Aeroengine Sample Data Imbalance Blade Crack Diagnosis Method and System

Abstract

The present invention discloses a method and system for diagnosing blade cracks with unbalanced sample data of an aero-engine, including: obtaining three-dimensional tip clearance signals of blades, and establishing a deep generation network diagnosis model based on a conditional variational autoencoder network; constructing a deep focal loss function based on a small number of fault class sample sets; combining the deep focal loss function with the deep generation network diagnosis model to obtain a dynamically weighted conditional variational autoencoder network, and expanding a small number of fault class samples in the unbalanced class sample set; mixing the small number of fault class sample sets with the original sample set to form a training set, using a feature extraction network to extract features from the training set, and performing feature fusion in a tensor fusion layer, and finally establishing a complex non-linear mapping relationship between the fused features and the fault types to complete the training of the sample unbalance diagnosis model; inputting the three-dimensional tip clearance samples of the blades to be diagnosed into the established sample unbalance diagnosis model to obtain the health state.

Description

Technical Field

[0001] The present invention belongs to the technical field of aero-engine fault diagnosis, and particularly relates to a method and system for diagnosing blade cracks with unbalanced sample data of an aero-engine. Background Art

[0002] With the development of science and technology, complex mechanical equipment such as aero-engines and large wind turbines is constantly developing towards the direction of being large-scale, complex, and high-speed. Due to the very complex structure and environment of the aero-engine itself, various components of the aero-engine are prone to failures. The turbine blade is the core part of the aero-engine. Once a failure occurs, it will have a serious impact on the healthy operation of the whole machine. With the extensive and in-depth application of artificial intelligence technology in the field of fault diagnosis, there is a new method to directly obtain effective operation information reflecting the health state of the equipment from the monitoring data, and then accurately identify its health state. Therefore, using artificial intelligence technology to establish an intelligent fault diagnosis model has become an important means to ensure the safe operation of mechanical equipment.

[0003] In engineering practice, since the turbine blade is in a normal state for a long time during operation, there is a serious imbalance in the distribution of the monitored normal data and fault data, forming an unbalanced monitoring data set. The diagnostic models based on traditional intelligent algorithms are all established under the assumption of uniform sample distribution, making it difficult to obtain the state information of a small number of fault class samples from the unbalanced data set, resulting in greater limitations of the trained diagnostic model, lower recognition ability for the health states of a small number of fault classes, and poor overall diagnostic ability. Therefore, how to improve the diagnostic accuracy of a small number of fault class samples under the condition of unbalanced data sets is a widely studied problem. Summary of the Invention

[0004] To overcome the limitation that traditional fault diagnosis methods can only effectively diagnose the situation where the number of samples in different health states is basically the same, the present invention provides a method and system for diagnosing blade cracks with unbalanced sample data of an aero-engine. The present invention can improve the diversity and generalization of the sample set, weaken the influence of sample imbalance on the diagnostic result, and further improve the diagnostic accuracy of the sample imbalance diagnostic model under the condition of unbalanced sample distribution.

[0005] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0006] A method for diagnosing blade cracks with unbalanced sample data of an aero-engine, comprising the following steps:

[0007] Obtain the three-dimensional tip clearance signal of the blade, and establish a deep generation network diagnostic model based on a conditional variational autoencoder network;

[0008] Construct a deep focal loss function based on a small number of fault class sample sets;

[0009] Combine the depth focusing loss function with the depth generation network diagnosis model to obtain a dynamically weighted conditional variational encoding network, and expand the minority fault class samples in the imbalanced class sample set;

[0010] Mix the minority fault class sample set with the original sample set to form a training set, use the feature extraction network to extract features from the training set, and perform feature fusion in the tensor fusion layer. Finally, establish a complex non-linear mapping relationship between the fused features and the fault types to complete the training of the sample imbalance diagnosis model;

[0011] Input the three-dimensional blade tip clearance sample to be diagnosed into the established sample imbalance diagnosis model to obtain the health status.

[0012] As a further improvement of the present invention, obtaining the three-dimensional blade tip clearance signal of the blade includes:

[0013] Collect the three-dimensional blade tip clearance signal of the turbine blade using the turbine blade fault simulation platform;

[0014] Collect the three-dimensional blade tip clearance signals in different crack states of the blade to form a sample set There are a total of X health states; among them, x m ∈R N×3 is the m-th three-dimensional blade tip clearance signal sample, which contains three-dimensional parameters consists of N data points, y m ∈R C×1 represents the health status label vector of the m-th sample, C is the number of health states, and M is the total number of samples.

