Lifting system fault diagnosis method based on depth map convolutional generative adversarial network

Abstract

The invention discloses a hoisting system fault diagnosis method based on a depth map convolutional generative adversarial network, and the method comprises the steps: collecting a vibration signal from a hoisting system under a complex working condition, and selecting a fault data set as a training and testing sample; setting and initializing internal parameters of a depth map convolutional generative adversarial network algorithm; performing fine tuning on the depth map convolutional generative adversarial network algorithm by using a training sample to obtain an optimal parameter; performing small sample fault diagnosis on the hoisting system; and generating a fault detection result, evaluating the diagnosis effect of the depth map convolutional generative adversarial network algorithm, and verifying the fault diagnosis of the depth map convolutional generative adversarial network algorithm. The small sample data can be generated and enhanced conveniently, and effective execution of small sample fault diagnosis of the hoisting system under complex working conditions is improved.

Description

Technical Field

[0001] The present invention relates to the field of hoisting system fault diagnosis, and in particular to a hoisting system fault diagnosis method based on a deep graph convolution generative adversarial network. Background Art

[0002] As a crucial mechanical equipment in automated production, the lifting system may suffer minor damage in harsh working environments, resulting in system failures. Therefore, it is crucial to develop accurate intelligent fault diagnosis methods. The fault diagnosis method based on deep learning significantly improves the diagnosis efficiency through automatic feature extraction and pattern recognition, but it requires a large amount of training data and high-performance hardware, resulting in reduced diagnostic accuracy. Since mechanical systems have fewer failures under normal operating conditions and data collection is difficult, existing methods perform poorly in small sample conditions and cannot effectively meet the high-precision diagnosis needs under limited data conditions.

[0003] For small sample data scenarios, current algorithms such as domain adaptive algorithms (DAA), convolutional neural networks (CNN), graph neural networks (GNN) and other methods, but these methods generally face problems such as overfitting, insufficient feature extraction capabilities and high computational complexity, resulting in reduced diagnostic accuracy under small amounts of data. In addition, although advanced small sample learning methods such as transfer learning (TL) and meta-learning (ML) can help solve the problem of data scarcity, they are often limited by problems such as label scarcity, model complexity and long training time, and are difficult to cope with challenges in practical applications. Summary of the invention

[0004] In order to overcome the shortcomings and deficiencies of the prior art, the present invention provides a lifting system fault diagnosis method based on a deep graph convolutional generative adversarial network.

[0005] The technical solution adopted by the present invention is a crane system fault diagnosis method based on a deep graph convolution generative adversarial network, the method comprising:

[0006] Step S1: collecting vibration signals from the lifting system under complex working conditions, and selecting fault data sets as training and testing samples;

[0007] Step S2: setting and initializing the internal parameters of the deep graph convolutional generative adversarial network algorithm, including the connectivity coefficient α, the interaction coefficient β, the mean μ and the variance σ;

[0008] Step S3: fine-tuning the deep graph convolutional generative adversarial network algorithm using training samples to obtain optimal parameters;

[0009] Step S4: Then input the test sample into the trained deep graph convolutional generative adversarial network algorithm to perform small sample fault diagnosis on the lifting system;

[0010] Step S5: generating fault detection results and evaluating the diagnostic effect of the deep graph convolutional generative adversarial network algorithm;

[0011] Step 6: Verify the fault diagnosis of the deep graph convolutional generative adversarial network algorithm by comparing it with existing fault diagnosis methods and small sample learning methods.

[0012] Furthermore, the deep graph convolutional generative adversarial network algorithm includes: reconstruction of graph structure, establishment of channel information interaction mechanism, optimization of generation model, optimization of discrimination model, and fault classification using softmax function;

[0013] Furthermore, the reconstruction of the graph structure: using the features of the nodes to reconstruct the adjacency matrix of the graph, and using the k-nearest neighbor algorithm to determine similar node connections based on the similarity between the node features, for a node v on the graph i (i=1,2,…,n), original node feature x i and x j With another node v j The cosine similarity expression of (j=1,2,…,i-1,i+1,…,n) is:

[0014]

[0015] Among them, s ij represents the cosine similarity between node i and node j, r i and r j Represents the feature vectors of node i and node j respectively, r i T r j is the vector r i and r j The dot product between ||r i || 2 Represents vector r i L 2 norm, ||r j || 2 Represents vector r j L 2 Norm.

[0016] Connect each node to its k most similar nodes, generating a graph structure containing a new adjacency matrix.

