Aero-engine blade defect detection method based on autoencoder continuous learning

Abstract

An aero-engine blade defect detection method based on autoencoder continuous learning comprises the steps that a basic autoencoder model comprising an encoder and a decoder is constructed, and an LoRA module is preset for defect detection tasks of different types or working conditions; for a certain type of defect detection tasks needing to be learned at present, network parameters of a basic auto-encoder model and a corresponding LoRA module are determined through parameter management, and a mixed training set is constructed through knowledge replay; on the mixed training set, carrying out potential space adversarial training on the current basic auto-encoder model and the LoRA module; and reconstructing the to-be-detected blade image and calculating reconstruction errors by using the models trained by all kinds of defect detection tasks and the LoRA module, taking the minimum value in the reconstruction errors as a final abnormal score, comparing the final abnormal score with a preset defect judgment threshold, and judging whether the to-be-detected blade has defects or not. The detection sensitivity is improved, and the adaptability is high.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computer vision and artificial intelligence, and relates to an aero-engine blade defect detection method based on self-encoder continuous learning, which can continuously detect defects of aero-engine blades in a dynamically changing production environment. BACKGROUND

[0002] Aero-engine blades are core high-temperature components that determine the performance, reliability and safety of engines. Surface defects such as cracks, scratches, corrosion and dents may occur during manufacturing and service, so it is crucial to accurately and reliably detect defects. At present, automatic visual detection methods based on deep learning have become the mainstream research direction due to their high precision and efficiency. However, existing technologies still face two core challenges when dealing with real industrial production environments:

[0003] (1) Conflict between catastrophic forgetting and continuous learning:

[0004] In actual production, production lines need to handle continuously changing blade models, production batches, or adapt to changes in imaging environments caused by changes in lighting and camera positions. Training a dedicated model for each new situation will result in huge data, computing and storage overhead. If you try to directly fine-tune the model with new task data on a single model, the model will quickly forget the feature knowledge learned for the old task, resulting in a cliff-like decline in its detection ability for old task blades. This phenomenon is called "catastrophic forgetting". This makes it difficult for traditional models to accumulate and iterate knowledge within a single architecture.

[0005] (2) Trade-off between detection sensitivity and generalization ability:

[0006] Many mainstream unsupervised defect detection methods, especially reconstruction-based models (such as various autoencoders), learn the feature distribution of a large number of normal samples to make abnormality judgments. However, the strong learning and generalization ability of such models is a "double-edged sword". On the one hand, they can model normal samples well, but on the other hand, for those subtle or hidden defects that are very similar to normal samples and are on the decision boundary, the model may "generalize" them as normal patterns, resulting in insignificant reconstruction error and missed detection. This is unacceptable in industrial quality inspection scenarios that require high recall rates.

[0007] Therefore, there is an urgent need for a new type of intelligent defect detection method. This method should not only be able to continuously learn the detection tasks of new types of blades and effectively overcome the problem of catastrophic forgetting, but also be able to improve the model's sensitivity to the identification of small and boundary defects, so as to build a unified detection model that can stably memorize historical knowledge and is highly sensitive to various defects and has strong adaptability. Summary of the Invention

[0008] The purpose of this invention is to address the problems in the prior art by providing a method for detecting defects in aero-engine blades based on continuous learning of an autoencoder. This method enables the defect detection model to continuously learn new tasks while maintaining historical knowledge, thereby improving its sensitivity to the detection of various defects and achieving flexible adaptation to unknown samples in the task.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] Firstly, a method for detecting defects in aero-engine blades based on continuous learning of an autoencoder is provided, including:

[0011] Construct a basic autoencoder model containing an encoder and a decoder, and pre-define an independent and trainable low-rank adaptive LoRA module for defect detection tasks of different types or working conditions.

[0012] For a specific type of defect detection task that needs to be learned, the network parameters of the basic autoencoder model and the corresponding low-rank adaptive LoRA module are determined through parameter management, and a hybrid training set is constructed through knowledge replay. On the hybrid training set, the current basic autoencoder model and the corresponding low-rank adaptive LoRA module are trained in the latent space adversarial mode.

[0013] The basic autoencoder model trained by all types of defect detection tasks and the corresponding low-rank adaptive LoRA module are used to reconstruct the image of the blade to be detected and calculate the reconstruction error. The minimum value of the reconstruction error is taken as the final anomaly score. The final anomaly score is compared with the preset defect judgment threshold to determine whether there is a defect in the blade to be detected.

