An aero-engine assembly quality anomaly detection and feature inference method based on an autoencoder

Abstract

This invention relates to the field of aero-engine performance optimization technology, specifically to a method for detecting and inferring aero-engine assembly quality anomalies based on autoencoders. The method includes acquiring aero-engine assembly and testing datasets; constructing a stacked autoencoder corresponding to each training dataset; each stacked autoencoder consisting of an encoder and a decoder; constructing a Bagging stacked autoencoder model based on each trained stacked autoencoder; setting a confidence level to determine a dynamic threshold for anomalies; and performing anomaly detection on the aero-engine assembly and testing data based on the Bagging stacked autoencoder model and the dynamic threshold. This invention achieves efficient and low-dimensional representation of aero-engine assembly quality data, effectively reducing data reconstruction errors and improving the robustness and generalization ability of feature learning.

Description

Technical Field

[0001] This invention relates to the field of aero-engine performance optimization technology, specifically to a method for detecting and inferring features of aero-engine assembly quality anomalies based on an autoencoder. Background Technology

[0002] Aero engines are highly complex and precise mechanical devices that provide flight power to aircraft, involving numerous disciplines and integrating many cutting-edge technologies. They have extremely stringent requirements for reliability and environmental adaptability, resulting in high investment costs, high development risks, and long development cycles in their development process. They are hailed as the "jewel in the crown of modern industry." The aero engine industry is a concentrated reflection of a nation's strong industrial foundation, advanced technological level, and powerful comprehensive national strength. It is also an important strategic guarantee for national security and great power status. my country is investing significant resources in promoting high-quality development in the aero engine field.

[0003] In actual production, the assembly and testing processes of aero-engines are two crucial stages in their manufacturing. The assembly quality directly determines the engine's operational performance, which in turn requires testing. Only aero-engines that meet the testing performance specifications can be put into actual use; otherwise, they will be disassembled and reassembled until they meet the testing requirements. The aero-engine assembly process involves numerous parts, requiring not only accurate positioning but also ensuring that the relative positions of all components remain within standard ranges. Furthermore, the reliability of connections and assembly precision must meet strict specifications, making the aero-engine assembly process extremely complex. However, aero-engine assembly relies heavily on manual operation, and workers often cannot accurately complete such high-precision assembly tasks. Even minor assembly feature errors can lead to a degradation in aero-engine testing performance, affecting the overall operational performance. Currently, there is no systematic assembly specification guideline for aero-engine assembly. The assembly quality characteristics are generally determined based on worker experience, resulting in significant variations in assembly quality and large fluctuations in testing performance results. Aero-engines that do not meet testing requirements need to be disassembled and reassembled, severely wasting production resources and increasing production costs. To improve the assembly precision and efficiency of aero-engines, from the perspective of test performance, the system completes the detection of aero-engine assembly quality anomalies, identifies assembly abnormalities, and infers the assembly range of quality characteristics corresponding to specific test performance indicators. This provides more refined and scientific assembly guidance, thereby increasing the first-pass yield, reducing aero-engine production costs, and enhancing the market competitiveness of aero-engine manufacturers.

[0004] While existing technologies have made some progress in the field of quality anomaly detection, unavoidable problems and challenges remain. Traditional quality anomaly detection methods, such as statistical methods and control charts, rely on rules derived from experience or manually selected features. These methods typically perform well when the feature space is small and the data distribution is relatively simple. However, their performance often falls short when faced with complex, high-dimensional data. The aero-engine assembly process involves numerous complex and interdependent features, making it difficult for traditional methods to capture the relationships between aero-engine assembly quality, resulting in low detection accuracy and poor robustness. With the popularization of machine learning, algorithms based on supervised and unsupervised learning have been widely introduced into anomaly detection tasks. For example, support vector machines and decision trees have shown outstanding anomaly detection performance in specific scenarios. However, these methods typically require a large amount of labeled data for training, which is often difficult and costly to obtain in real-world scenarios. Furthermore, supervised learning methods have poor generalization ability to anomaly samples and are easily affected by the distribution of the training set; while unsupervised learning relies on distance metrics or clustering assumptions, making it difficult to adapt to complex, non-linear, high-dimensional data from aero-engine assembly. Based on the aforementioned research shortcomings, autoencoders, as a deep learning algorithm, effectively extract latent features by learning low-dimensional feature representations of input data and mapping them to the output. They can then accurately identify anomalies by reconstructing errors. Autoencoders are adaptable to anomaly detection in complex high-dimensional data and exhibit strong generalization ability, making them widely used in quality anomaly detection. However, they still suffer from susceptibility to noise interference and overfitting, requiring optimization and improvement to overcome these shortcomings and enhance their performance in anomaly detection. Summary of the Invention

[0005] To address the problems existing in the aforementioned background technology, this invention proposes a method for detecting assembly quality anomalies and inferring features in aero-engines based on autoencoders. This invention is based on a stacked autoencoder model and incorporates the core idea of ​​ensemble learning. It matches the encoding space of the stacked autoencoder with test performance, and constructs a Bagging stacked autoencoder model by optimizing the loss function. This achieves efficient and low-dimensional representation of aero-engine assembly quality data, effectively reducing data reconstruction errors and improving the robustness and generalization ability of feature learning.

