Deep learning model automatic recommendation method for aero-engine damage detection

Abstract

The invention discloses a deep learning model automatic recommendation method for aircraft engine damage detection. The method comprises the following steps: designing a metric element system for an aircraft engine damage detection model; integrating an automatic metric element extraction technology and combining with manual annotation to obtain metric element data of the public model; abnormal values and missing values are processed, text features are coded, and normalization and standardization preprocessing is carried out; a random forest regression model is constructed to carry out static performance evaluation on the aero-engine damage detection model, and feature importance analysis is carried out; the Bayesian optimization algorithm is improved, alpha beta-BO is provided, and model configuration capable of achieving the optimal performance is searched. Static recommendation of aero-engine damage detection model configuration can be realized, effective prediction of the Top-1 accuracy of the model is realized, the model debugging work is greatly simplified, the energy consumption required in the model training process is greatly reduced, and the requirement of a developer for quickly constructing a deep learning model under different hardware limitations is met.

Description

Technical Field

[0001] The present invention belongs to the field of AutoML automated machine learning, and specifically relates to a method for automatically recommending deep learning models for aircraft engine damage detection. Background Art

[0002] With the rapid development of deep learning technology, various complex models have emerged one after another, and people's requirements for model accuracy are constantly increasing. This trend is particularly evident in the field of aircraft engine damage detection. Accurate damage detection is one of the key links to ensure flight safety. It can help engineers detect potential problems in a timely manner and take necessary maintenance measures. However, after researchers develop a new network structure, the performance of the model often needs to be further improved by fine-tuning multiple hyperparameters. These hyperparameters include batch size, data augmentation, epoch, etc. Each hyperparameter setting may have a significant impact on the final performance of the model. Therefore, finding the best hyperparameter combination is crucial to improving model performance.

[0003] However, this process usually requires a lot of experiments and tuning, which consumes a lot of researchers' time and computing resources. In order to simplify this process, automated machine learning (AutoML) methods have emerged. AutoML methods aim to automatically search and optimize the hyperparameters and architectures of deep learning models, thereby reducing the workload of researchers. However, most existing AutoML methods focus on searching for neural network architectures rather than optimizing hyperparameters. These methods perform well when processing small data sets, but when faced with complex tasks such as aircraft engine damage detection, they often require unacceptable energy consumption. Model training and optimization on large-scale data sets are not only time-consuming, but also have extremely high requirements for computing resources, which limits the existing AutoML methods in practical applications.

[0004] In addition, studies have shown that the performance of the same neural network may vary significantly when running on different hardware devices. This difference may be caused by factors such as the hardware architecture, computing power, and memory bandwidth. However, current research has not yet systematically revealed the relationship between hardware devices and neural network structures. This problem poses a huge challenge to researchers when selecting and optimizing hardware, making it difficult to determine which hardware device can most effectively run a specific neural network model. This not only affects the performance of the model, but may also lead to a waste of resources and increased energy consumption. Therefore, researchers need to develop more efficient and intelligent automated tuning methods that can not only work effectively on large-scale data sets, but also consider the mutual influence of hardware and network structure. This will provide strong support for the development and optimization of deep learning models in aircraft engine damage detection and promote the further development of artificial intelligence technology in the aviation field. Summary of the invention

[0005] In view of the problems existing in the prior art, the present invention proposes an automatic recommendation method for deep learning models for aircraft engine damage detection to meet people's demand for model performance; the method proposed in the present invention comprehensively utilizes the prior knowledge of a large number of researchers, quantifies the various indicators of the deep learning model in a metric way, and constructs a systematic metric meta-system. The system can capture the relationship between model performance and various components, and automatically find the optimal component combination and parameter settings through optimization algorithms to improve model performance. By implementing an automatic recommendation method for deep learning models for aircraft engine damage detection, the model performance can be improved by adjusting the model parameters on the basis of the original model to meet the needs of developers and help them quickly build deep learning models under different hardware constraints.

