Gas turbine state recognition method based on multi-scale void variational graph convolutional network

Abstract

The invention relates to a gas turbine state recognition method based on a multi-scale void variational graph convolutional network, and the method comprises the steps: obtaining the working condition data of a gas turbine under a fault working condition, and enabling the working condition data to comprise a plurality of sensor signals which are mutually interfered and have redundant information; performing multi-scale associated feature extraction and coding on the working condition data through a multi-scale cavity variational graph convolution model to obtain hidden variable features; and decoding the hidden variable features into reconstructed data, and based on the reconstructed data, performing fault task classification by applying a pre-trained neural network to obtain a gas turbine state matched with the working condition data. The problem of poor accuracy of gas turbine state recognition of fault states of different parts under information interference is solved, and remarkable progress is achieved in the aspect of improving the running robustness and fault recognition precision of a gas turbine system.

Description

Technical Field

[0001] The present application relates to the field of gas turbine control, and in particular to a method, system, computer device and computer-readable storage medium for gas turbine state recognition based on a multi-scale atrous variational graph convolutional network. Background Art

[0002] As a thermal machine with excellent characteristics, gas turbines are widely used in the fields of shipbuilding, electric power, aviation industry, etc., and have gradually become important core power equipment. With the continuous increase in the number and scale of gas turbine equipment, its safety has received widespread attention.

[0003] During the service life of the equipment, gas turbine failures often occur, which not only threatens the safety of personnel, but also may lead to serious consequences such as production interruption, energy waste and environmental pollution. In the actual operation and maintenance management of gas turbine equipment, the monitoring process is affected by the complex structural units such as the intake section, low-pressure compressor, high-pressure system, combustion chamber, turbine, etc., resulting in the overlapping of different fault types of key components and difficulty in feature extraction.

[0004] In the process of component monitoring, although sensors can be arranged independently for key components, the collected gas path signals or vibration signals still contain signal components of different component objects and environmental noise components, and there may be correlations between sensors, which interferes with the fault state identification of the intelligent diagnosis model. Therefore, how to identify and distinguish the fault states of different components under information interference is an important task in fault diagnosis research on gas turbines, and it is also the difficulty of current diagnostic technology. Summary of the invention

[0005] The embodiments of the present application provide a method, system, computer device and computer-readable storage medium for identifying the state of a fuel engine based on a multi-scale void variational graph convolutional network, so as to at least solve the problem of poor accuracy in identifying the state of a fuel engine by identifying the fault states of different components under information interference.

[0006] In a first aspect, an embodiment of the present application provides a method for identifying a state of a gas engine based on a multi-scale hole variational graph convolutional network, the method comprising:

[0007] Acquiring operating data of the gas turbine under a fault condition, wherein the operating data includes a plurality of sensor signals that interfere with each other and have redundant information;

[0008] Through a multi-scale atrous variational graph convolution model, multi-scale correlation features are extracted and encoded for the working condition data to obtain latent variable features;

[0009] Furthermore, the latent variable features are decoded into reconstructed data, and based on the reconstructed data, a pre-trained neural network is applied to perform fault task classification to obtain a combustion engine state that matches the operating condition data.

[0010] In some embodiments, the multi-scale correlation features of the operating condition data are extracted and encoded through a multi-scale atrous variational graph convolution model, and the latent variable features obtained include:

[0011] Through a multi-scale atrous convolutional attention network, features with different weights are extracted from the working condition data at multiple scales to obtain a multi-scale feature representation;

[0012] Based on the correlation between the working condition data, a graph structure is constructed, and the multi-scale feature representation is processed based on the graph structure through a graph convolution network to obtain the spatial correlation between the sensor data and obtain an optimized multi-scale feature representation;

[0013] The optimized multi-scale feature representation is dimensionality reduced by a variable self-encoder to obtain the latent variable feature.

[0014] In some embodiments, the multi-scale dilated convolution attention network is used to extract features of the working condition data at multiple scales with different weights, and the multi-scale feature representation obtained includes:

[0015] By using a plurality of convolution kernels with different void ratios, convolution operations are performed on the working condition data at different scales to obtain feature maps of a plurality of different scales;

[0016] Fusing multiple feature maps of different scales to obtain a fused feature map, wherein the fused feature map contains information of the sensor signal at different scales;

[0017] Through the attention mechanism, the key features in the fusion feature map are enhanced and irrelevant redundant features are suppressed to obtain the multi-scale feature representation.

[0018] In some embodiments, the multi-scale feature representation is obtained by enhancing key features in the fusion feature map and suppressing irrelevant redundant features through an attention mechanism, and includes:

[0019] Extracting global information of the fused feature map through global average pooling;

[0020] Processing the global information through multiple fully connected layers to obtain weight vectors corresponding to each piece of global information, wherein each weight vector corresponds to a channel of the fused feature map;

[0021] The weight vector is multiplied element-by-element by the fused feature map to enhance key features in the fused feature map and suppress irrelevant redundant features, thereby obtaining the multi-scale feature representation.

[0022] In some embodiments, based on the correlation between the working condition data, a graph structure is constructed, and based on the graph structure, the multi-scale feature representation is processed by a graph convolution network to obtain the spatial correlation between the sensor data, and the optimized high-dimensional feature representation includes:

[0023] Obtaining the correlation coefficient between each sensor data in the working condition data;

[0024] Constructing a graph structure corresponding to the working condition data by taking each sensor data as a node, taking the correlation between the sensor data as an edge, and using the correlation coefficient as a weight;

[0025] Through the graph convolutional network, the multi-scale feature representation is convolved using the association information and timing characteristics between the nodes in the graph structure to obtain an optimized feature representation with enhanced feature expression capabilities.

[0026] In some embodiments, the graph convolution network is used to perform a convolution operation on the optimized feature representation by utilizing the association information and timing characteristics between the nodes in the graph structure, and the optimized feature representation after the feature expression capability is enhanced includes:

[0027] Multiple graph convolution layers are used to update the feature representation of each node by aggregating the information of neighboring nodes, where the graph convolution layer implements a basic convolution operation through the following formula:

[0028]

[0029] Among them, H (l) is the feature matrix of the l-th layer nodes, is the normalized adjacency matrix, W (l) is the weight matrix of the lth layer, σ is the activation function;

[0030] By stacking multiple graph convolutional layers, the range of neighbors corresponding to each node that can aggregate information is expanded, and a higher-level feature representation is extracted to obtain a multi-scale feature representation corresponding to each node, wherein the multi-scale feature representation includes the original features of the sensor data and the association information between the neighbor nodes corresponding to the sensor data.

[0031] In some embodiments, the optimized multi-scale feature representation is dimensionality reduced by a variable self-decomposition encoder to obtain the latent variable features including:

[0032] The optimized multi-scale feature representation is subjected to feature extraction by a variational autoencoder to obtain a latent space probability distribution, wherein the latent space probability distribution is used to represent the distribution of each sensor data in the latent space;

[0033] The latent space probability distribution is sampled to obtain a latent variable representation, and reconstructed operating condition data that is similar to the original operating condition data is obtained based on the latent variable representation through a decoding network.

