Main bearing damage diagnosis method based on flow-thermal coupling dynamic topology and physical information graph autoencoder
By using a method based on a dynamic topology and physical information graph autoencoder with fluid-thermal coupling, the problem of lack of physical basis for features and false alarms in the damage diagnosis of offshore wind turbine main bearings is solved. This method enables effective capture of local overheating and adaptive early warning throughout the entire life cycle, improving the accuracy and robustness of diagnosis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUAZHONG UNIV OF SCI & TECH
- Filing Date
- 2026-05-06
- Publication Date
- 2026-07-24
AI Technical Summary
Existing technologies for diagnosing damage to offshore wind turbine main bearings suffer from several problems, including a lack of physical basis for characteristic features, false alarms under extreme operating conditions, difficulty in distinguishing fault signals from changes in normal operating conditions under varying operating conditions, and mismatch of thresholds throughout the entire life cycle. These issues lead to frequent false alarms and difficulty in predicting early damage to the wind turbine.
A method based on dynamic topology and physical information graph autoencoder with fluid-thermal coupling is adopted. Through one-dimensional residual time series feature extraction, dynamic fluid-thermal topology graph construction, combination of graph encoder and decoder, and constraint of physical consistency loss function, damage diagnosis of main bearing is realized, and the early warning threshold is dynamically adjusted to adapt to the degradation of the bearing throughout its entire life cycle.
It enables effective capture of local overheating chain reactions under extreme operating conditions, reduces false alarm rate, improves the robustness and engineering applicability of diagnostic methods, adapts to the damage early warning needs throughout the bearing's entire life cycle, and improves the accuracy of early fault identification.
Smart Images
Figure CN122451733A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of offshore wind power monitoring technology, and in particular relates to a method for diagnosing main bearing damage based on a fluid-thermal coupling dynamic topology and physical information graph autoencoder. Background Technology
[0002] As a critical load-bearing component of the transmission system, the main bearing of an offshore wind turbine operates under extreme environments such as variable operating conditions, strong impacts, and high humidity and salinity for extended periods. Its health directly impacts the safety and economic efficiency of the entire turbine. Currently, condition monitoring based on SCADA systems has become a mainstream low-cost maintenance method. In recent years, deep learning models such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory Networks (LSTMs) have been gradually introduced into main bearing fault diagnosis, attempting to achieve early warning by mining hidden patterns in multi-dimensional sensor data. Some studies have also attempted to construct Graph Neural Networks (GNNs) to model the spatial relationships between sensors and improve state reconstruction capabilities.
[0003] Existing methods for diagnosing damage to offshore wind turbine main bearings face the following key technical bottlenecks: First, to adapt to convolutional or cyclic network structures, multi-source sensor data is typically forced to be tiled or reshaped into a two-dimensional matrix, completely ignoring the actual physical spatial layout and thermal resistance network logic within the turbine. This results in the model extracting features lacking physical basis and making it difficult to capture chain reactions caused by local overheating. Second, purely data-driven models are highly dependent on the distribution of training data, which can easily produce "prediction illusions" that violate thermodynamic laws under extreme sea conditions, leading to false alarms or missed alarms. Third, existing models cannot perceive the dynamic fluctuations in Nusselt number and convective heat transfer coefficient caused by changes in main shaft speed in real time, resulting in severe residual fluctuations under varying operating conditions and making it difficult to distinguish between fault signals and normal operating condition changes. Finally, most diagnostic methods use fixed thresholds or statistical variance limits, which cannot adapt to the natural performance degradation caused by microscopic osmosis, increased surface roughness, etc., throughout the bearing's entire life cycle, leading to frequent false alarms in older turbines and difficulty in predicting early, minor damage in new turbines. Summary of the Invention
[0004] To address the aforementioned technical problems, this invention proposes a main bearing damage diagnosis method based on a fluid-thermal coupling dynamic topology and physical information graph autoencoder, thereby resolving the issues present in the prior art.
