Thrust bearing tile temperature prediction method based on sgmd decomposition and se reconstruction combined with avoabi lstm / informer

Abstract

The thrust bearing tile temperature prediction method based on SGMD decomposition and SE reconstruction combined with AVOABiLSTM / Informer comprises the following steps: collecting original signals of thrust bearing guide tile temperature and temperature-related influence factor data; preprocessing the original data; applying SGMD to the obtained temperature sequence to obtain a plurality of initial single-component sequences; calculating sample entropy SE of each initial single-component obtained; based on a preset similarity threshold, superimposing and reconstructing similar single components to obtain a plurality of reconstructed modal components; dividing the reconstructed modal components into high-complexity components and low-complexity components according to the sample entropy values of the reconstructed modal components; inputting the high-complexity components into BiLSTM for training and prediction, and inputting the low-complexity components into an Informer model for training and prediction; using AVOA to optimize key hyperparameters of the BiLSTM and the Informer model; and fusing the obtained component prediction results in a point-by-point superposition manner to obtain the final guide tile temperature prediction value. The method can obtain high-precision guide tile temperature prediction results.

Description

Technical Field

[0001] This invention relates to the field of hydropower unit operation status prediction and intelligent monitoring, specifically to a method based on SGMD decomposition and SE reconstruction combined with AVOA. BiLSTM / Informer method for predicting thrust bearing bearing temperature. Background Technology

[0002] As a crucial component of hydroelectric generators, bearings bear axial forces and rotating parts. The temperature of the guide bearing not only directly reflects the bearing's lubrication and cooling status but also reveals potential problems in the lubrication system, cooling water system, unit load fluctuations, and abnormal vibrations. Guide bearing temperature signals typically exhibit non-stationarity, multi-scale characteristics, strong coupling, and nonlinear abrupt changes. Single threshold monitoring or single prediction models struggle to simultaneously capture short-term abrupt changes and long-term trends, resulting in insufficient prediction accuracy and robustness. Existing research often employs traditional statistical or single machine learning models, such as BP and SVR, or combined models based on simple decomposition methods. However, these methods often suffer from insufficient feature extraction, reliance on experience for parameter tuning, and inadequate consideration of abrupt changes and long-term dependencies when dealing with complex coupled engineering signals.

[0003] To improve prediction accuracy, researchers have recently proposed combining signal decomposition with machine learning models. This involves breaking down complex signals into several modal components through multi-scale decomposition, and then applying appropriate modeling methods to different components. However, key challenges in engineering applications remain: how to merge similar components after decomposition to reduce computational complexity, how to select appropriate models based on component complexity, and how to globally optimize model hyperparameters. Furthermore, the noise resistance and fidelity of the decomposition method itself directly affect subsequent modeling results. Therefore, this paper proposes a combined prediction scheme that balances decomposition accuracy, component complexity discrimination, differentiated modeling, and global parameter optimization. This scheme can better address the multi-scale and nonlinear characteristics of thrust bearing guide plate temperature signals, thereby providing reliable data support for unit operation safety and condition-based maintenance. Summary of the Invention

[0004] To address the issue of insufficient prediction accuracy caused by the multi-scale, nonlinear, and strongly coupled characteristics of thrust bearing guide bearing temperature signals, this invention provides a method based on symplectic geometric mode decomposition (SGMD) and sample entropy (SE) reconstruction, combined with African vulture optimization. A BiLSTM / Informer method for predicting thrust bearing bearing temperature. This method divides complex temperature signals into components of varying complexity through multi-scale decomposition and sample entropy-driven component reconstruction. It then employs BiLSTM and Informer models to specifically target these components, and uses AVOA for global optimization of key hyperparameters. Finally, it fuses the prediction results of each component point-by-point to obtain a high-precision prediction of the bearing bearing temperature.