[0014] As a preferred embodiment, the basic autoencoder model adopts the variational autoencoder (VAE) model. The encoder of the variational autoencoder (VAE) model maps the input blade image to a latent probability distribution, and the decoder samples from the latent probability distribution and reconstructs the blade image.

[0015] As a preferred approach, the independent and trainable low-rank adaptive LoRA module updates the model parameters in the following manner:

[0016] The original weight matrix W0 in the basic autoencoder model is frozen, and the weight update amount ΔW = BA is represented by the product of two low-rank matrices A and B, thereby adjusting the model behavior during forward propagation.

[0017] As a preferred embodiment, the step of determining the network parameters of the basic autoencoder model and the corresponding low-rank adaptive LoRA module through parameter management includes: when k=1, training the basic autoencoder model and the corresponding first low-rank adaptive LoRA module;

[0018] When k>1, freeze the network parameters of the basic autoencoder model, and only keep the low-rank adaptive LoRA module LoRA corresponding to the current k-th task. k This is a trainable state.

[0019] As a preferred embodiment, the construction of the hybrid training set through knowledge replay includes: when k>1, using the basic autoencoder model trained on the (k-1)th type of leaf defect detection task and the decoder in the corresponding low-rank adaptive LoRA module to generate pseudo-sample images for reproducing the normal features of the (k-1)th type of leaf; merging the pseudo-sample images with the real normal sample images of the current kth type of leaf to form a hybrid training set; when k=1, the training set only contains real normal sample images of the kth type of leaf.

[0020] As a preferred approach, the knowledge replay involves sampling a latent vector from a standard normal prior distribution and inputting the latent vector into the decoder corresponding to the (k-1)th type of blade defect detection task to generate the pseudo-sample image.

[0021] As a preferred embodiment, the step of performing latent space adversarial training on the current basic autoencoder model and the corresponding low-rank adaptive LoRA module on the hybrid training set includes:

[0022] The encoder of the current basic autoencoder model encodes the leaf images in the mixed training set into true latent vectors.

[0023] An adversarial generator is introduced, which perturbs the real latent vector to generate adversarial latent vectors. The current basic autoencoder model and the adversarial generator are alternately optimized so that the decoder of the current basic autoencoder model can distinguish between real latent vectors and adversarial latent vectors, generating a first reconstruction error for real latent vectors and a second reconstruction error for adversarial latent vectors. At the same time, the adversarial generator generates adversarial latent vectors that cause the decoder to generate the first reconstruction error.

[0024] Through latent spatial adversarial training, update the low-rank adaptive LoRA module LoRA corresponding to the current k-th task. kAnd the parameters of the adversarial generator.

[0025] As a preferred embodiment, the objective function of the latent space adversarial training includes:

[0026] The loss function of the current basic autoencoder model includes a basic loss to ensure reconstruction quality and latent space regularization, and an adversarial discriminative loss to maximize the difference in reconstruction error between the true latent vector and the adversarial latent vector.

[0027] The loss function of the adversarial generator minimizes the reconstruction error of the generated adversarial latent vector after reconstruction by the decoder.

[0028] As a preferred approach, the image of the blade to be detected is input with all the basic autoencoder models and corresponding low-rank adaptive LoRA modules trained for various defect detection tasks, without knowing its specific category or operating condition. Each trained basic autoencoder model and corresponding low-rank adaptive LoRA module independently reconstructs the image of the blade to be detected and calculates the reconstruction error. The minimum value among all reconstruction errors is taken as the final anomaly score. The reconstructed image generated by the minimum reconstruction error is selected, and the pixel-level residual map between the corresponding reconstructed image and the image of the blade to be detected is calculated to achieve pixel-level defect localization.

[0029] Secondly, a defect detection system for aero-engine blades based on autoencoder continuous learning is provided, including:

[0030] The model building and configuration module is used to build a basic autoencoder model containing an encoder and a decoder, and to pre-set an independent and trainable low-rank adaptive LoRA module for different types or working conditions of defect detection tasks.

