Method and device for detecting fan bearing of wind power generation equipment

Abstract

The invention discloses a method and a device for detecting a fan bearing of wind power generation equipment, which can realize continuous adaptation and stable diagnosis capability improvement of a fan bearing diagnosis model under variable working conditions through pre-training guided by physical prompt and incremental updating based on new working condition data.

Description

Technical Field

[0001] This invention relates to the field of wind turbine bearing testing technology for wind power generation equipment, and in particular to a method for testing wind turbine bearings of wind power generation equipment, a device for testing wind turbine bearings of wind power generation equipment, an electronic device, and a readable storage medium. Background Technology

[0002] As an important carrier of clean energy, the bearing assembly of wind turbine generator sets plays a crucial role in the overall reliability and power generation efficiency. Wind turbine bearings are in a complex and ever-changing operating environment for a long time, and are continuously affected by factors such as wind speed, load, temperature, and humidity. This results in bearing vibration signals exhibiting significant multi-condition characteristics, including non-stationarity, distributed drift, and multimodal features.

[0003] Traditional wind turbine bearing fault diagnosis methods are mostly based on data from laboratory fixed operating conditions or finite stability loading conditions for modeling and training. These methods are difficult to adapt effectively to the changes in data distribution caused by frequent switching of operating conditions in actual wind farm applications, which leads to a significant decrease in the generalization performance of the diagnostic model and a reduction in the accuracy of fault identification. Summary of the Invention

[0004] The present invention provides a method, apparatus, electronic device, and readable storage medium for testing wind turbine bearings in wind power generation equipment, in order to overcome or at least partially solve the above-mentioned problems.

[0005] To solve the above-mentioned technical problems, this application is implemented as follows: In a first aspect, embodiments of this application provide a method for testing wind turbine bearings in wind power generation equipment, including: Obtain the equivalent mass, damping coefficient, and time-varying equivalent stiffness of the wind turbine bearing; Read the structural parameters of the wind turbine bearing, and calculate the theoretical characteristic frequency of the fault at the target position of the wind turbine bearing under the target operating condition based on the structural parameters; By splicing together the theoretical fault characteristic frequency, the equivalent mass, the damping coefficient, and the equivalent stiffness, a physical cue vector is generated; Collect raw vibration signals under different working conditions; The target loss function is determined based on the physical cue vector and the original vibration signal, and a physical information diffusion model with cross-condition basic generation and diagnosis capabilities is pre-trained based on the target loss function. The vibration signal under the new working condition is received in real time and identified as online incremental data. Incremental training is performed on the physical information diffusion model based on the online incremental data to generate optimal incremental model parameters, and an incremental diagnostic model is generated based on the optimal incremental model parameters. When the current vibration signal of the target wind turbine bearing is received, the incremental diagnostic model is controlled to generate the detection results for the target wind turbine bearing.

[0006] Optionally, before the step of determining the target loss function based on the physical cue vector and the original vibration signal, the method further includes: The original vibration signal is subjected to wavelet threshold denoising and normalization operations to generate a preprocessed original vibration signal.

[0007] Optionally, the step of determining the target loss function based on the physical cue vector and the original vibration signal includes: Construct a deep learning framework that includes a feature extraction module composed of a convolutional neural network, a diffusion and inverse diffusion module based on a diffusion probability model architecture, and a classification module. The preprocessed original vibration signal is input into the feature extraction module to generate a high-dimensional feature vector. By fusing the high-dimensional feature vector with the physical cue vector, and performing forward diffusion noise addition and reverse diffusion denoising operations through the diffusion and de-diffusion modules, the reconstructed original feature vector is obtained. A physical correlation model between the fault impact load and the final vibration displacement is established based on the physical equation of vibration response using the equivalent mass, the damping coefficient, and the equivalent stiffness. Substitute the original feature vector into the physical association model, calculate the error between the original feature vector and the physical cue vector, and generate the physical constraint loss. Calculate the mean square error between the original feature vector and the high-dimensional feature vector to obtain the inverse diffusion reconstruction loss; The classification module determines the predicted probability of the fault type to which the original feature vector belongs, and the true label of the original vibration signal corresponding to the original feature vector; The original feature vector is input into the classification module, and the cross-entropy loss between the predicted probability and the true label is calculated to obtain the classification loss. The total loss function is constructed by combining the physical constraint loss, the inverse diffusion reconstruction loss, and the classification loss.

[0008] Optionally, the steps of pre-training a physical information diffusion model with cross-condition basic generation and diagnosis capabilities based on the target loss function include: The deep learning framework is pre-trained using the total loss function to obtain pre-trained model parameters; The parameters of the pre-trained model are solidified to generate a pre-trained physical information diffusion model.

[0009] Optionally, the step of performing incremental training on the physical information diffusion model based on the online incremental data to generate optimal incremental model parameters includes: Wavelet threshold denoising and normalization operations are performed on the online incremental data to generate preprocessed online incremental training and validation sets. Freeze the bottom-level parameters of the feature extraction module in the pre-trained physical information diffusion model, and adjust the top-level parameters of the feature extraction module, the diffusion and inverse diffusion module, and the classification module to obtain a fine-tuning parameter set; By fusing the online incremental training set with the physical cue vector, and performing incremental learning operations through the fine-tuned parameter set, the inverse diffusion reconstruction loss, frequency bias loss, dynamic bias loss, and classification loss of the online incremental data are calculated. An incremental loss function is constructed by combining the inverse diffusion reconstruction loss, the frequency bias loss, the dynamic bias loss, and the classification loss of the incremental data; Based on the reverse diffusion process of the pre-trained physical information diffusion model, random noise is input to generate virtual historical samples that are consistent with the historical fault type characteristics of the original vibration signal. The virtual historical samples are mixed with the online incremental training set at a preset ratio to form a hybrid training set; The incremental loss function is used to perform incremental training on the hybrid training set to obtain incrementally updated model parameters; The classification accuracy and loss value of the incrementally updated model parameters are calculated using the online validation set. When the classification accuracy and the loss value meet the preset conditions, the incremental training operation on the mixed training set is stopped, and the optimal incremental model parameters are generated.

[0010] Optionally, the step of controlling the incremental diagnostic model to generate detection results for the target wind turbine bearing includes: Perform wavelet threshold denoising and normalization operations on the current vibration signal to generate a preprocessed current vibration signal; The preprocessed current vibration signal is input into the feature extraction module of the incremental diagnostic model to generate the current high-dimensional feature vector. By fusing the current high-dimensional feature vector with the physical cue vector, a reconstruction operation is performed through the diffusion-inverse diffusion module of the incremental diagnostic model to obtain the reconstructed current feature vector; The reconstructed current feature vector is input into the classification module of the incremental diagnostic model, and the failure type prediction result of the target wind turbine bearing and the confidence level of the failure type prediction result are output. If the confidence level is lower than a preset threshold, then the reconstructed current feature vector is subjected to secondary optimization through the inverse diffusion process of the incremental diagnostic model to obtain a refined feature vector, and the physical constraint error of the refined feature vector and the physical correlation model is calculated. When the physical constraint error is less than a preset threshold, the refined feature vector is used to generate a corrected fault type prediction result.

[0011] Optionally, it also includes: When the physical constraint error is not less than a preset threshold, an early warning signal is generated, and the fault type prediction result is associated with and stored with real-time operating condition data to generate a fault condition association database.

[0012] Secondly, embodiments of this application provide a wind turbine bearing testing device for wind power generation equipment, characterized in that it includes: The bearing rotating component parameter acquisition module is used to acquire the equivalent mass, damping coefficient, and time-varying equivalent stiffness of the wind turbine bearing. The fault theoretical characteristic frequency calculation module is used to read the structural parameters of the wind turbine bearing and calculate the fault theoretical characteristic frequency of the target position of the wind turbine bearing under the target operating condition based on the structural parameters. The physical cue vector generation module is used to concatenate the fault theoretical characteristic frequency, the equivalent mass, the damping coefficient, and the equivalent stiffness to generate a physical cue vector; The raw vibration signal acquisition module is used to acquire raw vibration signals under different working conditions; The physical information diffusion model training module is used to determine the target loss function based on the physical cue vector and the original vibration signal, and to pre-train a physical information diffusion model with cross-condition basic generation and diagnosis capabilities based on the target loss function. The online incremental data acquisition module is used to receive vibration signals under new working conditions in real time and identify the vibration signals under new working conditions as online incremental data. The incremental diagnostic model generation module is used to perform incremental training on the physical information diffusion model based on the online incremental data, generate optimal incremental model parameters, and generate an incremental diagnostic model based on the optimal incremental model parameters. The detection result generation module is used to control the incremental diagnostic model to generate detection results for the target wind turbine bearing when the current vibration signal of the target wind turbine bearing is received.

[0013] Thirdly, embodiments of this application provide an electronic device including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the method described in the first aspect.

[0014] Fourthly, embodiments of this application provide a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method described in the first aspect.

[0015] Fifthly, embodiments of this application provide a chip, the chip including a processor and a communication interface, the communication interface being coupled to the processor, the processor being used to run programs or instructions to implement the method as described in the first aspect.