[0005] The technical solution adopted in this invention is as follows:

[0006] Based on SGMD decomposition and SE reconstruction, combined with AVOA The BiLSTM / Informer thrust bearing bearing temperature prediction method includes the following steps: Step 1: Collect raw temperature signals of the thrust bearing guide pads and data on temperature-related influencing factors; preprocess the raw data; Step 2: Apply symplectic geometric mode decomposition (SGMD) to the temperature sequence obtained in Step 1 to obtain several initial single-component sequences; Step 3: Calculate the sample entropy (SE) for each initial single component obtained in Step 2, and reconstruct similar single components by superimposing them based on a preset similarity threshold to obtain several reconstructed modal components; Step 4: Based on the sample entropy values ​​of each reconstructed modal component, divide them into high-complexity components and low-complexity components; Step 5: Input the high-complexity components into the Bidirectional Long Short-Term Memory (BiLSTM) network for training and prediction, and input the low-complexity components into the Informer model for training and prediction; Step 6: The key hyperparameters of the Bidirectional Long Short-Term Memory Network (BiLSTM) and the Informer model in Step 5 are optimized using the African Vulture Optimization Algorithm (AVOA). Step 7: Combine the predicted results of each component by point-by-point superposition to obtain the final predicted value of the guide tile temperature.

[0007] In step 1, the collected influencing factors include the following 13 items: Thrust bearing oil level in the oil sump and thrust external circulation oil inlet pipe temperature; Unit speed, active power, and head; Vibration of the lower frame in the X, Y, and Z directions; Thrust external circulation main oil pipe pressure; cooling oil flow rate, Total outlet flow rate of technical water supply cooling water; pressure and thrust of technical water supply pump outlet pipe; inlet temperature of external circulating cooling water; The selection of influencing factors is based on the analysis of the heat generation and heat dissipation mechanisms of the thrust bearing system. This mechanism analysis includes the relationship between oil film frictional heat generation and heat removal by lubricating oil / cooling water, and uses this to plot the influence of heat generation-related factors and heat dissipation-related factors on the guide bearing temperature; specifically including: The heat generation comes from two sources: first, the oil film formed between the thrust bearing and the sliding rotor, and the heat generated by the friction between the thrust bearing and the sliding rotor and the intermediate oil film during their relative movement. Q 1; Secondly, the oil film molecules generate heat through mutual friction in a laminar flow state.Q 2. Under laminar flow conditions and with an approximately linear velocity profile, the two values ​​are the same; adding them together yields the frictional heat generated by the bearing per unit time. P 0: (1); Under empirical correction, it is written as follows: (2); In formula (2): λ This refers to the viscosity coefficient of the lubricating oil. v N The linear velocity of the sliding rotor. P For the unit pressure of the guide tile, C These are the unit's factory parameters; This shows that heat source loss P 0 and λ , v N , P There is a positive correlation. As the lubricating oil temperature increases, the viscosity coefficient decreases, and the generation of oil film frictional heat changes accordingly. Therefore, the oil level in the thrust bearing oil groove and the temperature of the thrust external circulation oil inlet pipe are selected to reflect the changes in lubricating oil temperature and viscosity. The drastic fluctuations in unit speed and load (active power, head) during start-up and shutdown directly affect v N Due to friction conditions, the unit speed, active power, and head are included in the model. In addition, the vibration of the lower frame in the X, Y, and Z directions will change the unit pressure on the friction surface. P Therefore, triaxial vibration is taken as the influencing factor of vibration characteristics; The above eight parameters together characterize the generation mechanism of oil film frictional heat.

[0008] Heat dissipation occurs through two pathways: one is the heat carried away by the lubricating oil through oil circulation. Q 油 ; Q 油 Heat dissipation through pipes Q 管 Oil tank heat dissipation Q 槽 and thrust external circulation cooling water heat dissipation Q 水 ; Secondly, the heat dissipated by the sliding rotor and guide bearings. Q 滑 , Q 导 ; Q 管 , Q 滑 , Q 导, Q 槽 Heat is dissipated through radiation to the surroundings. 空 and heat dissipation through soaking lubricating oil Q 油 .

