Aero-engine gas path system unbalanced data fault diagnosis method based on space-time network

Abstract

The invention discloses an unbalanced data fault diagnosis method for an aero-engine gas path system based on a space-time network. The method comprises the steps of obtaining limited unbalanced data of gas path faults of an aero-engine in different operation states; constructing an STFD model for the unbalanced data fault diagnosis of the aero-engine gas path system based on the space-time network; the system comprises a spatio-temporal feature extraction network, a spatio-temporal feature distribution alignment module and a class balance label distributed sensing marginal module. And performing model training on the STFD model according to the source domain data set and the data training set, determining an optimal STFD model according to the data verification set based on the constructed model total loss function, and inputting the data test set into the optimal STFD model, thereby realizing fault prediction of the unbalanced data of the aero-engine gas path system. The problems that according to an existing unbalanced data fault diagnosis method, the features of the time dimension and the space dimension cannot be effectively extracted, feature information in limited data cannot be fully mined, and the problem of data imbalance cannot be solved in combination with an advanced algorithm are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of aviation engine automatic control, and in particular to a method for diagnosing imbalance data faults of an aviation engine gas path system based on a spatiotemporal network. Background Art

[0002] Aircraft engines are complex aerodynamic and thermodynamic systems. When engines operate for long periods in extremely harsh environments, such as high temperature, high pressure, and intense vibration, structural degradation of components is inevitable, increasing the risk of mechanical failure. As the primary propulsion system for aircraft, performance degradation or sudden failure of aircraft engines can not only cause significant economic losses but also potentially lead to air crashes. Therefore, ensuring their safe and reliable operation is of paramount importance.

[0003] Aircraft engines inevitably experience performance degradation throughout their lifecycle. Because the airflow system is directly related to the engine's compression, combustion, and exhaust processes, its performance degradation has a significant impact on the engine's overall efficiency and safety. Therefore, airflow fault diagnosis technology has become a crucial component in ensuring safe and reliable engine operation. In recent years, research on airflow system fault diagnosis has become a hot topic. Deep learning technology, owing to its ability to efficiently process high-dimensional time series data without requiring human intervention, has garnered widespread attention and has been widely applied to aircraft engine airflow system fault detection. Traditional fault diagnosis algorithms typically assume that training and test data follow the same distribution, achieving good results in fault detection. However, in actual flight, healthy data constitute the majority, while faulty data only accounts for a very small proportion, resulting in a significant imbalance in the collected data. This data imbalance makes traditional fault diagnosis algorithms difficult to apply in practical applications.

[0004] Unbalanced data distribution significantly impacts the attention mechanism of deep learning models. Due to the large number of samples in certain categories, the model classifier tends to favor these categories, resulting in insufficient recognition of the minority category. In this case, although the overall accuracy appears high, the actual performance of the classifier may be suboptimal, or even ineffective when dealing with imbalanced categories. While existing research on imbalanced data has made some progress, these studies have overlooked the coexistence of feature shift and label shift that often accompany data imbalance. This phenomenon hinders the extraction of effective features and reduces the model's fault diagnosis performance. In fact, under multiple operating conditions, the shift phenomenon may be exacerbated, placing higher demands on the model's adaptability and generalization capabilities. How to effectively adjust this shift to ensure that the model maintains good performance under different operating conditions has become a major challenge. Furthermore, finding the minimum decision boundary when addressing feature shift and label shift is also a difficult problem.

[0005] Deep learning typically relies on large amounts of data for training, enabling effective classification and differentiation by discovering inherent connections and differences within the data. However, in real-world applications, machine failures are often rare and unpredictable. Furthermore, collecting failure data is often challenging. Insufficient data can prevent the model from fully learning features that require good generalization, potentially leading to underfitting or overfitting, resulting in suboptimal model performance.

[0006] Current research on air path system fault diagnosis often assumes sufficient data. However, due to the structural complexity and high operating costs of air path systems, it is difficult to obtain sufficient flight fault data. Therefore, the main challenge is to fully extract effective features from data with limited and unbalanced data. In summary, there is an urgent need to provide an unbalanced data fault diagnosis method for aircraft engine air path systems based on spatiotemporal networks. This method can extract features in both time and space, fully exploit the feature information in the limited data, and combine advanced algorithms to solve the data imbalance problem, thereby solving the problems existing in existing research. Summary of the Invention

[0007] The present invention provides an aero-engine gas path system imbalance data fault diagnosis method based on a spatiotemporal network to overcome the above technical problems.

[0008] In order to achieve the above object, the technical solution of the present invention is:

[0009] A method for diagnosing imbalance data faults in an aircraft engine gas path system based on a spatiotemporal network specifically comprises the following steps:

[0010] S1: Obtain limited imbalance data of gas path faults of aircraft engines under different operating conditions, which includes source domain datasets and target domain datasets;

[0011] Without changing the original distribution of the finite imbalanced data, the target domain dataset is randomly divided into a data training set, a data validation set, and a data test set;

[0012] S2: Construct an STFD model for fault diagnosis of imbalanced data in aero-engine gas path systems based on a spatiotemporal network. The STFD model includes a spatiotemporal feature extraction network, a spatiotemporal feature distribution alignment module, and a class balance label distributed perception margin module.

