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

Abstract

This invention discloses a fault diagnosis method for lifting systems based on depth graph convolutional generative adversarial networks (DGANs). The method includes: collecting vibration signals from the lifting system under complex operating conditions and selecting a fault dataset as training and testing samples; setting and initializing the internal parameters of the DGAN algorithm; fine-tuning the DGAN algorithm using training samples to obtain optimal parameters; performing small-sample fault diagnosis on the lifting system; generating fault detection results; evaluating the diagnostic effect of the DGAN algorithm; and verifying the fault diagnosis capabilities of the DGAN algorithm. This invention facilitates the generation and enhancement of small-sample data, improving the effective execution of small-sample fault diagnosis for lifting systems under complex operating conditions.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis of lifting systems, and in particular to a fault diagnosis method for lifting systems based on depth graph convolutional generative adversarial networks. Background Technology

[0002] Crane systems, as crucial mechanical equipment in automated production, are susceptible to system failure due to minor damage caused by harsh working environments. Therefore, developing accurate and intelligent fault diagnosis methods is essential. Deep learning-based fault diagnosis methods significantly improve diagnostic efficiency through automatic feature extraction and pattern recognition; however, they require large amounts of training data and high-performance hardware, leading to reduced diagnostic accuracy. Since mechanical systems experience fewer failures under normal operating conditions, and data collection is difficult, existing methods perform poorly with small sample sizes and cannot effectively address the high-precision diagnostic needs under limited data conditions.

[0003] For scenarios with limited sample data, current algorithms such as Domain Adaptive Algorithm (DAA), Convolutional Neural Networks (CNN), and Graph Neural Networks (GNN) generally suffer from overfitting, insufficient feature extraction capabilities, and high computational complexity, leading to reduced diagnostic accuracy with limited data. Furthermore, while advanced few-sample learning methods such as transfer learning (TL) and meta-learning (ML) help address the data scarcity problem, they are often limited by label scarcity, model complexity, and long training times, making them difficult to meet the challenges of real-world applications. Summary of the Invention

[0004] To overcome the shortcomings and deficiencies of existing technologies, this invention provides a fault diagnosis method for lifting systems based on depth graph convolutional generative adversarial networks.

[0005] The technical solution adopted in this invention is a fault diagnosis method for lifting systems based on depth graph convolutional generative adversarial networks, which includes:

[0006] Step S1: Collect vibration signals from the lifting system under complex working conditions, and select the fault dataset as training and test samples;

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

[0008] Step S3: Fine-tune the depth map convolutional generative adversarial network algorithm using training samples to obtain the optimal parameters;

[0009] Step S4: Then, the test samples are input into the trained depth map convolutional generative adversarial network algorithm to perform small-sample fault diagnosis on the lifting system;

[0010] Step S5: Generate fault detection results and evaluate the diagnostic performance of the depth map convolutional generative adversarial network algorithm;

[0011] Step 6: Validate the fault diagnosis of the depth graph convolutional generative adversarial network algorithm by comparing it with existing fault diagnosis methods and few-shot learning methods.

[0012] Furthermore, the depth graph convolutional generative adversarial network algorithm includes: graph structure reconstruction, establishing a channel information interaction mechanism, optimizing the generative model, optimizing the discriminative model, and using the softmax function for fault classification;

[0013] Furthermore, the reconstruction of the graph structure involves: reconstructing the adjacency matrix of the graph using the features of the nodes, and using the k-nearest neighbor algorithm to determine similar node connections based on the similarity between node features. For a node in the graph... Original node features x i and x j With another node Another node , j =1,2,…, n The cosine similarity expression is:

[0014]

[0015] in, s ij Represents a node i and nodes j Cosine similarity between them r i and r j Representing nodes respectively i and nodes j eigenvectors, For vectors r i and r j dot product between Representing vectors r i L2 norm, Representing vectors r j The L2 norm.

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

[0017] Furthermore, the established channel information interaction mechanism allows retrieval of initial graph signals from the original and current branches, improving the interaction of features between different branches, and the graph structure reconstruction calculation is obtained.H o and In a total of 100 floors L In the network, the first l The expression for the feature interaction process of the layer is:

[0018]

[0019]

[0020] in, H l For the first l Layer feature representation, l =1, 2, ..., L , d This represents an activation function that can perform non-linear transformations on the transmitted features; P This represents the adjacency matrix involved in graph convolution. α These are the initial residual connectivity coefficients. β The interaction coefficient of channel information. H l-1 For the first l- Feature representation of layer 1, H 0 represents the initial feature. Indicates the first l The layer's weight matrix. Additionally... , , and They are respectively with H l , P , H 0 and The corresponding feature interaction values, and the updated weight matrix are shown below:

[0021]

[0022]

[0023] in, and The updated weight matrix, W l and The first of the two branches l The weight matrix of the layer, I n for n An identity matrix of order 1. The parameter increases with the number of network layers, and its expression is:

[0024]

[0025] Where θ = 0.5 represents control The hyperparameter of the incremental rate is determined by designing a linear feature fusion method. This method adaptively fuses structural information features and similarity information features into a new node representation, and its expression is:

[0026]

[0027] Where out represents the final output of the network, and f3 and f4 are two linear transformations. and Indicates the first L The layer yields two different feature representations.

