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 are used for driving self-adaptive updating of kinetic parameters in real time through working condition data and carrying out closed-loop constraint on neural network output in combination with a physical mechanism, so that the problem of poor variable working condition adaptability caused by fixed physical parameters in a traditional diagnosis method is effectively solved. And the physical interpretability of the diagnosis conclusion and the reliability in a complex operation environment are obviously improved.

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] With the large-scale application of wind power generation technology, wind turbine bearings, as core supporting components, directly affect the safety and efficiency of the unit. Currently, wind turbine bearing fault diagnosis mainly faces the following technical challenges: Mechanistic model diagnostic methods for related technologies typically rely on fixed dynamic parameters (such as stiffness and damping). However, wind turbine generators operate under constantly fluctuating, random conditions, and frequent changes in speed and load can cause nonlinear shifts in the equivalent dynamic characteristics of the bearings. Because traditional models cannot capture the dynamic evolution of these physical parameters in real time, the constraint accuracy of the physical equations under complex operating conditions is insufficient, leading to false alarms.

[0003] While deep learning-based diagnostic methods can handle complex vibration signals, they are often considered "black box" models, lacking clear physical evidence for their diagnostic results. In scenarios lacking massive amounts of labeled data, purely data-driven models struggle to learn the underlying dynamics of vibration. When operating conditions exceed the distribution range of the training set, the model's generalization ability rapidly declines, making it difficult to guarantee the physical consistency of diagnostic conclusions. 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: Collect the original vibration signal of the wind turbine bearing, as well as the operating condition data reflecting the operating status of the wind power generation equipment; Construct input samples that reflect the mapping relationship between the original vibration signal and the operating condition data; Based on the input samples, generate state estimation subnet prediction values ​​that reflect the vibration displacement of the wind turbine bearing caused by the original vibration signal; Real-time physical parameters, including the equivalent dynamic stiffness and equivalent damping parameters of the wind turbine bearing, are generated based on the operating condition data. The target loss function is determined by the predicted values ​​of the state estimation subnet and the real-time physical parameters. The target model is generated by training based on the target loss function; When target data containing target vibration signals and target operating condition data is obtained, the target model is used to output the detection results for the target wind turbine bearing based on the target data.

[0006] Optionally, before the step of generating the state estimation subnet prediction value reflecting the vibration displacement of the wind turbine bearing caused by the original vibration signal based on the input sample, the method further includes: A two-stream network is constructed, which consists of a state estimation subnet and a parameter adaptation subnet.

[0007] Optionally, the step of generating a state estimation subnet prediction value reflecting the vibration displacement of the wind turbine bearing caused by the original vibration signal based on the input sample includes: The input sample is input into the state estimation subnet, and the state estimation subnet is controlled to output the state estimation subnet prediction value that reflects the vibration displacement of the wind turbine bearing caused by the original vibration signal.

[0008] Optionally, the operating condition data includes at least real-time speed data and real-time power data, and the step of generating real-time physical parameters based on the operating condition data, including the equivalent dynamic stiffness and equivalent damping parameters of the wind turbine bearing, includes: The operating condition data is input into the parameter adaptive subnet, and the parameter adaptive subnet is controlled to output real-time physical parameters including equivalent stiffness parameters and equivalent damping parameters under the current operating condition based on the real-time speed data and real-time power data.

[0009] Optionally, the step of determining the target loss function using the predicted values ​​of the state estimation subnet and the real-time physical parameters includes: A two-degree-of-freedom dynamic equation is established based on the geometry of the wind turbine bearing. The external load excitation term characterizing the external impact force borne by the wind turbine bearing in real time is calculated using the real-time power data. Calculate the derivative of the predicted value of the state estimation subnet, and substitute the derivative, the real-time physical parameters, and the external load excitation term into the two-degree-of-freedom dynamic equation. Determine the physical consistency loss by calculating the residuals on both sides of the two-degree-of-freedom dynamic equation. Calculate the mean square error between the predicted values ​​of the state estimation subnet and the original vibration signal to determine the data fitting loss; Based on the degree of overflow of the residual relative to the preset physical law violation tolerance band, and combined with learnable slack variables, a soft reward / penalty loss is constructed to characterize the degree of deviation of the model from the physical law. Establish a composite total loss function consisting of the data fitting loss, the physical consistency loss, and the soft reward / penalty loss.

[0010] Optionally, the step of training and generating the target model based on the target loss function includes: Set initial weighting coefficients for the data fitting loss, the physical consistency loss, and the soft reward / penalty loss; Obtain the initial gradient norm of each loss term with respect to the network parameters in the data fitting loss, the physical consistency loss, and the soft reward / penalty loss; Adjust the values ​​of the physical consistency loss and soft reward / penalty loss in the initial weighting coefficients to obtain the target weighting coefficients for the first stage; The dual-stream network is preheated and iterated using the first-stage target weighting coefficients to obtain the preheating network weights that initially capture the distribution pattern of the original vibration signal. Using the initial gradient norm as a baseline, the gradient energy ratio of each loss term in the current iteration is calculated in real time using the gradient norm balancing algorithm, and the second-stage real-time weighting coefficients are obtained by dynamically correcting them over time. Using the real-time weighting coefficients of the second stage and the weights of the preheating network, the composite total loss function is jointly optimized and iterated to obtain candidate network parameters whose physical residuals stably fall into the preset physical law violation tolerance zone; When the total composite loss value corresponding to the candidate network parameters is determined to reach a preset convergence threshold, the candidate network parameters are frozen and the target model is generated.

[0011] Optionally, the step of using the target model to output the detection results for the target wind turbine bearing based on the target data includes: Extract the target vibration sequence and target working condition features from the target data that match the dimension of the target model input layer to obtain the input tensor to be diagnosed; The input tensor to be diagnosed is input into the state estimation subnet of the target model to obtain a predicted vibration displacement sequence that reflects the real-time dynamic response of the target bearing fan. Input the target working condition features in the input tensor to be diagnosed into the parameter adaptive subnet of the target model to obtain the real-time stiffness prediction value and real-time damping prediction value under the current working condition. Calculate the derivative of the predicted vibration displacement sequence, and substitute the derivative, the real-time stiffness prediction value, the real-time damping prediction value, and the external load term deduced from the target working condition characteristics into the two-degree-of-freedom dynamic equation based on the geometric structure of the target wind turbine bearing to obtain the target residual distribution sequence. By comparing the deviation of the real-time stiffness prediction value and the real-time damping prediction value from the benchmark value, a parameter degradation evaluation index reflecting the abnormal drift trend of physical parameters is obtained. The mean and peak values ​​of the target residual distribution sequence on the time axis are statistically analyzed, and a quantitative evaluation is performed in conjunction with the violation tolerance band of the physical law to obtain a physical credibility score that reflects the degree to which the model's predictive behavior conforms to the physical law. By combining the parameter degradation evaluation index and the physical reliability score, the fault logic is determined, and a complete detection result including fault type, severity and diagnostic reliability is output.