[0013] The spatiotemporal feature extraction network is used to extract spatiotemporal features from the limited imbalanced data in the source domain dataset and the target domain dataset respectively to obtain spatiotemporal time series data features;

[0014] The spatiotemporal feature distribution alignment module is used to perform cross-domain spatiotemporal feature alignment on the source domain dataset and the target domain dataset based on the spatiotemporal time series data characteristics to obtain spatiotemporal alignment time series features;

[0015] The distributed perception edge module of class balance labels is used to achieve fault prediction of imbalance data of aircraft engine gas path system based on spatiotemporal alignment of time series features;

[0016] S3: Based on the source domain dataset and the data training set, the STFD model is trained to obtain the trained STFD model:

[0017] S4: Based on the constructed model total loss function, determine whether the output of the trained STFD model converges according to the data validation set; if convergence is confirmed, the trained STFD model at this time is used as the optimal STFD model; otherwise, the parameter weights of the trained STFD model are adaptively updated based on the backpropagation method, and S3 is repeated;

[0018] S5: Input the data test set into the optimal STFD model to achieve fault prediction of imbalance data of the aircraft engine gas path system.

[0019] Furthermore, the spatiotemporal feature extraction network constructed in S2 includes a cascaded first-scale feature extraction network, a second-scale feature extraction network, a splicing layer, and a fully connected output layer;

[0020] The first scale feature extraction network includes a first spatial feature extraction branch and a first temporal feature extraction branch; the first spatial feature extraction branch includes a two-dimensional convolutional neural network, a first pooling layer, and a first fully connected layer connected in sequence; the two-dimensional convolutional neural network is used to perform a 2D convolution operation on a two-dimensional tiled data map of preset finite imbalanced data to obtain a first spatial feature map; the first pooling layer is used to perform a pooling operation on the first spatial feature map to obtain a second spatial feature map; the first fully connected layer is used to perform a fully connected operation on the second spatial feature map to obtain a third spatial feature map;

[0021] The first time feature extraction branch is specifically a recurrent neural network (RNN), which is used to extract time features from a preset two-dimensional tiled data graph of finite imbalanced data to obtain a first time feature graph. A recurrent neural network (RNN) is a type of neural network used to process sequence data. When processing sequence data, the recurrent neural network captures the time information in the sequence data through a loop structure, thereby better modeling. The basic idea of ​​the recurrent neural network (RNN) is to introduce a loop structure into the network so that the network can process sequence data of any length.

[0022] The second-scale feature extraction network includes a second spatial feature extraction branch and a second temporal feature extraction branch; the second spatial feature extraction branch includes a one-dimensional convolutional neural network, a normalization layer, a second pooling layer, a Dropout layer, and a second fully connected layer connected in sequence; the one-dimensional convolutional neural network is used to perform a 1D convolution operation on the third spatial feature map to obtain a fourth spatial feature map; the normalization layer is used to perform a normalization operation on the fourth spatial feature map to obtain a fifth spatial feature map; the second pooling layer is used to perform a pooling operation on the fifth spatial feature map to obtain a sixth spatial feature map; the Dropout layer is used to perform a random drop operation on the sixth spatial feature map according to a preset drop probability to obtain a seventh spatial feature map; the second fully connected layer is used to perform a fully connected operation on the seventh spatial feature map to obtain an eighth spatial feature map;

[0023] The second time feature extraction branch includes a one-dimensional ResNet18 network, a third pooling layer, and a third fully connected layer connected in sequence; the one-dimensional ResNet18 network is used to extract time features from the first time feature map to obtain a second time feature map; the third pooling layer is used to perform a pooling operation on the second time feature map to obtain a third time feature map; the third fully connected layer is used to perform a fully connected operation on the third time feature map to obtain a fourth time feature map;

[0024] The splicing layer is used to splice the fourth temporal feature map and the eighth spatial feature map to obtain a spliced ​​feature map;

[0025] The fully connected output layer is used to perform full connection operations on the concatenated feature maps to output spatiotemporal time series data features.

[0026] Furthermore, the spatiotemporal feature distribution alignment module in S2 is used to perform cross-domain spatiotemporal feature alignment of the source domain dataset and the target domain dataset based on the spatiotemporal time series data features, including the following steps:

[0027] S100: Constructing a spatiotemporal feature difference value function for cross-domain spatiotemporal feature alignment between a source domain dataset and a target domain dataset;

[0028] And the expression of the spatiotemporal feature difference value function is

[0029]

[0030] Where: d LMMD Represents the difference between the spatiotemporal features in the source domain dataset and the spatiotemporal features in the target domain dataset; and Representation sample With sample The probability of belonging to the air path fault category c of the aircraft engine and satisfying and D SWith D t represent the source domain dataset and the target domain dataset respectively; and The kernel function representing the spatiotemporal feature map in the source domain dataset; The kernel function representing the spatiotemporal feature map in the target domain dataset; The kernel function representing the spatiotemporal feature mapping between the source and target domain datasets; represents the value obtained after the spatiotemporal features in the source domain dataset or the target domain dataset pass through the activation function; C represents the number of fault categories; n s Represents the total number of samples in the source domain dataset; n t Represents the total number of samples in the target domain data used as the data training set;

[0031] S101: Based on the spatiotemporal feature difference value function, a cross-domain spatiotemporal feature alignment loss function is constructed. The cross-domain spatiotemporal feature alignment loss of the source domain dataset and the target domain dataset is optimized during the STFD model training process through the cross-domain spatiotemporal feature alignment loss function, thereby obtaining spatiotemporal alignment time series features.

