Shaft center trajectory image enhancement method and device, electronic equipment and storage medium

Abstract

The application relates to an axle center trajectory image enhancement method and device, electronic equipment and a storage medium, and relates to the technical field of fault diagnosis. The application comprises the following steps: performing multi-step degradation on an axle center trajectory image to convert the axle center trajectory image into axle center trajectory images containing different noise levels; constructing a denoising reconstruction network model based on a U-Net structure, fusing a multi-head attention and a time step embedding mechanism; obtaining an optimal hyperparameter combination of the denoising reconstruction network model, so that the denoising reconstruction network model learns to predict Gaussian noise contained in a noisy axle center trajectory image under different diffusion time steps; and using the denoising reconstruction network model to gradually predict and remove Gaussian noise in the noisy axle center trajectory image at each time step to obtain an enhanced axle center trajectory image. Through the scheme, a large number of, various, realistic fault axle center trajectory images can be unconditionally generated, fault features can be deeply mined and learned, and high-quality and high-fidelity fault samples can be ensured.

Description

Technical Field

[0001] This application relates to the field of fault diagnosis technology, specifically to a method, apparatus, electronic device, and storage medium for enhancing shaft trajectory images. Background Technology

[0002] Large rotating machinery (such as centrifugal compressors, axial compressors, steam turbines, and gas turbines) plays a crucial role in many key sectors, including energy, petrochemicals, and aviation. Their safety and reliability are a focus of attention for both academia and industry. With the rapid development of big data analytics and artificial intelligence technologies, intelligent monitoring and diagnostic methods based on massive amounts of historical operational data are gradually emerging, demonstrating enormous potential in improving the safety and reliability of large rotating machinery and reducing maintenance costs.

[0003] In fault diagnosis of rotating machinery, the shaft center trajectory is a crucial and intuitive indicator of the unit's operating status. The shaft center trajectory is composed of displacement signals collected by two mutually perpendicular eddy current sensors installed at the same cross-section, representing the rotor's eddy current pattern. When the unit deviates from its normal operating state and malfunctions, the shape of the shaft center trajectory undergoes a specific change; therefore, its shape characteristics are a key signal for fault diagnosis.

[0004] However, most existing intelligent diagnostic models rely on massive amounts of data for training. In actual engineering, rotating machinery operates in a healthy state for a long time with an extremely low probability of failure, resulting in very limited real fault data that can be obtained. This serious "data imbalance" makes it difficult for intelligent diagnostic models to fully learn fault characteristics, which severely restricts the effectiveness of intelligent diagnostic models in practical applications. Summary of the Invention

[0005] In view of this, this application provides a method, apparatus, electronic device and storage medium for enhancing shaft trajectory images. The main purpose is to solve the technical problem that most existing intelligent diagnostic models rely on massive amounts of data for training. However, in actual engineering, rotating machinery operates in a healthy state for a long time with an extremely low probability of failure, resulting in very limited real fault data that can be obtained. This serious "data imbalance" makes it difficult for intelligent diagnostic models to fully learn fault characteristics, which seriously restricts the effectiveness of intelligent diagnostic models in practical applications.

[0006] According to a first aspect of this application, a method for enhancing an axisymmetric trajectory image is provided, the method comprising: The original vibration signal of the rotating machinery is acquired, and the original vibration signal is preprocessed to obtain a preprocessed shaft center trajectory image. The preprocessed axisymmetric trajectory image is subjected to multi-step degradation processing through a perturbation process in the form of a multi-step Markov chain, so as to transform the preprocessed axisymmetric trajectory image into an axisymmetric trajectory image containing different noise levels. A denoising and reconstruction network model based on the U-Net structure is constructed, which integrates a multi-head attention mechanism and a temporal step embedding mechanism. The multi-head attention mechanism is used to capture the global spatial correlation features of the axisymmetric trajectory image, and the temporal step embedding mechanism is used to perceive the current diffusion time step and adaptively adjust the denoising intensity. Using the FID score between the generated image and the real image as the objective function, the Bayesian optimization algorithm is used to automatically search and tune the hyperparameters of the denoising and reconstruction network model to obtain the optimal hyperparameter combination. Based on the optimal hyperparameter combination, the denoising and reconstruction network model is trained using the noisy axisymmetric trajectory images containing different levels of noise, so that the denoising and reconstruction network model can learn to predict the Gaussian noise contained in the noisy axisymmetric trajectory images at different diffusion time steps. Starting with a pure Gaussian noise image, and using the noisy axis center trajectory images at different time steps as input, the trained denoising reconstruction network model is used to progressively predict and remove Gaussian noise from the noisy axis center trajectory images at each time step, resulting in an enhanced axis center trajectory image.

