A Fault Diagnosis Method for Multiple Components of a Rolling Mill

Abstract

The present invention discloses a method for diagnosing multiple components of a rolling mill, relating to the technical field of fault diagnosis. The method comprises the following steps: obtaining status data of multiple components of the rolling mill; converting the status data into an RGB image using a Gram angle difference field; converting the RGB image into three images of R, G, and B channels, and converting each channel image into a grayscale image; performing dimensionality reduction processing on each grayscale image, and obtaining a reduced-dimensionality RGB image based on the three grayscale images after dimensionality reduction processing; and inputting the reduced-dimensionality RGB image into an improved ShuffleNetV2 network model to obtain a fault diagnosis result for the rolling mill; the improved ShuffleNetV2 network model is obtained by training a basic ShuffleNetV2 network with a dual attention mechanism. This method can improve the efficiency of fault diagnosis while ensuring the accuracy of fault diagnosis when diagnosing multiple components of a rolling mill.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a method for diagnosing faults of multiple components of a rolling mill. Background Art

[0002] Rolling mills are core equipment in the production of steel and other metals. Bearings and gears are critical transmission components, and failures can cause mill downtime. Timely and accurate fault diagnosis and classification can provide early warning and scheduled repairs, avoiding unplanned downtime, maintaining smooth production line operation, and minimizing the economic losses caused by production interruptions. For example, a single hour of downtime due to a rolling mill failure in a large steel mill can result in the loss of several tons of steel production and corresponding processing profits.

[0003] Rolling mills operate under harsh conditions such as high temperatures, heavy loads, and strong vibrations. Bearings and gears are subjected to alternating loads, shock loads, and high temperatures, making them susceptible to various types of failures, including wear, fatigue cracks, and fractures. These complex operating conditions lead to diverse and interrelated causes of failures, necessitating professional diagnostic and classification to accurately identify the root cause and implement effective remedial measures.

[0004] However, the conventional method of performing fault diagnosis on multiple components of a rolling mill has the problem of low efficiency. Summary of the Invention

[0005] Based on this, it is necessary to provide a fault diagnosis method for multiple components of a rolling mill to address the above technical problems. When performing fault diagnosis on multiple components of a rolling mill, this method can improve the fault diagnosis efficiency while ensuring the accuracy of the fault diagnosis.

[0006] The present invention adopts the following technical solutions:

[0007] The present invention provides a method for diagnosing faults of multiple components of a rolling mill, comprising:

[0008] Acquiring status data of multiple components of a rolling mill;

[0009] The state data is converted into RGB image through Gram angle difference field;

[0010] Convert the RGB image into three images of R, G, and B channels, and convert the image of each channel into a grayscale image;

[0011] Perform dimensionality reduction processing on each grayscale image respectively, and obtain the reduced-dimensional RGB image based on the three grayscale images after dimensionality reduction processing;

[0012] The RGB image after dimension reduction is input into the improved ShuffleNetV2 network model to obtain the fault diagnosis results of the rolling mill; the improved ShuffleNetV2 network model is obtained by adding a dual attention mechanism to the basic ShuffleNetV2 network for training.

[0013] Optionally, dimensionality reduction is performed on each grayscale image separately, including:

[0014] For any grayscale image, get the mean of all data points in the grayscale image;

[0015] According to the values ​​and mean of each data point, the scatter matrix is ​​determined, and the eigenvalues ​​and eigenvectors of the scatter matrix are solved;

[0016] The largest front K The eigenvector corresponding to the eigenvalue is determined as the new eigenvector;

[0017] According to the new feature vector and the original grayscale image, the grayscale image after dimensionality reduction is determined.

[0018] Optionally, the improved ShuffleNetV2 network model includes an input module integrating an attention mechanism, a plurality of lightweight modules connected in series, and an output module integrating an attention mechanism; the dimensionally reduced RGB image is input into the improved ShuffleNetV2 network model to obtain a fault diagnosis result of the rolling mill, including:

[0019] The RGB image after dimension reduction is input into the improved ShuffleNetV2 network model, and the features in the RGB image after dimension reduction are extracted through the input module;

[0020] The convolution feature is obtained by sequentially performing convolution operations on the features in the RGB image after dimension reduction through multiple lightweight modules connected in series.

[0021] The convolution features are predicted through the output module to obtain the fault diagnosis results of the rolling mill.

[0022] Optionally, the input module includes an input layer, a convolution layer, an activation function, an attention mechanism layer, and a fully connected layer connected in series. Extracting features from the RGB image after dimensionality reduction through the input module includes:

[0023] The input layer receives the reduced-dimensional RGB image and performs convolution on the reduced-dimensional RGB image through the convolution layer to extract local features.

[0024] The nonlinear characteristics of local features are increased by activation functions to obtain nonlinear features;

[0025] Assign weights to different features in the nonlinear features through the attention mechanism layer;

[0026] The features output by the attention mechanism layer are fully connected through the fully connected layer to obtain the features in the RGB image after dimensionality reduction.

[0027] Optionally, the attention mechanism layer includes a channel domain attention module and a spatial domain attention module; the channel domain attention module and the spatial domain attention module are cascaded.

[0028] Optionally, the output module includes a convolution layer, an attention mechanism layer, and a fully connected layer connected in series. The output module performs feature prediction on the convolution feature to obtain a fault diagnosis result of the rolling mill, including:

[0029] The convolutional features are convolved through the convolution layer, and the weights are assigned to the convolved features through the attention mechanism layer;

[0030] The features output by the attention mechanism layer are mapped to various fault categories through the fully connected layer to obtain the fault diagnosis results of the rolling mill.

[0031] Optionally, the fault diagnosis result of the rolling mill includes the probability of each fault category occurring in each component of the rolling mill; the method further includes:

[0032] If the probability is greater than a preset fault threshold, the fault category corresponding to the probability is determined as the category of the rolling mill fault.

[0033] Optionally, the training process of the improved ShuffleNetV2 network model includes:

[0034] The dual attention mechanism is introduced into the basic ShuffleNetV2 network to obtain the initial ShuffleNetV2 network;

[0035] Obtain a training set and pre-train the initial ShuffleNetV2 network on the training set through transfer learning to obtain an initial ShuffleNetV2 network model; the weight parameters in the initial ShuffleNetV2 network model are weight parameters with general feature extraction capabilities;

[0036] The initial ShuffleNetV2 network model is iteratively trained using the training set to obtain an improved ShuffleNetV2 network model.

[0037] The present invention provides a multi-component fault diagnosis device for a rolling mill, comprising:

[0038] An acquisition module, used for acquiring status data of multiple components of the rolling mill;

[0039] A conversion module is used to convert the state data into an RGB image through a Gram angle difference field, and then convert the RGB image into three images of R, G, and B channels, and then convert the image of each channel into a grayscale image;

[0040] The dimensionality reduction module is used to perform dimensionality reduction processing on each grayscale image and obtain the RGB image after dimensionality reduction based on the three grayscale images after dimensionality reduction processing;

[0041] The classification module is used to input the reduced-dimensional RGB image into the improved ShuffleNetV2 network model to obtain the fault diagnosis results of the rolling mill; the improved ShuffleNetV2 network model is obtained by adding a dual attention mechanism to the basic ShuffleNetV2 network for training.