[0028] Furthermore, the optimization of the generative model involves: the generator encoding the acquired real images; encoding the real images and extracting their features during the input stage; for each mean and variance obtained, importing normally distributed noise to make the extracted features of the real images closer to a normally distributed feature vector; and then feeding the obtained feature vector into the generator to produce an image signal; x r The mean value is obtained by processing the input image as a real sample through the encoder. µ and variance s And construct randomized feature vectors using the standard normal distribution. c Its mathematical expression is:

[0029]

[0030] in, This indicates element-wise multiplication. µ and s These represent the mean and variance, respectively. c The final generated feature vector, N (0, I () represents a matrix with a mean of 0 and a covariance matrix that is the identity matrix. I The multidimensional normal distribution.

[0031] Let represent the generated data. The output of the discriminator can be represented as y∈[0,1], where the feature vector generated by the generator can be calculated using the following formula:

[0032]

[0033] in, G Represents a generator. This indicates the generation of data. g For generating functions, i g For generator parameters, zRepresenting latent variables, and X The generated image and the original image obtained by the discriminator are respectively expressed as follows:

[0034]

[0035] The output obtained by the discriminator can be represented as:

[0036]

[0037] D(x) represents the true sample distribution of the image, and the error function is expressed as:

[0038]

[0039] in, i d These are the parameters of the discriminator. p ( x ) is a sample x The true data distribution, i.e. the probability distribution of the true data. Samples generated by the generator To generate data The probability distribution. The discriminator's internal parameters represent the distribution of noise from the input generator. i d The expression is:

[0040]

[0041] in, i f These are the parameters of the discriminator's feature extraction part. i c These are the parameters of the discriminator's classification section.

[0042] By maximizing the entropy of the noise distribution, the generated sample images gradually approach the real sample images, and the network eventually reaches equilibrium. The expression is as follows:

[0043]

[0044] in, This indicates a maximization operation. For generator parameters, To generate data The probability distribution. The larger the value, the better the fitting ability of the generating network, and the higher the similarity between the generated image and the real image, indicating that the network proposed in this invention has the best image generation performance.

[0045] Furthermore, the optimization of the discriminant model: Let pdata (x) and p g (x) represent the real sample and image sample output by the optimized model, respectively. The loss function of the depth map convolutional generative adversarial network algorithm consists of a discriminator network and a generator network, and is expressed as:

[0046]

[0047] Where, p data (x) and p g (x) During training, the generator and discriminator networks iteratively approach each other. When training the discriminator network, real and generated images are mixed and input into the discriminator, and the network parameters are adjusted based on the discriminator's output. Simultaneously, when a random set of data is input, the generator network generates an image, and the discriminator provides the discrimination result. The difference between the generated and real images is calculated and fed back to the generator network to adjust its parameters. This iterative process continues until the network reaches stability. The discriminator network is optimized by minimizing the cross-entropy operation; the loss function expression is:

[0048]

[0049] The optimal solution expression for the loss function is:

[0050]

[0051] in, D ( x ) indicates the discriminator D For input data x The predicted output, p data ( x )and p g ( x () represent the real samples and image samples output by the optimized model, respectively. p g p represents the probability of generating data. data (x) and p g The values ​​of (x) are all real numbers between (0, 1) during actual training. When the generator network is fixed, the discriminator network has the optimal solution and achieves the best performance; the loss function is updated as follows:

[0052]

[0053] in, V ( D , G ) indicates the discriminator D and generator G loss function, It is a discriminator D For input x The output represents the probability that it is judged as real data. p data ( x )and p g ( x () represent the real samples and image samples output by the optimized model, respectively. p z For noise distribution, Indicates the measure of distribution p data ( x )and p g ( x The difference between p and y is the KL divergence. data (x)=p g When (x), the global minimum value can be calculated.

[0054] Furthermore, the fault classification using the softmax function involves normalizing the input samples using a whitening preprocessing method, expressed as:

[0055]

[0056] Where E[x(k)] represents the average value of each batch of training data, using two learning parameters. and An inverse transformation is performed to ensure the accuracy and stability of feature learning; the expression is:

[0057]

[0058] Finally, the feature weights for the scale are calculated using the softmax function, expressed as follows:

[0059]

[0060] in, The expression for calculating the output classification probability under each feature weight is as follows:

[0061]

[0062] in, Given input x The j Predicted probability of class Is with category k Relevant weights, x j Indicates input x In category j The following featuresK It is the total number of categories. Used to automatically update P in the training and testing datasets t (x) j .