[0031] The continuous learning and model optimization module is used to determine the network parameters of the basic autoencoder model and the corresponding low-rank adaptive LoRA module through parameter management for a certain type of defect detection task that needs to be learned. It constructs a hybrid training set through knowledge replay and performs latent space adversarial training on the hybrid training set for the current basic autoencoder model and the corresponding low-rank adaptive LoRA module.

[0032] The defect detection module with unknown tasks is used to reconstruct the image of the blade to be detected using the basic autoencoder model trained on all kinds of defect detection tasks and the corresponding low-rank adaptive LoRA module, and calculate the reconstruction error. The minimum value of the reconstruction error is taken as the final anomaly score. The final anomaly score is compared with the preset defect judgment threshold to determine whether there is a defect in the blade to be detected.

[0033] Compared with the prior art, the present invention has at least the following beneficial effects:

[0034] By freezing the parameters of the basic autoencoder model and configuring an independent low-rank adaptive LoRA module for each task, parameter isolation is achieved. This method physically prevents the destructive modification of old task knowledge by training for new tasks. Constructing a hybrid training set through knowledge replay further consolidates historical knowledge. Compared to pure replay methods, the defect detection method of this invention retains knowledge more stably and thoroughly. When learning a new task, this invention only needs to train and store a very lightweight (few parameters) low-rank adaptive LoRA module, rather than the entire model or decoder. This makes the continuous iteration cost of the model extremely low, significantly reducing training and storage costs, making it very suitable for resource-constrained industrial deployment environments. On the hybrid training set, latent space adversarial training is performed on the current basic autoencoder model and its corresponding low-rank adaptive LoRA module. This forces the model to learn a latent space representation that is more compact and has clearer boundaries for describing normal features, thereby improving sensitivity to small and boundary defects. The original latent space adversarial training mechanism of this invention significantly improves detection sensitivity, effectively solving the pain point of traditional autoencoders being insensitive to small and boundary defects, and improving detection recall and reliability. In the detection phase, the method of this invention reconstructs the image of the blade to be detected using a basic autoencoder model trained on all types of defect detection tasks in parallel, along with the corresponding low-rank adaptive LoRA module. The reconstruction error is then calculated, and the minimum value of the reconstruction error is used as the final anomaly score. This final anomaly score is compared with a preset defect judgment threshold to determine whether the blade to be detected has a defect. This invention's autoencoder-based continuous learning method for aero-engine blade defect detection enhances deployment flexibility and ease of use. Based on an ensemble reasoning mechanism for unknown tasks, it eliminates the step of determining the specific category of a sample during detection, simplifying the detection logic. It can seamlessly adapt to complex scenarios such as mixed production lines, exhibiting stronger robustness and practicality. This invention enables the detection model to continuously learn new tasks while maintaining historical knowledge, and improves its sensitivity to various defects, thereby constructing a unified detection model that can stably memorize historical knowledge and is highly sensitive to various defects, demonstrating strong adaptability.

[0035] Furthermore, the basic autoencoder model of this invention employs a variational autoencoder (VAE) model. The encoder of the VAE model maps the input leaf image to a latent probability distribution, and the decoder samples from this distribution and reconstructs the leaf image. Using a VAE model as the base model provides a probabilistic latent representation, enhancing the model's generative ability and robustness. Instead of a fixed latent vector, the encoder of the VAE model maps the input leaf image to a latent probability distribution, allowing the model to better capture the uncertainty and variability of the data. This probabilistic representation helps generate more diverse pseudo-samples, thus providing richer historical knowledge during knowledge replay and further improving the model's generalization ability to different types of defects.

[0036] Furthermore, this invention achieves pixel-level defect localization by calculating the pixel-level residual map between the reconstructed image and the image of the blade to be inspected. This pixel-level defect localization method can not only determine whether a defect exists on the blade, but also accurately pinpoint the specific location and extent of the defect. By calculating the pixel-level residual between the reconstructed image and the original image, abnormal areas can be visually displayed. This greatly improves the interpretability and practicality of the detection results, making subsequent defect analysis and processing more accurate and efficient. In industrial applications, this precise localization capability can help engineers quickly identify and repair problem areas, improving production efficiency and product quality. Attached Figure Description

[0037] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 The principle architecture diagram of the aero-engine blade defect detection method based on continuous learning of autoencoder in this invention embodiment;

[0039] Figure 2 Comparison chart of the method of this invention with various baseline methods on core performance metrics at the image level;

[0040] Figure 3 A comparison chart of the method of this invention with various baseline methods on pixel-level core performance metrics. Detailed Implementation

[0041] 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, and not all embodiments. Based on the embodiments of the present invention, those skilled in the art can obtain other embodiments without creative effort.