[0042] The present invention provides a computer-readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned method for diagnosing faults of multiple components of a rolling mill is implemented.

[0043] The present invention provides a computer device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned method for diagnosing the faults of multiple components of a rolling mill when executing the program.

[0044] At least one of the above technical solutions adopted by the present invention can achieve the following beneficial effects:

[0045] In the present invention, first, the status data of multiple components of the rolling mill are converted into RGB images through the Gram angle difference field. This conversion method can map one-dimensional status data to two-dimensional space, retain the time series information and features in the data, provide richer information for subsequent image processing and feature extraction, and help improve the accuracy of fault diagnosis; and when the RGB image is reduced in dimension, the image is converted into a three-channel image, and each grayscale image is subjected to dimensionality reduction processing separately. In this way, while retaining the main features of the image, the amount of data is reduced, the computational complexity is reduced, and thus the fault diagnosis efficiency is improved; further, ShuffleNetV2 itself is a lightweight network structure with efficient computing performance and good feature extraction capabilities. The addition of a dual attention mechanism to the basic network allows the model to automatically focus on key feature areas and channels when processing images and ignore irrelevant information. Through the dual attention mechanism, the model can more accurately capture the features in the image related to the rolling mill fault, enhance the feature representation capability, and thus improve the accuracy of fault diagnosis. At the same time, since the model can process data more specifically, the calculation of irrelevant information is reduced, which helps to improve the efficiency of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0047] Figure 1 A schematic flow chart of a method for diagnosing multiple components of a rolling mill provided by the present invention;

[0048] Figure 2 This is a structural diagram of the improved ShuffleNetV2 network model;

[0049] Figure 3 This is a schematic diagram of the structure of the lightweight module in the improved ShuffleNetV2 network model;

[0050] Figure 4 Flowchart of the fault diagnosis method for multiple components of a rolling mill using the dimensionality reduction method (RPCA) and dual CBAM-ShuffleNetV2 provided by the present invention;

[0051] Figure 5 The present invention provides a sampled public data set of bearings corresponding to the reduced-dimensional RGB image, wherein (a) is the reduced-dimensional RGB image corresponding to the bearing sample with the fault label H in the sampled public data set of bearings, (b) is the reduced-dimensional RGB image corresponding to the bearing sample with the fault label B7 in the sampled public data set of bearings, (c) is the reduced-dimensional RGB image corresponding to the bearing sample with the fault label IR7 in the sampled public data set of bearings, and (d) is the reduced-dimensional RGB image corresponding to the bearing sample with the fault label IR7 in the sampled public data set of bearings. The RGB image after dimension reduction corresponding to the bearing sample of the sampled bearing public data set with the fault label of B14 is shown in the figure. The RGB image after dimension reduction corresponding to the bearing sample of the sampled bearing public data set with the fault label of IR14 is shown in the figure. The RGB image after dimension reduction corresponding to the bearing sample of the sampled bearing public data set with the fault label of The RGB image after dimension reduction corresponding to the bearing sample of the sampled bearing public data set with the fault label B21 is shown in the figure. The RGB image after dimension reduction corresponding to the bearing sample of the sampled bearing public data set with the fault label IR21 is shown in the figure. The RGB image after dimension reduction corresponding to the bearing sample of the sampled bearing public data set with the fault label IR21 is shown in the figure. The RGB image after dimension reduction corresponding to the bearing sample;

[0052] Figure 6The present invention provides a dimensionality-reduced RGB image corresponding to a rolling mill bearing fault data set, wherein (a) is a dimensionality-reduced RGB image corresponding to the bearing data with a fault label of H in the rolling mill bearing fault data set, (b) is a dimensionality-reduced RGB image corresponding to the bearing data with a fault label of B1 in the rolling mill bearing fault data set, (c) is a dimensionality-reduced RGB image corresponding to the bearing data with a fault label of IR1 in the rolling mill bearing fault data set, and (d) is a dimensionality-reduced RGB image corresponding to the bearing data with a fault label of The RGB image after dimension reduction corresponding to the bearing data of the rolling mill bearing fault data set, (e) is the RGB image after dimension reduction corresponding to the bearing data with fault label B2 in the rolling mill bearing fault data set, (f) is the RGB image after dimension reduction corresponding to the bearing data with fault label IR2 in the rolling mill bearing fault data set, (g) is the RGB image after dimension reduction corresponding to the bearing data with fault label IR2 in the rolling mill bearing fault data set. The RGB image after dimension reduction corresponding to the bearing data;

[0053] Figure 7 A data processing flow chart provided by the present invention;

[0054] Figure 8 A schematic diagram of the variation of the accuracy and loss value of a public bearing dataset at different speeds as a function of the number of iterations provided by the present invention, wherein (a) is a curve showing the variation of the fault diagnosis accuracy of the public bearing dataset at speeds of 1730, 1750, 1772, and 1792 as a function of the number of iterations; (b) is a curve showing the variation of the loss value of the model at speeds of 1730, 1750, 1772, and 1792 as a function of the number of iterations;

[0055] Figure 9 The present invention provides a classification confusion matrix of a public bearing dataset at four speeds, wherein (a) is a classification confusion matrix diagram of the public bearing dataset at a speed of 1730, (b) is a classification confusion matrix diagram of the public bearing dataset at a speed of 1750, (c) is a classification confusion matrix diagram of the public bearing dataset at a speed of 1772, and (d) is a classification confusion matrix diagram of the public bearing dataset at a speed of 1792.

[0056] Figure 10 Schematic diagrams of the accuracy and loss values ​​of a rolling mill bearing fault dataset at three rotational speeds provided by the present invention, wherein (a) is a schematic diagram of the accuracy and loss values ​​of the rolling mill bearing fault dataset at a rotational speed of 600, (b) is a schematic diagram of the accuracy and loss values ​​of the rolling mill bearing fault dataset at a rotational speed of 1200, and (c) is a schematic diagram of the accuracy and loss values ​​of the rolling mill bearing fault dataset at a rotational speed of 1800;

[0057] Figure 11The present invention provides a confusion matrix of a rolling mill bearing fault data set at three speeds, wherein (a) is the classification confusion matrix of the rolling mill bearing fault data set at a speed of 600, (b) is the classification confusion matrix of the rolling mill bearing fault data set at a speed of 1200, and (c) is the classification confusion matrix of the rolling mill bearing fault data set at a speed of 1800.

