An aero-engine fault diagnosis method based on graph attention wavelet neural network

Abstract

This invention belongs to the field of aero-engine fault diagnosis technology and discloses an aero-engine fault diagnosis method based on a graph attention wavelet neural network. The method includes: acquiring vibration signals of various faults in the aero-engine under operating conditions; converting the acquired vibration signals of various faults into corresponding roadmap signals; generating graph wavelet basis sets using multiple wavelet kernel functions and using a cross-attention mechanism to model the importance of different graph wavelet basis sets; then obtaining the graph attention wavelet basis based on the importance of different graph wavelet basis sets; constructing a graph attention wavelet neural network using the graph attention wavelet basis and training the graph attention wavelet neural network using roadmap signals to obtain an aero-engine fault classification model. This invention can effectively enhance the noise resistance and generalization ability of the fault classification model and improve the accuracy of fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of aero-engine fault diagnosis technology, and specifically to an aero-engine fault diagnosis method based on graph attention wavelet neural network. Background Technology

[0002] Aircraft engines are a critical component of modern air transport, and their reliability and performance are essential for aircraft safety and operational efficiency. However, various malfunctions can occur in aircraft engines during operation, including mechanical failures, poor combustion, and sensor failures. Failure to diagnose and repair these malfunctions in a timely manner can lead to serious accidents and even endanger the lives of flight crew and passengers. Over the past few decades, aircraft engine fault diagnosis technology has been extensively researched and applied. Traditional diagnostic methods are primarily based on the experience and intuition of operators or rely on manual inspection and maintenance. However, these methods are subject to the subjectivity and limitations of human judgment and cannot meet the needs for rapid and accurate diagnosis.

[0003] To overcome the limitations of traditional methods, machine learning and artificial intelligence-based aero-engine fault diagnosis technologies have received widespread attention and research in recent years. These technologies, by collecting and analyzing large amounts of real-time data, including parameters such as engine vibration, temperature, and pressure, utilize machine learning algorithms and pattern recognition techniques to automatically identify and classify various engine faults, providing accurate diagnostic results and repair suggestions. However, due to the complexity and variability of aero-engines, accurately identifying and classifying faults remains a challenging task. Secondly, the difficulty in acquiring aero-engine fault data and the scarcity of available samples make it crucial to fully leverage the structural information of existing data and achieve accurate fault diagnosis with limited sample sizes. Furthermore, the harsh operating environment of aero-engines, such as high temperature, high pressure, and high speed, places higher demands on the noise resistance and stability of fault diagnosis technologies. Summary of the Invention

[0004] To address the aforementioned shortcomings in existing technologies, this invention provides a fault diagnosis method for aero-engines based on graph attention wavelet neural networks. This method solves the problem that existing intelligent fault diagnosis methods fail to fully exploit the structural information of fault data and cannot achieve effective and stable fault diagnosis of aero-engines and their subsystems under conditions of small samples and high noise.

[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:

[0006] This invention proposes a fault diagnosis method for aero-engines based on graph attention wavelet neural networks, comprising the following steps:

[0007] S1. Acquire vibration signals of various faults in the aero-engine under operating conditions;

[0008] S2. Convert the acquired vibration signals of various faults into corresponding road map signals;

[0009] S3. Use multiple wavelet kernel functions to generate graph wavelet basis sets, and use the cross-attention mechanism to model the importance of different graph wavelet basis sets. Then, obtain the graph attention wavelet basis set according to the importance of different graph wavelet basis sets.

[0010] S4. Construct a graph attention wavelet neural network using graph attention wavelet basis, and train the graph attention wavelet neural network using road map signals to obtain an aero-engine fault classification model;

[0011] S5. The acquired vibration signal of the aero-engine is converted into the corresponding road map signal, and the aero-engine fault classification result is obtained through the trained aero-engine fault classification model.

[0012] Furthermore, step S1 also includes:

[0013] The vibration signals of various faults are sliced ​​using a sliding window to generate sample signals of consistent length.

[0014] Furthermore, step S1 also includes:

[0015] Based on the principle of small sample size, vibration signals of various faults are divided into training sets and test sets, and the training sets and test sets are randomly shuffled.

[0016] Furthermore, step S2 specifically includes:

[0017] Each sampling point in the vibration signals of various faults is taken as a vertex of the graph, and then edges are constructed between adjacent vertices according to the sampling order to generate the road map signal corresponding to each vibration signal.

[0018] Furthermore, the number of vertices in the road map signal is the number of sampling points of the vibration signal after slicing.