[0032] and

[0033] Where: L 时空特征对齐 Loss function representing cross-domain spatiotemporal feature alignment; Represents the spatiotemporal feature difference value of the ξth cross-domain spatiotemporal feature alignment; Indicates that the loss weight will be dynamically adjusted according to the difference in spatiotemporal features during model training. Represents the temporal feature difference value between the first temporal feature graph of the source domain dataset extracted based on the spatiotemporal feature extraction network and the first temporal feature graph of the target domain dataset; represents a temporal feature difference value between a fourth temporal feature map of a source domain dataset extracted based on a spatiotemporal feature extraction network and a fourth temporal feature map of a target domain dataset; Represents the spatial feature difference value between the third spatial feature map of the source domain dataset extracted based on the spatiotemporal feature extraction network and the third spatial feature map of the target domain dataset; Represents the spatial feature difference value between the eighth spatial feature map of the source domain dataset extracted based on the spatiotemporal feature extraction network and the eighth spatial feature map of the target domain dataset.

[0034] Furthermore, the class balance label distributed perception edge module described in S2 is used to implement fault prediction of imbalance data of the aircraft engine gas path system based on the spatiotemporal alignment time series features, specifically including the following steps:

[0035] S31: The spatiotemporal time series data features output by the spatiotemporal feature extraction network are recorded as x i , then the labeled examples of the limited imbalanced data are denoted as D c and D c =(x i ,y i ), where y i represents the fault class label of the labeled example and y i ∈{1,...C};

[0036] And construct the decision boundary minimum distance function from the fault category sample to the fault category label, and use the decision boundary minimum distance function as the optimal inter-class interval model for each fault category;

[0037] And the expression of the optimal inter-class interval model is

[0038]

[0039] Where: γ c represents the minimum distance of the decision boundary; γ(x i ,y i ) represents the distance function between the labeled example and the label edge; K represents the design parameter; n c Represents the total number of samples in the data training set;

[0040] S32: Define the output logarithm Z of the spatiotemporal feature extraction network = [z1,…,z C ] and z i =log e x i , and based on the optimal inter-class interval model, define the label distribution aware marginal loss function;

[0041] And the label distribution aware marginal loss function L LDAM-HG The expression is

[0042]

[0043] Where: z c represents the output logarithm of the c-th spatiotemporal feature extraction network and c∈{1,...,C}; Represents the fault category label y i The output logarithm of the corresponding spatiotemporal feature extraction network;

[0044] Label distribution aware marginal loss function L LDAM-HG Perform smoothing to obtain a smooth loss function with label distribution-aware margins;

[0045] And the smooth loss function L LDAM The expression is

[0046]

[0047] S33: Based on the smooth loss function L LDAM , construct the class balance label distributed perception marginal module; and based on the class balance label distributed perception marginal module, realize the fault prediction of the imbalance data of the aircraft engine gas path system; and the class balance label distributed perception marginal module L CBDM The expression is

[0048]

[0049] Where: Where: β m Indicates the optimal parameters of the STFD model found by the grid search method during each model training process and β m ∈[0,1]; n c Represents the total number of samples in the data training set; represents the optimal model parameters corresponding to the data training set; z i represents the output logarithm of the spatiotemporal feature extraction network of the i-th sample; represents the spatiotemporal feature extraction network output value of the cth sample; Represents the spatiotemporal feature extraction network output value of the i-th sample.

[0050] Furthermore, the total loss function of the model constructed in S4 is formulated as

[0051]

[0052] Where: Loss all Represents the total loss function of the model.

[0053] Beneficial effects: The present invention provides a method for diagnosing imbalanced data in an aero-engine gas path system based on a spatio-temporal network, and constructs an STFD model for diagnosing imbalanced data in an aero-engine gas path system based on a spatio-temporal network. Specifically, a spatio-temporal feature extraction network is used to extract spatio-temporal features from limited imbalanced data to obtain spatio-temporal time series data features. The present invention maximizes the potential information of the data by simultaneously capturing temporal and spatial features. The network aims to extract multi-level spatio-temporal features from limited data to obtain more fault-related information. A spatio-temporal feature distribution alignment module is used to perform cross-domain spatio-temporal feature alignment of the source domain dataset and the target domain dataset based on the spatio-temporal time series data features. By performing dual alignment in time and space, the alignment at the fine-grained level of the model is more detailed, effectively solving the offset problem caused by imbalanced data, and improving classification accuracy while reducing the difficulty of subsequent classification tasks. A class-balanced label distributed perception margin module is used to implement fault prediction of imbalanced data in an aero-engine gas path system based on the spatio-temporal alignment time series features. Model parameters are updated and optimized during model training based on the constructed model total loss function, greatly improving fault diagnosis accuracy while ensuring flight safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0055] Figure 1 This is a flow chart of a method for diagnosing imbalanced data in an aero-engine gas path system based on a spatiotemporal network according to the present invention;

[0056] Figure 2 It is a technical circuit diagram of the method described in this embodiment;

[0057] Figure 3 This is a specific flow chart of the time-space network in this embodiment;

[0058] Figure 4 Schematic diagram of distribution alignment in this embodiment;

[0059] Figure 5 This is a comparison diagram of the confusion matrix of the fault diagnosis results in this embodiment;

[0060] Figure 6 This is a radar comparison diagram of the fault diagnosis results in this embodiment;

[0061] Figure 7 This is a T-SNE comparison chart of the experimental results in this embodiment;

[0062] Figure 8 2 is a comparison chart of the ablation experiment results in this embodiment. DETAILED DESCRIPTION

[0063] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0064] This embodiment provides a method for diagnosing imbalanced data faults in an aircraft engine gas path system based on a spatiotemporal network. Figures 1 to 2 As shown, the specific steps include:

[0065] S1: Using preset sensors to obtain limited imbalance data of gas path faults of aircraft engines under different operating conditions, which includes a source domain dataset and a target domain dataset;

[0066] Without changing the original distribution of the limited imbalanced data, the target domain dataset is randomly divided into a data training set, a data validation set, and a data test set. In this embodiment, the AGPS data collected by the sensor contains not only temporal features but also spatial features. If only the spatial features at a specific time point are focused on, the overall behavior of the system cannot be fully reflected. Similarly, if only the time series features of a certain spatial location are analyzed, it will also be difficult to capture the complete system dynamics. In fact, the spatial behavior at a certain moment will have a significant impact on the behavior of the adjacent spatial locations at the next moment. Therefore, in this embodiment, a spatiotemporal feature extraction network is designed to maximize the potential information of the data by simultaneously capturing temporal and spatial features. The network aims to extract multi-level spatiotemporal features from limited data and obtain more fault-related information.

[0067] S2: Construct an STFD model for fault diagnosis of imbalanced data in aero-engine gas path systems based on a spatiotemporal network. The STFD model includes a spatiotemporal feature extraction network, a spatiotemporal feature distribution alignment module, and a class balance label distributed perception margin module.

[0068] The spatiotemporal feature extraction network is used to extract spatiotemporal features from the limited imbalanced data in the source domain dataset and the target domain dataset respectively to obtain spatiotemporal time series data features;

[0069] In a specific embodiment, Figure 3 As shown, the constructed spatiotemporal feature extraction network includes a cascaded first-scale feature extraction network, a second-scale feature extraction network, a splicing layer, and a fully connected output layer;

[0070] The first scale feature extraction network includes a first spatial feature extraction branch and a first temporal feature extraction branch; the first spatial feature extraction branch includes a two-dimensional convolutional neural network, a first pooling layer, and a first fully connected layer connected in sequence; the two-dimensional convolutional neural network is used to perform a 2D convolution operation on a two-dimensional tiled data map of preset finite imbalanced data to obtain a first spatial feature map; the first pooling layer is used to perform a pooling operation on the first spatial feature map to obtain a second spatial feature map; the first fully connected layer is used to perform a fully connected operation on the second spatial feature map to obtain a third spatial feature map;

[0071] The first time feature extraction branch is specifically a recurrent neural network, which is used to extract time features from a preset two-dimensional tiled data graph of finite imbalanced data to obtain a first time feature graph;

[0072] Specifically, in this embodiment, a simple recurrent neural network is used to perform preliminary extraction of temporal features, obtain shallow temporal features, and pass the extracted shallow temporal features to the fully connected block for processing; then, the output of the fully connected block is passed to the second temporal feature extraction branch and used as the input of the one-dimensional ResNet18 network, that is, the temporal features are further extracted through the preset four residual block structures;

[0073] The calculation formula of the residual block is:

[0074] y=F(x,W i )+x

[0075] Where: x represents the input feature; F represents the residual function; W i Represents the convolution kernel parameters; y represents the temporal features of the output; then, through the global pooling layer and the fully connected layer, while retaining the global information, the risk of overfitting is effectively reduced, and finally the shallow temporal features, namely the fourth temporal feature map, are obtained;

[0076] The second-scale feature extraction network includes a second spatial feature extraction branch and a second temporal feature extraction branch; the second spatial feature extraction branch includes a one-dimensional convolutional neural network, a normalization layer, a second pooling layer, a Dropout layer, and a second fully connected layer connected in sequence; the one-dimensional convolutional neural network is used to perform a 1D convolution operation on the third spatial feature map to obtain a fourth spatial feature map; the normalization layer is used to perform a normalization operation on the fourth spatial feature map to obtain a fifth spatial feature map; the second pooling layer is used to perform a pooling operation on the fifth spatial feature map to obtain a sixth spatial feature map; the Dropout layer is used to perform a random drop operation on the sixth spatial feature map according to a preset drop probability to obtain a seventh spatial feature map; the second fully connected layer is used to perform a fully connected operation on the seventh spatial feature map to obtain an eighth spatial feature map;

[0077] The second time feature extraction branch includes a one-dimensional ResNet18 network, a third pooling layer, and a third fully connected layer connected in sequence; the one-dimensional ResNet18 network is used to extract time features from the first time feature map to obtain a second time feature map; the third pooling layer is used to perform a pooling operation on the second time feature map to obtain a third time feature map; the third fully connected layer is used to perform a fully connected operation on the third time feature map to obtain a fourth time feature map;

[0078] The splicing layer is used to splice the fourth temporal feature map and the eighth spatial feature map to obtain a spliced ​​feature map;

[0079] The fully connected output layer is used to perform a fully connected operation on the spliced ​​feature map to output the spatiotemporal time series data features; wherein, the extraction of spatial features in this embodiment includes: first, extracting spatial features through three 2D convolutional layers, and then processing through the ReLU activation function and the maximum pooling layer. This step can capture nonlinear spatial feature associations and enhance the model fitting ability. The result is output through the fully connected layer and then input into three 1D convolutional layers. After batch normalization and global average pooling processing, the aggregation of effective spatial feature information is achieved. At the same time, the generalization ability of the model is improved by adding the Dropout layer. Finally, the second fully connected layer outputs the result, which is the eighth spatial feature map;

[0080] Specifically, the two-dimensional convolutional neural network model:

[0081]

[0082] Where: f represents input data; g represents convolution kernel; (i, j) represents output features;

[0083] ReLU activation function: R(X) = max(0,X);