[0058] Figure 12 The present invention provides a clustering result diagram of a rolling mill bearing fault data set at three speeds, wherein (a) is a data sample diagram of the rolling mill bearing fault data set at a speed of 600, (b) is a cluster diagram of the rolling mill bearing fault data set at a speed of 600, (c) is a data sample diagram of the rolling mill bearing fault data set at a speed of 1200, (d) is a cluster diagram of the rolling mill bearing fault data set at a speed of 1200, (e) is a data sample diagram of the rolling mill bearing fault data set at a speed of 1800, and (f) is a cluster diagram of the rolling mill bearing fault data set at a speed of 1800;

[0059] Figure 13 The present invention provides a dimensionality-reduced RGB image corresponding to a sampled public gear data set, wherein (a) is a dimensionality-reduced RGB image corresponding to gear data with a fault label of healthy gear in the sampled public gear data set, (b) is a dimensionality-reduced RGB image corresponding to gear data with a fault label of gear defect in the sampled public gear data set, (c) is a dimensionality-reduced RGB image corresponding to gear data with a fault label of tooth root wear in the sampled public gear data set, (d) is a dimensionality-reduced RGB image corresponding to gear data with a fault label of tooth surface wear in the sampled public gear data set, and (e) is a dimensionality-reduced RGB image corresponding to gear data with a fault label of broken tooth in the sampled public gear data set;

[0060] Figure 14 The present invention provides a reduced-dimensional RGB image corresponding to a rolling mill gear fault dataset, wherein (a) is a reduced-dimensional RGB image corresponding to gear data with a fault label of healthy gear in the rolling mill gear fault dataset, (b) is a reduced-dimensional RGB image corresponding to gear data with a fault label of tooth root wear in the rolling mill gear fault dataset, (c) is a reduced-dimensional RGB image corresponding to gear data with a fault label of tooth surface wear in the rolling mill gear fault dataset, and (d) is a reduced-dimensional RGB image corresponding to gear data with a fault label of broken teeth in the rolling mill gear fault dataset;

[0061] Figure 15The present invention provides a curve showing how the accuracy and loss value of a sampled public gear dataset under different vibration conditions vary with the number of iterations. (a) shows a curve showing how the fault diagnosis accuracy of the sampled public gear dataset varies with the number of iterations at vibration frequencies of 20 and 30, and (b) shows a curve showing how the loss value of the model varies with the number of iterations at vibration frequencies of 20 and 30.

[0062] Figure 16 The confusion matrix of a sampled public gear dataset provided by the present invention under two vibration conditions, wherein (a) is the confusion matrix of the sampled public gear dataset under a vibration frequency of 20, and (b) is the confusion matrix of the sampled public gear dataset under a vibration frequency of 30;

[0063] Figure 17 The present invention provides a curve showing the accuracy and loss value of a rolling mill gear fault dataset at three speeds as a function of the number of iterations. (a) shows the curve showing the accuracy and loss value of a rolling mill gear fault dataset at a speed of 600 as a function of the number of iterations. (b) shows the curve showing the accuracy and loss value of a rolling mill gear fault dataset at a speed of 1200 as a function of the number of iterations. (c) shows the curve showing the accuracy and loss value of a rolling mill gear fault dataset at a speed of 1800 as a function of the number of iterations.

[0064] Figure 18 The present invention provides a confusion matrix of a rolling mill gear fault dataset at three speeds, wherein (a) is the classification confusion matrix of the rolling mill gear fault dataset at a speed of 600, (b) is the classification confusion matrix of the rolling mill gear fault dataset at a speed of 1200, and (c) is the classification confusion matrix of the rolling mill gear fault dataset at a speed of 1800.

[0065] Figure 19 The present invention provides a clustering result diagram of a rolling mill gear fault dataset at three speeds, wherein (a) is a data sample diagram of the rolling mill gear fault dataset at a speed of 600, (b) is a cluster diagram of the rolling mill gear fault dataset at a speed of 600, (c) is a data sample diagram of the rolling mill gear fault dataset at a speed of 1200, (d) is a cluster diagram of the rolling mill gear fault dataset at a speed of 1200, (e) is a data sample diagram of the rolling mill gear fault dataset at a speed of 1800, and (f) is a cluster diagram of the rolling mill gear fault dataset at a speed of 1800;

[0066] Figure 20 A schematic diagram of a computer device for implementing a method for diagnosing multiple components of a rolling mill provided by the present invention. DETAILED DESCRIPTION

[0067] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present invention and corresponding drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0068] The present invention provides a method for diagnosing the faults of multiple components of a rolling mill, which can realize the fault diagnosis of multiple components of the rolling mill at the same time.

[0069] The technical solutions provided by various embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0070] Figure 1 The present invention is a flowchart of a method for diagnosing a fault of multiple components of a rolling mill, which specifically includes the following steps:

[0071] S101, obtaining status data of multiple components of a rolling mill.

[0072] Among them, the multiple components of the rolling mill may include bearings and gears of the rolling mill, and vibration signals of the bearings and gears of the rolling mill are mainly collected through sensors, and the vibration signals are used as status data.

[0073] S102, converting the state data into an RGB image through a Gram angle difference field.

[0074] The Gramian Angular Difference Field (GADF) is a method for converting one-dimensional time series data (such as vibration signals) into a two-dimensional image. It uses trigonometric functions (cosine and sine) to encode the values ​​in the time series and the relationships between them. The core idea is to map the time series values ​​into polar coordinate space and then construct an image by calculating the Gramian matrix.

[0075] Specifically, the maximum value in the vibration signal is calculated, and each value in the vibration signal is mapped to the interval [0, 1]. This value is then converted to an angle in polar coordinates (using the inverse cosine function). The Gram difference field matrix is ​​constructed by calculating the cosine of the difference between the angles. To obtain an RGB image, a three-channel matrix is ​​constructed. Here, the Gram difference field matrix is ​​simply copied three times, serving as the values ​​for the R, G, and B channels, respectively. These are then combined to form a three-dimensional array representing the RGB image.

[0076] S103, converting the RGB image into three images of R, G, and B channels, and converting the image of each channel into a grayscale image; performing dimensionality reduction processing on each grayscale image respectively, and obtaining a reduced-dimensional RGB image based on the three grayscale images after dimensionality reduction processing.

[0077] The preset conversion function can be used to convert the RGB image into three images of R, G, and B channels, and the image of each channel can be converted into a grayscale image.

[0078] Optionally, dimensionality reduction processing is performed on each grayscale image, including: for any grayscale image, obtaining the mean of all data points in the grayscale image; determining the scatter matrix based on the value and mean of each data point, and solving the eigenvalue and eigenvector of the scatter matrix; reducing the maximum front K The eigenvector corresponding to the eigenvalue is determined as a new eigenvector; and the grayscale image after dimensionality reduction is determined according to the new eigenvector and the original grayscale image.

[0079] Specifically, consider each grayscale image as an image matrix X, which is expressed as: .

[0080] in, ; A column vector representing the image matrix; Represents the number of column vectors in the image matrix.

[0081] Find the mean of the image matrix , scatter matrix and covariance matrix ,as follows:

[0082] (1);

[0083] (2);

[0084] (3).

[0085] Solve the eigenvalues ​​and eigenvectors of the covariance matrix: It can be found that the divergence matrix is ​​the covariance matrix multiplied by (total data volume - 1). Therefore, their eigenvalues ​​and eigenvectors are the same. So we can solve the eigenvalues ​​and eigenvectors of the divergence matrix by solving the characteristic equation , find the eigenvalue and the corresponding eigenvectors ,in Represents the number of eigenvectors. The eigenvectors of the covariance matrix are expressed as .

[0086] Arrange the eigenvectors of the covariance matrix according to the size of the eigenvalue and take the largest one The eigenvectors are used as row vectors to form a new eigenmatrix .

[0087] The image data is converted into the dimensional space constructed by the new feature matrix, that is, the principal component matrix of the new training image is , and get the grayscale image after dimensionality reduction.

[0088] The three grayscale images after dimensionality reduction are merged into a new image according to the original decomposition method. This image is the RGB image after dimensionality reduction.

