Engine fault detection method based on shared potential space, electronic equipment and readable storage medium

Abstract

The invention provides an engine fault detection method based on a shared potential space, electronic equipment and a readable storage medium, and the method comprises the following steps: S1, collecting the time sequence vibration data of an engine in real time, and constructing a subgraph and a global graph; s2, performing feature extraction on the sub-graphs and the global graph to obtain corresponding sub-graph features and global graph features; s3, performing feature updating on the sub-graph features and the global graph features according to a preset adaptive updating strategy to obtain sub-graph aggregation features and updated global graph features; s4, inputting the sub-graph aggregation features and the updated global graph features into a variational auto-encoder I, and extracting potential vectors sharing a potential space; s5, performing classification based on the potential vector of the shared potential space, wherein a classification result is a fault detection result of the engine; through the method, the characteristics of the sub-graph and the global graph can be fully extracted, the distribution difference between the sub-graph and the global graph is reduced through adaptive updating, and the engine fault detection accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the field of fault detection, and in particular to an engine fault detection method based on a shared potential space, an electronic device and a readable storage medium. Background Art

[0002] The engine is the heart of a car. Its working condition directly affects the vehicle's power, economy and emission performance. Accurately grasping the real-time operating status of the engine and timely diagnosing potential engine failures are of great significance to extending the engine's service life and reducing vehicle operating costs.

[0003] In recent years, with the rapid development of sensor technology, signal processing theory, artificial intelligence and other fields, engine status monitoring and fault diagnosis systems with higher intelligence and richer functions have become a research hotspot. When the engine is working, its internal components will generate mechanical vibrations. The vibration anomalies caused by different fault types have different frequency distribution characteristics. Therefore, the engine fault diagnosis can be achieved by transforming the time domain vibration signal collected by the vibration sensor, extracting the time and frequency domain features, and then combining the machine learning classification algorithm.

[0004] Engine fault diagnosis based on machine learning usually separates the steps of feature extraction and fault classification. Sensitive feature quantities reflecting the fault state, such as time domain statistical features, frequency domain features, time-frequency domain features, etc., are extracted from the vibration monitoring data collected by the sensor to form a feature vector. Then, a suitable machine learning classification model is selected, such as support vector machine, decision tree, artificial neural network, etc., and the feature vector is input to classify and identify the engine state. This diagnostic method has a clear process, a moderate number of training samples required, and a relatively stable application effect, but its classification performance depends on the quality of artificially designed features and has limited generalization ability.

[0005] The fault diagnosis method based on deep learning uses deep learning models such as convolutional neural networks and long short-term memory neural networks to automatically learn the fault diagnosis rules directly from the original collected data, reducing the manual feature extraction process. This method can mine the inherent laws of the data, does not rely on domain knowledge and diagnostic experience, and has stronger learning and generalization capabilities. However, it requires a large number of labeled samples for training. Training with a small number of samples will result in limited model accuracy, and high-precision complex models are difficult to directly apply to on-board diagnostic systems.

[0006] Therefore, in order to solve the problem of small number of engine fault samples and thus low model detection accuracy, it is urgent to propose an engine fault detection method for small samples. Summary of the invention

[0007] In view of this, in order to improve the accuracy of engine fault detection in the few-shot scenario, the present invention proposes an engine fault detection method, an electronic device, and a readable storage medium based on a shared latent space.

[0008] An engine fault detection method based on a shared latent space provided by the present invention includes the following steps:

[0009] S1. Real-time collect the time-series vibration data of the engine, and construct a subgraph and a global graph according to the time-series vibration data;

[0010] S2. Respectively input the subgraph and the global graph into different graph convolutional networks for feature extraction to obtain corresponding subgraph features and global graph features;

[0011] S3. Respectively update the subgraph features and the global graph features according to a preset adaptive update strategy to obtain subgraph aggregation features and updated global graph features;

[0012] S4. Input the subgraph aggregation features and the updated global graph features into variational autoencoder I to obtain a shared latent space of the subgraph and the global graph, and extract the latent vectors of the shared latent space;

[0013] S5. Classify based on the latent vectors of the shared latent space, and the classification result is the engine fault detection result.