[0084] Among them, for any input value X, when X is greater than 0, the output is X itself; when X is less than or equal to 0, the output is 0, and finally the pooling operation is added, that is, the maximum pooling layer processing;

[0085] One-dimensional convolutional neural network model: X out =X in *W+b

[0086] Where: X in Represents the input matrix; W represents the convolution kernel; b represents the bias term; * represents the convolution symbol;

[0087] Batch normalization is:

[0088] Where: X represents input; μ, σ 2Respectively represent the mean and variance of the data in the batch; γ, β represent learnable parameters; ε represents a very small number to prevent the denominator from being zero;

[0089] Finally, the pooling operation is added, that is, the pooling operation is performed through the second pooling layer. A dropout layer with a dropout rate of 50% is also added to the network;

[0090] Dropout layer: Y = M⊙X;

[0091] Where: M represents a binary (0 or 1) matrix of the same shape as X; and the elements in M ​​are randomly selected to be 0 or 1 according to the preset drop probability;

[0092] Finally, the outputs of the temporal feature extraction network and the spatial feature extraction network are fused through the concatenate function. The fused features are further input into the fully connected layer to generate the output result.

[0093] Concatenate function of the concatenation layer (axis=1):

[0094] C0=Concatenate(A,B)

[0095] Where: C0 represents the spatiotemporal time series data feature; A represents the fourth time feature map; B represents the eighth spatial feature map;

[0096] The spatiotemporal feature distribution alignment module is used to perform cross-domain spatiotemporal feature alignment on the source domain dataset and the target domain dataset based on the spatiotemporal time series data characteristics to obtain spatiotemporal alignment time series features;

[0097] In a specific embodiment, the spatiotemporal feature distribution alignment module is used to perform cross-domain spatiotemporal feature alignment on a source domain dataset and a target domain dataset based on spatiotemporal time series data features, including the following steps:

[0098] S100: Constructing a spatiotemporal feature difference value function for cross-domain spatiotemporal feature alignment of the source domain dataset and the target domain dataset, specifically including:

[0099] S1001: Obtain the difference between the source domain dataset and the target domain dataset, which is expressed as

[0100]

[0101] Where: Indicates the category difference between the source domain dataset and the target domain dataset; and Represent the feature maps of the source domain dataset and the target domain dataset samples respectively;

[0102] For a sample x i , with probability vector And calculate the weight for:

[0103]

[0104] Where: Indicates p i The probability vector of category C in ; Note that for the labeled sample set D s The sample in p i Can be regarded as a one-hot vector, and for the unlabeled sample set D t For samples in , the predicted distribution of the network output is used as the pseudo label;

[0105] S1002: Based on step S1001, a spatiotemporal feature difference value function is constructed, and its expression is:

[0106]

[0107] Where: d LMMD Represents the difference between the spatiotemporal features in the source domain dataset and the spatiotemporal features in the target domain dataset; and Representation sample With sample The probability of belonging to the air path fault category c of the aircraft engine and satisfying and D s With D t represent the source domain dataset and the target domain dataset respectively; and The kernel function representing the spatiotemporal feature map in the source domain dataset; The kernel function representing the spatiotemporal feature map in the target domain dataset; The kernel function representing the spatiotemporal feature mapping between the source and target domain datasets; represents the value obtained after the spatiotemporal features in the source domain dataset or the target domain dataset pass through the activation function; C represents the number of fault categories; n s Represents the total number of samples in the source domain dataset; n t Represents the total number of samples in the target domain data used as the data training set;

[0108] S101: If Figure 4 As shown in the figure, a cross-domain spatiotemporal feature alignment loss function is constructed based on the spatiotemporal feature difference value function. The cross-domain spatiotemporal feature alignment loss of the source domain dataset and the target domain dataset is optimized during the STFD model training process through the cross-domain spatiotemporal feature alignment loss function, thereby obtaining the spatiotemporal alignment time series features.

[0109] and

[0110] Where: L 时空特征对齐 Loss function representing cross-domain spatiotemporal feature alignment; Represents the spatiotemporal feature difference value of the ξth cross-domain spatiotemporal feature alignment; Indicates that the loss weight will be dynamically adjusted according to the difference in spatiotemporal features during model training. Represents the temporal feature difference value between the first temporal feature graph of the source domain dataset extracted based on the spatiotemporal feature extraction network and the first temporal feature graph of the target domain dataset; represents a temporal feature difference value between a fourth temporal feature map of a source domain dataset extracted based on a spatiotemporal feature extraction network and a fourth temporal feature map of a target domain dataset; Represents the spatial feature difference value between the third spatial feature map of the source domain dataset extracted based on the spatiotemporal feature extraction network and the third spatial feature map of the target domain dataset; represents the spatial feature difference value between the eighth spatial feature map of the source domain dataset extracted based on the spatiotemporal feature extraction network and the eighth spatial feature map of the target domain dataset;

[0111] In this embodiment, after the source domain dataset and the target domain dataset pass through the spatiotemporal feature extraction network, the cross-domain spatiotemporal features of the shallow and deep networks are aligned through LMMD; for the temporal feature alignment, the losses generated by the shallow and deep networks are respectively recorded as LMMD1, that is, and LMMD2 In terms of alignment in the spatial dimension, the losses generated by the shallow and deep networks are respectively recorded as LMMD3, that is, and LMMD4 This embodiment uses dual alignment in time and space to achieve more detailed alignment at the fine-grained model level, effectively resolving the offset problem caused by unbalanced data. This alignment method reduces the difficulty of subsequent classification tasks while also improving classification accuracy.