[0012] Secondly, embodiments of this application provide a wind turbine bearing testing device for wind power generation equipment, comprising: The training data acquisition module is used to collect the original vibration signal of the wind turbine bearing and the operating condition data reflecting the operating status of the wind power generation equipment. An input sample construction module is used to construct input samples that reflect the mapping relationship between the original vibration signal and the operating condition data; The state estimation subnet prediction value generation module is used to generate state estimation subnet prediction values ​​that reflect the vibration displacement of the wind turbine bearing caused by the original vibration signal based on the input sample. A real-time physical parameter generation module is used to generate real-time physical parameters based on the operating condition data, including the equivalent dynamic stiffness and equivalent damping parameters of the wind turbine bearing. The target loss function determination module is used to determine the target loss function based on the predicted values ​​of the state estimation subnet and real-time physical parameters. The target model generation module is used to train and generate a target model based on the target loss function. The detection result generation module is used to output the detection results for the target wind turbine bearing based on the target data containing the target vibration signal and the target operating condition data when the target data is obtained.

[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: This invention effectively solves the problem of poor adaptability to changing operating conditions caused by fixed physical parameters in traditional diagnostic methods by driving the adaptive update of dynamic parameters in real time through operating condition data and combining physical mechanisms to impose closed-loop constraints on the neural network output. This significantly improves the physical interpretability of diagnostic conclusions and the reliability in complex operating environments.

[0017] Furthermore, in this embodiment of the invention, by dynamically deriving the equivalent dynamic stiffness and damping that evolves with rotational speed and load using real-time operating data, the lack of accuracy in variable operating condition constraints caused by fixed parameters in traditional mechanistic models is effectively compensated. At the same time, by substituting the adaptive physical parameters and the predicted vibration response into the dynamic equation to construct a composite loss function, the deep learning model must follow physical conservation laws while fitting vibration characteristics, thereby eliminating the "black box" attribute of purely data-driven models. Under the condition of not relying on massive amounts of labeled data, the physical consistency and generalization ability of diagnostic results under complex and unknown operating conditions are ensured. Attached Figure Description

[0018] 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 4 This is a schematic diagram of a computer-readable medium provided in an embodiment of the present invention. Detailed Implementation

[0019] 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.

[0020] 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.

[0021] 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.

[0022] 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: Collect the original vibration signal of the wind turbine bearing and the operating condition data reflecting the operating status of the wind power generation equipment; Step 102: Construct an input sample that reflects the mapping relationship between the original vibration signal and the operating condition data; Step 103: Generate state estimation subnet prediction values ​​based on the input samples, reflecting the vibration displacement of the wind turbine bearing caused by the original vibration signal; Step 104: Generate real-time physical parameters based on the operating data, including the equivalent dynamic stiffness and equivalent damping parameters of the wind turbine bearing. Step 105: Determine the target loss function using the predicted values ​​of the state estimation subnet and the real-time physical parameters; Step 106: Train and generate the target model based on the target loss function; Step 107: When target data containing target vibration signals and target operating condition data is obtained, the target model is used to output the detection results for the target wind turbine bearing based on the target data.

[0023] In this embodiment of the invention, the original vibration signal of the wind turbine bearing and the operating condition data reflecting the operating status of the wind power generation equipment can be collected to simultaneously obtain the vibration data reflecting the micro-response of the bearing and the background parameters of the macro-operating environment, providing basic materials for constructing the "operating condition-response" mapping relationship. The raw vibration signal can refer to the time-series electrical signal directly collected by the accelerometer installed on the bearing housing without in-depth processing, which contains all the impact and noise information during the operation of the bearing.

[0024] Operating condition data refers to a set of parameters that describe the current external operating conditions of wind power generation equipment. It is usually extracted from the SCADA system and is used to define the physical boundary environment of bearing operation.

[0025] In this embodiment of the invention, an input sample reflecting the mapping relationship between the original vibration signal and the working condition data can be constructed to align and fuse heterogeneous vibration signals and working condition data on the time axis, forming a feature tensor that can be read by a neural network. The mapping relationship can refer to the causal correspondence between a specific operating condition (such as speed and power) and the specific vibration response generated by the bearing under that condition.

[0026] The input samples are data fragments that have been standardized, timestamped, and concatenated, and serve as the direct input units for model training and inference.

[0027] In this embodiment of the invention, a state estimation subnet prediction value reflecting the vibration displacement of the wind turbine bearing caused by the original vibration signal can be generated based on the input sample, so as to utilize the powerful function fitting capability of the neural network to mathematically simulate the vibration displacement of the bearing under specific time and operating conditions. The state estimation subnet is a branch of the neural network specifically responsible for learning the mapping relationship between time, operating conditions, and vibration displacement; it acts as the "observer" in the physical information neural network. The predicted value of the state estimation subnet is the theoretical displacement value calculated by the neural network based on the input samples.

[0028] Vibration displacement refers to the distance offset of a bearing component from its equilibrium position during vibration, and it is the most critical state variable in the dynamic equation.

[0029] In this embodiment of the invention, real-time physical parameters including the equivalent dynamic stiffness and equivalent damping parameters of the wind turbine bearing can be generated based on the operating condition data, thereby giving the model the ability to perceive the evolution of physical characteristics with operating conditions and breaking the limitation of fixed physical parameters in traditional methods. Equivalent dynamic stiffness is the ability of a bearing system to resist elastic deformation, and it will dynamically drift with changes in oil film thickness caused by rotational speed and changes in contact area caused by load.

[0030] The equivalent damping parameter is the ability of a bearing system to consume energy and suppress vibration, reflecting the dynamic characteristics of lubrication and internal friction.

[0031] Real-time physical parameters are dynamic constant values ​​derived in real time by a neural network based on current operating conditions such as rotational speed and power, and are consistent with the current physical scenario.

[0032] In this embodiment of the invention, the target loss function can be determined by the predicted value of the state estimation subnet and the real-time physical parameters, so as to establish a comprehensive evaluation index that integrates "data bias" and "physical law bias", guiding the model to optimize in a direction that conforms to both measured data and mechanistic laws. The objective loss function is a mathematical formula used to measure the quality of a model's predictions. It includes the data fitting error and the physical equation residuals and serves as the "guide" for model training. It is usually composed of multiple weighted losses to ensure that the model achieves an optimal balance under multiple constraints.

[0033] In this embodiment of the invention, a target model can be trained and generated based on the target loss function, so as to minimize the loss function through an iterative optimization algorithm, fix the network weights, and finally produce a stable model with intelligent diagnostic capabilities. The target model is the final neural network system that has been trained and is capable of directly performing fault identification and state assessment tasks.

[0034] In this embodiment of the invention, when target data containing target vibration signals and target operating condition data is obtained, the target model can be used to output the detection results for the target wind turbine bearing based on the target data, and the model can be applied to the actual production scenario to generate a diagnostic conclusion containing fault information in real time. Target data can refer to unknown state signals and operating parameters that are collected in real time during the inference phase and are to be detected.