[0006] This invention provides a method for detecting and inferring features of aero-engine assembly quality anomalies based on autoencoders. First, aero-engine assembly and test data are collected and organized, and preprocessed to ensure data integrity. Second, based on the core idea of ​​ensemble learning, a Bagging stacked autoencoder model is constructed to learn the features of the aero-engine assembly data, minimizing data reconstruction errors. Then, based on the Bagging stacked autoencoder model, a dynamic threshold strategy is proposed for anomaly detection, identifying whether a sample is abnormal by comparing the reconstruction error with the dynamic threshold. Finally, the loss function is optimized by matching the encoding space of the Bagging stacked autoencoder model with test performance values. By using specific test performance indicators as input to the decoder, the corresponding aero-engine assembly quality features can be obtained, enabling the inference of the assembly range of quality features.

[0007] The first objective of this invention is to provide a method for detecting and inferring features of assembly quality anomalies in aero-engines based on an autoencoder, comprising:

[0008] Obtain the assembly and testing dataset of the aero-engine, which includes the assembly quality dataset and the corresponding test performance dataset;

[0009] Multiple training datasets were generated from the test dataset using random sampling with replacement.

[0010] Construct a stacked autoencoder for each training dataset; each stacked autoencoder consists of an encoder and a decoder;

[0011] Train the corresponding stacked autoencoder using each training dataset until convergence, and obtain each trained stacked autoencoder.

[0012] During training, assembly quality data is used as input to the stacked autoencoder; test performance data is used as the feature representation of the code and as input data to the decoder.

[0013] Construct a Bagging stacked autoencoder model based on each trained stacked autoencoder;

[0014] Set confidence levels to determine dynamic thresholds for outlier data;

[0015] The assembly and testing data of the aero-engine to be tested are used to detect anomalies in the assembly quality of the aero-engine based on the Bagging stacked autoencoder model and dynamic thresholds.

[0016] Preferably, the loss function for training each stacked autoencoder includes:

[0017] The error between the input data and the output of the stacked autoencoder is used as the reconstruction error;

[0018] The error between the encoding of the stacked autoencoder and the test performance data is taken as the encoding error;

[0019] The sum of reconstruction error and encoding error is used as the loss function of the stacked autoencoder.

[0020] Preferably, the dynamic threshold determination step for the abnormal data includes:

[0021] The test dataset is used to generate a validation dataset through random sampling with replacement; both the training and validation sets contain normal data.

[0022] The Bagging stacked autoencoder model is reconstructed using the validation set, and the reconstruction error of the validation set is obtained.

[0023] The dynamic threshold for outlier data is determined by setting a confidence level based on the reconstruction error of the validation set.

[0024] Preferably, constructing the Bagging stacked autoencoder model includes:

[0025] The average value of the output data of each trained stacked autoencoder is used as the final output of the model.

[0026] Preferably, it further includes: using the set test performance of the aero-engine as the input of the decoder of each stacked autoencoder in the Bagging stacked autoencoder model to infer the assembly range of the aero-engine assembly quality characteristics.

[0027] Preferably, when inferring the assembly range of aero-engine assembly quality characteristics, the process includes:

[0028] The set test performance of the aero-engine is discretized to obtain multiple discrete values;

[0029] Discrete values ​​are used as inputs to the decoders of each stacked autoencoder in the Bagging stacked autoencoder model to reconstruct the corresponding assembly quality feature values.

[0030] The reconstructed assembly quality feature values ​​are fitted to obtain the assembly quality feature change curve;

[0031] The assembly range of aero-engine assembly quality characteristics is inferred based on the assembly quality characteristic variation curve.

[0032] The second objective of this invention is to provide a computer program product, including a computer program that, when executed by a processor, implements a method for detecting and inferring features of aero-engine assembly quality anomalies based on an autoencoder.

[0033] A third objective of this invention is to provide an electronic device comprising:

[0034] Processor; and

[0035] Memory for storing the executable instructions of the processor;

[0036] The processor is configured to perform a method for detecting and inferring features of aero-engine assembly quality anomalies based on an autoencoder by executing the executable instructions.