[0112] The distributed perception edge module of class balance labels is used to achieve fault prediction of imbalance data of aircraft engine gas path system based on spatiotemporal alignment of time series features;

[0113] In a specific embodiment, the method for constructing the class-balanced label distributed perception margin module specifically includes the following steps:

[0114] S31: The spatiotemporal time series data features output by the spatiotemporal feature extraction network are recorded as x i , then the labeled examples of the limited imbalanced data are denoted as D c and D c =(x i ,y i), where y i represents the fault class label of the labeled example and y i ∈1,…C}; define the output logarithm Z of the spatiotemporal feature extraction network = [z1,…,z C ] and z i =log e x i , then p is the predicted probability vector of class C

[0115] In this embodiment, the weight of each fault category c is Defined as:

[0116]

[0117] Where: β∈[0,1] is a preset hyperparameter related to the entire dataset. For the so-called class balance, this embodiment defines a reweighting factor α c :

[0118]

[0119] This embodiment maintains the weight α c It can adapt to the consistency of different data sets and rewrite it as:

[0120]

[0121] Where: Represents the weight of the i-th sample, which is due to α c The β in the equation is a hyperparameter that needs to be manually adjusted. In this embodiment, the grid search method is used to determine the β value as β in order to improve the accuracy of the model. m ;

[0122] This embodiment constructs a decision boundary minimum distance function from fault category samples to fault category labels, and uses the decision boundary minimum distance function as the optimal inter-class interval model for each fault category;

[0123] And the expression of the optimal inter-class interval model is

[0124]

[0125] Where: γ c Represents the minimum distance of the decision boundary, and thus the minimum boundary γ of all fault categories can be obtained min =min{γ1,...,γ C};γ(x i ,y i ) represents the distance function between the labeled example and the label edge; K represents the design parameter; n cRepresents the total number of samples in the data training set;

[0126] S32: Based on the optimal inter-class margin model, define the label distribution aware marginal loss function;

[0127] And the label distribution aware marginal loss function L LDAM-HG The expression is

[0128]

[0129] Where: z c represents the output logarithm of the c-th spatiotemporal feature extraction network and c∈{1,...,C}; Represents the fault category label y i The output logarithm of the corresponding spatiotemporal feature extraction network;

[0130] Since the loss in this embodiment is non-smooth and difficult to optimize, the label distribution aware marginal loss function L LDAM-HG Perform smoothing to obtain a smooth loss function with label distribution-aware margins;

[0131] And the smooth loss function L LDAM The expression is

[0132]

[0133] S33: In order to improve the generalization ability of minority classes and reduce boundary vulnerability, this embodiment proposes a class-balanced label distributed perception margin (CBDM) method that combines reweighting and regularization, namely the class-balanced label distributed perception margin module; this embodiment is based on the smooth loss function L LDAM Construct a class balance label distributed perception margin module; and based on the class balance label distributed perception margin module, realize the fault prediction of the imbalance data of the aircraft engine gas path system; in CBDM, each class of samples has its own reweighted weight factor for adjusting the boundary, which can improve the vulnerability problem of the boundary, among which the class balance label distributed perception margin module L CBDM The expression is

[0134]

[0135] Where: Where: β m Indicates the optimal parameters of the STFD model found by the grid search method during each model training process and β m ∈[0,1]; n c Represents the total number of samples in the data training set; represents the optimal model parameters corresponding to the data training set; z i represents the output logarithm of the spatiotemporal feature extraction network of the i-th sample; represents the spatiotemporal feature extraction network output value of the cth sample; Represents the spatiotemporal feature extraction network output value of the i-th sample;

[0136] S3: Based on the source domain dataset and the data training set, the STFD model is trained to obtain the trained STFD model:

[0137] S4: Based on the constructed model total loss function, determine whether the output of the trained STFD model converges according to the data validation set; if convergence is confirmed, the trained STFD model at this time is used as the optimal STFD model; otherwise, the parameter weights of the trained STFD model are adaptively updated based on the backpropagation method, and S3 is repeated;

[0138] Specifically, the total loss function of the constructed model is expressed as

[0139]

[0140] Where: Loss all Represents the total loss function of the model; The initial value is set to 0.25, The value is dynamically adjusted according to the loss value during training; the inverse of each loss value in LMMD is taken to calculate its reverse loss value, and the reversed loss value is normalized to ensure that the sum of the weights is 1, thus obtaining the adjusted value; current The value and the adjusted value are smoothed by weighted average (each is given a weight of 0.5), and finally the adjusted value is normalized again. value to ensure that the sum is 1, thus determining the final value;

[0141] S5: Input the data test set into the optimal STFD model to achieve fault prediction of imbalance data of the aircraft engine gas path system.

[0142] This embodiment also includes a simulation verification study of the aircraft engine gas path system imbalance data fault diagnosis method based on the spatiotemporal network to verify its effectiveness:

[0143] The JT9D engine was selected as the experimental object. The failure mode of this engine is caused by the performance degradation of five important components, including the fan, low-pressure compressor (LPC), high-pressure compressor (HPC), combustor, high-pressure turbine (HPT), and low-pressure turbine (LPT). To ensure variable operating conditions, this embodiment divides the engine operating state into four stages: low speed to acceleration stage (S1), acceleration to high speed stage (S2), high speed to deceleration stage (S3), and deceleration to low speed stage (S4). Eight sensors are used to monitor the engine operating state with a single engine working cycle as the node. The sensor sampling frequency is set to 50Hz, and relevant data is collected to form the experimental data set.