[0017] Furthermore, the channel information interaction mechanism is established, which allows the initial graph signal to be retrieved from the original and current branches, and is used to improve the interaction of features between different branches. The H 0 and H' 0, in a network with a total number of layers of l, the feature interaction process expression of the lth (l = 1, 2, ..., L) layer is:

[0018]

[0019] Among them, H l is the feature representation of the lth layer, δ represents the activation function, which can perform nonlinear transformation on the transmitted features; P represents the adjacency matrix involved in the graph convolution, α is the initial residual connectivity coefficient, β is the channel information interaction coefficient, and H l-1 is the feature representation of the l-1th layer, H 0 is the initial feature representation, represents the weight matrix of the lth layer. In addition, H′ l , P′, H′ 0 and They are respectively l , P, H 0 as well as The corresponding feature interaction values ​​and the updated weight matrix are expressed as follows:

[0020]

[0021] in, and is the updated weight matrix, W l and W′ l are the weight matrices of the lth layer in the two branches, I n is the n-order identity matrix, γ l is a parameter that increases with the number of network layers, and its expression is:

[0022]

[0023] Among them, θ = 0.5 means control γ l The hyperparameter of the incremental rate is expressed as follows:

[0024] out=f 3 (H L )+f 4 (H′ L )

[0025] Among them, out represents the final output of the network, f 3 and f 4 are two linear transformations, H L and H′ L Represents two different feature representations obtained at layer L.

[0026] Furthermore, the optimization of the generative model is as follows: the generator encodes the collected real image, encodes the real image in the input stage and extracts its features, and for each mean and variance obtained, introduces noise with normal distribution to make the extracted features of the real image closer to the feature vector that satisfies the normal distribution, and then sends the obtained feature vector to the generator to generate an image signal; x r As the real sample of the input image, its mean μ and variance σ are obtained by the encoder, and the standard normal distribution is used to construct the randomized feature vector c, whose mathematical expression is:

[0027]

[0028] in, represents element-by-element multiplication, μ and σ represent the mean and variance respectively, c is the final generated eigenvector, and N(0,I) represents a multidimensional normal distribution with a mean of 0 and a covariance matrix of the identity matrix I.

[0029] Represents the sample obtained by the generator, and the output of the discriminator can be expressed as y∈[0,1], where the feature vector generated by the generator can be calculated by the following formula:

[0030]

[0031] Among them, G represents the generator, represents the sample obtained by the generator, g is the generating function, θ g is the generator parameter, z represents the latent variable, and X are the generated image and original image obtained by the discriminator respectively, and the expression is:

[0032]

[0033] The output obtained by the discriminator can be expressed as:

[0034]

[0035] D(x) represents the true sample distribution of the image, and the error function expression is:

[0036]

[0037] Among them, θ d is the parameter of the discriminator, and p(x) is the true data distribution of sample x, that is, the probability distribution of the true data. The samples generated by the generator, To generate data The probability distribution of . represents the distribution of the input generator noise, the internal parameters θ of the discriminator d The expression is:

[0038] θ d =(θ f ,θ c )

[0039] Among them, θ f is the parameter of the feature extraction part of the discriminator, θ c are the parameters of the classification part of the discriminator.

[0040] By maximizing the entropy of the noise distribution, the generated sample image is constantly close to the real sample image, and the network finally reaches a balanced expression:

[0041]

[0042] Among them, max represents the maximization operation, θ g is the generator parameter, is the output of the discriminator, representing the generated data The probability of belonging to the real data, To generate data The probability distribution of . The larger the value of , the better the fitting ability of the generated network, and the higher the similarity between the generated image and the real image, which indicates that the network proposed in the present invention has the best image generation performance.

[0043] Furthermore, the optimization of the discriminant model: let p data (x) and p g (x) represent the real samples and image samples output by the optimized model respectively. The loss function of the deep graph convolutional generative adversarial network algorithm consists of the discriminant network and the generative network, which is expressed as:

[0044]

[0045] Among them, p data (x) and p g (x) During the training process, the two networks are infinitely close to each other, and the generative network and the discriminative network are interactively iterated. When training the discriminative network, the real image is mixed with the generated image and input into the discriminator, and the parameters of the network are adjusted according to the output of the discriminator. At the same time, when a set of data is randomly input, the generative network generates an image, and the discriminator gives a discriminative result. The difference between the generated image and the real image is calculated and fed back to the generative network to adjust the parameters, and interactive iteration is performed until the network reaches stability. The optimization of the discriminative network is achieved by minimizing the cross entropy operation. The loss function expression is:

[0046]

[0047] The optimal solution expression of the loss function is:

[0048]

[0049] Where D(x) represents the predicted output of the discriminator D for the input data x, p data (x) and p g (x) represent the real samples and image samples output by the optimized model, respectively, and p g represents the probability of generating data, D * (x) is the optimal solution of the loss function. data (x) and p g In actual training, the value of (x) is a real number between (0,1). When the generator network is fixed, the discriminator network has the optimal solution and obtains the best performance; the loss function is updated as:

[0050]

[0051] Among them, V(D,G) represents the loss function of the discriminator D and the generator G, It is the output of the discriminator D for the input x, indicating the probability that it is judged to be true data. data (x) and p g (x) represent the real samples and image samples output by the optimized model, respectively, and p z is the noise distribution, Denotes the measure distribution p data (x) and p g (x), namely, KL divergence. data (x) = p g (x), the global minimum can be calculated.

[0052] Furthermore, the softmax function is used for fault classification: the input samples are normalized using a whitening preprocessing method, and the expression is:

[0053]

[0054] Among them, E[x(k)] represents the average value of each batch of training data, and two learning parameters γ and ξ are used for inverse transformation to ensure the accuracy and stability of feature learning. The expression is

[0055]

[0056] Finally, the softmax function is used to calculate the feature weight of the scale, and the expression is:

[0057]

[0058] in, The output classification probability calculation expression under each feature weight is:

[0059]

[0060] Among them, P t (x) j is the predicted probability of the jth class given an input x, η(k) is the weight associated with class k, and x j represents the features of input x under category j, K is the total number of categories, Used to automatically update P in the training and test datasets t (x) j .