[0005] To achieve the above objectives, this invention provides a method for diagnosing main bearing damage based on a fluid-thermal coupling dynamic topology and physical information graph autoencoder, comprising: One-dimensional residual time-series features are extracted based on multi-source SCADA time-series data to obtain the time-series feature vectors of each node; The real-time Nusselt number is calculated based on the real-time angular velocity of the main shaft. The edge weights of the directed adjacency matrix are dynamically calculated based on the real-time Nusselt number, the temperature gradient gating function, and the thermal resistance parameters. Combined with the temporal feature vectors of each node, a dynamic heat flow topology graph is constructed. A graph encoder is used to spatially aggregate the features of neighboring nodes along the directed edges in the dynamic heat flow topology graph, and then a graph decoder is used for dimensionality reduction and reconstruction to obtain the predicted temperature of the main bearing at the next moment. Based on the mean square error between predicted and actual temperatures and the physical consistency loss constructed from the lumped parameter energy conservation equation, the total loss function is calculated and the model training is constrained to obtain the trained state reconstruction model. Based on the residual between the predicted temperature and the actual temperature output by the state reconstruction model after training, and the theoretical degradation factor calculated based on the bearing fatigue life equation, the warning threshold is dynamically adjusted, and a damage warning is triggered when the residual exceeds the warning threshold.
[0006] Optionally, the process of extracting one-dimensional residual time-series features from multi-source SCADA time-series data to obtain the time-series feature vectors of each node includes: The missing values in the multi-source SCADA time series data are filled in using a piecewise cubic Hermite interpolation polynomial to obtain complete time series data; Based on the complete time series data, the univariate sequence of each sensor node is normalized to obtain the normalized univariate sequence. A one-dimensional causal convolutional neural network is used to extract deep temporal features from each normalized univariate sequence. The input of the current layer is added to the output after two convolutions and ReLU activation through residual connections to obtain the temporal feature vector of each node.
[0007] Optionally, the expression for calculating the real-time Nusselt number based on the real-time angular velocity of the spindle is: ; In the formula, For feature size; The fluid's kinematic viscosity; It is the Prandtl number (a constant); For real-time Nusselt numbers; The real-time angular velocity of the main axis; t For the transient moment, is an empirical proportionality constant, m is the Reynolds number exponent, and n is the Prandtl number exponent.
[0008] Optionally, the process of dynamically calculating the edge weights of the directed adjacency matrix based on the real-time Nusselt number, temperature gradient gating function, and thermal resistance parameters, and constructing a dynamic heat flow topology graph by combining the temporal feature vectors of each node, includes: Based on the inherent structural thermal resistance between node i and node j, and the temperature components in the temporal characteristic vectors of node i and node j at time t, the temperature gradient gating function value is calculated. Calculate the convection enhancement factor based on the real-time Nusselt number; The physical prior weights are calculated based on the temperature gradient gating function value, the convection enhancement factor, and the reciprocal of the thermal conductivity resistance. Meanwhile, the Pearson correlation coefficient is calculated based on the temporal feature vectors of node i and node j, and used as the data-driven implicit association weight. The physical prior weights and the data-driven implicit association weights are weighted and summed using a preset physical prior proportion coefficient to obtain the edge weights of the directed adjacency matrix. A dynamic heat flow topology graph is constructed using sensor nodes as vertices and the edge weights as the strength of the directed edges.
[0009] Optionally, the process of using a graph encoder to process the features of neighboring nodes along the directed edges in the dynamic heat flow topology graph to obtain the predicted temperature of the main bearing at the next moment includes: The in-degree inverse matrix is calculated after adding self-loops to the directed adjacency matrix based on the dynamic heat flow topology graph. Matrix multiplication is performed based on the directed adjacency matrix and the temporal feature vectors of each node to obtain aggregated neighbor features; The hidden state of the nodes is obtained by performing matrix multiplication based on the in-degree inverse matrix and the aggregated neighbor features, multiplying it with the learnable weights and then performing nonlinear activation. The hidden variables of the main bearing node are obtained based on the hidden states of several nodes. Based on the latent variables of the main bearing nodes, a multilayer perceptron decoder is used for dimensionality reduction mapping to obtain the predicted temperature of the main bearing at the next moment.
