Method and device for enhancing operation fault data of hydroelectric generating set

Abstract

The invention discloses a hydroelectric generating set operation fault data enhancement method and device, and the method comprises the steps: firstly collecting a set vibration signal, selecting a time-frequency transformation method to convert a one-dimensional vibration signal into a two-dimensional time-frequency image, enhancing the feature dimension of the signal, constructing a diffusion feature migration model, gradually disturbing the data distribution to Gaussian noise through forward diffusion, and carrying out the recognition of the Gaussian noise. The method comprises the following steps of: performing inverse denoising to generate simulation data highly similar to a real fault sample, realizing relevance learning and migration sharing of fault features among different working conditions in combination with an adversarial feature migration architecture, and finally evaluating an enhancement effect by calculating similarity among samples, and inputting enhanced data into a fault diagnosis model to verify precision improvement. Through the combination of time-frequency transformation and a diffusion model, sample scarcity and working condition barriers are broken through, a remarkable effect is shown in the aspects of expanding the fault sample scale and enriching the sample dimension, the similarity of generated data and a real sample is improved, the diagnosis precision is improved, and the model generalization ability is remarkably enhanced.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a method and device for enhancing operational fault data of a hydropower unit. Background Art

[0002] With advances in hydropower unit design and manufacturing technology and improvements in intelligent operation and maintenance management, modern hydropower units are now able to achieve long-term, efficient, and stable operation. However, this high-reliability operation and maintenance model also brings new technical challenges, such as the scarcity of fault samples. While advanced monitoring systems and rapid response mechanisms can intervene promptly in the early stages of faults, effectively avoiding major accidents, this also results in the number of fault samples that can actually be collected being far less than normal operating data. For intelligent diagnostic models that rely on big data training, the sample imbalance problem seriously restricts the model's generalization ability and diagnostic accuracy. Traditional data augmentation methods (such as oversampling and adversarial generative networks) often have difficulty maintaining the physical meaning of fault characteristics when applied to hydropower unit fault data, resulting in distorted generated samples and affecting model training results. Therefore, a data augmentation method that is more consistent with the fault characteristics of hydropower units is urgently needed to expand the effective training samples and improve the robustness of the diagnostic model.

[0003] There may also be insufficient adaptability across operating conditions. With the large-scale grid connection of renewable energy sources such as wind power and photovoltaics, hydropower units have taken on more frequent peak-shaving and frequency-regulating tasks, and their operating conditions have become more complex and variable. Frequent starts and stops and variable load operation exacerbate the wear of mechanical components and make the vibration characteristics of the units more non-stationary. Fault signals are easily drowned out by background noise, increasing the difficulty of feature extraction. In addition, the fault modes under different operating conditions vary significantly. Models trained on a single operating condition are difficult to adapt to multi-operating condition scenarios, resulting in reduced diagnostic accuracy. Existing transfer learning methods still have problems in cross-operating condition fault diagnosis, such as insufficient feature decoupling and insufficient domain adaptability, which limits the engineering applicability of the model.

[0004] In summary, existing technologies in the field of hydropower unit fault diagnosis still face two core challenges: sample scarcity and operating condition adaptability. Improving the generalization capability of diagnostic models through effective data augmentation and cross-operating condition knowledge transfer, without relying on massive amounts of fault data, has become a key technical issue that needs to be addressed. Summary of the Invention

[0005] The purpose of this application is to provide a method and device for enhancing the operating fault data of a hydropower unit in order to overcome the existing technical defects. By constructing an intelligent model to deeply mine and expand the limited fault data, a more comprehensive and diversified fault data set is constructed to enhance the operating fault data of the hydropower unit, thereby improving the adaptability and accuracy of the fault diagnosis model under different working conditions.

[0006] The purpose of this application is achieved through the following technical solutions:

[0007] In a first aspect, the present application proposes a method and device for enhancing operational fault data of a hydropower unit, the method comprising:

[0008] Step S1, collecting the vibration signal of the unit;

[0009] Step S2: Select a time-frequency transformation method to enhance the data features of the unit vibration signal and convert it into a two-dimensional time-frequency image;

[0010] Step S3: Based on the two-dimensional time-frequency image, the constructed diffusion feature migration model is used to perform forward diffusion and reverse denoising processes to generate simulated fault data, and the adversarial feature migration architecture is combined to perform fault feature migration and sharing between different working conditions;

[0011] Step S4: Calculate the similarity between samples to evaluate the enhancement effect, and input the enhanced data into the fault diagnosis model to verify the diagnosis accuracy.