[0009] According to thermodynamic formulas, the heat carried away by a fluid per unit time... P With traffic Q P The following relationship exists: (3); In formula (3): C P Δ is the specific heat capacity of the fluid; T For temperature rise; γ Let be the fluid density. Equation (3) shows that... Q 油 and Q 水 All of them are positively correlated with their fluid flow rate.

[0010] Given that the lubricating oil circulation intensity is affected by pipeline pressure, the thrust external circulation total oil pipe pressure and cooling oil flow rate are used to evaluate the oil cooling heat dissipation capacity. Meanwhile, in order to comprehensively characterize the heat exchange performance of the water cooling system, the total outlet flow rate of the technical water supply cooling water, the outlet pipe pressure of the technical water supply pump, and the inlet temperature of the thrust external circulation cooling water were selected. These parameters together reflect the influence of cooling water flow rate, pressure, and inlet temperature on the oil heat exchange efficiency.

[0011] The above five factors fully cover the two-stage heat dissipation process of lubricating oil cooling and cooling water cooling.

[0012] In step 1, in order to ensure the input quality of subsequent SGMD decomposition and model training, the original monitoring data must be systematically preprocessed.

[0013] First, outlier detection and removal are performed on the original sequence using the interquartile range (ICM) method: ICM calculation. If the original observed values x satisfy or Then it is marked as an exception. IQR It is four parts of the positional distance; Q 1 、Q 3 represents the first and third quartiles, respectively.

[0014] Outliers can be replaced with nearest-neighbor valid values ​​or by interpolation to maintain sequence continuity. For imputation of missing values, linear interpolation is used for short-term consecutive missing values: (4); In equation (4): This is the estimated value after interpolation or filling. t、 1 t 2 represents the interpolation endpoint time.

[0015] For long-term or periodic deficiencies, the daily average value under the same working conditions can be used to fill the gaps. (5); In formula (5): Representing history M At the same time τ The observed values. To adapt to different modeling granularities, it is necessary to resample and aggregate the high-frequency raw data, and the aggregated values ​​after resampling. : (6); In formula (6): t i To fall into the resampling window T The original moment, n The aggregation operator can be set to the mean or median to enhance its robustness against anomalies, representing the number of samples within the window.

[0016] Finally, all numerical features are standardized to facilitate neural network training by using Min-Max normalization: (7); In equation (7): where Normalized values; The minimum and maximum values ​​used for normalization provide high-quality input data for subsequent decomposition and modeling.

[0017] In step 2, several initial single-component sequences, i.e., initial SGC components, are obtained through phase space reconstruction, symplectic geometric matrix transformation, and eigenvalue decomposition to achieve multi-scale separation of the original non-stationary signal. Specific implementation steps include: S2.1: Phase Space Reconstruction: For a given original time series input signal x ={ x 1, x 2,… x n}, where n represents the length of the signal data; the trajectory matrix is ​​constructed according to the Takens embedding theorem. X for: (8); In equation (8): d For the embedding dimension; τ For delay time; m = n -( d -1) τ This indicates that the time series is mapped to... m Dimensional vector.

[0018] S2.2: Symplectic geometric matrix transformation: For the trajectory matrix X, perform autocorrelation analysis to obtain the covariance matrix. A = X T X ,use A Construct the Hamilton matrix M : (9); In equation (9): A T Let A be the transpose of matrix A.

[0019] make N = M 2 According to the definition, N Also known as the Hamiltonian matrix, construct a symplectic orthogonal matrix. Q : (10); In formula (10): B It is an upper triangular matrix. R It is a symmetric matrix; At this time the matrix A eigenvalues σ i for: (11); In equation (11): λ i It is a matrix B The eigenvalues ​​are denoted by d, where d is the embedding dimension.

[0020] Will σ i Sort in descending order to obtain the corresponding feature vectors. Q i Then reconstruct the trajectory matrix Z A series Z i Constructed, that is: (12); In equation (12): Z i = Q i S i Let be the initial single-component matrix, 1≤ i ≤ d ;S i = Q i T X T This is the matrix of change coefficients; Q i T , X T They represent Q i and X The transpose of .