[0019] Furthermore, step S3, which employs a cross-attention mechanism to model the importance of different graph wavelet bases and obtains the graph attention wavelet base based on their importance, specifically includes the following steps:

[0020] The query matrix is ​​obtained by multiplying the vertex features of the road map signal by the attention parameter matrix. The key matrix is ​​obtained by multiplying the graph wavelet basis in the graph wavelet basis set by different attention parameter matrices. The value matrix is ​​obtained by multiplying the graph wavelet basis set by the identity matrix. The query matrix is ​​multiplied by the transpose of the key matrix in the key matrix set to obtain the attention score. Then, the graph attention wavelet basis is obtained by weighted summation of different graph wavelet basis (value matrix set) according to the attention score.

[0021] Furthermore, the calculation formulas for the graph wavelet basis, query matrix, key matrix, value matrix, and graph attention wavelet basis are as follows:

[0022] ψ i =U·diag(g i (λ1),…,g i (λ n ))·U T

[0023] Q = ReLU(XW) Q )

[0024] K i =ReLU(ψ) i W i K )

[0025] V i =ψ i I

[0026]

[0027]

[0028] Among them, g i (x) represents the i-th wavelet kernel function, U is the eigenvector of the graph Laplacian matrix, and λ j Let ψ be the eigenvalue of the graph Laplace matrix. i Let X represent the graphical wavelet basis generated by the i-th wavelet kernel function, and W represent the input signal. Q W i K This represents the attention parameter matrix, where I is the identity matrix, Q is the query matrix, and K is the input matrix. i The key matrix (the i-th element of the key matrix group), V i Let Ψ be the value matrix (the i-th element of the value matrix group), k represents the number of wavelet kernel functions, and Ψ atten The diagram represents the attention wavelet basis.

[0029] Furthermore, the graph attention wavelet neural network described in step S4 specifically includes:

[0030] The layers are connected in sequence: first feature transformation layer, graph attention wavelet transform layer, nonlinear activation layer, second feature transformation layer, global average graph pooling layer, and softmax activation layer.

[0031] Furthermore, the calculation formula for the graph attention wavelet transform layer is as follows:

[0032]

[0033] in, Indicates the output signal, Ψ atten The graph represents the attention wavelet basis, and X represents the input signal.

[0034] The present invention has the following beneficial effects:

[0035] This invention transforms the original time-domain vibration signal into a graph signal, then generates a graph wavelet basis set using different wavelet kernel functions. Simultaneously, it uses a cross-attention mechanism to model the importance of different graph wavelet bases and obtains the graph attention wavelet base based on the importance of different graph wavelet bases. Finally, it uses the graph attention wavelet base to construct a graph attention wavelet neural network model to classify aero-engine fault signals. This effectively enhances the noise resistance and generalization ability of the fault classification model and improves the accuracy of fault diagnosis. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the process of the aero-engine fault diagnosis method based on graph attention wavelet neural network in Embodiment 1 of the present invention;

[0037] Figure 2 This is a schematic diagram illustrating the principle of the aero-engine fault diagnosis method based on graph attention wavelet neural network in Embodiment 1 of the present invention. Detailed Implementation

[0038] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0039] Example 1

[0040] Reference Figure 1 and Figure 2This is a schematic diagram of the process and principle of an aero-engine fault diagnosis method based on a graph attention wavelet neural network according to an embodiment of the present invention. This embodiment first transforms the original time-domain vibration signal into a graph signal, then generates graph wavelet basis sets using different wavelet kernel functions. Simultaneously, a cross-attention mechanism is used to model the importance of different graph wavelet bases and obtain the graph attention wavelet base based on their importance. Finally, a graph attention wavelet neural network is constructed using the graph attention wavelet base to classify aero-engine fault signals.

[0041] The aero-engine fault diagnosis method based on graph attention wavelet neural network provided in this embodiment of the invention includes the following steps S1 to S5:

[0042] S1. Acquire vibration signals of various faults in the aero-engine under operating conditions.

[0043] In an optional embodiment of the present invention, data acquisition is performed via a propulsion reducer. During data acquisition, a vibration acceleration sensor is magnetically attached to the engine surface and connected to a signal acquisition card via a connecting cable. The vibration acquisition instrument used is the DH5922D from Donghua Testing, and the sensor used is the CA7002A from Huishi Technology. Vibration acquisition client software is used to acquire vibration signals from the engine surface at a sampling frequency of 25 kHz and a sampling time of 10 seconds. Simultaneously, the software records the vibration signals generated during this process in real time. A total of eight types of fault data, including normal conditions, are collected.

[0044] In an optional embodiment of the present invention, after data collection, the following is further included:

[0045] The vibration signals of various faults are sliced ​​using a sliding window to generate sample signals of consistent length.