[0007] According to a second aspect of this application, an axial trajectory image enhancement apparatus is provided, the apparatus comprising: The acquisition module is used to acquire the original vibration signal of the rotating machinery and preprocess the original vibration signal to obtain the preprocessed shaft center trajectory image; The processing module is used to perform multi-step degradation processing on the preprocessed axisymmetric trajectory image through a perturbation process in the form of a multi-step Markov chain, so as to transform the preprocessed axisymmetric trajectory image into an axisymmetric trajectory image containing different noise levels. The building module is used to construct a denoising and reconstruction network model based on the U-Net structure and integrating a multi-head attention mechanism and a temporal step embedding mechanism. The multi-head attention mechanism is used to capture the global spatial correlation features of the axisymmetric trajectory image, and the temporal step embedding mechanism is used to perceive the current diffusion time step and adaptively adjust the denoising intensity. The tuning module is used to automatically search and tune the hyperparameters of the denoising and reconstruction network model using the FID score between the generated image and the real image as the objective function, and to obtain the optimal combination of hyperparameters. The training module is used to train the denoising and reconstruction network model based on the optimal hyperparameter combination and using the noisy axisymmetric trajectory images containing different levels of noise, so that the denoising and reconstruction network model learns to predict the Gaussian noise contained in the noisy axisymmetric trajectory images at different diffusion time steps. The enhancement module is used to start from a pure Gaussian noise image, take the noisy axis center trajectory images at different time steps as input, and use the trained denoising reconstruction network model to gradually predict and remove Gaussian noise in the noisy axis center trajectory images at each time step to obtain the enhanced axis center trajectory image.

[0008] According to a third aspect of this application, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are used to cause a computer to perform the method of the first aspect described above.

[0009] According to a fourth aspect of this application, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method of the first aspect described above.

[0010] By employing the above technical solutions, this application provides a method, apparatus, electronic device, and storage medium for enhancing shaft center trajectory images. Compared with existing technologies, this application can obtain a preprocessed shaft center trajectory image by acquiring the original vibration signal of rotating machinery and preprocessing the original vibration signal; through a multi-step Markov chain-like perturbation process, the preprocessed shaft center trajectory image is subjected to multi-step degradation processing to transform it into shaft center trajectory images containing different levels of noise; a denoising and reconstruction network model based on a U-Net structure is constructed, integrating a multi-head attention mechanism and a temporal step embedding mechanism; wherein, the multi-head attention mechanism is used to capture the global spatial correlation features of the shaft center trajectory image, and the temporal step embedding mechanism is used for perception. The current diffusion time step is used to adaptively adjust the denoising intensity. Using the FID score between the generated image and the real image as the objective function, a Bayesian optimization algorithm is employed to automatically search and optimize the hyperparameters of the denoising reconstruction network model to obtain the optimal hyperparameter combination. Based on the optimal hyperparameter combination, the denoising reconstruction network model is trained using noisy axisymmetric trajectory images with different noise levels, enabling the model to learn to predict Gaussian noise contained in the noisy axisymmetric trajectory images at different diffusion time steps. Starting from a pure Gaussian noise image, and using noisy axisymmetric trajectory images at different time steps as input, the trained denoising reconstruction network model is used to progressively predict and remove Gaussian noise from the noisy axisymmetric trajectory images at each time step, resulting in an enhanced axisymmetric trajectory image.

[0011] The scheme described in this application involves adding Gaussian noise to the preprocessed shaft center trajectory image through a multi-step Markov chain, generating noisy images with varying levels of noise. During the inference phase, starting from pure Gaussian noise, a pre-trained denoising network is used to progressively remove noise, unconditionally generating a large number of diverse and realistic fault shaft center trajectory images. These generated images can be directly added to the training set of the intelligent diagnostic model, greatly alleviating the "data imbalance" problem caused by the scarcity of real fault samples.

[0012] This application's multi-head attention mechanism captures the global spatial correlation features of trajectory images, ensuring that the generated trajectory conforms to real fault patterns in its macroscopic shape. The application's time-step embedding mechanism senses the diffusion time step and adaptively adjusts the denoising intensity, enabling the model to accurately predict Gaussian noise at different noise scales, ensuring the accuracy of the reverse reconstruction process from pure noise to a clear fault trajectory. This allows the model to deeply mine and learn the essential fault modes even with only a small number of real samples.

[0013] This application uses the FID score between the generated image and the real image as the objective function and employs a Bayesian optimization algorithm to automatically search for the optimal hyperparameters. This ensures that even with limited real data, the denoising and reconstruction network can still achieve the optimal generation state, guaranteeing high quality and high fidelity of the generated faulty samples.

