Rotating machine fault diagnosis method and system

Abstract

The invention provides a rotating machine fault diagnosis method and system. The method comprises the following steps: collecting a vibration signal of a rotating machine to be detected; on the basis of the vibration signals, the fault type of the rotating machine to be detected is judged through a trained VLFFD model; wherein the VLFFD model is composed of a backbone network and a classification header network based on dot product similarity, the backbone network comprises a one-dimensional convolution feedforward network and a feature fusion module, and the feature fusion module is a feature fusion network based on a multivariable self-attention mechanism and a grouping query attention mechanism. According to the scheme, the adaptability and generalization ability of mechanical fault identification under different sampling signals can be enhanced, and the accuracy of rotary mechanical fault identification is guaranteed.

Description

Technical Field

[0001] This invention belongs to the field of deep learning technology, and in particular relates to a method and system for diagnosing faults in rotating machinery. Background Technology

[0002] Rotating machinery plays a crucial role in industrial production and engineering applications, with key rotating components widely used in systems such as sensor robots, vehicles, and wind turbines. The stable operation of rotating machinery is essential for ensuring uninterrupted industrial processes, and its failure can lead to significant economic losses. Fault diagnosis of rotating machinery typically involves assessing the condition of the equipment through sensor signals. Effectively collecting operational data in real-world industrial environments has accelerated the development of data-driven fault diagnosis methods. These data-driven methods possess efficient data mining and pattern recognition capabilities, enabling them to effectively utilize monitoring data for fault diagnosis and identify potential fault sources.

[0003] Currently, data-driven deep learning methods have become a research hotspot in the field of fault diagnosis. To address the noise interference problem of gearbox vibration signals under extreme conditions, Wang et al. proposed a VMTransformer model that integrates variational mode decomposition (VMD) and multi-channel Transformer for fault diagnosis of planetary gears in wind turbines. To address the challenges of fault identification under mixed operating conditions such as noise and small sample sizes, Sun et al. introduced an end-to-end multi-scale residual network with a parallel attention mechanism. For small sample and high-noise scenarios, Zhang et al. developed a novel lightweight convolutional neural network. This model uses an interactive convolutional module with dilated layers, maintaining strong noise resistance while using minimal parameters. For imbalanced open set fault detection, Zuo et al. proposed a novel contrastive conditional adversarial autoencoder. This method utilizes multiple classes of known health and fault data to construct an end-to-end unified model, overcoming the limitation of traditional methods that rely on the availability of fault samples.

[0004] Meanwhile, deep learning-based fault diagnosis methods can effectively address fault diagnosis tasks in different scenarios. To address the limitations of cross-domain fault diagnosis methods in effectively utilizing complementary information from multi-source heterogeneous data, Zhou et al. proposed a novel multi-source heterogeneous information fusion model based on the Swin-Transformer framework. This model achieves efficient multi-source heterogeneous data fusion and distribution alignment. To address the challenges faced by transfer learning methods under varying operating conditions in aligning deep features and handling decision boundaries specific to the task, Li et al. introduced a transfer diagnosis method based on maximum classifier difference and deep feature alignment. Experiments on different datasets demonstrated that this method improved accuracy by 2%–10%. To address the problem that current domain adaptation methods neglect the discriminativeness and intrinsic structure of features within the target domain, Li et al. proposed a two-layer discriminative domain adaptation network. This network adjusts domain-level features through adversarial training and designs class-level and sample-level discriminative modules to achieve domain adaptation. Experimental results on bearing and gearbox datasets demonstrate its effectiveness and superiority. To address the challenges of fault diagnosis under incomplete multi-source domains, Zhang et al. proposed a cross-domain bilateral transfer learning method. This method relaxes the constraint of consistent fault categories between the source and target domains and cleverly integrates knowledge from multiple incomplete source domains.

[0005] The above methods propose different fault diagnosis approaches for various complex industrial scenarios (including background noise, small sample sizes, and overlapping operating conditions), and have achieved excellent research results. However, Since these methods are mainly based on deep learning, they require preprocessing of monitoring data with different variables and signal lengths, or manual adjustment of model parameters. In actual fault diagnosis, these methods have poor adaptability and generalization ability. Summary of the Invention

[0006] In view of this, embodiments of the present invention provide a method and system for diagnosing rotating machinery faults, which solves the problem that current mechanical fault diagnosis has poor adaptability and generalization ability for different signals.