[0037] The fourth objective of this invention is to provide a system for detecting and inferring features of assembly quality anomalies in aero-engines based on an autoencoder, comprising:

[0038] The data acquisition module is used to acquire the assembly and testing dataset of the aero-engine. The assembly and testing dataset includes the assembly quality dataset and the corresponding test performance dataset. Multiple training datasets are generated from the assembly and testing dataset using a random sampling method with replacement.

[0039] The encoder model building module is used to construct a stacked autoencoder corresponding to each training dataset. Each stacked autoencoder consists of an encoder and a decoder. The stacked autoencoder corresponding to each training dataset is trained until convergence, resulting in each trained stacked autoencoder. During training, assembly quality data is used as the input to the stacked autoencoder, and test performance data is used as the feature representation of the encoder and as the input data to the decoder. A Bagging stacked autoencoder model is constructed based on each trained stacked autoencoder.

[0040] The anomaly detection module is used to set a confidence level to determine the dynamic threshold of abnormal data; it uses the assembly and testing data of the aero-engine to be tested to perform anomaly detection on the aero-engine assembly quality based on the dynamic threshold according to the Bagging stacked autoencoder model.

[0041] The present invention has at least the following beneficial effects:

[0042] This invention provides a method for detecting assembly quality anomalies and inferring features in aero-engines based on autoencoders. This invention achieves anomaly detection and fault diagnosis of assembly quality from the perspective of test performance, and infers assembly-related information that meets test performance requirements. Based on a stacked autoencoder model and incorporating the core idea of ​​ensemble learning, this invention matches the encoding space of the stacked autoencoder with test performance. By optimizing the loss function, a Bagging stacked autoencoder model is constructed, achieving efficient and low-dimensional representation of aero-engine assembly quality data. This effectively reduces data reconstruction errors and improves the robustness and generalization ability of feature learning. The modeling process of this invention is simple and efficient, and it has good applicability to anomaly detection and feature inference methods in other manufacturing scenarios. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the structure of an autoencoder.

[0044] Figure 2 This is a schematic diagram of a stacked autoencoder.

[0045] Figure 3 This is a framework diagram of the Bagging stacked autoencoder model of the present invention.

[0046] Figure 4 This is a framework diagram for aero-engine assembly quality anomaly detection and feature inference based on a Bagging stacked autoencoder.

[0047] Figure 5 This is a flowchart of the inference process for assembly quality features of aero-engines based on a Bagging stacked autoencoder.

[0048] Figure 6 This is a structural diagram of the stacked autoencoder in this example.

[0049] Figure 7 This is a training error curve for the stacked autoencoder in this example.

[0050] Figure 8 This is a comparison chart of the dynamic thresholds of the stacked autoencoder and the Bagging stacked autoencoder model in this example.

[0051] Figure 9 This is a scatter plot of the reconstruction error of the Bagging stacked autoencoder model in this example on the training, validation, and test sets.

[0052] Figure 10 This is a comparison chart of the F1 scores of the stacked autoencoder model in this example and the Bagging stacked autoencoder model.

[0053] Figure 11This is an assembly range inference diagram for the assembly quality characteristics of the aero-engine in this example. Detailed Implementation

[0054] In order to illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following detailed description is provided in conjunction with the embodiments.

[0055] Based on the current state of anomaly detection research, the method proposed in this invention can better capture the potential features of high-dimensional complex data, enhance the accuracy of data reconstruction, and thus more accurately identify anomalies. The model built based on the ensemble learning method can effectively reduce the impact of noise and improve the model's sensitivity and robustness to anomalies. The proposed method embeds test performance into the coding space, improving the model's generalization ability while realizing the inference of assembly quality features. This method is applicable to assembly quality anomaly detection in various complex environments and has excellent adaptability and application range.

[0056] The purpose of this invention is to optimize the assembly quality of aero-engines. This invention provides a method for detecting and inferring assembly quality anomalies in aero-engines based on autoencoders. This method integrates multiple stacked autoencoders using the Bagging method and matches the encoding space of the stacked autoencoders with test performance to construct a Bagging stacked autoencoder model, which learns the feature representation of aero-engine assembly quality. Then, a dynamic threshold strategy is proposed to achieve the detection of aero-engine assembly quality anomalies. Finally, by utilizing the characteristics of the decoder and inputting specific test performance indicators, the assembly intervals of various quality characteristics of the aero-engine are inferred.

[0057] The main contents of this invention include:

[0058] Acquire assembly quality and test performance data of aero-engines, and perform simple data preprocessing, such as handling outliers and missing values;

[0059] Multiple training sets are generated by random sampling with replacement. The Bagging ensemble method is used to train multiple stacked autoencoder models with different performance based on different training sets. The encoding space of the stacked autoencoder is matched with the test performance. The loss function of the stacked autoencoder is optimized. The average value of the reconstructed output data of all individual learners is taken as the final output result to construct the Bagging stacked autoencoder model.