[0034] In a second aspect, an embodiment of the present application provides a combustion engine state recognition system based on a multi-scale void variational graph convolutional network, the system comprising: a data acquisition module and a data processing module, wherein:

[0035] The data acquisition module is used to obtain the operating data of the gas turbine under a fault condition, wherein the operating data includes a plurality of sensor signals that interfere with each other and have redundant information;

[0036] The data processing module is used to extract and encode multi-scale correlation features of the operating condition data through a multi-scale atrous variational graph convolution model to obtain latent variable features;

[0037] Furthermore, the latent variable features are decoded into reconstructed data, and based on the reconstructed data, a pre-trained neural network is applied to perform fault task classification to obtain a combustion engine state that matches the operating condition data.

[0038] In a third aspect, an embodiment of the present application provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method described in the first aspect above when executing the computer program.

[0039] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect above.

[0040] Compared with the related art, the engine state recognition method based on the multi-scale void variational graph convolution network provided in the embodiment of the present application obtains the operating data of the gas turbine under fault conditions, wherein the operating data includes a variety of sensor signals that interfere with each other and have redundant information; through the multi-scale void variational graph convolution model, the operating data is subjected to multi-scale correlation feature extraction and encoding to obtain latent variable features; and the latent variable features are decoded into reconstructed data, and based on the reconstructed data, the pre-trained neural network is applied to classify the fault tasks to obtain the engine state matching the operating data. The problem of poor accuracy in engine state recognition of the fault states of different components under information interference is solved, and significant progress has been made in improving the robustness of engine system operation and the accuracy of fault recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0042] Figure 1 It is a flow chart of a method for identifying a state of a gas engine based on a multi-scale hole variational graph convolutional network according to an embodiment of the present application;

[0043] Figure 2 is a schematic diagram of the structure of a multi-scale dilated convolutional attention network according to an embodiment of the present application;

[0044] Figure 3 is a schematic diagram of a multi-scale multi-scale atrous variational graph convolution model according to an embodiment of the present application;

[0045] Figure 4 It is a structural block diagram of a combustion engine state recognition system based on a multi-scale hole variational graph convolutional network according to an embodiment of the present application;

[0046] Figure 5 It is a gearbox failure mode test scenario diagram and architecture diagram according to an embodiment of the present application;

[0047] Figure 6 is a schematic diagram of changes in accuracy and loss during a training process according to an embodiment of the present application;

[0048] Figure 7 is a schematic diagram of the time domain waveform of the gear vibration signal under different states according to an embodiment of the present application;

[0049] Figure 8 is a schematic diagram of changes in accuracy and loss during a training process according to an embodiment of the present application;

[0050] Fig. 9 is a schematic diagram of a confusion matrix of a GM-VAE model at different noise levels according to an embodiment of the present application;

[0051] Fig.10 It is a schematic diagram of the visualization results of the GM-VAE model according to the embodiment of the present application using feature dimensionality reduction to 2 dimensions at noise levels I and III;

[0052] Fig.11 It is a normalized partial time domain waveform diagram of the unit speed collected according to an embodiment of the present application;

[0053] Fig.12 is a schematic diagram of changes in accuracy and loss during a training process according to an embodiment of the present application;

[0054] Fig.13 It is a schematic diagram of the test set classification confusion matrix and the visualization result after the dimensionality reduction of the sample hidden layer features according to the embodiment of the present application;

[0055] Fig.14 It is a schematic diagram of the internal structure of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is described and illustrated below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. Based on the embodiments provided in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0057] Obviously, the drawings described below are only some examples or embodiments of the present application. For ordinary technicians in this field, the present application can also be applied to other similar scenarios based on these drawings without creative work. In addition, it can also be understood that although the efforts made in this development process may be complicated and lengthy, for ordinary technicians in this field related to the content disclosed in this application, some changes in design, manufacturing or production based on the technical content disclosed in this application are just conventional technical means, and should not be understood as insufficient content disclosed in this application.

[0058] Reference to "embodiments" in this application means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those of ordinary skill in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0059] Unless otherwise defined, the technical terms or scientific terms involved in this application should be understood by people with ordinary skills in the technical field to which this application belongs. The words "one", "a", "a", "the" and the like involved in this application do not indicate a quantitative limitation, and may represent the singular or plural. The terms "include", "comprise", "have" and any of their variations involved in this application are intended to cover non-exclusive inclusions; for example, a process, method, system, product or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units that are not listed, or may also include other steps or units inherent to these processes, methods, products or devices. The words "connect", "connected", "coupled" and the like involved in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The "multiple" involved in this application refers to two or more. "And / or" describes the association relationship of associated objects, indicating that there may be three relationships, for example, "A and / or B" can represent: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the objects before and after are in an "or" relationship. The terms "first", "second", "third", etc. involved in this application are only used to distinguish similar objects and do not represent a specific ordering of the objects.

[0060] Since the 20th century, gas turbines, as a thermal machinery with excellent characteristics, have been widely used in the fields of shipbuilding, electric power, aviation industry, etc., and have gradually become important core power equipment. With the continuous increase in the number and scale of gas turbine equipment, its safety assurance has received widespread attention. During the service of the equipment, gas turbine failure events occur from time to time, which not only threatens the safety of personnel, but also may lead to serious consequences such as production interruption, energy waste and environmental pollution. In the actual operation and maintenance management of gas turbine equipment, the monitoring process is affected by the complex structural units such as the intake section, low-pressure compressor, high-pressure system, combustion chamber, turbine, etc., resulting in the overlapping of different fault types of key components and difficulty in feature extraction. In the process of component monitoring, although sensors can be arranged independently for key components, there are still signal components and environmental noise components of different component objects in the collected gas path signals or vibration signals, and there may be correlation between sensors, which interferes with the fault state recognition of the intelligent diagnosis model. Therefore, how to identify and distinguish the fault states of different components under information interference is an important task in the fault diagnosis research of gas turbines, and it is also the difficulty of current diagnosis technology.

[0061] At present, fault diagnosis methods are mainly divided into two methods: traditional feature extraction and artificial intelligence network learning. In the traditional feature extraction method, the identification of fault types is based on the physical characteristics of the monitoring signal. In 2018, Li et al. proposed a method combining adaptive multi-scale filtering and hierarchical permutation entropy, using the Laplace score method to refine the extracted entropy features, and input them into the support vector machine to identify different fault degrees of the gearbox. In 2020, Mao et al. proposed an improved singular spectrum decomposition method, which introduced the mode mixing index and over-decomposition index to select the optimal embedding dimension, and combined with the frequency-weighted energy slicing bispectral technology to achieve effective rolling bearing weak fault symptom extraction. In 2022, Song et al. proposed a support vector machine classification method based on sparse scaled convex hull, which reduced the dimension of the sample set features through the random forest algorithm, and used sparse technology to adjust the convex hull scaling coefficient of the sample set to achieve high-performance fault identification of rolling bearings. Fault identification based solely on traditional features is suitable for fault diagnosis of simple rotating parts. When faced with monitoring data of complex gas turbine equipment, the identification results are easily affected by monitoring signal noise and human experience. Such problems can be effectively avoided by combining artificial intelligence network learning. In 2019, Zhong et al. studied a transfer learning method based on convolutional neural networks and support vector machines for gas turbine fault diagnosis, which has excellent diagnostic performance under small sample conditions. In 2020, Zhao et al. developed a new deep learning method called deep residual shrinkage network, which extracts features from high-noise vibration signals and achieves high-precision fault diagnosis by inserting soft thresholding as a nonlinear transformation layer into the deep structure. In 2021, Castro-Cros et al. compared and verified the superiority of different intelligent network learning methods in gas turbine object applications, demonstrating its great application potential in gas turbine fault diagnosis.