[0061] Beneficial effects:

[0062] This application proposes a crane system fault diagnosis method based on deep graph convolution generative adversarial network. The present invention designs a new deep graph convolution generative adversarial network algorithm to effectively obtain all small sample fault information, thereby facilitating the generation and enhancement of small sample data. This method improves the effective execution of small sample fault diagnosis of crane systems under complex working conditions. The details are as follows:

[0063] 1. Small sample fault information acquisition: The present invention can effectively acquire small sample fault information, and utilize the generation capability of generative adversarial networks to enhance small sample data to make up for the problem of data scarcity. This method can not only help the model learn more fault features, but also avoid overfitting problems caused by insufficient data, thus improving the accuracy of diagnosis.

[0064] 2. Solve the overfitting problem: The present invention uses the adversarial training mechanism of the generative adversarial network to enable the model to learn more general features under limited samples, avoiding excessive reliance on training data, thereby improving the accuracy and robustness of the model in practical applications.

[0065] 3. Enhanced feature extraction capability: The present invention can not only extract deep-level features implicit in the graph structure, but also enhance the model’s ability to recognize fault modes through the graph convolution layer, thereby improving diagnostic accuracy.

[0066] 4. Improve diagnostic efficiency under complex working conditions: By combining the advantages of deep graph convolution and generative adversarial networks, the present invention can not only effectively deal with fault diagnosis problems under complex working conditions, but also improve performance in complex environments by enhancing learned features.

[0067] 5. Reduce computational complexity: The present invention generates more data samples by introducing a generative adversarial network, reducing the dependence on a large amount of original data. At the same time, it reduces the computational complexity through graph convolution and improves the efficiency of model training.

[0068] 6. Addressing the problem of label scarcity: This invention uses a generative adversarial network to expand the training set by synthesizing new data samples without relying on a large amount of labeled data, thereby alleviating the problem caused by label scarcity. This enables the model to maintain a high diagnostic accuracy with a small number of labels. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is a general step flow chart of the present invention;

[0070] Figure 2 The architecture diagram of the deep graph convolution generative adversarial network algorithm of the present invention. DETAILED DESCRIPTION

[0071] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application may be combined with each other. The present application is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0072] The present invention proposes a hoisting system fault diagnosis method based on deep graph convolution generative adversarial network to solve the problems of low fault diagnosis accuracy and poor diagnostic efficiency of the current hoisting system under complex working conditions, especially under small sample data conditions. Through the innovative deep graph convolution generative adversarial network algorithm, the effective capture of small sample fault information is achieved, and the generation capability of small sample data is enhanced, thereby improving the accuracy and robustness of hoisting system fault diagnosis under complex working conditions.

[0073] like Figure 1 As shown, a crane system fault diagnosis method based on deep graph convolution generative adversarial network includes:

[0074] Step S1: collecting vibration signals from the lifting system under complex working conditions, and selecting fault data sets as training and testing samples;

[0075] Specifically, this step includes first collecting vibration signals from different parts of the lifting system through sensors (such as accelerometers, vibration sensors, etc.). Since the lifting system usually works in a harsh environment, the collected data may contain a variety of noise and interference, so the signal needs to be preprocessed to remove noise and ensure the accuracy of the data. Next, select representative fault data sets from these vibration signals as samples for training and testing. Fault data should include different types of failure modes (such as bearing failure, gear failure, etc.) and cover different working conditions to improve the generalization ability of the model. According to the diversity of the data, it is divided into training set and test set.

[0076] Step S2: setting and initializing the internal parameters of the deep graph convolutional generative adversarial network algorithm, including the connectivity coefficient α, the interaction coefficient β, the mean μ and the variance σ;

[0077] Specifically, this step involves first setting a series of internal parameters for the Deep Graph Convolutional Generative Adversarial Network (DGC-GAN). These parameters include the connectivity coefficient α, which is used to control the feature transfer between the generator and the discriminator; the interaction coefficient β, which is used to adjust the degree of feature interaction between different parts of the network; the mean μ and variance σ, which are used to standardize the input data. Parameter initialization is a very important step in the training process. Appropriate initialization can speed up the convergence of the network and avoid gradient vanishing or gradient exploding.

[0078] Step S3: fine-tuning the deep graph convolutional generative adversarial network algorithm using training samples to obtain optimal parameters;

[0079] Specifically, this step includes fine-tuning the deep graph convolutional generative adversarial network algorithm using the training samples collected in step S1. This process continuously optimizes the parameters in the network (such as convolution kernels, weight matrices, etc.) through the back-propagation algorithm, so that the generator can generate more realistic pseudo data and the discriminator can more accurately distinguish between real and generated data. During the fine-tuning process, the network will gradually adjust the parameters according to the feedback of the loss function until the loss is minimized. At this stage, parameter updates are achieved through the gradient descent method, and the optimization goal is to improve the performance of the generator and the discriminator.