[0014] Further, the steps of updating the subgraph features and the global graph features through the adaptive update strategy are as follows:

[0015] S31. There are n subgraph features in total. Respectively input the n subgraph features and the global graph features into a multi-channel variational autoencoder and variational autoencoder II to obtain corresponding latent vectors of n subgraph latent spaces and latent vectors of the global graph latent space;

[0016] The multi-channel variational autoencoder is composed of n variational autoencoders;

[0017] S32. Determine the subgraph attention coefficient matrix according to the latent vectors of the n subgraph latent spaces;

[0018] Determine the global graph attention coefficient matrix according to the latent vectors of the global graph latent space;

[0019] S33. Update the global graph features according to the subgraph attention coefficient matrix to obtain updated global graph features;

[0020] Update the n subgraph features according to the global graph attention coefficient matrix to obtain subgraph aggregation features.

[0021] Further, the sub - graph attention coefficient matrix is determined by the following method:

[0022] S3211. Multiply the latent vectors of the n sub - graph latent spaces with the latent vectors of the n sub - graph latent spaces respectively to obtain an interaction matrix composed of n×n interaction features;

[0023] S3212. Input the interaction matrix into a one - dimensional adaptive pooling layer for node feature compression to obtain a sub - graph compressed feature matrix;

[0024] S3213. Pass the sub - graph compressed feature matrix through the Sigmoid function to obtain an attention coefficient matrix a 11 ;

[0025] S3214. Multiply the attention coefficient matrix a 11 by the transpose matrix of the attention coefficient matrix a 11 to obtain the sub - graph attention coefficient matrix a1.

[0026] Further, the global graph feature is updated by the following method to obtain the updated global graph feature:

[0027] The global graph includes a global graph Euclidean distance matrix and global graph node features; multiply the sub - graph attention coefficient matrix by the global graph Euclidean distance matrix to obtain the updated global graph Euclidean distance matrix; input the global graph feature and the updated global graph Euclidean distance matrix into a graph convolutional network to obtain the updated global graph feature.

[0028] Further, the global graph attention coefficient matrix is determined by the following method:

[0029] S3221. Input the latent vector of the global graph latent space into a one - dimensional adaptive pooling layer for node feature compression to obtain a global graph compressed feature matrix of size n×1;

[0030] S3222. Pass the global graph compressed feature matrix through the Sigmoid function to obtain the global graph attention coefficient matrix a2.

[0031] Further, the sub - graph feature is updated by the following method to obtain the sub - graph aggregation feature:

[0032] Multiply and add the global graph attention coefficient matrix with the n sub - graph features to obtain the sub - graph aggregation feature, and the calculation formula is as follows:

[0033]

[0034] where, H r represents the sub - graph aggregation feature, Hi represents the sub-graph feature, i = 1, 2, ..., n, a2 represents the global graph attention coefficient matrix, a2 = [a 21 ,a 22 ,…,a 2n ].

[0035] Furthermore, the multi-channel variational autoencoder and variational autoencoder II both use the following loss function:

[0036]

[0037] in, represents the loss function, θ represents the decoder parameters, represents the encoder parameters, x represents the real sub-graph features or the real global graph features, and p θ (x|z) represents the true posterior distribution of the latent vector in the latent space of the subgraph or the true posterior distribution of the latent vector in the latent space of the global graph. The general expression of the latent vector is z=μ+εσ, where μ and σ represent the output of the encoder, ε represents a random number that conforms to the standard normal distribution, J represents the total number of dimensions of the latent space, j represents the index of the latent variable dimension, G represents the total number of Monte Carlo sampling, and g represents the index of the number of Monte Carlo sampling;

[0038] When the loss function When it is the largest, it outputs the latent vectors of n sub-graph latent spaces or the latent vectors of the global graph latent space.

[0039] Furthermore, the subgraph and global graph are constructed by the following method:

[0040] Divide the collected time series vibration data into n segments according to the preset time window size;

[0041] Construct sub-graphs: Determine n sub-graphs based on n segments of time-series vibration data:

[0042] The number of vibration data in the i-th time series vibration data is used as the number of subgraph nodes, i = 1, 2, ..., n, and the vibration data in the i-th time series vibration data is used as the subgraph node feature X i ; Calculate the Euclidean distance between the features of each subgraph node, and combine the Euclidean distances between the features of each subgraph node into a Euclidean distance matrix A i , the Euclidean distance matrix A i is a symmetric matrix; the Euclidean distance matrix A i and subgraph node feature X i Combine and get subgraph G i =(A i ,X i ); The construction method of the remaining n-1 subgraphs is similar;

[0043] Construct a global graph: Determine a global graph based on n segments of time-series vibration data:

[0044] Take each time window as a global graph node, and take the vibration data within each segment of time-series vibration data as the global graph node feature X q ; Calculate the Euclidean distance between n global graph nodes and combine them into a Euclidean distance matrix A q , where the Euclidean distance matrix A q is a symmetric matrix; Combine the Euclidean distance matrix A q with the global graph node features to obtain the global graph G q =(A q , X q ), X q =[X1, X2, …, X n .