[0010] Optionally, the process of obtaining the trained state reconstruction model includes: The mean square error between the predicted temperature and the actual temperature is calculated to obtain the data reconstruction loss; The heat storage term is calculated based on the predicted temperature change rate and the equivalent heat capacity of the main bearing. The mechanical heat generation power is calculated based on the spindle speed and friction torque. The convective heat dissipation power is calculated based on the convective heat transfer coefficient, heat dissipation area and the difference between the predicted temperature and the ambient temperature. The heat source term is obtained by subtracting the convective heat dissipation power from the mechanical heat generation power. The square of the L2 norm of the difference between the heat storage term and the heat source term is calculated to obtain the physical consistency loss. The weighted physical consistency loss is added to the data reconstruction loss to obtain the total loss function; Based on the total loss function, backpropagation is used to update the network parameters until convergence, resulting in the trained state reconstruction model.
[0011] Optionally, the process of dynamically adjusting the early warning threshold based on the residual between the predicted temperature and the actual temperature output by the reconstructed state model after training, and based on the theoretical degradation factor calculated from the bearing fatigue life equation, includes: The trained state reconstruction model is used to reconstruct historical health period data to obtain a health residual sequence, and the residual mean and residual standard deviation are calculated based on the health residual sequence. The trained state reconstruction model is used to reconstruct the SCADA data at the current moment to obtain the current predicted temperature. The current residual is calculated based on the current predicted temperature and the actual temperature, and the current residual is smoothed by exponential weighted moving average to obtain the smoothed residual. Based on the basic rated dynamic load of the bearing, the equivalent dynamic load under each load level and the corresponding actual number of operating cycles, the total cumulative damage is calculated using the L10 fatigue life equation and Miner linear cumulative damage theory. The theoretical decay factor is calculated based on the negative exponential function of the total cumulative damage. A dynamic early warning threshold is obtained by linearly combining the residual mean, the residual standard deviation, the mechanism decay factor, and the preset adaptive adjustment coefficient.
[0012] Optionally, the dynamic adjustment expression for the warning threshold is: ; In the formula, These represent the mean and standard deviation of the residuals under historical health data, respectively. These are the first adaptive adjustment coefficient and the second adaptive adjustment coefficient, respectively. This is the warning threshold; It is a mechanism of decline factor.
[0013] The present invention also provides a computer, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computer program.
[0014] The present invention also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described thereon.
[0015] Compared with the prior art, the present invention has the following advantages and technical effects: This invention embeds Nusselt number, temperature gradient gating function, and thermal resistance parameters into the graph topology construction process, enabling the dynamic heat transfer topology graph to respond in real time to fluctuations in convective heat transfer intensity caused by changes in spindle speed. This solves the problem of existing methods neglecting physical spatial layout and thermal resistance network logic, resulting in a lack of physical basis for features. A graph encoder aggregates neighbor node features along directed edges, combined with a graph decoder to reconstruct the predicted temperature, effectively capturing local overheating chain reactions. By transforming the lumped parameter energy conservation equation into a physically consistent loss constraint model for training, the prediction results are forced to follow thermodynamic laws, avoiding the "prediction illusion" that violates physical logic produced by purely data-driven models under extreme conditions, significantly reducing false alarms and missed alarm rates. Based on the bearing fatigue life equation, a theoretical degradation factor is calculated and the warning threshold is dynamically adjusted, achieving adaptive adjustment of residual tolerance throughout the entire life cycle. This resolves the contradiction between frequent false alarms in old fans and the difficulty in predicting minor damage in new fans, improving the robustness and engineering practicality of the diagnostic method at different service stages. Attached Figure Description
[0016] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of step 1 of an embodiment of the present invention; Figure 2 This is a schematic diagram of step 2 in an embodiment of the present invention; Figure 3 This is a schematic diagram of step 3 in an embodiment of the present invention; Figure 4 This is a schematic diagram of step 4 in an embodiment of the present invention; Figure 5 This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation
[0017] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.