[0014] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0015] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating an axis trajectory image enhancement method provided in an embodiment of this application is shown. Figure 2 A complete flowchart of an attention diffusion optimization model provided in an embodiment of this application is shown; Figure 3This application provides an embodiment of an axisymmetric trajectory diagram generated from X and Y direction data. Figure 4 This paper illustrates a U-Net structure diagram provided in an embodiment of this application; Figure 5 A structural diagram of a multi-head attention mechanism provided in an embodiment of this application is shown; Figure 6 A structural diagram of a time step embedding mechanism provided in an embodiment of this application is shown; Figure 7 This paper illustrates a flowchart of Bayesian parameter optimization for an Attention-Diffusion model provided in an embodiment of this application. Figure 8 This paper illustrates a graph showing the change in FID score during a Bayesian optimization process, as provided in an embodiment of this application. Figure 9 This paper presents a comparison chart of the generation effects of various generation models provided in an embodiment of this application; Figure 10 A schematic diagram of the structure of an axis trajectory image enhancement device provided in an embodiment of this application is shown. Detailed Implementation

[0018] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of this application, including various details to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0019] The following description, with reference to the accompanying drawings, describes a method, apparatus, electronic device, and storage medium for enhancing axisymmetric trajectory images according to embodiments of this application.

[0020] To address the problem that most existing intelligent diagnostic models rely on massive amounts of data for training, while in actual engineering, rotating machinery operates in a healthy state for extended periods with an extremely low probability of failure, resulting in a very limited availability of real-world fault data, this severe "data imbalance" makes it difficult for intelligent diagnostic models to fully learn fault characteristics, significantly limiting their effectiveness in practical applications. This application provides a method for enhancing shaft center trajectory images, such as... Figure 1 As shown, the method includes: Step 101: Obtain the original vibration signal of the rotating machinery and preprocess the original vibration signal to obtain the preprocessed shaft center trajectory image.

[0021] like Figure 2 As shown, the axisymmetric trajectory image enhancement method involves four core modules: a forward diffusion module, a denoising and reconstruction network module, a backward diffusion module, and a Bayesian parameter optimization module. These four core modules are tightly coupled in terms of function and data flow, forming a complete process from image noise addition and neural network modeling to image restoration and parameter tuning optimization. This structure not only ensures the model's robustness and accuracy under small sample conditions but also possesses good generalization ability and training stability.

[0022] The original vibration signals originated from multiple industrial rotating machinery units in operation within a certain group, encompassing typical equipment such as generators, centrifugal compressors, and steam turbines. It is understood that the rotating machinery may also include at least one of axial compressors or gas turbines.

[0023] To improve the image quality and fault detection capability of subsequent model training, the original vibration signal needs to undergo the following preprocessing steps before entering the diffusion model training: First, the original vibration signal is denoised to obtain a synthesized axis trajectory map of the denoised signal, thus removing discrete points and jitter interference from the original vibration signal; for example... Figure 3 The comparison of the synthesized axis trajectory graphics before and after noise reduction is shown. The original trajectory has obvious jitter and discrete points, while the denoised trajectory is smoother, has a closed outline, and presents a clear geometric structure. Next, the synthesized axisymmetric trajectory of the denoised signal is image scaled and grayscaled to make the image size and number of channels meet the model input requirements; Finally, the processed signal-synthesized shaft center trajectory images are labeled, organized, and stored to construct shaft center trajectory images containing different fault types, thereby improving the image quality and fault discrimination ability of subsequent model training. These fault types may include dynamic-static rubbing, rotor imbalance, oil film whirl, misalignment, and other fault types.

[0024] Step 102: Through a multi-step Markov chain-like perturbation process, the preprocessed axisymmetric trajectory image is subjected to multi-step degradation processing to transform the preprocessed axisymmetric trajectory image into an axisymmetric trajectory image containing different levels of noise.

[0025] This step corresponds to the function of the forward diffusion module. Specifically, the original, clear axisymmetric trajectory map is degraded in multiple steps through a perturbation process in the form of a multi-step Markov chain. Each step adds small noise to the image, and after a sufficient number of steps, the axisymmetric trajectory map is gradually transformed into a pure noise image with an approximately Gaussian distribution. This process generates axisymmetric trajectory maps containing different levels of noise, providing training samples for the reverse process to learn and gradually recover the original image.

[0026] Step 103: Construct a denoising and reconstruction network model based on the U-Net structure, which integrates multi-head attention mechanism and temporal step embedding mechanism; wherein, the multi-head attention mechanism is used to capture the global spatial correlation features of the axisymmetric trajectory image, and the temporal step embedding mechanism is used to perceive the current diffusion time step and adaptively adjust the denoising intensity.

[0027] This step corresponds to the denoising and reconstruction network module. The constructed denoising and reconstruction network model is based on the U-Net structure and integrates multi-head attention and temporal embedding mechanisms, enabling the model to simultaneously possess multi-scale feature modeling capabilities, global spatial structure awareness capabilities, and dynamic temporal awareness capabilities. Figure 4-6 The U-Net structure, multi-head attention mechanism, and temporal step embedding mechanism are demonstrated respectively.