[0045] Correspondingly, the present invention also provides an electronic device, including:

[0046] A memory and a processor, the memory is used to store a computer program, and when the computer program is executed by the processor, the above-mentioned engine fault detection method based on a shared latent space is implemented.

[0047] Correspondingly, the present invention also provides a readable storage medium, in which computer instructions are stored, and when the computer instructions are executed by a processor, the above-mentioned engine fault detection method based on a shared latent space is implemented.

[0048] Advantages of the present invention: The present invention constructs subgraphs and a global graph according to the time-series vibration data of the engine, can collect feature information of different dimensions of the engine, and provides data support for improving the accuracy of engine fault detection; The present invention extracts the feature information of the subgraphs and the global graph through a graph convolutional network, can improve the data processing efficiency, and provides a basis for performing efficient engine fault detection; The present invention adaptively updates the subgraph features and the global graph features, reduces the distribution difference between the subgraphs and the global graph, and maps the updated global graph features and the subgraph aggregation features to the same latent space, which can ensure that the latent representations of the global graph and the subgraphs are aligned in the same continuous latent space, enabling the global information to assist the local representation, making the local features more robust and less susceptible to noise, and at the same time allowing the local information to supplement the global graph features, making the global representation more delicate; Further, it can realize more effective utilization of global information, and enhance the understanding and judgment of local features, thereby improving the accuracy of engine fault detection in the case of a small number of samples. Description of the Drawings

[0049] The present invention will be further described below with reference to the drawings and embodiments:

[0050] Figure 1 Flow chart of an engine fault detection method based on a shared latent space provided by an embodiment of the present invention.

[0051] Figure 2 Schematic diagram of a sub - graph and a global graph of an embodiment of the present invention.

[0052] Figure 3 Schematic diagram of adaptively updating the sub - graph and the global graph in an embodiment of the present invention. Detailed implementation manners

[0053] The present invention will be further described below with reference to the accompanying drawings of the specification:

[0054] See Figure 1 , an engine fault detection method based on a shared latent space provided by the present invention includes the following steps:

[0055] S1. Collect the time - series vibration data of the engine in real time, and construct a sub - graph and a global graph according to the time - series vibration data;

[0056] S2. Input the sub - graph and the global graph into different graph convolutional networks respectively for feature extraction, and obtain corresponding sub - graph features and global - graph features;

[0057] S3. Update the sub - graph features and the global - graph features respectively according to a preset adaptive update strategy to obtain sub - graph aggregation features and updated global - graph features;

[0058] S4. Input the sub - graph aggregation features and the updated global - graph features into variational auto - encoder Ⅰ to obtain the shared latent space of the sub - graph and the global graph, and extract the latent vectors of the shared latent space;

[0059] S5. Classify based on the latent vectors of the shared latent space, and the classification result is the engine fault detection result.

[0060] Through the above - mentioned method, signal feature information can be quickly obtained, the information in the samples can be utilized more fully, and the distribution difference of the graph structures between the sub - graph and the global graph can be reduced through adaptive update, improving the detection accuracy.

[0061] In this embodiment, in step S1, the time - series vibration data of the engine is collected in real time, and a sub - graph and a global graph are constructed according to the time - series vibration data; among them, the time - series vibration data refers to the vibration data collected sequentially according to time; the vibration data is collected by existing collection methods, which will not be elaborated here;

[0062] The process of constructing the sub - graph and the global graph is as Figure 2 shown, and the steps are as follows:

[0063] Divide the collected time series vibration data into n segments according to the preset time window size, that is, into n time windows. The size of the time window is set according to demand or experience, and is not limited here, preferably 5 seconds;

[0064] Construct subgraphs: Determine n subgraphs based on n segments of time-series vibration data, that is, generate a subgraph corresponding to each time window. The following takes the i-th time window as an example:

[0065] The number of vibration data in the i-th time series vibration data is used as the number of subgraph nodes, i = 1, 2, ..., n, and the vibration data in the i-th time series vibration data is used as the subgraph node feature X i ; Calculate the Euclidean distance between the features of each subgraph node. Calculating the Euclidean distance is a prior art and will not be described here; and combine the Euclidean distances between the features of each subgraph node into a Euclidean distance matrix A i , the Euclidean distance matrix A i is the adjacency matrix of subgraph i, the Euclidean distance matrix A i It is still a symmetric matrix; the Euclidean distance matrix A i and subgraph node feature X i Combine and get subgraph G i =(A i ,X i ); The construction method of the remaining n-1 subgraphs is similar;

[0066] Constructing a global graph: Determine a global graph based on n segments of time-series vibration data:

[0067] Each time window is taken as a global graph node, and the vibration data in each time series vibration data is taken as the feature X of the global graph node. q ; Calculate the Euclidean distances between n global graph nodes and combine them into a Euclidean distance matrix A q , the Euclidean distance matrix A q is a symmetric matrix; the Euclidean distance matrix A q Combined with the global graph node features, we get the global graph G q =(A q ,X q ), X q =[X1,X2,…,X n ].

[0068] Constructing a subgraph based on vibration data can reduce the complexity of calculation, and constructing a global graph can provide global information, thereby better representing node characteristics; combining the subgraph and the global graph can not only efficiently process large-scale data, but also provide global and local feature information, thereby better supporting vibration analysis and fault diagnosis.

[0069] In this embodiment, in step S2, the sub-graph and the global graph are respectively input into different graph convolutional networks for feature extraction to obtain corresponding sub-graph features and global graph features;

[0070] For each of the n sub-graphs, n graph convolutional networks are respectively used for feature extraction to obtain n sub-graph features. The convolution process of the sub-graph is represented by the following formula:

[0071]

[0072] where H i represents the i-th sub-graph feature, A i represents the adjacency matrix of the i-th sub-graph. The adjacency matrix of the i-th sub-graph is a matrix composed of the Euclidean distances of each node in the i-th sub-graph. X i represents the i-th sub-graph node feature, W i represents the weight matrix corresponding to each feature of the i-th sub-graph, and Relu (Rectified Linear Unit, ReLU, linear rectifier function) represents the activation function. represents the symmetric normalization operation on the adjacency matrix of the i-th sub-graph;

[0073] For the global graph, one graph convolutional network is used for feature extraction to obtain one global graph feature. The convolution process of the global graph is represented by the following formula:

[0074]

[0075] where H q represents the global graph feature, A q represents the adjacency matrix of the global graph. The adjacency matrix of the global graph is a matrix composed of the Euclidean distances of each node in the global graph. represents the symmetric normalization operation on the adjacency matrix of the global graph, D represents the degree matrix, represents the expanded degree matrix during the normalization process, X q represents the global graph node feature, W q represents the weight matrix corresponding to each feature of the global graph;

[0076] The graph convolutional networks used are all existing structures and will not be elaborated here;

[0077] By mining the local and global features of the engine vibration signal through sub-graph convolution and global graph convolution, it is possible to gradually deepen the abstraction from the original signal to the local and global features, improving the hierarchy and expression ability of feature representation.

[0078] In this embodiment, in step S3, the subgraph features and the global graph features are respectively updated according to a preset adaptive update strategy to obtain subgraph aggregation features and updated global graph features. The feature extraction and adaptive update processes of the subgraph and the global graph are as Figure 3 shown, Figure 3 the Z in h1 , Z h2 , …, Z hn respectively correspond to the latent vectors of the 1st to the nth subgraph latent spaces, and Z q represents the latent vector of the global graph latent space. The VAE encoder refers to the variational auto-encoder (VAE). The VAE encoders 1 to VAE encoder n are multi-channel variational auto-encoders, and the multi-channel variational auto-encoder consists of n variational auto-encoders;

[0079] The steps of updating the subgraph features and the global graph features through the adaptive update strategy are as follows:

[0080] S31. There are n subgraph features in total. The n subgraph features and the global graph feature are respectively input into the multi-channel variational auto-encoder and the variational auto-encoder II to obtain the corresponding latent vectors of the n subgraph latent spaces and the latent vector of the global graph latent space; correspondingly Figure 3 the process of inputting the output features of the graph convolution into the VAE encoder.