[0042] This invention presents a method for detecting defects in aero-engine blades based on continuous learning of an autoencoder, which overcomes the limitations of traditional methods and aims to simultaneously solve two core challenges in industrial defect detection:

[0043] First, the detection model suffers from catastrophic forgetting when faced with serialization tasks such as new models and new operating conditions;

[0044] Secondly, there is a problem with insufficient sensitivity in detecting boundary and hidden defects that are only slightly different from normal samples.

[0045] Please see Figure 1 The present invention relates to an aero-engine blade defect detection method based on continuous learning of autoencoders. It innovatively integrates a parameter-efficient low-rank adaptive LoRA module for knowledge isolation and preservation, and a latent space adversarial training mechanism to improve the model's discriminative ability, based on a variational autoencoder (VAE).

[0046] Specifically, the aero-engine blade defect detection method based on autoencoder continuous learning in this embodiment of the invention includes:

[0047] Construct a basic autoencoder model containing an encoder and a decoder, and pre-define an independent and trainable low-rank adaptive LoRA module for defect detection tasks of different types or working conditions.

[0048] For a specific type of defect detection task that needs to be learned, the network parameters of the basic autoencoder model and the corresponding low-rank adaptive LoRA module are determined through parameter management, and a hybrid training set is constructed through knowledge replay. On the hybrid training set, the current basic autoencoder model and the corresponding low-rank adaptive LoRA module are trained in the latent space adversarial mode.

[0049] The basic autoencoder model trained by all types of defect detection tasks and the corresponding low-rank adaptive LoRA module are used to reconstruct the image of the blade to be detected and calculate the reconstruction error. The minimum value of the reconstruction error is taken as the final anomaly score. The final anomaly score is compared with the preset defect judgment threshold to determine whether there is a defect in the blade to be detected.

[0050] In one possible implementation, the basic autoencoder model employs a variational autoencoder (VAE) model, comprising an encoder E and a decoder D. In this embodiment of the invention, the network parameters of the basic autoencoder model are frozen during subsequent learning, serving as a shared knowledge base for all tasks. The encoder of the variational autoencoder (VAE) model maps the input blade image to a latent probability distribution, and the decoder samples from the latent probability distribution and reconstructs the blade image.

[0051] In one possible implementation, for the N tasks to be learned, N independent, trainable low-rank adaptive LoRA modules are pre-configured, denoted as {LoRA1, LoRA2, ..., LoRA...}. N}

[0052] Each LoRA i The module contains a pair of tiny low-rank matrices (A i B i The low-rank adaptive LoRA module updates model parameters in the following way to learn specific knowledge for the i-th task:

[0053] The original weight matrix W0 in the basic autoencoder model is frozen, and the weight update amount ΔW = BA is represented by the product of two low-rank matrices A and B, thereby adjusting the model behavior during forward propagation.

[0054] Construct a potential space adversarial generator G adv It is used to improve the model's discriminative ability during training.

[0055] In one possible implementation, for the k-th type of blade task to be learned (k is an integer greater than or equal to 1), the network parameters of the basic autoencoder model and the corresponding low-rank adaptive LoRA module are determined through parameter management, including:

[0056] When k=1, the basic autoencoder model and the corresponding first low-rank adaptive LoRA module are trained to train a high-performance basic autoencoder model, and the parameters of the basic variational autoencoder (VAE) model are frozen after training.

[0057] When k>1, freeze the network parameters of the basic autoencoder model, and only keep the low-rank adaptive LoRA module LoRA corresponding to the current k-th task. k and the adversarial generator G adv This is the trainable state. The pre-trained base variational autoencoder (VAE) model and all historical low-rank adaptive LoRA modules {LoRA1, ..., LoRA} are frozen. k-1 The parameters of}.

[0058] In one possible implementation, constructing a hybrid training set through knowledge replay includes:

[0059] When k>1, the basic autoencoder model trained on the (k-1)th type of leaf defect detection task and the decoder in the corresponding low-rank adaptive LoRA module are used to generate pseudo sample images for reproducing the normal features of the (k-1)th type of leaf; the pseudo sample images are merged with the real normal sample images of the current kth type of leaf to form a mixed training set.