[0007] In a first aspect of the present invention, a method for diagnosing faults in rotating machinery is provided, comprising: Collect vibration signals from the rotating machinery under test; Based on the vibration signal, the fault type of the rotating machinery under test is determined by the trained VLFFD model; The VLFFD model consists of a backbone network and a classification head network based on dot product similarity. The backbone network includes a one-dimensional convolutional feedforward network and a feature fusion module. The feature fusion module is a feature fusion network based on multivariate self-attention mechanism and grouped query attention mechanism.

[0008] In a second aspect of the present invention, a rotating machinery fault diagnosis system is provided, comprising: The signal acquisition module is used to acquire the vibration signals of the rotating machinery under test. The fault identification module is used to determine the fault type of the rotating machinery under test based on the vibration signal and the trained VLFFD model. The VLFFD model consists of a backbone network and a classification head network based on dot product similarity. The backbone network includes a one-dimensional convolutional feedforward network and a feature fusion module. The feature fusion module is a feature fusion network based on multivariate self-attention mechanism and grouped query attention mechanism.

[0009] In a third aspect of the present invention, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor executes the computer program to implement the steps of the method as described in the first aspect of the present invention.

[0010] In a fourth aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the method provided in the first aspect of the present invention.

[0011] In this embodiment of the invention, a feature fusion network based on multivariate self-attention and group query attention is introduced into the VLFFD model to integrate variable-related features and local fault features. A one-dimensional convolutional feedforward network is used to enhance the extraction of local vibration features. A classification head network based on dot product similarity is used to achieve dynamic fault identification. This not only improves the accuracy of rotating machinery fault identification, but also enhances the adaptability and generalization ability in the fault diagnosis process, meets the fault identification needs under different variables and different signal length scenarios, and improves the fault identification capability under multi-scale signal features. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating a method for diagnosing faults in rotating machinery according to an embodiment of the present invention. Figure 2 This is another schematic flowchart of a rotating machinery fault diagnosis method provided in one embodiment of the present invention; Figure 3 A schematic diagram illustrating the principle of a feature fusion module provided in one embodiment of the present invention; Figure 4 A schematic diagram illustrating the principle of a one-dimensional convolutional feedforward network provided in one embodiment of the present invention; Figure 5 A schematic diagram illustrating the principle of a classification head network provided in one embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of a rotating machinery fault diagnosis system according to an embodiment of the present invention; Figure 7 This is a schematic diagram of the structure of an electronic device provided in one embodiment of the present invention. Detailed Implementation

[0014] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0015] It should be understood that the terms "comprising" and other similar expressions in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, system, or apparatus that includes a series of steps or units and is not limited to the listed steps or units. Furthermore, "first" and "second" are used to distinguish different objects and are not intended to describe a specific order.

[0016] Please see Figure 1 A flowchart illustrating a method for diagnosing faults in rotating machinery, provided by an embodiment of the present invention, includes: S101. Collect vibration signals from the rotating machinery under test; Vibration sensors can be used to collect vibration signals from the selected machinery under test, and these vibration signals can be used as the output of the VLFFD model to determine the fault category.

[0017] S102. Based on the vibration signal, determine the fault type of the rotating machinery under test using the trained VLFFD model; The VLFFD (Variable Association Features and Local Feature Fusion Fault Diagnosis) model is trained using fault diagnosis datasets with different data dimensions. The trained VLFFD model is then used to determine the fault type of the rotating machinery under test. For example, rotating machinery datasets can include the CWRU dataset, the HUST gear dataset, and the DIRG dataset.

[0018] The VLFFD model consists of a backbone network and a classification head network based on dot product similarity. The backbone network includes a one-dimensional convolutional feedforward network and a feature fusion module. The feature fusion module is a feature fusion network based on multivariate self-attention mechanism and grouped query attention mechanism.

[0019] The backbone network refers to the main part of the model, which is mainly responsible for extracting features from the original input for subsequent classification and detection tasks. The classification head network is the classification network at the end of the model, which is used to integrate the extracted features and map them to class labels.