[0081] Inputting the subgraph features into the multi-channel variational auto-encoder means inputting the n subgraph features into the n variational auto-encoders respectively;

[0082] The variational auto-encoder in this application can be an existing feature extraction structure such as a one-dimensional convolutional neural network and a multi-layer perceptron. Since the number of parameters of the multi-layer perceptron is small, the multi-layer perceptron is preferably selected. The variational auto-encoder in this embodiment consists of two multi-layer perceptrons, that is, after the subgraph features or the global graph features are input into the first multi-layer perceptron, the output of the first multi-layer perceptron is input into the second multi-layer perceptron;

[0083] Both the multi-channel variational auto-encoder and the variational auto-encoder II adopt the following loss function:

[0084]

[0085] Among them, represents the loss function, θ represents the decoder parameters, represents the encoder parameters, x represents the real subgraph features or the real global graph features, p θ(x|z) represents the true posterior distribution of the latent vectors in the sub-graph latent space or the true posterior distribution of the latent vectors in the global graph latent space. The general form of the hidden vector is z = μ + εσ, where μ and σ represent the outputs of the encoder, and ε represents a random number conforming to the standard normal distribution. J represents the total number of dimensions of the latent space, j represents the index of the latent variable dimension, G represents the total number of Monte Carlo samplings, g represents the index of the Monte Carlo sampling times, and p θ (x|z) g represents the true posterior distribution of the g-th sampling;

[0086] When the loss function is maximized, the latent vectors in the n sub-graph latent spaces or the latent vectors in the global graph latent space are output; among them, the latent vectors perform a compressed and abstract representation of the input data, capturing the important features of the input data and helping to learn the latent structure and variation rules of the data;

[0087] By maximizing the loss function, it is possible to minimize the Kullback-Leibler divergence (relative entropy or information divergence), thereby making the distribution of the latent vectors closer to the preset prior distribution, and also improving the quality of the generated data and the generalization ability of the model;

[0088] By processing the sub-graph features and global graph features through a variational autoencoder, it is possible to reduce the dimension of the feature space, reduce redundant information, and denoise the signal through the characteristics of the latent space, improving the quality of the data.

[0089] S32. Determine the sub-graph attention coefficient matrix according to the latent vectors in the n sub-graph latent spaces, and the steps are as follows:

[0090] S3211. Multiply the latent vectors in the n sub-graph latent spaces with the latent vectors in the n sub-graph latent spaces respectively to obtain an interaction matrix composed of n×n interaction features; corresponding Figure 3 to the process of product interaction of the outputs of the VAE encoder;

[0091] Specifically, multiply the latent vector in the first sub-graph latent space with the latent vectors in the first to n sub-graph latent spaces respectively, and so on. This way can effectively capture the mutual relationships and interactions between sub-graphs, enhancing the model's ability to model complex associations between sub-graphs;

[0092] S3212. Input the interaction matrix into a one-dimensional adaptive pooling layer (Adaptive Pooling 1D) for node feature compression to obtain a sub-graph compressed feature matrix of size n×1; corresponding Figure 3The process of one-dimensional adaptive pooling after product interaction; the n parameters therein respectively represent the node features regenerated for each subgraph; the one-dimensional adaptive pooling layer is used to perform an adaptive pooling operation on the input data, which is an existing structure and will not be elaborated here.

[0093] S3213. Pass the subgraph compressed feature matrix through the Sigmoid function to obtain an attention coefficient matrix a 11 , the attention coefficient matrix a 11 has a size of n×1;

[0094] S3214. Multiply the attention coefficient matrix a 11 by the transpose matrix of the attention coefficient matrix a 11 to obtain a subgraph attention coefficient matrix a1, and the subgraph attention coefficient matrix a1 has a size of n×n; the subgraph attention coefficient matrix a1 represents the feature relationship between each subgraph.

[0095] Determine the global graph attention coefficient matrix according to the latent vector of the global graph latent space, and the steps are as follows:

[0096] S3221. Input the latent vector Z q of the global graph latent space into the one-dimensional adaptive pooling layer for node feature compression to obtain a global graph compressed feature matrix with a size of n×1; the n parameters in the global graph compressed feature matrix respectively represent the n node features of the regenerated global graph; inputting the latent vector into the one-dimensional adaptive pooling layer can reduce the feature dimension and the computational complexity.