[0062] As a powerful data feature learning and sample generation model, the variational autoencoder (VAE) obtains the deep state representation of the input signal through the generative model and maps it to the latent variable space. It has the ability to learn the potential distribution of fault features from complex data, and provides new ideas and methods for gas turbine fault diagnosis. In 2021, Dewangan et al. proposed an intelligent fault diagnosis scheme based on deep convolutional variational autoencoder. By extracting discriminant features and combining data reduction and random sampling techniques, it achieved high accuracy on rotating machinery data sets. Yan et al. proposed an optimized stacked variational denoising autoencoder method for intelligent fault diagnosis of rolling bearings. By integrating variational autoencoders and denoising autoencoders, robust fault features were extracted. Experiments have shown that this method has high recognition accuracy in rolling bearing fault diagnosis and is superior to other deep learning methods [8]. However, there are still the following problems when using variational autoencoders for gas turbine fault diagnosis: the gas turbine has a complex structure, the monitoring process of the components is interfered by itself or other components, the internal fault modes are mixed and there is a large amount of redundant noise, which makes it difficult to extract fault features; at the same time, the sensor data of the gas turbine usually has correlations, and the traditional variational autoencoder fault diagnosis method often cannot make full use of this correlation information, which may lead to the model's lack of perception of the dependency between sensors, thereby limiting the performance of fault diagnosis.

[0063] In view of this, the embodiment of the present application adopts a method for gas engine state recognition based on a multi-scale void variational graph convolutional network. Based on the variational autoencoder (VAE) structure, this method further integrates a multi-scale void convolutional attention mechanism and a graph convolutional network to improve the model's expressiveness and robustness for complex data.

[0064] Figure 1 is a flow chart of a method for identifying a combustion engine state based on a multi-scale hole variational graph convolutional network according to an embodiment of the present application, such as Figure 1 As shown, the method comprises the following steps:

[0065] S101, acquiring operating data of a gas turbine under a fault condition, wherein the operating data includes a plurality of sensor signals that interfere with each other and have redundant information;

[0066] Among them, in the process of gas turbine component monitoring, although sensors can be independently arranged for its key components, the collected gas path signals or vibration signals still contain signal components of different component objects and environmental noise components; at the same time, there may be correlations between sensors, which interferes with the fault state identification of the intelligent diagnosis model.

[0067] Specifically, these sensor signals include but are not limited to: temperature data such as exhaust temperature, intake temperature, combustion chamber temperature, pressure data such as intake pressure, combustion chamber pressure and exhaust pressure, flow data such as gas flow and coolant flow, and speed data such as rotor speed and compressor speed.

[0068] It should be understood that the operating data in this embodiment mainly comes from a plurality of different types of sensors, which record the operation of the gas turbine under abnormal conditions. These operating data usually include a variety of sensor signals, which interfere with and are redundant with each other.

[0069] S102, extracting and encoding multi-scale correlation features of the working condition data through a multi-scale atrous variational graph convolution model to obtain latent variable features;

[0070] It should be noted that before the above working condition data is input into the multi-scale atrous variational graph convolution, the working condition data can be preprocessed for subsequent feature extraction and model training;

[0071] In this embodiment, the specific means used for preprocessing include but are not limited to data cleaning and data segmentation; among them, data cleaning is to perform preprocessing operations such as noise filtering, data interpolation, and normalization on the collected sensor signals to ensure the consistency and integrity of the data; data segmentation is used to divide the operating condition data into multiple small segments according to the time window.

[0072] Furthermore, the multi-scale edge-controlled variational graph convolution model provided by the technical solution of the present application further introduces a graph convolution network and multi-scale hole convolution on the basis of the variational self-decomposition encoder, which significantly improves the performance of the model in the gas turbine fault diagnosis task;

[0073] Specifically, this step includes the following subdivision steps:

[0074] Step 1: Through the multi-scale dilated convolutional attention network, the working condition data is extracted with different weights at multiple scales to obtain multi-scale feature representation;

[0075] Among them, the monitoring signal of the gas turbine contains high-frequency carrier components and multiple low-frequency modulated signal components, and fault information often exists in the low-frequency modulated signal components. At the same time, affected by the environment and structure, the signal redundant noise is serious, causing useful features to be submerged by noise. When using the network model for fault diagnosis, it is necessary to pay attention to different types of signal components at the same time, that is, to obtain signal features from multi-scale perspectives. Therefore, the multi-scale void convolution attention mechanism is introduced to obtain feature information at different scales and angles, further improving the model's robustness to noise.

[0076] In this step, the multi-scale dilated convolution is combined with the attention mechanism to form a multi-scale dilated convolution attention mechanism, which allows the model to extract features from data at different scales and enhances attention to important features through the attention mechanism.

[0077] Figure 2 is a schematic diagram of the structure of a multi-scale hole convolutional attention network according to an embodiment of the present application, combined with Figure 2 Specifically, step 1 includes the following steps:

[0078] Step 1.1, through multiple convolution kernels with different void ratios, convolution operations are performed on the working condition data at different scales to obtain feature maps of multiple scales;

[0079] Before step S1, a multi-scale dilated convolutional network is constructed, and convolution kernels with different dilation rates are pre-set to extract features at different scales.

[0080] Specifically, three convolution kernels with different dilation rates can be designed, namely r=1, r=2, and r=3. These convolution kernels are used to capture feature information of different scales and perform convolution operations on the input sensor signals. Each dilated convolution kernel acts on the input data to generate a corresponding feature map. In addition, after each layer of convolution, the Relu activation function can be used to improve the nonlinear expression ability, and a normalization layer can be added to stabilize the training process.

[0081] It can be understood that dilated convolution is a common convolution method, also known as expanded convolution or dilated convolution, which is a process of expanding the convolution kernel by adding some holes (zero values) between the convolution kernel elements. For larger objects that occupy more pixels in the original data, there are often discontinuous convolution kernels, which may eventually result in not all points in the feature map being able to participate in the calculation, which will also lose the continuity of information. Multi-scale dilated convolution is introduced to form a cyclic structure, which can ensure that all pixels can participate in the calculation in each cycle.

[0082] Furthermore, in the process of obtaining feature maps using dilated convolution, it is necessary to calculate the receptive field size under different dilation rates to ensure that different convolution kernels can cover different areas of the input signal and effectively extract feature information.

[0083] Among them, the receptive field refers to the area range in the input space corresponding to a pixel on the output feature map. Specifically, the calculation formula for the receptive field size is as follows:

[0084]

[0085] Among them, RF l+1 is the receptive field size corresponding to the current feature map, that is, the target receptive field to be calculated. RF l is the receptive field size corresponding to the feature map of the previous layer, f l+1 is the convolution kernel size of the previous convolution layer, and the last multiplication term represents the stride product of the previous convolution layer. When the stride is greater than 1, the size of the receptive field will grow exponentially.