[0063] Beneficial effects:

[0064] This application proposes a fault diagnosis method for lifting systems based on depth graph convolutional generative adversarial networks (GANs). The invention designs a novel GAN ​​algorithm that effectively acquires all small-sample fault information, thus facilitating the generation and enhancement of small-sample data. This method improves the effective execution of small-sample fault diagnosis in lifting systems under complex operating conditions. Specifically:

[0065] 1. Acquisition of Small Sample Fault Information: This invention can effectively acquire small sample fault information. By utilizing the generative capabilities of generative adversarial networks, it enhances the small sample data to compensate for the problem of data scarcity. This method not only helps the model learn more fault features but also avoids overfitting caused by insufficient data, thereby improving the accuracy of diagnosis.

[0066] 2. Solving the overfitting problem: This invention uses the adversarial training mechanism of generative adversarial networks to enable the model to learn more general features with limited samples, avoiding excessive dependence on training data, thereby improving the accuracy and robustness of the model in practical applications.

[0067] 3. Enhanced feature extraction capability: This invention can not only extract deep-level features hidden in graph structures, but also enhance the model's ability to identify fault modes through graph convolutional layers, thereby improving diagnostic accuracy.

[0068] 4. Improve diagnostic efficiency under complex working conditions: By combining the advantages of depth map convolution and generative adversarial networks, this invention can not only effectively address the fault diagnosis problem under complex working conditions, but also improve performance in complex environments by enhancing the learned features.

[0069] 5. Reduced computational complexity: This invention introduces generative adversarial networks to generate more data samples, reducing the dependence on a large amount of raw data. At the same time, it reduces computational complexity and improves the efficiency of model training by using graph convolution.

[0070] 6. Addressing Label Scarcity: This invention utilizes generative adversarial networks 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 allows the model to maintain high diagnostic accuracy even with a limited number of labels. Attached Figure Description

[0071] Figure 1This is a flowchart illustrating the overall steps of the present invention;

[0072] Figure 2 This is an architecture diagram of the depth graph convolutional generative adversarial network algorithm of the present invention. Detailed Implementation

[0073] It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. The following describes the application in further detail with reference to the accompanying drawings and specific embodiments.

[0074] This invention proposes a fault diagnosis method for lifting systems based on depth graph convolutional generative adversarial networks (GANs) to address the problems of low accuracy and poor efficiency in fault diagnosis of current lifting systems under complex operating conditions, especially with small sample data. Through an innovative depth graph convolutional GAN ​​algorithm, it effectively captures fault information from small samples, enhances the generation capability of small sample data, and thus improves the accuracy and robustness of fault diagnosis for lifting systems under complex operating conditions.

[0075] like Figure 1 As shown, a fault diagnosis method for lifting systems based on depth graph convolutional generative adversarial networks is proposed. This method includes:

[0076] Step S1: Collect vibration signals from the lifting system under complex working conditions, and select the fault dataset as training and test samples;

[0077] Specifically, this step involves first collecting vibration signals from different parts of the lifting system using sensors (such as accelerometers and vibration sensors). Since lifting systems typically operate in harsh environments, the collected data may contain various types of noise and interference; therefore, signal preprocessing is necessary to remove noise and ensure data accuracy. Next, representative fault datasets are selected from these vibration signals as training and testing samples. The fault data should include different types of fault modes (such as bearing failure, gear failure, etc.) and cover different operating states to improve the model's generalization ability. Based on the diversity of the data, it is divided into training and testing sets.

[0078] Step S2: Set and initialize the internal parameters of the depth graph convolutional generative adversarial network algorithm, including connectivity coefficient α, interaction coefficient β, mean µ, and variance σ;

[0079] 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, used to control feature transfer between the generator and discriminator; the interaction coefficient β, used to adjust the degree of feature interaction between different parts of the network; and the mean μ and variance σ, used to standardize the input data. Parameter initialization is a crucial step in the training process; proper initialization can accelerate the network's convergence and avoid gradient vanishing or exploding.

[0080] Step S3: Fine-tune the depth map convolutional generative adversarial network algorithm using training samples to obtain the optimal parameters;

[0081] Specifically, this step involves fine-tuning the depth map convolutional generative adversarial network (GAN) algorithm using the training samples collected in step S1. This process continuously optimizes the network parameters (e.g., convolutional kernels, weight matrices) through backpropagation, enabling the generator to produce more realistic pseudo-data and the discriminator to more accurately distinguish between real and generated data. During fine-tuning, the network gradually adjusts its parameters based on feedback from the loss function until the loss is minimized. In this stage, parameter updates are achieved using gradient descent, with the optimization goal of improving the performance of both the generator and discriminator.