[0016] The embodiments of the present invention have the following advantages: In this embodiment of the invention, the equivalent mass, damping coefficient, and time-varying equivalent stiffness of the wind turbine bearing are obtained; the structural parameters of the wind turbine bearing are read, and the theoretical fault characteristic frequency of the target position of the wind turbine bearing under the target operating condition is calculated based on the structural parameters; the theoretical fault characteristic frequency, the equivalent mass, the damping coefficient, and the equivalent stiffness are concatenated to generate a physical cue vector; original vibration signals under different operating conditions are collected; a target loss function is determined based on the physical cue vector and the original vibration signals, and a system with cross-operating condition basic generation and diagnosis capabilities is pre-trained based on the target loss function. A physical information diffusion model is used to receive vibration signals under new operating conditions in real time and identify these signals as online incremental data. Incremental training is performed on the physical information diffusion model based on this online incremental data to generate optimal incremental model parameters, and an incremental diagnostic model is generated based on these optimal parameters. When the current vibration signal of the target wind turbine bearing is received, the incremental diagnostic model is controlled to generate detection results for the target wind turbine bearing. This allows for continuous adaptation and improved stable diagnostic capabilities of the wind turbine bearing diagnostic model under varying operating conditions through physical cue-guided pre-training and incremental updates based on new operating condition data. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the steps of a wind turbine bearing testing method for wind power generation equipment provided in an embodiment of the present invention; Figure 2 This is a structural block diagram of a wind turbine bearing testing device for wind power generation equipment provided in an embodiment of the present invention; Figure 3 This is a hardware structure block diagram of an electronic device provided in an embodiment of the present invention; Figure 4This is a schematic diagram of a computer-readable medium provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] Reference Figure 1 The diagram illustrates a flowchart of a wind turbine bearing testing method for a wind power generation device provided in an embodiment of the present invention, which may specifically include the following steps: Step 101: Obtain the equivalent mass, damping coefficient, and time-varying equivalent stiffness of the wind turbine bearing; Step 102: Read the structural parameters of the wind turbine bearing, and calculate the theoretical characteristic frequency of the fault at the target position of the wind turbine bearing under the target operating condition based on the structural parameters; Step 103: Concatenate the fault theoretical characteristic frequency, the equivalent mass, the damping coefficient, and the equivalent stiffness to generate a physical cue vector; Step 104: Collect raw vibration signals under different working conditions; Step 105: Determine the target loss function based on the physical cue vector and the original vibration signal, and pre-train a physical information diffusion model with cross-condition basic generation and diagnosis capabilities based on the target loss function; Step 106: Receive the vibration signal under the new working condition in real time, and determine the vibration signal under the new working condition as online incremental data; Step 107: Perform incremental training on the physical information diffusion model based on the online incremental data to generate optimal incremental model parameters, and generate an incremental diagnostic model based on the optimal incremental model parameters; Step 108: When the current vibration signal of the target wind turbine bearing is received, control the incremental diagnostic model to generate the detection results for the target wind turbine bearing.

[0022] In this embodiment of the invention, the equivalent mass, damping coefficient, and time-varying equivalent stiffness of the wind turbine bearing can be obtained to acquire the core physical parameters of the wind turbine bearing vibration response, providing basic constraints for subsequent physical information guidance.

[0023] Equivalent mass refers to the numerical value of the equivalent lumped mass when the bearing system is simplified to a single-degree-of-freedom model, reflecting the inertial characteristics of the bearing and related components.

[0024] The damping coefficient is a parameter that reflects the energy dissipation capacity of a bearing system during vibration. It is usually determined by material damping, lubricating oil film damping, and other factors.

[0025] Equivalent stiffness over time refers to the stiffness degradation of a bearing during operation due to fault evolution such as wear, fatigue, and crack propagation. It is represented by a time-dependent function and is used to capture the dynamic development of faults.

[0026] In this embodiment of the invention, the structural parameters of the wind turbine bearing can be read, and the theoretical characteristic frequency of the fault at the target position of the wind turbine bearing under the target operating condition can be calculated based on the structural parameters. The theoretical characteristic frequencies of different fault types under the current operating condition can be derived based on the bearing geometry, providing frequency domain prior knowledge for physical indications.

[0027] Structural parameters include bearing inner diameter, outer diameter, rolling element diameter, number of rolling elements, pitch circle diameter, contact angle, and other geometric dimensions.

[0028] The target location refers to the specific part of the bearing to be diagnosed, such as the inner ring, outer ring, rolling elements, or cage.

[0029] The target operating condition refers to the current or specified combination of operating conditions, which typically includes speed, load, temperature, etc.

[0030] The theoretical characteristic frequency of a fault is the frequency of repeated impact pulses that should appear in the frequency spectrum for a specific fault (such as inner ring fault, outer ring fault, or rolling element fault) calculated based on the bearing kinematics.

[0031] In this embodiment of the invention, the fault theoretical characteristic frequency, the equivalent mass, the damping coefficient, and the equivalent stiffness can be concatenated to generate a physical cue vector, thereby integrating the scattered physical prior information into a unified vector form, which is convenient for input into a deep learning model as conditional guidance.

[0032] The physical hint vector is a numerical vector composed of the above physical parameters in a certain order, used to constrain the model generation process to conform to the actual physical laws of bearings.

[0033] In this embodiment of the invention, raw vibration signals under different operating conditions can be collected to obtain raw data of real wind turbine bearings under various operating conditions, which can be used as the basic input for model learning.

[0034] The raw vibration signal is the time-domain acceleration or velocity signal directly collected by the wind turbine bearing sensor, and is the raw measurement data without any processing.

[0035] Different operating conditions refer to different operating states with different combinations of parameters such as wind speed, load, rotational speed, and temperature.

[0036] In this embodiment of the invention, a target loss function can be determined based on the physical cue vector and the original vibration signal, and a physical information diffusion model with cross-condition basic generation and diagnosis capabilities can be pre-trained based on the target loss function. In this way, a generative model that can stably generate reasonable vibration features and has preliminary diagnostic capabilities under various conditions can be pre-trained through a combination of physical constraints and data-driven methods.

[0037] The objective loss function can be an optimization objective that comprehensively considers multiple losses such as data reconstruction error, physical consistency constraint, and fault classification error.

[0038] The physical information diffusion model is a generative model based on the diffusion probability model and incorporating physical cue constraints. It can gradually reconstruct vibration characteristics that conform to physical laws from noise.

[0039] Cross-condition basic generation and diagnosis capability means that the model has learned the distribution patterns under multiple operating conditions during the pre-training stage, and can generate reasonable signals and initially distinguish fault types.

[0040] In this embodiment of the invention, vibration signals under new operating conditions can be received in real time and identified as online incremental data to capture new operating condition data that exceed the historical range during wind turbine operation, providing input for continuous model adaptation.

[0041] Online incremental data refers to newly collected vibration signals and their corresponding operating conditions that have not been pre-trained during the real-time operation of the wind turbine.

[0042] In this embodiment of the invention, incremental training of the physical information diffusion model can be performed based on the online incremental data to generate optimal incremental model parameters, and an incremental diagnostic model can be generated based on the optimal incremental model parameters. This allows for targeted updates to the existing model using new data, enabling it to adapt to new working conditions while preserving existing knowledge as much as possible, resulting in an updated diagnostic model that can be deployed immediately.

[0043] Incremental training is a training process that continues to optimize existing model parameters using only new data (usually combined with a historical knowledge preservation mechanism).

[0044] The optimal incremental model parameters are the set of model parameters that are optimal after incremental training and when the validation metrics (such as classification accuracy and loss value) are at their best.

[0045] The incremental diagnostic model is the final model instantiated using the optimal incremental model parameters, and it can handle multi-condition fault diagnosis tasks, including new operating conditions.

[0046] In this embodiment of the invention, when the current vibration signal of the target wind turbine bearing is received, the incremental diagnostic model can be controlled to generate detection results for the target wind turbine bearing. This allows the continuously updated diagnostic model to identify faults in the currently acquired vibration signal and provide diagnostic conclusions.

[0047] The current vibration signal is the real-time vibration data collected at the current moment for the target wind turbine bearing.

[0048] The detection results are fault type judgments, confidence levels, or other relevant diagnostic information output by the incremental diagnostic model.

[0049] In this embodiment of the invention, the equivalent mass, damping coefficient, and time-varying equivalent stiffness of the wind turbine bearing are obtained; the structural parameters of the wind turbine bearing are read, and the theoretical fault characteristic frequency of the target position of the wind turbine bearing under the target operating condition is calculated based on the structural parameters; the theoretical fault characteristic frequency, the equivalent mass, the damping coefficient, and the equivalent stiffness are concatenated to generate a physical cue vector; original vibration signals under different operating conditions are collected; a target loss function is determined based on the physical cue vector and the original vibration signals, and a system with cross-operating condition basic generation and diagnosis capabilities is pre-trained based on the target loss function. A physical information diffusion model is used to receive vibration signals under new operating conditions in real time and identify these signals as online incremental data. Incremental training is performed on the physical information diffusion model based on this online incremental data to generate optimal incremental model parameters, and an incremental diagnostic model is generated based on these optimal parameters. When the current vibration signal of the target wind turbine bearing is received, the incremental diagnostic model is controlled to generate detection results for the target wind turbine bearing. This allows for continuous adaptation and improved stable diagnostic capabilities of the wind turbine bearing diagnostic model under varying operating conditions through physical cue-guided pre-training and incremental updates based on new operating condition data.

[0050] Based on the above embodiments, modified embodiments of the above embodiments are proposed. It should be noted that, in order to keep the description brief, only the differences from the above embodiments are described in the modified embodiments.