[0097] S3222. Pass the global graph compressed feature matrix through the Sigmoid function to obtain the global graph attention coefficient matrix a2, and the global graph attention coefficient matrix a2 has a size of n×1, a2 = [a 21 , a 22 , …, a 2n ]; inputting the features compressed by the one-dimensional adaptive pooling layer into the Sigmoid function can normalize the output value to the [0,1] interval and generate a low-dimensional representation with probabilistic interpretation, thereby reducing the complexity of the adaptive update of the subgraph features and the global graph features.

[0098] S33. Update the global graph features according to the subgraph attention coefficient matrix to obtain the updated global graph features, and the method is as follows:

[0099] The global graph includes a global graph Euclidean distance matrix and global graph node features; multiplying the subgraph attention coefficient matrix by the global graph Euclidean distance matrix to obtain an updated global graph Euclidean distance matrix; inputting the global graph features and the updated global graph Euclidean distance matrix into a graph convolutional network to obtain updated global graph features, and the calculation formula is as follows:

[0100]

[0101] Among them, H q represents the global graph features, H q ′ represents the updated global graph features, Relu represents the activation function, represents symmetric normalization operation on the updated global graph Euclidean distance matrix, D represents the degree matrix, a1 represents the subgraph attention coefficient matrix, A q represents the global graph Euclidean distance matrix, W q represents the weight matrix of the global graph;

[0102] Since the global graph has the greatest relevance to the task of engine fault detection, inputting the updated local graph Euclidean distance matrix and global graph features into the graph convolutional network can optimize the global features;

[0103] Multiplying the subgraph attention coefficient matrix by the local graph Euclidean distance matrix can adaptively update the node connection relationship of the global graph according to the node features of each subgraph. Inputting the global graph features and the updated global graph Euclidean distance matrix into the graph convolutional network can realize adaptive update of the node features of the global graph according to the features of each subgraph.

[0104] Updating the n subgraph features according to the global graph attention coefficient matrix to obtain subgraph aggregation features, and the method is as follows:

[0105] Multiplying and adding the global graph attention coefficient matrix by the n subgraph features to obtain subgraph aggregation features, and the calculation formula is as follows:

[0106]

[0107] Among them, H r represents the subgraph aggregation features, H i represents the subgraph features, i = 1, 2, …, n, a2 represents the global graph attention coefficient matrix, a2 = [a 21 , a 22 , …, a 2n .

[0108] Through the above method, the information in the data can be fully utilized to reduce the distribution difference of the graph structure between the subgraph and the whole graph.

[0109] In this embodiment, in step S4, the sub-graph aggregation feature and the updated global graph feature are input into the variational autoencoder I to obtain the shared latent space of the sub-graph and the global graph, and the latent vector of the shared latent space is extracted; extracting the latent vector of the latent space is prior art and will not be elaborated here.

[0110] The input of the global graph feature can provide richer context information, which helps the model to identify and filter out noise; while the input of the sub-graph feature can focus on the local area and reduce the influence of global noise on local features; inputting the sub-graph and the global graph into the same variational autoencoder can ensure that the latent representations of the global graph and the sub-graph are aligned in the same continuous latent space, enabling the global information to assist the local representation, making the local features more robust and less susceptible to noise, and at the same time allowing the local information to supplement the global features, making the global representation more delicate, so as to fully learn the data with a small amount of data.

[0111] In this embodiment, in step S5, classification is performed based on the latent vector of the shared latent space, and the classification result is the fault detection result of the engine.

[0112] An existing classification method is used to classify the latent vector, which is not limited here. Preferably, the latent vector of the shared latent space is input into a fully connected layer and then input into the Softmax function (normalized exponential function) for classification. This process is prior art and will not be elaborated here.

[0113] The latent vector of the shared latent space has fully learned the local information and global information in the engine time series data. Classifying based on the latent vector of the shared latent space can improve the accuracy of fault detection.

[0114] Correspondingly, the present invention also provides an electronic device, including:

[0115] A memory and a processor, where the memory is used to store a computer program, and when the computer program is executed by the processor, the above-mentioned engine fault detection method based on the shared latent space is implemented.

[0116] The device for executing the engine fault detection method based on the shared latent space may further include: an input device and an output device.