[0086] Through the above steps, the dilated convolution is used to extract features of the original working condition data at multiple scales. The step size can be adjusted by adjusting the expansion rate, thereby adjusting the receptive field, so that the model can obtain richer contextual information without increasing parameters.

[0087] Step 1.2, multiple feature maps of different scales are fused to obtain a fused feature map, where the fused feature map contains information of sensor signals at different scales;

[0088] Specifically, the feature maps generated under different dilation rates are superimposed to form a multi-scale feature representation, which can capture important information of the signal at different scales.

[0089] Step 1.3, through the attention mechanism, enhance the key features in the fusion feature map and suppress irrelevant redundant features to obtain a multi-scale feature representation.

[0090] Among them, the attention mechanism is after the feature map is generated to further improve the representation ability of the feature map or the degree of attention of the model to key information;

[0091] In this embodiment, an attention mechanism is introduced after the convolutional layer to enhance the model's attention to important features based on channel attention or spatial attention (such as SE module or CBAM module). The process specifically includes the following steps:

[0092] Step 1.3a, extract the global information of the fused feature map through global average pooling;

[0093] Specifically, the core of the attention mechanism is to calculate the importance of each feature map element and assign corresponding weights. Through such weight allocation, the network can focus on the most important features for the current task.

[0094] Therefore, firstly, the global information of the feature map is extracted by global average pooling (GAP) or global max pooling (GMP). Assuming that the dimension of the feature map is H×W×C (height, width, number of channels), the result after global pooling is 1×1×C, that is, each channel generates a value;

[0095] Step 1.3b, process the global information through multiple fully connected layers to obtain the weight vectors corresponding to each global information, each weight vector corresponds to a channel of the fused feature map;

[0096] Among them, the global pooling result is passed through one or more fully connected layers (or through convolution operations) to generate a weight vector W, the size of which is C1×1×C, corresponding to each channel of the feature map.

[0097] Step 1.3c, multiply the weight vector by the fused feature map element by element to enhance the key features in the fused feature map and suppress irrelevant redundant features to obtain a multi-scale feature representation.

[0098] Combined with the above steps, this step is to re-multiply the weight vector W with each channel of the fusion feature map (element-by-element product). At this time, each channel of the feature map will be weighted and adjusted according to its importance.

[0099] Furthermore, the weighted feature map emphasizes the important feature areas and weakens the noise and unimportant parts. The final result is the enhanced feature map processed by the attention mechanism.

[0100] Through the above step Step 1.3, an attention mechanism is introduced after the convolution layer. Based on the channel attention mechanism, the weights of the convolution kernel can be dynamically adjusted so that the model can better focus on the signal features related to the fault.

[0101] Corresponding to the scheme of the present application, in order to address the phenomenon that the monitoring signals of complex mechanical equipment contain serious noise, which leads to difficulty in pattern recognition during fault diagnosis, the multi-scale dilated convolution is combined with the attention mechanism through the above-mentioned step Step 1 to form a multi-scale dilated convolution attention mechanism. This embodiment combines the multi-scale dilated convolution with the attention mechanism to establish a multi-scale dilated convolution attention mechanism, thereby effectively processing the noise and redundant information in the gas turbine monitoring signal, and significantly improving the robustness and diagnostic accuracy of the model, especially when the signal noise is large.

[0102] Step 2, based on the correlation between the working condition data, construct the working condition data graph structure;

[0103] Among them, considering that the data generated by gas turbine multi-sensors usually have complex spatial and temporal correlations, traditional data processing methods often have difficulty capturing these complex dependencies. Multi-sensor data usually contain spatial correlations, that is, the reading of one sensor may be affected by the reading of its neighboring sensors.

[0104] In the present application, a graph convolutional network (GCNs) is used to effectively capture the spatial correlation between multi-sensor data of a gas turbine through a graph convolution operation, thereby improving the accuracy of signal processing.

[0105] At the same time, the multi-sensor signals of gas turbines often have time series characteristics. GCNs can also aggregate information in the time dimension through graph convolution operations. By adding graph convolution modules to the model, the performance of signal processing can be effectively improved, providing more accurate results for fault diagnosis tasks.

[0106] Specifically, step 2 includes the following subdivision steps:

[0107] Step 2.1, obtain the correlation coefficient between each sensor data in the working condition data, use each sensor data as an edge, use the correlation between sensor data as an edge, and use the correlation coefficient as a weight to construct a graph structure corresponding to the working condition data;

[0108] Specifically, in the graph structure, each node represents a sensor data in the graph structure, and the edge represents the correlation between sensors. Furthermore, the weight of the edge can be defined based on the correlation between sensor data or other metrics (such as Pearson correlation coefficient, mutual information, etc.). The higher the weight, the stronger the correlation between the data of the two sensors.

[0109] The correlation coefficient (such as the Pearson correlation coefficient) between each pair of sensors is calculated to determine the weight of the edge in the graph; further, these calculation results are used to construct the adjacency matrix A of the graph. The adjacency matrix is ​​a matrix that represents the connection relationship between nodes. In addition, the weighted adjacency matrix can also be normalized to facilitate subsequent graph convolution operations.

[0110] Step 2.2, process the multi-scale feature representation based on the working condition data graph structure through the graph convolution network to obtain the spatial correlation between sensor data and obtain the optimized multi-scale feature representation.

[0111] It should be noted that in the graph convolutional network, the convolution operation is performed on the nodes of the graph structure, which updates the feature representation of each node by aggregating the information of neighboring nodes. Furthermore, GCN realizes the transmission and fusion of information by weighted summing the features of each node with the features of its neighboring nodes.

[0112] In an exemplary embodiment, the graph convolution layer implements a basic convolution operation through the following formula:

[0113]

[0114] Among them, H (l) is the feature matrix of the l-th layer nodes, is the normalized adjacency matrix, W (l) is the weight matrix of the lth layer, σ is the activation function;

[0115] Graph Convolutional Network is a deep learning model for processing graph structured data. Unlike traditional Convolutional Neural Networks (CNNs) which are suitable for processing regular grid structured data (such as images), GCNs can effectively process the relationship and local structure between nodes. In graph convolutional networks, we consider a graph that contains a set of nodes and the connection relationships between nodes. Each node has a feature vector that represents the attributes or features of the node. The goal of GCNs is to learn the representation of each node so that the node representation can capture the relationship between nodes and the global structure of the graph. The core idea of ​​GCNs is to update the representation of each node through the information of neighboring nodes.

[0116] Specifically, for each node, GCNs aggregates its features with the features of its neighboring nodes and fuses the information of neighboring nodes by learning weights. Each node is able to comprehensively consider the features of the nodes related to it and generate an updated representation. By stacking multiple graph convolutional layers, GCNs are able to gradually expand the receptive field of the node and extract more advanced feature representations. Each graph convolutional layer updates the representation of the node and passes the updated representation to the next layer. In this way, information can propagate in the network, and the representation of the node is gradually enriched and refined.

[0117] It can be understood that the multi-scale dilated convolution attention mechanism will first perform feature extraction and noise reduction on the sensor data, which can improve the quality of the features and provide a cleaner and richer input for subsequent graph convolution operations.