[0006] In order to achieve the above technical objectives, the present invention proposes the following technical solutions:

[0007] A deep learning model automatic recommendation method for aircraft engine damage detection, comprising:

[0008] S1. Design a metric system for aircraft engine damage detection model based on the characteristics of deep learning model structure;

[0009] S2. Integrate the automatic extraction technology of metric metadata and extract the metric metadata of the public model in combination with the manual annotation process to obtain a metric metadata information set; the public model is: a public deep learning model for aircraft engine damage detection;

[0010] S3, data preprocessing: first, data cleaning is performed to process the outliers and missing values ​​in the metric metadata extracted in step S2; then, the text features are encoded, and finally, the data is normalized and standardized;

[0011] S4. Construct a random forest regression model to evaluate the static performance of the aircraft engine damage detection model and perform metric importance analysis.

[0012] S5. Based on the characteristics of the metric metadata dataset, the Bayesian optimization algorithm is improved, denoted as αβ-BO; the model configuration that can achieve the best performance is searched through αβ-BO.

[0013] Furthermore, step S1 specifically includes the following steps:

[0014] S11. Collect and analyze existing publicly available deep learning models for aircraft engine damage detection, and sort out metrics that may affect model performance;

[0015] S12, classify the sorted metric meta-models into five categories: model structure, training hyperparameters, framework, data set, and hardware;

[0016] S13. Apply the initially designed metric system to actual cases, verify the validity of each metric through correlation analysis, and eliminate the metrics that are not related to the Top-1 accuracy based on the verification results to form the final metric system.

[0017] More specifically, the final metric system in step S13 is composed of the five categories described in step S12, a total of 20 metric elements, specifically including:

[0018] Model structure: normalization layer, initialization method, convolution layer, skip connection, activation function, pooling layer, feedforward neural network, attention mechanism, output function;

[0019] Training hyperparameters: learning rate strategy, optimizer, number of parameters, regularization method, epoch, batch size, learning rate, data augmentation method;

[0020] Frame: Frame;

[0021] Dataset: dataset;

[0022] Hardware: Type of hardware.

[0023] Furthermore, step S2 specifically includes the following steps:

[0024] S21, searching gitee and github model storage websites to collect model repository links related to the target task; the target task is an aircraft engine damage detection task;

[0025] S22, extracting the code address and description document from each model repository link, and removing the data that does not contain the code address or description document;

[0026] S23, traverse the README files and code files in the code repository, perform formatted data matching based on the KMP algorithm, and extract metric metadata, including: framework, dataset, epoch, batch size, learning rate, and data enhancement method;

[0027] S24, traverse the experimental setting part of the description document, extract the metric information, including: normalization layer, initialization method, convolution layer, skip connection, activation function, pooling layer, feedforward neural network, attention mechanism, output function, learning rate strategy, optimizer, parameter quantity, regularization method, hardware type, and data set, framework, learning rate information; compare and verify the extracted data set, framework, and learning rate metric information with the extracted in the previous step to check whether they are the same or synonymous. If they are different, manually extract the metric information;

[0028] S25. Manually fill in the missing measurement metadata.

[0029] Furthermore, step S3 specifically includes the following steps:

[0030] S31, fill the missing values ​​in the epoch, batch size, learning rate, and hardware type metrics in the metric information set obtained in step S2 using the same type mean interpolation method, and remove the remaining data with missing values, duplicate data, and abnormal data;

[0031] S32, convert the units in the parameter measurement unit into megabytes, and convert the learning rate into decimals;

[0032] S33, digitizing the text data using one-hot coding, binary coding, or polynomial coding;

[0033] S34. Use maximum and minimum normalization to normalize each metric metadata.

[0034] Furthermore, step S4 specifically includes the following steps:

[0035] S41, calculate the top-1 accuracy between each aircraft engine component and its corresponding model, then calculate the Spearman correlation coefficient of the top-1 accuracy, and eliminate redundant features based on the hierarchical clustering method;

[0036] S42, constructing a random forest regression model using the metric metadata set after data preprocessing in step S3, training the model and evaluating its performance;

[0037] S43. Use the Bayesian optimization algorithm to tune the hyperparameters of the random forest regression model, run multiple optimization processes, record the results and performance indicators after each optimization, and find the optimal hyperparameter combination;

[0038] S44. Calculate the Permutation Importance based on the optimal random forest regression model to obtain the importance ranking of each metric.