[0035] The test results are the final report output by the model, which includes decision-making information such as the health status of the bearing and the type of fault.

[0036] This invention effectively solves the problem of poor adaptability to changing operating conditions caused by fixed physical parameters in traditional diagnostic methods by driving the adaptive update of dynamic parameters in real time through operating condition data and combining physical mechanisms to impose closed-loop constraints on the neural network output. This significantly improves the physical interpretability of diagnostic conclusions and the reliability in complex operating environments.

[0037] 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.

[0038] In an optional embodiment of the present invention, the original vibration signal of the wind turbine bearing and the operating condition data reflecting the operating status of the wind power generation equipment are collected; the operating condition data includes at least real-time speed data and real-time power data. The original vibration signal is processed by variational mode decomposition algorithm, and the working condition data and vibration features are time-stamped and feature-stitched to construct an input sample that reflects the mapping relationship between the original vibration signal and the working condition data; A two-stream network is constructed, which consists of a state estimation subnet and a parameter adaptation subnet. The input sample is input into the state estimation subnet, and the state estimation subnet is controlled to output the state estimation subnet prediction value that reflects the vibration displacement of the wind turbine bearing caused by the original vibration signal. The operating condition data is input into the parameter adaptive subnet, and the parameter adaptive subnet is controlled to adaptively output real-time physical parameters including equivalent stiffness parameters and equivalent damping parameters under the current operating condition according to the changes in the instantaneous speed data and instantaneous power data. A two-degree-of-freedom dynamic equation is established based on the bearing geometry, and the external load excitation term characterizing the external impact force borne by the wind turbine bearing in real time is calculated using the real-time power data. The derivative of the predicted value of the state estimation subnet is calculated using automatic differentiation, and the physical consistency loss is determined by substituting the derivative, the real-time physical parameters, and the external load excitation terms into the two-degree-of-freedom dynamic equation and calculating the residuals on both sides of the two-degree-of-freedom dynamic equation. Calculate the mean square error between the predicted values ​​of the state estimation subnet and the original vibration signal to determine the data fitting loss; Based on the degree of overflow of the residual relative to the preset physical law violation tolerance band, and combined with learnable slack variables, a soft reward / penalty loss is constructed to characterize the degree of deviation of the model from the physical law. Establish a composite total loss function consisting of the data fitting loss, the physical consistency loss, and the soft reward / penalty loss; The target model is generated by training based on the target loss function; The reasoning and diagnosis unit inputs the target vibration signal collected in real time and the target working condition data corresponding to the target vibration signal into the target model, and controls the target model to generate a diagnosis report that includes fault type, abnormal drift trend of physical parameters and physical reliability score.

[0039] Optionally, in embodiments of the present invention, a dual-stream network may be constructed before the step of generating a state estimation subnet prediction value reflecting the vibration displacement of the wind turbine bearing caused by the original vibration signal based on the input sample.

[0040] A two-stream network is a deep learning architecture with two independent and parallel sub-network branches, similar to two parallel pipelines that process different tasks and eventually converge at the loss function.

[0041] In its specific implementation, the dual-stream network comprises a state estimation subnet and a parameter adaptation subnet.

[0042] The purpose of constructing the state estimation subnet is to establish a computational branch specifically responsible for fitting the time series characteristics of vibration signals, in order to extract and reconstruct the motion state of the bearing from the data. The state estimation subnet is the first branch in the two-stream network. Its core function is to act as a nonlinear function mapper to predict the vibration response (such as displacement) of the bearing in real time based on the input time or operating condition characteristics.

[0043] The purpose of constructing a parameter adaptive subnet is to establish a computational branch specifically responsible for sensing environmental conditions and mapping physical properties, so that the model has the ability to discriminate the mechanism parameters of the mechanical system in real time. The parameter adaptive subnet is the second branch in the two-stream network. Its core function is to learn the evolution relationship between operating condition data (speed, power) and physical mechanism parameters (stiffness, damping), so that the physical parameters are no longer preset rigid values.

[0044] Optionally, the step of generating a state estimation subnet prediction value reflecting the vibration displacement of the wind turbine bearing caused by the original vibration signal based on the input sample includes: The input sample is input into the state estimation subnet, and the state estimation subnet is controlled to output the state estimation subnet prediction value that reflects the vibration displacement of the wind turbine bearing caused by the original vibration signal.

[0045] In this embodiment of the invention, preprocessed multidimensional feature data can be fed into a specific neural network branch to initiate a nonlinear mapping process from features to physical states. Input samples can refer to composite data vectors formed by splicing the original vibration characteristics with the corresponding operating conditions (speed, power, etc.) at the time, which serve as the initial signals for model perception.

[0046] The input operation refers to the forward propagation initiation action of neural network computation, which assigns digitized features to the input layer neurons of the state estimation subnet.

[0047] In this embodiment of the invention, the core kinematic features of a bearing can be abstracted and reconstructed from noisy input data by utilizing a multi-layer nonlinear activation mechanism within the subnet. The state estimation subnet is a neural network module specifically designed to simulate the motion behavior of bearings. It establishes a mathematical relationship between complex input signals and clean motion states by learning from a large number of samples.

[0048] In this embodiment of the invention, a physical variable that can represent the real-time operating attitude of the bearing can be generated, providing a benchmark for subsequent alignment with the physical dynamics equations. Vibration displacement describes the instantaneous distance by which a bearing component deviates from its equilibrium position, and it is the most direct and critical physical quantity observed in dynamic analysis.

[0049] The state estimation subnet predicts the value of the bearing as an estimate of "theoretically how much displacement should be" under specific operating conditions.

[0050] This invention, through its embodiment, introduces input samples of fused operating condition information into a specific state estimation subnet, achieving a high-precision conversion from high-dimensional nonlinear signals to specific physical state quantities (vibration displacement), providing a standard source of state observation for subsequent residual verification using physical equations.

[0051] Optionally, the operating condition data includes at least real-time speed data and real-time power data, and the step of generating real-time physical parameters based on the operating condition data, including the equivalent dynamic stiffness and equivalent damping parameters of the wind turbine bearing, includes: The operating condition data is input into the parameter adaptive subnet, and the parameter adaptive subnet is controlled to output real-time physical parameters including equivalent stiffness parameters and equivalent damping parameters under the current operating condition based on the real-time speed data and real-time power data.

[0052] In this embodiment of the invention, macroscopic physical variables reflecting the operating environment of the equipment can be introduced into the model to provide an external driving source for the dynamic identification of physical parameters. Real-time rotational speed data refers to the number of revolutions the bearing shaft makes per unit time at the current moment, which directly affects the thickness of the lubricating oil film inside the bearing and its dynamic stiffness.

[0053] Real-time power data refers to the current output power of a wind turbine generator, which indirectly reflects the magnitude of the external load on the bearing and is an important reference for calculating equivalent damping.