[0015] As a further improvement of the present invention, the depth generation network diagnosis model based on the conditional variational encoding network includes a data generation layer, a feature extraction layer, a tensor fusion layer, and a classification layer; specifically, a tensor fusion layer is introduced after the feature extraction layer, and a classification layer is added at the output end to form a depth generation network diagnosis model, and the collected three-dimensional blade tip clearance signal is used as the input.

[0016] As a further improvement of the present invention, construct a depth focusing loss function based on the minority fault class sample set; including:

[0017] Construct a depth focusing loss function that focuses on the learning of minority fault class samples, which consists of L c and L r two parts. L c considers the losses provided by the minority fault class samples and the majority normal class samples to the network, and weights the sample losses. L r is mainly used to punish the misclassification probability, as follows:

[0018]

[0019] where M is the number of samples, y i is the label vector of the samples, p i is the predicted probability vector of the samples, is the dynamic focusing parameter, and α, β are adjustable parameters.

[0020] As a further improvement of the present invention, a training set is formed by mixing the minority fault class sample set with the original sample set. The feature extraction network is used to extract features from the training set, and feature fusion is performed in the tensor fusion layer. Finally, a complex non-linear mapping relationship between the fused features and the fault types is established to complete the training of the sample imbalance diagnosis model; including:

[0021] Mix the minority fault class sample set with the original sample set to form a training set, and eliminate the unbalanced distribution of samples by generating fault samples;

[0022] Use the feature extraction layer to extract features from the training sample set to obtain feature information of different dimensions;

[0023] Use the tensor fusion layer to fuse features of different dimensions, and the obtained fused features contain feature information within different feature dimensions and their mutual correlations;

[0024] Take the fused features as the input of the classification layer, and continuously optimize the error distance between the output of the classification layer and the blade fault types;

[0025] Iteratively optimize the sample imbalance diagnosis model composed of the data generation layer, feature extraction layer, tensor fusion layer, and classification layer in sequence.

[0026] As a further improvement of the present invention, constructing the feature extraction layer includes:

[0027] Mix the generated fault sample set with the unbalanced sample set to form a training sample set, and eliminate the unbalanced distribution of samples by generating fault samples; perform fault feature extraction on the training sample set, and take the extracted feature f m as the output of the feature extraction layer. The extraction process is as follows:

[0028] f m = g θ (x m )

[0029] where: f m is the feature output after the training sample x m passes through the feature extraction layer, g θ is the transformation function of the feature extraction layer, and θ is the set of parameters to be optimized in the feature extraction layer.

[0030] As a further improvement of the present invention, constructing a tensor fusion layer includes:

[0031] Joint modeling of the dynamic characteristics among multi-dimensional features,

[0032] Fusing the features of different dimensions of the extracted three-dimensional tip clearance, and the fusion process is as follows:

[0033]

[0034] In the formula: is the outer product operation of tensors, z is the input tensor, k is the number of tensors, r is the rank of the tensor, and w is the weight matrix.

[0035] As a further improvement of the present invention, constructing a classification layer includes:

[0036] Classifying the output features of the tensor fusion layer to obtain the prediction probability that the sample x m belongs to the c-th health state as:

[0037]

[0038] In the formula, w and b are the weight matrix and the bias term respectively;

[0039] Based on the Adam optimization algorithm, optimize the parameters to be optimized in the feature extraction layer, the tensor fusion layer, and the classification layer in turn. The output of the classification layer needs to minimize the following objective function:

[0040]

[0041] In the formula: M is the number of input samples, is the probability that the sample x m is predicted to be in the c-th health state, and Ι(·) is the indicator function.

[0042] An aero-engine sample data unbalanced blade crack diagnosis system includes:

[0043] An acquisition module, used to acquire the three-dimensional tip clearance signal of the blade, establish a deep generation network diagnosis model based on the conditional variational autoencoder network; construct a deep focal loss function based on a small number of fault class sample sets;

[0044] An expansion module, used to combine the deep focal loss function with the deep generation network diagnosis model to obtain a dynamic weighted conditional variational autoencoder network, and expand the small number of fault class samples in the unbalanced class sample set;

[0045] A training module, which is used to mix a small number of fault class sample sets with the original sample set to form a training set, extract features from the training set using a feature extraction network, perform feature fusion in a tensor fusion layer, and finally establish a complex non-linear mapping relationship between the fused features and the fault types to complete the training of the sample imbalance diagnosis model;

[0046] A diagnosis module, which is used to input the three-dimensional blade tip clearance samples to be diagnosed into the established sample imbalance diagnosis model to obtain the health status.