[0039] Furthermore, step S5 specifically includes the following steps:

[0040] S51, using the random forest regression model obtained in step S4 as the objective function and the Gaussian process as the proxy model to construct an improved Bayesian optimization model; the Gaussian process is defined as:

[0041] f(x)~gp(m(x),k(x,x′));

[0042] Among them, m(x) is the mean function; k(x,x') is the kernel function, which is set to the covariance function.

[0043] S52, using an improved search strategy to find the next exploration point;

[0044] S53, test different parameter combinations on the data set, retrain and verify the model, and obtain the optimal parameter settings;

[0045] S54, according to the result of the metric element importance analysis in step S4, select the five most important metric elements as recommended metric elements;

[0046] S55. Use αβ-BO to find the optimal configuration of the five most important metrics. The termination condition is set to the model has converged or the number of searches has reached a threshold. After each search, check whether the termination condition is met.

[0047] S56. The optimal configuration of the metric elements when the termination condition is reached is reversely decoded according to the encoding process to obtain the best combination of the metric elements.

[0048] More specifically, the search strategy in step S52 is:

[0049] A random search is performed with a probability of Ω to prevent local convergence. The formula for Ω is defined as follows:

[0050] Ω=1-K×P;

[0051] The variable K represents the number of repeated random explorations. If an exploration is not random, K is reduced to 1. P represents the probability of random exploration input by the user.

[0052] The exploration point is found using an acquisition function with a probability of 1-Ω, and the acquisition function γEI is designed as:

[0053]

[0054] Among them, f bestis the currently observed optimal target value, σ(x) is the standard deviation of the Gaussian process at position x, P(f(x)|x,D) is the probability distribution of the target value, and α and β are the set proportional coefficients.

[0055] More specifically, the acquisition function γEI in step S52 is further deduced to make its calculation process more suitable for high-dimensional data, specifically including:

[0056] Define a vector Substituting vector Z for f(x), we have

[0057]

[0058] in, is the cumulative distribution function of the multivariate normal distribution; using the cumulative distribution function CDF of the standard normal distribution, this formula can be further derived into the following form:

[0059]

[0060] The acquisition function γEI derived above is used to search for the next exploration point.

[0061] More specifically, step S55 specifically includes:

[0062] S551, initialize the model, calculate the random search probability according to Ω, and if it is determined to be a random search, randomly select a point in the unexplored area as the next exploration point;

[0063] S552, if random search is not performed, the value of the acquisition function γEI is calculated, and the point with the maximum value of the acquisition function γEI is selected as the next exploration point;

[0064] S553. Check whether the acquisition function frequently obtains the maximum value within a certain range. If so, it is considered that the termination condition is met to obtain the optimal configuration and the search ends; otherwise, a new exploration point is added to update the model, and the above steps are repeated until the model converges or the search number threshold is reached.

[0065] Based on the above technical solution, the present invention has the following beneficial effects:

[0066] The present invention uses automated search and optimization algorithms to significantly improve the performance of the aircraft engine damage detection model and save a lot of time and resources required by traditional parameter adjustment methods.

[0067] The deep learning model is quantified through the metric metasystem, which improves the interpretability of the deep learning model; the data is scientifically preprocessed and a random forest regression model is constructed to obtain model performance at a low cost.

[0068] The present invention uses an improved Bayesian optimization algorithm to find the optimal solution in a high-dimensional parameter space, significantly improving the effect and efficiency of the search task. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is a flow chart of the overall steps of the present invention;

[0070] Figure 2 This is a schematic diagram of the process of extracting metric elements in the present invention;

[0071] Figure 3 Schematic diagram of data preprocessing and construction of random forest regression model in the present invention;

[0072] Figure 4 A schematic diagram of parameters recommended by the improved Bayesian optimization algorithm used in the present invention. DETAILED DESCRIPTION

[0073] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0074] Although the steps in the present invention are arranged with numbers, they are not used to limit the order of the steps. Unless the order of the steps is clearly stated or the execution of a certain step requires other steps as a basis, the relative order of the steps can be adjusted. It is understood that the term "and / or" used in this article involves and covers any and all possible combinations of one or more of the associated listed items.