[0012] In one possible implementation, the step of selecting a time-frequency transformation method to enhance data features of the unit vibration signal and converting it into a two-dimensional time-frequency image includes:

[0013] Wavelet transform is selected as the time-frequency transformation method, and the frequency components and energy distribution of vibration signals under different working conditions are generated into time-frequency diagrams, in which the differences in time-frequency characteristics between normal signals and fault signals are displayed through two-dimensional images.

[0014] In one possible implementation, the steps of constructing the diffusion feature migration model include:

[0015] Perform a forward diffusion process to gradually add noise to the original data distribution in molecular form, and iteratively transform the original data distribution into a standard Gaussian distribution;

[0016] The inverse process learning model is trained to simulate the inverse process of the forward diffusion process, restore data from Gaussian distribution noise and generate simulated samples.

[0017] In one possible implementation, the adversarial feature transfer architecture includes a feature extractor and a domain discriminator, where an adversarial effect is generated between the feature extractor and the domain discriminator until a dynamic equilibrium state is reached;

[0018] A feature extractor is used to extract the same data features and update the feature extraction parameters according to the inverted gradient signal;

[0019] Domain discriminator, used to blur the boundary between the target domain and the source domain distribution.

[0020] In one possible implementation, a gradient reversal layer is provided between the feature extractor and the domain discriminator for performing a back-propagation process, inverting the gradient signal generated by the domain classification loss to obtain an inverted gradient signal and sending it to the feature extractor.

[0021] In a possible implementation, step S3 further includes:

[0022] Based on the fully connected layer and activation function, a fault classifier is established to classify the features after dynamic balance optimization.

[0023] In a possible implementation, step S4 includes:

[0024] The distance between the real sample and the generated sample is calculated through the covariance matrix to obtain the similarity between the real samples, the similarity between the generated samples, and the similarity between the real and generated samples;

[0025] The samples after data enhancement are input into the fault diagnosis model for verification.

[0026] In one possible implementation, the device includes:

[0027] Acquisition module, used to collect vibration signals of the unit;

[0028] The feature enhancement module is used to select the time-frequency transformation method to enhance the data features of the unit vibration signal and convert it into a two-dimensional time-frequency image;

[0029] The migration sharing module is used to generate simulated fault data based on two-dimensional time-frequency images by performing forward diffusion and reverse denoising processes using the constructed diffusion feature migration model. It also combines the adversarial feature migration architecture to perform fault feature migration and sharing between different working conditions.

[0030] The verification module is used to calculate the similarity between samples to evaluate the enhancement effect, and input the enhanced data into the fault diagnosis model to verify the diagnosis accuracy.

[0031] The above-mentioned main scheme of this application and its further options can be freely combined to form multiple schemes, all of which are schemes that can be adopted and protected by this application; and in this application, (non-conflicting options) can also be freely combined with each other and with other options. After understanding the scheme of this application, those skilled in the art will understand that there are many combinations based on existing technology and common knowledge, all of which are technical solutions to be protected by this application, and they are not exhaustive here.

[0032] The present application discloses a method for enhancing operational fault data of a hydropower unit. First, the vibration signal of the unit is collected, and a time-frequency transformation method is selected to convert the one-dimensional vibration signal into a two-dimensional time-frequency image to enhance the signal feature dimension. Then, a diffusion feature migration model is constructed, and the data distribution is gradually disturbed to Gaussian noise through forward diffusion. Then, reverse denoising is performed to generate simulated data that is highly similar to the real fault sample. The adversarial feature migration architecture is combined to realize the correlation learning and migration sharing of fault features between different working conditions. Finally, the enhancement effect is evaluated by calculating the similarity between samples, and the enhanced data is input into the fault diagnosis model to verify the accuracy improvement. By combining the time-frequency transformation with the diffusion model, the sample scarcity and working condition barriers are broken through, and significant effects are shown in expanding the scale of fault samples and enriching the sample dimensions, improving the similarity between the generated data and the real samples, improving the diagnosis accuracy, and significantly enhancing the generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work.

[0034] Figure 1 A flow chart of a method for enhancing operational fault data of a hydropower unit proposed in an embodiment of the present application is shown.

[0035] Figure 2a These are the time-frequency transformation results of normal data under four working conditions.

[0036] Figure 2b These are the time-frequency transformation results of the inner race fault under four working conditions.

[0037] Figure 2c These are the time-frequency transformation results of rolling element faults under four working conditions.

[0038] Figure 3a This is the generation result of the real sample corresponding to the inner circle fault data.