[0054] In this embodiment of the invention, a nonlinear functional relationship between working conditions and physical mechanisms can be learned using a neural network, transforming macroscopic working condition indicators into microscopic physical performance indicators. The parameter adaptive subnet is a neural network module specifically designed to learn the evolution of the physical properties of mechanical systems. It can adjust its internal mapping relationship in real time according to changes in input conditions, thereby achieving parameter generation that "changes with operating conditions".

[0055] The embodiments of the present invention can produce equivalent parameter values ​​that conform to the physical characteristics of the bearing at the current moment, providing physical quantity support for constructing accurate dynamic constraint equations. The equivalent stiffness parameter is a physical quantity that describes the bearing's ability to resist deformation under specific loads and speeds.

[0056] The equivalent damping parameter is a physical quantity that describes the ability of a bearing system to dissipate vibration energy under specific operating conditions.

[0057] Real-time physical parameters are combinations of stiffness and damping that are instantaneously predicted by the subnet based on the current operating conditions. They are dynamic characteristic quantities that describe the current physical nature of the bearing.

[0058] In this embodiment of the invention, a real-time mapping between operating conditions and dynamic parameters is established through a parameter adaptive subnet, enabling the diagnostic model to accurately perceive the dynamic drift of bearing physical characteristics under different speeds and powers. This effectively solves the model mismatch problem caused by fixed physical parameter settings in traditional diagnostic methods.

[0059] Optionally, the step of determining the target loss function using the predicted values ​​of the state estimation subnet and the real-time physical parameters includes: A two-degree-of-freedom dynamic equation is established based on the geometry of the wind turbine bearing. The external load excitation term characterizing the external impact force borne by the wind turbine bearing in real time is calculated using the real-time power data. Calculate the derivative of the predicted value of the state estimation subnet, and substitute the derivative, the real-time physical parameters, and the external load excitation term into the two-degree-of-freedom dynamic equation. Determine the physical consistency loss by calculating the residuals on both sides of the two-degree-of-freedom dynamic equation. Calculate the mean square error between the predicted values ​​of the state estimation subnet and the original vibration signal to determine the data fitting loss; Based on the degree of overflow of the residual relative to the preset physical law violation tolerance band, and combined with learnable slack variables, a soft reward / penalty loss is constructed to characterize the degree of deviation of the model from the physical law. Establish a composite total loss function consisting of the data fitting loss, the physical consistency loss, and the soft reward / penalty loss.

[0060] In this embodiment of the invention, by establishing a two-degree-of-freedom dynamic equation, a framework of underlying physical mechanisms governing bearing operation can be constructed, transforming the diagnostic task from a purely data-driven task into one constrained by physical laws. The two-degree-of-freedom dynamic equation is a mathematical differential equation that describes the vibration law of a bearing in two orthogonal directions (usually horizontal and vertical), and represents the energy balance relationship of the mechanical system.

[0061] In this embodiment of the invention, the macroscopic power in SCADA can be transformed into the right-hand side of the dynamic equation by calculating the external load excitation term, thus providing a real-time and authentic energy input source for the equation. The external load excitation term is the external excitation force acting on the bearing system, reflecting the pressure or impact force applied to the bearing housing due to changes in the fan load.

[0062] In this embodiment of the invention, the physical consistency loss can be determined to verify whether the motion state output by the neural network violates Newton's second law, and the "physical violation" can be transformed into an optimizable error signal. Automatic differentiation is a computer algorithm used to accurately calculate the first derivative (velocity) and second derivative (acceleration) of the displacement sequence output by a neural network with respect to time.

[0063] The residual is the deviation between the force balance term on the left side of the dynamic equation and the excitation term on the right side.

[0064] Physical consistency loss is an indicator that quantifies the residual. The larger the value, the less the prediction result conforms to physical laws.

[0065] In this embodiment of the invention, by determining the data fitting loss, it can be ensured that the displacement output by the model closely approximates the actual observation signal collected by the sensor in terms of both numerical and waveform characteristics. Mean squared error is a statistic that measures the difference between predicted and measured values; it is the average of the sum of the squares of the differences between the two.

[0066] Data fitting loss is a loss term that characterizes the similarity between the predicted curve and the measured curve.

[0067] In this embodiment of the invention, a soft reward-penalty loss can be constructed to provide a flexible buffer for reasonable errors caused by model simplification or environmental noise, while severely penalizing serious physical violations. The pre-defined physical law violation tolerance zone is a small interval that allows physical residuals to exist, representing the model's tolerance for minor physical mismatches caused by force majeure.

[0068] A learnable slack variable is a coefficient that automatically adjusts during training and can dynamically absorb noise errors that cannot be explained by physical equations.

[0069] In this embodiment of the invention, a composite total loss function can be established to integrate multi-dimensional evaluation indicators into a unified objective, enabling model training to simultaneously take into account observation data, physical mechanisms, and environmental noise. The composite total loss function is a weighted sum of the losses from data fitting, physical constraints, and soft rewards / penalties. It is the ultimate goal that guides the model to perform deep fusion optimization.

[0070] This invention achieves physical closed-loop verification of neural network prediction behavior by establishing dynamic equations and introducing a composite loss function with tolerance bandwidth. This ensures high-precision fitting of diagnostic results to measured signals and filters out random noise interference using physical mechanisms, significantly enhancing the robustness of the diagnostic model under complex and variable conditions.

[0071] Optionally, the step of training and generating the target model based on the target loss function includes: Set initial weighting coefficients for the data fitting loss, the physical consistency loss, and the soft reward / penalty loss; Obtain the initial gradient norm of each loss term with respect to the network parameters in the data fitting loss, the physical consistency loss, and the soft reward / penalty loss; Adjust the values ​​of the physical consistency loss and soft reward / penalty loss in the initial weighting coefficients to obtain the target weighting coefficients for the first stage; The dual-stream network is preheated and iterated using the first-stage target weighting coefficients to obtain the preheating network weights that initially capture the distribution pattern of the original vibration signal. Using the initial gradient norm as a baseline, the gradient energy ratio of each loss term in the current iteration is calculated in real time using the gradient norm balancing algorithm, and the second-stage real-time weighting coefficients are obtained by dynamically correcting them over time. Using the real-time weighting coefficients of the second stage and the weights of the preheating network, the composite total loss function is jointly optimized and iterated to obtain candidate network parameters whose physical residuals stably fall into the preset physical law violation tolerance zone; When the total composite loss value corresponding to the candidate network parameters is determined to reach a preset convergence threshold, the candidate network parameters are frozen and the target model is generated.

[0072] In this embodiment of the invention, by setting initial weighting coefficients, an initial magnitude balance point is established for loss terms of different dimensions, preventing model instability in the early stages of training due to excessive differences in the values ​​of various losses. The initial weighting coefficients are manually set weighting constants that are pre-assigned to the losses of data fitting, physical constraints, and soft rewards / penalties.