[0051] In an optional embodiment of the present invention, the equivalent mass, damping coefficient, and time-varying equivalent stiffness of the wind turbine bearing can be obtained. Read the structural parameters of the wind turbine bearing, and calculate the theoretical characteristic frequency of the fault at the target position of the wind turbine bearing under the target operating condition based on the structural parameters; By splicing together the theoretical fault characteristic frequency, the equivalent mass, the damping coefficient, and the equivalent stiffness, a physical cue vector is generated; A physical correlation model between the fault impact load and the final vibration displacement is established based on the physical equation of vibration response using the equivalent mass, the damping coefficient, and the equivalent stiffness. Collect raw vibration signals under different working conditions; Perform wavelet threshold denoising and normalization operations on the original vibration signal to generate a preprocessed original vibration signal. Construct a deep learning framework that includes a feature extraction module composed of a convolutional neural network, a diffusion and inverse diffusion module based on a diffusion probability model architecture, and a classification module. The preprocessed original vibration signal is input into the feature extraction module to generate a high-dimensional feature vector. By fusing the high-dimensional feature vector with the physical cue vector, and performing forward diffusion noise addition and reverse diffusion denoising operations through the diffusion and de-diffusion modules, the reconstructed original feature vector is obtained. Substitute the original feature vector into the physical association model, calculate the error between the original feature vector and the physical cue vector, and generate the physical constraint loss. Calculate the mean square error between the original feature vector and the high-dimensional feature vector to obtain the inverse diffusion reconstruction loss; The classification module determines the predicted probability of the fault type to which the original feature vector belongs, and the true label of the original vibration signal corresponding to the original feature vector; The original feature vector is input into the classification module, and the cross-entropy loss between the predicted probability and the true label is calculated to obtain the classification loss. The total loss function is constructed by combining the physical constraint loss, the inverse diffusion reconstruction loss, and the classification loss. The deep learning framework is pre-trained using the total loss function to obtain pre-trained model parameters; Solidify the parameters of the pre-trained model to generate a pre-trained physical information diffusion model; The vibration signal under the new working condition is received in real time and identified as online incremental data. Wavelet threshold denoising and normalization operations are performed on the online incremental data to generate preprocessed online incremental training and validation sets. Freeze the bottom-level parameters of the feature extraction module in the pre-trained physical information diffusion model, and adjust the top-level parameters of the feature extraction module, the diffusion and inverse diffusion module, and the classification module to obtain a fine-tuning parameter set; By fusing the online incremental training set with the physical cue vector, and performing incremental learning operations through the fine-tuned parameter set, the inverse diffusion reconstruction loss, frequency bias loss, dynamic bias loss, and classification loss of the online incremental data are calculated. An incremental loss function is constructed by combining the inverse diffusion reconstruction loss, the frequency bias loss, the dynamic bias loss, and the classification loss of the incremental data; Based on the reverse diffusion process of the pre-trained physical information diffusion model, random noise is input to generate virtual historical samples that are consistent with the historical fault type characteristics of the original vibration signal. The virtual historical samples are mixed with the online incremental training set at a preset ratio to form a hybrid training set; The incremental loss function is used to perform incremental training on the hybrid training set to obtain incrementally updated model parameters; The classification accuracy and loss value of the incrementally updated model parameters are calculated using the online validation set. When the classification accuracy and the loss value meet the preset conditions, the incremental training operation on the mixed training set is stopped, and the optimal incremental model parameters are generated.

[0052] The optimal incremental model parameters are fixed, and an incremental diagnostic model is generated based on the optimal incremental model parameters; Receive the current vibration signal of the target wind turbine bearing, perform wavelet threshold denoising and normalization operations on the current vibration signal, and generate a preprocessed current vibration signal; The preprocessed current vibration signal is input into the feature extraction module of the incremental diagnostic model to generate the current high-dimensional feature vector. By fusing the current high-dimensional feature vector with the physical cue vector, a reconstruction operation is performed through the diffusion-inverse diffusion module of the incremental diagnostic model to obtain the reconstructed current feature vector; The reconstructed current feature vector is input into the classification module of the incremental diagnostic model, and the failure type prediction result of the target wind turbine bearing and the confidence level of the failure type prediction result are output. If the confidence level is lower than a preset threshold, then the reconstructed current feature vector is subjected to secondary optimization through the inverse diffusion process of the incremental diagnostic model to obtain a refined feature vector, and the physical constraint error between the refined feature vector and the current feature vector is calculated. When the physical constraint error is less than a preset threshold, the refined feature vector is used to generate a corrected diagnostic result; When the physical constraint error is not less than a preset threshold, an early warning signal is generated, and the fault type prediction result is associated with and stored with real-time operating condition data to generate a fault condition association database.

[0053] Optionally, before the step of determining the target loss function based on the physical cue vector and the original vibration signal, the method further includes: The original vibration signal is subjected to wavelet threshold denoising and normalization operations to generate a preprocessed original vibration signal.

[0054] In this embodiment of the invention, before determining the target loss function based on the physical cue vector and the original vibration signal, wavelet threshold denoising and normalization operations can be performed on the original vibration signal to generate a preprocessed original vibration signal. This removes environmental noise interference from the original vibration signal and unifies the data scale, enabling subsequent model training and diagnosis processes to obtain cleaner, more scale-consistent, and high-quality input signals.

[0055] Wavelet thresholding denoising refers to the process of using wavelet transform to decompose vibration signals into wavelet coefficients of different scales, retaining or shrinking wavelet coefficients larger than a set threshold, setting wavelet coefficients smaller than the threshold to zero or performing soft / hard thresholding and then performing inverse transform, thereby effectively removing high-frequency noise while preserving fault impact characteristics as much as possible.

[0056] Normalization is a linear transformation operation that maps the denoised vibration signal values ​​to a uniform interval (such as [0,1] or [-1,1]). It usually uses the minimum-maximum normalization formula to make signals under different working conditions, different acquisition devices, or different amplitude ranges comparable and have a consistent numerical distribution, which facilitates the stable convergence of deep learning models.

[0057] The preprocessed original vibration signal refers to the vibration signal after wavelet threshold denoising and normalization. This signal has significantly reduced the impact of noise and achieved numerical scale standardization, and can be used as a reliable input for subsequent physical cue vector fusion, feature extraction and loss function calculation.

[0058] In this embodiment of the invention, by performing wavelet threshold denoising and normalization operations on the original vibration signal before determining the target loss function, a preprocessed original vibration signal is generated. This provides the physical information diffusion model with high-quality input data that has lower noise and uniform scale, thereby improving the model stability and cross-condition generation and diagnosis performance during the pre-training stage.

[0059] Optionally, the step of determining the target loss function based on the physical cue vector and the original vibration signal includes: Construct a deep learning framework that includes a feature extraction module composed of a convolutional neural network, a diffusion and inverse diffusion module based on a diffusion probability model architecture, and a classification module. The preprocessed original vibration signal is input into the feature extraction module to generate a high-dimensional feature vector. By fusing the high-dimensional feature vector with the physical cue vector, and performing forward diffusion noise addition and reverse diffusion denoising operations through the diffusion and de-diffusion modules, the reconstructed original feature vector is obtained. A physical correlation model between the fault impact load and the final vibration displacement is established based on the physical equation of vibration response using the equivalent mass, the damping coefficient, and the equivalent stiffness. Substitute the original feature vector into the physical association model, calculate the error between the original feature vector and the physical cue vector, and generate the physical constraint loss. Calculate the mean square error between the original feature vector and the high-dimensional feature vector to obtain the inverse diffusion reconstruction loss; The classification module determines the predicted probability of the fault type to which the original feature vector belongs, and the true label of the original vibration signal corresponding to the original feature vector; The original feature vector is input into the classification module, and the cross-entropy loss between the predicted probability and the true label is calculated to obtain the classification loss. The total loss function is constructed by combining the physical constraint loss, the inverse diffusion reconstruction loss, and the classification loss.

[0060] In this embodiment of the invention, a deep learning framework can be constructed that includes a feature extraction module composed of a convolutional neural network, a diffusion and anti-diffusion module based on a diffusion probability model architecture, and a classification module, in order to build an end-to-end physical information-guided generative diagnostic network architecture, providing structured support for subsequent feature extraction, diffusion process, and loss calculation.

[0061] A deep learning framework refers to a neural network structure composed of multiple functional modules, used to process vibration signals and perform generation and classification tasks.

[0062] The feature extraction module of a convolutional neural network refers to the network part that automatically extracts multi-scale spatiotemporal features from time-series vibration signals using operations such as convolutional layers and pooling layers. It typically includes multiple one-dimensional or two-dimensional convolutions to capture local patterns and global context.

[0063] The diffusion-inverse diffusion module based on the diffusion probability model architecture refers to a module that follows the principle of the denoising diffusion probability model (DDPM). It models complex data distributions through a forward diffusion process that gradually adds noise and a reverse process that gradually denoises, and incorporates physical constraints in the process.

[0064] The classification module refers to the network tail consisting of fully connected layers or softmax layers, which is used to map the reconstructed or extracted features to the fault category probability distribution.

[0065] In this embodiment of the invention, the preprocessed original vibration signal can be input into the feature extraction module to generate a high-dimensional feature vector, so as to extract an abstract high-level representation from the original time-domain signal, which is convenient for subsequent fusion and diffusion operations with physical cues.