[0047] Compared with the prior art, the present invention has the following advantages:

[0048] To overcome the limitation that traditional fault diagnosis methods can only effectively diagnose the situation where the number of samples in different health states is basically the same, the present invention combines a deep generation model with cost-sensitive learning, dynamically assigns different weight values to different classes of samples in the class-imbalanced sample set, focuses on the learning of a small number of fault class samples, can effectively establish a decision boundary between the majority of normal samples and the minority of fault class samples, while generating a small number of fault class samples, can improve the diversity and generalization of the sample set, weaken the influence of sample imbalance on the diagnosis result, and further improve the diagnosis accuracy of the intelligent diagnosis model under the condition of uneven sample distribution. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The following further elaborates on the present invention in detail in conjunction with the accompanying drawings and specific embodiments;

[0050] Figure 1 is a flowchart of the method for diagnosing blade cracks with imbalanced aero-engine sample data;

[0051] Figure 2 is a schematic diagram of the DWCAVE-DFL model;

[0052] Figure 3 is a schematic diagram of the sample imbalance diagnosis model of the present invention;

[0053] Figure 4 is an effect diagram of the deep focus loss function;

[0054] Figure 5 is the three-dimensional feature clustering of the generated sample set;

[0055] Figure 6 is the distribution diagram of the diagnosis results;

[0056] Figure 7 is a system diagram of the aero-engine sample data imbalance blade crack diagnosis. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0057] To enable those skilled in the art to better understand the solution of the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0058] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily need to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments of the present invention described here can be implemented in an order other than those illustrated or described here. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0059] Currently, the solutions to the problem of diagnosing unbalanced dataset faults are mainly considered from two perspectives: (1) data-level methods, which reduce or eliminate sample imbalance by adjusting the distribution of training samples and then use intelligent algorithms for classification and recognition. The main methods include oversampling, undersampling, and hybrid sampling, etc.; (2) algorithm-level methods, which do not change the distribution of training sample data and improve the intelligent algorithm to pay more attention to minority fault class samples to adapt to the classification of unbalanced data. The main methods include cost-sensitive learning and ensemble learning, etc. Among them, the sample data generated by traditional data generation methods has poor generalization and is prone to error samples, and the improvement effect on the performance of the diagnostic algorithm is limited. At the same time, the Conditional Variational Autoencoder (CVAE) is a new deep generative model, which effectively avoids the disadvantages of traditional methods. The method of the present invention combines two methods at the data level and the algorithm level, uses them as a breakthrough for intelligent diagnosis of unbalanced data, strengthens the learning ability of traditional deep generative models for minority fault class samples, adaptively extracts the blade fault state feature information contained in the unbalanced dataset, and performs intelligent diagnosis based on these features.

[0060] The present invention combines the proposed deep focal loss function with the conditional variational coding network to solve the class imbalance problem of two types of samples, normal class and fault class, and applies it to the crack fault diagnosis of aero-engine blades.

[0061] Refer toFigure 1 , the first object of the present invention is to provide a method for diagnosing blade cracks with unbalanced aero-engine sample data, including the following steps:

[0062] Obtain the three-dimensional tip clearance signal of the blade and establish a deep generative network diagnosis model based on the conditional variational autoencoder network;

[0063] Construct a deep focal loss function based on the minority fault class sample set;

[0064] Combine the deep focal loss function with the deep generative network diagnosis model to obtain a dynamically weighted conditional variational autoencoder network, and expand the minority fault class samples in the unbalanced class sample set;

[0065] Mix the minority fault class sample set with the original sample set to form a training set, use the feature extraction network to extract features from the training set, and perform feature fusion in the tensor fusion layer. Finally, establish a complex non-linear mapping relationship between the fused features and the fault types to complete the training of the sample unbalance diagnosis model;

[0066] Input the three-dimensional tip clearance sample of the blade to be diagnosed into the established sample unbalance diagnosis model to obtain the health status.