[0028] The U-Net structure may include an encoder, a decoder, and skip connections connecting the corresponding layers of the encoder and decoder. The encoder can be used to downsample the input noisy axisymmetric trajectory image to extract multi-scale features, and the decoder can be used to upsample the features to restore the image resolution.

[0029] Specifically, the multi-head attention mechanism is used to capture the global spatial correlation features of the axisymmetric trajectory image, and its processing is as follows: like Figure 5 As shown, the input feature matrix First, it goes through three learnable weight matrices. , , After performing a linear mapping, the query matrices are obtained respectively. Key matrix Sum matrix Subsequently, the dot product of the query matrix and the transpose of the key matrix can be calculated. This measures the correlation between features at different locations and divides it by a scaling factor. Scale normalization is performed to avoid the dot product result being too large, which could lead to gradient instability. The attention weights are then obtained through the Softmax function, and finally, the attention weights are compared with the value matrix. Multiply to obtain the output features that incorporate global context information. In this way, the multi-head attention mechanism enables the model to focus on the spatial relationships between different regions when reconstructing the axisymmetric trajectory image, thereby enhancing its ability to perceive the overall trajectory shape.

[0030] Specifically, the time-step embedding mechanism can be used to sense the current diffusion time step and adaptively adjust the denoising intensity. The processing procedure is as follows: like Figure 6 As shown, obtain the current diffusion time step. The sine and cosine functions are used to map the data to a high-dimensional embedding space, generating a time embedding vector. The sine and cosine functions are used to encode time location information in different dimensions, and their encoding formulas are as follows:

[0031]

[0032] In the formula, For position encoding, For diffusion time step, For the dimension index of the embedded vector, For embedded dimensions, Even-numbered dimensions For odd-numbered dimensions, It is a sine function. It is a cosine function.

[0033] The sine and cosine terms are used to encode temporal location information in different dimensions, enabling similar embedding representations for close time steps and clear distinctions between time steps that are far apart. The generated temporal embedding vectors are injected into the residual blocks of the denoising and reconstruction network model. In this way, the model can perceive the diffusion stage of the current image and adaptively adjust the denoising process according to the noise intensity at different time steps.

[0034] This module not only improves the reconstruction quality of the axis trajectory image, but also enhances the robustness and controllability of the model for images at different degradation stages.

[0035] Step 104: Using the FID score between the generated image and the real image as the objective function, the Bayesian optimization algorithm is used to automatically search and tune the hyperparameters of the denoising and reconstruction network model to obtain the optimal hyperparameter combination.

[0036] This step corresponds to the Bayesian parameter optimization module. To further improve the image quality and stability of the model in the axisymmetric trajectory generation task, this embodiment automatically searches and tunes several key hyperparameters (such as learning rate, batch size, optimizer momentum factor, etc.) based on the image generation evaluation metric (Fréchet Inception Distance, FID). Traditional manual parameter tuning not only relies on experience but is also inefficient, making it difficult to obtain optimal results with limited experimental resources. Therefore, this application introduces a Bayesian optimization strategy to construct a global hyperparameter optimization mechanism to achieve adaptive improvement in generation quality.

[0037] Figure 7The flowchart of the Bayesian optimization of the Attention-Diffusion model is shown, which automatically finds the optimal parameter combination during the training process. This optimization process not only significantly shortens the parameter search time, but also effectively improves the clarity and distribution diversity of the generated images. The process specifically includes the following sub-steps: A Gaussian process surrogate model is constructed to simulate the mapping relationship between hyperparameter combinations and generated image quality (i.e., FID score). The image generation evaluation metric FID comprehensively measures the distance between the generated image and the real image in the feature space distribution, effectively reflecting the image's realism and diversity. The Bayesian optimization objective function is defined as the FID score of the image generated by the model trained under this hyperparameter combination, i.e.:

[0038] In the formula, The optimal combination of hyperparameters is... For the hyperparameter combination to be optimized, This is the hyperparameter space.

[0039] Based on the Gaussian process surrogate model, within a preset parameter space, the next hyperparameter combination to be evaluated is selected through a data acquisition function (such as ExpectedImprovement). The denoising and reconstruction network model can be trained using the hyperparameter combination to be evaluated, and the FID score between the generated image and the real image can be calculated as the observation value. In each round of Bayesian optimization, the parameter combination is sampled once, the generated image is trained and the FID is calculated as feedback for the "performance evaluation" of the current point.

[0040] The Gaussian process surrogate model is updated using observed values ​​to make its posterior distribution estimation of the objective function more accurate.

[0041] The above steps are executed iteratively until the preset number of iterations or the convergence condition is reached. The hyperparameter combination corresponding to the lowest FID score is output as the optimal hyperparameter combination.