[0018] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.
[0019] Example 1 The purpose of this invention is to overcome the problems of missing spatial topology expression, missing physical logic, and mismatch of thresholds throughout the entire life cycle in the existing technology. In particular, for the problems of high false alarm rate of abnormal detection and difficulty in identifying weak damage in offshore wind turbine main bearings under variable operating conditions, a main bearing damage diagnosis method based on flow-thermal coupling dynamic topology and physical information graph autoencoder is proposed.
[0020] This invention proposes a main bearing damage diagnosis method based on fluid-thermal coupling dynamic topology and physical information graph autoencoder. By embedding fluid mechanics and heat transfer mechanisms into the bottom layer of a neural network, a physical information-driven digital twin image is constructed, which effectively solves the technical problems of low accuracy and poor robustness in early fault identification of main bearings under complex working conditions.
[0021] like Figure 5 As shown, this embodiment provides a main bearing damage diagnosis method based on a fluid-thermal coupling dynamic topology and physical information graph autoencoder, including the following steps: Step S1: Extraction of one-dimensional residual temporal features from multi-source SCADA data (1D-ResNet).
[0022] like Figure 1 As shown, the SCADA system of the wind turbine records multi-dimensional time-series data such as wind speed, power, and temperature of various components. This data is characterized by low sampling frequency (typically a 10-minute average), large fluctuations across operating conditions, and completely different physical dimensions of each sensor. Directly performing cross-variable spatial fusion can easily cause "catastrophic interference" in the feature dimensions. Therefore, before establishing spatial physical connections, the data is first "purified" on the time axis to extract the dynamic evolution patterns of each sensor node.
[0023] S1-1 collects multi-source SCADA timing data from the acquisition unit.
[0024] To address missing values caused by sensor communication delays, a piecewise cubic Hermite interpolation polynomial (PCHIP) is used to fill in the missing values. This not only completes the data but also ensures the physical continuity of the first derivative of the sequence. S1-2 inputs each normalized univariate sequence independently into a one-dimensional residual neural network (1D-ResNet).
[0025] Deep temporal features are extracted using one-dimensional causal convolution. The formula for calculating a single layer is as follows: ; In the formula: The first Layer and first The univariate temporal feature vector output by the layer; It is a non-linear activation function; This is a one-dimensional causal convolution operation, designed to prevent future data from being leaked into historical features; It is a one-dimensional convolution kernel weight matrix; This is the bias vector.
[0026] Step S2: Dynamic topology mapping of fluid-thermal coupling modulated by Nusselt number.