[0075] like Figure 1 As shown, a deep learning model automatic recommendation method for aircraft engine damage detection proposed by the present invention is given, which specifically includes:

[0076] S1. Design a metric system for aircraft engine damage detection model based on the characteristics of deep learning model structure;

[0077] In this embodiment, step S1 obtains all the metrics by analyzing the model structure of the deep learning model; then the metrics are divided into five categories, and finally each metric is tested to see whether it is valid, and the final metric system is determined; as a preferred implementation, step S1 specifically includes:

[0078] S11. Collect and analyze existing publicly available deep learning models for aircraft engine damage detection, and sort out metrics that may affect model performance;

[0079] S12, classify the sorted metric meta-models into five categories: model structure, training hyperparameters, framework, data set, and hardware;

[0080] S13. Apply the preliminarily designed metric system to actual cases. In this embodiment, six popular deep learning models, namely ResNet, Mobilenet, Bi-LSTM, CycleGAN, Maxvit, and EfficientNet, are selected as actual cases. The validity of each metric is verified through correlation analysis, and the metrics that are not related to the Top-1 accuracy are eliminated according to the verification results to form the final metric system.

[0081] In this embodiment, the Pearson and Spearman correlation coefficients of each metric element and the Top-1 accuracy are calculated, and the metrics elements with both correlation coefficients less than 0.3 are eliminated to form a final metric element system, which consists of the five categories described in step S12, a total of 20 metrics elements, specifically including:

[0082] Model structure: normalization layer, initialization method, convolution layer, skip connection, activation function, pooling layer, feedforward neural network, attention mechanism, output function;

[0083] Training hyperparameters: learning rate strategy, optimizer, number of parameters, regularization method, epoch, batch size, learning rate, data augmentation method;

[0084] Frame: Frame;

[0085] Dataset: dataset;

[0086] Hardware: Type of hardware.

[0087] S2. Integrate automated metric extraction technology and combine it with manual annotation process to extract metric metadata of public models;

[0088] In this embodiment, step S2 extracts information through web crawler technology and automatic extraction of metric elements, and multiple checks are performed to ensure the accuracy of the data. As a preferred implementation, Figure 2 As shown, step S2 specifically includes:

[0089] S21, searching gitee and github model storage websites to collect model repository links related to the target task; the target task is an aircraft engine damage detection task;

[0090] S22, extracting the code address and description document from each model repository link, and removing the data that does not contain the code address or description document;

[0091] S23, traverse the README files and code files in the code repository, perform formatted data matching based on the KMP algorithm, and extract metric metadata, including: framework, dataset, epoch, batch size, learning rate, and data enhancement method;

[0092] S24, traverse the experimental setting part of the description document, extract the metric information, including: normalization layer, initialization method, convolution layer, skip connection, activation function, pooling layer, feedforward neural network, attention mechanism, output function, learning rate strategy, optimizer, parameter quantity, regularization method, hardware type, and data set, framework, learning rate; compare and verify the extracted data set, framework, and learning rate metric information with the extracted in the previous step to check whether they are the same or synonymous. If they are different, manually extract the metric information;

[0093] In this embodiment, the README file (a file that describes how the code runs) only contains the number of epochs used to train the model and the commands for training the model (including batch size and learning rate); the code file includes the import package, such as importtorch, where torch is its framework, and the data enhancement method also has a special import function in the code, which is easy to extract. However, other model details cannot be extracted from the README file and code file. The experimental settings section of the description document will describe the model structure, training hyperparameters, and hardware types. In contrast, the three metrics of data set, framework, and learning rate are all extractable information, so they can be used for verification. For example: there are import torch and importtensorflow in the code, but one of them may have been only introduced but not used in the code. At this time, look for a description based on a certain framework in the description document to check whether they are the same or synonymous.

[0094] S25. Manually fill in the missing measurement metadata.

[0095] S3, data preprocessing: first process the outliers and missing values ​​in the metric metadata extracted in step S2, then encode the text features, and finally normalize and standardize the data;

[0096] In this embodiment, step S3 digitizes the text data so that the data can be input into the model calculation, and the calculation removes redundant features based on the hierarchical clustering results, and then constructs a random forest regression model; as a preferred implementation, Figure 3 As shown, step S3 specifically includes:

[0097] S31, fill the missing values ​​in the epoch, batch size, learning rate, and hardware type metrics in the metric information set obtained in step S2 using the same type mean interpolation method, and remove the remaining data with missing values, duplicate data, and abnormal data;

[0098] S32, convert the units in the parameter measurement unit into megabytes, and convert the learning rate into decimals;

[0099] S33, digitizing the text data using one-hot coding, binary coding, or polynomial coding;

[0100] S34. Use maximum and minimum normalization to normalize each metric metadata.