[0082] Step S4: Then, the test samples are input into the trained depth map convolutional generative adversarial network algorithm to perform small-sample fault diagnosis on the lifting system;

[0083] Specifically, this step involves inputting the test samples selected in step S1 into a trained deep graph convolutional generative adversarial network (GAN). Using the trained network model, the system performs small-sample fault diagnosis on the input test samples. Because GANs have the ability to enhance small-sample data, even during the testing 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 from the graph structure through graph convolution operations and combines this with the generated sample data to perform accurate fault identification and classification.

[0084] Step S5: Generate fault detection results and evaluate the diagnostic performance of the depth map convolutional generative adversarial network algorithm;

[0085] Specifically, this step involves processing the data through a depth graph convolutional generative adversarial network (GAN), after which the system generates fault detection results. These results are typically a diagnostic label or a predicted value for a set of fault modes, such as "normal," "minor fault," or "serious fault." Next, the algorithm's performance is evaluated based on the diagnostic results, primarily by comparing the actual fault conditions with the network's predicted results to calculate performance metrics such as accuracy, precision, and recall. This step assesses the network's diagnostic capabilities, verifying whether the algorithm can accurately identify different types of faults and possesses high diagnostic efficiency.

[0086] Step S6: Verify the fault diagnosis of the depth graph convolutional generative adversarial network algorithm by comparing it with existing fault diagnosis methods (the specific diagnosis methods should be specified here) and few-shot learning methods.

[0087] Specifically, this step involves comparing and validating the diagnostic results of the deep graph convolutional generative adversarial network (DBGAN) algorithm with existing fault diagnosis methods and few-shot learning methods. Existing fault diagnosis methods include distance probability graph convolutional networks (e.g., multi-scale deep graph convolutional networks, multi-channel residual multi-scale clustering graph convolutional networks, and multi-channel domain adaptive graph convolutional networks), while few-shot learning methods include model-agnostic meta-learning, task ranking meta-learning, attention-based deep meta-transfer learning, and meta-learning generative adversarial networks. By comparing the performance of different methods, the advantages and effectiveness of the proposed algorithm under few-shot conditions can be verified. This validation process includes the model's performance on different datasets and its applicability in practical applications, ensuring that the algorithm has strong generalization ability and practical value.

[0088] The depth graph convolutional generative adversarial network algorithm includes: graph structure reconstruction, establishment of channel information interaction mechanism, optimization of generative model, optimization of discriminative model, and fault classification using softmax function.

[0089] like Figure 2 As shown, in the data generation stage: Generator G The generator is responsible for generating "pseudo-data" that resembles real data. It receives random noise from the input and uses a network to generate pseudo-data similar to the target data. The discriminator... DThe task is to distinguish between generated and real data. It determines whether the input data is real and gives a "true" or "false" result. The discriminator optimizes its parameters to improve its ability to distinguish generated data. The generator and discriminator are trained adversarially. The generator optimizes how to generate more realistic data, while the discriminator optimizes how to more accurately distinguish between real and fake data. Finally, the generator and discriminator continuously compete until the data generated by the generator is as close to 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 transformations. In the deep graph convolutional network stage: the input graph data is fed into the deep graph convolutional network for processing. The graph convolutional network extracts features of the graph structure by learning about nodes and edges. Convolutional layers perform convolution operations on the input data through sliding windows. In graph convolution, the convolution kernel convolves with the features of nodes and their neighbors to extract features of the local graph structure. The parameters of each convolution operation are learned through training. Pooling layers in graph convolutional networks are used for downsampling, reducing feature dimensionality while preserving important feature information. Each convolutional block is followed by a global max pooling layer, which selects the most salient features in the entire graph. Each layer of the graph convolutional network convolves each node in the graph through the connections between nodes, learning deeper graph features. Stacking multiple graph convolutional layers allows for the extraction of increasingly complex graph structure features. The final convolutional output is then passed to the next stage. In the fusion stage: different features are combined through a fusion operation, and the outputs of multiple convolutional layers are passed through a 1×1 convolution (each convolution having a feature count of...). f The features are fused using graph convolutions (e.g., 1×1 convolutions). This convolution operation reduces computation while maintaining feature diversity. In the prediction phase: the features after graph convolution and fusion are passed to a fully connected layer for further processing, ultimately yielding the prediction result. 1×1 convolutions are then used for feature extraction, gradually reducing the feature dimensionality until the final output is obtained. Finally, the prediction result of the output layer is the final prediction for fault diagnosis. After computation across the entire network, the output indicates the fault type or system state of each node.