[0027] like Figure 2 As shown, after step 1, the pure temporal characteristics of each isolated component (sensor node) within the nacelle are obtained. The key next step is to spatially connect these isolated nodes. Existing technologies often simply stack them into a mesh image, which completely violates the physical topology of the wind turbine. In reality, the heat transfer efficiency of the main bearing is not constant but dynamically affected by airflow disturbances within the nacelle (forced convection driven by the main shaft rotation). Therefore, this step applies the dimensionless criteria of fluid mechanics and heat transfer (Nusser number). This is directly injected into the bottom layer of the graph neural network, allowing the "connection weights" of the graph to "breathe" in real time as the wind turbine speed increases or decreases. Constructing a dynamic heat flow topology: ; The node set V represents the sensor features output from step 1. Define the weights of the directed physical adjacency matrix pointing from node i to node j. : ; Among them, the real-time Nusselt number characterizes the intensity of convective heat transfer. The calculation is as follows: ; In the formula: For a moment node To the node The directed heat transfer topology weights; For nodes The inherent structural thermal resistance (TDR) ); This is a temperature gradient gating function. It only gates if heat originates from a high temperature. Flowing to low temperature Only when the heat transfer edge weight is activated does it strictly follow the second law of thermodynamics; Main spindle real-time angular velocity ( The increase in rotational speed leads to an increase in the turbulent kinetic energy of the flow field, resulting in... Increase, thereby dynamically amplifying the heat transfer weight between network nodes; For feature size; The fluid's kinematic viscosity; It is the Prandtl number (a constant); Pearson correlation coefficient (as a supplement to purely data-driven implicit associations). This represents the physical prior proportion coefficient. Let t be the real-time angular velocity of the main axis, and t be time (transient moment). and It becomes a real-time (transient) quantity, suitable for dynamic thermal analysis scenarios where the spindle speed changes over time; This is an empirical scaling constant, referring to the dimensionless scaling coefficient (a purely mathematical constant, usually without units) preceding the entire correlation. Its sole function is to scale the Nusselt number globally. The numerical value of the formula ensures that the calculated result matches the experimental measurement value precisely on the order of magnitude. The value is determined entirely by fitting experimental data and has no independent physical meaning (e.g., it does not represent the Reynolds number, Prandtl number, time, or any geometric / physical property parameter). It is only used to compensate for other terms in the formula. and Experimental-specific effects such as geometry, surface roughness, and flow details that were not fully covered.
[0028] m is the Reynolds number exponent, referring to the rotational Reynolds number (m). The power exponent of m is a dimensionless exponent in a purely mathematical sense (usually without units). It quantitatively reflects the dependence of forced convection intensity on the Nusselt number: the larger the value of m, the more significant the effect of rotational flow on heat transfer enhancement. Typical values are: m≈0.5 in laminar flow, m≈0.6~0.7 in transition flow, and m≈0.8 in turbulent flow. The value of m is determined by experimental fitting and is used only for characterization. The nonlinear effects; n is the exponent of the Prandtl number, which refers to the Prandtl number ( The power exponent of is also a dimensionless exponent in a purely mathematical sense (usually without units). It quantitatively reflects the corrective effect of fluid thermal properties (the ratio of momentum diffusion to thermal diffusion) on heat transfer performance: n reflects the relative influence of viscosity and thermal conductivity on boundary layer thickness. Typical value range: n≈0.3~0.4 for liquid media, n≈0.3 for gaseous media. The value of n is also determined by experimental fitting and is only used to correct for the contribution of Pr.
[0029] Step S3: Physical Information Graph Neural Network Autoencoder (PIGNN-AE) State Reconstruction.
[0030] like Figure 3As shown, having obtained the temporal evolution characteristics of a single node (step 1) and the physical heat flow distribution channels between nodes (step 2), step 3 utilizes an algorithm capable of simultaneously processing graph topology and node attributes to establish a digital mirror image (normal behavior model) of the wind turbine in its "healthy state." A graph autoencoder is the optimal vehicle for performing this spatiotemporal fusion and state reconstruction.
[0031] Using a directed graph convolutional network as an encoder, the influence of surrounding heat source nodes is spatially aggregated towards the main bearing node along the dynamic directed edges established in step 2: ; In the formula: For graph networks The hidden state of the layer's nodes; The dynamic adjacency matrix for adding self-loops (output of step 2); This is the in-degree matrix; These are learnable weights.
[0032] The graph decoder reduces the dimensionality of the fused latent variables and outputs the predicted temperature of the main bearing at the next time step. .
[0033] Step S4: Constraints the physical loss function of the combined thermodynamic partial differential equation.