[0020] In the feature fusion module, the multivariate attention mechanism can be used to extract the correlation features between signals, and the group query attention mechanism can be used to capture the local abnormal features within each signal, thereby achieving feature fusion.

[0021] One-dimensional convolutional feedforward networks (Conv1DFFN) can effectively preserve and filter local anomaly information in fused fault features. By using a fault classification head based on dot product similarity, the similarity between the category label and the category patch can be calculated to predict the fault category.

[0022] In this embodiment, the VLFFD model based on the backbone network and the classification head network can accurately identify the fault categories of rotating machinery and has good adaptability and generalization ability under different variables and signal lengths.

[0023] In one embodiment, step S102 includes: S201. Extract fault correlation features between vibration signals through the multivariate attention mechanism in the feature fusion module, extract local abnormal features within each vibration signal through the group query attention mechanism, and perform feature fusion on the fault correlation features and local abnormal features. S202. Based on a one-dimensional convolutional feedforward network, filter out local anomalies in the fused features and output the final fault features; S203. Using a fault classification head network, calculate the dot product similarity between fault features and actual category vectors to predict the fault category of rotating machinery.

[0024] The working principle of the feature fusion module based on the parallel self-attention mechanism is as follows: Figure 3 As shown, the feature fusion module can extract and fuse fault-related features and local anomaly features.

[0025] Specifically, the vibration signal is position-encoded through a patch embedding layer, and a category patch is generated and inserted at the end of the sample label.

[0026] Vibration signals are position-encoded using a patch embedding layer and can be mapped to patches. The position encoding employs a learnable position encoding layer. Simultaneously, an additional patch C is generated. c As a category patch, it is inserted at the end of the sample tag for use in the fault identification head network.

[0027] The input structure of the VLFFD model is shown in formulas (1) and (2), where, This represents the embedded vector. It is a patch of the original vibration signal. This is a category patch.

[0028] (1) (2) A multivariate self-attention mechanism is designed to extract correlation features and global fault information between different variables. The multivariate self-attention block models the interactions between variable sequences, formally described as shown in equations (3) and (4). Indicates the associated features, and It represents the average of Q and K generated for each variable, LN indicates layer normalization, and Dropout indicates dropping layers. This represents the sequence of variables for each input.

[0029] (3) (4) Grouped Query Attention (GQA) is introduced to extract local fault features from each vibration signal. The GQA mechanism models the correlation features between each patch in each variable and reflects local anomalies through a sharing mechanism, as shown in equations (5) and (6). Contains H / G query headers, and It consists of a shared key matrix and a value matrix. The final matrix is ​​obtained by concatenating the outputs of each group. Output As shown in formula (7), where LN represents layer normalization and Dropout represents dropping layers, This represents the sequence of variables for each input.

[0030] (5) (6) (7) (8) Finally, in order to effectively extract fault features from vibration signals with different quantities and data lengths, the correlation features between variables are fused with the local anomaly features of the monitoring signals. The fusion method is shown in formula (8), where, This indicates the fault characteristics after fusion. Indicates fault association characteristics, This represents local anomalous features within a single signal. Optionally, the entire attention mechanism consists of 8 attention heads with a feature dimension of 512, 4 query groups, and a dropout rate of 0.1.

[0031] By focusing on the fusion characteristics of vibration signals and the inter-channel correlation characteristics between different vibration signals, the parallel self-attention mechanism of the VLFFD model can effectively handle fault diagnosis tasks under monitoring signals with different variables and lengths.

[0032] In another embodiment, the feedforward layer network operates as follows: Figure 4 As shown, in order to enhance the focus on key features in vibration signals, a feedforward layer network based on one-dimensional convolution is designed to capture local features of vibration signals more effectively through local perception and weight sharing mechanisms.

[0033] Optionally, a feedforward network based on one-dimensional convolutional kernels is constructed, wherein the first convolutional layer of the feedforward network maps the input from the original dimension to a high-dimensional feature space, and the second convolutional layer of the feedforward network maps the high-dimensional features back to the original dimension. Construct a gate network layer to filter features of different importance levels; In the feedforward layer network, a residual network structure is used to connect the fused features with the fault features mapped through the feedforward layer network and the gate network layer, and output the final fault features.