[0090] To accurately extract limited feature information and improve the fault diagnosis accuracy of lifting systems, this invention proposes a novel deep learning architecture—Depth Graph Convolutional Generative Adversarial Network (DGAN). First, the original data... By inputting two different nonlinear fully connected layers, the initial features of the graph structure branches and node similarity branches are obtained, and these features are used as the initial inputs to the depth graph convolutional module and the channel information interaction mechanism, as shown below:

[0091] (1)

[0092] In this design, f1 and f2 are two linear transformations, and δ(·) is the activation function. The nonlinear transformation layer compresses the dimension of the initial feature R to a smaller value, thus significantly reducing the number of network parameters without affecting classification accuracy. Simultaneously, gradient descent is used to calculate the loss function, optimizing both the generative and discriminative models. Compared to existing graph convolutional networks, the deep graph convolutional generative adversarial network designed in this invention demonstrates stronger capabilities in extracting more comprehensive high-order semantic features, thereby improving fault classification accuracy. The specific construction process of the deep graph convolutional generative adversarial network framework proposed in this invention is as follows:

[0093] Graph Structure Reconstruction: Existing diagnostic methods based on graph convolutional networks primarily consider the graph's topology, resulting in a large number of low-dimensional nodes and a limited number of neighboring nodes, thus hindering sufficient information acquisition during feature aggregation. Furthermore, multiple graph convolutions can destroy the structural information between similar nodes, leading to node attribute convergence and loss of feature diversity. To overcome these problems, this invention uses node features to reconstruct the graph's adjacency matrix and employs the k-nearest neighbor algorithm to determine similar node connections based on the similarity between node features. For a node in the graph... Its original node features x i and x j With another node , j =1,2,…, n The cosine similarity of can be expressed as:

[0094] (2)

[0095] in, s ij Represents a node i and nodes j Cosine similarity between them r i and r j Representing nodes respectively i and nodes j eigenvectors, For vectors r i and r j dot product between Representing vectors r i L2 norm, Representing vectors r j The L2 norm.

[0096] Accordingly, each node is connected to its k most similar nodes to generate a graph structure containing a new adjacency matrix, which overcomes the problem of node similarity in traditional graph convolutional network methods that destroys the original graph structure information. Therefore, the deep graph convolutional generative adversarial network proposed in this invention can extract high-order semantic feature information and obtain more accurate node representations.

[0097] Establishing a channel information interaction mechanism: To maintain the diversity of node features, this invention designs a channel information interaction mechanism and incorporates residual connections between branches. This mechanism allows retrieval of initial graphic signals from the original and current branches, thereby improving the interaction of features between different branches. Based on H0 calculated by equation (1) and In a total of 100 floors l In the network, the first l ( l The feature interaction process of layers (=1,2,…,L) can be represented as:

[0098] (3)

[0099] (4)

[0100] in, H l For the first l Layer feature representation, d This represents an activation function that can perform non-linear transformations on the transmitted features; P This represents the adjacency matrix involved in graph convolution. α These are the initial residual connectivity coefficients. β The interaction coefficient of channel information. H l-1 For the first l- Feature representation of layer 1, H 0 represents the initial feature. Indicates the first l The layer's weight matrix. Additionally... , , and They are respectively with H l , P , H 0 and The corresponding feature interaction values, and the updated weight matrix are shown below:

[0101] (5)

[0102] (6)

[0103] In the formula, α is the initial residual connectivity coefficient, In is the n-order identity matrix, and β is the channel information interaction coefficient. W l and The first of the two branches l The weight matrix of the layer. The parameter that increases with the number of network layers is calculated using the following formula:

[0104] (7)

[0105] Where θ = 0.5 is the control... Hyperparameters of the incremental rate. To rationally integrate channel feature information, a linear feature fusion method is designed. This method can adaptively fuse structural information features and similarity information features into the new node representation. The specific calculation equation is as follows:

[0106] (8)

[0107] Where out represents the final output of the network, and f3 and f4 are two linear transformations. and Indicates the first L The layer yields two different feature representations.

[0108] Optimization of the generative model: The generator encodes the acquired real images. To ensure high-quality image generation, the real images are encoded and their features are extracted during the input stage. For each mean and variance obtained, normally distributed noise is introduced to make it closer to a normally distributed feature vector. The resulting feature vector is then fed into the generator to produce the image signal. Here, x... r As real samples of the input image, their mean µ and variance σ are obtained through encoder processing, and a randomized feature vector c is constructed using a standard normal distribution. Its mathematical expression is:

[0109] (9)

[0110] in, This indicates element-wise multiplication. µ and s These represent the mean and variance, respectively. c The final generated feature vector, N (0, I () represents a matrix with a mean of 0 and a covariance matrix that is the identity matrix. I The multidimensional normal distribution.