[0039] Figure 3b This is the generation result of the real sample corresponding to the rolling element failure data.

[0040] Figure 3c This is the generation result of the generated sample corresponding to the inner circle fault data.

[0041] Figure 3d The generated results of the generated samples corresponding to the rolling element fault data.

[0042] Figure 4This is a diagram of the adversarial feature migration architecture provided in an embodiment of the present application.

[0043] Figure 5 The average diagnostic accuracy results corresponding to four different working conditions. DETAILED DESCRIPTION

[0044] The following describes the embodiments of the present application through specific examples. Those skilled in the art can easily understand the other advantages and effects of the present application from the content disclosed in this specification. The present application can also be implemented or applied through other different specific embodiments. The details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that the following embodiments and features in the embodiments can be combined with each other unless they conflict.

[0045] Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of this application.

[0046] In existing technologies, with the large-scale integration of renewable energy sources such as wind and solar power, the structure of power grid systems is becoming increasingly complex. As hydropower units play a crucial role in regulating grid balance, their operational tasks are also becoming increasingly demanding. Frequent start-stop operations not only exacerbate the wear of their mechanical components but also make the vibration characteristics of the units more variable under different operating conditions. Fault signals are often drowned out by this complex and variable background noise, posing a significant challenge to the accurate extraction of fault features. Given that hydropower units exhibit distinct fault modes under different operating conditions, fault data from a single operating condition is unable to fully reflect all potential fault types of the unit. This operating condition barrier not only limits the generalization capabilities of fault diagnosis models but also increases the difficulty and complexity of fault diagnosis. Therefore, how to effectively integrate, transfer, and share fault knowledge across operating condition boundaries with limited data resources has become a key research topic in the current field of hydropower unit fault diagnosis.

[0047] Therefore, in order to solve the above-mentioned technical problems, the embodiment of the present application proposes a method and device for enhancing the operating fault data of a hydropower unit. By constructing an intelligent model to deeply mine and expand the limited fault data, a more comprehensive and diversified fault data set is constructed to enhance the operating fault data of the hydropower unit, thereby improving the adaptability and accuracy of the fault diagnosis model under different working conditions. It is described in detail below.

[0048] Please refer to Figure 1 , Figure 1 A flow chart of a method for enhancing operational fault data of a hydropower unit proposed in an embodiment of the present application is shown, which includes the following steps:

[0049] Step S1: collecting vibration signals of the unit.

[0050] Vibration sensors are installed at key locations on hydropower units to monitor and record vibration signals during operation in real time. These signals are stored as time series and contain characteristic information about the unit's mechanical state, providing the raw data foundation for subsequent fault analysis. Data collection must cover different operating conditions to ensure data diversity.

[0051] Step S2: Select a time-frequency transformation method to enhance the data features of the unit vibration signal and convert it into a two-dimensional time-frequency image.

[0052] By selecting a time-frequency analysis method, the one-dimensional vibration signal is converted into a two-dimensional time-frequency image, visually displaying the signal's energy distribution characteristics in both time and frequency dimensions. While normal signals exhibit uniform energy distribution in the time-frequency image, faulty signals (such as bearing inner race or rolling element faults) exhibit significant characteristic differences, such as high-frequency spikes or periodic streaks. This step enhances the visualization of signal characteristics, providing easier-to-understand data input for subsequent models.

[0053] Step S2 includes:

[0054] Wavelet transform is selected as the time-frequency transformation method, and the frequency components and energy distribution of vibration signals under different working conditions are generated into time-frequency diagrams, in which the differences in time-frequency characteristics between normal signals and fault signals are displayed through two-dimensional images.

[0055] Wavelet transform is used as the time-frequency analysis method for the vibration signals collected from the model unit test bench (including normal, inner ring fault, rolling element fault and other categories, covering four typical working conditions). Compared with the short-time Fourier transform, the wavelet transform has multi-resolution characteristics and can more accurately capture the transient fault characteristics of non-stationary vibration signals. The one-dimensional time domain vibration signal is mapped to the time-frequency joint domain through wavelet transform to generate a two-dimensional time-frequency diagram. Figure 2a are the time-frequency transformation results of normal data under four working conditions, Figure 2b is the time-frequency transformation result of the inner race fault under four working conditions, Figure 2c These are the time-frequency transformation results of rolling element faults under four working conditions.