[0021] S2.3: Diagonal averaging: Initial single component Z i (1≤ i ≤ d The diagonal averaging method can be used to transform it into a set of lengths. n The time series, therefore d Group length is n The sum of the time series is the original time series. x The expression for the diagonal mean transformation method is: (13); In equation (13): It is a one-dimensional time series; Embedding dimension d Conjugate; Time series are mapped to m-dimensional conjugates; This is the summation index in the diagonal transformation matrix. Z i The elements in are defined as z ij ,1≤ i ≤ m, 1≤ j ≤ d ; d = min ( m , d ); m = max ( m , d ), m < d hour, z ij = z ij ,otherwise z ij =z ji , z ij express z ij Conjugate; z ji Representation matrix Z i The element after transpose.

[0022] This gives us a new set of lengths. n The time series, i.e., the reconstruction matrix Z The reconstructed independent components were obtained by diagonal averaging. d Each reconstructed independent component is shown in equation (14) below: (14); In equation (14): Represents the initial symplectic geometric components; Y i Represents the initial single signal component, 1≤ i ≤ d .

[0023] In step 3, each initial single component obtained in step 2 is denoted as SGC. i (0) (1≤ i ≤ n ), length is N The sample entropy (SE) is calculated to measure component complexity. Based on a preset sample entropy similarity threshold, similar components are superimposed and reconstructed to obtain several reconstructed modal components, preserving important scale information while reducing the number of components and computational burden. Specifically, this includes: Sample entropy is defined as: (15); In equation (15): r For similarity tolerance; B m ( r ), B m+1 ( r ) are SGC in r Next Match m points and m +1 point probability; If the difference or similarity of their sample entropy (SE) meets a preset threshold, the two components are superimposed and reconstructed into one component; after reconstruction, the sample entropy (SE) is calculated again, and the component is divided into high complexity and low complexity categories according to the classification threshold. Specifically: For any two components SGC a(0) SGC b (0) Calculate its sample entropy SE a SE b Define the difference Given a similarity threshold SE threshold ,when At that time, the two components are added together point by point according to the time interval to form a new component. ; The final reconstruction K Each component Calculate the entropy SE of each sample according to equation (15). i Given a classification threshold SE class_threshold Using a discriminant: (16); And can be pressed SE Sort by size from largest to smallest and then renumber them, so that the high-complexity components are... The low-complexity component is .

[0024] In step 4, the high-complexity components mainly include short-term mutations and high-frequency components, such as... Figure 1 The decomposition and recombination results shown in the SGMD-SE-based diagram indicate that the three components SGC1, SGC2, and SGC3 are high-complexity components; the low-complexity components mainly exhibit stable or periodic trends and long-range dependencies. For example... Figure 1 As shown, the SGC4 component is a low-complexity component.

[0025] In step 5, the high-complexity components are input into a bidirectional long short-term memory network (BiLSTM) for training and prediction in order to capture local dynamics and short-term mutations. The hyperparameters of the Bidirectional Long Short-Term Memory (BiLSTM) network include: number of hidden layer neurons, number of network layers, number of fully connected layer neurons, random inactivation rate, and learning rate. Low-complexity components are input into the Informer model for training and prediction to efficiently capture long-range dependencies and overall trends.

[0026] The hyperparameters of the Informer model include: embedding dimension, number of attention heads, number of network layers, number of neurons in fully connected layers, random inactivation rate, and learning rate.

[0027] In step 6, the African Vulture Optimization Algorithm (AVOA) is used to globally optimize the key hyperparameters of BiLSTM and Informer from step 5. The AVOA algorithm iteratively updates the population within a predefined search space and selects the optimal parameter combination through fitness evaluation to improve the model's generalization ability and prediction accuracy.