[0118] Furthermore, the graph convolutional network processes the extracted features and captures the spatial correlation between sensors through the graph structure, which will further enrich the feature representation and make it more suitable for gas turbine fault diagnosis tasks.

[0119] By combining these two modules, the model can simultaneously capture the multi-scale characteristics of sensor data and the complex correlations between sensors, thereby fully expressing the characteristics of the data at different levels, ultimately improving the accuracy and robustness of gas turbine fault diagnosis.

[0120] Step 4, through the variable self-encoder, the optimized multi-scale feature representation is encoded by dimension reduction to obtain the latent variable features.

[0121] Among them, the traditional encoder (AE) is a data compression model. In the data compression process, the original data X is encoded by the encoding layer and output as a low-dimensional variable Z, and then the latent variable Z is decoded by the decoding layer to output the reconstructed data X'. In this process, the latent space of AE generates discrete points, which makes AE produce invalid content when decoding certain points in the latent space, and also makes the AE latent space lack in interpretability and usability.

[0122] The Variational Autoencoder (VAE) is a generative deep network structure based on variational Bayesian inference, which can encode input data into probability distribution and constrain its latent space distribution to force it to be close to the standard normal distribution, thereby normalizing the latent space structure. Therefore, its encoding of the latent space is continuous, which can better contain the information and noise of the input signal, and learn the potential attribute features from the probability distribution of the latent variable space, and use it to generate new data so that it can be better used for fault diagnosis.

[0123] In this embodiment, obtaining latent variable features by varying the self-decomposition encoder includes the following steps:

[0124] Step 4.1, optimize the multi-scale feature representation, extract features through the encoder network of the variational autoencoder, and obtain the latent space probability distribution, where the dependent variable probability distribution is used to represent the distribution of each sensor data in the latent space;

[0125] In one example, a variational autoencoder (VAE) is an approximate inference model that fits the true distribution by optimizing the lower bound of the evidence. Suppose there is a batch of data samples {x1, ...x n}, and its whole is described by X. If we want to describe it according to {x1, …x n} to get the distribution p(x) of X. If this is possible, then directly sampling based on p(x) will give us all possible X (including {x1, …x n}), this is the ultimate ideal generative model. However, this process is often difficult, so it is necessary to use hidden variables Z.

[0126] Furthermore, we assume that data can be generated according to a certain distribution, which contains an unobservable continuous variable z. To generate new data, we can do it in two steps: First, from the prior distribution Find a value z in i ; Then, using the conditional probability distribution To generate xi The variational autoencoder introduces a recognition model and generative models Use neural networks to learn the parameters φ of the recognition model and the parameters θ of the generation model, and finally achieve an approximation of the true distribution. Recognition model It can be seen as a probabilistic encoder, generating a model It can be regarded as a probabilistic decoder.

[0127] At the same time, a constraint is imposed on the distribution of the latent space z, forcing it to be close to the standard normal distribution, ensuring local and global regularization of the latent space, where local regularization is due to variance control and global regularization is controlled by the mean. This prevents the encoding of the model in the latent space from moving away from each other and encourages as many distributions as possible to "overlap" to meet the continuity and integrity conditions of the latent space.

[0128] Step 4.2, sample the latent space probability distribution to obtain the latent variable representation, and reconstruct the reconstructed working condition data that is similar to the original working condition data based on the latent variable representation through the decoding network of the variational autoencoder.

[0129] Specifically, in combination with the above example, in order to make the probability distribution and As similar as possible, this can be achieved by minimizing the KL divergence between the two distributions. KL divergence measures the distance between two distributions. The smaller the value, the closer the two are, and the larger the value, the greater the difference between the two. The working process of the variational autoencoder is: first, the encoder is used to encode the input into a probability distribution in the latent space, then a point in the latent space is sampled from the distribution, and finally the sampled point is used to reconstruct the signal through the decoder and calculate the reconstruction error. The loss function of the variational autoencoder is:

[0130] L VAE =L recon +L Kl

[0131] In the above formula, the first term on the right is the reconstruction loss, and the second term is the KL loss. For the reconstruction error term, since the random sampling of the latent variable z will make the back propagation of the network difficult, the reparameterization technique is used to solve this problem. Assuming the prior distribution p θ (z) is a normal distribution. We can sample ε from the standard normal distribution N(0, 1) and use the encoder to fit the distribution parameters μ and σ. Then, when obtaining the latent variables, we can reconstruct z=μ+ε×σ as the latent variables. The reparameterization method enables the VAE network to achieve sampling, derivatives, error back propagation, etc.

[0132] Through the above step S102, based on the variational autoencoder model, a multi-scale dilated variational graph convolutional network model is constructed, and graph convolution and multi-scale dilated convolution attention mechanisms are introduced into the encoder of the original variational autoencoder to encode the input vibration signal, realize feature extraction and dimensionality reduction, and obtain the compressed latent variable z.

[0133] Figure 3 is a schematic diagram of a multi-scale multi-scale hole variational graph convolution model according to an embodiment of the present application, such as Figure 3 As shown, the original input data x is encoded by the network EM to obtain the hidden layer distribution information μ and σ 2 , the hidden variable z is obtained under the resampling operation, and the sample is generated after input into the decoding network DM, and the correction parameters Δμ and Δσ of the posterior distribution of the hidden variable are output. 2 , and update the distribution parameters. At the same time, the multi-scale information of the decoding network is obtained for the classification recognition network CM, so as to carry out sample recognition analysis. Among them, the multi-scale void attention mechanism can further improve the feature extraction ability and reduce noise interference. The graph convolution structure can effectively capture the correlation between different sensor information, thereby improving the model's diagnostic accuracy and robustness to noise.

[0134] Finally, the loss function of the i-th sample during network model training can be expressed as:

[0135]

[0136] Where: L (i) Represents the loss function of the i-th sample; is the cross entropy loss; is the mean square loss of the reconstructed signal; Represents the KL divergence loss of latent variables and independent multivariate standard normal distributions.

[0137] S103, decoding the latent variable features into reconstructed data, and based on the reconstructed data, applying a pre-trained neural network classification network to perform fault task classification to obtain a gas turbine state that matches the operating condition data.

[0138] After the above step S102, the latent variable feature z is obtained. Among the dependent variable features, z learns to contain feature information related to the fault. After decoding and reconstructing it to obtain features similar to the original operating condition data, the reconstructed features are input into a classifier constructed by another neural network to perform the fault classification task, and finally the engine state matching the operating condition data is obtained.

[0139] It should be noted that the latent variable features are gradually reconstructed into an output X similar to the original input signal through a series of deconvolution (upsampling) layers; the goal of the deconvolution network is to minimize the difference between the input signal and the reconstructed signal. Specifically, the mean square error MSE can be used as the reconstruction loss; it can be understood that the reconstruction process can ensure that the latent variable retains important features of the original signal while compressing information.

[0140] In addition, the reconstructed data is further feature processed through a series of fully connected layers and activation functions (such as ReLU). The last output layer uses a softmax activation function to output the probability distribution of various types of faults, i.e., engine state recognition. The specific training process of the classifier has no effect on the core invention of this application, so it will not be repeated in this embodiment.