[0073] In this embodiment of the invention, by obtaining the initial gradient norm, the intensity of the "push" generated by each loss function on the network parameters at the initial moment is extracted and used as the original physical reference benchmark for subsequent automatic weight adjustment. The gradient norm is a metric that measures the strength of the impact of the loss function on the update of model parameters; the larger the norm, the stronger the willingness of the loss term to modify the model parameters.

[0074] In this embodiment of the invention, the first-stage target weighting coefficient is obtained by adjustment. Following the strategy of "starting with the easy and then moving to the difficult", the loss ratio is intervened to create a primary learning environment for the model that focuses on data fitting. The first-stage target weighting coefficients are a combination of coefficients after the physical constraint weights have been weakened, aiming to allow the model to first learn to fit the waveform and then learn the physical laws.

[0075] In this embodiment of the invention, the preheating network weights are obtained through preheating iteration, which completes the initial "basic literacy" of the model, enabling it to quickly grasp the temporal distribution characteristics of vibration signals without being disturbed by strict physical constraints. Warm-up iterations can refer to a small number of low-constraint loop computations performed in the early stages of training.

[0076] The preheating network weights are the parameter states stored inside the neural network after preheating training, which provide a good starting point for subsequent complex physical mechanism fusion.

[0077] In this embodiment of the invention, by calculating the real-time weighting coefficients in the second stage and introducing an automated feedback mechanism, the "competition" problem between physical residuals and data errors in the gradient update process is solved, and dynamic balance is achieved. Gradient norm balancing algorithms are intelligent algorithms (such as GradNorm) that automatically adjust the weight ratio of each loss term based on the real-time performance of its gradient, ensuring that no single loss term dominates or annihilates the others.

[0078] Gradient energy percentage reflects the update activity of a certain loss term relative to the initial baseline at the current moment.

[0079] In this embodiment of the invention, candidate network parameters are obtained through joint optimization iteration. Under the guidance of dynamic weights, the model is driven to deeply integrate physical laws and observation data to find a solution that satisfies all constraints. Joint optimization iteration is an advanced training process in which physical constraints and data fitting simultaneously affect the model parameters.

[0080] Candidate network parameters are the set of internal parameters of the model that have achieved a good balance between physical consistency and data accuracy when training is nearing completion.

[0081] In this embodiment of the invention, by determining convergence and generating a target model, a final quality check is performed. When the model performs stably and meets the physical conservation requirements, learning is stopped and the function is locked. The convergence threshold is a preset small value. When the decrease in the loss function is less than this value, the model is considered to have found the optimal solution.

[0082] The freeze operation refers to the operation of stopping network parameter updates, turning the model into a production state where it can directly perform inference tasks.

[0083] In this embodiment of the invention, by introducing an initial gradient norm benchmark and a gradient norm balancing algorithm, an automatic evolution from "pure data-driven warm-up" to "physical-data dynamic trade-off" is achieved. This effectively solves the gradient imbalance and convergence difficulties that easily occur in the training of physical information neural networks, and ensures that the final model can smoothly and stably converge to the global optimal state that conforms to physical laws.

[0084] Optionally, the step of using the target model to output the detection results for the target wind turbine bearing based on the target data includes: Extract the target vibration sequence and target working condition features from the target data that match the dimension of the target model input layer to obtain the input tensor to be diagnosed; The input tensor to be diagnosed is input into the state estimation subnet of the target model to obtain a predicted vibration displacement sequence that reflects the real-time dynamic response of the target bearing fan. Input the target working condition features in the input tensor to be diagnosed into the parameter adaptive subnet of the target model to obtain the real-time stiffness prediction value and real-time damping prediction value under the current working condition. Calculate the derivative of the predicted vibration displacement sequence, and substitute the derivative, the real-time stiffness prediction value, the real-time damping prediction value, and the external load term deduced from the target working condition characteristics into the two-degree-of-freedom dynamic equation based on the geometric structure of the target wind turbine bearing to obtain the target residual distribution sequence. By comparing the deviation of the real-time stiffness prediction value and the real-time damping prediction value from the benchmark value, a parameter degradation evaluation index reflecting the abnormal drift trend of physical parameters is obtained. The mean and peak values ​​of the target residual distribution sequence on the time axis are statistically analyzed, and a quantitative evaluation is performed in conjunction with the violation tolerance band of the physical law to obtain a physical credibility score that reflects the degree to which the model's predictive behavior conforms to the physical law. By combining the parameter degradation evaluation index and the physical reliability score, the fault logic is determined, and a complete detection result including fault type, severity and diagnostic reliability is output.

[0085] In this embodiment of the invention, the input tensor to be diagnosed is obtained by extracting features, so as to transform the real-time unstructured raw data into a standard mathematical format that the model can recognize, and complete the format alignment before inference. The target vibration sequence is a segment of the vibration signal of the bearing under test, acquired in real time.

[0086] The input tensor to be diagnosed is a multidimensional array that integrates vibration and operating condition characteristics after standardization and dimensional transformation.

[0087] In this embodiment of the invention, by obtaining a predicted vibration displacement sequence, the macroscopic motion trajectory of the bearing under the current working condition can be reconstructed using the state prediction capability of the model. The predicted vibration displacement sequence is a continuously changing time-varying predicted value of the bearing center displacement output by the state estimation subnet, and it forms the basis for evaluating the vibration amplitude.

[0088] In this embodiment of the invention, by obtaining real-time stiffness and damping prediction values, the internal dynamic health parameters of the bearing can be identified in real time online without disassembling the bearing. The real-time stiffness / damping prediction is the current physical property calculated by the parameter adaptive subnet based on real-time operating conditions, which directly reflects the mechanical decay state of the bearing.

[0089] In this embodiment of the invention, by obtaining the target residual distribution sequence, physical laws are introduced as a check criterion during the inference stage to quantify the degree of matching between the currently observed motion state and known mechanical laws. The target residual distribution sequence is the distribution of the error generated on both sides of the two-degree-of-freedom equation over time after substituting the displacement derivative, parameters, and load into the equation.

[0090] In this embodiment of the invention, by obtaining a parameter degradation evaluation index, early damage that has not yet appeared in the vibration signal but has occurred at the physical property level can be discovered through the direct offset of physical parameters. The reference values ​​are the standard values ​​of stiffness and damping of the bearing at the beginning of its entire life cycle (healthy state).

[0091] Abnormal parameter drift trends are the deviation paths of real-time predicted values ​​from healthy baseline values, indicating bearing material fatigue or lubrication failure.

[0092] In this embodiment of the invention, a physical credibility score is obtained to provide quality assurance for the diagnostic results and to identify “serious nonsense” generated by the model due to abnormal operating condition disturbances. The physical reliability score is calculated based on the degree of overlap between the residual and the tolerance band; the higher the score, the more the model output conforms to physical logic.