[0117] The processor, the memory, the input device and the output device can be connected through a bus or other means.

[0118] As a non-volatile computer-readable storage medium, the memory can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the program instructions / modules corresponding to the engine fault detection method based on the shared latent space in the embodiments of the present application. The processor executes various functional applications and data processing of the server by running the non-volatile software programs, instructions, and modules stored in the memory, that is, implements the engine fault detection method based on the shared latent space in the above method embodiments.

[0119] The memory may include a program storage area and a data storage area. Among them, the program storage area can store an operating system and application programs required for at least one function; the data storage area can store data created according to the use of the electronic device, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other non-volatile solid-state storage devices. In some embodiments, the memory may optionally include a memory remotely set relative to the processor, and these remote memories can be connected to the electronic device through a network. Examples of the above networks include, but are not limited to, the Internet, an enterprise intranet, a local area network, a mobile communication network, and combinations thereof.

[0120] The input device can receive input digital or character information, and generate signals related to the user settings and function control of the electronic device. The output device may include a display device such as a display screen.

[0121] The one or more modules are stored in the memory and, when executed by the one or more processors, execute the engine fault detection method based on the shared latent space in any of the above method embodiments.

[0122] The above product can execute the method provided in the embodiments of the present application, and has the corresponding functional modules and beneficial effects of the execution method. For technical details not described in detail in this embodiment, reference can be made to the method provided in the embodiments of the present application.

[0123] The electronic device in the embodiments of the present application exists in various forms, including but not limited to:

[0124] (1) Ultra-mobile personal computer devices: Such devices belong to the category of personal computers, have computing and processing functions, and generally also have the characteristic of mobile Internet access. Such terminals include: PDA, MID, and UMPC devices, etc.

[0125] (2) Other airborne electronic devices with data interaction functions, such as in-vehicle device installed on a vehicle.

[0126] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0127] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the essence of the above technical solution or the part that contributes to the related technology can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0128] Correspondingly, the present invention also provides a readable storage medium, in which one or more programs including execution instructions are stored. The execution instructions can be read and executed by an electronic device (including but not limited to a computer, a server, or a network device, etc.) to be used for executing the steps of any one of the above engine fault detection methods based on a shared latent space of this application.

[0129] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the purpose and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. An engine fault detection method based on a shared latent space, characterized in that: It includes the following steps: S1. Collect the sequential vibration data of the engine in real time, and construct a subgraph and a global graph according to the sequential vibration data; S2. Input the subgraph and the global graph into different graph convolutional networks respectively for feature extraction to obtain corresponding subgraph features and global graph features; S3. Update the subgraph features and the global graph features respectively according to a preset adaptive update strategy to obtain subgraph aggregation features and updated global graph features; S4. Input the subgraph aggregation features and the updated global graph features into a variational autoencoder Ⅰ to obtain a shared latent space of the subgraph and the global graph, and extract the latent vectors of the shared latent space; S5. Classify based on the latent vectors of the shared latent space, and the classification result is the fault detection result of the engine.

2. The engine fault detection method based on a shared latent space according to claim 1, wherein: The steps of updating the subgraph features and the global graph features through the adaptive update strategy to obtain subgraph aggregation features and updated global graph features are as follows: S31. There are n subgraph features in total. Input the n subgraph features and the global graph features into a multi-channel variational autoencoder and a variational autoencoder Ⅱ respectively to obtain corresponding latent vectors of n subgraph latent spaces and latent vectors of the global graph latent space; The multi-channel variational autoencoder is composed of n variational autoencoders; S32. Determine a subgraph attention coefficient matrix according to the latent vectors of the n subgraph latent spaces; Determine a global graph attention coefficient matrix according to the latent vectors of the global graph latent space; S33. Update the global graph features according to the subgraph attention coefficient matrix to obtain updated global graph features; Update the n subgraph features according to the global graph attention coefficient matrix to obtain subgraph aggregation features.

3. The engine fault detection method based on a shared latent space according to claim 2, wherein: Determine the subgraph attention coefficient matrix by the following method: S3211. Multiply the latent vectors of the n subgraph latent spaces with the latent vectors of the n subgraph latent spaces respectively to obtain an interaction matrix composed of n×n interaction features; S3212. Input the interaction matrix into a one-dimensional adaptive pooling layer for node feature compression to obtain a subgraph compression feature matrix; S3213. Pass the sub-graph compression feature matrix through the Sigmoid function to obtain an attention coefficient matrix a 11 ; S3214. Multiply the attention coefficient matrix a 11 by the transposed matrix of the attention coefficient matrix a 11 to obtain a sub-graph attention coefficient matrix a1.