[0141] Through the above steps S101 to S103, the present application scheme combines multi-scale dilated convolution with the attention mechanism to form a multi-scale dilated convolution attention mechanism. This mechanism allows the model to extract features from data at different scales, while enhancing the focus on important features through the attention mechanism, thereby effectively processing the noise and redundant information in the gas turbine monitoring signal and improving the model's ability to identify fault features.

[0142] Furthermore, in order to better handle the complex spatial and temporal correlations in the multi-sensor data of the gas turbine, the present application introduces a graph convolutional network (GCN). GCN can effectively capture the correlation between sensors, and by building a graph structure of sensors and utilizing the capabilities of graph neural networks, it enhances the expressiveness of feature representation, thereby improving the accuracy of fault diagnosis.

[0143] By integrating the variational autoencoder, multi-scale dilated convolution and graph convolutional network into a multi-scale dilated variational graph convolutional network model, not only the capabilities of VAE in feature learning and data generation are fully utilized, but also the model's capture of signal details is enhanced through multi-scale dilated convolution, and the complex correlation between sensor data is processed through GCN. This integration method significantly improves the performance of the model in gas turbine fault diagnosis tasks.

[0144] It should be noted that the steps shown in the above process or the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0145] The present application also provides a gas turbine state recognition system based on a multi-scale hole variational graph convolutional network. Figure 4is a structural block diagram of a combustion engine state recognition system based on a multi-scale void variational graph convolutional network according to an embodiment of the present application, such as Figure 4 As shown, the system includes: a data acquisition module 40 and a data processing module 41, wherein:

[0146] The data acquisition module 40 is used to obtain the operating data of the gas turbine under a fault condition, wherein the operating data includes a plurality of sensor signals that interfere with each other and have redundant information;

[0147] The data processing module 41 is used to extract and encode multi-scale correlation features of the working condition data through a multi-scale atrous variational graph convolution model to obtain latent variable features;

[0148] Furthermore, the latent variable features are decoded into reconstructed data, and based on the reconstructed data, a pre-trained neural network is applied to perform fault task classification to obtain the engine state that matches the operating condition data.

[0149] Through the above system, by integrating the variational autoencoder, multi-scale dilated convolution and graph convolution network into a multi-scale dilated variational graph convolution network model, not only the capabilities of VAE in feature learning and data generation are fully utilized, but also the model's capture of signal details is enhanced through multi-scale dilated convolution, and the complex correlation between sensor data is processed through GCN. This integration method significantly improves the performance of the model in gas turbine fault diagnosis tasks.

[0150] In order to further verify the technical effect of the present invention, the present invention also verifies the effectiveness of the proposed method through gear fault simulation tests and real gas turbine fault data. The experimental results show that the model has made significant progress in improving robustness and recognition accuracy. Specifically, the test process and results are as follows:

[0151] 1. Gear failure simulation test

[0152] Under high-speed and heavy-load conditions, the gearbox is a failure-prone component of gas turbines. Safety accidents caused by gearbox transmission system failures occur from time to time, causing serious loss of life and economic risks. The operation and maintenance management of gas turbine units regards the gearbox as an important monitoring object, and its fault diagnosis plays a key role in the safe operation and maintenance of gas turbines. In order to verify the effectiveness of the multi-scale atrous variational graph convolution model in the operation and maintenance management of key components of gas turbines, a gearbox transmission system fault simulation experiment was carried out.

[0153] 1.1 Dataset Introduction

[0154] A preset fault simulation experiment for the gearbox was carried out. The test bench consists of drive motor, transmission belt, transmission shaft, gearbox, brake and other components. Figure 5is a gearbox failure mode test scenario diagram and architecture diagram according to an embodiment of the present application, such as Figure 5 As shown in the figure, the gear failure state in the gearbox transmission system is simulated by presetting gear cracks and wear faults.

[0155] When conducting the gearbox preset fault simulation experiment, the parameters of each gear are set as shown in Table 1:

[0156] Table 1: Test bench gear parameters

[0157]

[0158] In Table 1, four vibration acceleration sensors are installed near the input shaft and output shaft of the gearbox, one in the vertical direction and one in the horizontal direction. The sensor layout is as follows: Figure 5 During the experiment, 8 types of gear failure modes were preset: normal state, 1-4mm single tooth crack failure, 2mm double tooth crack failure, 0.5mm single tooth wear and 0.5mm double tooth wear. The crack width is the tooth width, the crack position is the tooth root crack, and the wear position is the tooth surface. Some of the failure modes are as follows Figure 6 As shown, Figure 6 It is a schematic diagram of the changes in accuracy and loss during the training process according to an embodiment of the present application; wherein, the rotation speed is set to 1000r / min, a group of data with a time length of approximately 115s is collected for each fault mode, and the sampling frequency is 10240Hz.

[0159] Figure 7 : is a schematic diagram of the time domain waveform of the gear vibration signal under different states according to the embodiment of the present application, and the original vibration signal collected when the gear in the gearbox is normal and when different fault modes occur is as follows Figure 7 As shown, (a) normal state, (b) 1mm crack fault, (c) 2mm crack fault, (d) 2mm double-tooth crack fault, (e) 3mm crack fault, (f) 4mm crack fault, (g) 0.5mm wear fault, (h) 0.5mm double-tooth wear fault; it can be seen from the time domain vibration signal that the signal waveforms are different but different fault modes cannot be intuitively identified.

[0160] Furthermore, in order to verify the effectiveness of the proposed multi-scale atrous variational graph convolutional network model, the collected different preset fault monitoring data were used as model input to carry out gearbox fault identification. In the specific implementation process, the input data sequence was divided into training set and test set in a ratio of 4:1. At the same time, a sliding window with a length of 128 was used to perform interval non-overlapping sliding sampling on the 4-channel data. The interval length was set to 220, and a sample with a dimension of 4×128 was constructed. On this basis, the total number of samples and test samples obtained are shown in Table 2. The samples were normalized to establish the model input sample set, so as to verify the target model.

[0161] Table 2: Total number of samples for different failure modes

[0162]

[0163] In order to simulate the various noise interferences in the actual operation of the gas turbine gearbox and verify the effectiveness of the model diagnosis in a strong noise environment, the experiment set three noise levels from high to low, and added Gaussian white noise with a mean of 0 and different standard deviations to the original signal. The noise level settings are shown in Table 3.

[0164] Table 3 Different noise level settings

[0165]

[0166] 1.2 Experimental Results Analysis

[0167] Figure 8 is a schematic diagram of the change in accuracy and loss during the training process according to an embodiment of the present application, such as Figure 8 As shown in the figure, the GM-VAE model converges after 5 rounds of training at noise level I and reaches the highest accuracy, while it takes 100 rounds of training to reach convergence accuracy at noise level III. The difficulty of model training increases significantly with the increase of noise level, and the noise level also has a great impact on model accuracy. In order to further verify the improvement of the accuracy of the GM-VAE model at different noise levels, other mainstream fault diagnosis models are selected for comparison, including AE, CNN and LSTM. The network layers and hyperparameter settings of the three models are consistent with GM-VAE, and the gearbox fault data is used for training. The final results are shown in Table 4.