[0080] Step S4: Then input the test sample into the trained deep graph convolutional generative adversarial network algorithm to perform small sample fault diagnosis on the lifting system;

[0081] Specifically, this step includes inputting the test sample selected in step S1 into the trained deep graph convolutional generative adversarial network. Through the trained network model, the system performs small sample fault diagnosis on the input test sample. Since the generative adversarial network has the ability to enhance small sample data, even in the test phase, when the number of samples is small, the network can still obtain high-quality diagnostic results from limited training data. The network extracts deep features in the graph structure through graph convolution operations, and combines the generated sample data for accurate fault identification and classification.

[0082] Step S5: generating fault detection results and evaluating the diagnostic effect of the deep graph convolutional generative adversarial network algorithm;

[0083] Specifically, this step includes that after being processed by the deep graph convolution generative adversarial network, the system will generate a fault detection result. The result is usually a diagnostic label or a set of predicted values ​​of fault modes, such as "normal", "minor fault", "serious fault", etc. Next, the effect of the algorithm is evaluated based on the diagnostic results, mainly by comparing the actual fault situation with the predicted results output by the network, and calculating performance indicators such as diagnostic accuracy, precision, and recall. This step is an evaluation of the network's diagnostic capabilities, which is used to verify whether the algorithm can accurately identify different types of faults and has a high diagnostic efficiency.

[0084] Step S6: The fault diagnosis of the deep graph convolutional generative adversarial network algorithm is verified by comparing it with the existing fault diagnosis method (the specific diagnosis method should be clarified here) and the small sample learning method.

[0085] Specifically, this step includes comparing and verifying the diagnosis results of the deep graph convolutional generative adversarial network algorithm with existing fault diagnosis methods and small sample learning methods. Existing fault diagnosis methods (such as: distance probability graph convolutional network, multi-scale deep graph convolutional network, multi-channel residual multi-scale clustering graph convolutional network, and multi-channel domain adaptive graph convolutional network) and small sample learning methods (such as: model agnostic meta-learning, task ranking meta-learning, attention-based deep meta-transfer learning, and meta-learning generative adversarial network). By comparing the performance of different methods, the advantages and effects of the proposed algorithm under small sample conditions can be verified. This verification process includes the performance of the model on different data sets and its applicability in practical applications to ensure that the algorithm has strong promotion capabilities and practical value.

[0086] The deep graph convolution generative adversarial network algorithm includes: reconstruction of graph structure, establishment of channel information interaction mechanism, optimization of generation model, optimization of discrimination model, and fault classification using softmax function.

[0087] like Figure 2As shown in the figure, in the data generation stage: the generator G is responsible for generating "pseudo data" similar to real data. The generator receives random noise from the input and generates pseudo data similar to the target data through the network. The task of the discriminator D is to distinguish between the generated data and the real data. It will determine whether the input data is real and give a "true" or "false" result. The discriminator optimizes its parameters to maximize the ability to discriminate the generated data. Adversarial training is performed through the generator and the discriminator. The generator optimizes how to generate more realistic data, and the discriminator optimizes how to distinguish true and false data more accurately. Finally, the generator and the discriminator continue to play against each other until the data generated by the generator is as close to the real data as possible and the accuracy of the discriminator is as high as possible. The generator uses noise as input and generates pseudo data through multiple layers of convolution and nonlinear transformation. In the deep graph convolutional network stage: the input graph data is sent to the deep graph convolutional network for processing. The graph convolutional network extracts the features of the graph structure by learning nodes and edges. The convolutional layer performs convolution operations on the input data through a sliding window. In graph convolution, the convolution kernel is convolved with the features of the node and its neighbors to extract the features of the local graph structure. The parameters of each convolution operation are learned through training. Pooling layers are used in graph convolution networks for downsampling, reducing the dimension of features and retaining important feature information. Each convolution block is followed by a global maximum pooling layer to select the most significant features in the entire graph. Each layer of the graph convolution network will convolve each node in the graph through the connection relationship between nodes to learn deeper graph features. After stacking multiple layers of graph convolution layers, more and more complex graph structure features can be extracted. The final convolution output will be further passed to the next stage. In the fusion stage: different features are combined through fusion operations, and the outputs of multiple convolution layers are fused through 1×1 convolutions (the number of features for each convolution f=1024). This convolution operation can maintain the diversity of features while reducing the amount of calculation. In the prediction stage: the features after graph convolution and fusion are passed to the fully connected layer for further processing, and finally the prediction results are obtained. Continue to use 1×1 convolutions for feature extraction, gradually reducing the feature dimension until the final output is obtained. Finally, the prediction result of the output layer is the final prediction of fault diagnosis. After calculation by the entire network, the output indicates the fault type or system status of each node.