[0060] When k=1, the training set contains only real normal sample images of the k-th type of leaf.

[0061] In one possible implementation, knowledge replay generates the pseudo-sample image by sampling a latent vector from a standard normal prior distribution and inputting the latent vector into the decoder corresponding to the (k-1)th type of blade defect detection task.

[0062] Specifically, this includes loading the historical decoder: when k>1, loading the LoRA from the base decoder and the previous task. k-1 The historical decoder, composed of modules, (k-1) ;

[0063] Generating pseudo-samples: Randomly sampling a batch of latent vectors {z} from a standard normal distribution. noise}, and input into Decoder (k-1) In the process, a batch of pseudo-sample images that can represent the normal characteristics of the (k-1)th type of leaf are generated.

[0064] Constructing a hybrid training set: using pseudo-sample images Compared with the real normal sample image x of the current k-th task (k) Merge to form a hybrid training set

[0065] In one possible implementation, the steps of performing latent spatial adversarial training on the current base autoencoder model and its corresponding low-rank adaptive LoRA module on a hybrid training set include:

[0066] The encoder of the current basic autoencoder model encodes the leaf images in the mixed training set into true latent vectors.

[0067] An adversarial generator is introduced, which perturbs the real latent vector to generate adversarial latent vectors. The current basic autoencoder model and the adversarial generator are alternately optimized so that the decoder of the current basic autoencoder model can distinguish between real latent vectors and adversarial latent vectors, generating a first reconstruction error for real latent vectors and a second reconstruction error for adversarial latent vectors. At the same time, the adversarial generator generates adversarial latent vectors that cause the decoder to generate the first reconstruction error.

[0068] Through latent spatial adversarial training, update the low-rank adaptive LoRA module LoRA corresponding to the current k-th task. k And the parameters of the adversarial generator.

[0069] Specifically, this includes encoding: from the mixed training set Sampling is performed using the encoder and the current low-rank adaptive LoRA module. k The true latent vector z is obtained.

[0070] Generate adversarial examples: Input the real latent vector z into the adversarial generator G adv The adversarial potential vector z is obtained. adv =G adv (z).

[0071] Adversarial optimization: Update LoRA through alternating training. k and G adv The parameters. On the one hand, optimize LoRA. k This minimizes the reconstruction error of the decoder on the true latent vector z, while minimizing the error on the adversarial latent vector z. adv On the one hand, the reconstruction error should be maximized; on the other hand, the adversarial generator G should be optimized. adv This generates the adversarial potential vector z adv To minimize the reconstruction error of the decoder.

[0072] In one possible implementation, the objective function for latent spatial adversarial training includes:

[0073] The loss function of the current basic autoencoder model includes a basic loss to ensure reconstruction quality and latent space regularization, and an adversarial discriminative loss to maximize the difference in reconstruction error between the true latent vector and the adversarial latent vector.

[0074] The loss function of the adversarial generator minimizes the reconstruction error of the generated adversarial latent vector after reconstruction by the decoder.

[0075] In one possible implementation, the image of the blade to be detected, without knowing its specific category or operating condition, is input into the basic autoencoder model and its corresponding low-rank adaptive LoRA module trained for all types of defect detection tasks. Each trained basic autoencoder model and its corresponding low-rank adaptive LoRA module independently reconstructs the image of the blade to be detected and calculates the reconstruction error. The minimum value among all reconstruction errors is taken as the final anomaly score. The reconstructed image generated with the minimum reconstruction error is selected, and the pixel-level residual map between the corresponding reconstructed image and the image of the blade to be detected is calculated to locate the defect region, achieving pixel-level defect localization. Specifically, determining whether the blade to be detected has a defect includes the following steps:

[0076] Integrated parallel inference: x from a leaf image to be detected test (The task category is unknown) are input in parallel into all N "expert models". The i-th expert model consists of "basic autoencoder model + LoRA". i "constitute.

[0077] Parallel reconstruction: Each expert model independently performs encoding and decoding, outputting a reconstructed image. And calculate its reconstruction error compared with the original image.

