Rolling bearing fault diagnosis method based on optimized GhostNet

Abstract

The invention requests to protect a rolling bearing fault diagnosis method based on optimized GhostNet. According to the method, a continuous wavelet transform technology is innovatively adopted, and a one-dimensional vibration signal is converted into a two-dimensional time-frequency image, so that time-domain and frequency-domain information of the signal is fully fused. In order to construct a lightweight and efficient diagnosis model, a core Ghost module of a GhostNet network is improved, original point convolution operation is replaced by conventional convolution, and a'ghost 'feature map is generated through linear transformation, so that the feature extraction capability is effectively improved while the model parameter quantity and the calculation complexity are remarkably reduced. According to the method provided by the invention, verification is carried out on a public bearing data set, and extremely high fault identification accuracy can still be kept even under the condition that the parameter quantity is greatly reduced. Compared with a traditional method, the method does not need complex artificial feature extraction, the model is lighter and easier to deploy in edge computing equipment, and an efficient and lightweight solution is provided for intelligent fault diagnosis of the rolling bearing.

Description

Technical Field

[0001] This invention relates to the field of rolling bearing technology, specifically a rolling bearing fault diagnosis method based on optimized GhostNet. Background Technology

[0002] As a core component of rotating machinery systems, the operating condition of rolling bearings directly affects the safety and stability of the entire equipment. Failures can not only lead to equipment downtime but also cause serious safety accidents and substantial economic losses. Therefore, developing efficient and accurate rolling bearing fault diagnosis technology is crucial.

[0003] Traditional fault diagnosis methods rely heavily on signal processing techniques and expert experience, requiring complex manual feature extraction, such as time-domain statistical features and frequency-domain spectral peaks. These methods often exhibit poor robustness and insufficient generalization ability when facing complex operating conditions and strong noise interference. With the development of artificial intelligence technology, deep learning-based fault diagnosis methods have become a research hotspot due to their powerful automatic feature learning capabilities. Among them, Convolutional Neural Networks (CNNs), with their outstanding performance in image recognition, are widely used in fault diagnosis. However, standard CNN models typically contain a large number of parameters and complex calculations, resulting in large models, slow training and inference speeds, and difficulty in meeting the needs of real-time diagnosis in industrial settings and edge deployment.

[0004] To address the issue of lightweight models, researchers have proposed several lightweight network architectures, such as MobileNet, ShuffleNet, and GhostNet. GhostNet introduces a "Ghost module" that utilizes inexpensive linear transformations to generate a large number of "ghost" feature maps from a small number of intrinsic feature maps, thereby significantly reducing the model's computational complexity and parameter count while maintaining performance. However, the original Ghost module uses 1×1 point convolutions to generate intrinsic feature maps, resulting in a limited receptive field and insufficient ability to extract spatial information from input features, thus limiting the model's performance ceiling.

[0005] Although existing technologies have explored the combination of signal imaging and lightweight networks (see CN114724112A), their improvements have mostly focused on the network architecture level or the introduction of attention mechanisms, failing to address the limitations of the lightweight convolutional module itself. In particular, the inherent flaw of the original Ghost module, which relies on 1×1 point convolutions to generate features, severely restricts further improvements in model performance.

[0006] This invention addresses the insufficient feature extraction capability of lightweight models in existing technologies by proposing a rolling bearing fault diagnosis method based on an optimized GhostNet. By replacing the 1×1 point convolutions in the Ghost module with regular k×k (k>1) convolutions, this invention effectively expands the receptive field and enhances the model's ability to extract time-frequency features of vibration signals. Simultaneously, this invention retains the lightweight principle of generating "ghost" feature maps using the Ghost module, ensuring the model's high efficiency. Experimental results show that the proposed optimization method significantly reduces model parameters and computational load while substantially improving the accuracy of fault diagnosis, overcoming the shortcomings of existing technologies. Summary of the Invention

[0007] This invention aims to address the technical problems of insufficient feature extraction capability and low accuracy of existing lightweight neural networks in rolling bearing fault diagnosis. A rolling bearing fault diagnosis method based on an optimized GhostNet is proposed. The technical solution of this invention is as follows:

[0008] A rolling bearing fault diagnosis method based on optimized GhostNet includes the following steps:

[0009] Step 1: Collect the vibration signal of the rolling bearing and process the collected signal into segments to obtain a series of signal segments of fixed length;

[0010] Step 2: Perform continuous wavelet transform on each signal segment obtained in Step 1 to generate the corresponding two-dimensional time-frequency image, so as to fuse the time-domain and frequency-domain features of the signal;

[0011] Step 3: Construct an improved GhostNet neural network model; The improved Ghost neural network first generates intrinsic feature maps through conventional convolution operations, then performs a linear transformation on the intrinsic feature maps to generate "ghost" feature maps, and finally concatenates the intrinsic feature maps and "ghost" feature maps for output;

[0012] Step 4: Use the time-frequency images generated in Step 2 as input data, and train the model constructed in Step 3 using the training set with fault labels;

[0013] Step 5: Use the trained model to diagnose faults in the new time-frequency images and output the corresponding diagnostic results.