[0042] like Figure 8 As shown, in the early stages of training, the model parameters are randomly initialized, resulting in significant fluctuations in the initial FID value. The model gradually converges during the middle of the search process, reaching a minimum value of 122.88 in the 105th round. At this point, the generated image exhibits optimal performance in terms of structural integrity, boundary sharpness, and shape stability. Based on the optimization results of the 105th round, the final training parameter configuration of the Attention-Diffusion model used for image generation is shown in Table 1.

[0043] Table 1. Bayesian optimization parameter configuration and optimization results.

[0044] Step 105: Based on the optimal hyperparameter combination, train the denoising and reconstruction network model using noisy axis center trajectory images with different noise levels, so that the denoising and reconstruction network model can learn to predict the Gaussian noise contained in the noisy axis center trajectory images at different diffusion time steps.

[0045] This step corresponds to the model training phase. Based on the optimal parameter combination determined by Bayesian optimization, a complete training process for the Attention-Diffusion model can be constructed. Specific parameter configurations can be as follows: based on the Bayesian optimization results, the optimal learning rate is 0.00015, the batch size is 9, the gradient accumulation steps are 3, and the Adam optimizer momentum factor is set to (0.9, 0.99). After multiple verifications, the entire model training takes 6000 steps, which can control the overall training cost while maintaining convergence of generation quality.

[0046] Step 106: Starting from the pure Gaussian noise image, and using the noisy axis center trajectory images at different time steps as input, the trained denoising and reconstruction network model is used to predict and remove Gaussian noise from the noisy axis center trajectory images at each time step to obtain the enhanced axis center trajectory image.

[0047] This step corresponds to the inverse diffusion module. After model training is complete, the key goal of the generation phase is to construct the inverse process, that is, starting from pure Gaussian noise, gradually "denoising" to finally recover the axisymmetric trajectory image that conforms to the true distribution. This process is called inverse diffusion, and its essence is a backsampling process of a conditionally driven Markov chain parameterized by a neural network.

[0048] Specifically, during the inverse diffusion process, the model takes noisy images at different time steps as input and uses the U-Net network to predict the Gaussian noise contained in the image at that time. The process can be represented as follows:

[0049] In the formula, , , , These can be represented as Gaussian noise predicted by the model at the corresponding time step. , , , These can represent the state of noisy images or images after gradual denoising at different diffusion time steps. It can represent the current diffusion time step. This can represent a denoising network used for noise prediction.

[0050] By continuously predicting and removing noise components at each time step, the model can gradually recover from a Gaussian noise image to a clear axisymmetric trajectory image, thus completing the reverse generation process.

[0051] To evaluate the effectiveness of the diffusion model in enhancing shaft center trajectory maps for rotating machinery fault diagnosis, Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) were selected as comparative models. Under the premise of approximately similar network structure complexity, the performance of the three methods in the shaft center trajectory map generation task was analyzed. Figure 9 It can be seen that the trajectory data generated by the diffusion model is closer to the real data in terms of contour, detail and structure, while the generation effect of GAN and VAE has certain defects.

[0052] To further quantify and evaluate the performance of different generative models, the FID (Fixed Indicator Value) of the three models under different fault types was calculated. The FID value measures the distribution distance between the generated data and the real data; a lower FID indicates higher quality generated data. As shown in Table 2, the FID value of the diffusion model is lower than that of GAN and VAE in all fault types, indicating that its generated data has the best quality. Overall, the diffusion model has a significant advantage in maintaining the structural clarity, reasonable distribution, and detail fidelity of trajectory images.

[0053] Table 2 Comparison of FID scores for each generative model

[0054] In summary, this application proposes a shaft center trajectory image enhancement method that integrates attention mechanisms, Bayesian optimization strategies, and diffusion models. This method achieves high-quality generation of shaft center trajectory images and solves the problems of model training difficulties and insufficient diagnostic accuracy caused by small sample sizes and data imbalance in rotating machinery fault diagnosis. This structure not only ensures the robustness and accuracy of the model under small sample conditions but also possesses good generalization ability and training stability.

[0055] Based on the above Figure 1 The specific implementation of the method shown in this embodiment provides an axisymmetric trajectory image enhancement device, such as... Figure 10 As shown, the device includes: an acquisition module 31, a processing module 32, a construction module 33, an optimization module 34, a training module 35, and an enhancement module 36; The acquisition module 31 is used to acquire the original vibration signal of the rotating machinery and preprocess the original vibration signal to obtain a preprocessed shaft center trajectory image; wherein the rotating machinery includes at least one of a centrifugal compressor, an axial compressor, a steam turbine, or a gas turbine; The processing module 32 is used to perform multi-step degradation processing on the preprocessed axisymmetric trajectory image through a multi-step Markov chain perturbation process, so as to transform the preprocessed axisymmetric trajectory image into an axisymmetric trajectory image containing different noise levels. Module 33 is used to construct a denoising and reconstruction network model based on the U-Net structure and integrating a multi-head attention mechanism and a temporal step embedding mechanism; wherein, the multi-head attention mechanism is used to capture the global spatial correlation features of the axisymmetric trajectory image, and the temporal step embedding mechanism is used to perceive the current diffusion time step and adaptively adjust the denoising intensity. The tuning module 34 is used to automatically search and tune the hyperparameters of the denoising and reconstruction network model using the FID score between the generated image and the real image as the objective function, and to obtain the optimal combination of hyperparameters. Training module 35 is used to train the denoising reconstruction network model based on the optimal hyperparameter combination using the noisy axisymmetric trajectory images containing different noise levels, so that the denoising reconstruction network model learns to predict the Gaussian noise contained in the noisy axisymmetric trajectory images at different diffusion time steps. The enhancement module 36 is used to start from a pure Gaussian noise image, take the noisy axis center trajectory images at different time steps as input, and use the trained denoising reconstruction network model to gradually predict and remove Gaussian noise in the noisy axis center trajectory images at each time step to obtain the enhanced axis center trajectory image.