[0060] Based on the dynamic threshold strategy, the judgment threshold for assembly quality anomalies is obtained. By comparing it with the reconstruction error of the sample, the task of assembly quality anomaly detection is achieved.

[0061] The required test performance values ​​of the aero-engine are used as input to the decoder in the model to infer the assembly range of the aero-engine's quality characteristics.

[0062] To achieve the above objectives, a method for detecting and inferring assembly quality anomalies in aero-engines based on an autoencoder includes:

[0063] S1. Obtain the assembly and testing dataset of the aero-engine, which includes the assembly quality dataset and the corresponding test performance dataset.

[0064] Multiple training datasets were generated from the test dataset using random sampling with replacement.

[0065] S2. Construct a stacked autoencoder for each training dataset; each stacked autoencoder consists of an encoder and a decoder;

[0066] S3. Train the corresponding stacked autoencoder using each training dataset until convergence, and obtain each trained stacked autoencoder.

[0067] The loss function for training each stacked autoencoder includes:

[0068] The error between the input data and the output of the stacked autoencoder is used as the reconstruction error;

[0069] The error between the encoding of the stacked autoencoder and the test performance data is taken as the encoding error;

[0070] The sum of reconstruction error and encoding error is used as the loss function of the stacked autoencoder.

[0071] During training, assembly quality data is used as input to the stacked autoencoder; test performance data is used as the feature representation of the code and as input data to the decoder.

[0072] S4. Construct a Bagging stacked autoencoder model based on each trained stacked autoencoder;

[0073] Set confidence levels to determine dynamic thresholds for outlier data;

[0074] When constructing a Bagging stacked autoencoder model, the following steps are included:

[0075] The average value of the output data of each trained stacked autoencoder is used as the final output of the model.

[0076] The dynamic threshold determination step for the abnormal data includes:

[0077] The test dataset is used to generate a validation dataset through random sampling with replacement; both the training and validation sets contain normal data.

[0078] The Bagging stacked autoencoder model is reconstructed using the validation set, and the reconstruction error of the validation set is obtained.

[0079] The dynamic threshold for outlier data is determined by setting a confidence level based on the reconstruction error of the validation set.

[0080] S5. The assembly and testing data of the aero-engine to be tested are used to detect anomalies in the assembly quality of the aero-engine based on the Bagging stacked autoencoder model and dynamic thresholds.

[0081] It also includes: using the set test performance of the aero-engine as the input of the decoder of each stacked autoencoder in the Bagging stacked autoencoder model to infer the assembly range of the aero-engine assembly quality characteristics.

[0082] When inferring the assembly range for obtaining the assembly quality characteristics of an aero-engine, the following is included:

[0083] The set test performance of the aero-engine is discretized to obtain multiple discrete values;

[0084] Discrete values ​​are used as inputs to the decoders of each stacked autoencoder in the Bagging stacked autoencoder model to reconstruct the corresponding assembly quality feature values.

[0085] The reconstructed assembly quality feature values ​​are fitted to obtain the assembly quality feature change curve;

[0086] The assembly range of aero-engine assembly quality characteristics is inferred based on the assembly quality characteristic variation curve.

[0087] To illustrate the method for detecting and inferring features of aero-engine assembly quality anomalies based on an autoencoder provided by the present invention, it is described in conjunction with the accompanying drawings.

[0088] In the context of aero-engine manufacturing, aero-engine assembly quality data includes the geometric dimensions of components, assembly errors, etc., denoted as... ,in Indicates the first One assembly quality characteristic, This indicates the number of features included in the aero-engine assembly feature set. Aero-engine test data refers to the various state values ​​recorded by sensors during the test run, denoted as... ,in Indicates the first The test performance of an aero engine This represents the number of performance indicators from aero-engine testing. Assembly and testing data collected from the production site are compiled into an aero-engine assembly and testing dataset. Duplicate records are removed, missing and outlier values ​​are handled, and data standardization and normalization are achieved to improve data quality and reduce the difficulty of data analysis. This is specifically for datasets containing... Based on a dataset of aero-engine assembly tests, this invention combines the core idea of ​​ensemble learning and proposes an autoencoder-based method for detecting aero-engine assembly quality anomalies and inferring features, as detailed below.

[0089] Ensemble learning, which combines multiple weak learners into a single strong learner using specific ensemble strategies, effectively improves the model's prediction accuracy and generalization ability, and compensates for the limitation of a single model in fully capturing data features. As one of the most classic ensemble learning methods, Bagging obtains a strong learner from the original dataset through random sampling with replacement. To build a new dataset A learner, and The results are integrated according to a combination strategy to output the final result. For classification problems, the combination strategy is generally a voting method, where the category with the most votes is the final output. For regression problems, the combination strategy is generally an averaging method or a weighted method, calculating the average or weighted average as the model output. The Bagging method performs exceptionally well, exhibiting strong generalization ability and playing a significant role in reducing the variance of model predictions.