[0056] The horizontal axis represents the time series, the vertical axis represents the frequency components, and the pixel color depth represents the signal energy intensity. Under normal signals, the time-frequency plot shows evenly distributed low-frequency energy, with minimal variation in energy distribution under different operating conditions, indicating stable normal operation. However, under fault signals, the time-frequency plots reveal significant characteristic differences: Inner race faults: Under certain operating conditions, a sudden increase in high-frequency components indicates impact vibration caused by bearing inner race damage. Rolling element faults: Periodic energy accumulation bands are visible in the time-frequency plot for Condition 3, corresponding to periodic shocks caused by rolling element defects.

[0057] By converting the collected unit vibration signals into time-frequency images, the signal's information dimension is greatly enhanced, significantly enhancing the fault signature. This method overcomes the limitations of traditional one-dimensional vibration signals in fault feature extraction, providing more comprehensive and accurate data support for the precise classification of fault samples and subsequent training of diagnostic models.

[0058] Step S3: Based on the two-dimensional time-frequency image, the constructed diffusion feature migration model is used to perform forward diffusion and reverse denoising processes to generate simulated fault data, and the adversarial feature migration architecture is combined to perform fault feature migration and sharing between different working conditions.

[0059] Based on two-dimensional time-frequency images, the one-dimensional vibration signal is first converted into a two-dimensional time-frequency image to enhance the information dimension of the signal features. A diffusion feature migration model is then constructed, and a forward diffusion process is performed to gradually add noise to the data, causing the data distribution to gradually approach a Gaussian distribution. This is followed by an inverse denoising process, where the model is trained to learn to restore the data structure from Gaussian noise, generating simulated fault data that is highly similar to real fault samples.

[0060] At the same time, combined with an adversarial feature transfer architecture, the adversarial effect between the feature extractor and the domain discriminator is leveraged to learn, transfer, and share fault features across different operating conditions. The feature extractor extracts key fault features from the simulated data, while the domain discriminator blurs the differences in feature distributions between different operating conditions. Through the backpropagation mechanism of the gradient reversal layer, the feature extractor learns universal feature representations applicable to different operating conditions, thereby achieving effective transfer and sharing of fault features across different operating conditions.

[0061] A diffusion feature transfer model is constructed to expand fault samples, generating simulated samples that are highly similar to real fault data, addressing data scarcity. This model also enables cross-operating feature transfer, learning the correlation between fault features under different operating conditions (such as loads and speeds), and improving model generalization capabilities.

[0062] The steps for constructing the diffusion feature migration model include:

[0063] Perform a forward diffusion process to gradually add noise to the original data distribution in molecular form, and iteratively transform the original data distribution into a standard Gaussian distribution;

[0064] The inverse process learning model is trained to simulate the inverse process of the forward diffusion process, restore data from Gaussian distribution noise and generate simulated samples.

[0065] The forward diffusion process can gradually add noise to the data, and transform the original fault data distribution into a standard Gaussian distribution by gradually adding noise. At each step m (total number of steps L), according to the fixed step size φ m∈(0,1)-directional data X m-1 Add Gaussian noise: Where X0 represents the original data, m represents the number of noise adding steps, L represents the total number of noise adding steps, φ represents the noise adding step length, X m Indicates the noise data corresponding to the noise adding step number m, l(X m |X m-1 ) represents the data distribution of single-step noise addition, l(X 1:L |X0) means that the original data X0 is gradually noised to X L After L steps of iteration, the original data X0 is completely destroyed into Gaussian noise, and its joint distribution is: At the same time, linear or cosine scheduling is usually used to adjust {φ m}, balance the noise intensity and convergence speed, and preserve the overall structure of the data distribution by controlling the noise ratio.

[0066] After the forward diffusion process, the reverse generation process is performed to restore the data from the noise. The training model recovers the original data distribution from the noise and generates simulated data that is highly similar to the real fault samples. Define the reverse distribution r θ , predict the denoised mean and variance at each step through the neural network: where r θ Denotes the distribution q(X t-1 ∣X t ) fitted inverse distribution, l(X m-1 |X m ,X0) represents the distribution inferred from X0.

[0067] By using the trained inverse process learning model, the data structure is gradually restored from the completely disturbed Gaussian distribution noise, thereby realizing data generation and constructing a highly flexible data generation model: Where Q represents the loss function of denoising during data generation, η represents the true value noise at each step in the denoising process, and η θ Indicates that the denoising process follows the distribution r θ (X m-1 |X m ) is the prediction noise constructed.

[0068] Generate samples based on the fault data corresponding to the first working condition. Figure 3a and Figure 3b is the generation result of the real sample corresponding to the inner ring and rolling element fault data, Figure 3c and Figure 3d These are the generated samples corresponding to the inner ring and rolling element fault data. Compared with the real samples, the generated samples are slightly lower in clarity, but can still effectively simulate the real samples to a certain extent. Their fault characteristics are similar to those of the real samples, and key information is retained.