[0111] in addition, Let represent the generated data. The output of the discriminator can be represented as y∈[0,1], where the feature vector generated by the generator can be calculated using the following formula:

[0112] (10)

[0113] Where G represents the generator, This indicates the generation of data. g For generating functions, i g For generator parameters, z Representing latent variables, X and X are the generated image and the original image obtained by the discriminator, respectively, and their expressions are:

[0114] (11)

[0115] Then, the output obtained by the discriminator can be expressed as:

[0116] (12)

[0117] Where D(x) represents the true sample distribution of the image, and its error function can be calculated as:

[0118] (13)

[0119] in, i d These are the parameters of the discriminator. p ( x ) is a sample x The true data distribution, i.e. the probability distribution of the true data. Samples generated by the generator To generate data The probability distribution. The discriminator's internal parameter θ represents the distribution of noise from the input generator. d The expression is:

[0120] (14)

[0121] in, i f These are the parameters of the discriminator's feature extraction part. i c These are the parameters of the discriminator's classification section.

[0122] By maximizing the entropy of the noise distribution, the generated sample images gradually approach the real sample images, and eventually the network reaches equilibrium. The specific formula is as follows:

[0123] (15)

[0124] in, This indicates a maximization operation. For generator parameters, To generate data The probability distribution. The larger the value, the better the fitting ability of the generating network, and the higher the similarity between the generated image and the real image, indicating that the network proposed in this invention has the best image generation performance.

[0125] Optimization of the discriminant model: Let p data ( x )and p g ( x ) represent the real sample and image sample output by the optimized model, respectively. Therefore, the loss function of the depth map convolutional generative adversarial network algorithm proposed in this invention consists of a discriminant network and a generator network, and is expressed as:

[0126] (16)

[0127] Among them, among them, D ( x ) indicates the discriminator D For input data x The predicted output, p data ( x )and p g ( x () represent the real samples and image samples output by the optimized model, respectively. p g This indicates the probability of generating data. p data ( x )and p g ( x During training, the network and the discriminator network iterate and interact during the optimization of formula (16). When training the discriminator network, real and generated images are mixed and input into the discriminator, and the network parameters are adjusted based on the discriminator's output. Simultaneously, when a set of data is randomly input, the network generates an image, and the discriminator provides the discrimination result. The difference between the generated image and the real image is calculated and fed back to the network to adjust its parameters, thus iterating and interacting until the network reaches stability. The optimization of the discriminator network is achieved by minimizing the cross-entropy operation, and its loss function is described as:

[0128] (17)

[0129] Therefore, the optimal solution of formula (17) can be expressed as:

[0130] (18)

[0131] in, p data ( x )and p g ( x The values ​​of ) are all real numbers between (0,1) in actual training. When the generator network is fixed, the discriminator network has the optimal solution and obtains the best performance. Therefore, equation (17) can be updated to:

[0132] (19)

[0133] in, V ( D , G ) indicates the discriminator D and generator G loss function, It is a discriminator D For input x The output represents the probability that it is judged as real data. p data ( x )and p g ( x () represent the real samples and image samples output by the optimized model, respectively. p z For noise distribution, Indicates the measure of distribution p data ( x )and p g ( x The difference between p and y is the KL divergence. data (x)=p g When (x), the global minimum value can be calculated.

[0134] Fault classification is performed using the softmax function: The input samples are normalized using a whitening preprocessing method, specifically defined as follows:

[0135] (20)

[0136] In the formula, E[x (k) The value represents the average of each batch of training data, using two learning parameters. and An inverse transformation is performed to ensure the accuracy and stability of feature learning, namely:

[0137] (twenty one)

[0138] The feature weights of the scale are calculated using the softmax function, as shown in equation (22):

[0139] (twenty two)

[0140] In the formula, The output classification probability under each feature weight is calculated by the following formula (23):

[0141] (twenty three)

[0142] in, Given input x The j Predicted probability of class Is with category k Relevant weights, x j Indicates input x In category j The following features K It is the total number of categories. Used to automatically update P in the training and testing datasets t (x) j .

[0143] This application effectively acquires 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 in lifting systems under complex working conditions. Specifically, it includes:

[0144] 1. Acquisition of Fault Information from Small Samples: Traditional deep learning methods often rely on large amounts of training data. However, in mechanical equipment such as lifting systems, acquiring fault data is difficult, especially under normal operating conditions where fault samples are scarce. The method proposed in this application, based on a depth map convolutional generative adversarial network (GAN), can effectively acquire fault information from small samples. By leveraging the generative capabilities of GANs, it enhances the small sample data to compensate for the data scarcity problem. This method not only helps the model learn more fault features but also avoids overfitting caused by insufficient data, thus improving diagnostic accuracy.

[0145] 2. Addressing Overfitting: Existing methods often encounter overfitting issues in scenarios with small sample sizes, especially in deep neural networks, where models tend to memorize training data and lack good generalization ability. The method in this application utilizes the adversarial training mechanism of generative adversarial networks to enable the model to learn more general features with limited samples, avoiding excessive reliance on training data and thus improving the model's accuracy and robustness in practical applications.