[0028] The hyperparameter optimization process of the African Vulture Optimization Algorithm (AVOA) includes the following steps: a. Initialization phase: Set relevant parameters and initialize the vulture population within the problem search space; b. Calculate fitness: Evaluate the fitness values ​​of individuals in the population and identify the best and second-best performing individuals; c. Selecting the individual's movement direction: Based on the individual's state and the target optimization direction, decide whether the vulture will approach the best or second-best individual to simulate its dynamic foraging behavior: (17); In equation (17): R i t For the first t During the nth iteration, the population is at the th... i The vulture moves towards either the best or second-best vulture. Bestv 1 t , Bestv 2 t The first t The optimal and second-best individuals in the next iteration; L 1. L 2∈(0,1), where 2 is a parameter given before the search and the sum of the two parameters is 1; P i The selection probability of an individual in the current population; t This represents the current iteration number.

[0029] d. Calculate the current hunger level of individuals in the population: (18); In equation (18): F i t The vulture's hunger level; r 1 t A random number within the interval (0,1); z t and h t A random number within the interval (-1, 1); t This represents the current iteration number; T This represents the maximum number of iterations. w Users can define parameters to control the probability of the algorithm entering the exploration phase; e. Based on hunger level F i t To determine the stage an individual has entered: when Fi t When the temperature is ≥1, vultures are in a state of satiation and tend to fly to greater distances in search of food; When | F i t When |≤1, the AVOA algorithm enters the development stage. In this stage, if 0.5≤| F i t If |≤1, the algorithm enters development phase 1, employing an encirclement strategy or a rotational flight strategy to perform a fine-grained search of the target region; if | F i t If |≤0.5, then proceed to Development Phase 2, employing a centralized or aggressive acquisition strategy to accelerate the optimization process. Depending on the phase, individuals adopt corresponding position update strategies to ensure the algorithm's optimization effectiveness.

[0030] f. Iterative optimization: Repeat steps b to e until the maximum number of iterations is reached or the optimization termination condition is met, and output the final optimization result.

[0031] In step 7, each component from step 5 is predicted separately to obtain the predicted value of each component at each prediction time. At each time point, the predicted values ​​of each component are summed point by point to obtain the final predicted value of the conductor tile temperature. The details are as follows: For each reconstructed component, its corresponding model is used to make predictions, resulting in predictions at each time step. t component prediction values Final predicted value The result is obtained by point-by-point superposition and fusion: (19); In equation (19): k This represents the total number of reconstructed components.

[0032] Finally, the fusion results are inversely normalized back to the original temperature scale, and evaluation indicators are calculated. The selected evaluation indicators are: (20); (twenty one); (twenty two); (twenty three); in: The number of samples; The actual value; These are the model's predicted values; The mean of the true values, i.e. .

[0033] This invention provides a method based on SGMD decomposition and SE reconstruction combined with AVOA. The BiLSTM / Informer thrust bearing bearing temperature prediction method has the following technical advantages: 1) This invention effectively separates the different scale components of the temperature signal through multi-scale decomposition of SGMD and component reconstruction driven by SE, thereby reducing mode aliasing and improving the signal-to-noise ratio of subsequent modeling.

[0034] 2) This invention adopts differentiated modeling based on component complexity (BiLSTM handles high-complexity components and Informer handles low-complexity components), which can simultaneously capture short-term mutations and long-term dependencies, and improve the ability to model complex coupled dynamics.

[0035] 3) This invention introduces AVOA to perform global optimization of key hyperparameters, reducing reliance on manual parameter tuning and enhancing the model's generalization ability and robustness.

[0036] 4) This invention uses a point-by-point superposition fusion strategy to achieve lossless synthesis of component prediction. The method is simple to implement, highly repeatable, and easy to deploy in engineering.

[0037] 5) In the verification of actual monitoring data, the present invention significantly reduced the prediction error and improved the fitting degree (in the embodiments, the RMSE, MAE, MAPE and R² indicators are all better than the traditional single model and several ablation models), providing reliable data support for thrust bearing condition assessment and preventive maintenance. Attached Figure Description

[0038] The present invention will be further described below with reference to the accompanying drawings and examples; Figure 1 This is a decomposition and recombination diagram based on SGMD-SE.