[0014] Furthermore, the signal segment length in step 1 is N sampling points, where N is an integer greater than 0.

[0015] Furthermore, the mathematical model used in the continuous wavelet transform in step 2 is described as follows:

[0016]

[0017] Where a is the scale parameter, b is the translation parameter, t represents the time variable, ψ is the complex wavelet basis function, and x(t) is the input vibration signal.

[0018] Furthermore, the improved Ghost neural network in step 3 includes the following sub-steps:

[0019] Step 3.1: Given an input feature map X, generate m intrinsic feature maps Y' using a conventional convolution operation with a kernel size of k×k;

[0020] Step 3.2: Apply s-1 inexpensive linear transformations Φ to each intrinsic feature map in Y' to generate m(s-1) "ghost" feature maps;

[0021] Step 3.3: Concatenate Y' with the generated "ghost" feature map along the channel dimension to obtain the final output feature map Y.

[0022] Furthermore, in step 3.1, k is an integer greater than 1.

[0023] Furthermore, in step 3.2, the inexpensive linear transformation Φ is a depthwise separable convolution or a 1×1 convolution.

[0024] Furthermore, the mathematical expression for the splicing operation in step 3.3 is:

[0025]

[0026] in, This represents m intrinsic feature maps generated through regular convolution. Indicates passage A set of m(s-1) "ghost" feature maps generated by applying s-1 inexpensive linear transformations. This represents the concatenation operation along the channel dimension, ultimately outputting a feature map. .

[0027] Furthermore, in step 4, the model training process uses the cross-entropy loss function to calculate the loss and uses the Adam optimizer to update the parameters.

[0028] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the rolling bearing fault diagnosis method based on optimized GhostNet as described in any one of the claims.

[0029] A non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the rolling bearing fault diagnosis method based on optimized GhostNet as described in any one of the claims.

[0030] The advantages and beneficial effects of this invention are as follows:

[0031] The rolling bearing fault diagnosis method proposed in this invention innovatively improves the Ghost module of the GhostNet network. By replacing the original 1×1 point convolution with a k×k conventional convolution, the receptive field of the network is significantly expanded, enhancing the model's feature extraction capability for the input signal (time-frequency image), and solving the problem of insufficient feature extraction caused by the small receptive field of the original Ghost module. In the signal processing stage, this invention uses continuous wavelet transform technology to convert the one-dimensional vibration signal into a two-dimensional time-frequency image, which can more comprehensively preserve and present the time-domain and frequency-domain information of the signal, providing richer input features for subsequent deep learning models. In order to achieve lightweight model while ensuring high accuracy, this invention retains the core idea of ​​the Ghost module to generate "ghost" feature maps, generating a large number of redundant features through inexpensive linear transformation operations, which greatly reduces the number of model parameters and computational complexity, making the model easier to deploy and run on resource-constrained edge devices.

[0032] Fault diagnosis experiments show that the optimized GhostNet model constructed in this invention exhibits superior performance on publicly available bearing datasets. Compared to the original GhostNet and other lightweight networks (such as MobileNetV2 and ShuffleNetV2), the model in this invention achieves higher fault identification accuracy while significantly reducing the number of parameters and model size, demonstrating an excellent balance between accuracy and efficiency. This invention provides an effective technical solution for achieving high-precision, lightweight intelligent fault diagnosis.

[0033] The core innovation of this invention is mainly reflected in the improved Ghost module structure defined in claim 4 and its dependent claims 5-7.

[0034] The original GhostNet module uses 1×1 point convolutions to generate eigenmaps, which reduces computation, but its receptive field is extremely small, making it difficult to effectively capture spatial context information in images, especially limiting its ability to extract widely distributed fault features (such as the time-frequency trajectories of periodic impacts) in time-frequency maps. This invention creatively replaces the 1×1 point convolution with a k×k (k>1) conventional convolution (such as a 3×3 convolution), significantly expanding the receptive field and enhancing the ability to model local spatial patterns.