[0069] The adversarial feature transfer architecture includes a feature extractor and a domain discriminator, which produce an adversarial effect between the feature extractor and the domain discriminator until a dynamic equilibrium state is reached;

[0070] A feature extractor is used to extract the same data features and update the feature extraction parameters according to the inverted gradient signal;

[0071] Domain discriminator, used to blur the boundary between the target domain and the source domain distribution.

[0072] Figure 4 This is a diagram of the adversarial feature migration architecture provided in an embodiment of the present application. The architecture integrates two main components: a feature extractor and a domain discriminator. The feature extractor and the domain discriminator cooperate with each other and produce an adversarial effect to achieve the learning, migration and sharing of fault feature knowledge between different working conditions, thereby achieving a dynamic equilibrium state. The feature extractor extracts representative and discriminative features from the input data. These features can reflect the essential characteristics of the data. For example, in the operational fault data of a hydropower unit, the feature extractor can extract key information such as frequency components and energy distribution related to different fault types. The parameter update of the feature extractor is based on the inverted gradient signal. During the adversarial process, the gradient signal generated by the domain discriminator is inverted through the gradient inversion layer and passed to the feature extractor. The feature extractor adjusts its own parameters based on these inverted gradient signals. In this way, the feature extractor can learn a general feature representation that can effectively characterize the characteristics of the source domain data and adapt to the characteristics of the target domain data, thereby improving the migration and generalization capabilities of the features.

[0073] The domain discriminator can distinguish whether the input features come from the source domain or the target domain, that is, determine whether the features come from the data of the original working condition or data from other working conditions. It discriminates the domain affiliation of the features by learning the distribution differences between the source domain and the target domain. In the adversarial process, the domain discriminator and the feature extractor compete with each other. The feature extractor attempts to extract data features that can blur the boundary between the source domain and the target domain, making it difficult for the domain discriminator to distinguish the source of the features; while the domain discriminator strives to learn the differences between the source domain and the target domain features in order to accurately classify the domain affiliation of the features. This adversarial mechanism prompts the feature extractor to continuously optimize the feature extraction process and extract more transferable features, thereby reducing the difference in the distribution of source domain and target domain features and realizing the transfer and sharing of fault feature knowledge between different working conditions.

[0074] Through continuous iteration of the adversarial process, the feature extractor and domain discriminator gradually reach a dynamic equilibrium state. In this state, the features extracted by the feature extractor are able to adapt to the distribution of both the source and target domains to a certain extent, while the domain discriminator's ability to discriminate between features cannot be further improved, that is, it cannot accurately distinguish whether the features come from the source or target domain. Reaching a dynamic equilibrium state means that the adversarial feature transfer architecture has successfully learned the correlation between fault features across different operating conditions, achieving effective transfer and sharing of fault feature knowledge. At this point, the extracted features retain the key information of the original fault sample while being able to adapt to data changes under different operating conditions.

[0075] A gradient reversal layer is set between the feature extractor and the domain discriminator for back propagation. The gradient signal generated by the domain classification loss is reversed to obtain the reversed gradient signal and sent to the feature extractor.

[0076] Backpropagation process: During training, the domain discriminator determines whether the features extracted by the feature extractor are from the source domain or the target domain, and calculates the domain classification loss. The gradient reversal layer performs almost no operation during forward propagation and outputs the same as the input, but during backpropagation, it reverses the gradient (changes the sign of the gradient):

[0077]

[0078] Among them, G bo and G co represents the loss of boundary distribution discriminator and conditional distribution discriminator, N e and N k represents the number of samples in the source domain e and the target domain k, D e and D k G represents the distance between the conditional distribution and the boundary distribution in the source domain e and the target domain k. d represents the domain discrimination loss, H bo represents the boundary distribution loss discriminator, H c represents the fault classifier, d i represents the label input, N represents the total number of wrong labels, and Represents the conditional distribution loss value and conditional distribution discriminant value of label j, Indicates that label j is in X i The conditional probability distribution on , δ represents the weight factor for achieving dynamic balance, and Represents the distance between the marginal probability distribution and the conditional probability distribution.

[0079] When the gradient signal of the domain classification loss passes through the gradient reversal layer, the direction of the gradient is reversed, that is, the gradient value is multiplied by -1. In this way, the gradient signal originally used to optimize the domain discriminator is now used to update the parameters of the feature extractor.