[0088] In order to accurately extract a small amount of feature information and improve the fault diagnosis accuracy of the lifting system, this paper proposes a new deep learning architecture - deep graph convolution generative adversarial network. Two different nonlinear fully connected layers are input to obtain the initial features of the graph structure branch and the node similarity branch, which are used as the initial input of the deep graph convolution module and the channel information interaction mechanism, as shown below:

[0089] H 0 =δ(f1 (R)),H′ 0 =δ(f 2 (R))(1)

[0090] Among them, f 1 and f 2 are two linear transformations, δ(·) is the activation function, and the dimension of the initial feature R can be compressed to a smaller value through the nonlinear transformation layer, thereby greatly reducing the number of network parameters without affecting the classification accuracy. At the same time, the gradient descent method is used to calculate the loss function to achieve the optimization of the generation model and the discriminant model. Compared with the existing graph convolutional network, the deep graph convolutional generative adversarial network designed in the present invention shows a stronger ability in extracting more comprehensive high-order semantic features, thereby improving the fault classification accuracy. The specific construction process of the deep graph convolutional generative adversarial network framework proposed in the present invention is as follows:

[0091] Reconstruction of graph structure: Existing diagnostic methods based on graph convolutional networks mainly consider the topological structure of the graph, resulting in a large number of low-dimensional nodes and limited neighboring nodes, which hinders the acquisition of sufficient information during feature aggregation. In addition, multiple graph convolutions will destroy the structural information between similar nodes, resulting in the convergence of node attributes and the loss of feature diversity. In order to overcome these problems, the present invention uses the features of the nodes to reconstruct the adjacency matrix of the graph, and uses the k-nearest neighbor algorithm to determine similar node connections based on the similarity between node features. For a node v on the graph, i (i=1,2,…,n), its original node feature x i and x j With another node v j The cosine similarity of (j=1,2,…,i-1,i+1,…,n) can be expressed as:

[0092]

[0093] Among them, s ij represents the cosine similarity between node i and node j, r i and r j Represents the feature vectors of node i and node j respectively, r i T r j is the vector r i and r j The dot product between ||r i || 2 Represents vector r i L 2 norm, ||r j || 2 Represents vector r j L 2 Norm.

[0094] Accordingly, each node is connected to its k most similar nodes to generate a graph structure containing a new adjacency matrix, which is used to overcome the problem that the traditional graph convolutional network method destroys the node similarity of the original graph structure information. Therefore, the deep graph convolutional generative adversarial network proposed in the present invention can extract high-order semantic feature information and obtain more accurate node representation.

[0095] Establishing a channel information interaction mechanism: In order to maintain the diversity of node features, the present invention designs a channel information interaction mechanism and incorporates the residual connections between branches. This mechanism allows the retrieval of initial graph signals from the original and current branches, thereby improving the interaction of features between different branches. Based on H calculated by formula (1), 0 and H′ 0 , in a network with a total number of layers of l, the feature interaction process of the lth (l = 1, 2, ..., L) layer can be expressed as:

[0096]

[0097] Among them, H l is the feature representation of the lth layer, δ represents the activation function, which can perform nonlinear transformation on the transmitted features; P represents the adjacency matrix involved in the graph convolution, α is the initial residual connectivity coefficient, β is the channel information interaction coefficient, and H l-1 is the feature representation of the l-1th layer, H 0 is the initial feature representation, represents the weight matrix of the lth layer. In addition, H′ l , P′, H′ 0 and They are respectively l , P, H 0 as well as The corresponding feature interaction values ​​and the updated weight matrix are expressed as follows:

[0098]

[0099] Where α is the initial residual connectivity coefficient, In is the n-order unit matrix, β is the channel information interaction coefficient, and W l and W′ l are the weight matrices of the lth layer in the two branches respectively. l It is a parameter that increases with the number of network layers, and its calculation formula is as follows:

[0100]

[0101] Among them, θ = 0.5 is the control γ lIn order to reasonably fuse channel feature information, a linear feature fusion method is designed, which can adaptively fuse structural information features and similarity information features into the new node representation. The specific calculation equation is:

[0102] out=f 3 (H L )+f 4 (H′ L )(8)

[0103] Among them, out represents the final output of the network, f 3 and f 4 are two linear transformations, H L and H′ L Represents two different feature representations obtained at layer L.

[0104] Optimization of the generative model: The generator encodes the real image collected. In order to ensure the generation of high-quality images, the real image is encoded and its features are extracted at the input stage. For each mean and variance obtained, noise with normal distribution is introduced to make it closer to the feature vector that satisfies the normal distribution. Then the obtained feature vector is sent to the generator to generate the image signal. r As the real sample of the input image, its mean μ and variance σ are obtained by the encoder, and the standard normal distribution is used to construct the randomized feature vector c, whose mathematical expression is:

[0105]

[0106] in, represents element-by-element multiplication, μ and σ represent the mean and variance respectively, c is the final generated eigenvector, and N(0,I) represents a multidimensional normal distribution with a mean of 0 and a covariance matrix of the identity matrix I.