[0066] High-dimensional feature vectors refer to fixed-dimensional dense vectors obtained after processing by the feature extraction module. They contain deep semantic information of vibration signals, such as impact modes, frequency components, and amplitude distribution.

[0067] In this embodiment of the invention, the high-dimensional feature vector and the physical cue vector can be fused, and forward diffusion noise addition and reverse diffusion denoising operations can be performed through the diffusion and reverse diffusion modules to obtain the reconstructed original feature vector. This introduces physical prior guidance during the diffusion process, so that the features generated by the model conform to the physical laws of bearings and achieve high-quality reconstruction.

[0068] Fusion refers to combining high-dimensional feature vectors with physical cue vectors through methods such as concatenation, addition, or cross-attention to form conditional inputs.

[0069] Forward diffusion noise addition refers to the process of gradually adding Gaussian noise to the original feature vector according to a preset schedule until it approaches a pure noise distribution.

[0070] Inverse diffusion denoising refers to an iterative process in which the model starts from a noisy state and gradually predicts and removes noise to reconstruct the original feature vector.

[0071] The reconstructed original feature vector refers to the reconstruction result that is close to the original high-dimensional feature vector after a complete reverse diffusion process.

[0072] In this embodiment of the invention, the equivalent mass, the damping coefficient, and the equivalent stiffness can be used to establish a physical correlation model between the fault impact load and the final vibration displacement based on the vibration response physical equation, so as to transform the dynamic principle into a computable constraint form and to quantify the physical rationality of the reconstruction features.

[0073] The vibration response physical equation refers to the bearing vibration dynamics model described by a second-order differential equation, typically in the form of: ,in, The equivalent mass of the rotating components of the bearing. The damping coefficient is... The equivalent stiffness varies over time (reflecting bearing wear, cracks, and other fault conditions). For vibration displacement, This refers to the impact load caused by the fault (such as the pulse load when the inner race fails). This is environmental noise.

[0074] The physical correlation model refers to the mapping relationship established using the above equations from the fault impact load to the vibration displacement response, which is used for subsequent error calculation.

[0075] In this embodiment of the invention, the reconstructed original feature vector can be substituted into the physical association model to calculate the error between the reconstructed original feature vector and the physical cue vector, thereby generating a physical constraint loss. This forces the features learned by the model to satisfy the bearing dynamics prior, thus avoiding the generation of pseudo-features that violate physical laws.

[0076] Physical constraint loss refers to a quantitative indicator calculated through a physical correlation model. It typically uses mean squared error or a custom distance metric to reflect the degree of deviation between the reconstructed features and the physical cue.

[0077] In this embodiment of the invention, the mean square error between the reconstructed original feature vector and the high-dimensional feature vector can be calculated to obtain the inverse diffusion reconstruction loss, so as to measure the fidelity of the diffusion model in data reconstruction and ensure that the inverse diffusion process can effectively restore the original feature distribution.

[0078] Inverse diffusion reconstruction loss refers to the reconstruction error term calculated based on mean squared error (MSE), which is used to optimize the denoising capability of the diffusion model.

[0079] In this embodiment of the invention, the predicted probability of the fault type to which the reconstructed original feature vector belongs, and the true label of the original vibration signal tag corresponding to the reconstructed original feature vector can be determined by the classification module, so as to obtain the model's prediction output and supervision signal on the fault classification task.

[0080] Predicted probability refers to the softmax normalized probability vector of various fault types output by the classification module.

[0081] Real labels refer to the fault category unique-heat codes or category indexes pre-labeled on the original vibration signals, which are used for supervised learning.

[0082] In this embodiment of the invention, the reconstructed original feature vector can be input into the classification module to calculate the cross-entropy loss between the predicted probability and the true label, thereby obtaining the classification loss and guiding the model to achieve accurate fault type differentiation based on the reconstructed features.

[0083] Cross-entropy loss is a commonly used classification loss function that measures the difference between the predicted probability distribution and the true label distribution.

[0084] In this embodiment of the invention, the total loss function can be constructed by combining the physical constraint loss, the inverse diffusion reconstruction loss, and the classification loss to form a comprehensive loss for multi-objective optimization, which is used to simultaneously constrain physical consistency, data fidelity, and classification performance.

[0085] The total loss function refers to the overall optimization objective obtained by summing or dynamically balancing the above three types of losses according to their weights. It typically takes the form of:

[0086] in, This is the reconstruction loss due to the reverse diffusion process. For physical constraint loss, Classify the losses based on the faults.

[0087] In this embodiment of the invention, a deep learning framework is constructed, the preprocessed original vibration signal is input into the feature extraction module to generate a high-dimensional feature vector, the high-dimensional feature vector and the physical cue vector are fused to perform diffusion and inverse diffusion operations to obtain the reconstructed original feature vector, a physical association model is established and the physical constraint loss is calculated, the inverse diffusion reconstruction loss is calculated, the classification loss is calculated through the classification module, and the total loss function is constructed by combining the three types of losses. In this way, a target loss function that takes into account physical consistency, data reconstruction quality and fault classification accuracy can be determined, providing a reliable multi-task optimization objective for the pre-training of the physical information diffusion model.

[0088] Optionally, the steps of pre-training a physical information diffusion model with cross-condition basic generation and diagnosis capabilities based on the target loss function include: The deep learning framework is pre-trained using the total loss function to obtain pre-trained model parameters; The parameters of the pre-trained model are solidified to generate a pre-trained physical information diffusion model.

[0089] In this embodiment of the invention, the total loss function can be used to perform pre-training operations on the deep learning framework to obtain pre-trained model parameters. Through a multi-objective optimization process, the model can simultaneously learn the ability to reconstruct data distribution, the consistency of physical laws, and the accuracy of fault classification, thereby establishing a model foundation with preliminary cross-working condition generation and diagnosis capabilities.

[0090] Pre-training refers to the end-to-end parameter optimization training process of a deep learning framework using historical multi-condition datasets and a total loss function. It typically employs gradient descent algorithms such as the Adam optimizer, updating network weights by minimizing the total loss function through multiple iterations.

[0091] Pre-trained model parameters refer to the set of all trainable parameters of a deep learning framework obtained after pre-training, including the weights and biases of the feature extraction module, the diffusion and inverse diffusion module, and the classification module.

[0092] In this embodiment of the invention, the pre-trained model parameters can be fixed to generate a pre-trained physical information diffusion model, so as to fix and save the optimized parameters to form a stable and reusable basic model instance, which can be used as a starting point for subsequent deployment and incremental updates.

[0093] Solidifying the pre-trained model parameters means saving the trained model parameters to a file or memory and stopping any further random initialization or significant modification of them, ensuring that the model retains the cross-condition knowledge acquired during the pre-training phase in subsequent use.

[0094] The pre-trained physical information diffusion model refers to a complete generative diagnostic model instantiated based on the parameters of the pre-trained model. This model incorporates physical cue constraints and can generate reasonable vibration characteristics from noise under various working conditions and initially achieve fault type prediction.

[0095] In this embodiment of the invention, a pre-trained model parameter is obtained by performing a pre-training operation on a deep learning framework using a total loss function, and the pre-trained model parameter is then solidified to generate a pre-trained physical information diffusion model. This enables the acquisition of a basic physical information diffusion model with stable generation capability and preliminary diagnostic performance under multiple working conditions, providing a reliable knowledge starting point for subsequent online incremental adaptation.

[0096] Optionally, the step of performing incremental training on the physical information diffusion model based on the online incremental data to generate optimal incremental model parameters includes: Wavelet threshold denoising and normalization operations are performed on the online incremental data to generate preprocessed online incremental training and validation sets. Freeze the bottom-level parameters of the feature extraction module in the pre-trained physical information diffusion model, and adjust the top-level parameters of the feature extraction module, the diffusion and inverse diffusion module, and the classification module to obtain a fine-tuning parameter set; By fusing the online incremental training set with the physical cue vector, and performing incremental learning operations through the fine-tuned parameter set, the inverse diffusion reconstruction loss, frequency bias loss, dynamic bias loss, and classification loss of the online incremental data are calculated. An incremental loss function is constructed by combining the inverse diffusion reconstruction loss, the frequency bias loss, the dynamic bias loss, and the classification loss of the incremental data; Based on the reverse diffusion process of the pre-trained physical information diffusion model, random noise is input to generate virtual historical samples that are consistent with the historical fault type characteristics of the original vibration signal. The virtual historical samples are mixed with the online incremental training set at a preset ratio to form a hybrid training set; The incremental loss function is used to perform incremental training on the hybrid training set to obtain incrementally updated model parameters; The classification accuracy and loss value of the incrementally updated model parameters are calculated using the online validation set. When the classification accuracy and the loss value meet the preset conditions, the incremental training operation on the mixed training set is stopped, and the optimal incremental model parameters are generated.

[0097] In this embodiment of the invention, wavelet threshold denoising and normalization operations can be performed on the online incremental data to generate preprocessed online incremental training and validation sets. This removes noise interference from the vibration signals under new operating conditions and unifies the data scale, ensuring that the incremental learning process uses high-quality, consistent input data.

[0098] The preprocessed online incremental training set refers to the set of vibration signals under new working conditions after wavelet threshold denoising and normalization, which are used as training samples for actual model parameter updates.

[0099] The online validation set refers to an independent subset of preprocessed online incremental data, used for real-time monitoring of model performance and early stopping detection during incremental training.