[0034] A feedforward network based on one-dimensional convolutional kernels is constructed to replace the original feedforward network based on fully connected layers. The first convolutional layer is shown in Equation (9), where, H 1 indicates output. Conv1D This represents a one-dimensional convolutional layer. express GELU Activation function This represents the input to the feedforward layer.

[0035] (9) The first convolutional layer maps the input from its original dimension to a high-dimensional feature space, which helps the model capture complex features. The second convolutional layer is shown in Equation (10), where, H 2 indicates output. Conv1D LN represents a one-dimensional convolutional layer. H 1 represents the input to the feedforward layer. The second convolutional layer maps the high-dimensional features back to the original dimensions. This two-layer mapping mechanism based on one-dimensional convolution helps the model understand the input data from different perspectives.

[0036] (10) The design uses a gate network layer to filter patches containing important features. As shown in Equation (11), the gate network layer obtains the importance level between each patch through a linear layer and uses a sigmoid activation function to implement weight allocation, where... The dimension is the feature dimension.

[0037] (11) Design the feedforward layer network as shown in equations (12) and (13). Utilize the residual network structure to... The fault features are connected to the fault features after feature mapping through Conv1DFFN and gate network to achieve the final fault feature output. Conv1DFFN represents a feedforward layer network based on one-dimensional convolution kernel.

[0038] (12) (13) Optionally, the one-dimensional convolution kernel has a kernel size of 3, 4096 intermediate channels, and a dropout rate of 0.1.

[0039] Compared to the global mapping of fully connected layers, the Conv1DFFN feedforward layer network based on one-dimensional convolutional kernels can effectively preserve and filter local anomalies in the fused fault features, making it suitable for fault diagnosis tasks in monitoring signal scenarios with different variables and lengths.

[0040] In one embodiment, the fault classification head based on dot product similarity works as follows: Figure 5 As shown, different mechanical equipment has different monitoring signals in different dimensions, and the types of mechanical equipment are different, resulting in different numbers of categories. In order to achieve unified fault diagnosis for different machines, a classification head network based on dot product similarity is used for fault identification.

[0041] For the VLFFD backbone network of the encoder-only paradigm, which consists of a feature fusion module and a feedforward network based on one-dimensional convolution, a classification head network based on dot product similarity is used to achieve fault classification. Optionally, class labels for different classification tasks are constructed, the predicted class is obtained by calculating the dot product similarity between the fault features and the class vectors of the class labels, and the actual fault class is obtained by classifying using the SoftMax function.

[0042] Design category tags separately for each task The model achieves dynamic adaptation to multi-class classification tasks. Class labels are special labels in the Transformer model that embed actual class information from different datasets into a high-dimensional space. The generation process is shown in formula (14). It follows a normal distribution with a standard deviation of 0.2, where T represents the t-th dataset and i represents the class.

[0043] (14) The predicted category is obtained by calculating the dot product similarity with the actual category vector.

[0044] In one embodiment, the classification head network further includes a cross-attention mechanism, based on which the category labels of the backbone network output features are updated. By employing the cross-attention mechanism, the category patches are updated using the feature representations of the backbone network output sequence, thereby achieving dynamic classification.

[0045] Introducing cross-attention mechanism to update category patches C c As shown in formula (15), where Q 1 is generated from the input sequence y. K 2 and V 2 by C c generate.

[0046] (15) The update process of the category label is as shown in formulas (16), (17) and (18). Linear Layer Indicates a linear layer. This indicates cross attention. X output This represents the output sequence from the VLFFD backbone network, while This indicates a category patch. (Updated) It is a feature map that encapsulates the characteristics of the fused vibration signal. A residual structure is introduced to prevent degradation. This indicates the final output category patch.

[0047] (16) (17) (18) The actual category is obtained by calculating the dot product similarity and then applying the result to the SoftMax function. The calculation formulas for the output are shown in formulas (19) and (20). Represents the probability distribution of categories. Indicates category label, This indicates the final output category patch.

[0048] (19) (20) In this way, the VLFFD model can handle multiple classification tasks simultaneously without designing a separate classification head for each task. It can effectively and flexibly identify fault categories across different machines, effectively adapting to mechanical equipment monitoring systems with varying variables and data lengths.

[0049] 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.