[0089] The above-mentioned dimensionality reduction method is a novel approach for reducing the dimensionality of two-dimensional image data. It reduces the dimensionality of RGB images, enhancing the data's features and making it more suitable for convolutional neural network recognition. This is particularly true for lightweight networks, which process larger data more slowly, thus failing to meet their desired characteristics of fewer parameters, less computational effort, and shorter inference times. The dimensionality reduction method proposed in this invention can significantly reduce the size of the data itself, making it more suitable for lightweight network applications.

[0090] S104, inputting the RGB image after dimension reduction into the improved ShuffleNetV2 network model to obtain the fault diagnosis result of the rolling mill; the improved ShuffleNetV2 network model is obtained by adding a dual attention mechanism to the basic ShuffleNetV2 network for training.

[0091] like Figure 2 As shown, Figure 2 It is a structural diagram of the improved ShuffleNetV2 network model; specifically, the improved ShuffleNetV2 network model includes an input module integrating an attention mechanism, multiple lightweight modules connected in series in sequence, and an output module integrating an attention mechanism; the reduced-dimensional RGB image is input into the improved ShuffleNetV2 network model to obtain a fault diagnosis result of the rolling mill, including: inputting the reduced-dimensional RGB image into the improved ShuffleNetV2 network model, extracting features in the reduced-dimensional RGB image through the input module; performing convolution operations on the features in the reduced-dimensional RGB image through multiple lightweight modules connected in series in sequence to obtain convolution features; performing feature prediction on the convolution features through the output module to obtain a fault diagnosis result of the rolling mill.

[0092] It should be noted that Figure 2 The improved ShuffleNetV2 network model is illustrated in six modules, wherein the first module includes an input module, the second module, the third module, the fourth module and the fifth module each include a lightweight module, and the sixth module includes an output module. Figure 2The example uses four lightweight modules. In actual applications, there is no limit on the number of lightweight modules in the improved ShuffleNetV2 network model.

[0093] Optionally, the input module includes an input layer, a convolution layer, an activation function, an attention mechanism layer and a fully connected layer connected in series in sequence; extracting features in the reduced-dimensional RGB image through the input module includes: receiving the reduced-dimensional RGB image through the input layer, and performing a convolution operation on the reduced-dimensional RGB image through the convolution layer to extract local features; increasing the nonlinear characteristics of the local features through the activation function to obtain nonlinear features; assigning weights to different features in the nonlinear features through the attention mechanism layer; performing a fully connected operation on the features output by the attention mechanism layer through the fully connected layer to obtain features in the reduced-dimensional RGB image.

[0094] Optionally, the attention mechanism layer includes a channel-domain attention module and a spatial-domain attention module, which are cascaded. The channel-domain attention module and the spatial-domain attention module are convolutional block attention modules (CBAMs).

[0095] Optionally, the output module includes a convolution layer, an attention mechanism layer and a fully connected layer connected in series in sequence; the convolution features are predicted by the output module to obtain the fault diagnosis results of the rolling mill, including: convolution processing of the convolution features by the convolution layer, and assigning weights to the convolved features by the attention mechanism layer; mapping the features output by the attention mechanism layer to various fault categories by the fully connected layer to obtain the fault diagnosis results of the rolling mill.

[0096] Optionally, the fault diagnosis result of the rolling mill includes the probability of each fault category occurring in each component of the rolling mill; this embodiment includes: if the probability is greater than a preset fault threshold, then determining the fault category corresponding to the probability as the category of the rolling mill fault.

[0097] Optionally, the training process of the improved ShuffleNetV2 network model includes: introducing a dual attention mechanism into the basic ShuffleNetV2 network to obtain an initial ShuffleNetV2 network; obtaining a training set, and pre-training the initial ShuffleNetV2 network on the training set through transfer learning to obtain an initial ShuffleNetV2 network model; the weight parameters in the initial ShuffleNetV2 network model are weight parameters with general feature extraction capabilities; and iteratively training the initial ShuffleNetV2 network model through the training set to obtain an improved ShuffleNetV2 network model.

[0098] Transfer learning can be simply defined as: by pre-training the initial ShuffleNetV2 network on the large-scale ImageNet training set, weight parameters with universal feature extraction capabilities are obtained. This model, through learning from massive amounts of data, already possesses a multi-level feature representation system, and can still achieve high recognition accuracy for small sample data sets. Because the weight parameters have strong feature transfer capabilities, they can be directly introduced to extract features from other similar data sets; while improving the efficiency of network model development, it can also enhance the model's generalization performance and convergence efficiency. Especially for rolling mill simulation platforms, the number of feature samples in the bearing and gear data obtained is relatively small, and the model needs to converge quickly to achieve fault diagnosis and classification of rolling mill bearings and gears. Using a transfer learning strategy that transfers similar data to achieve rapid model convergence and thus improve the model's recognition accuracy is an excellent way to achieve rapid model convergence and improve the model's recognition accuracy.

[0099] The design goal of CBAM is to solve the problem that traditional convolutional neural networks are unable to represent multi-scale, morphological transformation and direction-sensitive information during feature extraction.

[0100] CBAM effectively improves the performance bottlenecks of traditional convolutional neural networks in terms of adaptability to multi-scale geometric features, generalization of topological structures, and robustness of directional perception through a dual-modal attention collaboration mechanism. Its technical implementation includes two core components: the channel-domain attention module realizes the dynamic recalibration of feature channel weights and optimizes the efficiency of cross-channel information integration; the spatial-domain attention module realizes adaptive allocation of spatial weights through learnable convolution kernels and strengthens the representation strength of local key areas. The two modules adopt a cascaded architecture design and can be flexibly integrated into feature extraction nodes at all levels of the deep network. Through a step-by-step feature distillation mechanism, they achieve adaptive enhancement of multi-level heterogeneous features and realize refined modeling of complex spatial structures. This type of design significantly improves the network's discriminative expression ability for heterogeneous features, which meets the optimization requirements of the mechanical field for model robustness and generalization performance.

[0101] The input layer receives raw image data, while the convolutional layer convolves the input data with convolution kernels to extract local features. Subsequently, activation functions (such as ReLU) are introduced to increase nonlinearity and enhance the model's expressiveness. Fully connected layers downsample feature maps to reduce computational complexity and improve model generalization. This is achieved by selecting the maximum or average value within a pooling window. This helps extract the most important features. Lightweight modules typically consist of multiple stacked convolutional and fully connected layers to extract progressively higher-level features. Deeper features can represent more complex patterns. Finally, fully connected layers transform the extracted feature maps into the network's final output. Fully connected and convolutional layers offer significant advantages in data feature extraction and are particularly well-suited for processing two-dimensional images.

[0102] The improved ShuffleNetV2 network model adds CBAM attention mechanism modules to the first and last convolutional layers. The improved network can better utilize its different depth hierarchical structures to extract features of different dimensions of the input two-dimensional image data.