[0100] In this embodiment of the invention, the underlying parameters of the feature extraction module in the pre-trained physical information diffusion model can be frozen, and the top-level parameters of the feature extraction module, the diffusion and anti-diffusion module, and the classification module can be adjusted to obtain a fine-tuning parameter set. This protects the general underlying feature representations learned during the pre-training stage, while only making adaptive adjustments to the high-level semantics and generation logic for new working conditions.

[0101] Low-level parameters refer to the low-level network parameters near the input end in the feature extraction module. They are usually responsible for learning low-level features such as general edges, textures, and local patterns. These parameters have strong cross-condition transfer capabilities after pre-training, so they are frozen in incremental learning to prevent catastrophic forgetting.

[0102] Top-level parameters refer to the high-level parameters in the network that are close to the output end. These include the high-level convolutional layers of the feature extraction module, the noise prediction network of the diffusion and inverse diffusion module, and the weights and biases of the classification module. These parameters are more task-specific, so they can be adjusted during incremental training to adapt to new working conditions.

[0103] The fine-tuning parameter set refers to the set of parameters that are allowed to be updated during incremental learning, and is usually limited to the weights and biases of the top convolutional layers of the network, the noise prediction network of the diffusion module, and the classification head.

[0104] In this embodiment of the invention, the online incremental training set and the physical cue vector can be fused together, and incremental learning operations can be performed through the fine-tuning parameter set to calculate the inverse diffusion reconstruction loss, frequency bias loss, dynamic bias loss, and classification loss of the online incremental data, so as to simultaneously evaluate data fidelity, frequency domain consistency, dynamic physical rationality, and classification accuracy in each forward propagation.

[0105] Inverse diffusion reconstruction loss refers to the incremental data reconstruction error calculated based on the mean square error, which is used to optimize the denoising capability of the diffusion model under new operating conditions.

[0106] Frequency deviation loss refers to the loss term obtained by comparing the spectral components of the reconstructed signal or feature with the fault theoretical characteristic frequencies contained in the physical cue vector, and is used to ensure that the generated signal meets the theoretical expectations at key frequency points.

[0107] Dynamic deviation loss refers to the dynamic response error calculated after substituting the reconstructed features into the physical equation of vibration response. It is used to constrain the model output to meet the physical dynamic laws of bearing equivalent mass, damping, and time-varying stiffness.

[0108] Classification loss refers to the cross-entropy loss calculated on incremental data under new operating conditions, which is used to maintain or improve the model's ability to distinguish faults under new data.

[0109] In this embodiment of the invention, an incremental loss function can be constructed by combining the inverse diffusion reconstruction loss, the frequency deviation loss, the dynamic deviation loss, and the classification loss of the incremental data to form a comprehensive optimization objective for the new working condition, while taking into account the generation quality, physical consistency, and diagnostic performance.

[0110] The incremental loss function is the overall loss obtained by weighting or adaptively balancing the above four types of losses, and is used to guide parameter updates during incremental training.

[0111] In this embodiment of the invention, random noise can be input into the reverse diffusion process of the pre-trained physical information diffusion model to generate virtual historical samples that are consistent with the historical fault type characteristics of the original vibration signal. This allows the generation capability of the pre-trained model to synthesize high-quality virtual old data, thereby reproducing historical knowledge in incremental training and suppressing catastrophic forgetting.

[0112] Virtual historical samples refer to synthetic vibration features or signals generated from pure noise through a reverse diffusion process. These samples are highly consistent with real samples in historical datasets in terms of fault type characterization, frequency components, and physical properties.

[0113] Historical fault type characteristics refer to the typical vibration modes, impact characteristics, and spectral distribution of normal states and various faults (such as inner race, outer race, rolling element, and cage faults) learned in the historical data of the pre-training stage.

[0114] In this embodiment of the invention, the virtual historical samples can be mixed with the online incremental training set at a preset ratio to form a hybrid training set, so as to learn new working condition features and historical knowledge simultaneously in each round of incremental training, thereby achieving a balance between knowledge retention and adaptation to new scenarios.

[0115] A hybrid training set refers to a training dataset composed of virtual historical samples and preprocessed online incremental training sets in a certain ratio (such as 1:1 or dynamically adjusted according to the degree of forgetting).

[0116] In this embodiment of the invention, the incremental loss function can be used to perform incremental training on the mixed training set to obtain incrementally updated model parameters, so as to gradually optimize and fine-tune the parameter set through gradient descent, so that the model can obtain good performance on both new and old data.

[0117] Incremental update of model parameters refers to the current set of model parameters obtained after several rounds of incremental training, representing the latest state of the model on the current batch of incremental data.

[0118] In this embodiment of the invention, the classification accuracy and loss value of the incrementally updated model parameters can be calculated by combining the online validation set, so as to evaluate the generalization performance and convergence of the current parameters on unseen new working condition validation data in real time.

[0119] Classification accuracy refers to the proportion of times a model correctly predicts the type of fault on an online validation set, usually expressed as a percentage.

[0120] The loss value refers to the average value of the incremental loss function calculated on the online validation set, which is used to determine the trend of overfitting or underfitting.

[0121] In this embodiment of the invention, when the classification accuracy and the loss value meet preset conditions, the incremental training operation on the mixed training set can be stopped, and the optimal incremental model parameters can be generated. This early stopping mechanism avoids overfitting and selects the parameter state with the best overall performance under the new working conditions.

[0122] Preset conditions refer to pre-defined stopping criteria, such as classification accuracy no longer improving after a certain number of iterations, loss value starting to rise, or reaching the maximum number of iterations.

[0123] The optimal incremental model parameters refer to the set of optimal model parameters that meet the preset conditions, and are used to generate the incremental diagnostic model in the future.

[0124] In this embodiment of the invention, by preprocessing online incremental data, freezing the underlying parameters and fine-tuning the top-level parameters, integrating physical cues to perform incremental learning and calculate multiple types of losses, constructing an incremental loss function, generating virtual historical samples, performing mixed training, performing incremental training, and generating optimal incremental model parameters based on early stopping judgment on the validation set, the physical information diffusion model can quickly and stably adapt to new working condition data and obtain optimal parameter updates suitable for real-time diagnosis without forgetting historical knowledge.

[0125] Optionally, the step of controlling the incremental diagnostic model to generate detection results for the target wind turbine bearing includes: Perform wavelet threshold denoising and normalization operations on the current vibration signal to generate a preprocessed current vibration signal; The preprocessed current vibration signal is input into the feature extraction module of the incremental diagnostic model to generate the current high-dimensional feature vector. By fusing the current high-dimensional feature vector with the physical cue vector, a reconstruction operation is performed through the diffusion-inverse diffusion module of the incremental diagnostic model to obtain the reconstructed current feature vector; The reconstructed current feature vector is input into the classification module of the incremental diagnostic model, and the failure type prediction result of the target wind turbine bearing and the confidence level of the failure type prediction result are output. If the confidence level is lower than a preset threshold, then the reconstructed current feature vector is subjected to secondary optimization through the inverse diffusion process of the incremental diagnostic model to obtain a refined feature vector, and the physical constraint error of the refined feature vector and the physical correlation model is calculated. When the physical constraint error is less than a preset threshold, the refined feature vector is used to generate a corrected fault type prediction result.

[0126] In this embodiment of the invention, wavelet threshold denoising and normalization operations can be performed on the current vibration signal to generate a preprocessed current vibration signal, thereby removing environmental noise from the real-time acquired signal and unifying the numerical scale, ensuring consistent data quality and minimizing noise interference in the input incremental diagnostic model.

[0127] The current vibration signal refers to the raw time-domain vibration data of the target wind turbine bearing collected in real time by the sensor at the current moment.

[0128] The preprocessed current vibration signal refers to the current vibration signal after wavelet threshold denoising and normalization. This signal has reduced the impact of noise and achieved scale standardization, and can be used as a reliable input for model inference.

[0129] In this embodiment of the invention, the preprocessed current vibration signal can be input into the feature extraction module of the incremental diagnostic model to generate a current high-dimensional feature vector, so as to extract an abstract high-level feature representation from the real-time vibration signal, which is convenient for subsequent physical fusion and reconstruction processes.

[0130] The current high-dimensional feature vector refers to a fixed-dimensional dense vector obtained after processing by the feature extraction module of the incremental diagnostic model. It contains the deep spatiotemporal features of the current vibration signal, such as impact mode, frequency distribution and amplitude characteristics.

[0131] In this embodiment of the invention, the current high-dimensional feature vector and the physical cue vector can be fused, and a reconstruction operation can be performed through the diffusion and inverse diffusion module of the incremental diagnostic model to obtain the reconstructed current feature vector. This allows the diffusion model to reconstruct the current feature representation that conforms to the physical laws of bearings from the noise using physical prior guidance.

[0132] Fusion refers to combining the current high-dimensional feature vector with the physical cue vector through methods such as concatenation, addition, or cross-attention to form a conditional input.

[0133] The reconstruction operation refers to the reverse denoising process performed by the diffusion and inverse diffusion module of the incremental diagnostic model during the inference phase, which gradually recovers high-quality feature vectors from noise or the initial state.

[0134] The reconstructed current feature vector refers to the reconstruction result of the current high-dimensional feature vector obtained after a complete reverse diffusion reconstruction process. This vector has incorporated physical constraints and has a high fidelity.