[0090] An autoencoder is an unsupervised learning algorithm based on a neural network architecture for representing data features. It obtains a low-dimensional representation (encoding) of high-dimensional data through a neural network, and then uses the encoding to reconstruct the data. An autoencoder consists of an encoder and a decoder, and its structure is as follows: Figure 1 As shown. The encoder receives input data and maps it to a low-dimensional representation, completing the "encoding" of high-dimensional data. The decoder then converts the encoded data back to an approximate representation of the original data, reconstructing the input data. To improve the accuracy of data reconstruction, the network structures of the encoder and decoder are generally mirror-symmetric. The loss function of the autoencoder is the difference between the input data and the reconstructed data. During forward propagation, the encoder utilizes a function... The input data is encoded, and the decoder uses a function. The encoded data is reconstructed into the output. During backpropagation, the network weights are updated using the gradient descent algorithm to minimize the loss function, thus improving the input data. and reconstructing data They are close to each other. As a variant of autoencoders, stacked autoencoders are deep learning models composed of multiple autoencoders stacked in series, with a structure as follows: Figure 2As shown, the hidden layer of the autoencoder serves as the input layer of another autoencoder. Each layer learns features at different levels, thus forming a multi-layer network structure. This enables the stacked autoencoder to have a better feature interpretation capability and to extract more effective information such as the assembly quality features of aero-engines.

[0091] For complex aero-engine assembly and testing data, stacked autoencoders are used as the basic model for anomaly detection and feature inference. However, single stacked autoencoders often lack stable robustness and generalization ability, leading to inaccurate results in anomaly detection and feature inference. Leveraging the advantages of ensemble learning, the Bagging method is used to integrate the outputs of multiple stacked autoencoders, reducing data reconstruction errors. Therefore, this section proposes a Bagging stacked autoencoder model based on ensemble learning to achieve anomaly detection of aero-engine assembly quality features, thereby further enabling range inference of assembly quality features.

[0092] Let the sample representation of the aero-engine assembly test dataset be as follows: ,in This indicates assembly quality data. This represents test performance data. The framework of the Bagging stacked autoencoder model is as follows: Figure 3 As shown, the dataset is generated from the original dataset using a random sampling method with replacement. There are 1 training dataset; using these training datasets, the corresponding... A stacked autoencoder, respectively Using stacked autoencoder models as individual learners in the Bagging method, the output of each stacked autoencoder can be represented as... ,in Pick The task of reconstructing the assembly quality data of aero-engines is obviously a regression problem. The ensemble strategy is to calculate the average value of the outputs of all stacked autoencoders as the final result and construct a Bagging stacked autoencoder model. The final output of the model is shown in Equation (1).

[0093] (1)

[0094] In constructing a stacked autoencoder as an individual learner, in order to infer the range of aero-engine assembly quality feature values ​​from aero-engine test performance, the encoding dimension is kept consistent with the number of aero-engine test performance indicators. The loss function of the stacked autoencoder is optimized to be the sum of reconstruction error and encoding error. While realizing the recovery of input data by the stacked autoencoder, test performance information is embedded into the encoding of the autoencoder to improve feature reasoning ability. The calculation formula is shown in Equation (2).

[0095] (2)

[0096] in, This represents the reconstruction error, which is the error between the input data and the output result; This represents the encoding error, specifically the error value between the encoding of the stacked autoencoder and the test performance.

[0097] The logical framework for detecting assembly quality anomalies and inferring features in aero-engines based on the constructed Bagging stacked autoencoder model is as follows: Figure 4 As shown, during model training, the dataset is divided into training, validation, and test sets using random sampling with replacement. The training and validation sets contain only normal data, while the test set contains both normal and abnormal data. The Bagging stacked autoencoder, trained on the training set, can learn to reconstruct normal data, but struggles to reconstruct abnormal data. According to the autoencoder principle, the reconstruction error of abnormal data related to aero-engine assembly quality is significantly greater than that of normal data. Therefore, setting an accurate error threshold is beneficial for improving the efficiency of anomaly detection. This invention proposes a dynamic threshold strategy for anomaly detection. The reconstruction error of the model on the validation set obtained through random sampling is calculated, and a confidence level is set to determine the dynamic threshold for abnormal data. For example, the reconstruction error of 99% of normal data is less than [a certain threshold]. ,but This serves as the threshold for judging abnormal data at a 99% confidence level. By adjusting the confidence level, different dynamic thresholds are obtained to distinguish between normal and abnormal data. The confidence level depends on the actual situation and is generally set between 98% and 99.5%. This ensures that the Bagging stacked autoencoder model can accurately identify abnormal data, avoid mistaking normal data for abnormal data, and improve the accuracy of detecting abnormalities in aero-engine assembly quality.

