parameter data in the intermediate state, partial afterburner state, and full afterburner state, the classification accuracy of the improved model in these three states was lower than that in the other three states.

[0076] Table 1. Accuracy of aero-engine condition recognition based on CVAE-LSTMA and VAE-LSTM

[0077] The above detailed embodiments are a description of the present invention. It should not be considered that the specific embodiments of the present invention are limited to these descriptions. For those skilled in the art, several simple deductions and substitutions can be made without departing from the concept of the present invention, and all of these should be considered to fall within the protection scope of the present invention.

Claims

1. A method for online identification of test status of an aero-engine, characterized in that, Includes the following steps: The first step is to obtain the minimum set of characteristic parameters based on the technical indicators clearly specified in the general specifications for flight testing of aero engines and the parameters that are most relevant when the corresponding state is reached; The second step is to preprocess the flight parameter data corresponding to the minimum feature parameter set and establish an online identification model for the test status of aero-engines. The third step is to collect sample datasets of aero-engines during actual flight, input the data into the online identification model of aero-engine test status, and determine the test status of the aero-engine based on the output of the online identification model of aero-engine test status.

2. The method according to claim 1, characterized in that, The first step specifically includes: Based on the technical indicators clearly specified in the general specifications for aero-engine flight testing and the parameters most relevant when the corresponding state is reached, a preliminary set of original feature parameters required for online identification of aero-engine test states is formed. Principal component analysis is used to reduce the dimensionality of the feature parameter set and reduce its attributes to form the minimum feature parameter set.

3. The method according to claim 2, characterized in that, Principal component analysis is used to reduce the dimensionality and attributes of the feature parameter set. The specific process includes: Prepare the original feature parameter set; Calculate the mean of the original feature parameter set; Calculate the covariance matrix C of the original feature parameter set; Calculate the eigenvalues ​​and corresponding eigenvectors of the covariance matrix C; Calculate the contribution of each principal component; Arrange the eigenvectors of the principal components in descending order of contribution to form the feature matrix U; The principal component is determined based on the defined cumulative contribution threshold; The sample feature matrix X is transformed using the feature matrix U to obtain the minimum set of feature parameters.

4. The method according to claim 3, characterized in that, The formula used to calculate the contribution of each principal component is as follows: P k The contribution of each principal element.

5. The method according to claim 3, characterized in that, The formula used to determine the principal component sample set is: , This is the mean of the sample dataset.

6. The method according to claim 1, characterized in that, The second step specifically includes: The flight parameter data corresponding to the minimum feature parameter set are preprocessed by outlier removal and normalization correction methods, and known engine test states are labeled by sample labels to form the sample dataset required for model training. A conditional variational autoencoder containing a decoder, encoder, and conditional input module is constructed, and a convolutional module is introduced into the conditional variational autoencoder to obtain a convolutional-conditional variational autoencoder. The local feature patterns of the aero-engine sample dataset are extracted by optimizing the loss function. Using the sample dataset as the input to the encoder in the convolutional-conditional variational autoencoder, the output of the encoder is the local feature pattern of the aero-engine sample dataset. A long short-term memory network model is constructed, and a multi-self-attention mechanism is introduced to assign weights to the hidden states output by the long short-term memory network model, forming a multi-self-attention mechanism-long short-term memory network model. The output of the convolutional-conditional variational autoencoder is used as the input of the multi-self-attention mechanism-long short-term memory network model. The output of the multi-self-attention mechanism-long short-term memory network model is the long-term dependency relationship and different spatial and temporal information in the aero-engine sample dataset. The output of the multi-self-attention mechanism-long short-term memory network model is used as the input of the decoder in the convolutional-conditional variational autoencoder. The output of the decoder is the final prediction result of the aero-engine test state. Thus, the online recognition model of aero-engine test state is trained.

7. The method according to claim 6, characterized in that, The formula used for preprocessing the flight parameter data corresponding to the minimum feature parameter set using outlier removal and normalization correction is as follows: X max X represents the maximum value of the sample elements. min It represents the minimum value of the sample elements.

8. The method according to claim 6, characterized in that, The formula used to guide the extraction of local feature patterns from aero-engine sample datasets by optimizing the loss function is as follows: In the formula, Represents the reconstruction loss, which is the difference between the model output and the input; KL represents the Kullback-Leible divergence, which is the difference between the variational distribution in the latent space and the prior distribution.

9. The method according to claim 6, characterized in that, The formula used to assign weights to the hidden states output by the Long Short-Term Memory network model, incorporating a multiple self-attention mechanism, is as follows: w and b are the weight matrix and bias of the final linear layer, respectively.

10. An online identification device for the test status of an aero-engine, characterized in that, The apparatus for performing the method according to any one of claims 1 to 9, the apparatus comprising: The processing module is used to obtain a minimum set of characteristic parameters based on the technical specifications clearly defined in the engine's general specifications and the main parameters directly related to achieving the corresponding state. A module is established to preprocess the flight parameter data corresponding to the minimum feature parameter set and to build an online identification model for aero-engine test conditions. The determination module is used to collect sample datasets of aero-engines during actual flight, input the data into the online identification model of aero-engine test status, and determine the test status of aero-engines based on the output of the online identification model of aero-engine test status.

Images