[0146] 3. Enhanced Feature Extraction Capability: While traditional Convolutional Neural Networks (CNNs) have achieved significant results in image processing, they have limitations when handling graph-structured data. Deep Graph Convolutional Networks can effectively process graph-structured data, utilizing the connections between nodes to learn features at a deeper level. Combined with Generative Adversarial Networks (GANs), the method in this application can not only extract deep-level features hidden in graph structures but also enhance the model's ability to identify fault modes through graph convolutional layers, thereby improving diagnostic accuracy.

[0147] 4. Improve diagnostic efficiency under complex operating conditions: Lifting systems may encounter multiple fault modes under complex operating conditions, and due to changes in environmental factors, fault manifestations exhibit strong nonlinearity and complexity. The method proposed in this application, by combining the advantages of depth map convolution and generative adversarial networks, can not only effectively address the fault diagnosis problem under complex operating conditions, but also improve performance in complex environments by enhancing learned features.

[0148] 5. Reduced computational complexity: Although many deep learning methods have strong feature extraction capabilities, their computational complexity and training time are often very high. The method in this application introduces a generative adversarial network to generate more data samples, reducing the dependence on a large amount of raw data. At the same time, it reduces computational complexity and improves the efficiency of model training by using graph convolution.

[0149] 6. Addressing Label Scarcity: Label scarcity is a common problem in few-shot learning, especially in mechanical equipment fault diagnosis, where obtaining labels is costly. This application's solution utilizes generative adversarial networks (GANs) to expand the training set by synthesizing new data samples without relying on a large amount of labeled data, thereby alleviating the label scarcity problem. This allows the model to maintain high diagnostic accuracy even with a limited number of labels.

[0150] Through these innovative methods, the fault diagnosis scheme of this application not only overcomes the problems mentioned in the background art, but also provides more accurate and efficient fault diagnosis capabilities in practical applications, meeting the high-precision diagnosis needs of complex mechanical equipment such as lifting systems under small sample data conditions.

[0151] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A fault diagnosis method for lifting systems based on depth graph convolutional generative adversarial networks, characterized in that... The method includes: Step S1: Collect vibration signals from the lifting system under complex working conditions, and select the fault dataset as training and test samples; Step S2: Set and initialize the internal parameters of the depth graph convolutional generative adversarial network algorithm, including connectivity coefficient α, interaction coefficient β, mean µ, and variance σ; Step S3: Fine-tune the depth map convolutional generative adversarial network algorithm using training samples to obtain the optimal parameters; Step S4: Then, the test samples are input into the trained depth map convolutional generative adversarial network algorithm to perform small-sample fault diagnosis on the lifting system; Step S5: Generate fault detection results and evaluate the diagnostic performance of the depth map convolutional generative adversarial network algorithm; Step S6: Verify the fault diagnosis of the depth map convolutional generative adversarial network algorithm using existing fault diagnosis methods and few-shot learning methods; The depth graph convolutional generative adversarial network algorithm includes: graph structure reconstruction, establishment of channel information interaction mechanism, optimization of generative model, optimization of discriminative model, and fault classification using softmax function; The established channel information interaction mechanism allows retrieval of initial graph signals from the original and current branches, improving the interaction of features between different branches. The graph structure reconstruction calculation yields... H 0 and In a total of 100 floors L In the network, the first l The expression for the feature interaction process of a layer is: in, H l For the first l Layer feature representation, =1, 2, ..., L , δ This represents the activation function, used to perform non-linear transformations on the transmitted features; P This represents the adjacency matrix involved in graph convolution. α These are the initial residual connectivity coefficients. β The interaction coefficient of channel information. H l-1 For the first - Feature representation of layer 1, H 0 represents the initial feature. Indicates the first The weight matrix of the layer, , , and They are respectively with H l , P , H 0 and The corresponding feature interaction values, and the updated weight matrix are represented as follows: in, and The updated weight matrix, W l and The first of the two branches The weight matrix of the layer, I n for n An identity matrix of order 1. The parameter increases with the number of network layers, and its expression is: Where θ = 0.5 represents control The hyperparameter of the incremental rate is determined by designing a linear feature fusion method. This method adaptively fuses structural and similarity information features into the new node representation, and its expression is: Where out represents the final output of the network. f 3 and f 4 represents two linear transformations. and Indicates the first L The layer yields two different feature representations.

2. The method for fault diagnosis of lifting systems based on depth graph convolutional generative adversarial networks as described in claim 1, characterized in that, The reconstruction of the graph structure utilizes node features to reconstruct the graph's adjacency matrix and employs the k-nearest neighbor algorithm to determine similar node connections based on the similarity between node features. For a node in the graph... , i =1,2,…, n Original node features x i and x j With another node , j =1,2,…, n The cosine similarity expression is: in, s ij Represents a node i and nodes j Cosine similarity between them r i and r j Representing nodes respectively i and nodes j eigenvectors, For vectors r i and r j dot product between Representing vectors r i L2 norm, Representing vectors r j The L2 norm; Connect each node to its k most similar nodes to generate a graph structure containing a new adjacency matrix.