[0050] Figure 6 This is a schematic diagram of a rotating machinery fault diagnosis system provided in an embodiment of the present invention. The system includes: The signal acquisition module 610 is used to acquire the vibration signal of the rotating machinery under test; The fault identification module 620 is used to determine the fault type of the rotating machinery under test based on the vibration signal and the trained VLFFD model. The VLFFD model consists of a backbone network and a classification head network based on dot product similarity. The backbone network includes a one-dimensional convolutional feedforward network and a feature fusion module. The feature fusion module is a feature fusion network based on multivariate self-attention mechanism and grouped query attention mechanism.

[0051] In one embodiment, the fault identification module 620 includes: The feature extraction unit is used to extract fault correlation features between vibration signals through the multivariate attention mechanism in the feature fusion module, extract local abnormal features within each vibration signal through the group query attention mechanism, and perform feature fusion on the fault correlation features and local abnormal features. The feature filtering unit is used to filter local anomalies in the fused features based on a one-dimensional convolutional feedforward network and output the final fault features. The fault prediction unit is used to predict the fault category of rotating machinery by calculating the dot product similarity between fault features and actual category vectors through the fault classification head network.

[0052] Optionally, the fault identification module 620 further includes The position encoding unit is used to encode the position of the vibration signal through the patch embedding layer and generate a category patch that is inserted at the end of the sample marker.

[0053] The step of filtering local anomalies in the fused features based on a one-dimensional convolutional feedforward network and outputting the final fault features includes: Construct a feedforward network based on one-dimensional convolutional kernels. The first convolutional layer of the feedforward network maps the input from the original dimension to a high-dimensional feature space, and the second convolutional layer of the feedforward network maps the high-dimensional features back to the original dimension. Construct a gate network layer to filter features of different importance levels; In the feedforward layer network, a residual network structure is used to connect the fused features with the fault features mapped through the feedforward layer network and the gate network layer, and output the final fault features.

[0054] Optionally, the step of calculating the dot product similarity between fault features and actual category vectors through a fault classification head network to predict the fault category of rotating machinery includes: We construct category labels for different classification tasks, obtain the predicted category by calculating the dot product similarity between the fault features and the category vectors of the category labels, and obtain the actual fault category by classifying using the SoftMax function.

[0055] In one embodiment, the classification head network further includes a cross-attention mechanism, based on which the category labels of the backbone network output features are updated.

[0056] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system and modules described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0057] Figure 7 This is a schematic diagram of an electronic device according to an embodiment of the present invention. The electronic device is used for mechanical fault type identification. Figure 7 As shown, the electronic device 70 of this embodiment includes a memory 710, a processor 720, and a system bus 730. The memory 710 includes an executable program 7101 stored thereon. As those skilled in the art will understand, Figure 7 The electronic device structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0058] The following is combined with Figure 7 A detailed introduction to each component of the electronic device: The memory 710 can be used to store software programs and modules. The processor 720 executes various functional applications and data processing of the electronic device by running the software programs and modules stored in the memory 710. The memory 710 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the electronic device (such as cached data), etc. In addition, the memory 710 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0059] The memory 710 contains an executable program 7101 for a fault identification method. This executable program 7101 can be divided into one or more modules / units, which are stored in the memory 710 and executed by the processor 720 to perform tasks such as identifying rotating machinery fault types. Each module / unit can be a series of computer program instruction segments capable of performing a specific function, describing the execution process of the executable program 7101 in the electronic device 70. For example, the executable program 7101 can be divided into functional modules such as a signal acquisition module and a fault identification module.

[0060] The processor 720 is the control center of the electronic device. It connects various parts of the electronic device via various interfaces and lines. By running or executing software programs and / or modules stored in the memory 710, and by calling data stored in the memory 710, it performs various functions and processes data, thereby monitoring the overall status of the electronic device. Optionally, the processor 720 may include one or more processing units; preferably, the processor 720 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, application programs, etc., and the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into the processor 720.

[0061] The system bus 730 is used to connect various functional components within the computer, transmitting data, address, and control information. Its type can be, for example, a PCI bus, an ISA bus, or a CAN bus. Instructions from the processor 720 are transmitted to the memory 710 via the bus, and the memory 710 sends data back to the processor 720. The system bus 730 handles the data and instruction exchange between the processor 720 and the memory 710. Of course, the system bus 730 can also connect to other devices, such as network interfaces and display devices.