[0080] The feature extractor parameters are updated through a parameter update process. The inverted gradient signal is passed to the feature extractor, which then updates its own feature extraction parameters based on these inverted gradient signals. In this way, the feature extractor can learn a universal feature representation that effectively characterizes the characteristics of the source domain data and adapts to the characteristics of the target domain data. The goal of the feature extractor is to extract data features that blur the boundary between the source and target domains, making it difficult for the domain discriminator to distinguish the source of the features. With continuous iteration, the feature extractor will gradually optimize its parameters and extract more transferable features.

[0081] Adversarial Effect: The gradient reversal layer creates an adversarial effect between the feature extractor and the domain discriminator. The feature extractor attempts to extract features that can deceive the domain discriminator, while the domain discriminator strives to learn the differences between source and target domain features in order to accurately classify the domain of the features. This adversarial mechanism forces the feature extractor to continuously optimize the feature extraction process and reduce the difference in feature distribution between the source and target domains.

[0082] Dynamic equilibrium: Through continuous iterative adversarial processes, the feature extractor and domain discriminator gradually reach a stable dynamic equilibrium. In this state, the features extracted by the feature extractor are able to adapt to the distribution of both the source and target domains to a certain extent, and the domain discriminator's ability to distinguish between features cannot be further improved.

[0083] Reaching a dynamic equilibrium state means that the adversarial feature transfer architecture has successfully learned the correlations between fault features across different operating conditions, enabling the effective transfer and sharing of fault feature knowledge. At this point, the extracted features retain the key information of the original fault sample while adapting to data variations under different operating conditions, providing a more robust and versatile feature representation for subsequent fault diagnosis. Through the action of the gradient reversal layer and adversarial training, the feature extractor is able to extract more transferable and robust features, thereby improving the diagnostic accuracy and generalization ability of the fault diagnosis model under different operating conditions.

[0084] Step S3 further includes:

[0085] Based on the fully connected layer and activation function, a fault classifier is established to classify the features after dynamic balance optimization.

[0086] The fully connected layer connects neurons, allowing each neuron to connect to all neurons in the previous layer. It comprehensively considers the information of each part of the input feature and establishes a mapping relationship between the feature and the fault category. The activation function converts the mapping result into a probability distribution, which represents the probability that the input feature belongs to each fault category, thereby achieving category classification. The classification loss function is: The total loss function is: G = G c -ε((1-δ)G bo +δG co ), G c represents the training goal of the fault classifier, G represents the learning goal of the adversarial feature transfer architecture, F represents the fault category, represents the probability corresponding to fault category f, H g represents the feature extractor, and ε represents the inversion layer parameter.

[0087] Step S4: Calculate the similarity between samples to evaluate the enhancement effect, and input the enhanced data into the fault diagnosis model to verify the diagnosis accuracy.

[0088] The covariance matrix is ​​used to calculate the distance between real and generated samples, quantitatively assessing their similarity and ultimately determining the quality and authenticity of the augmented data. The smaller the similarity, the closer the generated sample is to the real sample, and the better the augmentation effect. Furthermore, by calculating the similarity between real samples, generated samples, and real and generated samples, we comprehensively analyze the feature distribution and information retention of the augmented data, ensuring that the generated simulated fault data effectively reflects the characteristics of the real fault.

[0089] To verify the improved performance of the fault diagnosis model achieved with augmented data, the augmented data was fed into the model for training and testing. The effectiveness of the data augmentation method was demonstrated by comparing the diagnostic accuracy before and after augmentation. Improved diagnostic accuracy indicates that the simulated data generated by the diffusion feature transfer model has successfully enriched the features and number of fault samples, enhancing the model's generalization and fault identification capabilities, thus demonstrating the effectiveness of the proposed method in improving fault diagnosis accuracy.

[0090] Step S4 includes:

[0091] The distance between the real sample and the generated sample is calculated through the covariance matrix to obtain the similarity between the real samples, the similarity between the generated samples, and the similarity between the real and generated samples;

[0092] The samples after data enhancement are input into the fault diagnosis model for verification.