[0078] Minimum error aggregation: Select the minimum value from all N reconstruction errors as the final outlier score, i.e.:

[0079]

[0080] Defect identification: The final anomaly score S(x) is determined. test The score is compared with a preset defect threshold. If the score is higher than the threshold, the blade is determined to have a defect. Please refer to [link / reference]. Figure 2 and Figure 3 The figure quantitatively demonstrates the significant advantages of the method of the present invention in terms of average performance and forgetting metric compared with other methods, proving the effectiveness of the detection process. Figure 2 The average performance (Image AUROC AP) and forgetting metric (Image FM) of each method on the image classification task are specifically presented. The results intuitively demonstrate that the present invention achieves the lowest possible knowledge forgetting while maintaining high detection accuracy. Figure 3 Specifically, the average performance (Pixel AUPR AP) and forgetting metric (PixelFM) of each method on the defect segmentation task are presented. The results demonstrate that the present invention also outperforms existing technologies in pixel-level defect localization capabilities while maintaining optimal performance stability.

[0081] To verify the claimed beneficial effects of the method in the embodiments of the present invention, a comprehensive performance comparison was conducted between the method of the embodiments of the present invention and various existing technologies (baseline methods) in a real industrial scenario continuous learning task for detecting surface defects on aero-engine blades. Specific quantitative comparison results are as follows: Figure 2 and Figure 3 As shown.

[0082] Please see Figure 2This figure illustrates the performance comparison at the image level. In the figure, "Average Performance (Image AUROCAP)" represents the model's average classification accuracy across all learned tasks, a key metric for measuring overall model performance; a higher value is better. "Image Forgetting Metric (Image FM)" precisely quantifies the degree to which the model forgets historical knowledge after learning a new task; a lower value indicates stronger knowledge retention. Figure 2 It can be clearly seen that the AP value of the method of the present invention (marked as 'Ours' in the figure) is the highest, proving that its overall detection accuracy is the best; at the same time, its FM value is significantly lower than all other comparative methods, close to zero, which strongly proves the superior ability of the present invention in suppressing catastrophic forgetting.

[0083] Please see Figure 3 The figure further illustrates the performance comparison on the more challenging pixel-level (i.e., defect localization) task. Pixel-level "Average Performance (Pixel AUPR AP)" and "Pixel Forgetting Metric (Pixel FM)" are core metrics for measuring the model's ability to accurately segment defect regions and maintain this stability. As shown in the figure, the aero-engine blade defect detection method based on autoencoder continuous learning in this embodiment of the invention also achieves the best performance on the pixel-level AUPR metric and maintains the lowest forgetting metric (FM). This indicates that the method of the present invention can not only accurately determine whether the blade has a defect, but also accurately locate the defect, and this accurate localization ability does not degrade with learning new tasks.

[0084] In summary, Figure 2 and Figure 3 The quantitative experimental results, from both image and pixel dimensions, jointly and powerfully demonstrate the effectiveness of the technical solution proposed in the embodiments of the present invention. Compared with the prior art, it has significant and quantifiable advantages in effectively mitigating catastrophic forgetting and maintaining a high level of detection and positioning accuracy, fully verifying its beneficial effects.

[0085] This invention presents a novel aero-engine blade defect detection method based on autoencoder continuous learning. This method addresses the catastrophic forgetting problem inherent in existing detection models when facing tasks involving multiple engine models, sequential processing, or varying lighting conditions, as well as the challenge of insufficient sensitivity to detect minute and boundary defects. The method proposes a novel technical solution: the main framework consists of a basic autoencoder model with fixed parameters and a highly lightweight, independently configured low-rank adaptive LoRA module for each task. When learning a new task, only the corresponding LoRA module is trained, fundamentally suppressing knowledge forgetting through parameter isolation strategies, and supplemented by generative replay to consolidate historical knowledge. Furthermore, this invention innovatively employs a latent space adversarial training mechanism. By engaging in adversarial games in the latent space, the model is forced to learn a more compact and accurate description of normal features, thereby significantly improving its ability to identify minute defects. During defect detection, this invention adopts a task-unknown ensemble inference strategy, calculating the reconstruction error of all learned tasks in parallel and taking the minimum value as the final discrimination score. The defect detection method of this invention, through the above design, not only completely alleviates catastrophic forgetting and greatly reduces training and storage costs, but also significantly improves the detection rate of difficult defects, enabling a single model to efficiently and robustly adapt to dynamically changing industrial production environments. Furthermore, thanks to its efficient knowledge iteration and retention capabilities, the aero-engine blade defect detection method based on autoencoder continuous learning of this invention can be widely extended to other precision manufacturing fields that require flexible responses to production line changes, such as intelligent quality inspection of automotive parts, semiconductors, and high-end medical devices.