[0101] S4. Construct a random forest regression model to evaluate the static performance of the aircraft engine damage detection model and perform metric importance analysis.

[0102] In this embodiment, step S4 constructs a random forest regression model based on the preprocessed data to perform static performance evaluation on the aircraft engine damage detection model as a preferred implementation method, such as Figure 3 As shown, step S4 specifically includes:

[0103] S41, calculate the top-1 accuracy between each aircraft engine component and its corresponding model, then calculate the Spearman correlation coefficient of the top-1 accuracy, and eliminate redundant features based on the hierarchical clustering method;

[0104] S42, constructing a random forest regression model using the metric metadata set after data preprocessing in step S3, training the model and evaluating its performance;

[0105] S43. Use the Bayesian optimization algorithm to tune the hyperparameters of the random forest regression model, run multiple optimization processes, record the results and performance indicators after each optimization, and find the optimal hyperparameter combination;

[0106] S44. Calculate the Permutation Importance based on the optimal random forest regression model to obtain the importance ranking of each metric.

[0107] S5. Based on the characteristics of the metric metadata dataset, the Bayesian optimization algorithm is improved, denoted as αβ-BO, to search for the model configuration that can achieve the best performance;

[0108] In this embodiment, step S5 improves the classic Bayesian optimization algorithm, based on the exploration and utilization capabilities of the data set balancing algorithm, and uses the Bayesian optimization algorithm to perform hyperparameter tuning to efficiently obtain the optimal configuration; as a preferred implementation method, Figure 4 As shown, step S5 specifically includes:

[0109] S51, using the random forest regression model obtained in step S4 as the objective function and the Gaussian process as the proxy model to construct an improved Bayesian optimization model; the Gaussian process is defined as:

[0110] f(x)~gp(m(x),k(x,x′));

[0111] Among them, m(x) is the mean function; k(x,x') is the kernel function, which is set to the covariance function;

[0112] S52, using an improved search strategy to find the next exploration point;

[0113] More specifically, in this embodiment, the search strategy is:

[0114] A random search is performed with a probability of Ω to prevent local convergence. The formula for Ω is defined as follows:

[0115] Ω=1-K×P;

[0116] The variable K represents the number of repeated random explorations. If an exploration is not random, K is reduced to 1 to avoid local convergence. P represents the probability of random exploration input by the user.

[0117] The exploration point is found using an acquisition function with a probability of 1-Ω, and the acquisition function γEI is designed as:

[0118]

[0119] Among them, f best is the currently observed optimal target value, σ(x) is the standard deviation of the Gaussian process at position x, P(f(x)|x,D) is the probability distribution of the target value, and α and β are the set proportional coefficients.

[0120] Furthermore, the acquisition function γEI is further derived to make its calculation process more suitable for high-dimensional data, including:

[0121] Define a vector Substituting vector Z for f(x), we have

[0122]

[0123] in, is the cumulative distribution function of the multivariate normal distribution; using the cumulative distribution function CDF of the standard normal distribution, this formula can be further derived into the following form:

[0124]

[0125] The acquisition function γEI derived above is used to search for the next exploration point.

[0126] S53, test different parameter combinations on the data set, retrain and verify the model, and obtain the optimal parameter settings;

[0127] S54, according to the result of the metric element importance analysis in step S4, select the five most important metric elements as recommended metric elements;

[0128] S55. Use αβ-BO to find the optimal configuration of the five most important metrics. The termination condition is set to the model has converged or the number of searches has reached a threshold. After each search, check whether the termination condition is met.