[0093] In this embodiment of the invention, by outputting complete detection results, all information from the data layer, physical layer, and confidence layer is integrated to provide a final judgment with decision-making reference value. Fault logic determination involves comprehensively considering the intensity of parameter drift and residual fluctuation to determine the causal relationship of the fault occurrence.

[0094] A complete test result is a comprehensive medical report that integrates "what the fault is", "how serious it is", and "how reliable the conclusion is".

[0095] This invention achieves dual diagnosis of the health status of wind turbine bearings by comparing physical parameter drift and calculating physical residuals in real time during the inference stage, which includes both "numerical prediction" and "mechanism verification". This not only improves the accuracy of fault identification under varying operating conditions, but also solves the problem that deep learning diagnostic results are difficult for engineers to trust by providing a physical credibility score.

[0096] 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.

[0097] In practical applications, with the acceleration of global energy transition and the advancement of "dual carbon" goals, the installed capacity of wind power generation continues to grow rapidly. Large wind turbine units generally use gearboxes to convert the low speed of the rotor to the high speed of the generator. Bearings, as key supporting components, are subjected to high alternating loads, complex stresses, and drastic temperature changes over long periods, making them highly susceptible to failure. Industry statistics show that bearing failure is one of the main causes of unplanned wind turbine downtime, causing not only significant economic losses but also potentially triggering catastrophic accidents. Therefore, achieving early, accurate, and intelligent fault diagnosis is of great significance for ensuring grid security and reducing operation and maintenance costs.

[0098] Current diagnostic methods are mainly divided into two categories: traditional methods based on signal processing and data-driven methods based on deep learning. The former relies on expert experience to extract vibration features, but under highly non-stationary conditions caused by random wind speed fluctuations, the characteristic frequencies are easily modulated or submerged in noise, significantly reducing diagnostic reliability. The latter (such as CNN, LSTM, Transformer, etc.) can automatically learn fault mapping relationships and improve efficiency, but it faces three major bottlenecks: first, the model lacks physical interpretability, making it difficult to gain engineering trust; second, it heavily relies on a large number of labeled fault samples, while fault data is scarce in actual wind fields and the categories are extremely imbalanced; and third, the model has weak generalization ability, and the distribution of training and testing data shifts due to differences in operating conditions, resulting in poor cross-wind field transfer performance.

[0099] To integrate the advantages of physical priors and data-driven approaches, Physical Information Neural Networks (PINNs) have been introduced into the field of fault diagnosis. By embedding dynamic equations into loss functions, PINNs guide models to approximate real system behavior under conditions of few or no labels. However, direct application of traditional PINNs still has significant limitations: First, they assume system parameters (such as stiffness and damping) to be fixed constants, failing to characterize the time-varying dynamic characteristics of wind turbine bearings caused by wear, temperature changes, and load fluctuations. Second, the use of hard mean square error to force a simplified physical model ignores the approximation inherent in the modeling itself, easily leading to overregulation and weakening the ability to fit measured data. Third, external excitations are usually simplified to analytical functions (such as sine waves), while actual wind turbine loads are formed by the coupling of aerodynamics, gravity, and unbalanced forces to create broadband stochastic excitations, making accurate pre-setting difficult.

[0100] In view of this, this invention proposes an intelligent fault diagnosis method for wind turbine bearings based on a learnable adaptive mechanism PINN. By establishing a parameter learnable mechanism and introducing a soft reward and punishment strategy under real-world operating conditions, this method solves the problems of rigid parameters, overly stringent physical constraints, and difficulty in adapting to changing operating conditions in existing PINN models, providing technical support for the current field of wind turbine operation, maintenance, and intelligent fault diagnosis technology.

[0101] To address the technical problems of fixed and rigid physical parameters, poor adaptability to varying operating conditions, and insufficient generalization ability due to overly rigid physical constraints in the application of Physical Information Neural Networks (PINN) for wind turbine bearing fault diagnosis, this paper proposes a learnable adaptive mechanism-based intelligent fault diagnosis method for wind turbine bearings. This method no longer treats structural parameters in the dynamic equations as constants, but instead constructs them as learnable variables that change with the state. By introducing a loss function with a soft reward / penalty mechanism, the model is allowed to find the optimal balance between physical laws and observed data, thereby achieving high-precision and robust fault diagnosis of wind turbine bearings under complex operating conditions.

[0102] Specifically, firstly, a two-stream neural network architecture consisting of a main network and a parameter subnetwork is constructed. The main network is responsible for fitting the vibration response of the bearing system, while the parameter subnetwork is responsible for adaptively outputting the physical parameters (stiffness and damping) at the current moment based on the input state. Next, based on the bearing dynamics mechanism, a dynamic residual equation containing nonlinear contact force and time-varying parameters is derived, and a soft reward-penalty loss function based on Sigmoid weighting and relaxation variables is designed. Secondly, using the collected actual wind turbine vibration acceleration and speed signals, a multi-stage adaptive weighted training strategy is adopted to optimize the network, synchronously updating the network weights and physical parameters. Finally, based on the residual distribution of the reconstructed signal and the actual signal after training, combined with the evolution trend of physical parameters, a health index is constructed and a dynamic threshold is set to realize the identification of bearing fault types and the quantitative assessment of their severity.

[0103] Furthermore, in a specific implementation, the following steps may be included: Step 1: Multi-source data acquisition and preprocessing.

[0104] Vibration acceleration sensors and speed sensors installed on the main bearing housing and high-speed shaft of the gearbox of the wind turbine are used to synchronously collect long-term series of operating data through a SCADA (Supervisory Control and Data Acquisition) system. The data includes high-frequency vibration signals, real-time speed signals, generator power, and nacelle temperature.

[0105] The collected raw data underwent cleaning and standardization. First, null values ​​and outliers caused by sensor malfunctions or packet loss were removed. Second, a noise reduction algorithm based on Variational Mode Decomposition (VMD) was used to filter out background noise, retaining frequency band components containing fault impact characteristics. Next, the vibration signal was normalized using Z-score normalization to eliminate amplitude dimension differences and accelerate subsequent network convergence. Finally, an overlapping sliding window technique was used to segment the long sequence signal into fixed-length sample segments, each containing a vibration sequence. The corresponding speed sequence And operating condition labels.

[0106] Step 2: Build a physical information neural network with a two-stream architecture.

[0107] Design a parallel neural network architecture that includes a "state estimation subnet" and a "parameter adaptation subnet".

[0108] (1) State estimation subnet (State-Net): A fully connected neural network (DNN) or a one-dimensional convolutional neural network (1D-CNN) is used as the backbone. The input is the time coordinate. and operating characteristics (speed) The output is the predicted displacement response of the system. Using Automatic Differentiation (AD) technology, the output is... Regarding input Perform differentiation and calculate the first derivative. (Velocity) and second derivative (Acceleration). This process is entirely based on backpropagation of the computational graph, avoiding truncation errors caused by numerical differencing.