[0135] In this embodiment of the invention, the reconstructed current feature vector can be input into the classification module of the incremental diagnostic model to output the fault type prediction result of the target wind turbine bearing and the confidence level of the fault type prediction result, so as to directly give the fault diagnosis conclusion and its reliability assessment based on the reconstructed features.

[0136] The fault type prediction result refers to the specific fault category judgment output by the classification module, such as normal state, inner ring fault, outer ring fault, rolling element fault, cage fault, or compound fault.

[0137] Confidence level refers to the softmax probability value or normalized confidence score of the classification module for predicting the fault type. It is used to quantify the reliability of the diagnostic results and is usually expressed as a percentage or in the range [0,1].

[0138] In this embodiment of the invention, when the confidence level is lower than a preset threshold, a secondary optimization is performed on the reconstructed current feature vector through the inverse diffusion process of the incremental diagnostic model to obtain a refined feature vector. The physical constraint error of the refined feature vector and the physical correlation model is calculated so that when the initial diagnostic confidence level is insufficient, the denoising capability of the diffusion model is further utilized to refine the features, and the rationality is verified by the physical model.

[0139] Secondary optimization refers to iterating several additional steps or adjusting the denoising strategy during the reverse diffusion process to perform more refined noise removal and physical constraint enhancement on the reconstructed features.

[0140] A refined feature vector refers to the current feature vector after secondary optimization, which has lower noise and stronger physical consistency.

[0141] Physical constraint error refers to the quantitative deviation calculated after substituting the refined feature vector into the physical correlation model. It is usually expressed as mean square error or a custom distance metric, reflecting the degree of conformity between the feature vector and the bearing dynamics law.

[0142] In this embodiment of the invention, when the physical constraint error is less than a preset threshold, the refined feature vector is used to generate a corrected fault type prediction result. In this way, if the physical verification passes, the more reliable refined features are used to reclassify the fault type and output a diagnostic conclusion with higher confidence.

[0143] The corrected fault type prediction result refers to the fault category judgment and its corresponding confidence level obtained by recalculating through the classification module based on the refined feature vector. This result is more physically reasonable and accurate in low-confidence scenarios.

[0144] In this embodiment of the invention, the current vibration signal is preprocessed, a high-dimensional feature vector is generated by the input feature extraction module, a reconstruction operation is performed by fusing physical cues to obtain the reconstructed current feature vector, a fault type prediction result and confidence level are output by the classification module, a secondary optimization is performed at low confidence levels to obtain a refined feature vector and calculate the physical constraint error, and when the error is less than a threshold, the refined features are used to generate a corrected fault type prediction result. This enables physical consistency verification and feature refinement of low confidence results during real-time diagnosis, thereby improving the reliability and accuracy of wind turbine bearing fault detection.

[0145] Optionally, it also includes: When the physical constraint error is not less than a preset threshold, an early warning signal is generated, and the fault type prediction result is associated with and stored with real-time operating condition data to generate a fault condition association database.

[0146] In this embodiment of the invention, when the physical constraint error is not less than a preset threshold, an early warning signal can be generated, and the fault type prediction result can be associated with and stored with real-time operating condition data to generate a fault condition association database. This allows for timely notification of maintenance personnel to perform manual intervention in the event of a physical consistency verification failure. At the same time, it accumulates the association data between the diagnostic process and actual operating conditions, providing long-term support for subsequent model optimization and fault mode analysis.

[0147] The physical constraint error is not less than the preset threshold, which means that the deviation value calculated after substituting the refined feature vector into the physical correlation model exceeds the preset tolerance limit. This indicates that the physical rationality of the current diagnostic result is insufficient and there may be model uncertainty or data anomaly.

[0148] Warning signals refer to alarm prompts issued by the system, usually in the form of sound and light, log records, remote notifications or push notifications from the operation and maintenance platform, used to remind operation and maintenance personnel to conduct on-site inspections or further verifications of the bearing status.

[0149] The fault type prediction result refers to the fault category judgment and its confidence level output by the incremental diagnostic model at the current inference stage, which is still retained as reference information even when the confidence level is low or the physical verification fails.

[0150] Real-time operating data refers to the set of parameters of the wind turbine bearing operating environment at the current moment, including wind speed, load, rotational speed, temperature, humidity, etc., which directly correspond to the actual operating status at the vibration signal acquisition time point.

[0151] Associated storage refers to the operation of binding and saving fault type prediction results, confidence level, physical constraint error, early warning trigger flags and corresponding real-time operating data by timestamp or unique identifier.

[0152] The fault condition association database refers to the structured data storage formed by the accumulation of multiple diagnostic events. It is used to record the correspondence between fault prediction and operating condition parameters, and supports subsequent data mining, fault mode recognition, and supplementation of new operating condition knowledge.

[0153] In this embodiment of the invention, a fault condition association database is generated by generating an early warning signal when the physical constraint error is not less than a preset threshold and associating the fault type prediction result with real-time operating condition data. This enables timely early warnings to be issued and fault-operating condition association data to be continuously accumulated in low-reliability diagnostic scenarios, providing a data foundation for the long-term reliable operation and maintenance of wind turbine bearings and future model iterations.

[0154] To enable those skilled in the art to better understand the embodiments of the present invention, an example is used below to illustrate the embodiments of the present invention.

[0155] In practical applications, wind turbine bearings, as a crucial mechanical component of wind turbine units, directly affect the overall performance, safety, and power generation revenue of the wind farm. Due to the complex operating environment of wind turbines, factors such as wind speed, load, temperature, air humidity, and rotational speed continuously change, resulting in significant multi-condition characteristics during actual operation. Traditional wind turbine bearing fault diagnosis methods are mostly based on data collected under single laboratory conditions or stable loading conditions, making them ill-suited to the frequently changing, non-stationary, multi-modal, and multi-scale characteristics of real wind farms. Existing models generally employ offline training schemes, where network parameters are fixed after a one-time training phase. When new operating conditions are introduced, the entire model needs to be retrained. This is not only extremely time-consuming but also suffers from the "catastrophic forgetting" problem, where learning new operating conditions can impair the diagnostic capabilities of the original model, leading to a decline in overall performance.

[0156] Online incremental learning, as an intelligent method that progressively learns data distribution and continuously adapts to changing scenarios, has shown great potential in the field of mechanical equipment monitoring. However, traditional incremental learning methods mostly focus on structural modifications to classification models or sample replay mechanisms. These methods cannot solve the inherent distribution drift problem of multi-condition vibration data of wind turbine bearings, nor can they effectively expand the ability to express differences between fault modes. Physical mechanism diffusion models, as a rapidly developing generative framework in recent years, can reconstruct complex distribution signals with high quality through progressive denoising and possesses good multi-modal generation capabilities, stability, and controllability. Existing physical information DDPM (Distributed Derivative Processing Model) is mostly still in the offline training stage, unable to cope with the constantly emerging new operating conditions, loads, and fault types during long-term wind turbine operation, and lacks the ability for adaptive updates and continuous learning in online incremental learning.

[0157] In view of this, this invention proposes an online incremental learning diagnostic method for wind turbine bearing physical information DDPM oriented to multiple operating conditions. This method, while ensuring physical consistency, adapts to dynamic changes in multiple operating conditions, supports learning of new scenarios, avoids catastrophic forgetting, and is a generative incremental learning method for wind turbine bearings that can update diagnostic capabilities in real time. It establishes cross-operating condition generation capabilities through a diffusion model guided by physical information, and combines this with an incremental learning strategy to achieve continuous adaptation to new scenarios in the diagnostic model update, providing technical support for the intelligent, long-term, and reliable operation of wind turbine bearings. In specific implementation, firstly, a multi-operating-condition physical information description system for wind turbine bearings is constructed. Physical cues are generated using information such as wind turbine dynamics models, structural parameters, speed ranges, and fault characteristic frequencies to constrain the generation process of the diffusion model. Next, a physical information-guided diffusion model framework is designed to improve the interpretability and physical rationality of the generated signals. Secondly, an offline pre-training strategy based on physical cues is proposed, obtaining basic generation capabilities through the fusion learning of multi-operating-condition samples. Furthermore, an online incremental learning strategy is proposed, which enables adaptive learning under new operating conditions through real-time wind farm data input, allowing for rapid adaptation to new operating states, fault modes, or environmental changes. Finally, real-time fault diagnosis and result feedback under multiple operating conditions ensure the long-term stable operation of the diagnostic system. Specific steps include: Step 1: Construct a physical information description system for wind turbine bearings under multiple operating conditions.

[0158] 1) Dynamic model parameters: Based on bearing dynamics theory and considering influencing factors such as load, temperature, and speed under multiple operating conditions, the physical equations for the vibration response of wind turbine bearings are established. The vibration signals of wind turbine bearings mainly consist of normal operating vibration, fault impact vibration, and environmental noise. Their dynamic equations can be expressed as follows:

[0159] in, The equivalent mass of the rotating components of the bearing. The damping coefficient is... The equivalent stiffness varies over time (reflecting bearing wear, cracks, and other fault conditions). For vibration displacement, This refers to the impact load caused by the fault (such as the pulse load when the inner race fails). This equation represents environmental noise. It clarifies the physical relationship between fault characteristics and vibration signals, providing a theoretical basis for constructing physical constraints for subsequent models.