[0034] like Figure 4 As shown, although step 3 completes the prediction, a neural network that relies purely on data fitting will produce severe "illusions" and output abnormal temperature curves when faced with unseen extreme sea conditions (such as sudden internet outages or typhoons). To give the AI a "bottom line" to withstand unknown environments, this step forces the lumped parameter energy conservation equation to be written into the network's loss function. This is like setting a physical red line for the AI: no matter how distorted the data is, every degree of temperature rise predicted by the model must perfectly match the mechanical friction work and convective heat dissipation input to the system.
[0035] Define the total loss function .
[0036] in Mean squared error. Physical consistency loss. Defined as: ; In the formula: The equivalent heat capacity of the main bearing system ( ); The time-discrete derivative for AI temperature prediction; The main bearing friction torque calculated from SCADA load ( ); spindle speed ( The product of the two represents the mechanical heating power; The dynamic surface convective heat transfer coefficient and heat dissipation area are affected by the rotational speed; To predict bearing temperature and engine room ambient temperature.
[0037] Step S5 is based on Adaptive threshold early warning driven by fatigue mechanism decay factors.
[0038] Through the above four steps, a highly accurate residual (actual temperature - reconstructed temperature) is obtained. A sudden increase in the residual indicates potential damage. In industry, methods such as exponentially weighted moving average (EWMA) are often used to smooth the residual, supplemented by fixed threshold alarms.
[0039] However, as wind turbines age, microscopic wear inevitably occurs on the bearing surfaces, leading to increased temperature fluctuations during normal operation. Adhering rigidly to fixed thresholds would result in older wind turbines experiencing daily false alarms. Therefore, this step incorporates design principles from mechanical fatigue design specifications. The lifespan equation allows macroscopic mechanical lifespan loss to inversely adjust the microscopic AI early warning threshold, achieving "strict early warning for new air ducts and leniency for older air ducts."
[0040] S5-1 calculates and smooths the residuals between the predicted and actual temperatures to obtain the smoothed residuals. .
[0041] S5-2 combines Miner's linear cumulative damage theory with the ISO bearing fatigue life equation.
[0042] ; Computerized recession factor : ; In the formula: It is a non-linear decay factor; This is the scaling factor; For the bearing in the first equivalent dynamic load The actual number of loops executed; The basic rated dynamic load of the main bearing; The life index of roller bearings (usually taken as...) ).
[0043] S5-3 sets dynamic early warning threshold limits.
[0044] ; In the formula: The mean and standard deviation of the residuals under the health history data; This is the adaptive adjustment coefficient.
[0045] Diagnostic judgment: As the bearing ages ( (Reduce), and AI will automatically tighten the residual tolerance upper limit. When Furthermore, if the damage continues to exceed the confidence time window, an early damage warning will be triggered.
[0046] Existing patents for equipment monitoring based on graph neural networks mostly use constants for the weight matrices of node edges based on spatial Euclidean distance. This solution (step 2) hard-codes dimensionless criteria (Nu number and Reynolds number) and thermodynamic temperature gradients from fluid mechanics into graph weights. The model can perceive that "the faster the rotational speed, the stronger the convection, and the greater the heat transfer weight between nodes." This design, which directly sublimates the industrial fluid dynamics mechanism into graph network topology rules, is highly innovative.
[0047] Faced with unpredictable extreme ocean loads, pure deep learning is highly susceptible to logical breakdown. This solution (step 4) transforms the lumped-parameter heat conservation equation into a loss function penalty term ( This is like adding a "physical baseline" to AI, forcing every degree of temperature rise calculated by the neural network to be balanced with the mechanical friction work and convective heat dissipation input into the system, thus solving the "black box unexplainability" problem in the wind power industry.
[0048] This invention breaks away from relying on pure mathematical statistics (such as...) The traditional approach of setting alarm red lines (step 5). This involves the absolute authority in the field of mechanical design— The fatigue life equation is introduced as a priori factor for degradation, realizing a closed-loop industrial digital intelligence system that allows for "tolerance for errors in new bearings and strict monitoring of old bearings" throughout the entire life cycle, greatly ensuring the feasibility of on-site deployment.