[0107] in addition, Represents the sample obtained by the generator, and the output of the discriminator can be expressed as y∈[0,1], where the feature vector generated by the generator can be calculated by the following formula:

[0108]

[0109] Among them, G represents the generator, represents the sample obtained by the generator, g is the generating function, θ g is the generator parameter, z represents the latent variable, and X are the generated image and original image obtained by the discriminator respectively, and the expression is:

[0110]

[0111] Then, the output obtained by the discriminator can be expressed as:

[0112]

[0113] Among them, D(x) represents the true sample distribution of the image, and its error function can be calculated as:

[0114]

[0115] Among them, θ d is the parameter of the discriminator, and p(x) is the true data distribution of sample x, that is, the probability distribution of the true data. The samples generated by the generator, To generate data The probability distribution of . represents the distribution of the input generator noise, the internal parameters θ of the discriminator d The expression is:

[0116] θ d =(θ f ,θ c ) (14)

[0117] Among them, θ f is the parameter of the feature extraction part of the discriminator, θ c are the parameters of the classification part of the discriminator.

[0118] By maximizing the entropy of the noise distribution, the generated sample image is constantly close to the real sample image, and the network finally reaches a balance. The specific formula is as follows:

[0119]

[0120] Among them, max represents the maximization operation, θ g is the generator parameter, is the output of the discriminator, representing the generated data The probability of belonging to the real data, To generate data The probability distribution of . The larger the value of , the better the fitting ability of the generated network, and the higher the similarity between the generated image and the real image, which indicates that the network proposed in the present invention has the best image generation performance.

[0121] Optimization of the discriminant model: Let p data (x) and p g (x) represent the real samples and image samples output by the optimized model respectively. Therefore, the loss function of the deep graph convolutional generative adversarial network algorithm proposed in the present invention is composed of the discriminant network and the generative network, which is expressed as:

[0122]

[0123] Among them, D(x) represents the predicted output of the discriminator D for the input data x, p data (x) and p g (x) represent the real samples and image samples output by the optimized model, respectively, and p g represents the probability of generating data, D * (x) is the optimal solution of the loss function. data (x) and p g (x) are infinitely close to each other during the training process, and the generative network and the discriminative network are interactively iterated in the process of optimizing formula (16). When training the discriminative network, the real image is mixed with the generated image and input into the discriminator, and the parameters of the network are adjusted according to the output of the discriminator. At the same time, when a set of data is randomly input, the generative network generates an image and the discriminator gives a discriminative result. The difference between the generated image and the real image is calculated and fed back to the generative network to adjust the parameters, thereby interactively iterating until the network reaches stability. The optimization of the discriminative network is achieved by minimizing the cross entropy operation, and its loss function is described as:

[0124]

[0125] Therefore, the optimal solution of formula (17) can be expressed as:

[0126]

[0127] Among them, p data (x) and p g In actual training, the value of (x) is a real number between (0,1). When the generator network is fixed, the discriminator network has the optimal solution and obtains the best performance. Therefore, formula (17) can be updated as:

[0128]

[0129] Among them, V(D,G) represents the loss function of the discriminator D and the generator G, It is the output of the discriminator D for the input x, indicating the probability that it is judged to be true data. data (x) and p g (x) represent the real samples and image samples output by the optimized model, respectively, and p z is the noise distribution, Denotes the measure distribution p data (x) and p g (x), namely, KL divergence. data (x) = p g(x), the global minimum can be calculated.

[0130] Fault classification using softmax function: The input samples are normalized using the whitening preprocessing method, which is specifically defined as follows:

[0131]

[0132] In the formula, E[x (k) ] represents the average value of each batch of training data, and two learning parameters γ and ξ are used for inverse transformation, thereby ensuring the accuracy and stability of feature learning, namely:

[0133]

[0134] The softmax function is used to calculate the feature weight of the scale, as shown in formula (22):

[0135]

[0136] In the formula, The output classification probability under each feature weight is calculated by the following formula (23):

[0137]

[0138] Among them, P t (x) j is the predicted probability of the jth class given an input x, η(k) is the weight associated with class k, and x j represents the features of input x under category j, K is the total number of categories, Used to automatically update P in the training and test datasets t (x) j .

[0139] This application effectively obtains all small sample fault information, thereby facilitating the generation and enhancement of small sample data and improving the effective execution of small sample fault diagnosis of the lifting system under complex working conditions. Specifically including:

[0140] 1. Acquisition of small sample fault information: Traditional deep learning methods often rely on a large amount of training data. However, in mechanical equipment such as lifting systems, it is difficult to obtain fault data, especially in normal operation, when there are fewer fault samples. The method based on deep graph convolution generative adversarial network proposed in this application can effectively obtain small sample fault information, and utilize the generation ability of generative adversarial network to enhance small sample data to make up for the problem of data scarcity. This method can not only help the model learn more fault features, but also avoid overfitting problems caused by insufficient data, thereby improving the accuracy of diagnosis.

[0141] 2. Solve the overfitting problem: Existing methods often encounter overfitting problems in small sample data scenarios, especially in deep neural networks, where the model tends to remember the training data but does not have good generalization capabilities. The method of this application uses the adversarial training mechanism of the generative adversarial network to encourage the model to learn more general features under limited samples, avoiding excessive dependence on training data, thereby improving the accuracy and robustness of the model in practical applications.