[0067] First, in view of the fact that the traditional conditional variational autoencoder network does not consider the loss weight ratio of a small number of sample classes when dealing with class-imbalanced data, making it difficult to effectively generate minority fault class samples, the present invention proposes a deep focal loss function, which enables the network to focus on the learning of minority fault class samples by assigning dynamic weights to different sample losses, realizing the overall optimization of the class-imbalanced sample set. Secondly, combine the deep focal loss function with the conditional variational autoencoder network to perform data augmentation on the problem of unbalanced crack fault data samples based on three-dimensional tip clearance, eliminating the influence of sample class (normal class and fault class) imbalance on the diagnosis result. Finally, use a multi-layer neural network to extract the features of samples in different health states and identify the cracks for fault recognition.

[0068] The experimental results show that the method proposed in this paper exhibits good diagnostic performance under sample class imbalance, improving the diagnostic accuracy rate, and is significantly better than the traditional fault diagnosis methods for class-imbalanced samples.

[0069] The following is a detailed description in combination with specific embodiments.

[0070] Embodiment 1

[0071] The present invention provides a method for diagnosing blade cracks with unbalanced aero-engine sample data, including the following steps:

[0072] Step 1) Use the turbine blade fault simulation platform to obtain the three-dimensional tip clearance signal of the blade, and establish a deep generative network diagnosis model based on the conditional variational autoencoder network. This model consists of a data generation layer, a feature extraction layer, a tensor fusion layer, and a classification layer, etc.

[0073] In the deep structure network diagnosis model based on the conditional variational autoencoder network established in Step 1), in order to jointly model the dynamic characteristics between multi-dimensional features and obtain feature information containing the inter-correlation within and between different feature dimensions, a tensor fusion layer is introduced after the feature extraction layer. In addition, in order to realize the fault diagnosis function, a classification layer is added at the output end to form a deep generative network diagnosis model, and the collected three-dimensional tip clearance signal is used as the input.

[0074] Step 2) Considering that the imbalanced sample set has a certain impact on the generative model and makes it difficult to effectively generate data of a small number of fault classes, a deep focal loss function is proposed to dynamically weight the samples of a small number of fault classes, increasing their proportion in the network loss. At the same time, it also takes into account the role of the majority of normal class samples in network optimization, avoiding the defect that the traditional loss function only optimizes the samples of a small number of fault classes and ignores the role of the majority of positive class samples.

[0075] In Step 2), a new deep focal loss function is proposed, which can provide a greater gradient optimization for the network, avoiding the problem that the network learning process becomes slow or even stops prematurely due to gradient vanishing, and is conducive to the network focusing on the feature learning of samples of a small number of fault classes.

[0076] Step 3) Combine the loss function proposed in Step 2) with the conditional variational autoencoder network to expand the samples of a small number of fault classes in the imbalanced class sample set, so as to eliminate the influence of sample imbalance on the subsequent diagnosis process.

[0077] In Step 3), combining the deep focal loss function with the conditional variational autoencoder network can strengthen the network's learning of samples of a small number of fault classes, and also take into account the optimization of the majority of normal class samples, avoiding the deficiency that the traditional generative network cannot take into account both samples of a small number of fault classes and the majority of normal class samples, and effectively improving the network feature learning performance.

[0078] Step 4) Mix the sample set of a small number of fault classes obtained in Step 3) with the original sample set to form a training set, use the feature extraction network to extract features from the training set, and perform feature fusion in the tensor fusion layer. Finally, establish a complex non-linear mapping relationship between the fused features and the fault types to complete the training of the diagnosis model.

[0079] In step 4), the generated set of minority fault class samples is mixed with the original sample set to form a training set, and the feature extraction layer is used to extract features from the training set to obtain feature information of different dimensions. In order to explore the correlation between information of different dimensions, the tensor fusion layer is used to fuse features of different dimensions. The fused features are used as the input of the classification layer, and the training of the diagnostic model is completed by minimizing the error between the output of the classification layer and the blade fault type.

[0080] Example 2

[0081] Step 1: Use the turbine blade fault simulation platform to collect the three-dimensional tip clearance signal of the turbine blade, and establish a deep generative network diagnostic model based on the conditional variational autoencoder network. This model mainly consists of a data generation layer, a feature extraction layer, a tensor fusion layer, a classification layer, etc.

[0082] In step 1), the established data generation layer simultaneously includes three groups of DWCAVE-DFL networks, which respectively learn the signals of the three dimensions of the three-dimensional tip clearance. The DWCAVE-DFL network consists of an encoding layer, a sampling layer, and a decoding layer. The encoding layer is used to learn the probability distribution of the features of the learning samples, the sampling layer samples the sample probability distribution to obtain a set of latent variables, and the decoder reconstructs the input samples through the latent variables.