[0049] The present invention also provides a computer, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computer program.
[0050] The present invention also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described thereon.
[0051] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A method for diagnosing main bearing damage based on a fluid-thermal coupling dynamic topology and physical information graph autoencoder, characterized in that, Includes the following steps: One-dimensional residual time-series features are extracted based on multi-source SCADA time-series data to obtain the time-series feature vectors of each node; The real-time Nusselt number is calculated based on the real-time angular velocity of the main shaft. The edge weights of the directed adjacency matrix are dynamically calculated based on the real-time Nusselt number, the temperature gradient gating function, and the thermal resistance parameters. Combined with the temporal feature vectors of each node, a dynamic heat flow topology graph is constructed. A graph encoder is used to spatially aggregate the features of neighboring nodes along the directed edges in the dynamic heat flow topology graph, and then a graph decoder is used for dimensionality reduction and reconstruction to obtain the predicted temperature of the main bearing at the next moment. Based on the mean square error between predicted and actual temperatures and the physical consistency loss constructed from the lumped parameter energy conservation equation, the total loss function is calculated and the model training is constrained to obtain the trained state reconstruction model. Based on the residual between the predicted temperature and the actual temperature output by the state reconstruction model after training, and the theoretical degradation factor calculated based on the bearing fatigue life equation, the warning threshold is dynamically adjusted, and a damage warning is triggered when the residual exceeds the warning threshold.
2. The main bearing damage diagnosis method based on a fluid-thermal coupling dynamic topology and physical information graph autoencoder according to claim 1, characterized in that, The process of extracting one-dimensional residual time-series features from multi-source SCADA time-series data to obtain the time-series feature vectors of each node includes: The missing values in the multi-source SCADA time series data are filled in using a piecewise cubic Hermite interpolation polynomial to obtain complete time series data; Based on the complete time series data, the univariate sequence of each sensor node is normalized to obtain the normalized univariate sequence. A one-dimensional causal convolutional neural network is used to extract deep temporal features from each normalized univariate sequence. The input of the current layer is added to the output after two convolutions and ReLU activation through residual connections to obtain the temporal feature vector of each node.
3. The main bearing damage diagnosis method based on a fluid-thermal coupling dynamic topology and physical information graph autoencoder according to claim 1, characterized in that, The expression for calculating the real-time Nusselt number based on the real-time angular velocity of the spindle is as follows: ； In the formula, For feature size; The fluid's kinematic viscosity; It is the Prandtl number (a constant); For real-time Nusselt numbers; The real-time angular velocity of the main axis; t For the transient moment, is an empirical proportionality constant, m is the Reynolds number exponent, and n is the Prandtl number exponent.
4. The main bearing damage diagnosis method based on a fluid-thermal coupling dynamic topology and physical information graph autoencoder according to claim 3, characterized in that, The process of dynamically calculating the edge weights of the directed adjacency matrix based on the real-time Nusselt number, temperature gradient gating function, and thermal resistance parameters, and combining this with the temporal feature vectors of each node, to construct a dynamic heat flow topology graph includes: Based on the inherent structural thermal resistance between node i and node j, and the temperature components in the temporal characteristic vectors of node i and node j at time t, the temperature gradient gating function value is calculated. Calculate the convection enhancement factor based on the real-time Nusselt number; The physical prior weights are calculated based on the temperature gradient gating function value, the convection enhancement factor, and the reciprocal of the thermal conductivity resistance. Meanwhile, the Pearson correlation coefficient is calculated based on the temporal feature vectors of node i and node j, and used as the data-driven implicit association weight. The physical prior weights and the data-driven implicit association weights are weighted and summed using a preset physical prior proportion coefficient to obtain the edge weights of the directed adjacency matrix. A dynamic heat flow topology graph is constructed using sensor nodes as vertices and the edge weights as the strength of the directed edges.