4. The method for engine fault detection based on a shared latent space according to claim 2 or 3, characterized in that: Update the global graph features by the following method to obtain updated global graph features: The global graph includes a global graph Euclidean distance matrix and global graph node features; Multiply the subgraph attention coefficient matrix with the global graph Euclidean distance matrix to obtain an updated global graph Euclidean distance matrix; Input the global graph features and the updated global graph Euclidean distance matrix into a graph convolutional network to obtain updated global graph features.

5. The engine fault detection method based on a shared latent space according to claim 2, wherein: Determine the global graph attention coefficient matrix by the following method: S3221. Input the latent vectors of the global graph latent space into a one-dimensional adaptive pooling layer for node feature compression to obtain a global graph compression feature matrix; S3222. Pass the global graph compression feature matrix through the Sigmoid function to obtain a global graph attention coefficient matrix a2.

6. The engine fault detection method based on a shared latent space according to claim 2 or 5, characterized in that: Update the subgraph features by the following method to obtain subgraph aggregation features: Multiply and add the global graph attention coefficient matrix with the n sub-graph features to obtain the sub-graph aggregation feature. The calculation formula is as follows: Among them, H r represents the subgraph aggregation feature, and H i represents the subgraph feature, where i = 1, 2, …, n, and a2 represents the global graph attention coefficient matrix, a2 = [a 21 , a 22 , …, a 2n .

7. The method for engine fault detection based on a shared latent space according to claim 2, characterized in that: Both the multi-channel variational autoencoder and the variational autoencoder II adopt the following loss function: Among them, represents the loss function, θ represents the decoder parameters, represents the encoder parameters, x represents the true subgraph features or the true global graph features, p θ (x|z) represents the true posterior distribution of the latent vectors in the subgraph latent space or the true posterior distribution of the latent vectors in the global graph latent space. The general expression form of the hidden vector is z = μ + εσ, where μ and σ represent the outputs of the encoder, ε represents a random number that conforms to the standard normal distribution, J represents the total number of dimensions of the latent space, j represents the index of the latent variable dimension, G represents the total number of Monte Carlo samplings, and g represents the index of the Monte Carlo sampling times; When the loss function is at its maximum, output the latent vectors of the latent spaces of n subgraphs or the latent vectors of the global graph latent space.

8. The method for engine fault detection based on a shared latent space according to claim 1, wherein: Construct sub-graphs and a global graph through the following method: Divide the collected time-series vibration data into n segments according to a preset time window size; Construct sub-graphs: Determine n sub-graphs based on the n segments of time-series vibration data; Take the number of vibration data in the i-th segment of time-series vibration data as the number of subgraph nodes, where i = 1, 2, …, n, and take the vibration data in the i-th segment of time-series vibration data as the subgraph node feature X i ; Calculate the Euclidean distance between each pair of subgraph node features, and combine the Euclidean distances between each pair of subgraph node features into a Euclidean distance matrix A i , where the Euclidean distance matrix A i is a symmetric matrix; Combine the Euclidean distance matrix A i and the subgraph node feature X i to obtain the subgraph G i =(A i , X i ); The construction methods of the remaining n - 1 subgraphs are similar Construct a global graph: Determine a global graph based on the n segments of time-series vibration data; Take each time window segment as a global graph node, and take the vibration data within each segment of time series vibration data as the global graph node feature X q ; Calculate the Euclidean distances between n global graph nodes and combine them into a Euclidean distance matrix A q , where the Euclidean distance matrix A q is a symmetric matrix; Combine the Euclidean distance matrix A q with the global graph node features to obtain the global graph G q = (A q , X q ), X q = [X1, X2, …, X n .

9. An electronic device, characterized in that: including: A memory and a processor, where the memory is used to store a computer program, and when the computer program is executed by the processor, it implements the engine fault detection method based on a shared latent space according to any one of claims 1-8.

10. A readable storage medium, characterized in that: Computer instructions are stored in the readable storage medium, and when the computer instructions are executed by a processor, they implement the engine fault detection method based on a shared latent space according to any one of claims 1-8.

Images