[0129] As a preferred implementation, step S55 specifically includes:

[0130] S551, initialize the model, calculate the random search probability according to Ω, and if it is determined to be a random search, randomly select a point in the unexplored area as the next exploration point;

[0131] S552, if random search is not performed, the value of the acquisition function γEI is calculated, and the point with the maximum value of the acquisition function γEI, that is, the point with the maximum AC value, is selected as the next exploration point;

[0132] S553. Check whether the acquisition function frequently obtains the maximum value within a certain range. If so, it is considered that the termination condition is met to obtain the optimal configuration and the search ends; otherwise, a new exploration point is added to update the model, and the above steps are repeated until the model converges or the search number threshold is reached.

[0133] S56. The optimal configuration of the metric elements when the termination condition is reached is reversely decoded according to the encoding process to obtain the best combination of the metric elements.

[0134] In summary, the present invention proposes a method for automatically recommending deep learning models for aircraft engine damage detection, and constructs a systematic metric system. The system can capture the relationship between model performance and various components, and automatically find the optimal component combination and parameter settings through optimization algorithms, thereby improving model performance. On the basis of the original model, the model performance is improved by adjusting the model parameters, and the needs of developers can be met, helping them to quickly build deep learning models under different hardware constraints.

[0135] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0136] The above-mentioned embodiments only express several implementation modes of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the attached claims.

Claims

1. A deep learning model automatic recommendation method for aircraft engine damage detection, characterized in that: The specific steps include: S1. Design a metric system for aircraft engine damage detection model based on the characteristics of deep learning model structure; S2. Integrate the automatic extraction technology of metric metadata and combine it with the manual annotation process to extract the metric metadata of the public model to obtain the metric metadata information set; The public model is: a public deep learning model for aircraft engine damage detection; S3, data preprocessing: first, data cleaning is performed to process the outliers and missing values ​​in the metric metadata extracted in step S2; then, the text features are encoded, and finally, the data is normalized and standardized; S4. Construct a random forest regression model to evaluate the static performance of the aircraft engine damage detection model and perform metric importance analysis. S5. Based on the characteristics of the metric metadata dataset, the Bayesian optimization algorithm is improved and denoted as αβ-BO; The αβ-BO is used to search for the model configuration that achieves the best performance.

2. The method for automatic recommendation of deep learning models for aircraft engine damage detection according to claim 1, characterized in that: Step S1 specifically includes the following steps: S11. Collect and analyze existing publicly available deep learning models for aircraft engine damage detection, and sort out metrics that may affect model performance; S12, classify the sorted metric meta-models into five categories: model structure, training hyperparameters, framework, data set, and hardware; S13. Apply the initially designed metric system to actual cases, verify the validity of each metric through correlation analysis, and eliminate the metrics that are not related to the Top-1 accuracy based on the verification results to form the final metric system.

3. The method for automatic recommendation of deep learning models for aircraft engine damage detection according to claim 2, characterized in that: The final metric system in step S13 is composed of the five categories and 20 metric elements in total described in step S12, specifically including: Model structure: normalization layer, initialization method, convolution layer, skip connection, activation function, pooling layer, feedforward neural network, attention mechanism, output function; Training hyperparameters: learning rate strategy, optimizer, number of parameters, regularization method, epoch, batch size, learning rate, data augmentation method; Frame: Frame; Dataset: dataset; Hardware: Type of hardware.

4. The method for automatic recommendation of deep learning models for aircraft engine damage detection according to claim 3 is characterized in that: Step S2 specifically includes the following steps: S21, searching gitee and github model storage websites to collect model repository links related to the target task; the target task is an aircraft engine damage detection task; S22, extracting the code address and description document from each model repository link, and removing the data that does not contain the code address or description document; S23, traverse the README files and code files in the code repository, perform formatted data matching based on the KMP algorithm, and extract metric metadata, including: framework, dataset, epoch, batch size, learning rate, and data enhancement method; S24, traverse the experimental setting part of the description document, extract the metric information, including: normalization layer, initialization method, convolution layer, skip connection, activation function, pooling layer, feedforward neural network, attention mechanism, output function, learning rate strategy, optimizer, parameter quantity, regularization method, hardware type, and data set, framework, learning rate information; compare and verify the extracted data set, framework, and learning rate metric information with the extracted in the previous step to check whether they are the same or synonymous. If they are different, manually extract the metric information; S25. Manually fill in the missing measurement metadata.