[0056] In specific application scenarios, the acquisition module 31 can be used to perform noise reduction processing on the original vibration signal to obtain a noise-reduced signal synthesis axis trajectory map, so as to remove discrete points and jitter interference in the original vibration signal; The synthesized axisymmetric trajectory of the denoised signal is subjected to image scaling and grayscale processing to make the image size and number of channels meet the model input requirements; The processed signal-synthesized shaft center trajectory map is labeled, organized, and stored to construct shaft center trajectory images containing different fault types.

[0057] In specific application scenarios, the construction module 33 can be used in the U-Net structure, including the encoder, decoder, and jump connections connecting the corresponding layers of the encoder and decoder. The encoder is used to downsample the input noisy axisymmetric trajectory image to extract multi-scale features, and the decoder is used to upsample the features to restore the image resolution.

[0058] In specific application scenarios, module 33 can be used in the processing of the multi-head attention mechanism, specifically including: The input feature matrix is ​​linearly mapped through a learnable weight matrix to obtain the query matrix, key matrix, and value matrix. Calculate the dot product of the query matrix and the transpose of the key matrix, normalize the dot product based on the scaling factor, and determine the attention weight of the normalized dot product using the Softmax function. The attention weights are multiplied by the value matrix to obtain output features that fuse global context information, thereby enhancing the denoising and reconstruction network model's ability to perceive the overall shape of the axis trajectory image.

[0059] In specific application scenarios, the construction module 33 can be used in the processing of the time step embedding mechanism, specifically including: The current diffusion time step is obtained, and the diffusion time step is mapped to a high-dimensional embedding space using sine and cosine functions to generate a time embedding vector. The sine and cosine functions are used to encode time position information in different dimensions, so that diffusion time steps that are close to each other have similar embedding representations, while diffusion time steps that are far apart are distinguished. The generated temporal embedding vector is injected into the residual block of the denoising and reconstruction network model, so that the denoising and reconstruction network model adaptively adjusts the denoising process according to the noise intensity of the current time step.

[0060] In specific application scenarios, the tuning module 34 can be used to construct a Gaussian process proxy model, which is used to simulate the mapping relationship between hyperparameter combinations and generated image quality. Based on the Gaussian process proxy model, select the hyperparameter combination to be evaluated within the preset parameter space; The denoising and reconstruction network model is trained using the hyperparameter combination to be evaluated, and the FID score between the generated image and the real image is calculated as an observation. The Gaussian process surrogate model is updated using the observed values, and the above steps are iteratively executed until a preset number of iterations or a convergence condition is reached. The hyperparameter combination corresponding to the lowest FID score is then output as the optimal hyperparameter combination.

[0061] It should be noted that other corresponding descriptions of the functional units involved in the axis trajectory image enhancement device provided in this embodiment can be found in [reference needed]. Figure 1 The corresponding descriptions in [the document] will not be repeated here.

[0062] Based on the above, Figure 1 Accordingly, this embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method. Figure 1 The method shown.

[0063] Based on this understanding, the technical solution of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause a computer device (such as personal computer, server, or network device, etc.) to execute the methods of various implementation scenarios of this application.

[0064] Based on the above, Figure 1 The method shown, and Figure 10 To achieve the above objectives, the present application also provides an electronic device, comprising a storage medium and a processor; the storage medium for storing a computer program; and the processor for executing the computer program to implement the above-described virtual device embodiments. Figure 1 The method shown.

[0065] Optionally, the aforementioned physical devices may also include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, input units such as a keyboard, etc., and optional user interfaces may also include USB interfaces, card reader interfaces, etc. The network interface may optionally include standard wired interfaces, wireless interfaces (such as Wi-Fi interfaces), etc.

[0066] Those skilled in the art will understand that the physical device structure provided in this embodiment does not constitute a limitation on the physical device, and may include more or fewer components, or combine certain components, or have different component arrangements.