[0109] (2) Param-Net: This is a lightweight, fully connected network. The input is the current operating condition vector (including slow variables such as speed and temperature), and the output is the estimated physical parameters at the current moment. To ensure the rationality of the physical meaning, the Softplus activation function is used in the output layer of Param-Net to force the output stiffness and damping values ​​to be non-negative, i.e. .

[0110] This dual-stream architecture transforms the traditional "black box" model into a "gray box" model, which utilizes the strong fitting ability of neural networks to approximate complex vibration signals, while retaining the interpretability of the physical model through explicit parameter output ports.

[0111] Step 3: Construct a physical dynamics model with learnable parameters.

[0112] Abandoning the traditional PINN approach of treating physical parameters as constants, a nonlinear two-degree-of-freedom bearing dynamics model based on varying operating conditions is established. Assuming the bearing system is a nonlinear vibration system comprising an inner ring, outer ring, and rolling elements, its governing equations are:

[0113] in, Indicates vibration displacement. For equivalent quality, The damping coefficient is time-varying. The time-varying stiffness coefficient, It is a nonlinear Hertzian contact force. External incentives.

[0114] It is important to note the definition of parameters. It is not merely a scalar quantity, but a function related to system states (such as rotational speed, temperature, and degree of aging). This technique does not directly pre-determine... and Instead of setting the value of , it is set as a variable to be learned. Specifically, a parameter evolution equation is introduced:

[0115] in, It is a learnable nonlinear mapping function. These are the parameters of the mapping function. This allows the coefficients in the physical equations to be dynamically adjusted according to the operating state of the fan, thus truly reflecting the changes in physical properties caused by changes in the lubricating oil film and clearance wear.

[0116] Step four: Design an adaptive loss function for the soft reward and punishment mechanism.

[0117] To address the issue of overly rigid physical constraints in traditional PINN, this technique proposes a composite loss function that includes "data-driven loss," "physical consistency loss," and "soft constraint regularization term."

[0118] The total loss function is defined as:

[0119] (1) Data-driven loss ): Used to measure the difference between network predictions and sensor measurements. Calculated only at sparse monitoring points. Expressed in mean squared error (MSE) form.

[0120] (2) Loss of physical consistency ( ): Calculated using automatic differentiation and the output of Param-Net Substitute into the dynamic equation and calculate the equation residuals. .

[0121]

[0122]

[0123] here, The number of collocation points, which can be randomly sampled throughout the time domain without corresponding measured labels.

[0124] (3) Soft reward and punishment mechanism ( ): Considering that simplified models cannot perfectly describe real systems, forced A value of zero can lead to overfitting. Therefore, a learnable slack variable is introduced. And a reward / punishment function based on physical priors.

[0125] Set a tolerance band for physical errors allowed by the project. .

[0126] When physical residual Less than At this time, a "soft reward" is given, the loss weight is reduced, and minor physical violations are allowed in exchange for better data fit; When physical residual Greater than At this time, a "strong penalty" is imposed, and the loss weight increases exponentially, forcing the network to revert to the physical manifold.

[0127] In mathematical expression, soft constraints are adopted in the following form:

[0128] in, It is a hyperparameter that controls the steepness of reward and punishment. This mechanism simulates the cognitive process of human experts: allowing for local, inexplicable perturbations, provided that the physical laws are generally valid.

[0129] Step 5: Multi-stage adaptive weighted training strategy.

[0130] To address the training difficulties (gradient competition) caused by inconsistent magnitudes of the terms in the loss function, a dynamic weight adjustment strategy is adopted for training.

[0131] Phase 1: Warm-up Training. Initialize weights. Use only The state estimation subnet is trained to quickly fit the low-frequency trend of the observed data, obtain a better initial value guess, and avoid getting trapped in local minima.

[0132] Phase Two: Physical constraints are introduced. Gradually increasing... The value of is set to enable parameter updates for Param-Net. At this point, the network begins searching for solutions that satisfy the dynamic equations while fitting the data.

[0133] Phase 3: Adaptive Balancing. A gradient norm balancing algorithm is introduced. At the end of each training epoch, the gradient norm balancing algorithm is calculated. and Regarding the gradient norm of network parameters. Dynamically adjusted. and This ensures that the magnitudes of the two gradients remain consistent.

[0134]

[0135] in It is the first The average gradient norm of the loss term. This ensures that in the later stages of training, physical constraints and data fitting can be optimized in sync, preventing one from dominating and causing the other to fail to decrease.

[0136] Step 6: Fault diagnosis and health status assessment.

[0137] After training, the samples to be diagnosed are input into the model. This technique does not directly use the classifier to output the fault category, but instead performs a dual diagnosis based on "physical violation degree" and "parameter anomaly drift".

[0138] (1) Residual-based diagnosis: Calculate the physical consistency loss value of the test sample after convergence. If the dynamic residual of a sample is still high after sufficient parameter adjustment, it indicates that the vibration mode of the sample has seriously deviated from the physical manifold of the normal bearing, and it is determined that a fault has occurred.

[0139] (2) Diagnosis based on parameter drift: Extracting the stiffness from the Param-Net output and damping Sequence. The stiffness of a normal bearing should be stable near its design value. When a bearing experiences spalling or cracking, the equivalent stiffness will undergo periodic abrupt changes or an overall decrease; when lubrication fails or wear is severe, the damping coefficient will fluctuate abnormally.

[0140] Constructing feature vectors .

[0141] Using Support Vector Data Description (SVDD) or a simple K-Means clustering algorithm, the health domain and the fault domain are divided in the feature space.

[0142] If the feature vector If the fault falls into the fault domain, it can be further distinguished as an inner ring fault (high characteristic frequency), an outer ring fault (low characteristic frequency), or a rolling element fault based on the magnitude of stiffness reduction and the energy distribution of the residual spectrum.

[0143] 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.

[0144] 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 training data acquisition module 201 is used to collect the original vibration signal of the wind turbine bearing and the operating condition data reflecting the operating status of the wind power generation equipment. The input sample construction module 202 is used to construct input samples that reflect the mapping relationship between the original vibration signal and the working condition data; The state estimation subnet prediction value generation module 203 is used to generate a state estimation subnet prediction value based on the input sample, which reflects the vibration displacement of the wind turbine bearing caused by the original vibration signal. The real-time physical parameter generation module 204 is used to generate real-time physical parameters based on the operating condition data, including the equivalent dynamic stiffness and equivalent damping parameters of the wind turbine bearing. The target loss function determination module 205 is used to determine the target loss function through the predicted values ​​of the state estimation subnet and the real-time physical parameters. Target model generation module 206 is used to train and generate a target model based on the target loss function; The detection result generation module 207 is used to output the detection result for the target wind turbine bearing based on the target data when target data containing target vibration signal and target working condition data is obtained, using the target model.

[0145] 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.

[0146] In addition, embodiments of the present invention also provide an electronic device, such as... Figure 3As 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.

[0147] The communication interface is used for communication between the aforementioned terminal and other devices.

[0148] 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.

[0149] 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.

[0150] 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.