[0083] Step 2: Construct a deep focal loss function (Deep Focal Loss), which focuses on the learning of minority fault class samples. It consists of L c and L r and consists of two parts. L c considers the losses provided by minority fault class samples and majority normal class samples for the network, and can also realize the weighting of sample losses, so that minority fault class samples have a higher loss weight ratio. L r is mainly used to punish the misclassification probability and improve the prediction probability of the correct class. Its formula is as follows:

[0084]

[0085] where M is the number of samples, y i is the label vector of the sample, p i is the prediction probability vector of the sample, γ * is the dynamic focusing parameter, and α, β are adjustable parameters.

[0086] In step 2), the constructed deep focal loss function can provide a greater loss weight for the network, effectively avoid the disappearance of gradients in the later stage of network optimization, resulting in a decrease in the network learning rate and falling into a local optimum. At the same time, it can effectively focus on the feature learning of minority fault class samples and optimize the overall learning effect of unbalanced set samples.

[0087] Step 3: Combine the depth focusing function with the conditional variational autoencoder network to obtain a dynamic weighted conditional variational autoencoder network (DWCAVE-DFL), which can augment a small number of fault class samples in an imbalanced class sample set, improve the diversity and generalization of the sample set, and eliminate the impact of imbalance on the subsequent diagnosis process.

[0088] In the above step 3), the depth focusing function is combined with the conditional variational autoencoder network. By using the sample loss value generated by the depth focusing function, the conditional variational autoencoder network focuses on learning a small number of fault class samples, effectively overcoming the deficiency that the traditional conditional variational autoencoder network does not consider the loss weight ratio of a small number of fault samples when dealing with imbalanced data and is difficult to effectively generate a small number of fault class samples.

[0089] Step 4: Use the DWCAVE-DFL in step 3) to generate fault class samples, mix them with the imbalanced sample set to form a training set, perform feature learning on the training set through a feature extraction layer, fuse the learned features in a tensor fusion layer, and finally establish a complex non-linear mapping relationship between the features and the fault types to complete the training of the diagnostic model.

[0090] In the above step 4), the distribution of the fault class samples generated by the DWCAVE-DFL in step 3) is more consistent with the distribution of the fault class samples in the imbalanced sample set. Mix them with the imbalanced sample set to form a training sample set, and eliminate the imbalanced distribution of the samples through the generated fault samples. Use the feature extraction layer to extract features from the training sample set to obtain feature information of different dimensions. In order to further establish the correlation between information in different dimensions, use the tensor fusion layer to fuse features of different dimensions, and the obtained fused features contain feature information within different feature dimensions and their mutual correlations. Use the fused features as the input of the classification layer, and complete the training of the diagnostic model by continuously optimizing the error distance between the output of the classification layer and the blade fault types.

[0091] Step 5: Use the deep generative network diagnostic model determined in step 4) to perform intelligent diagnosis of blade cracks under sample imbalance.

[0092] Embodiment 3

[0093] The present invention will be further described in detail below with reference to the accompanying drawings and experiments.

[0094] Refer to Figure 2 , a method for diagnosing blade cracks with imbalanced aero-engine sample data, comprising the following steps:

[0095] Step 1: Collect three-dimensional tip clearance signals in different degrees of blade crack states to form a sample set There are a total of X health states. Among them, x m ∈R N×3 is the m-th three-dimensional tip clearance signal sample, which contains three-dimensional parameters consisting of N data points, y m ∈R C×1 represents the health state label vector of the m-th sample, C is the number of health states, and M is the total number of samples.

[0096] Step 2: Refer to Figure 3 , construct a dynamic weighted variational autoencoder (DWCAVE-DFL) based on the deep focal loss function, learn the distribution of the input three-dimensional tip clearance signal samples, and obtain the latent variables through sampling layer, which are used as the input of the decoding layer for sample reconstruction. In addition, in order to prevent the overfitting of the network and accelerate the convergence rate of the network, batch normalization technology (Batch normalization, BN) is used between network layers. The loss value of DWCAVE-DFL consists of L KL , L G and L DFL in three parts, that is, the loss function to be optimized by DWCAVE-DFL is as follows

[0097] L = L KL + L G + L DFL

[0098] The specific description is as follows:

[0099] Since the true probability distribution of the imbalanced sample set cannot be calculated, L KL is used to measure the difference between the true probability distribution of the sample and the probability distribution of the latent variable, and its expression is

[0100]

[0101] where μ is a set of means and σ is a set of variance values.