[0067] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages the hardware and software resources of the aforementioned physical device, supporting the operation of information processing programs and other software and / or programs. The network communication module is used to enable communication between the various components within the storage medium, as well as communication with other hardware and software in the information processing physical device.

[0068] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platform, or it can be implemented by hardware. Compared with the prior art, the technical solution of this application can obtain the original vibration signal of rotating machinery and preprocess the original vibration signal to obtain a preprocessed shaft center trajectory image; through a multi-step Markov chain perturbation process, the preprocessed shaft center trajectory image is subjected to multi-step degradation processing to transform the preprocessed shaft center trajectory image into a shaft center trajectory image containing different noise levels; a denoising reconstruction network model based on U-Net structure is constructed, which integrates multi-head attention mechanism and time step embedding mechanism; wherein, the multi-head attention mechanism is used to capture the global spatial correlation features of the shaft center trajectory image, and the time step embedding mechanism is used to perceive the current diffusion time step and adaptively adjust the denoising strength. The algorithm uses the FID score between the generated image and the real image as the objective function, and employs a Bayesian optimization algorithm to automatically search and tune the hyperparameters of the denoising reconstruction network model to obtain the optimal hyperparameter combination. Based on the optimal hyperparameter combination, the denoising reconstruction network model is trained using noisy axis center trajectory images with different levels of noise, so that the denoising reconstruction network model can learn to predict the Gaussian noise contained in the noisy axis center trajectory images at different diffusion time steps. Starting from a pure Gaussian noise image, the noisy axis center trajectory images at different time steps are used as input, and the trained denoising reconstruction network model is used to gradually predict and remove the Gaussian noise in the noisy axis center trajectory images at each time step to obtain the enhanced axis center trajectory image.

[0069] The scheme described in this application involves adding Gaussian noise to the preprocessed shaft center trajectory image through a multi-step Markov chain, generating noisy images with varying levels of noise. During the inference phase, starting from pure Gaussian noise, a pre-trained denoising network is used to progressively remove noise, unconditionally generating a large number of diverse and realistic fault shaft center trajectory images. These generated images can be directly added to the training set of the intelligent diagnostic model, greatly alleviating the "data imbalance" problem caused by the scarcity of real fault samples.

[0070] This application's multi-head attention mechanism captures the global spatial correlation features of trajectory images, ensuring that the generated trajectory conforms to real fault patterns in its macroscopic shape. The application's time-step embedding mechanism senses the diffusion time step and adaptively adjusts the denoising intensity, enabling the model to accurately predict Gaussian noise at different noise scales, ensuring the accuracy of the reverse reconstruction process from pure noise to a clear fault trajectory. This allows the model to deeply mine and learn the essential fault modes even with only a small number of real samples.

[0071] This application uses the FID score between the generated image and the real image as the objective function and employs a Bayesian optimization algorithm to automatically search for the optimal hyperparameters. This ensures that even with limited real data, the denoising and reconstruction network can still achieve the optimal generation state, guaranteeing high quality and high fidelity of the generated faulty samples.

[0072] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the term "comprising" or any other variations thereof is 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 a process, method, article, or apparatus. Without further limitations, 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 the element.

[0073] The above are merely specific embodiments of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to these embodiments, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for enhancing an axisymmetric trajectory image, characterized in that, The method includes: The original vibration signal of the rotating machinery is acquired, and the original vibration signal is preprocessed to obtain a preprocessed shaft center trajectory image. The preprocessed axisymmetric trajectory image is subjected to multi-step degradation processing through a perturbation process in the form of a multi-step Markov chain, so as to transform the preprocessed axisymmetric trajectory image into an axisymmetric trajectory image containing different noise levels. A denoising and reconstruction network model based on the U-Net structure is constructed, which integrates a multi-head attention mechanism and a temporal step embedding mechanism. The multi-head attention mechanism is used to capture the global spatial correlation features of the axisymmetric trajectory image, and the temporal step embedding mechanism is used to perceive the current diffusion time step and adaptively adjust the denoising intensity. Using the FID score between the generated image and the real image as the objective function, the Bayesian optimization algorithm is used to automatically search and tune the hyperparameters of the denoising and reconstruction network model to obtain the optimal hyperparameter combination. Based on the optimal hyperparameter combination, the denoising and reconstruction network model is trained using the noisy axisymmetric trajectory images containing different levels of noise, so that the denoising and reconstruction network model can learn to predict the Gaussian noise contained in the noisy axisymmetric trajectory images at different diffusion time steps. Starting with a pure Gaussian noise image, and using the noisy axis center trajectory images at different time steps as input, the trained denoising reconstruction network model is used to progressively predict and remove Gaussian noise from the noisy axis center trajectory images at each time step, resulting in an enhanced axis center trajectory image.