[0160] 2) Multi-condition Fault Data Acquisition and Preprocessing. Data on various fault types of the wind turbine bearings under different operating conditions (wind speed: 3-25 m / s, load: 0.5-2.0 times rated load, temperature: -40℃-80℃) were collected, including normal state, inner ring fault, outer ring fault, rolling element fault, and cage fault. Each fault type corresponds to a different fault severity (crack width: 0.1-1.0 mm). The collected vibration signals were preprocessed: First, wavelet threshold denoising was used to remove environmental noise, using the following formula:

[0161] in, These are wavelet coefficients. These are the denoised wavelet coefficients. The threshold is then used. The denoised signal is subsequently normalized, mapping the data to the [0,1] interval, as shown in the formula:

[0162] in, The original signal value, , These are the minimum and maximum values ​​of the signal, respectively. These are the normalized signal values. The preprocessed dataset is divided into a historical dataset (containing various basic operating conditions and fault types) and an online incremental dataset (data on subsequently added operating conditions or fault types).

[0163] Structural parameters, including core structural parameters such as bearing inner diameter, outer diameter, number of rolling elements, and rolling element diameter, are used to derive the characteristic frequencies of different fault types.

[0164] Speed ​​range: Covering the entire speed range of the fan operation (500-2000 r / min), the real-time fault characteristic frequency is calculated based on the speed, using the following formula:

[0165]

[0166] in, For bearing speed, For the number of rolling elements, The diameter of the rolling element, The bearing pitch circle diameter, Contact angle, , These are the fault characteristic frequencies of the inner and outer rings, respectively.

[0167] Fault feature mapping: Establishing a physical mapping relationship between fault types (inner ring fault, outer ring fault, rolling element fault, cage fault) and the frequency components and amplitude characteristics of vibration signals.

[0168] The aforementioned physical information is transformed into numerical physical cue vectors that can be input into the model. These vectors are used to constrain the diffusion and de-diffusion processes of the subsequent diffusion model, ensuring that the feature data generated by the model conforms to the physical laws governing bearing operation.

[0169] Step 2: Design a diffusion model framework guided by physical information.

[0170] The overall model framework includes a feature extraction module, a diffusion-reverse diffusion module, a physical constraint module, and a classification module. The feature extraction module uses a convolutional neural network (CNN) to convert the preprocessed vibration signal into a high-dimensional feature vector; the diffusion-reverse diffusion module follows the core logic of DDPM, gradually adding Gaussian noise to the original feature vector through a T-step diffusion process.

[0171] The diffusion process formula is:

[0172] in, Let be the eigenvector after the t-th diffusion step. The noise figure at step t (satisfying) ), Indicates a Gaussian distribution. It is an identity matrix.

[0173] The inverse diffusion process reconstructs the original features step by step from random noise, as shown in the formula:

[0174] in, The mean value of the reverse diffusion process (fitted by a neural network). The variance (fixed) (Or learn through the network). The physical constraint module transforms the dynamic equations constructed in step one into loss terms, and ensures that the features learned by the model have physical rationality by constraining the feature vectors reconstructed during the backdiffusion process to satisfy the physical equations; the classification module uses a fully connected layer to map the feature vectors reconstructed by backdiffusion to the fault type label.

[0175] Step 3: Pre-train the physical information DDPM model.

[0176] Physical cue fusion integrates preprocessed multi-condition data with physical cue vectors to form a "data + physical cue" input pair. Each sample is associated with corresponding dynamic parameters, structural parameters, speed range, and fault characteristic frequencies.

[0177] Model pre-training utilizes pre-processed historical datasets to pre-train the model, with the optimization objective being the total loss function. Its expression is:

[0178] in, This is the reconstruction loss due to the reverse diffusion process. For physical constraint loss, Classify the losses based on the faults.

[0179] Calculated using mean squared error (MSE): ( (This refers to the original feature vector that is ultimately reconstructed).

[0180] The error between the vibration signal corresponding to the reconstructed feature and the theoretical output of the physical equation is obtained by calculating: ; Cross-entropy loss is used:

[0181] in, Number of fault types For real labels, (For model prediction probabilities) and The weighting coefficients for physical loss and classification loss are respectively (determined through cross-validation). , .

[0182] Model solidification. The pre-training process uses the Adam optimizer, with a learning rate set to... The iteration count is 500 rounds. When the classification accuracy of the validation set is stable at over 95%, the pre-training is stopped and the pre-trained model parameters are saved.

[0183] Step 4: Online incremental data processing and incremental prediction.

[0184] During wind turbine operation, vibration signals from new operating conditions (such as wind speed and temperature exceeding the historical dataset range) or new fault types (such as composite faults not included in historical data) are received in real time. The incremental data is processed according to the preprocessing steps (wavelet denoising and normalization) to obtain an online incremental training set. To verify the performance of the incremental model, an online validation set was constructed using a subset of incremental data. Considering the real-time requirements of online scenarios, incremental data is received using a sliding window approach, with a window size of 100 sampling points and a sliding step of 50 sampling points, to ensure that the model can quickly respond to the input of new data.

[0185] Physics-guided incremental fine-tuning: The underlying parameters of the feature extraction module in the pre-trained model are frozen, and only the parameters of the top-level feature extraction module, the diffusion-reverse diffusion module, the physics consistency discrimination module, and the classification module are fine-tuned to avoid interference with the underlying basic physical features; real-time physics cue vectors are fused with incremental data, and the model optimization objective is updated to the incremental loss function. :

[0186] in, , , , The losses are reconstruction loss, frequency bias loss, dynamics bias loss, and classification loss for incremental data, calculated in the same way as in the pre-training stage. An adaptive learning rate strategy is adopted, with an initial learning rate of [missing value]. The learning rate decays by 10% every 50 iterations, with a total of 200 iterations to ensure rapid model convergence. A forgetting suppression mechanism based on DDPM generation capability. To further improve the retention of historical bearing signal knowledge, a virtual historical sample library is constructed using the generation capability of DDPM: based on the inverse diffusion process of a pre-trained model, random noise is input to generate virtual samples consistent with the characteristics of historical fault types. The virtual sample generation process satisfies physical constraint loss. This ensures that the virtual samples match the physical characteristics of real historical samples. During incremental training, the virtual samples are... With online incremental training set Mixed at a 1:1 ratio, forming a hybrid bearing signal training set. During incremental training, the model simultaneously learns the features of new data and virtual historical samples. By replicating historical knowledge through virtual samples, it further suppresses catastrophic forgetting. The generation frequency of virtual samples is consistent with the frequency of incremental data reception. For each batch of incremental data (100 sampling points), a corresponding number of virtual historical samples are generated, ensuring the dynamic balance of the mixed training set. This approach avoids consuming additional storage resources (virtual samples are generated in real time and used immediately, requiring no long-term storage) while effectively preserving historical knowledge.

[0187] Real-time updates and parameter fixing of the incremental model. After each iteration of incremental training, an online validation set is calculated. The classification accuracy and loss value are compared. If the accuracy improves and the loss value decreases compared to the previous round, the current model parameters are saved as temporary update parameters. If the accuracy does not improve or the loss value increases after 10 consecutive iterations, the current incremental training is stopped, adopting an early stopping strategy to avoid overfitting and solidifying the current optimal model parameters. The model update adopts the "parameter incremental replacement" method, replacing only the parameters of the fine-tuning part (top-level feature extraction module, diffusion-anti-diffusion module, classification module), while the bottom feature extraction parameters remain unchanged, ensuring that the update process is fast and efficient, meeting the online real-time requirements (single round incremental training time controlled within 0.5 seconds, adapting to the response requirements of real-time diagnosis of wind turbine bearings).

[0188] Step 5: Real-time diagnosis and result feedback of faults under multiple operating conditions.

[0189] The solidified incremental model is deployed to the wind turbine control system, receiving pre-processed bearing vibration signals in real time. After inputting these signals into the model, high-dimensional features are extracted by the feature extraction module, reconstructed by the diffusion-reverse diffusion module, and physical constraints are applied for verification. Finally, the classification module outputs the fault type prediction results (normal state, inner ring fault, outer ring fault, rolling element fault, cage fault, and newly added compound fault, etc.) and confidence levels (0-100%). If the confidence level is below 85%, a secondary diagnostic process is triggered: the features are optimized again through the reverse diffusion process of DDPM, and the physical constraint error is recalculated. If the error is less than a set threshold (0.05), the corrected diagnostic result is output; if the error is still greater than the threshold, an early warning signal is issued, prompting maintenance personnel to conduct manual verification. Simultaneously, the diagnostic results are associated with real-time operating data (wind speed, load, temperature) and stored to form a fault-operating condition association database, providing data support for subsequent incremental learning.

[0190] It should be noted that, for the sake of simplicity, the method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions involved are not necessarily essential to the embodiments of the present invention.