5. The main bearing damage diagnosis method based on a fluid-thermal coupling dynamic topology and physical information graph autoencoder according to claim 1, characterized in that, The process of using a graph encoder to process the features of neighboring nodes along the directed edges of the dynamic heat flow topology graph to obtain the predicted temperature of the main bearing at the next moment includes: The in-degree inverse matrix is calculated after adding self-loops to the directed adjacency matrix based on the dynamic heat flow topology graph. Matrix multiplication is performed based on the directed adjacency matrix and the temporal feature vectors of each node to obtain aggregated neighbor features; The hidden state of the nodes is obtained by performing matrix multiplication based on the in-degree inverse matrix and the aggregated neighbor features, multiplying it with the learnable weights and then performing nonlinear activation. The hidden variables of the main bearing node are obtained based on the hidden states of several nodes. Based on the latent variables of the main bearing nodes, a multilayer perceptron decoder is used for dimensionality reduction mapping to obtain the predicted temperature of the main bearing at the next moment.
6. The main bearing damage diagnosis method based on a fluid-thermal coupling dynamic topology and physical information graph autoencoder according to claim 1, characterized in that, The process of obtaining the trained state reconstruction model includes: The mean square error between the predicted temperature and the actual temperature is calculated to obtain the data reconstruction loss; The heat storage term is calculated based on the predicted temperature change rate and the equivalent heat capacity of the main bearing. The mechanical heat generation power is calculated based on the spindle speed and friction torque. The convective heat dissipation power is calculated based on the convective heat transfer coefficient, heat dissipation area and the difference between the predicted temperature and the ambient temperature. The heat source term is obtained by subtracting the convective heat dissipation power from the mechanical heat generation power. The square of the L2 norm of the difference between the heat storage term and the heat source term is calculated to obtain the physical consistency loss. The weighted physical consistency loss is added to the data reconstruction loss to obtain the total loss function; Based on the total loss function, backpropagation is used to update the network parameters until convergence, resulting in the trained state reconstruction model.
7. The main bearing damage diagnosis method based on a fluid-thermal coupling dynamic topology and physical information graph autoencoder according to claim 1, characterized in that, The process of dynamically adjusting the early warning threshold based on the residual between the predicted temperature and the actual temperature output by the trained state reconstruction model, and based on the theoretical degradation factor calculated from the bearing fatigue life equation, includes: The trained state reconstruction model is used to reconstruct historical health period data to obtain a health residual sequence, and the residual mean and residual standard deviation are calculated based on the health residual sequence. The trained state reconstruction model is used to reconstruct the SCADA data at the current moment to obtain the current predicted temperature. The current residual is calculated based on the current predicted temperature and the actual temperature, and the current residual is smoothed by exponential weighted moving average to obtain the smoothed residual. Based on the basic rated dynamic load of the bearing, the equivalent dynamic load under each load level and the corresponding actual number of operating cycles, the total cumulative damage is calculated using the L10 fatigue life equation and Miner linear cumulative damage theory. The theoretical decay factor is calculated based on the negative exponential function of the total cumulative damage. A dynamic early warning threshold is obtained by linearly combining the residual mean, the residual standard deviation, the mechanism decay factor, and the preset adaptive adjustment coefficient.
8. The main bearing damage diagnosis method based on a fluid-thermal coupling dynamic topology and physical information graph autoencoder according to claim 7, characterized in that, The dynamic adjustment expression for the warning threshold is: ； In the formula, These represent the mean and standard deviation of the residuals under historical health data, respectively. These are the first adaptive adjustment coefficient and the second adaptive adjustment coefficient, respectively. This is the warning threshold; It is a mechanism of decline factor.
9. A computer comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in claim 1.
10. A storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in claim 1.