[0035] Although conventional convolutions are widely used in CNNs, replacing 1×1 convolutions with larger kernels in the Ghost module is not a conventional technique in this field. The reasons are: 1) GhostNet was designed for extreme lightweightness, and introducing k×k convolutions might be considered to undermine its low parameter advantage; 2) Traditional lightweight network designs tend to reduce rather than increase kernel size. Therefore, those skilled in the art, in the pursuit of lightweight design, typically do not actively increase kernel size to improve performance, and may even consider this a violation of lightweight principles. This invention demonstrates through experiments that, with proper design (such as controlling the number of intrinsic feature maps m and the generation factor s), even with the introduction of k×k convolutions, the overall model parameter count and computational cost are still significantly lower than the original GhostNet and other mainstream lightweight networks, achieving a win-win situation of accuracy and efficiency. Attached Figure Description

[0036] Figure 1 This is a flowchart of a preferred embodiment of the rolling bearing fault diagnosis method based on optimized GhostNet provided by the present invention;

[0037] Figure 2 A schematic diagram of the improved Ghost module structure;

[0038] Figure 3 This is a graph showing the accuracy versus loss during the training process of the model in this invention.

[0039] Figure 4 This is a confusion matrix diagram of the model of the present invention;

[0040] Figure 5 A comparison chart of the accuracy rates of different models.

[0041] Figure 6 Comparison chart of model parameters and model size. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0043] The technical solution of the present invention to solve the above-mentioned technical problems is:

[0044] like Figure 1 This is a flowchart of a rolling bearing fault diagnosis method based on optimized GhostNet in a specific embodiment. The method includes:

[0045] Step 1: Collect vibration signals from the rolling bearing and segment the collected signals to obtain a series of signal segments of fixed length. This step prepares data for subsequent time-frequency transformation and model training.

[0046] Step 2: Perform continuous wavelet transform on each signal segment obtained in Step 1 to generate the corresponding two-dimensional time-frequency image. This process can clearly display fault features such as periodic impacts in one-dimensional time-domain signals on the time-frequency plane, which is beneficial for deep learning models to capture key information.

[0047] Step 3: Construct the improved GhostNet neural network model. The core of this embodiment lies in the improvement of the Ghost module. For example... Figure 2 As shown, the improved Ghost module first uses a 3×3 regular convolution (replacing the original 1×1 point convolution) to generate the intrinsic feature map, then uses a 1×1 convolution (as a cheap linear transformation) to generate the "ghost" feature map, and finally concatenates the two for output. This design significantly improves feature extraction capabilities while maintaining the advantage of lightweight design;

[0048] Step 4: Using the time-frequency images generated in Step 2 as input data, train the model constructed in Step 3 using the training set with fault labels. In this embodiment, the cross-entropy loss function is used to measure the difference between the predicted values ​​and the true labels, and the Adam optimizer is used to iteratively update the network parameters;

[0049] Step 5: Use the trained model to diagnose faults in the new time-frequency images and output the corresponding diagnostic results.

[0050] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0051] In one embodiment of the present invention, the rolling bearing fault diagnosis method based on optimized GhostNet is verified. The following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0052] In this embodiment, to verify the performance of the proposed optimized GhostNet model in rolling bearing fault diagnosis, the Case Western Reserve University (CWRU) Open Bearing Database was used as the validation dataset. The experiment selected acceleration vibration signals of the drive-end bearing under 0HP load, with a signal sampling frequency of 12kHz. The experiment included 10 different bearing states: normal, inner ring fault, outer ring fault, and rolling element fault. Each fault type included three fault sizes: 0.007 inches, 0.014 inches, and 0.021 inches. Specific parameters are shown in Table 1.

[0053] Table 1. Parameters of the Case Western Reserve University Bearing Dataset

[0054]

[0055] In this embodiment, the original vibration signal is first segmented, with 1024 sampling points in each segment. Then, a continuous wavelet transform is performed on each signal segment using the complex Morlet wavelet basis function to generate a two-dimensional time-frequency image of size 224×224. According to Table 1, 300 time-frequency images are generated for each state. The dataset is divided into a training set and a test set, with 240 images for each class used as the training set and 60 images for the test set, for a total of 2400 training images and 600 test images.

[0056] In this embodiment, the optimized GhostNet model is built around the improved Ghost module. The model training environment is: PyTorch deep learning framework, NVIDIA GPU. The training parameters are set as follows: batch size of 32, initial learning rate of 0.0001, optimizer of Adam, and training epochs of 50.

[0057] This embodiment verifies the effectiveness of the optimized GhostNet model by comparing it with MobileNetV2, ShuffleNetV2, and the original GhostNet model. Evaluation metrics include model accuracy, model size, and number of model parameters.

[0058] Experimental results are as follows Figures 3 to 6 As shown. Figure 3 The training accuracy and loss curves of the model of this invention are shown. It can be seen that as the number of training rounds increases, the accuracy of both the training set and the test set rises rapidly and tends to stabilize, eventually approaching 100%; while the loss value continues to decrease and approaches 0, indicating that the model has good convergence. Figure 4 The confusion matrix on the test set shows that the model misclassifies only one 0.021-inch rolling element fault sample as a 0.007-inch rolling element fault, while all other samples are correctly classified, demonstrating that the model has extremely high diagnostic accuracy.