3. The method for fault diagnosis of lifting systems based on depth graph convolutional generative adversarial networks as described in claim 1, characterized in that, The optimization of the generative model involves the generator encoding the acquired real images. During the input stage, the real images are encoded and their features are extracted. For each mean and variance obtained, normally distributed noise is introduced to make the extracted real image features closer to a normally distributed feature vector. The resulting feature vector is then fed into the generator to produce an image signal. r The mean value is obtained by processing the input image as a real sample through the encoder. µ and variance σ And construct randomized feature vectors using the standard normal distribution. c The expression is: in, This indicates element-wise multiplication. µ and σ These represent the mean and variance, respectively. c The final generated feature vector, N (0, I () represents a matrix with a mean of 0 and a covariance matrix that is the identity matrix. I The multidimensional normal distribution; The output of the discriminator is represented as y∈[0,1], where the feature vector generated by the generator is expressed as: Where G represents the generator, Indicates the generation of data. g For generating functions, θ g For generator parameters, z Representing latent variables, X and X are the generated image and the original image obtained by the discriminator, respectively, and their expressions are: The output obtained by the discriminator is represented as follows: D ( x ) indicates the discriminator D For input data x The predicted output, the error function expression is: in, θ d These are the parameters of the discriminator. p ( x ) is a sample x The true data distribution, i.e., the probability distribution of the true data. To generate data The probability distribution; The discriminator's internal parameter θ represents the distribution of noise from the input generator. d The expression is: in, θ f These are the parameters of the discriminator's feature extraction part. θ c These are the parameters of the discriminator's classification section; By maximizing the entropy of the noise distribution, the generated sample images continuously approach the real sample images, and the network reaches equilibrium as expressed by the following expression: in, This indicates a maximization operation. For generator parameters, To generate data The probability distribution; The larger the value, the better the fitting ability of the generator network, and the higher the similarity between the generated image and the real image.

4. The method for fault diagnosis of lifting systems based on depth graph convolutional generative adversarial networks as described in claim 1, characterized in that, The optimization of the discriminant model is to let p data (x) and p g (x) represent the real sample and image sample output by the optimized model, respectively. The loss function of the depth map convolutional generative adversarial network algorithm consists of a discriminator network and a generator network, and is expressed as: in, D ( x ) indicates the discriminator D For input data x The predicted output, p data (x) and p g (x) During training, the generator and discriminator networks iteratively approach each other. When training the discriminator network, real and generated images are mixed and input into the discriminator, and the network parameters are adjusted based on the discriminator's output. When a random set of data is input, the generator network generates an image, and the discriminator provides the discrimination result. The difference between the generated image and the real image is calculated and fed back to the generator network to adjust its parameters. This iterative process continues until the network reaches stability. The discriminator network is optimized by minimizing the cross-entropy operation; the loss function expression is: The optimal solution expression for the loss function is: in, Discriminator D For input data x The predicted output, p data ( x )and p g ( x () represent the real samples and image samples output by the optimized model, respectively. p g p represents the probability of generating data. 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 follows: in, Indicates the relationship between real samples p data ( x Random variables obtained by sampling from ) x Expectation is calculated using JS, which represents the Jensen-Shannon divergence, a symmetric version of the KL divergence, used to measure the difference between two distributions. p data ( x )and p g ( x The difference between ) V ( D , G ) indicates the discriminator D and generator G loss function, It is a discriminator D For input x The output represents the probability of being judged as real data. p data ( x )and p g ( x () represent the real samples and image samples output by the optimized model, respectively. p z For noise distribution, Indicates the measure of distribution p data ( x )and p g ( x The difference, KL represents the divergence, when p data (x)=p g When (x), calculate the global minimum value.

5. The method for fault diagnosis of lifting systems based on depth graph convolutional generative adversarial networks as described in claim 1, characterized in that, The method uses the softmax function for fault classification and a whitening preprocessing method to normalize the input samples. The expression is as follows: in, E [ x (k) The value represents the average of each batch of training data, using two learning parameters. and Perform inverse transform. This represents the normalized training data. x (k) Represents the first of the original data k For each sample, Var represents variance calculation, and the expression is: in, This represents the normalized target value. Indicates the scaling factor. Indicates the bias term. The feature weights for the scale are calculated using the softmax function, expressed as follows: in, The output of the softmax function represents the th j The probability of a class The expression for calculating the output classification probability under each feature weight is as follows: in, The output of the softmax function represents the th j The probability of a class Indicates the first in the input sample k One element, Given input x The j The predicted probability of a class, where e represents the natural constant. Is with category k Relevant weights, x j Indicates input x In category j The following features k Indicates the number of categories or feature values. K It is the total number of categories. Used to automatically update P in the training and testing datasets t (x) j .

Images