[0103] In a specific embodiment, Figure 3 As shown, Figure 3 This is a structural diagram of the lightweight module in the improved ShuffleNetV2 network model, specifically including: starting from the input of the lightweight module, it includes two branches, one branch includes a convolutional layer, a batch normalization layer, an activation function, a deep convolutional layer, a batch normalization layer, a convolutional layer, a batch normalization layer, and an activation function connected in series, and the other branch includes a deep convolutional layer, a batch normalization layer, a convolutional layer, a batch normalization layer, and an activation function connected in series. The output of the last activation function of both branches is connected to the fully connected layer, the output of the fully connected layer is the input of the anonymous function, and the output of the anonymous function is the output of the lightweight module.

[0104] The improved ShuffleNetV2 network model in the present invention is a double (Double) CBAM-ShuffleNetV2 network model, such as Figure 4 As shown, Figure 4 The flowchart of the fault diagnosis method for multiple components of a rolling mill using the dimensionality reduction method (RPCA) and dual CBAM-ShuffleNetV2 provided by the present invention is as follows:

[0105] (1) Data acquisition: Public bearing and gear datasets were used. Bearing vibration data were obtained from the roller bearing seat of the rolling mill test platform in the laboratory. Gear vibration data were obtained from the external connecting shaft of the driven wheel of the reduction gearbox. 20 seconds of stable data were selected as the original dataset of the rolling mill bearings and gears.

[0106] (2) Data processing: The selected public dataset and the original dataset of bearings and gears obtained from the rolling mill platform are converted into two-dimensional RGB image data using the Gram angle difference field. Then, RPCA is used to reduce the dimension of the RGB image data and highlight its features. The data are given fault labels for different faults and divided into training and testing datasets.

[0107] (3) Network training: The processed image data is input into the dual CABM-ShuffleNetV2 network. After data enhancement, the pre-trained module on the ImageNet dataset is migrated to the model training module. The number of classification nodes in the last layer is changed to the number of classification nodes in the input training set to adapt to the network's classification of the dataset. Finally, after each training, the weight of the model with the highest accuracy is saved and returned.

[0108] (4) Fault diagnosis and classification: The weights and test sets that return the highest accuracy are input into the network for fault classification. Visual confusion matrices and cluster diagrams are drawn based on the classification results to further illustrate the classification effect. The feasibility of the proposed model framework is verified using public datasets, and the framework is applied using experimental data on rolling mill bearings and gears.

[0109] In one implementation, to test the performance of the improved ShuffleNetV2 network model, verify its effectiveness, and apply it to rolling mill bearing and gear fault diagnosis and classification, experiments were conducted on two datasets: a publicly available dataset of bearings from a university and a rolling mill bearing fault dataset, and a publicly available dataset of gears from another university and a rolling mill gear fault dataset. Two comparisons were also conducted: one with other deep learning-based methods to assess the feasibility and applicability of the model framework for bearing and gear fault diagnosis, and the other with existing methods validated on bearing and gear datasets to evaluate the superiority of the model framework for rolling mill bearing and gear fault diagnosis.

[0110] All network models were trained using Python 3.8 under the PyTorch framework, using an Intel Core i5-7300HQ CPU @ 2.5 GHz and a GTX1050 (4GB) GPU under Windows 11.

[0111] Case 1: Verifying the feasibility and effectiveness of the rolling mill fault diagnosis model framework for bearing fault diagnosis

[0112] (1) Data division

[0113] To validate the effectiveness of the proposed framework for bearing diagnosis, datasets were used from a public bearing dataset and the research group's rolling mill bearing fault dataset. The public bearing dataset was sourced from a university's bearing data center. The test platform used consisted of a motor, a sensor, a power meter, and electronic monitoring equipment, primarily for signal acquisition and detection of the bearing supporting the motor. The test bearing was a deep groove ball bearing, and the fault was a man-made fault, resulting from single-point fault machining using electric sparks. The dataset consisted of a healthy (H) and three man-made faults: an inner race fault (IR), a rolling element fault (B), and an outer race fault (OR). Each fault state had three fault sizes: 0.007, 0.014, and 0.021 inches. Four operating conditions were selected: 0HP, 1HP, 2HP, and 3HP.

[0114] The present invention uses a sampling frequency of 12 kHz, a motor load of 0.75 kW, and operating conditions at four different speeds. Three different fault diameters corresponding to the inner ring, outer ring, and rolling element (6 o'clock direction) are selected as fault data samples, plus a healthy state data sample, for a total of ten categories. Data processing generates 600 sample image data for each category, which are divided into training and test sets in an 8:2 ratio. The data partitioning method in Table 1 is used to verify the feasibility of the proposed rolling mill fault diagnosis model framework for bearing classification.

[0115] Table 1. Sample bearing fault types in the public bearing dataset

[0116]

[0117] Among them, such as Figure 5 As shown, Figure 5 This is the RGB image after dimension reduction corresponding to the sampled bearing public dataset. Figure 5 Figure (a) is the RGB image after dimension reduction corresponding to the bearing sample with fault label H in the sampled public bearing dataset. Figure 5 Figure (b) is the RGB image after dimension reduction corresponding to the bearing sample with fault label B7 in the sampled public bearing dataset. Figure 5 Figure (c) is the RGB image after dimension reduction corresponding to the bearing sample with fault label IR7 in the sampled public bearing dataset. Figure 5 Figure (d) shows the fault labels of the sampled bearing public dataset. The RGB image after dimension reduction corresponding to the bearing sample, Figure 5 Figure (e) is the RGB image after dimension reduction corresponding to the bearing sample with fault label B14 in the sampled public bearing dataset. Figure 5 Figure (f) is the RGB image after dimension reduction corresponding to the bearing sample with fault label IR14 in the sampled public bearing dataset. Figure 5 Figure (g) shows the fault labels of the sampled bearing public dataset. The RGB image after dimension reduction corresponding to the bearing sample, Figure 5 Figure (h) is the RGB image after dimension reduction corresponding to the bearing sample with fault label B21 in the sampled public bearing dataset. Figure 5 Figure (i) is the RGB image after dimension reduction corresponding to the bearing sample with fault label IR21 in the sampled public bearing dataset. Figure 5 Figure (j) shows the fault labels of the sampled bearing public dataset. The RGB image after dimension reduction corresponding to the bearing sample.

[0118] Among them, the rolling mill bearing fault data set comes from a rolling mill vibration simulation experimental platform of a certain research group. The experimental platform used includes a motor, a horizontal and vertical exciter, a planetary gearbox, a reducer, an operating table and an electronic acquisition device. The two acceleration sensors for collecting the vibration signals of the rolling mill bearings are located in the vertical and horizontal positions of the upper working roll. The four acceleration sensors for collecting the vibration signals of the rolling mill gears are located in the vertical and horizontal positions of the driving wheel and the driven wheel of the reducer. It mainly collects and detects signals from the gears connecting the roller's external bearings and the reducer. The failures of the bearings and gears are man-made failures, and single-point failure processing is performed by electric sparks. In order to collect data close to the public data set, the control variable method is used. The rolling mill bearing fault data set also consists of health (H) and three man-made faults, namely inner ring fault (IR), rolling element fault (B) and outer ring fault ( )composition.

[0119] The bearing operating conditions used were 1HP and 2HP, and three speeds of 600, 1200, and 1800. The bearing fault categories sampled at each speed included six fault states and one healthy state, for a total of seven categories. Data processing generated 400 sample images for each category, which were then divided into training and test sets with an 8:2 ratio. The data partitioning method in Table 2 was used to verify the effectiveness of the proposed rolling mill fault diagnosis model framework for rolling mill bearing classification.