[0046] Specifically, a sliding window is a visual concept for slicing time-series signals. The window slides gradually from one end of the signal to the other with a step size *s*, simultaneously capturing a signal segment of length *k* equal to the window width, until the remaining signal length is less than the window width. Here, the sample length equals the sliding window width, and the number of samples equals (original vibration signal length - sliding window width) | sliding window step size + 1. The "|" signifies integer division.

[0047] In this embodiment, a sliding window with a width of 1000 and a step size of 1000 is used to slice the original vibration signal. 250 samples with a shape of 1×1000 are obtained for each fault type, and a total of 2000 samples are obtained for 8 fault types.

[0048] In an optional embodiment of the present invention, after data sampling, the embodiment further includes:

[0049] Based on the principle of small sample size, vibration signals of various faults are divided into training sets and test sets, and the training sets and test sets are randomly shuffled.

[0050] Specifically, considering the scarcity of fault data in real industrial scenarios, the collected data on eight types of aero-engine faults (including normal data) were divided according to the principle of small sample size. 20% of the samples for each fault were selected as the training set, and 80% were selected as the test set. Then, the training set and the test set were randomly shuffled to increase the randomness of the samples.

[0051] In this embodiment, a total of 8 fault categories (including normal data) under variable speed of aero-engine were collected. After slicing the original vibration data, each fault data contained 250 samples. Considering the scarcity of fault data in real industrial scenarios, 50 samples were selected as the training set and 200 samples were selected as the test set according to the small sample principle. Then the training set and the test set were randomly shuffled to increase the randomness of the samples.

[0052] S2. Convert the acquired vibration signals of various faults into corresponding road map signals.

[0053] In an optional embodiment of the present invention, step S2 specifically includes:

[0054] Each sampling point in the vibration signals of various faults is taken as a vertex of the graph, and then edges are constructed between adjacent vertices according to the sampling order to generate the road map signal corresponding to each vibration signal.

[0055] The number of vertices in the road map signal is the number of sampling points of the vibration signal after slicing.

[0056] Specifically, a graph consists of a finite, non-empty set of vertices and a set of edges between vertices, typically represented as G = (V, E, A), where G represents a graph, V is the set of vertices in graph G, E is the set of edges in graph G, and A is the adjacency matrix of the graph. The rule for converting a time-domain signal into a path graph is: each sample point of the time-domain signal is treated as a vertex of the graph, and edges are constructed between adjacent vertices according to the sampling order. In this embodiment, a time-domain sample is constructed as a graph, where the vertices of the graph correspond to the sample points of the time-domain sample, and the number of vertices is equal to the number of sample points of the time-domain sample, which is numerically equal to the sliding window width k.

[0057] In this embodiment, for time-domain samples with a shape of 1×1000, the sample sampling points are constructed as vertices of the path graph, and then edges are constructed between adjacent vertices according to the sampling order. Thus, each path graph contains 1000 vertices and 999 edges. In this embodiment, all samples (including the test set and the training set) are transformed from time-domain signals into path graphs, resulting in a total of 2000 path graphs.

[0058] S3. Use multiple wavelet kernel functions to generate graph wavelet basis sets, and use the cross-attention mechanism to model the importance of different graph wavelet basis sets. Then, obtain the graph attention wavelet basis set according to the importance of different graph wavelet basis sets.

[0059] In an optional embodiment of the present invention, the wavelet kernel function is a function that describes the variation law of graph frequency response, the graph wavelet basis is a basis for graph wavelet transform, and the cross-attention mechanism is a deep learning model that captures the correlation or importance between multiple inputs. This embodiment uses 10 wavelet kernel functions to generate a graph wavelet basis set and utilizes the cross-attention mechanism to model the importance of different graph wavelet bases. Then, it obtains the attention graph wavelet based on the importance of different graph wavelet bases. This not only considers the sensitivity of different wavelet kernel functions to different fault types of data but also effectively improves the generalization ability of the entire model.

[0060] In this embodiment, step S3 uses multiple wavelet kernel functions to generate graph wavelet basis sets, and utilizes a cross-attention mechanism to model the importance of different graph wavelet basis sets. Then, the graph attention wavelet is obtained based on the importance of different graph wavelet basis sets, specifically including the following steps:

[0061] The query matrix is ​​obtained by multiplying the vertex features of the road map signal by the attention parameter matrix. The key matrix is ​​obtained by multiplying the graph wavelet basis in the graph wavelet basis set by different attention parameter moments. The value matrix is ​​obtained by multiplying the graph wavelet basis set by the identity matrix. The query matrix is ​​multiplied by the transpose of the key matrix in the key matrix set to obtain the attention score. Then, the graph attention wavelet basis is obtained by weighted summation of different graph wavelet basis (value matrix set) according to the attention score.