[0102] L G is used to calculate the difference between the input sample x and the reconstructed sample , and its expression is

[0103]

[0104] L DFLThen a weighted loss value focusing on a small number of fault-class samples is provided, which can dynamically weight the losses of the small number of fault-class samples, increasing their proportion in the network loss. At the same time, it also takes into account the role of the majority of normal-class samples in network optimization, avoiding the defect that the traditional loss function only optimizes the small number of fault-class samples and ignores the majority of normal classes. Considering that the network becomes more and more accurate in identifying a small number of fault-class samples and a majority of normal-class samples during the optimization process, and its division boundary is constantly changing, the dynamic focusing parameter γ inital is the initial focusing parameter value, and its expression is

[0105]

[0106] where M is the number of samples, y i is the label vector of the sample, p i is the predicted probability vector of the sample, and α, β are adjustable parameters.

[0107] Step 3: Refer to Figure 3 , construct a feature extraction layer, mix the fault sample set generated in Step 2 with the unbalanced sample set to form a training sample set, and eliminate the unbalanced distribution of samples through the generated fault samples. Extract fault features from the training sample set, and use the extracted feature f m as the output of the feature extraction layer. The extraction process is as follows:

[0108] f m = g θ (x m )

[0109] where: f m is the feature output after the training sample x m passes through the feature extraction layer, g θ is the transformation function of the feature extraction layer, and θ is the set of parameters to be optimized in the feature extraction layer;

[0110] Step 4: Refer to Figure 3 , construct a tensor fusion layer. In order to realize the joint modeling of the dynamic characteristics between multi-dimensional features,

[0111] fuse the features of different dimensions of the three-dimensional tip clearance extracted in Step 3 to obtain a fusion feature that includes the feature information within and the mutual correlation between different dimensions of features. The fusion process is as follows:

[0112]

[0113] where: is the outer product operation of tensors, z is the input tensor, k is the number of tensors, r is the rank of the tensor, and w is the weight matrix.

[0114] Step 5: Refer to Figure 3 , construct a classification layer to classify the output features of the tensor fusion layer, and then obtain the sample x m The predicted probability that the sample x

[0115]

[0116] belongs to the c-th health state is:

[0117] In the formula, w and b are the weight matrix and the bias term respectively. Figure 2 Step 6: Refer to

[0118]

[0119] Based on the Adam optimization algorithm, optimize the parameters to be optimized in the feature extraction layer, the tensor fusion layer, and the classification layer in sequence. The following objective function needs to be minimized through the output of the classification layer: where M is the number of input samples, m is the probability that the sample x

[0120] is predicted to be in the c-th health state, and Ι(·) is the indicator function.

[0121] Step 7: Repeat steps 2 to 6 iteratively to optimize the sample imbalance diagnosis model composed of the data generation layer, the feature extraction layer, the tensor fusion layer, and the classification layer;

[0122] Step 8: Input the three-dimensional blade tip clearance sample into the established sample imbalance diagnosis model, and output the corresponding health state through the classification layer.

[0123] Apply the proposed method to the simulation of the intelligent diagnosis experiment for the sample imbalance of turbine blade cracks to further verify the effectiveness of the present invention.

[0124] Table 1 Unbalanced vibration signal sample set

[0125]

[0126] The unbalanced dataset shown in Table 1 was used to construct a diagnostic model to verify the feasibility of the method of the present invention. To quantify the effect of the present invention on the sample imbalance diagnosis task, in addition to using the diagnostic accuracy to measure the diagnostic effect, two unbalanced classification evaluation indices, G-mean and F-score, were also selected to quantify the diagnostic results of different health states. To exclude the interference of random factors, the experiment was repeated 10 times, and the calculated diagnostic results are shown in Table 2. The method of the present invention achieved a diagnostic accuracy of 95.97% on the dataset. At the same time, the G-mean and F-score indices of the method of the present invention reached 0.9533 and 0.9599 respectively, indicating that the method of the present invention has high diagnostic accuracy on the unbalanced sample set and verifying the feasibility of the method of the present invention in solving the problem of sample imbalance.