[0142] 3. Enhanced feature extraction capabilities: Although traditional convolutional neural networks (CNNs) have achieved remarkable results in the field of image processing, they have certain limitations when processing graph-structured data. Deep graph convolutional networks can effectively process graph-structured data and use the connection relationship between nodes to learn features at a deeper level. Combined with generative adversarial networks, the method of this application can not only extract deep features implicit in the graph structure, but also enhance the model's ability to recognize fault modes through graph convolution layers, thereby improving diagnostic accuracy.

[0143] 4. Improve diagnostic efficiency under complex working conditions: The lifting system may encounter multiple failure modes under complex working conditions, and due to changes in environmental factors, the failure manifestations are highly nonlinear and complex. The method proposed in this application combines the advantages of deep graph convolution and generative adversarial networks, which can not only effectively deal with the fault diagnosis problem under complex working conditions, but also improve the performance in complex environments by enhancing the learned features.

[0144] 5. Reduce computational complexity: Although many deep learning methods have strong feature extraction capabilities, their computational complexity and training time are often very high. The method of this application generates more data samples by introducing a generative adversarial network, reducing the reliance on a large amount of raw data. At the same time, it reduces computational complexity through graph convolution and improves the efficiency of model training.

[0145] 6. Dealing with label scarcity: Label scarcity is a common problem in small sample learning, especially in mechanical equipment fault diagnosis, where the cost of obtaining labels is high. The solution of this application uses a generative adversarial network to expand the training set by synthesizing new data samples without relying on a large amount of labeled data, thereby alleviating the problem caused by label scarcity. This enables the model to maintain a high diagnostic accuracy with a small number of labels.

[0146] Through these innovative methods, the fault diagnosis solution of the present application not only overcomes the problems mentioned in the background technology, but also can provide more accurate and efficient fault diagnosis capabilities in practical applications, meeting the high-precision diagnosis requirements of complex mechanical equipment such as lifting systems under small sample data conditions.

[0147] Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A crane system fault diagnosis method based on deep graph convolutional generative adversarial network, characterized by , the method comprises: Step S1: collecting vibration signals from the lifting system under complex working conditions, and selecting fault data sets as training and testing samples; Step S2: setting and initializing the internal parameters of the deep graph convolutional generative adversarial network algorithm, including the connectivity coefficient α, the interaction coefficient β, the mean μ and the variance σ; Step S3: fine-tuning the deep graph convolutional generative adversarial network algorithm using training samples to obtain optimal parameters; Step S4: Then input the test sample into the trained deep graph convolutional generative adversarial network algorithm to perform small sample fault diagnosis on the lifting system; Step S5: generating fault detection results and evaluating the diagnostic effect of the deep graph convolutional generative adversarial network algorithm; Step S6: Verify the fault diagnosis of the deep graph convolutional generative adversarial network algorithm by comparing it with the existing fault diagnosis method and small sample learning method.

2. The method for diagnosing a hoisting system fault based on a deep graph convolutional generative adversarial network according to claim 1, characterized in that: The deep graph convolution generative adversarial network algorithm includes: reconstruction of graph structure, establishment of channel information interaction mechanism, optimization of generation model, optimization of discrimination model, and fault classification using softmax function.

3. The method for diagnosing a hoisting system fault based on a deep graph convolutional generative adversarial network according to claim 2, characterized in that: The reconstruction of the graph structure uses the features of the nodes to reconstruct the adjacency matrix of the graph, and uses the k-nearest neighbor algorithm to determine similar node connections based on the similarity between node features. For a node v on the graph i (i=1,2,…,n), original node feature x i and x j With another node v j The cosine similarity expression of (j=1,2,…,i-1,i+1,…,n) is: Among them, s ij represents the cosine similarity between node i and node j, r i and r j Represent the feature vectors of node i and node j respectively, is the vector r i and r j The dot product between ||r i ||2 represents vector r i The L2 norm of ||r j ||2 represents vector r j The L2 norm of ; Connect each node to the k most similar nodes, generating a graph structure containing a new adjacency matrix.

4. The method for diagnosing a hoisting system fault based on a deep graph convolutional generative adversarial network according to claim 2, characterized in that: The channel information interaction mechanism is established, which allows the initial graph signal to be retrieved from the original and current branches, and is used to improve the interaction of features between different branches. The H0 and H'0 calculated by the reconstruction of the graph structure, in a network with a total number of layers of l, the feature interaction process expression of the lth (l = 1, 2, ..., L) layer is: Among them, H l is the feature representation of the lth layer, δ represents the activation function, which is used to perform nonlinear transformation on the transferred features; P represents the adjacency matrix involved in the graph convolution, α is the initial residual connectivity coefficient, β is the channel information interaction coefficient, and H l-1 is the feature representation of the l-1th layer, H0 is the initial feature representation, represents the weight matrix of the lth layer, H′ l , P′, H′0 and They are respectively l , P, H0, and The corresponding feature interaction values, the updated weight matrix is ​​expressed as: in, and is the updated weight matrix, W l and W l ′ are the weight matrices of the lth layer in the two branches, I n is the n-order identity matrix, γ l is a parameter that increases with the number of network layers, and its expression is: Among them, θ = 0.5 means control γ l The hyperparameter of the incremental rate is expressed as follows: out=f3(H L )+f4(H′ L ) Among them, out represents the final output of the network, f3 and f4 are two linear transformations, and H L and H′ L Represents two different feature representations obtained at layer L.