5. The method for automatic recommendation of deep learning models for aircraft engine damage detection according to claim 3, characterized in that: Step S3 specifically includes the following steps: S31, fill the missing values ​​in the epoch, batch size, learning rate, and hardware type metrics in the metric information set obtained in step S2 using the same type mean interpolation method, and remove the remaining data with missing values, duplicate data, and abnormal data; S32, convert the units in the parameter measurement unit into megabytes, and convert the learning rate into decimals; S33, digitizing the text data using one-hot coding, binary coding, or polynomial coding; S34. Use maximum and minimum normalization to normalize each metric metadata.

6. The method for automatic recommendation of deep learning models for aircraft engine damage detection according to claim 1, characterized in that: Step S4 specifically includes the following steps: S41, calculate the top-1 accuracy between each aircraft engine component and its corresponding model, then calculate the Spearman correlation coefficient of the top-1 accuracy, and eliminate redundant features based on the hierarchical clustering method; S42, constructing a random forest regression model using the metric metadata set after data preprocessing in step S3, training the model and evaluating its performance; S43. Use the Bayesian optimization algorithm to tune the hyperparameters of the random forest regression model, run multiple optimization processes, record the results and performance indicators after each optimization, and find the optimal hyperparameter combination; S44. Calculate the Permutation Importance based on the optimal random forest regression model to obtain the importance ranking of each metric.

7. The method for automatic recommendation of deep learning models for aircraft engine damage detection according to claim 1, characterized in that: Step S5 specifically includes the following steps: S51, using the random forest regression model obtained in step S4 as the objective function and the Gaussian process as the proxy model to construct an improved Bayesian optimization model; the Gaussian process is defined as: f(x)~gp(m(x),k(x,x')); Among them, m(x) is the mean function; k(x,x') is the kernel function, which is set to the covariance function; S52, using an improved search strategy to find the next exploration point; S53, testing different parameter combinations on the metric metadata set, retraining and validating the model, and obtaining the optimal parameter settings; S54, according to the result of the metric element importance analysis in step S4, select the five most important metric elements as recommended metric elements; S55. Use αβ-BO to find the optimal configuration of the five most important metrics. The termination condition is set to the model has converged or the number of searches has reached a threshold. After each search, check whether the termination condition is met. S56. The optimal configuration of the metric elements when the termination condition is reached is reversely decoded according to the encoding process to obtain the best combination of the metric elements.

8. The method for automatic recommendation of deep learning models for aircraft engine damage detection according to claim 7, characterized in that: In step S52, the search strategy is specifically: A random search is performed with a probability of Ω to prevent local convergence. The formula for Ω is defined as follows: Ω=1-K×P; The variable K represents the number of repeated random explorations. If an exploration is not random, K is reduced to 1. P represents the probability of random exploration input by the user. The exploration point is found using an acquisition function with a probability of 1-Ω, and the acquisition function γEI is designed as: Among them, f best is the currently observed optimal target value, σ(x) is the standard deviation of the Gaussian process at position x, P(f(x)|x,D) is the probability distribution of the target value, and α and β are the set proportional coefficients.

9. The method for automatic recommendation of deep learning models for aircraft engine damage detection according to claim 8, characterized in that: The acquisition function γEI is further deduced to make its calculation process more suitable for high-dimensional data, specifically including: Define a vector Substituting vector Z for f(x), we have in, is the cumulative distribution function of the multivariate normal distribution; using the cumulative distribution function CDF of the standard normal distribution, this formula can be further derived into the following form: The acquisition function γEI derived above is used to search for the next exploration point.

10. The method for automatic recommendation of deep learning models for aircraft engine damage detection according to claim 7, characterized in that: Step S55 specifically includes: S551, initialize the model, calculate the random search probability according to Ω, and if it is determined to be a random search, randomly select a point in the unexplored area as the next exploration point; S552, if random search is not performed, the value of the acquisition function γEI is calculated, and the point with the maximum value of the acquisition function γEI is selected as the next exploration point; S553. Check whether the acquisition function frequently obtains the maximum value within a certain range. If so, it is considered that the termination condition is met to obtain the optimal configuration and the search ends; otherwise, a new exploration point is added to update the model, and the above steps are repeated until the model converges or the search number threshold is reached.

Images