2. The axisymmetric trajectory image enhancement method according to claim 1, characterized in that, The preprocessing of the original vibration signal to obtain the preprocessed axis trajectory image specifically includes: The original vibration signal is denoised to obtain a synthesized axis trajectory map of the denoised signal, so as to remove discrete points and jitter interference in the original vibration signal. The synthesized axisymmetric trajectory of the denoised signal is subjected to image scaling and grayscale processing to make the image size and number of channels meet the model input requirements; The processed signal-synthesized shaft center trajectory map is labeled, organized, and stored to construct shaft center trajectory images containing different fault types.

3. The axisymmetric trajectory image enhancement method according to claim 1, characterized in that, The U-Net structure includes an encoder, a decoder, and jump connections connecting the corresponding layers of the encoder and decoder. The encoder is used to downsample the input noisy axisymmetric trajectory image to extract multi-scale features, and the decoder is used to upsample the features to restore the image resolution.

4. The axis trajectory image enhancement method according to claim 1, characterized in that, The processing steps of the multi-head attention mechanism specifically include: The input feature matrix is ​​linearly mapped through a learnable weight matrix to obtain the query matrix, key matrix, and value matrix. Calculate the dot product of the query matrix and the transpose of the key matrix, normalize the dot product based on the scaling factor, and determine the attention weight of the normalized dot product using the Softmax function. The attention weights are multiplied by the value matrix to obtain output features that fuse global context information, thereby enhancing the denoising and reconstruction network model's ability to perceive the overall shape of the axis trajectory image.

5. The axisymmetric trajectory image enhancement method according to claim 1, characterized in that, The processing steps of the time step embedding mechanism specifically include: The current diffusion time step is obtained, and the diffusion time step is mapped to a high-dimensional embedding space using sine and cosine functions to generate a time embedding vector. The sine and cosine functions are used to encode time position information in different dimensions, so that diffusion time steps that are close to each other have similar embedding representations, while diffusion time steps that are far apart are distinguished. The generated temporal embedding vector is injected into the residual block of the denoising and reconstruction network model, so that the denoising and reconstruction network model adaptively adjusts the denoising process according to the noise intensity of the current time step.

6. The axisymmetric trajectory image enhancement method according to claim 1, characterized in that, The process involves using the FID score between the generated image and the real image as the objective function, and employing a Bayesian optimization algorithm to automatically search and fine-tune the hyperparameters of the denoising and reconstruction network model to obtain the optimal hyperparameter combination. Specifically, this includes: A Gaussian process proxy model is constructed to simulate the mapping relationship between hyperparameter combinations and generated image quality. Based on the Gaussian process proxy model, select the hyperparameter combination to be evaluated within the preset parameter space; The denoising and reconstruction network model is trained using the hyperparameter combination to be evaluated, and the FID score between the generated image and the real image is calculated as an observation. The Gaussian process surrogate model is updated using the observed values, and the above steps are iteratively executed until a preset number of iterations or a convergence condition is reached. The hyperparameter combination corresponding to the lowest FID score is then output as the optimal hyperparameter combination.

7. The axisymmetric trajectory image enhancement method according to claim 1, characterized in that, The rotating machinery includes at least one of a centrifugal compressor, an axial compressor, a steam turbine, or a gas turbine.

8. A device for enhancing images of axial trajectory, characterized in that, include: The acquisition module is used to acquire the original vibration signal of the rotating machinery and preprocess the original vibration signal to obtain the preprocessed shaft center trajectory image; The processing module is used to perform multi-step degradation processing on the preprocessed axisymmetric trajectory image through a perturbation process in the form of a multi-step Markov chain, so as to transform the preprocessed axisymmetric trajectory image into an axisymmetric trajectory image containing different noise levels. The building module is used to construct a denoising and reconstruction network model based on the U-Net structure and integrating a multi-head attention mechanism and a temporal step embedding mechanism. The multi-head attention mechanism is used to capture the global spatial correlation features of the axisymmetric trajectory image, and the temporal step embedding mechanism is used to perceive the current diffusion time step and adaptively adjust the denoising intensity. The tuning module is used to automatically search and tune the hyperparameters of the denoising and reconstruction network model using the FID score between the generated image and the real image as the objective function, and to obtain the optimal combination of hyperparameters. The training module is used to train the denoising and reconstruction network model based on the optimal hyperparameter combination and using the noisy axisymmetric trajectory images containing different levels of noise, so that the denoising and reconstruction network model learns to predict the Gaussian noise contained in the noisy axisymmetric trajectory images at different diffusion time steps. The enhancement module is used to start from a pure Gaussian noise image, take the noisy axis center trajectory images at different time steps as input, and use the trained denoising reconstruction network model to gradually predict and remove Gaussian noise in the noisy axis center trajectory images at each time step to obtain the enhanced axis center trajectory image.

9. An electronic device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the axisymmetric trajectory image enhancement method according to any one of claims 1 to 7.

10. A 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 axis trajectory image enhancement method according to any one of claims 1 to 7.

Images