[0127] Table 2 shows the diagnostic effects of different methods

[0128]

[0129] In addition, three different diagnostic methods were selected to further verify the effectiveness of the present invention. Method 1 was to use CAVE based on the focal loss function (FL) for the diagnosis of unbalanced samples. This method also considered the importance degree between samples. Its diagnostic accuracy on the dataset was only 90.32%, while the G-mean and F-score were 0.8889 and 0.9040 respectively, which were lower than those of the method of the present invention, indicating that the DFL loss function proposed by the present invention is superior to the FL loss function. Method 2 was only based on the traditional CAVE for the diagnosis of unbalanced samples, which did not consider the importance degree between samples. The diagnostic accuracy of this method on the dataset was 80.23%, and the G-mean and F-score were 0.7769 and 0.8042 respectively, which were significantly lower than those of the method of the present invention. Method 3 was the traditional upsampling method to resample the unbalanced data to eliminate the imbalance between various fault samples. The diagnostic accuracy of this method on the dataset was 89.65%, and the G-mean and F-score were 0.7100 and 0.7396 respectively, which were also significantly lower than those of the method of the present invention.

[0130] Figure 4 It is the effect diagram of the deep focal loss function proposed by the present invention. It can be seen that as γ* increases, a larger loss weight can be weighted for the minority fault class samples with small probability predictions, making the weight in the overall loss increase.

[0131] Figure 5It is a clustering graph of three-dimensional features. A training set is composed of generated fault class samples and unbalanced set samples, and after feature learning on the training set, the t-SNE reduction method is used to reduce the dimension of the obtained high-dimensional features to obtain a three-dimensional feature clustering graph. It can be seen that although there is a certain confusion in identifying different health states by the proposed method, the boundaries of different health states are clear.

[0132] Figure 6 It is a distribution graph of the diagnosis results for different health states. It can be seen that the method of the present invention realizes the accurate diagnosis of unbalanced fault samples.

[0133] From the specific processing process of the above diagnosis examples and the experimental results compared with the other three diagnosis methods, it shows that the method of the present invention can effectively overcome the influence of sample imbalance on fault diagnosis and improve the diagnosis accuracy of the diagnosis model.

[0134] As Figure 7 shown, the present invention also provides a blade crack diagnosis system for unbalanced aero-engine sample data, including:

[0135] An acquisition module, used to acquire the three-dimensional tip clearance signal of the blade, establish a deep generation network diagnosis model based on a conditional variational autoencoder network; construct a deep focus loss function based on a small number of fault class sample sets;

[0136] An expansion module, used to combine the deep focus loss function with the deep generation network diagnosis model to obtain a dynamically weighted conditional variational autoencoder network, and expand a small number of fault class samples in the unbalanced class sample set;

[0137] A training module, used to mix a small number of fault class sample sets with the original sample set to form a training set, extract features from the training set using a feature extraction network, and perform feature fusion at a tensor fusion layer, and finally establish a complex non-linear mapping relationship between the fused features and the fault types to complete the training of the sample imbalance diagnosis model;

[0138] A diagnosis module, used to input the three-dimensional tip clearance sample of the blade to be diagnosed into the established sample imbalance diagnosis model to obtain the health state.

[0139] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: modifications or equivalent replacements can still be made to the specific embodiments of the present invention, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for diagnosing blade cracks with unbalanced sample data of an aero-engine, characterized in that, Including: Obtain the three-dimensional tip clearance signal of the blade and establish a deep generative network diagnosis model based on the conditional variational autoencoder network; Construct a deep focal loss function based on a small number of fault class sample sets; Combine the deep focal loss function with the deep generative network diagnosis model to obtain a dynamically weighted conditional variational autoencoder network, and expand the small number of fault class samples in the imbalanced class sample set; Mix the small number of fault class sample sets with the original sample set to form a training set, use the feature extraction network to extract features from the training set, and perform feature fusion in the tensor fusion layer. Finally, establish a complex non-linear mapping relationship between the fused features and the fault types to complete the training of the sample imbalance diagnosis model; Input the three-dimensional tip clearance sample of the blade to be diagnosed into the established sample imbalance diagnosis model to obtain the health status; The deep generative network diagnosis model based on the conditional variational autoencoder network includes a data generation layer, a feature extraction layer, a tensor fusion layer, and a classification layer; specifically, a tensor fusion layer is introduced after the feature extraction layer, and a classification layer is added at the output end to form a deep generative network diagnosis model, and the collected three-dimensional tip clearance signal is used as the input; Construct a deep focal loss function based on a small number of fault class sample sets; including: Construct a deep focus loss function for learning a small number of faulty class samples, which consists of L c and L r in two parts. L c It includes a small number of faulty class samples and a large number of normal class samples, and weights the sample losses L r to penalize the misclassification probability as follows: where M is the number of samples, y i is the label vector of the sample, p i is the predicted probability vector of the sample, is the dynamic focusing parameter, α , β is the adjustable parameter; Mix the small number of fault class sample sets with the original sample set to form a training set, use the feature extraction network to extract features from the training set, and perform feature fusion in the tensor fusion layer. Finally, establish a complex non-linear mapping relationship between the fused features and the fault types to complete the training of the sample imbalance diagnosis model; including: Mix the small number of fault class sample sets with the original sample set to form a training set, and eliminate the imbalanced distribution of samples by generating fault samples; Use the feature extraction layer to extract features from the training sample set to obtain feature information of different dimensions; Use the tensor fusion layer to fuse features of different dimensions, and the obtained fused features contain feature information within different feature dimensions and their mutual correlations; Use the fused features as the input of the classification layer, and continuously optimize the error distance between the output of the classification layer and the blade fault types; Iteratively optimize the sample imbalance diagnosis model composed of the data generation layer, the feature extraction layer, the tensor fusion layer, and the classification layer in sequence.