[0144] In the source domain dataset, there are 240 samples in the healthy state and 36 samples in each fault state, resulting in a significant imbalance of 6.67:1. In contrast, the target domain dataset contains 400 samples per category. The data dimensions of both the source and target domains are [1, 8, 500], meeting the requirements of this problem. During the experiment, the target domain dataset was partitioned according to a preset ratio while maintaining the original distribution to obtain the training, validation, and test sets.

[0145] This embodiment uses existing software to perform computer simulations. Compared with the unbalanced data algorithm in existing research, the results are as follows: Figures 5 to 8 As shown. Figures 5 to 8 It can be seen that the established fault diagnosis model improves the influence of the limited unbalanced aircraft engine gas path system data under variable operating conditions on the fault diagnosis accuracy; therefore, this embodiment can conclude that the proposed aircraft engine gas path system imbalance data fault diagnosis method based on spatiotemporal network is effective.

[0146] To summarize, the method described in this embodiment: first, a time-space feature extraction network and multi-domain adaptive technology are used to perform feature extraction and distribution alignment. At the same time, in order to solve the problem of class imbalance, a class-balanced label distributed perception margin method is adopted to enhance the generalization ability of the minority class. Then, the model parameters of the STFD model are updated and optimized through the constructed total loss function; finally, the effectiveness of the method is verified on the JT9D aircraft engine data. Therefore, the method described in this embodiment fully mines the feature information in the limited data by extracting features in the two dimensions of time and space, and then combines advanced algorithms to solve the data imbalance problem, thereby achieving more efficient fault diagnosis and predictive maintenance of aircraft engines.

[0147] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for diagnosing imbalanced data in an aircraft engine gas path system based on a spatiotemporal network, characterized in that: The specific steps include: S1: Obtain limited imbalance data of gas path faults of aircraft engines under different operating conditions, which includes source domain datasets and target domain datasets; Without changing the original distribution of the finite imbalanced data, the target domain dataset is randomly divided into a data training set, a data validation set, and a data test set; S2: Construct an STFD model for fault diagnosis of imbalanced data in aero-engine gas path systems based on a spatiotemporal network. The STFD model includes a spatiotemporal feature extraction network, a spatiotemporal feature distribution alignment module, and a class balance label distributed perception margin module. The spatiotemporal feature extraction network is used to extract spatiotemporal features from the limited imbalanced data in the source domain dataset and the target domain dataset respectively to obtain spatiotemporal time series data features; The spatiotemporal feature distribution alignment module is used to perform cross-domain spatiotemporal feature alignment on the source domain dataset and the target domain dataset based on the spatiotemporal time series data characteristics to obtain spatiotemporal alignment time series features; The distributed perception edge module of class balance labels is used to achieve fault prediction of imbalance data of aircraft engine gas path system based on spatiotemporal alignment of time series features; S3: Based on the source domain dataset and the data training set, the STFD model is trained to obtain the trained STFD model: S4: Based on the constructed model total loss function, determine whether the output of the trained STFD model converges according to the data validation set; If convergence is confirmed, the trained STFD model is used as the optimal STFD model. Otherwise, the parameter weights of the trained STFD model are adaptively updated based on the back propagation method, and S3 is repeated. S5: Input the data test set into the optimal STFD model to achieve fault prediction of imbalance data of the aircraft engine gas path system.

2. The method for diagnosing imbalanced data of an aircraft engine gas path system based on a spatiotemporal network according to claim 1, characterized in that: The spatiotemporal feature extraction network constructed in S2 includes a cascaded first-scale feature extraction network, a second-scale feature extraction network, a splicing layer, and a fully connected output layer; The first scale feature extraction network includes a first spatial feature extraction branch and a first temporal feature extraction branch; the first spatial feature extraction branch includes a two-dimensional convolutional neural network, a first pooling layer, and a first fully connected layer connected in sequence; the two-dimensional convolutional neural network is used to perform a 2D convolution operation on a two-dimensional tiled data map of preset finite imbalanced data to obtain a first spatial feature map; the first pooling layer is used to perform a pooling operation on the first spatial feature map to obtain a second spatial feature map; the first fully connected layer is used to perform a fully connected operation on the second spatial feature map to obtain a third spatial feature map; The first time feature extraction branch is specifically a recurrent neural network, which is used to extract time features from a preset two-dimensional tiled data graph of finite imbalanced data to obtain a first time feature graph; The second scale feature extraction network includes a second spatial feature extraction branch and a second temporal feature extraction branch; The second spatial feature extraction branch includes a one-dimensional convolutional neural network, a normalization layer, a second pooling layer, a Dropout layer, and a second fully connected layer connected in sequence; the one-dimensional convolutional neural network is used to perform a 1D convolution operation on the third spatial feature map to obtain a fourth spatial feature map; the normalization layer is used to perform a normalization operation on the fourth spatial feature map to obtain a fifth spatial feature map; the second pooling layer is used to perform a pooling operation on the fifth spatial feature map to obtain a sixth spatial feature map; the Dropout layer is used to perform a random drop operation on the sixth spatial feature map according to a preset drop probability to obtain a seventh spatial feature map; the second fully connected layer is used to perform a fully connected operation on the seventh spatial feature map to obtain an eighth spatial feature map; The second time feature extraction branch includes a one-dimensional ResNet18 network, a third pooling layer, and a third fully connected layer connected in sequence; the one-dimensional ResNet18 network is used to extract time features from the first time feature map to obtain a second time feature map; the third pooling layer is used to perform a pooling operation on the second time feature map to obtain a third time feature map; the third fully connected layer is used to perform a fully connected operation on the third time feature map to obtain a fourth time feature map; The splicing layer is used to splice the fourth temporal feature map and the eighth spatial feature map to obtain a spliced ​​feature map; The fully connected output layer is used to perform full connection operations on the concatenated feature maps to output spatiotemporal time series data features.