[0151] This application embodiment also provides a chip, which includes a processor and a communication interface. The communication interface is coupled to the processor. The processor is used to run programs or instructions to implement the various processes of the above-described wind turbine bearing detection method embodiment and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0152] 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.

[0153] 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.

[0154] 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.

[0155] 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: Collect the original vibration signal of the wind turbine bearing, as well as the operating condition data reflecting the operating status of the wind power generation equipment; Construct input samples that reflect the mapping relationship between the original vibration signal and the operating condition data; Based on the input samples, generate state estimation subnet prediction values ​​that reflect the vibration displacement of the wind turbine bearing caused by the original vibration signal; Real-time physical parameters, including the equivalent dynamic stiffness and equivalent damping parameters of the wind turbine bearing, are generated based on the operating condition data. The target loss function is determined by the predicted values ​​of the state estimation subnet and the real-time physical parameters. The target model is generated by training based on the target loss function; When target data containing target vibration signals and target operating condition data is obtained, the target model is used to output the detection results for the target wind turbine bearing based on the target data.

2. The method according to claim 1, characterized in that, Before the step of generating the state estimation subnet prediction value reflecting the vibration displacement of the wind turbine bearing caused by the original vibration signal based on the input sample, the method further includes: A two-stream network is constructed, which consists of a state estimation subnet and a parameter adaptation subnet.

3. The method according to claim 2, characterized in that, The step of generating a state estimation subnet prediction value based on the input sample that reflects the vibration displacement of the wind turbine bearing caused by the original vibration signal includes: The input sample is input into the state estimation subnet, and the state estimation subnet is controlled to output the state estimation subnet prediction value that reflects the vibration displacement of the wind turbine bearing caused by the original vibration signal.

4. The method according to claim 3, characterized in that, The operating condition data includes at least real-time speed data and real-time power data. The step of generating real-time physical parameters based on the operating condition data, including the equivalent dynamic stiffness and equivalent damping parameters of the wind turbine bearing, includes: The operating condition data is input into the parameter adaptive subnet, and the parameter adaptive subnet is controlled to output real-time physical parameters including equivalent stiffness parameters and equivalent damping parameters under the current operating condition based on the real-time speed data and real-time power data.

5. The method according to claim 4, characterized in that, The step of determining the target loss function using the predicted values ​​of the state estimation subnet and the real-time physical parameters includes: A two-degree-of-freedom dynamic equation is established based on the geometry of the wind turbine bearing. The external load excitation term characterizing the external impact force borne by the wind turbine bearing in real time is calculated using the real-time power data. Calculate the derivative of the predicted value of the state estimation subnet, and substitute the derivative, the real-time physical parameters, and the external load excitation term into the two-degree-of-freedom dynamic equation. Determine the physical consistency loss by calculating the residuals on both sides of the two-degree-of-freedom dynamic equation. Calculate the mean square error between the predicted values ​​of the state estimation subnet and the original vibration signal to determine the data fitting loss; Based on the degree of overflow of the residual relative to the preset physical law violation tolerance band, and combined with learnable slack variables, a soft reward / penalty loss is constructed to characterize the degree of deviation of the model from the physical law. Establish a composite total loss function consisting of the data fitting loss, the physical consistency loss, and the soft reward / penalty loss.

6. The method according to claim 5, characterized in that, The step of training and generating the target model based on the target loss function includes: Set initial weighting coefficients for the data fitting loss, the physical consistency loss, and the soft reward / penalty loss; Obtain the initial gradient norm of each loss term with respect to the network parameters in the data fitting loss, the physical consistency loss, and the soft reward / penalty loss; Adjust the values ​​of the physical consistency loss and soft reward / penalty loss in the initial weighting coefficients to obtain the target weighting coefficients for the first stage; The dual-stream network is preheated and iterated using the first-stage target weighting coefficients to obtain the preheating network weights that initially capture the distribution pattern of the original vibration signal. Using the initial gradient norm as a baseline, the gradient energy ratio of each loss term in the current iteration is calculated in real time using the gradient norm balancing algorithm, and the second-stage real-time weighting coefficients are obtained by dynamically correcting them over time. Using the real-time weighting coefficients of the second stage and the weights of the preheating network, the composite total loss function is jointly optimized and iterated to obtain candidate network parameters whose physical residuals stably fall into the preset physical law violation tolerance zone; When the total composite loss value corresponding to the candidate network parameters is determined to reach a preset convergence threshold, the candidate network parameters are frozen and the target model is generated.

7. The method according to claim 5, characterized in that, The step of using the target model to output the detection results for the target wind turbine bearing based on the target data includes: Extract the target vibration sequence and target working condition features from the target data that match the dimension of the target model input layer to obtain the input tensor to be diagnosed; The input tensor to be diagnosed is input into the state estimation subnet of the target model to obtain a predicted vibration displacement sequence that reflects the real-time dynamic response of the target bearing fan. Input the target working condition features in the input tensor to be diagnosed into the parameter adaptive subnet of the target model to obtain the real-time stiffness prediction value and real-time damping prediction value under the current working condition. Calculate the derivative of the predicted vibration displacement sequence, and substitute the derivative, the real-time stiffness prediction value, the real-time damping prediction value, and the external load term deduced from the target working condition characteristics into the two-degree-of-freedom dynamic equation based on the geometric structure of the target wind turbine bearing to obtain the target residual distribution sequence. By comparing the deviation of the real-time stiffness prediction value and the real-time damping prediction value from the benchmark value, a parameter degradation evaluation index reflecting the abnormal drift trend of physical parameters is obtained. The mean and peak values ​​of the target residual distribution sequence on the time axis are statistically analyzed, and a quantitative evaluation is performed in conjunction with the violation tolerance band of the physical law to obtain a physical credibility score that reflects the degree to which the model's predictive behavior conforms to the physical law. By combining the parameter degradation evaluation index and the physical reliability score, the fault logic is determined, and a complete detection result including fault type, severity and diagnostic reliability is output.

8. A wind turbine bearing testing device for wind power generation equipment, characterized in that, include: The training data acquisition module is used to collect the original vibration signal of the wind turbine bearing and the operating condition data reflecting the operating status of the wind power generation equipment. An input sample construction module is used to construct input samples that reflect the mapping relationship between the original vibration signal and the operating condition data; The state estimation subnet prediction value generation module is used to generate state estimation subnet prediction values ​​that reflect the vibration displacement of the wind turbine bearing caused by the original vibration signal based on the input sample. A real-time physical parameter generation module is used to generate real-time physical parameters based on the operating condition data, including the equivalent dynamic stiffness and equivalent damping parameters of the wind turbine bearing. The target loss function determination module is used to determine the target loss function based on the predicted values ​​of the state estimation subnet and real-time physical parameters. The target model generation module is used to train and generate a target model based on the target loss function. The detection result generation module is used to output the detection results for the target wind turbine bearing based on the target data containing the target vibration signal and the target operating condition data when the target data is obtained.

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