[0120] Table 2 Classification of bearing status data sets

[0121]

[0122] like Figure 6 As shown, Figure 6 This is the RGB image after dimension reduction corresponding to the rolling mill bearing fault dataset. Figure 6 Figure (a) is the RGB image after dimension reduction corresponding to the bearing data with fault label H in the rolling mill bearing fault dataset. Figure 6Figure (b) is the dimensionality-reduced RGB image corresponding to the bearing data with fault label B1 in the rolling mill bearing fault dataset. Figure 6 Figure (c) is the RGB image after dimension reduction corresponding to the bearing data with fault label IR1 in the rolling mill bearing fault dataset. Figure 6 Figure (d) shows the fault labels in the rolling mill bearing fault dataset. The RGB image after dimension reduction corresponding to the bearing data, Figure 6 Figure (e) is the RGB image after dimension reduction corresponding to the bearing data with fault label B2 in the rolling mill bearing fault dataset. Figure 6 Figure (f) is the RGB image after dimension reduction corresponding to the bearing data with fault label IR2 in the rolling mill bearing fault dataset. Figure 6 Figure (g) shows the fault labels in the rolling mill bearing fault dataset. The RGB image after dimension reduction corresponding to the bearing data.

[0123] (2) Bearing data processing

[0124] Since convolutional neural networks have more advantages in processing image data, and color images have more layers and can better display data features, GADF is used in the present invention to convert the one-dimensional vibration signal into two-dimensional color, and then RPCA is used to reduce the dimension of the two-dimensional color image. While reducing noise, the data features are more prominent, which makes it easier for the network to capture the features, and indirectly improves the training speed of the network.

[0125] Due to the limitation of data volume, half-overlap sampling method is used to capture more data features when using GADF. The specific data processing process is as follows: Figure 7 shown.

[0126] (3) Comparative analysis of bearing data experiments

[0127] like Figure 8 As shown, Figure 8 This is a diagram showing how the accuracy and loss of the public bearing dataset change with the number of iterations at different speeds. The four speeds are: 1730, 1750, 1772, and 1792. Figure 8 Figure (a) shows the curve of the fault diagnosis accuracy of the bearing public data set at speeds of 1730, 1750, 1772 and 1792 as the number of iterations changes. Figure 8 Figure (b) shows the curve of the loss value of the model changing with the number of iterations at speeds of 1730, 1750, 1772 and 1792 for the public bearing dataset. Figure 8 This fully demonstrates the feasibility of the model framework under variable speed.

[0128] In order to further illustrate the feasibility and classification effect of dual CABM-ShuffleNetV2 for bearing classification, the following Figure 9 The following figure shows the classification confusion matrix for a public bearing dataset at four speeds: (a) for a public bearing dataset at a speed of 1730, (b) for a public bearing dataset at a speed of 1750, (c) for a public bearing dataset at a speed of 1772, and (d) for a public bearing dataset at a speed of 1792. Each node in the horizontal and vertical coordinates of the classification confusion matrix corresponds to a set fault type, and the elements on the diagonal represent the overlap between the correct label and the predicted label. It can be seen that the number of misclassifications is very small, indicating that the feasibility of this model framework for bearing fault diagnosis is guaranteed.

[0129] To further verify the feasibility of the model for bearing fault diagnosis, we compared it with other convolutional neural networks of the same type, namely ResNeXt with transfer learning module, ResNeXt without transfer learning module, residual network TL-ResNet with transfer learning but without grouping, and shallow networks GoogLeNet and AlexNet. To clearly highlight the feasibility of the model, we selected several common evaluation criteria in the field of deep learning fault diagnosis, such as accuracy. , precision , recall rate 、 Score, the expressions of several indicators are as follows:

[0130] ;

[0131] ;

[0132] ;

[0133] ;

[0134] in: 、 They are i The number of correct predictions in each category; 、 They are i The number of incorrect predictions in each category.

[0135] The comparison of each model parameter is shown in Table 3.

[0136] Table 3 Parameters of different network models for a certain university bearing

[0137]

[0138] To further illustrate the superiority of the model, the present invention also compares some existing networks that have been studied on bearing datasets, as shown in Table 4. The existing networks include: UFE-ResNet50, L-CNN, IS-DATN, ADA, and GGRU-1DCNN. From the accuracy alone, it can be clearly seen that the model of the present invention is superior to bearing diagnosis.

[0139] Table 4 Comparison of accuracy of different algorithms

[0140]

[0141] (4) Application in rolling mill bearing diagnosis

[0142] like Figure 10 As shown, Figure 10 The figure is a schematic diagram of the accuracy and loss value of the rolling mill bearing fault dataset at three speeds, among which (a) is a schematic diagram of the accuracy and loss value of the rolling mill bearing fault dataset at a speed of 600, (b) is a schematic diagram of the accuracy and loss value of the rolling mill bearing fault dataset at a speed of 1200, and (c) is a schematic diagram of the accuracy and loss value of the rolling mill bearing fault dataset at a speed of 1800; it shows that very good accuracy and small loss can be achieved at three different speeds, which fully demonstrates the effectiveness of the model framework for rolling mill bearing diagnosis at variable speeds.

[0143] In order to further illustrate the effectiveness and classification effect of dual CABM-ShuffleNetV2 on rolling mill bearing diagnosis, the present invention draws the following Figure 11 、 Figure 12 The classification confusion matrix and clustering result diagram under three speeds are shown, where: Figure 11 Figure (a) is the classification confusion matrix of the rolling mill bearing fault dataset at a speed of 600. Figure 11 Figure (b) is the classification confusion matrix of the rolling mill bearing fault dataset at a speed of 1200. Figure 11 Figure (c) is the classification confusion matrix of the rolling mill bearing fault dataset at a speed of 1800. Figure 12 Figure (a) is a sample diagram of the rolling mill bearing fault data set at a speed of 600. Figure 12 Figure (c) is a sample diagram of the rolling mill bearing fault data set at a speed of 1200. Figure 12 Figure (e) is a sample diagram of the rolling mill bearing fault data set at a speed of 1800. Figure 12 Figure (b) is the cluster diagram of the rolling mill bearing fault data set at a speed of 600. Figure 12Figure (d) is the cluster diagram of the rolling mill bearing fault data set at a speed of 1200. Figure 12 Figure (f) is the cluster diagram of the rolling mill bearing fault data set at a speed of 1800. Figure 12 It can be seen that the number of misclassified items is very small, and the clustering diagram also shows that the clustering effect is obvious, indicating that this model framework is effective for fault diagnosis of rolling mill bearings and the classification effect is very significant.

[0144] Case 2: Verifying the feasibility and effectiveness of the rolling mill fault diagnosis model framework for gear classification

[0145] (1) Data division

[0146] To verify the effectiveness of the proposed framework for gear diagnosis, datasets were used from a public gear dataset and a rolling mill gear fault dataset. The public gear dataset was sourced from another university's gear database. The experiment used the Drivetrain Dynamic Simulator (DDS) platform, a gear testing platform. This platform consists of a motor, a motor controller, a planetary gearbox, a reduction gearbox, a brake, and a brake controller. Its primary function is to collect and detect gear signals. The test gears were standard components, and the gear faults were man-made, using single-point fault machining (EDM) using electric sparks. Four fault categories were identified: gear defect, tooth breakage, tooth root wear, and tooth surface wear. In addition, five categories were considered for the normal state, as shown in Table 5. The gear operating conditions were selected as either 0HP (no load) or 1HP (load).