3. The method for diagnosing imbalanced data in an aircraft engine gas path system based on a spatiotemporal network according to claim 2, characterized in that: The spatiotemporal feature distribution alignment module in S2 is used to perform cross-domain spatiotemporal feature alignment of the source domain dataset and the target domain dataset based on the spatiotemporal time series data features, including the following steps: S100: Constructing a spatiotemporal feature difference value function for cross-domain spatiotemporal feature alignment between a source domain dataset and a target domain dataset; And the expression of the spatiotemporal feature difference value function is Where: d LMMD Represents the difference between the spatiotemporal features in the source domain dataset and the spatiotemporal features in the target domain dataset; and Representation sample With sample The probability of belonging to the air path fault category c of the aircraft engine and satisfying and D s With D t represent the source domain dataset and the target domain dataset respectively; and The kernel function representing the spatiotemporal feature map in the source domain dataset; The kernel function representing the spatiotemporal feature map in the target domain dataset; The kernel function representing the spatiotemporal feature mapping between the source and target domain datasets; represents the value obtained after the spatiotemporal features in the source domain dataset or the target domain dataset pass through the activation function; C represents the number of fault categories; n s Represents the total number of samples in the source domain dataset; n t Represents the total number of samples in the target domain data used as the data training set; S101: Based on the spatiotemporal feature difference value function, a cross-domain spatiotemporal feature alignment loss function is constructed, and the cross-domain spatiotemporal feature alignment loss of the source domain dataset and the target domain dataset is optimized during the STFD model training process through the cross-domain spatiotemporal feature alignment loss function, thereby obtaining spatiotemporal alignment time series features; and Where: L 时空特征对齐 Loss function representing cross-domain spatiotemporal feature alignment; Represents the spatiotemporal feature difference value of the ξth cross-domain spatiotemporal feature alignment; Indicates that the loss weight will be dynamically adjusted according to the difference in spatiotemporal features during model training. Represents the temporal feature difference value between the first temporal feature graph of the source domain dataset extracted based on the spatiotemporal feature extraction network and the first temporal feature graph of the target domain dataset; represents a temporal feature difference value between a fourth temporal feature map of a source domain dataset extracted based on a spatiotemporal feature extraction network and a fourth temporal feature map of a target domain dataset; Represents the spatial feature difference value between the third spatial feature map of the source domain dataset extracted based on the spatiotemporal feature extraction network and the third spatial feature map of the target domain dataset; Represents the spatial feature difference value between the eighth spatial feature map of the source domain dataset extracted based on the spatiotemporal feature extraction network and the eighth spatial feature map of the target domain dataset.

4. The method for diagnosing imbalanced data in an aircraft engine gas path system based on a spatiotemporal network according to claim 3, characterized in that: The distributed perception edge module of the class balance label described in S2 is used to implement fault prediction of imbalance data of the aircraft engine gas path system based on the spatiotemporal alignment time series features, and specifically includes the following steps: S31: The spatiotemporal time series data features output by the spatiotemporal feature extraction network are recorded as x i , then the labeled examples of the limited imbalanced data are denoted as D c and D c =(x i ,y i ), where y i represents the fault class label of the labeled example and y i ∈{1,...C}; And construct the decision boundary minimum distance function from the fault category sample to the fault category label, and use the decision boundary minimum distance function as the optimal inter-class interval model for each fault category; And the expression of the optimal inter-class interval model is Where: γ c represents the minimum distance of the decision boundary; γ(x i ,y i ) represents the distance function between the labeled example and the label edge; K represents the design parameter; n c Represents the total number of samples in the data training set; S32: Define the output logarithm Z of the spatiotemporal feature extraction network = [z1, ..., z C ] and z i =log e x i , and based on the optimal inter-class interval model, define the label distribution aware marginal loss function; And the label distribution aware marginal loss function L LDAM-HG The expression is Where: z c represents the output logarithm of the c-th spatiotemporal feature extraction network and c∈{1,...,C}; Represents the fault category label y i The output logarithm of the corresponding spatiotemporal feature extraction network; Label distribution aware marginal loss function L LDAM-HG Perform smoothing to obtain a smooth loss function with label distribution-aware margins; And the smooth loss function L LDAM The expression is S33: Based on the smooth loss function L LDAM , construct the class balance label distributed perception marginal module; and based on the class balance label distributed perception marginal module, realize the fault prediction of the imbalance data of the aircraft engine gas path system; and the class balance label distributed perception marginal module L CBDM The expression is Where: Where: β m Indicates the optimal parameters of the STFD model found by the grid search method during each model training process and β m ∈[0, 1]; n c Represents the total number of samples in the data training set; represents the optimal model parameters corresponding to the data training set; z i represents the output logarithm of the spatiotemporal feature extraction network of the i-th sample; represents the spatiotemporal feature extraction network output value of the cth sample; Represents the spatiotemporal feature extraction network output value of the i-th sample.

5. The method for diagnosing imbalanced data in an aircraft engine gas path system based on a spatiotemporal network according to claim 4, characterized in that: The total loss function of the model constructed in S4 is: Where: Loss all Represents the total loss function of the model.

Images