[0059] Figure 5 The comparison of accuracy among the four models is shown. Figure 5 As can be seen, the optimized GhostNet model proposed in this invention is comparable to the original GhostNet in terms of accuracy, both reaching a level close to 100%, which is significantly better than MobileNetV2 and ShuffleNetV2.

[0060] Figure 6The paper presents a comparison of the number of parameters and model size for four different models. It can be seen that the model of this invention exhibits a significant advantage in terms of lightweight design: its model size is only 61.20MB, and its number of parameters is 1.04M, far smaller than the original GhostNet (149.65MB, 3.91M), and even smaller than ShuffleNetV2 (68.17MB, 1.26M) and MobileNetV2 (156.16MB, 1.36M). This demonstrates that this invention, while maintaining high accuracy, greatly reduces the model's resource consumption, laying a solid foundation for the practical application of the model on industrial edge devices.

[0061] In summary, the rolling bearing fault diagnosis method proposed in this invention effectively solves the problem of insufficient feature extraction capability of lightweight models by improving the Ghost module and introducing time-frequency images, achieving an excellent balance between accuracy and efficiency.

[0062] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions.

[0063] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0064] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0065] The above embodiments should be understood as illustrative only and not as limiting the scope of protection of the present invention. After reading the description of the present invention, those skilled in the art can make various alterations or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A method for diagnosing rolling bearing faults based on optimized GhostNet, characterized in that, Includes the following steps: Step 1: Collect the vibration signal of the rolling bearing and process the collected signal into segments to obtain a series of signal segments of fixed length; Step 2: Perform continuous wavelet transform on each signal segment obtained in Step 1 to generate the corresponding two-dimensional time-frequency image, so as to fuse the time-domain and frequency-domain features of the signal; Step 3: Construct an improved GhostNet neural network model; The improved Ghost neural network first generates intrinsic feature maps through conventional convolution operations, then performs a linear transformation on the intrinsic feature maps to generate "ghost" feature maps, and finally concatenates the intrinsic feature maps and "ghost" feature maps for output; Step 4: Use the time-frequency images generated in Step 2 as input data, and train the model built in Step 3 using the training set with fault labels; Step 5: Use the trained model to diagnose faults in the new time-frequency images and output the corresponding diagnostic results.

2. The rolling bearing fault diagnosis method based on optimized GhostNet according to claim 1, characterized in that, The signal segment length in step 1 is N sampling points, where N is an integer greater than 0.

3. The rolling bearing fault diagnosis method based on optimized GhostNet according to claim 1, characterized in that, The mathematical model used in step 2 for the continuous wavelet transform is described as follows: Where a is the scale parameter, b is the translation parameter, t represents the time variable, ψ is the complex wavelet basis function, and x(t) is the input vibration signal.

4. The rolling bearing fault diagnosis method based on optimized GhostNet according to claim 1, characterized in that, The improved Ghost neural network in step 3 includes the following sub-steps: Step 3.1: Given an input feature map X, generate m intrinsic feature maps Y' using a conventional convolution operation with a kernel size of k×k; Step 3.2: Apply s-1 inexpensive linear transformations Φ to each intrinsic feature map in Y' to generate m(s-1) "ghost" feature maps; Step 3.3: Concatenate Y' with the generated "ghost" feature map along the channel dimension to obtain the final output feature map Y.

5. The rolling bearing fault diagnosis method based on optimized GhostNet according to claim 4, characterized in that, In step 3.1, k is an integer greater than 1.

6. The rolling bearing fault diagnosis method based on optimized GhostNet according to claim 4, characterized in that, In step 3.2, the inexpensive linear transformation Φ is a depthwise separable convolution or a 1×1 convolution.

7. The rolling bearing fault diagnosis method based on optimized GhostNet according to claim 4, characterized in that, The mathematical expression for the splicing operation in step 3.3 is: in, This represents m intrinsic feature maps generated through regular convolution. Indicates passage A set of m(s-1) "ghost" feature maps generated by applying s-1 inexpensive linear transformations. This represents the concatenation operation along the channel dimension, ultimately outputting a feature map. .

8. The rolling bearing fault diagnosis method based on optimized GhostNet according to claim 1, characterized in that, The model training process in step 4 uses the cross-entropy loss function to calculate the loss and the Adam optimizer to update the parameters.

9. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the rolling bearing fault diagnosis method based on optimized GhostNet as described in any one of claims 1 to 8.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the rolling bearing fault diagnosis method based on optimized GhostNet as described in any one of claims 1 to 8.

Images