2. A method for diagnosing blade cracks with unbalanced aero-engine sample data according to claim 1, characterized in that: Obtaining the three-dimensional tip clearance signal of the blade includes: Use the turbine blade fault simulation platform to collect the three-dimensional tip clearance signal of the turbine blade; Collect the three-dimensional tip clearance signals under different degrees of blade crack states to form a sample set , including a total of X healthy states; among them, x m ∈ R N×3 is the m th three-dimensional tip clearance signal sample, which contains three-dimensional parameters , and is composed of N data points, y m ∈ R C×1 represents the health state label vector of the m th sample, C is the number of health states, M is the total number of samples.

3. A method for diagnosing blade cracks with unbalanced sample data of an aeroengine according to claim 1, characterized in that: Constructing the feature extraction layer includes: Mix the generated fault sample set with the imbalanced sample set to form a training sample set and generate fault samples; perform fault feature extraction on the training sample set, and use the extracted features f m as the output of the feature extraction layer. The extraction process is as follows: Wherein: f m is the training sample x m is the feature output after passing through the feature extraction layer, g θ is the transformation function of the feature extraction layer, θ is the set of parameters to be optimized in the feature extraction layer.

4. A method for diagnosing blade cracks with unbalanced sample data of an aero-engine according to claim 1, characterized in that: Constructing the tensor fusion layer includes: Joint modeling of the dynamic characteristics between multi-dimensional features, Fuse the features of different dimensions of the extracted three-dimensional tip clearance, and the fusion process is as follows: In the formula: is the outer product operation of tensors, z is the input tensor, k is the number of tensors, r is the rank of the tensor, w is the weight matrix.

5. A method for diagnosing blade cracks with unbalanced sample data of an aeroengine according to claim 1, characterized in that: Constructing the classification layer includes: Classify the output features of the tensor fusion layer to obtain the sample x m belongs to the c predicted probability of the healthy state is: Wherein, w and b are the weight matrix and the bias term, respectively; Based on the Adam optimization algorithm, optimize the parameters to be optimized in the feature extraction layer, the tensor fusion layer, and the classification layer in sequence. The following objective function needs to be minimized through the output of the classification layer: Wherein: M is the number of input samples, is the sample x m is predicted to be the c probability of the \(i\)-th health state, and \(I(\cdot)\) is the indicator function.

6. An aero-engine sample data unbalanced blade crack diagnosis system, based on the aero-engine sample data unbalanced blade crack diagnosis method according to any one of claims 1 to 5, characterized in that, Including: An acquisition module for obtaining the three-dimensional tip clearance signal of the blade and establishing a deep generative network diagnosis model based on the conditional variational autoencoder network; Construct a deep focal loss function based on a small number of fault class sample sets; An expansion module for combining a depth focus loss function with a depth generation network diagnosis model to obtain a dynamically weighted conditional variational coding network for expanding a small number of fault class samples in an imbalanced class sample set; A training module for mixing a small number of fault class sample sets with the original sample set to form a training set, using a feature extraction network to extract features from the training set, performing feature fusion in a tensor fusion layer, and finally establishing a complex non-linear mapping relationship between the fused features and the fault types to complete the training of the sample imbalance diagnosis model; A diagnosis module for inputting the three-dimensional blade tip clearance sample to be diagnosed into the established sample imbalance diagnosis model to obtain the health status.

Images