[0098] Based on the Bagging stacked autoencoder model, the specified aero-engine test performance indicators are used as the encoded feature representations and as the input data for the decoder in the stacked autoencoder, thereby obtaining the reconstructed outputs of each stacked autoencoder. The average of all stacked autoencoder outputs is obtained using an ensemble method, which is the reconstructed result of the aero-engine assembly quality from the Bagging stacked autoencoder model. The process is as follows: Figure 5 As shown.

[0099] In summary, this invention preprocesses the collected aero-engine assembly and test data, organizing them into a systematic aero-engine assembly and test dataset. Combining the core idea of ​​ensemble learning, it uses stacked autoencoders as individual learners, representing the encoding of the stacked autoencoders as aero-engine test performance, and optimizes the loss function of the autoencoders, thereby constructing a Bagging stacked autoencoder model.

[0100] Then, a dynamic threshold strategy for anomaly detection is proposed. Based on the reconstruction error distribution of the validation set, a corresponding confidence level is set according to practical experience, thus obtaining a dynamic threshold at that confidence level. The reconstruction error of the test sample is compared with this threshold to identify anomalous data samples. It should be noted that the confidence level is set empirically based on actual conditions, and is generally small, ranging from 98% to 99.5%. This invention mainly selects an appropriate confidence value based on the detection effect in actual applications.

[0101] Finally, specific aero-engine test performance is used as input to the decoders of each stacked autoencoder in the model to infer the assembly range of the corresponding aero-engine assembly quality characteristics. This provides more refined guidance for assembly process standards, improves the assembly quality of aero-engines, increases assembly efficiency and first-time test pass rate, and enhances the efficiency and market competitiveness of aero-engine manufacturing enterprises.

[0102] To further illustrate the method for detecting and inferring features of aero-engine assembly quality anomalies based on an autoencoder provided by the present invention, specific embodiments are described in conjunction with the present invention.

[0103] This embodiment uses an aero-engine manufacturing plant as the research subject. Aero-engine assembly data and test data were collected through on-site measurements. The resulting aero-engine assembly and test dataset contained 1000 samples, each with two aero-engine test performance indicators and 18 aero-engine assembly quality characteristics. The data was segmented based on test performance; when the test performance indicator values ​​of a sample all exceeded the 80% threshold of the overall dataset, that sample was marked as outlier data, totaling 150 outlier samples. Using a random sampling method with replacement, the aero-engine assembly and test dataset was divided into a training set, a validation set, and a test set. The training set contained 800 normal data points, the validation set contained 200 normal data points, and the test set contained 50 normal data points and 150 outlier data points. During data preprocessing, the min-max normalization method was used to normalize the dataset samples.

[0104] In this example experiment, the stacked autoencoder consists of three layers, with the number of neurons and activation function of each linear layer as follows: Figure 6 As shown, the Bagging stacked autoencoder is constructed by integrating six independent stacked autoencoders. The hyperparameter settings are: 500 training iterations, a batch size of 32 for the input data, and a learning rate of 0.005.

[0105] The error curves of each stacked autoencoder during training are as follows: Figure 7 As shown in the figure, since the training datasets for each stacked autoencoder are generated using random sampling with replacement, all training datasets are different. Therefore, the stacked autoencoders exhibit different training error variation curves. However, as can be seen from the figure, after nearly 500 training rounds, the training error of the stacked autoencoders tends to stabilize and no longer fluctuates drastically, indicating that the stacked autoencoders have completed training, their performance has converged, and the parameters no longer change significantly. The figure also shows that stacked autoencoder 4 exhibits the strongest data reconstruction capability, with its loss function significantly lower than other encoders. Furthermore, stacked autoencoder 5 performs poorly, with the smallest decrease in error variation curve.

[0106] Following the proposed dynamic threshold strategy for anomaly detection, this example sets the confidence level to 98%. Based on the model's error distribution on the validation set, the dynamic threshold for anomaly detection is 1.266. However, the anomaly detection thresholds for each stacked autoencoder at the 98% confidence level are all greater than the dynamic threshold. The comparison results are as follows: Figure 8 As shown, the lower the dynamic threshold, the smaller the overall reconstruction error of the model. The Bagging stacked autoencoder model, by integrating multiple stacked autoencoders, can exhibit superior feature learning capabilities. Its overall error in reconstructing input data is significantly lower than that of a single stacked autoencoder, overcoming the problems of overfitting and underfitting of a single model. This further verifies the practicality and effectiveness of the proposed method in detecting assembly quality anomalies in aero-engines.