[0093] According to the similarity calculation principle, the distance between the two images is calculated by the covariance matrix to represent the difference between the two images. The similarity between the real samples, the similarity between the generated samples, and the similarity between the real and generated samples are calculated respectively: A = || β τ -β υ || 2 +Tr(∑ τ +∑ υ -2(∑ τ ∑ υ ) 1 / 2 ), where A represents similarity, (β τ ,∑ τ ) and (β υ ,∑ υ ) represent the real samples B τ and generate sample B υ The Gaussian distribution parameters of , β and ∑ represent the mean vector and covariance matrix. Table 1 shows the similarity calculation results between normal signals, inner race and rolling element fault samples:

[0094] Table 1

[0095] category Real samples Generate samples Real and generated samples Normal signal 20.27 6.77 10.85 Inner race fault 34.95 16.95 19.54 Rolling element failure 35.40 17.36 32.15

[0096] The smaller the similarity value, the closer the two images are. This is verified by the highest similarity value between real samples in Table 1. The lower similarity values ​​between generated samples reflect the fuzziness of the generated images and the difficulty in capturing differences between samples. The slightly lower similarity between real and generated samples compared to real samples indicates that the generated samples have a certain degree of similarity with the real samples and are not completely random. At the same time, this similarity value is higher than that between generated samples, indicating that the generated samples are not simple copies of the real samples but rather similar samples generated by learning their distribution. Furthermore, the similarity between normal signal samples is significantly lower than that between faulty samples. This is because normal signal samples do not contain fault information, resulting in less information and lower similarity.

[0097] The first working condition sample is used as the source domain for the diffusion feature transfer model learning, and the remaining three working conditions are used as target domains to learn the fault features between different working conditions, thereby achieving fault data enhancement and fault feature migration and sharing. The enhanced data processed by the diffusion feature transfer model is input into the fault diagnosis model. Figure 5 The average diagnostic accuracy results corresponding to four different working conditions.

[0098] Under different operating conditions, the fault diagnosis accuracy on the original dataset shows a decreasing trend. Specifically, under the first operating condition, the method of data enhancement using the diffusion feature transfer model shows a diagnostic accuracy close to 100%. However, when the operating condition changes to the fourth, this accuracy shows a significant decline, falling to around 93%. The main reason for this change is that as the load continues to increase, the speed gradually decreases accordingly, which makes the effective information contained in the sample increasingly scarce. Therefore, when the model attempts to learn fault features from these information-deficient samples, the acquired features are often incomplete, which leads to a decrease in diagnostic accuracy.

[0099] In addition, it should be noted that the original data set itself has the problem of sample imbalance, that is, the number of samples of various types of faults is not consistent. This imbalance makes some faults insufficiently representative in the data set, thereby increasing the difficulty of the model to identify these faults, and further affecting the accuracy of the diagnosis. In order to solve this problem, the present invention adopts a diffusion feature migration model. The model can cleverly learn the characteristics of the same fault from other working conditions and migrate them to the current working conditions, thereby breaking the barriers of sample features between different working conditions and realizing knowledge sharing of the same fault characteristics. Therefore, even in the face of insufficient data on a certain fault, the diagnostic model will not be restricted by the working conditions, and can extract fault features more deeply, realize the enhancement of hydropower unit operation fault data, and avoid a decrease in diagnostic accuracy.

[0100] Under the first three operating conditions, the diagnostic accuracy of data enhancement based on the diffusion feature migration model exceeded 99%. This result fully verifies the excellent effectiveness of the hydropower unit operation fault data enhancement method proposed in this invention in improving the accuracy of fault diagnosis.

[0101] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0102] First, by generating simulated fault data through the diffusion feature migration model, the scale of fault samples is effectively expanded, solving the problems of difficulty in obtaining fault data and insufficient sample quantity in actual engineering projects, and providing richer training data for the fault diagnosis model.

[0103] Second, converting the one-dimensional vibration signal into a two-dimensional time-frequency image increases the information dimension of the data, making the fault characteristics more obvious and easier to distinguish, which helps to improve the accuracy of fault diagnosis.

[0104] Third, combined with the adversarial feature transfer architecture, the knowledge transfer and sharing of fault features between different working conditions is realized, breaking the limitations of working condition differences on fault diagnosis and improving the adaptability and generalization ability of the model under different working conditions.

[0105] Fourth, the generated simulated data is highly similar to real data and contains key fault information, which enables the fault diagnosis model to learn more comprehensive and representative features during the training process, enhancing the model's robustness and fault identification capabilities.

[0106] Fifth, the sample input fault diagnosis model after data enhancement has been verified to show that this method can significantly improve the diagnosis accuracy and maintain a high fault diagnosis accuracy under different working conditions, providing a more reliable basis for the maintenance and overhaul of hydropower units.

[0107] Sixth, this solution has shown significant results in expanding the scale of fault samples and enriching the sample dimensions, effectively enhancing the fault identification capability of subsequent diagnostic models. It has great practical engineering application value and can provide strong support for fault diagnosis and health management of hydropower units.