[0039] Figure 2 Create a flowchart for the SGMD-SE-AVOA-BiLSTM / Informer model.

[0040] Figure 3 This is a schematic diagram of the thrust bearing lubricating oil cooling system.

[0041] Figure 4 This is a schematic diagram of the BiLSTM structure.

[0042] Figure 5 This is a schematic diagram of the Informer structure.

[0043] Figure 6 A comparison chart of the prediction results of each model.

[0044] Figure 7 This is a comparison chart of the errors of each model.

[0045] Figure 3 middle: . Detailed Implementation

[0046] Based on symplectic geometric mode decomposition (SGMD) and sample entropy (SE) reconstruction, combined with African vulture optimization A BiLSTM / Informer method for predicting thrust bearing guide tile temperature. This method first preprocesses the original temperature signal and its influencing factors. Then, it uses symplectic geometrical mode decomposition (SGMD) combined with sample entropy (SE) to achieve multi-scale decomposition and single-component reconstruction. Based on component complexity, the reconstructed components are divided into high-complexity and low-complexity categories. High-complexity components are modeled using a bidirectional long short-term memory network (BiLSTM), while low-complexity components are modeled using Informer. The key hyperparameters of both models are globally optimized using the African Vulture Optimization Algorithm (AVOA). Finally, the predicted values ​​of each component are fused using a point-by-point superposition method to obtain the predicted guide tile temperature. Examples show that this invention significantly outperforms traditional single-model and several ablation models in terms of RMSE, MAE, MAPE, and R², effectively improving the accuracy of thrust bearing guide tile temperature prediction and providing a basis for condition-based maintenance. The implementation process is as follows: Figure 2 As shown, it includes the following steps: Step 1: Data Acquisition and Preprocessing The raw temperature signal of the thrust bearing guide pad and data of temperature-related influencing factors were collected. The raw data were processed, including outlier detection and removal, missing value imputation, resampling and normalization with unified dimensions, to provide high-quality input data for subsequent decomposition and modeling.

[0047] Step 2: SGMD Multiscale Decomposition: The preprocessed temperature sequence obtained in step 1 is subjected to symplectic geometric mode decomposition (SGMD). Through phase space reconstruction, symplectic geometric matrix transformation and eigenvalue decomposition, several initial single-component sequences (initial SGC components) are obtained to achieve multi-scale separation of the original non-stationary signal.

[0048] Step 3: Sample Entropy Calculation and Component Reconstruction (SE Reconstruction): For each initial single component obtained in step 2, the sample entropy (SE) is calculated to measure the component complexity. Based on the preset sample entropy similarity threshold, similar components are superimposed and reconstructed to obtain several reconstructed modal components, which retains important scale information while reducing the number of components and computational burden.

[0049] Step 4: Component complexity partitioning: Based on the sample entropy values ​​of each modal component after reconstruction, the components are divided into two categories: high-complexity components and low-complexity components. High-complexity components mainly contain short-term mutations and high-frequency components, while low-complexity components mainly exhibit stable or periodic trends and have long-range dependencies.

[0050] Step 5: Differentiated Modeling with BiLSTM and Informer: High-complexity components are fed into a Bidirectional Long Short-Term Memory (BiLSTM) network for training and prediction to capture local dynamics and short-term mutations; low-complexity components are fed into an Informer model for training and prediction to efficiently capture long-term dependencies and overall trends.

[0051] Step 6: Optimize AVOA hyperparameters: The African Vulture Optimization Algorithm (AVOA) is used to globally optimize the key hyperparameters of BiLSTM and Informer in step 5. AVOA iteratively updates the population within a predefined search space and selects the optimal parameter combination through fitness evaluation to improve the model's generalization ability and prediction accuracy.

[0052] Step 7: Component prediction fusion (point-by-point overlay): For each component in step 5, a prediction is made separately to obtain the predicted value of each component at each prediction time. At each time step, the predicted values ​​of each component are summed point by point to obtain the final predicted value of the conductor tile temperature. .