[0107] Bagging stacked autoencoder models calculate the reconstruction error of test set samples to evaluate the model's accuracy in anomaly detection. For example... Figure 9The figure shows the scatter plot of reconstruction errors of the Bagging stacked autoencoder across various datasets. Blue dots represent training set samples, green dots represent validation set samples, and red dots represent test set samples. The red line is the dynamic threshold line for anomaly detection with a 98% confidence level. The reconstruction errors of the normal data and validation set data in the test set are almost all within the same range, mostly below the dynamic threshold line. However, the reconstruction errors of the anomalous data in the test set are significantly higher than the dynamic threshold line. This indicates that the Bagging stacked autoencoder model constructed in this section performs excellently in assembly quality anomaly detection, and the proposed dynamic threshold strategy is very reasonable in anomaly detection. It can accurately identify abnormal assembly quality data by judging reconstruction errors, demonstrating scientific validity and effectiveness.

[0108] The F1 score is an important indicator in statistics used to evaluate the accuracy of classification models. Its calculation formula is shown in equation (3).

[0109] (3)

[0110] in Precision is the ratio of the number of samples correctly predicted as positive to the total number of samples predicted as positive. The formula is shown in equation (4). Recall is the ratio of the number of samples that correctly predict the positive class to the total number of samples that are positive. The formula is shown in equation (5).

[0111] (4)

[0112] (5)

[0113] in, , and These are terms from the confusion matrix, representing the number of samples that were actually positive and predicted as positive, the number of samples that were actually negative and predicted as positive, and the number of samples that were actually positive and predicted as negative. The F1 score, as the harmonic mean of precision and recall, provides a more comprehensive evaluation of the model's overall performance and demonstrates outstanding performance in various scenarios, including those with imbalanced data. It can be used for the analysis and evaluation of the abnormality detection results in the aero-engine assembly quality in this experiment.

[0114] The F1 scores of all stacked autoencoder and bagging stacked autoencoder models are compared as follows: Figure 10As shown, the F1 scores of each stacked autoencoder are 0.801, 0.925, 0.889, 0.948, 0.918, and 0.933, respectively, while the F1 score of the Bagging stacked autoencoder model is 0.962. This indicates that the Bagging stacked autoencoder model based on ensemble learning can effectively improve the performance of a single stacked autoencoder, make full use of the output results of multiple stacked autoencoders, avoid the overfitting problem caused by a single stacked autoencoder, enhance the robustness of the ensemble model, and improve the accuracy of data reconstruction.

[0115] Based on the actual needs of aero-engine production, the performance indicators of aero-engine test runs need to be maintained at a relatively low level. Assuming that the performance indicators of aero-engine test runs are kept within the 10th to 50th percentile of the original test performance distribution, this example needs to infer the assembly range of the corresponding aero-engine assembly quality characteristic values ​​based on this standard. The decoder of the Bagging stacked autoencoder model can only input a single encoded value, for example, input encoding... This represents the 10th percentile of the overall performance index value, thus yielding the assembly quality characteristic value of the aero-engine. To achieve characteristic inference of performance from the 10th to 50th percentiles, the data is discretized, dividing the performance range into multiple discrete values ​​at 1% intervals. The discrete code is used as input to the decoder to obtain the corresponding aero-engine assembly quality feature values. The reconstructed assembly quality features are then fitted to obtain the assembly quality feature variation curve, as shown below. Figure 11 As shown, this enables the inference of quality characteristics.

[0116] The present invention provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method for detecting and inferring the assembly quality anomalies of aero-engines based on an autoencoder.

[0117] This invention provides an electronic device, comprising:

[0118] Processor; and

[0119] Memory for storing the executable instructions of the processor;

[0120] The processor is configured to execute the aforementioned method for detecting and inferring anomalies in aero-engine assembly quality based on an autoencoder by executing the executable instructions.

[0121] This invention provides a system for detecting and inferring features of assembly quality anomalies in aero-engines based on an autoencoder, comprising:

[0122] The data acquisition module is used to acquire the assembly and testing dataset of the aero-engine. The assembly and testing dataset includes the assembly quality dataset and the corresponding test performance dataset. Multiple training datasets are generated from the assembly and testing dataset using a random sampling method with replacement.

[0123] The encoder model building module is used to construct a stacked autoencoder corresponding to each training dataset. Each stacked autoencoder consists of an encoder and a decoder. The stacked autoencoder corresponding to each training dataset is trained until convergence, resulting in each trained stacked autoencoder. During training, assembly quality data is used as the input to the stacked autoencoder, and test performance data is used as the feature representation of the encoder and as the input data to the decoder. A Bagging stacked autoencoder model is constructed based on each trained stacked autoencoder.