[0191] Reference Figure 2 The diagram shows a structural block diagram of a wind turbine bearing testing device for wind power generation equipment provided in an embodiment of the present invention, which may specifically include the following modules: The bearing rotating component parameter acquisition module 201 is used to acquire the equivalent mass, damping coefficient, and time-varying equivalent stiffness of the wind turbine bearing. The fault theoretical characteristic frequency calculation module 202 is used to read the structural parameters of the wind turbine bearing and calculate the fault theoretical characteristic frequency of the target position of the wind turbine bearing under the target operating condition based on the structural parameters. The physical cue vector generation module 203 is used to concatenate the fault theoretical characteristic frequency, the equivalent mass, the damping coefficient, and the equivalent stiffness to generate a physical cue vector; The raw vibration signal acquisition module 204 is used to acquire raw vibration signals under different working conditions; The physical information diffusion model training module 205 is used to determine the target loss function based on the physical cue vector and the original vibration signal, and to pre-train a physical information diffusion model with cross-condition basic generation and diagnosis capabilities based on the target loss function. The online incremental data acquisition module 206 is used to receive new working condition vibration signals in real time and determine the new working condition vibration signals as online incremental data. The incremental diagnostic model generation module 207 is used to perform incremental training on the physical information diffusion model based on the online incremental data, generate optimal incremental model parameters, and generate an incremental diagnostic model based on the optimal incremental model parameters. The detection result generation module 208 is used to control the incremental diagnostic model to generate detection results for the target wind turbine bearing when the current vibration signal of the target wind turbine bearing is received.

[0192] As the device embodiment is basically similar to the method embodiment, the description is relatively simple, and relevant parts can be found in the description of the method embodiment.

[0193] In addition, embodiments of the present invention also provide an electronic device, such as... Figure 3 As shown, it includes a processor 301, a communication interface 302, a memory 303, and a communication bus 304, wherein the processor 301, the communication interface 302, and the memory 303 communicate with each other through the communication bus 304. Memory 303 is used to store computer programs; When the processor 301 executes the program stored in the memory 303, it implements the wind turbine bearing detection method of any of the wind power generation equipment described in the above embodiments: The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0194] The communication interface is used for communication between the aforementioned terminal and other devices.

[0195] The memory may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0196] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0197] like Figure 4 As shown, in another embodiment of the present invention, a computer-readable storage medium 401 is also provided, which stores instructions that, when run on a computer, cause the computer to execute the wind turbine bearing detection method for wind power generation equipment described in the above embodiment.

[0198] This application embodiment also provides a chip, which includes a processor and a communication interface. The communication interface and the processor are coupled. The processor is used to run programs or instructions to implement the various processes of the wind turbine bearing detection method embodiment of the above-described wind power generation equipment, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0199] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0200] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such 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 that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0201] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0202] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for testing wind turbine bearings in wind power generation equipment, characterized in that, include: Obtain the equivalent mass, damping coefficient, and time-varying equivalent stiffness of the wind turbine bearing; Read the structural parameters of the wind turbine bearing, and calculate the theoretical characteristic frequency of the fault at the target position of the wind turbine bearing under the target operating condition based on the structural parameters; By splicing together the theoretical fault characteristic frequency, the equivalent mass, the damping coefficient, and the equivalent stiffness, a physical cue vector is generated; Collect raw vibration signals under different working conditions; The target loss function is determined based on the physical cue vector and the original vibration signal, and a physical information diffusion model with cross-condition basic generation and diagnosis capabilities is pre-trained based on the target loss function. The vibration signal under the new working condition is received in real time and identified as online incremental data. Incremental training is performed on the physical information diffusion model based on the online incremental data to generate optimal incremental model parameters, and an incremental diagnostic model is generated based on the optimal incremental model parameters. When the current vibration signal of the target wind turbine bearing is received, the incremental diagnostic model is controlled to generate the detection results for the target wind turbine bearing.

2. The method according to claim 1, characterized in that, Before the step of determining the target loss function based on the physical cue vector and the original vibration signal, the method further includes: The original vibration signal is subjected to wavelet threshold denoising and normalization operations to generate a preprocessed original vibration signal.

3. The method according to claim 2, characterized in that, The step of determining the target loss function based on the physical cue vector and the original vibration signal includes: Construct a deep learning framework that includes a feature extraction module composed of a convolutional neural network, a diffusion and inverse diffusion module based on a diffusion probability model architecture, and a classification module. The preprocessed original vibration signal is input into the feature extraction module to generate a high-dimensional feature vector. By fusing the high-dimensional feature vector with the physical cue vector, and performing forward diffusion noise addition and reverse diffusion denoising operations through the diffusion and de-diffusion modules, the reconstructed original feature vector is obtained. A physical correlation model between the fault impact load and the final vibration displacement is established based on the physical equation of vibration response using the equivalent mass, the damping coefficient, and the equivalent stiffness. Substitute the original feature vector into the physical association model, calculate the error between the original feature vector and the physical cue vector, and generate the physical constraint loss. Calculate the mean square error between the original feature vector and the high-dimensional feature vector to obtain the inverse diffusion reconstruction loss; The classification module determines the predicted probability of the fault type to which the original feature vector belongs, and the true label of the original vibration signal corresponding to the original feature vector; The original feature vector is input into the classification module, and the cross-entropy loss between the predicted probability and the true label is calculated to obtain the classification loss. The total loss function is constructed by combining the physical constraint loss, the inverse diffusion reconstruction loss, and the classification loss.

4. The method according to claim 3, characterized in that, The steps for pre-training a physical information diffusion model with cross-condition basic generation and diagnosis capabilities based on the aforementioned objective loss function include: The deep learning framework is pre-trained using the total loss function to obtain pre-trained model parameters; The parameters of the pre-trained model are solidified to generate a pre-trained physical information diffusion model.

5. The method according to claim 4, characterized in that, The step of performing incremental training on the physical information diffusion model based on the online incremental data to generate optimal incremental model parameters includes: Wavelet threshold denoising and normalization operations are performed on the online incremental data to generate preprocessed online incremental training and validation sets. Freeze the bottom-level parameters of the feature extraction module in the pre-trained physical information diffusion model, and adjust the top-level parameters of the feature extraction module, the diffusion and inverse diffusion module, and the classification module to obtain a fine-tuning parameter set; By fusing the online incremental training set with the physical cue vector, and performing incremental learning operations through the fine-tuned parameter set, the inverse diffusion reconstruction loss, frequency bias loss, dynamic bias loss, and classification loss of the online incremental data are calculated. An incremental loss function is constructed by combining the inverse diffusion reconstruction loss, the frequency bias loss, the dynamic bias loss, and the classification loss of the incremental data; Based on the reverse diffusion process of the pre-trained physical information diffusion model, random noise is input to generate virtual historical samples that are consistent with the historical fault type characteristics of the original vibration signal. The virtual historical samples are mixed with the online incremental training set at a preset ratio to form a hybrid training set; The incremental loss function is used to perform incremental training on the hybrid training set to obtain incrementally updated model parameters; The classification accuracy and loss value of the incrementally updated model parameters are calculated using the online validation set. When the classification accuracy and the loss value meet the preset conditions, the incremental training operation on the mixed training set is stopped, and the optimal incremental model parameters are generated.

6. The method according to claim 5, characterized in that, The step of controlling the incremental diagnostic model to generate detection results for the target wind turbine bearing includes: Perform wavelet threshold denoising and normalization operations on the current vibration signal to generate a preprocessed current vibration signal; The preprocessed current vibration signal is input into the feature extraction module of the incremental diagnostic model to generate the current high-dimensional feature vector. By fusing the current high-dimensional feature vector with the physical cue vector, a reconstruction operation is performed through the diffusion-inverse diffusion module of the incremental diagnostic model to obtain the reconstructed current feature vector; The reconstructed current feature vector is input into the classification module of the incremental diagnostic model, and the failure type prediction result of the target wind turbine bearing and the confidence level of the failure type prediction result are output. If the confidence level is lower than a preset threshold, then the reconstructed current feature vector is subjected to secondary optimization through the inverse diffusion process of the incremental diagnostic model to obtain a refined feature vector, and the physical constraint error of the refined feature vector and the physical correlation model is calculated. When the physical constraint error is less than a preset threshold, the refined feature vector is used to generate a corrected fault type prediction result.

7. The method according to claim 5, characterized in that, Also includes: When the physical constraint error is not less than a preset threshold, an early warning signal is generated, and the fault type prediction result is associated with and stored with real-time operating condition data to generate a fault condition association database.

8. A wind turbine bearing testing device for wind power generation equipment, characterized in that, include: The bearing rotating component parameter acquisition module is used to acquire the equivalent mass, damping coefficient, and time-varying equivalent stiffness of the wind turbine bearing. The fault theoretical characteristic frequency calculation module is used to read the structural parameters of the wind turbine bearing and calculate the fault theoretical characteristic frequency of the target position of the wind turbine bearing under the target operating condition based on the structural parameters. The physical cue vector generation module is used to concatenate the fault theoretical characteristic frequency, the equivalent mass, the damping coefficient, and the equivalent stiffness to generate a physical cue vector; The raw vibration signal acquisition module is used to acquire raw vibration signals under different working conditions; The physical information diffusion model training module is used to determine the target loss function based on the physical cue vector and the original vibration signal, and to pre-train a physical information diffusion model with cross-condition basic generation and diagnosis capabilities based on the target loss function. The online incremental data acquisition module is used to receive vibration signals under new operating conditions in real time and determine the vibration signals under new operating conditions as online incremental data. The incremental diagnostic model generation module is used to perform incremental training on the physical information diffusion model based on the online incremental data, generate optimal incremental model parameters, and generate an incremental diagnostic model based on the optimal incremental model parameters. The detection result generation module is used to control the incremental diagnostic model to generate detection results for the target wind turbine bearing when the current vibration signal of the target wind turbine bearing is received.

9. An electronic device, characterized in that, It includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the method as described in claims 1-7.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the method as described in claims 1-7.

Images