[0108] The following describes a possible implementation of a device for enhancing operational fault data of a hydropower unit, which is used to perform the various steps and corresponding technical effects of the method for enhancing operational fault data of a hydropower unit as described in the above embodiments and possible implementations. The device includes:

[0109] Acquisition module, used to collect vibration signals of the unit;

[0110] The feature enhancement module is used to select the time-frequency transformation method to enhance the data features of the unit vibration signal and convert it into a two-dimensional time-frequency image;

[0111] The migration sharing module is used to generate simulated fault data based on two-dimensional time-frequency images by performing forward diffusion and reverse denoising processes using the constructed diffusion feature migration model. It also combines the adversarial feature migration architecture to perform fault feature migration and sharing between different working conditions.

[0112] The verification module is used to calculate the similarity between samples to evaluate the enhancement effect, and input the enhanced data into the fault diagnosis model to verify the diagnosis accuracy.

[0113] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A method for enhancing operational fault data of a hydropower unit, characterized in that: The method comprises: Step S1, collecting the vibration signal of the unit; Step S2: Select a time-frequency transformation method to enhance the data features of the unit vibration signal and convert it into a two-dimensional time-frequency image; Step S3: Based on the two-dimensional time-frequency image, the constructed diffusion feature migration model is used to perform forward diffusion and reverse denoising processes to generate simulated fault data, and the adversarial feature migration architecture is combined to perform fault feature migration and sharing between different working conditions; Step S4: Calculate the similarity between samples to evaluate the enhancement effect, and input the enhanced data into the fault diagnosis model to verify the diagnosis accuracy.

2. The method for enhancing the operating fault data of a hydropower unit according to claim 1, characterized in that: The steps of selecting a time-frequency transformation method to enhance the data features of the unit vibration signal and converting it into a two-dimensional time-frequency image include: Wavelet transform is selected as the time-frequency transformation method, and the frequency components and energy distribution of vibration signals under different working conditions are generated into time-frequency diagrams, in which the differences in time-frequency characteristics between normal signals and fault signals are displayed through two-dimensional images.

3. The method for enhancing the operating fault data of a hydropower unit according to claim 1, characterized in that: The steps for constructing the diffusion feature migration model include: Perform a forward diffusion process to gradually add noise to the original data distribution in molecular form, and iteratively transform the original data distribution into a standard Gaussian distribution; The inverse process learning model is trained to simulate the inverse process of the forward diffusion process, restore data from Gaussian distribution noise and generate simulated samples.

4. The method for enhancing the operating fault data of a hydropower unit according to claim 3, characterized in that: The adversarial feature transfer architecture includes a feature extractor and a domain discriminator, which produce an adversarial effect between the feature extractor and the domain discriminator until a dynamic equilibrium state is reached; A feature extractor is used to extract the same data features and update the feature extraction parameters according to the inverted gradient signal; Domain discriminator, used to blur the boundary between the target domain and the source domain distribution.

5. The method for enhancing the operating fault data of a hydropower unit according to claim 4, characterized in that: A gradient reversal layer is set between the feature extractor and the domain discriminator for back propagation. The gradient signal generated by the domain classification loss is reversed to obtain the reversed gradient signal and sent to the feature extractor.

6. The method for enhancing the operating fault data of a hydropower unit according to claim 3, characterized in that: Step S3 further includes: Based on the fully connected layer and activation function, a fault classifier is established to classify the features after dynamic balance optimization.

7. The method for enhancing the operating fault data of a hydropower unit according to claim 1, characterized in that: Step S4 includes: The distance between the real sample and the generated sample is calculated through the covariance matrix to obtain the similarity between the real samples, the similarity between the generated samples, and the similarity between the real and generated samples; The samples after data enhancement are input into the fault diagnosis model for verification.

8. A device for enhancing operational fault data of a hydropower unit, characterized in that: The device comprises: Acquisition module, used to collect vibration signals of the unit; The feature enhancement module is used to select the time-frequency transformation method to enhance the data features of the unit vibration signal and convert it into a two-dimensional time-frequency image; The migration sharing module is used to generate simulated fault data based on two-dimensional time-frequency images by performing forward diffusion and reverse denoising processes using the constructed diffusion feature migration model. It also combines the adversarial feature migration architecture to perform fault feature migration and sharing between different working conditions. The verification module is used to calculate the similarity between samples to evaluate the enhancement effect, and input the enhanced data into the fault diagnosis model to verify the diagnosis accuracy.

Images