[0124] The anomaly detection module is used to set a confidence level to determine the dynamic threshold of abnormal data; it uses the assembly and testing data of the aero-engine to be tested to perform anomaly detection on the aero-engine assembly quality based on the dynamic threshold according to the Bagging stacked autoencoder model.

[0125] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method of aeroengine assembly quality anomaly detection and feature inference based on autoencoder, characterized in that, The method comprises the following steps: obtaining an aero-engine test data set, wherein the test data set comprises an assembly quality data set and a corresponding test performance data set; generating multiple training data sets from the test data set by means of random sampling with replacement; constructing a corresponding stacked autoencoder for each training data set, wherein each stacked autoencoder is composed of an encoder and a decoder; training each training data set on the stacked autoencoder corresponding thereto until convergence, thereby obtaining a trained stacked autoencoder; during the training process, using the assembly quality data as the input of the stacked autoencoder, and using the test performance data as the encoded feature representation and the input data of the decoder; constructing a Bagging stacked autoencoder model according to each trained stacked autoencoder; setting a confidence level to determine a dynamic threshold for abnormal data; performing abnormality detection on the aero-engine assembly quality based on the dynamic threshold according to the Bagging stacked autoencoder model and the test data of the aero-engine to be detected.

2. The method of aeroengine assembly quality anomaly detection and feature inference based on autoencoder as claimed in claim 1, wherein, The loss function for training each stacked autoencoder comprises: using the error between the input data and the output result of the stacked autoencoder as the reconstruction error; using the error between the encoding of the stacked autoencoder and the test performance data as the encoding error; using the sum of the reconstruction error and the encoding error as the loss function of the stacked autoencoder.

3. The method of self-encoder based aeroengine assembly quality anomaly detection and feature inference as claimed in claim 2, wherein, The step of determining the dynamic threshold for abnormal data comprises: generating a validation data set from the test data set by means of random sampling with replacement; wherein the training set and the validation set both contain normal data; reconstructing the Bagging stacked autoencoder model using the validation set to obtain the reconstruction error of the validation set; determining the dynamic threshold for abnormal data according to the reconstruction error of the validation set by setting a confidence level.

4. The method of self-encoder based gas turbine engine assembly quality anomaly detection and feature inference of claim 1, wherein, When constructing the Bagging stacked autoencoder model, the method comprises: using the average value of the output data of each trained stacked autoencoder as the final output result of the model.

5. The method for aero-engine assembly quality anomaly detection and feature inference based on autoencoder of claim 1, wherein, The method further comprises: using the set aero-engine test performance as the input of the decoder of each stacked autoencoder in the Bagging stacked autoencoder model to infer the assembly interval of the aero-engine assembly quality feature.

6. The method of self-encoder based gas turbine engine assembly quality anomaly detection and feature inference of claim 5, wherein, When inferring the assembly interval of the aero-engine assembly quality feature, the method comprises: discretizing the set aero-engine test performance to obtain multiple discrete values; using the discrete values as the input of the decoder of each stacked autoencoder in the Bagging stacked autoencoder model to reconstruct the corresponding assembly quality feature values; fitting the reconstructed assembly quality feature values to obtain an assembly quality feature change curve; inferring the assembly interval of the aero-engine assembly quality feature according to the assembly quality feature change curve.

7. A computer program product comprising a computer program, characterized in that, The computer program, when executed by a processor, implements the method for aero-engine assembly quality abnormality detection and feature inference based on an autoencoder according to any one of claims 1 to 6.

8. An electronic device, comprising: The method comprises the following steps: a processor; and a memory for storing executable instructions of the processor; The processor is configured to execute the executable instructions to perform the method for aero-engine assembly quality anomaly detection and feature inference based on a self-encoder according to any one of claims 1 to 6.

9. A system for aeroengine assembly quality anomaly detection and feature inference based on autoencoders, characterized in that, The method comprises: a data acquisition module configured to acquire an assembly and test data set of an aero-engine, wherein the assembly and test data set comprises an assembly quality data set and a corresponding test performance data set; and generate a plurality of training data sets from the assembly and test data set by means of random sampling with replacement; an encoder model construction module configured to construct a stacked self-encoder corresponding to each training data set; each stacked self-encoder is composed of an encoder and a decoder; each training data set is used to train the stacked self-encoder corresponding thereto until convergence, thereby obtaining a trained stacked self-encoder; wherein, during the training process, the assembly quality data is used as the input of the stacked self-encoder; the test performance data is used as the encoded feature representation and as the input data of the decoder; and a Bagging stacked self-encoder model is constructed according to each trained stacked self-encoder; an anomaly detection module configured to set a confidence level to determine a dynamic threshold of abnormal data; and perform anomaly detection on the assembly quality of the aero-engine based on the dynamic threshold according to the Bagging stacked self-encoder model and the assembly and test data of the aero-engine to be detected.

Images