[0147] The present invention selects five types of fault sample data under two working conditions. 600 sample image data are generated for each category through data processing. The data are divided into training set and test set according to the ratio of 8:2. The data division method in Table 5 is used to verify the feasibility of the proposed rolling mill fault diagnosis model for gear classification.

[0148] Table 5 Gear fault type samples from the gear public dataset

[0149]

[0150] The rolling mill gear fault dataset includes three fault types: broken tooth, tooth root wear, and tooth surface wear, plus the normal state, for a total of four categories, as shown in Table 6. The gear operating conditions are the same as those for rolling mill bearings: 600, 1200, and 1800 speeds. Data processing generates 400 sample images for each category, which are then divided into training and test sets with an 8:2 ratio. The data partitioning method shown in Table 6 is used to verify the effectiveness of the proposed rolling mill fault diagnosis model for gear classification.

[0151] Table 6 Classification of gear state data sets

[0152]

[0153] (2) Gear data processing

[0154] The same method as bearing data processing is used, such as Figure 13 Shown and Figure 14 As shown, Figure 13 This is the RGB image after dimension reduction corresponding to the sampled gear public dataset. Figure 14 This is the RGB image after dimension reduction corresponding to the rolling mill gear fault dataset. Specifically, Figure 13 Figure (a) is the RGB image after dimension reduction corresponding to the gear data with the fault label of healthy gear in the sampled gear public dataset. Figure 13 Figure (b) is the RGB image after dimension reduction corresponding to the gear data with the fault label of gear defect in the sampled gear public dataset. Figure 13 Figure (c) is the RGB image after dimension reduction corresponding to the gear data with the fault label of tooth root wear in the sampled gear public dataset. Figure 13 Figure (d) is the RGB image after dimension reduction corresponding to the gear data with the fault label of tooth surface wear in the sampled gear public dataset. Figure 13 Figure (e) is the RGB image after dimension reduction corresponding to the gear data with the fault label of broken teeth in the sampled gear public dataset; Figure 14 Figure (a) is the RGB image after dimension reduction corresponding to the gear data with the fault label of healthy in the rolling mill gear fault dataset. Figure 14 Figure (b) is the RGB image after dimension reduction corresponding to the gear data with the fault label of tooth root wear in the rolling mill gear fault dataset. Figure 14 Figure (c) is the RGB image after dimension reduction corresponding to the gear data with the fault label of tooth surface wear in the rolling mill gear fault dataset. Figure 14 Figure (d) is the RGB image after dimensionality reduction corresponding to the gear data with the fault label of broken tooth in the rolling mill gear fault dataset.

[0155] (3) Comparative analysis of gear data experiments from another university

[0156] like Figure 15 As shown, Figure 15 The following is a curve showing the accuracy and loss of the sampled gear public dataset under different vibration conditions as a function of the number of iterations. The vibration conditions include vibration frequencies of 20 and 30. Figure 15 Figure (a) shows the curve of the fault diagnosis accuracy of the sampled gear public data set at vibration frequencies of 20 and 30 as the number of iterations changes. Figure 15 Figure (b) shows the curve of the loss value of the model changing with the number of iterations at vibration frequencies of 20 and 30 for the sampled gear public dataset.

[0157] Very good accuracy can be achieved at two different speeds, which fully demonstrates the feasibility of the model framework under variable speed.

[0158] In order to further illustrate the feasibility and classification effect of dual CABM-ShuffleNetV2 for gear classification, the following Figure 16 The following figure shows the confusion matrices for a publicly available dataset of sampled gears under two vibration conditions. (a) shows the confusion matrix for the dataset at a vibration frequency of 20, and (b) shows the confusion matrix for the dataset at a vibration frequency of 30. Each node in the confusion matrix corresponds to a fault type, and the elements on the diagonal represent the overlap between the correct label and the predicted label. The number of misclassifications is very small, demonstrating the feasibility of this model framework for bearing fault diagnosis.

[0159] Similar to the bearing comparison of the same type of network, the comparison of model parameters is shown in Table 7.

[0160] Table 7 Parameters of different network models for the sampled gear public dataset

[0161]

[0162] In order to further illustrate the superiority of the model for gear fault diagnosis, the present invention is compared with the existing network comparison on another university (SUFD) gear public dataset. The existing networks include: SAE-DNN, GRU, BIGRU, LFGRU, VGG; the results are shown in Table 8. Compared with the gear classification network used above, the proposed model has higher classification accuracy.

[0163] Table 8 Comparison of accuracy of different algorithms

[0164]

[0165] (4) Application in rolling mill gear diagnosis

[0166] like Figure 17 As shown, Figure 17 The curves of the accuracy and loss value of the rolling mill gear fault data set at three speeds as the number of iterations changes are shown. Figure 17 Figure (a) shows the curve of the accuracy and loss value of the rolling mill gear fault dataset at a speed of 600 as the number of iterations changes. Figure 17 Figure (b) shows the curve of the accuracy and loss value of the rolling mill gear fault dataset at a speed of 1200 as the number of iterations changes. Figure 17Figure (c) shows the curve of the accuracy and loss value of the rolling mill gear fault dataset at a speed of 1800 as the number of iterations changes. Very good accuracy and small loss can be achieved at three different speeds, which fully demonstrates the effectiveness of the model framework for rolling mill gear diagnosis at variable speeds.

[0167] In order to further illustrate the effectiveness and classification effect of dual CABM-ShuffleNetV2 on rolling mill gear diagnosis, the following graph is drawn: Figure 18 、 Figure 19 As shown, Figure 18 and Figure 19 They are respectively the classification confusion matrix and clustering result diagram of the rolling mill gear fault data set at three speeds, among which, Figure 18 Figure (a) is the classification confusion matrix of the rolling mill gear fault dataset at a speed of 600. Figure 18 Figure (b) is the classification confusion matrix of the rolling mill gear fault dataset at a speed of 1200. Figure 18 Figure (c) is the classification confusion matrix of the rolling mill gear fault dataset at a speed of 1800. Figure 19 Figure (a) is a sample diagram of the rolling mill gear fault data set at a speed of 600. Figure 19 Figure (c) is a sample diagram of the rolling mill gear fault dataset at a speed of 1200. Figure 19 Figure (e) is a sample diagram of the rolling mill gear fault data set at a speed of 1800. Figure 19 Figure (b) is the cluster diagram of the rolling mill gear fault data set at a speed of 600. Figure 19 Figure (d) is the cluster diagram of the rolling mill gear fault dataset at a speed of 1200. Figure 19 Figure (f) is the cluster diagram of the rolling mill gear fault data set at a speed of 1800. Figure 19 It can be seen that the number of misclassified items is very small, and the clustering diagram also shows that the clustering effect is obvious, indicating that this model framework is effective for fault diagnosis of rolling mill gears and the classification effect is very significant.