[0062] The calculation formulas for the graph wavelet basis, query matrix, key matrix, value matrix, and graph attention wavelet basis are as follows:

[0063] ψ i =U·diag(g i (λ1),…,g i (λ n ))·U T

[0064] Q = ReLU(XW) Q )

[0065] K i =ReLU(ψ) i W i K )

[0066] V i =ψ i I

[0067]

[0068]

[0069] Among them, g i (x) represents the i-th wavelet kernel function, U is the eigenvector of the graph Laplacian matrix, and λ j Let ψ be the eigenvalue of the graph Laplace matrix. i Let X represent the graphical wavelet basis generated by the i-th wavelet kernel function, and W represent the input signal. Q W i K This represents the attention parameter matrix, where I is the identity matrix, Q is the query matrix, and K is the input matrix. i The key matrix (the i-th element of the key matrix group), V i Let Ψ be the value matrix (the i-th element of the value matrix group), k represents the number of wavelet kernel functions, and Ψ atten The diagram represents the attention wavelet basis.

[0070] Specifically, this embodiment uses 10 wavelet kernel functions (including Heat, Abspline, Expwin, Held, Itersine, Mexicanhat, Meyer, Papasakis, Regular, and Simpletight, as shown in Table 1) to generate a wavelet basis set [ψ1,ψ2,……,ψ] of shape 10×1000×1000. 10 The wavelet basis set is multiplied by a parameter matrix of shape 10×1000×128. We obtain a bond matrix group [K1,K2,……,K] with a shape of 10×1000×128. 10 The elements of the wavelet basis set are multiplied by the identity matrix to obtain a value matrix [V1, V2, ..., V] of shape 10 × 1000 × 1000. 10 The graph signal x is multiplied by a parameter matrix W of shape 1×128. Q The query matrix Q with a shape of 1000×128 is obtained, and the specific calculation formula of the graph attention wavelet basis is as follows:

[0071]

[0072] In the formula, d i The key matrix K i Size of the dimension.

[0073] Table 1 Wavelet kernel function table

[0074]

[0075]

[0076] Compared with signals in traditional Euclidean space, the graph signals used in this embodiment have richer geometric structure information. Fully mining the structural information between data helps to extract data features, thereby achieving better classification results. The graph wavelet transform maps the graph signal from the vertex domain to the spectral domain, making the signal features more obvious and easier to extract. At the same time, the wavelet kernel filters the spectral domain signal, achieving the effect of enhancing features.

[0077] Traditional graph wavelet neural networks are constructed using hot kernel functions. After mapping the graph signal from the vertex domain to the spectral domain, the graph wavelet neural network uses wavelet kernels to filter the graph frequency. However, signals of different fault types often have different graph frequency characteristics, and a single wavelet kernel cannot effectively filter signals of different fault types. Therefore, this embodiment uses 10 kernel functions to generate graph wavelet basis sets and utilizes a cross-attention mechanism to model the importance of different graph wavelet basis sets and obtain the graph attention wavelet basis based on the importance of different graph wavelet basis sets. This takes into account the sensitivity of different kernel functions to signals of different fault types, and at the same time improves the generalization performance and classification ability of the present invention.

[0078] S4. Construct a graph attention wavelet neural network using graph attention wavelet basis, and train the graph attention wavelet neural network using road map signals to obtain an aero-engine fault classification model.

[0079] Specifically, the graph attention wavelet neural network includes: a first feature transformation layer, a graph attention wavelet transformation layer, a nonlinear activation layer, a second feature transformation layer, a global average graph pooling layer, and a softmax activation layer, connected in sequence.

[0080] The calculation formula for the graph attention wavelet transform layer is as follows:

[0081]

[0082] in, Indicates the output signal, Ψ atten The graph represents the attention wavelet basis, and X represents the input signal.

[0083] Specifically, for a graphical signal x, the graphical wavelet transform is defined as follows: Among them Ψ atten This is the graph attention wavelet basis obtained by applying an attention mechanism to the graph wavelet basis generated from the 10 wavelet kernel functions in step S3 of this embodiment. The graph attention wavelet neural network in this embodiment includes one graph attention wavelet layer and one softmax layer, as shown in the following formula:

[0084] grap attention wavelet layer:X1=ReLU(Ψ atten -1 X0W0)

[0085] softmax layer:O=softmax(X1W1)

[0086] In the formula, W0 and W1 are network learning parameters.