5. The method for diagnosing a hoisting system fault based on a deep graph convolutional generative adversarial network according to claim 2, characterized in that: The optimization of the generative model is that the generator encodes the collected real image, encodes the real image in the input stage and extracts its features, introduces noise with normal distribution to each mean and variance obtained, so that the features of the extracted real image are closer to the feature vector that satisfies the normal distribution, and sends the obtained feature vector to the generator to generate the image signal, and x r As the real sample of the input image, the mean μ and variance σ are obtained by the encoder, and the standard normal distribution is used to construct the randomized feature vector c, which is expressed as: in, represents element-by-element multiplication, μ and σ represent mean and variance respectively, c is the final generated eigenvector, N(0,I) represents a multidimensional normal distribution with mean 0 and covariance matrix as the identity matrix I; The output of the discriminator is represented as y∈[0,1], where the feature vector generated by the generator is expressed as: Among them, G represents the generator, represents the sample obtained by the generator, g is the generating function, θ g is the generator parameter, z represents the latent variable, and X are the generated image and original image obtained by the discriminator respectively, and the expression is: The output obtained by the discriminator is expressed as: D(x) represents the true sample distribution of the image, and the error function expression is: Among them, θ d is the parameter of the discriminator, p(x) is the true data distribution of sample x, that is, the probability distribution of the true data, The samples generated by the generator, To generate data The probability distribution of represents the distribution of the input generator noise, the internal parameters θ of the discriminator d The expression is: i d =(θ f ,i c ) Among them, θ f is the parameter of the feature extraction part of the discriminator, θ c are the parameters of the classification part of the discriminator; By maximizing the entropy of the noise distribution, the generated sample image is constantly close to the real sample image, and the network reaches a balanced expression: Among them, max represents the maximization operation, θ g is the generator parameter, is the output of the discriminator, representing the generated data The probability of belonging to the real data, To generate data The probability distribution of The larger the value of , the better the fitting ability of the generated network is, and the higher the similarity between the generated image and the real image is.

6. The method for diagnosing a hoisting system fault based on a deep graph convolutional generative adversarial network according to claim 2, characterized in that: The optimization of the discriminant model, let p data (x) and p g (x) represent the real samples and image samples output by the optimized model respectively. The loss function of the deep graph convolutional generative adversarial network algorithm consists of the discriminant network and the generative network, which is expressed as: Among them, p data (x) and p g (x) During the training process, the two networks are infinitely close to each other, and the generative network and the discriminative network are interactively iterated. When training the discriminative network, the real image is mixed with the generated image and input into the discriminator, and the parameters of the network are adjusted according to the output of the discriminator. When a set of data is randomly input, the generative network generates an image, and the discriminator gives the discrimination result. The difference between the generated image and the real image is calculated and fed back to the generative network to adjust the parameters, and interactive iteration is performed until the network reaches stability. The optimization of the discriminative network is realized by minimizing the cross entropy operation. The loss function expression is: The optimal solution expression of the loss function is: Where D(x) represents the predicted output of the discriminator D for the input data x, p data (x) and p g (x) represent the real samples and image samples output by the optimized model, respectively, and p g represents the probability of generating data, D * (x) is the optimal solution of the loss function, p data (x) and p g The value of (x) is a real number between (0,1) in actual training. When the generator network is fixed, the discriminator network has the optimal solution and obtains the best performance. The loss function is updated as: in, Represents the real sample p data The random variable x sampled in (x) is subjected to expectation operation. JS stands for Jensen-Shannon divergence, which is a symmetric version of KL divergence and is used to measure the two distributions, namely p data (x) and p g (x), V(D,G) represents the loss function of the discriminator D and the generator G, is the output of the discriminator D for the input x, indicating the probability of judging it as real data, p data (x) and p g (x) represent the real samples and image samples output by the optimized model, respectively, and p z is the noise distribution, Denotes the measure distribution p data (x) and p g (x), namely, KL divergence, when p data (x) = p g (x), calculate the global minimum.

7. The method for diagnosing a hoisting system fault based on a deep graph convolutional generative adversarial network according to claim 2, characterized in that: The softmax function is used for fault classification, and the whitening preprocessing method is used to normalize the input samples. The expression is: Among them, E[x (k) ] represents the average value of each batch of training data, and uses two learning parameters γ and ξ for inverse transformation. represents the normalized training data, x (k) represents the kth sample of the original data, Var represents the variance operation, and the expression is: Among them, y (k) represents the normalized target value, γ (k) represents the scaling factor, ξ (k) represents the bias term, The softmax function is used to calculate the feature weight of the scale, and the expression is: Among them, j is the output of the softmax function, indicating the probability of the jth class, The output classification probability calculation expression under each feature weight is: Among them, j is the output of the softmax function, indicating the probability of the jth class, x k represents the kth element in the input sample, P t (x) j is the predicted probability of the jth class given an input x, e represents a natural constant, η(k) is the weight associated with class k, and x j represents the features of input x under category j, k represents the number of categories or feature values, K is the total number of categories, Used to automatically update P in the training and test datasets t (x) j .

Images