[0168] The rolling mill multi-component fault diagnosis method of the present invention has the following advantages:

[0169] 1) Taking advantage of the superiority of convolutional neural networks in image data recognition, GADF is used to convert the original one-dimensional vibration signal into two-dimensional time-frequency image data.

[0170] 2) RPCA reduces the size of the data by performing noise reduction and feature enhancement on the two-dimensional time-frequency image data. This not only preserves more data information and enhances features, but also reduces the data size, significantly reducing the number of parameters, computational complexity, and inference time of the lightweight network.

[0171] 3) A dual CABM module is used to enhance the perception capability of the ShuffleNetV2 network model, and transfer learning is inserted into the network model to speed up the convergence of the model.

[0172] 4) The model was validated using validated public datasets for both bearing and gear cases. Accuracy was maintained across various operating conditions, and comparisons with other models using the same datasets demonstrated its superiority. Finally, the proposed model was applied to a rolling mill simulation platform. The experimental results demonstrate the effectiveness of the proposed model framework for multi-component fault diagnosis in rolling mills.

[0173] When applying the method for diagnosing the faults of multiple components of a rolling mill provided by the present invention, it is not necessary to Figure 1 The steps are executed in the order shown. The specific execution order of the steps can be determined according to needs, and the present invention does not limit this.

[0174] The above is a method for diagnosing a fault of multiple components of a rolling mill provided by one or more embodiments of the present invention. Based on the same idea, the present invention also provides a corresponding device for diagnosing a fault of multiple components of a rolling mill, which device includes:

[0175] An acquisition module, used for acquiring status data of multiple components of the rolling mill;

[0176] A conversion module is used to convert the state data into an RGB image through a Gram angle difference field, and then convert the RGB image into three images of R, G, and B channels, and then convert the image of each channel into a grayscale image;

[0177] The dimensionality reduction module is used to perform dimensionality reduction processing on each grayscale image and obtain the RGB image after dimensionality reduction based on the three grayscale images after dimensionality reduction processing;

[0178] The classification module is used to input the reduced-dimensional RGB image into the improved ShuffleNetV2 network model to obtain the fault diagnosis results of the rolling mill; the improved ShuffleNetV2 network model is obtained by adding a dual attention mechanism to the basic ShuffleNetV2 network for training.

[0179] The specific definitions of the multi-component rolling mill fault diagnosis device can be found in the definitions of the multi-component rolling mill fault diagnosis method described above and will not be further elaborated here. Each module in the multi-component rolling mill fault diagnosis device described above can be implemented in whole or in part via software, hardware, or a combination thereof. Each of these modules can be embedded in or independent of a processor in a computer device in hardware form, or stored in a computer device memory in software form, so that the processor can call and execute the corresponding operations of each module.

[0180] The present invention also provides a computer-readable storage medium, which stores a computer program, which can be used to execute the above Figure 1 A fault diagnosis method for multiple components of a rolling mill is provided.

[0181] The present invention also provides Figure 20 The structural diagram of the computer equipment shown in FIG. Figure 20 As shown in the figure, at the hardware level, the computer device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory. Of course, it may also include other hardware required for the business. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to achieve the above Figure 1 A fault diagnosis method for multiple components of a rolling mill is provided.

[0182] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware using a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes in the above-described method embodiments. Any reference to memory, storage, database, or other media used in the various embodiments provided herein may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can take various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0183] The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of the present invention.

Claims

1. A method for diagnosing faults of multiple components of a rolling mill, characterized in that: include: Acquiring status data of multiple components of a rolling mill; The state data is converted into RGB image through Gram angle difference field; Convert the RGB image into three images of R, G, and B channels, and convert the image of each channel into a grayscale image; Perform dimensionality reduction processing on each grayscale image respectively, and obtain the reduced-dimensional RGB image based on the three grayscale images after dimensionality reduction processing; The RGB image after dimension reduction is input into the improved ShuffleNetV2 network model to obtain the fault diagnosis result of the rolling mill; the improved ShuffleNetV2 network model is obtained by adding a dual attention mechanism to the basic ShuffleNetV2 network for training; the improved ShuffleNetV2 network model includes an input module fused with the attention mechanism, multiple lightweight modules connected in series, and an output module fused with the attention mechanism; the input module includes an input layer, a convolutional layer, an activation function, an attention mechanism layer, and a fully connected layer connected in series; The training process of the improved ShuffleNetV2 network model includes: introducing a dual attention mechanism into a basic ShuffleNetV2 network to obtain an initial ShuffleNetV2 network; obtaining a training set and pre-training the initial ShuffleNetV2 network on the training set through transfer learning to obtain an initial ShuffleNetV2 network model; the weight parameters in the initial ShuffleNetV2 network model are weight parameters with general feature extraction capabilities; and iteratively training the initial ShuffleNetV2 network model through the training set to obtain an improved ShuffleNetV2 network model.

2. The method according to claim 1, characterized in that The dimensionality reduction processing is performed on each grayscale image respectively, including: For any grayscale image, get the mean of all data points in the grayscale image; Determine a scatter matrix based on the values ​​and mean of each data point, and solve the eigenvalues ​​and eigenvectors of the scatter matrix; The largest front K The eigenvector corresponding to the eigenvalue is determined as the new eigenvector; According to the new feature vector and the original grayscale image, the grayscale image after dimensionality reduction is determined.

3. The method according to claim 1, characterized in that The RGB image after dimension reduction is input into the improved ShuffleNetV2 network model to obtain the fault diagnosis results of the rolling mill, including: Inputting the reduced-dimensional RGB image into the improved ShuffleNetV2 network model, and extracting features from the reduced-dimensional RGB image through the input module; Performing convolution operations on the features in the RGB image after dimension reduction in sequence by sequentially connecting multiple lightweight modules in series to obtain convolution features; The output module performs feature prediction on the convolution feature to obtain a fault diagnosis result of the rolling mill.

4. The method according to claim 3, characterized in that The extracting features from the RGB image after dimension reduction through the input module includes: The input layer receives the reduced-dimensional RGB image and performs convolution on the reduced-dimensional RGB image through the convolution layer to extract local features. Increasing the nonlinear characteristics of the local features through an activation function to obtain nonlinear features; Assign weights to different features in the nonlinear features through the attention mechanism layer; The features output by the attention mechanism layer are fully connected through the fully connected layer to obtain the features in the RGB image after dimension reduction.

5. The method according to claim 4, characterized in that The attention mechanism layer includes a channel domain attention module and a spatial domain attention module; the channel domain attention module and the spatial domain attention module are cascaded.

6. The method according to claim 3, characterized in that The output module includes a convolution layer, an attention mechanism layer, and a fully connected layer connected in series. The output module performs feature prediction on the convolution feature to obtain the fault diagnosis result of the rolling mill, including: The convolutional features are convolved through the convolution layer, and the weights are assigned to the convolved features through the attention mechanism layer; The features output by the attention mechanism layer are mapped to various fault categories through the fully connected layer to obtain the fault diagnosis results of the rolling mill.

7. The method according to claim 1, characterized in that The fault diagnosis result of the rolling mill includes the probability of each fault category occurring in each component of the rolling mill; the method further includes: If the probability is greater than a preset fault threshold, the fault category corresponding to the probability is determined as the category of the rolling mill fault.

Images