[0086] This invention also proposes an aero-engine blade defect detection system based on autoencoder continuous learning, comprising:

[0087] The model building and configuration module is used to build a basic autoencoder model containing an encoder and a decoder, and to pre-set an independent and trainable low-rank adaptive LoRA module for different types or working conditions of defect detection tasks.

[0088] The continuous learning and model optimization module is used to determine the network parameters of the basic autoencoder model and the corresponding low-rank adaptive LoRA module through parameter management for a certain type of defect detection task that needs to be learned. It constructs a hybrid training set through knowledge replay and performs latent space adversarial training on the hybrid training set for the current basic autoencoder model and the corresponding low-rank adaptive LoRA module.

[0089] The defect detection module with unknown tasks is used to reconstruct the image of the blade to be detected using the basic autoencoder model trained on all kinds of defect detection tasks and the corresponding low-rank adaptive LoRA module, and calculate the reconstruction error. The minimum value of the reconstruction error is taken as the final anomaly score. The final anomaly score is compared with the preset defect judgment threshold to determine whether there is a defect in the blade to be detected.

[0090] Another embodiment of the present invention also provides an electronic device, comprising:

[0091] A memory for storing at least one instruction; and a processor for executing the instructions stored in the memory to implement the aero-engine blade defect detection method based on autoencoder continuous learning.

[0092] Another embodiment of the present invention also proposes a computer-readable storage medium storing at least one instruction, which is executed by a processor in an electronic device to implement the aero-engine blade defect detection method based on autoencoder continuous learning.

[0093] For example, the instructions stored in the memory can be divided into one or more modules / units. These modules / units are stored in a computer-readable storage medium and executed by the processor to complete the autoencoder-based continuous learning-based aero-engine blade defect detection method described in this embodiment of the invention. The one or more modules / units can be a series of computer-readable instruction segments capable of performing specific functions, and these instruction segments describe the execution process of the computer program on the server.

[0094] The electronic device may be a smartphone, laptop, PDA, or cloud server, among other computing devices. It may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the electronic device may also include more or fewer components, or combinations of certain components, or different components; for example, it may also include input / output devices, network access devices, buses, etc.

[0095] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0096] The memory can be an internal storage unit of the server, such as a hard drive or RAM. Alternatively, it can be an external storage device, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory can include both internal and external storage units. The memory is used to store computer-readable instructions and other programs and data required by the server. It can also be used to temporarily store data that has been output or will be output.

[0097] It should be noted that the information interaction and execution process between the above-mentioned module units are based on the same concept as the method embodiment. For details on their specific functions and technical effects, please refer to the method embodiment section. They will not be repeated here.

[0098] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0099] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying the computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0100] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0101] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. An aero-engine blade defect detection method based on self-encoder continuous learning, characterized in that, The application relates to a method for defect detection of wind turbine blades. The method comprises the following steps: A basic autoencoder model comprising an encoder and a decoder is constructed, and a low-rank adaptive LoRA module independent and trainable for different types or working conditions of defect detection tasks is preset; For a certain type of defect detection task that needs to be learned, the network parameters of the basic autoencoder model and the corresponding low-rank adaptive LoRA module are determined through parameter management, and a mixed training set is constructed through knowledge replay; 2. The method of claim 1, wherein, The basic autoencoder model and the corresponding low-rank adaptive LoRA module trained for all types of defect detection tasks are used to reconstruct the blade image to be detected and calculate the reconstruction error, the minimum value in the reconstruction error is taken as the final anomaly score, and the final anomaly score is compared with a preset defect judgment threshold to determine whether the blade to be detected has defects.