[0087] Specifically, the input graph signal first undergoes a 1×16 feature transformation, a graph attention wavelet transform, and a nonlinear activation function (such as SiLu, ReLu, etc.), then a 16×8 feature transformation, global average graph pooling, and a softmax layer. Finally, the graph attention wavelet neural network is trained using the backpropagation algorithm to classify fault signals.

[0088] The graph attention wavelet neural network constructed in step S3 is trained using the road map signals in the training set given in step S2. Then, the parameters of the graph attention wavelet neural network are optimized using the backpropagation algorithm. The trained graph attention wavelet neural network is used as an aero-engine fault classification model.

[0089] S5. The acquired vibration signal of the aero-engine is converted into the corresponding road map signal, and the aero-engine fault classification result is obtained through the trained aero-engine fault classification model.

[0090] For the vibration signal to be measured in an aero-engine, firstly, according to step S2, the vibration signal is converted into a corresponding roadmap signal, and then the roadmap signal is input into the aero-engine fault classification model to obtain the corresponding aero-engine fault classification result.

[0091] In this embodiment, the roadmap signal from the test set given in step S2 is input into the trained aero-engine fault classification model to obtain the corresponding aero-engine fault classification results. Testing shows that the aero-engine fault classification model provided by this invention has excellent noise resistance and generalization ability, and can effectively improve the accuracy of fault diagnosis.

[0092] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0093] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0094] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0095] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

[0096] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for diagnosing aero-engine faults based on graph attention wavelet neural networks, characterized in that, Includes the following steps: S1. Acquire vibration signals of various faults in the aero-engine under operating conditions; S2. Convert the acquired vibration signals of various faults into corresponding road map signals; S3. Use multiple wavelet kernel functions to generate graph wavelet basis sets, and use the cross-attention mechanism to model the importance of different graph wavelet basis sets. Then, obtain the graph attention wavelet basis set according to the importance of different graph wavelet basis sets. Specifically, the following steps are included: The query matrix is ​​obtained by multiplying the vertex features of the roadmap signal by the attention parameter matrix. The key matrix is ​​obtained by multiplying the graph wavelet bases in the graph wavelet base set by different attention parameter matrices. The value matrix is ​​obtained by multiplying the graph wavelet base set by the identity matrix. The query matrix is ​​then multiplied by the transpose of the key matrices in the key matrix set to obtain the attention score. Finally, the different graph wavelet bases are weighted and summed based on the attention score to obtain the graph attention wavelet base. The calculation formula for the cross-attention mechanism is as follows: Q=ReLU(XW Q ) K i =ReLU(ψ i IN i K ) V i =ψ i I Where X represents the input signal, ψ i W represents the graph wavelet basis generated by the i-th wavelet kernel function. Q W i K Let I be the attention parameter matrix, k be the number of wavelet kernel functions, and Ψ be the number of wavelet kernel functions. atten Represents the attention wavelet basis of the diagram; S4. Construct a graph attention wavelet neural network using graph attention wavelet basis, and train the graph attention wavelet neural network using roadmap signals to obtain an aero-engine fault classification model; the graph attention wavelet neural network specifically includes: a first feature transformation layer, a graph attention wavelet transform layer, a nonlinear activation layer, a second feature transformation layer, a global average graph pooling layer, and a softmax activation layer connected in sequence. The calculation formula for the graph attention wavelet transform layer is as follows: in, Ψ represents the output signal after graph attention wavelet transform. atten This represents the attention wavelet basis, where X represents the input signal; S5. The acquired vibration signal of the aero-engine is converted into the corresponding road map signal, and the aero-engine fault classification result is obtained through the trained aero-engine fault classification model.

2. The aero-engine fault diagnosis method based on graph attention wavelet neural network according to claim 1, characterized in that, Step S1 also includes: The vibration signals of various faults are sliced ​​using a sliding window to generate sample signals of consistent length.

3. The aero-engine fault diagnosis method based on graph attention wavelet neural network according to claim 1 or 2, characterized in that, Step S1 also includes: Based on the principle of small sample size, vibration signals of various faults are divided into training sets and test sets, and the training sets and test sets are randomly shuffled.

4. The aero-engine fault diagnosis method based on graph attention wavelet neural network according to claim 1, characterized in that, Step S2 specifically includes: Each sampling point in the vibration signals of various faults is taken as a vertex of the graph, and then edges are constructed between adjacent vertices according to the sampling order to generate the road map signal corresponding to each vibration signal.

5. The aero-engine fault diagnosis method based on graph attention wavelet neural network according to claim 1 or 4, characterized in that, The number of vertices in the road map signal is the number of sampling points of the vibration signal after slicing.

Images