[0053] (1) Data sources and preprocessing: The data used in this embodiment comes from a mixed-flow reversible water pump. An online monitoring system for hydro-generator units. The original guide vane temperature is collected at a 1-second frequency. To balance noise filtering and trend preservation, the original data is resampled every 5 minutes to obtain modeling samples (example: 4896 sampling points), and then divided into training, validation, and test sets in an 8:1:1 ratio. Data preprocessing includes: outlier detection and removal (based on interquartile range or empirical thresholds), linear interpolation for short-term missing values, and imputation of long-term missing values ​​using daily averages of adjacent units under the same operating conditions; for numerical variables, Min... Max normalization is used to unify the units of measurement and facilitate neural network training. All measurement data are recorded with timestamps and measurement locations for easy mapping to mechanistic factors later.

[0054] (2) Derivation and selection of influencing factors: To enhance the predictive model's physical interpretability and improve its generalization ability, this invention selects 13 key influencing factors based on the heat generation and dissipation mechanisms of the guide vane. These factors cover two major processes: oil film frictional heat generation and heat removal by oil / cooling water. For example... Figure 3As shown, the heat generation mainly comes from two sources: first, the oil film formed between the thrust bearing and the sliding rotor, and the heat generated by the friction between the thrust bearing and the sliding rotor and the intermediate oil film during their relative movement. Q 1; Secondly, the oil film molecules generate heat through mutual friction in a laminar flow state. Q 2. Under laminar flow conditions and with an approximately linear velocity profile, the two values ​​are the same; adding them together yields the frictional heat generated by the bearing per unit time: (1); Under empirical correction, it can be written as: (2); in λ This refers to the viscosity coefficient of the lubricating oil. v N The linear velocity of the sliding rotor. P For the unit pressure of the guide tile, C These are the unit's factory parameters; it can be seen that the heat source loss... P 0 and λ , v N , P There is a positive correlation. As lubricating oil temperature increases, its viscosity coefficient decreases, and the generation of oil film frictional heat changes accordingly. Therefore, the oil level in the thrust bearing oil sump and the temperature of the thrust external circulation inlet pipe are selected to reflect the changes in lubricating oil temperature and viscosity. The drastic fluctuations in speed and load (active power, head) during unit start-up and shutdown directly affect… v N Due to friction conditions, unit speed, active power, and head are incorporated into the model; furthermore, the vibration of the lower frame in the X, Y, and Z directions will change the unit pressure on the friction surface. P Therefore, triaxial vibration is used as the vibration characteristic influencing factor. The above eight parameters together characterize the generation mechanism of oil film friction heat.

[0055] like Figure 3 As shown, heat dissipation mainly occurs through two pathways: one is the heat carried away by the lubricating oil through oil circulation. Q 油 ; Q 油 Heat dissipation through pipes Q 管 Oil tank heat dissipation Q 槽 and thrust external circulation cooling water heat dissipation Q 水 Secondly, the heat dissipated by the sliding rotor and guide bearings. Q 滑 , Q 导 ; Q 管 , Q 滑 ,Q 导 , Q 槽 Heat is dissipated through radiation to the surroundings. 空 and heat dissipation through soaking lubricating oil Q 油 According to thermodynamic formulas, the amount of heat carried away by a fluid per unit time... P With traffic Q P The following relationship exists: (3); In the formula: C P Δ is the specific heat capacity of the fluid; T For temperature rise; γ Let be the fluid density. It can be seen from equation (3) that... Q 油 and Q 水 All of these factors are positively correlated with the fluid flow rate. Given that the lubricating oil circulation intensity is affected by pipeline pressure, the total oil pipe pressure of the thrust external circulation and the cooling oil flow rate are used to evaluate the oil cooling heat dissipation capacity. Simultaneously, to comprehensively characterize the heat exchange performance of the water cooling system, the total outlet flow rate of the technical supply cooling water, the outlet pipe pressure of the technical supply water pump, and the inlet temperature of the thrust external circulation cooling water are selected. These parameters collectively reflect the influence of cooling water flow rate, pressure, and inlet temperature on the oil heat exchange efficiency. The above five factors comprehensively cover the two-stage heat dissipation process of lubricating oil cooling and cooling water cooling.