3. The method of claim 1, wherein, The basic autoencoder model adopts a variational autoencoder (VAE) model, the encoder of the variational autoencoder (VAE) model maps the input blade image to a latent probability distribution, and the decoder samples and reconstructs the blade image from the latent probability distribution. The independent and trainable low-rank adaptive LoRA module updates the model parameters in the following way:

4. The method of claim 1, wherein, The original weight matrix W0 in the basic autoencoder model is frozen, and the weight update amount △W=BA is represented by learning the product of two low-rank matrices A and B, so as to realize the adjustment of the model behavior in the forward propagation. When k>1, freeze the network parameters of the base autoencoder model, only keep the low-rank adaptive LoRA module LoRA corresponding to the current kth task k is in a trainable state.

5. The method of claim 1, wherein, The network parameters of the basic autoencoder model and the corresponding low-rank adaptive LoRA module are determined through parameter management, which comprises the following steps:

6. The method of claim 5, wherein, When k=1, the basic autoencoder model and the corresponding first low-rank adaptive LoRA module are trained.

7. The method of claim 1, wherein, When k>1, the decoder of the basic autoencoder model and the corresponding low-rank adaptive LoRA module trained for the k-1th type of blade defect detection task are used to generate a pseudo sample image for reproducing the normal features of the k-1th type of blade; the pseudo sample image and the real normal sample image of the kth type of blade are merged to form a mixed training set; when k=1, the training set only contains the real normal sample image of the kth type of blade. The knowledge replay is realized by sampling a latent vector from a standard normal prior distribution and inputting the latent vector into the decoder corresponding to the k-1th type of blade defect detection task to generate the pseudo sample image. The latent space adversarial training of the current basic autoencoder model and the corresponding low-rank adaptive LoRA module on the mixed training set comprises the following steps: The encoder of the current basic autoencoder model is used to encode the blade image in the mixed training set into a real latent vector; introducing an adversarial generator, the adversarial generator applying perturbations to the real latent vector to generate an adversarial latent vector; alternately optimizing the current base autoencoder model and the adversarial generator, so that the decoder of the current base autoencoder model can distinguish the real latent vector from the adversarial latent vector, and produce a first reconstruction error for the real latent vector and a second reconstruction error for the adversarial latent vector; at the same time, the adversarial generator generates an adversarial latent vector that can make the decoder produce the first reconstruction error; update the low-rank adaptive LoRA module LoRA corresponding to the current k-th task through latent space adversarial training k and the parameters of the adversarial generator.

8. The method of claim 7, wherein, The objective function of the latent space adversarial training includes: The loss function of the current base autoencoder model includes a base loss for ensuring reconstruction quality and latent space regularization, and an adversarial discriminant loss for maximizing the difference in reconstruction error between the real latent vector and the adversarial latent vector; The loss function of the adversarial generator minimizes the reconstruction error of the generated adversarial latent vector after being reconstructed by the decoder.

9. The method of claim 1, wherein, The to-be-detected blade image is input into the base autoencoder model and the corresponding low-rank adaptive LoRA module trained for all defect detection tasks without knowing the specific category or working condition of the to-be-detected blade image. Each trained base autoencoder model and corresponding low-rank adaptive LoRA module independently reconstructs the to-be-detected blade image and calculates the reconstruction error. The minimum value of all reconstruction errors is taken as the final anomaly score. The reconstructed image generated by the minimum reconstruction error is selected. The pixel-level residual map between the corresponding reconstructed image and the to-be-detected blade image is calculated to realize pixel-level defect positioning.

10. An aero-engine blade defect detection system based on self-encoder continual learning, characterized in that, It includes: A model construction and configuration module is used to construct a base autoencoder model including an encoder and a decoder, and to pre-set an independent and trainable low-rank adaptive LoRA module for different types or working conditions of defect detection tasks; A continuous learning and model optimization module is used to determine the network parameters of the base autoencoder model and the corresponding low-rank adaptive LoRA module for a certain type of defect detection task that needs to be learned at present by parameter management, and to build a mixed training set by knowledge replay; on the mixed training set, the latent space adversarial training is performed on the current base autoencoder model and the corresponding low-rank adaptive LoRA module; A task-unknown defect detection module is used to reconstruct the to-be-detected blade image by using all the base autoencoder models and corresponding low-rank adaptive LoRA modules trained for various defect detection tasks and to calculate the reconstruction error. The minimum value of the reconstruction error is taken as the final anomaly score. The final anomaly score is compared with the pre-set defect judgment threshold to determine whether the to-be-detected blade has defects.

Images