[0168] Table 4: Fault diagnosis results of different models for noisy monitoring signals

[0169]

[0170] From the recognition results, we can see that as the noise level increases, compared with other network models, GM-VAE has the highest diagnosis accuracy, especially in strong noise conditions, and the model performs better in complex signal environments. This verifies the effectiveness of the GM-VAE model in improving the accuracy of strong noise fault diagnosis. Fig. 9 : is a schematic diagram of the confusion matrix of the GM-VAE model at different noise levels according to an embodiment of the present application. The confusion matrix of the test results of the GM-VAE model at noise levels I and III is as follows: Fig. 9 As shown in the figure, the horizontal axis is the predicted label and the vertical axis is the label corresponding to the sample.

[0171] Fig.10It is a schematic diagram of the visualization results of the GM-VAE model according to an embodiment of the present application using feature dimensionality reduction to 2 dimensions at noise levels I and III.

[0172] 2. Gas Turbine Test

[0173] Gas turbine units have high reliability requirements and a long life cycle. The working status of important components of the equipment will change under different load or environmental conditions, resulting in a variety of possible fault types. By using fault type identification methods, the connection between important components and the working environment can be established, providing an important basis for maintenance decisions. Therefore, it is of great significance to carry out multi-unit fault identification for gas turbines. This section verifies the fault identification model by using multi-unit monitoring data of gas turbines.

[0174] 2.1 Introduction to field monitoring data

[0175] Table 5 is the field monitoring data of gas turbine units under different equipment conditions. The data is mainly from the unit speed monitoring. After a period of monitoring or an abnormal event, it is found through disassembly and boresight that the fault types involved in each unit include: normal state, blade shedding, turbine damage, blade block loss, and nozzle abnormality.

[0176] Table 5: On-site monitoring data of gas turbine unit failure

[0177]

[0178] Fig.11 It is a normalized partial time domain waveform diagram of the unit speed collected according to the embodiment of the present application, such as Fig.11 As shown in the figure, (a) normal operating speed, (b) blade shedding speed, (c) turbine damage speed, (d) blade falling speed, (e) nozzle abnormal speed; it is difficult to intuitively distinguish different fault states from the original monitoring time domain signal, and it is necessary to analyze the data and find information related to the fault type from the internal structure of the data. The multi-scale void variational graph convolutional network model is used for data analysis and processing, and the model verification is carried out.

[0179] Since the original data contains many invalid values ​​and outliers, the monitoring data is cleaned before constructing the model input sample, mainly including: removing outliers, null values ​​and invalid monitoring values. In order to eliminate the influence of dimension, normalization operation is performed on the cleaned data, and the data range is mapped to the interval [0, 1]. Then, a sliding window with a window size of 128 is used to randomly sample the data set to establish a sample set, and the training set and test set samples are divided in a ratio of 4:1. The test set does not appear in the training set, so as to verify the generalization ability of the model. On this basis, a total of 3936 training set samples and 984 test set samples were obtained.

[0180] 2.2 Analysis of test results

[0181] The relationship between the training accuracy and training loss of samples during training and the number of iterations is as follows: Fig.12 As shown in the figure, we can see that the training process is stable and the accuracy rises to more than 95% after 18 iterations of training. Fig.12 It is a schematic diagram of the changes in accuracy and loss during the training process according to an embodiment of the present application.

[0182] In order to verify the effectiveness of the GM-VAE model, we also built the following models for comparison with GM-VAE, namely AE, CNN, and LSTM. The number of network layers of the model is the same as the network structure in the previous gearbox simulation test. The real gas turbine data is used for training. The final results are shown in Table 6:

[0183] Table 6: Fault diagnosis results of different models for real gas turbine monitoring signals

[0184]

[0185] On the test set, the GM-VAE model performed best, achieving an accuracy of 99.39%, followed by the CNN model with an accuracy of 97.87%. The LSTM model had a test set accuracy of 94.28%, slightly lower than CNN and GM-VAE. The autoencoder (AE) had the lowest accuracy on the test set, at 92.33%. Considering the performance of the training set and the test set, the GM-VAE model showed the highest performance in the real gas turbine monitoring signal fault diagnosis task, indicating that GM-VAE can more effectively capture the complex patterns and potential structures of the signal when processing this type of time series data.

[0186] Fig.13 is a schematic diagram of the test set classification confusion matrix and the visualization result after the dimensionality reduction of the sample hidden layer features according to the embodiment of the present application; Fig.13 It can be seen that the model correctly classified most of the samples in the test set. The hidden layer distribution of the abnormal data of the unit caused by different fault types in the model has a high degree of separation in the feature space, and the data of the same fault type gradually cluster into the same cluster. The model has learned the internal fault features of the collected data well and mapped them to different fault types. The results verify the applicability of the model to the actual field monitoring data of gas turbine units.

[0187] 3. Ablation Test

[0188] To verify the effectiveness of the improved method, an ablation test is conducted as shown in Table 7. In Table 7, the data set of the gearbox simulation fault test under noise level III is used as the test data set. In the table, G represents the graph convolutional network structure, and M represents the multi-scale hole convolution attention mechanism. In the ablation test, we quantitatively analyzed the contribution of the multi-scale hole convolution attention mechanism (M) and the graph convolutional network (G) structure in the gas turbine fault diagnosis model. By comparing the training and test accuracy of different model configurations, we can clearly see the performance improvement of each structure relative to the baseline model (using only variational autoencoder VAE).

[0189] Table 7: Comparison of ablation test results

[0190]

[0191] When the graph convolutional network structure (G) is added to the baseline model, the training accuracy is improved from 84.58% to 89.21%, an increase of 4.63%. In terms of test accuracy, it is increased from 70.12% to 71.66%, an increase of 1.54%. This shows that the graph convolutional network can effectively utilize the spatial correlation between sensor data and enhance the model's ability to identify fault features. After introducing the multi-scale hole convolution attention mechanism (M), the training accuracy is further improved to 88.51%, an increase of 3.93% compared with the baseline model. The test accuracy is also improved, from 70.12% to 72.16%, an increase of 2.04%. This confirms the advantages of multi-scale hole convolution in processing noise and extracting signal details. When the graph convolutional network and the multi-scale hole convolution attention mechanism (i.e., GM-VAE model) are used simultaneously, the training accuracy reaches 94.56%, an increase of 10.98% compared with the baseline model. In terms of test accuracy, it increases from 70.12% to 73.52%, an increase of 3.4%. This result not only demonstrates the complementarity of the two structures, but also emphasizes the importance of combining multiple advanced techniques when performing fault diagnosis in complex systems.

[0192] The results of the ablation tests not only validated the effectiveness of each structure, but also highlighted their synergistic effect in improving the accuracy of gas turbine fault diagnosis. These findings provide effective guidance for future fault diagnosis in similar complex systems, especially when designing diagnostic models that can handle high noise and complex data.