[0062] In this embodiment of the invention, the executable program executed by the processor 720 included in the electronic device includes: Collect vibration signals from the rotating machinery under test; Based on the vibration signal, the fault type of the rotating machinery under test is determined by the trained VLFFD model.

[0063] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0064] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0065] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for diagnosing faults in rotating machinery, characterized in that, include: Collect vibration signals from the rotating machinery under test; Based on the vibration signal, the fault type of the rotating machinery under test is determined by the trained VLFFD model; The VLFFD model consists of a backbone network and a classification head network based on dot product similarity. The backbone network includes a one-dimensional convolutional feedforward network and a feature fusion module. The feature fusion module is a feature fusion network based on multivariate self-attention mechanism and grouped query attention mechanism.

2. The method according to claim 1, characterized in that, The method of determining the fault type of the rotating machinery under test using the trained VLFFD model includes: The fault correlation features between vibration signals are extracted by the multivariate attention mechanism in the feature fusion module, and the local anomaly features within each vibration signal are extracted by the group query attention mechanism. The fault correlation features and local anomaly features are then fused. The final fault features are output by filtering local anomalies in the fused features based on a one-dimensional convolutional feedforward network. By using a fault classification head network, the dot product similarity between fault features and actual category vectors is calculated to predict the fault category of rotating machinery.

3. The method according to claim 2, characterized in that, The feature fusion module extracts fault correlation features between vibration signals through a multivariate attention mechanism and extracts local anomaly features within each vibration signal through a grouped query attention mechanism. Before fusing the fault correlation features and local anomaly features, the process also includes... The vibration signal is positionally encoded by a patch embedding layer, and a category patch is generated and inserted at the end of the sample label.

4. The method according to claim 2, characterized in that, The process of filtering local anomalies in the fused features based on a one-dimensional convolutional feedforward network, before outputting the final fault features, includes: Construct a feedforward network based on one-dimensional convolutional kernels. The first convolutional layer of the feedforward network maps the input from the original dimension to a high-dimensional feature space, and the second convolutional layer of the feedforward network maps the high-dimensional features back to the original dimension. Construct a gate network layer to filter features of different importance levels; In the feedforward layer network, a residual network structure is used to connect the fused features with the fault features mapped through the feedforward layer network and the gate network layer, and output the final fault features.

5. The method according to claim 2, characterized in that, The method of predicting the fault category of rotating machinery by calculating the dot product similarity between fault features and actual category vectors through a fault classification head network includes: We construct category labels for different classification tasks, obtain the predicted category by calculating the dot product similarity between the fault features and the category vectors of the category labels, and obtain the actual fault category by classifying using the SoftMax function.

6. The method according to claim 1, characterized in that, The classification head network also includes a cross-attention mechanism, based on which the category labels of the backbone network output features are updated.

7. A fault diagnosis system for rotating machinery, characterized in that, include: The signal acquisition module is used to acquire the vibration signals of the rotating machinery under test. The fault identification module is used to determine the fault type of the rotating machinery under test based on the vibration signal and the trained VLFFD model. The VLFFD model consists of a backbone network and a classification head network based on dot product similarity. The backbone network includes a one-dimensional convolutional feedforward network and a feature fusion module. The feature fusion module is a feature fusion network based on multivariate self-attention mechanism and grouped query attention mechanism.

8. The system according to claim 7, characterized in that, The method of determining the fault type of the rotating machinery under test using the trained VLFFD model includes: The feature extraction unit is used to extract fault correlation features between vibration signals through the multivariate attention mechanism in the feature fusion module, extract local abnormal features within each vibration signal through the group query attention mechanism, and perform feature fusion on the fault correlation features and local abnormal features. The feature filtering unit is used to filter local anomalies in the fused features based on a one-dimensional convolutional feedforward network and output the final fault features. The fault prediction unit is used to predict the fault category of rotating machinery by calculating the dot product similarity between fault features and actual category vectors through the fault classification head network.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of a rotating machinery fault diagnosis method as described in any one of claims 1 to 6.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed, it implements the steps of a rotating machinery fault diagnosis method as described in any one of claims 1 to 6.

Images