[0193] In summary, the multi-scale atrous variational graph convolutional network model proposed in this paper has shown excellent performance in the field of gas turbine fault diagnosis. By introducing the multi-scale atrous convolution attention mechanism, the model can extract data features at different scales, enhance the expression ability of complex data, and improve the robustness and recognition accuracy of noisy monitoring signals through noise reduction. At the same time, the introduced graph convolution module effectively handles the correlation between sensors, captures the correlation information in the gas turbine data, and further improves the accuracy of fault diagnosis. In the gear fault simulation test and the verification of real gas turbine fault data, the model showed a high diagnostic accuracy, especially in the strong noise environment. The test accuracy in the gearbox simulation fault test of noise level III was 73.52%, and the test accuracy in the real gas turbine fault data set was 99.39%. Its performance is significantly better than other mainstream fault diagnosis models. And through ablation experiments, it is found that compared with the baseline model using only variational autoencoder (VAE), the test accuracy is improved by 1.54% after the introduction of the graph convolutional network structure, while the introduction of the multi-scale atrous convolution attention mechanism can improve it by 2.04%. When these two structures are used simultaneously, namely the GM-VAE model, the test accuracy reaches 73.52%, which is 3.4% higher than the baseline model. The synergistic effect of graph convolutional networks and multi-scale dilated convolutional attention mechanisms in processing complex and noisy interference signals, as well as their important role in improving the accuracy of gas turbine fault diagnosis, are demonstrated. These results show that the proposed model provides an effective solution for gas turbine fault diagnosis and is expected to play an important role in practical industrial applications.

[0194] In one embodiment, Fig.14 is a schematic diagram of the internal structure of an electronic device according to an embodiment of the present application, such as Fig.14 As shown, an electronic device is provided, which may be a server, and its internal structure diagram may be as shown in Fig.14 The electronic device includes a processor, a memory, a network interface and a database connected through a system bus. The processor of the electronic device is used to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the electronic device is used to store data. The network interface of the electronic device is used to communicate with an external terminal through a network connection. The computer program is executed by the processor to implement a method for identifying a gas engine state based on a multi-scale hole variational graph convolutional network.

[0195] Those skilled in the art will understand that Fig.14The structure shown in the figure is merely a block diagram of a partial structure related to the solution of the present application, and does not constitute a limitation on the electronic device to which the solution of the present application is applied. The specific electronic device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0196] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0197] Those skilled in the art should understand that the technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0198] The above embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A method for identifying the state of a gas turbine based on a multi-scale atrous variational graph convolutional network, characterized in that: The method comprises: Acquiring operating data of the gas turbine under a fault condition, wherein the operating data includes a plurality of sensor signals that interfere with each other and have redundant information; Through a multi-scale atrous variational graph convolution model, multi-scale correlation features are extracted and encoded for the working condition data to obtain latent variable features; Furthermore, the latent variable features are decoded into reconstructed data, and based on the reconstructed data, a pre-trained neural network is applied to perform fault task classification to obtain a combustion engine state that matches the operating condition data.

2. The method according to claim 1, characterized in that Through the multi-scale void variational graph convolution model, the multi-scale correlation features of the working condition data are extracted and encoded, and the latent variable features obtained include: Through a multi-scale atrous convolutional attention network, features with different weights are extracted from the working condition data at multiple scales to obtain a multi-scale feature representation; Based on the correlation between the working condition data, a graph structure is constructed, and the multi-scale feature representation is processed based on the graph structure through a graph convolution network to obtain the spatial correlation between the sensor data and obtain an optimized multi-scale feature representation; The optimized multi-scale feature representation is dimensionality reduced by a variable self-encoder to obtain the latent variable feature.

3. The method according to claim 2, characterized in that Through the multi-scale dilated convolution attention network, the working condition data is subjected to feature extraction with different weights at multiple scales, and the multi-scale feature representation obtained includes: By using a plurality of convolution kernels with different void ratios, convolution operations are performed on the working condition data at different scales to obtain feature maps of a plurality of different scales; Fusing multiple feature maps of different scales to obtain a fused feature map, wherein the fused feature map contains information of the sensor signal at different scales; Through the attention mechanism, the key features in the fusion feature map are enhanced and irrelevant redundant features are suppressed to obtain the multi-scale feature representation.

4. The method according to claim 3, characterized in that By using the attention mechanism, the key features in the fusion feature map are enhanced and irrelevant redundant features are suppressed, and the multi-scale feature representation obtained includes: Extracting global information of the fused feature map through global average pooling; Processing the global information through multiple fully connected layers to obtain weight vectors corresponding to each piece of global information, wherein each weight vector corresponds to a channel of the fused feature map; The weight vector is multiplied element-by-element by the fused feature map to enhance key features in the fused feature map and suppress irrelevant redundant features, thereby obtaining the multi-scale feature representation.

5. The method according to claim 2, characterized in that: Based on the correlation between the working condition data, a graph structure is constructed, and the multi-scale feature representation is processed based on the graph structure through a graph convolution network to obtain the spatial correlation between the sensor data, and the optimized high-dimensional feature representation is obtained, including: Obtaining the correlation coefficient between each sensor data in the working condition data; Constructing a graph structure corresponding to the working condition data by taking each sensor data as a node, taking the correlation between the sensor data as an edge, and using the correlation coefficient as a weight; Through the graph convolutional network, the multi-scale feature representation is convolved using the association information and timing characteristics between the nodes in the graph structure to obtain an optimized feature representation with enhanced feature expression capabilities.

6. The method according to claim 5, characterized in that By using the graph convolution network, the optimized feature representation is convolved using the association information and timing characteristics between the nodes in the graph structure, and the optimized feature representation after the feature expression capability is enhanced includes: Multiple graph convolutional layers are used to update the feature representation of each node by aggregating the information of neighboring nodes, where the graph convolutional layer implements a basic convolution operation through the following formula: Among them, H (l) is the feature matrix of the l-th layer nodes, is the normalized adjacency matrix, W (l) is the weight matrix of the lth layer, σ is the activation function; By stacking multiple graph convolutional layers, the range of neighbors corresponding to each node that can aggregate information is expanded, and a higher-level feature representation is extracted to obtain a multi-scale feature representation corresponding to each node, wherein the multi-scale feature representation includes the original features of the sensor data and the association information between the neighbor nodes corresponding to the sensor data.

7. The method according to claim 2, characterized in that The optimized multi-scale feature representation is dimensionality reduced by a variable self-decomposition encoder, and the latent variable features obtained include: The optimized multi-scale feature representation is subjected to feature extraction by a variational autoencoder to obtain a latent space probability distribution, wherein the latent space probability distribution is used to represent the distribution of each sensor data in the latent space; The latent space probability distribution is sampled to obtain a latent variable representation, and reconstructed operating condition data that is similar to the original operating condition data is obtained based on the latent variable representation through a decoding network.

8. A gas turbine state recognition system based on a multi-scale atrous variational graph convolutional network, characterized in that: The system comprises: a data acquisition module and a data processing module, wherein: The data acquisition module is used to obtain the operating data of the gas turbine under a fault condition, wherein the operating data includes a plurality of sensor signals that interfere with each other and have redundant information; The data processing module is used to extract and encode multi-scale correlation features of the operating condition data through a multi-scale atrous variational graph convolution model to obtain latent variable features; Furthermore, the latent variable features are decoded into reconstructed data, and based on the reconstructed data, a pre-trained neural network is applied to perform fault task classification to obtain a combustion engine state that matches the operating condition data.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images