[0056] (3) Implementation steps of SGMD decomposition and SE reconstruction: To separate the multi-scale components of the temperature signal while preserving its intrinsic structure, symplectic geometric mode decomposition (SGMD) is used for multi-scale decomposition. The specific implementation steps are as follows: 3.1: Phase Space Reconstruction For a given original input signal x ={ x 1, x 2,… x n The trajectory matrix can be constructed based on the Takens embedding theorem. X for: (8); In the formula: d For the embedding dimension; τ For delay time; m = n -( d -1) λ .

[0057] 3.2: Symplectic geometric matrix transformation: definition A = X T X Construct the Hamilton matrix M : (9); make N = M 2 According to the definition, N Also known as the Hamiltonian matrix, construct a symplectic orthogonal matrix. Q : (10); In the formula: B For an upper triangular matrix, we have b ij =0( i > j +1); R It is a symmetric matrix.

[0058] At this time the matrix A eigenvalues σ i for: (11); Will σ i Sort in descending order to obtain the corresponding feature vectors. Q i Then reconstruct the trajectory matrix Z A series Z i Constructed, that is: (12); In the formula: Z i = Q i S i For the initial single-component matrix, S i = Q i T X T This is the matrix of change coefficients.

[0059] 3.3: Diagonal averaging: Initial single component Z i (1≤ i ≤ d Perform diagonal averaging to transform it into a set of lengths... nThe time series, therefore d Group length is n The sum of the time series is the original time series. x The diagonal mean transformation matrix is: (13); In the formula: d = min ( m , d ); m = max ( m , d ), m < d hour, z ij = z ij ,otherwise z ij = z ji .

[0060] This gives us a new set of lengths. n The time series, i.e., the reconstruction matrix Z The reconstructed independent components were obtained by diagonal averaging. d Each reconstructed independent component is as follows: (14); 3.4: Sample Entropy (SE) Calculation and Component Reconstruction The sample entropy is calculated for each initial component to measure complexity. Sample entropy is defined as follows: (15); In the formula: r For similarity tolerance; B m ( r ), B m+1 ( r ) are SGC in r Next Match m points and m +1 point probability. For any two initial components, if their SE difference or similarity meets a preset threshold (in the example, SE is taken as...), then... threshold If the value is 0.10, then the two components are superimposed and reconstructed into one component; after reconstruction, SE is calculated again and based on the classification threshold (in the example, SE is taken as 0.10). class_threshold=0.30) The components are divided into two categories: high complexity and low complexity. In the embodiment, three high complexity components and one low complexity component are finally obtained, and the SE values ​​of each component are listed in Table 1 for reference.

[0061]

[0062] 3.5: Key points of BiLSTM and Informer modeling: After reconstruction, the components are divided into two categories based on sample entropy: high complexity and low complexity. High complexity components typically contain short-term mutations and high-frequency components, making them suitable for capturing local dynamics using recurrent neural network models. Low complexity components typically exhibit stable or periodic trends, making them suitable for capturing long-range dependencies using self-attention-based long sequence models. Based on this, this invention uses BiLSTM to process high complexity components and Informer to process low complexity components, with both models trained and predicted in parallel.

[0063] BiLSTM: Figure 4 The diagram shows the structure of a BiLSTM, and the gating equations for the LSTM cells are given for reproduction: (twenty four); in: For the Gate of Oblivion For input gate, The output gate is BiLSTM. BiLSTM consists of forward and backward LSTMs, and the final output is a concatenation or weighted mapping of the hidden states in both directions. This bidirectional structure enables BiLSTM to extract local dynamics and long-range dependencies in a sequence more comprehensively, and is especially suitable for modeling highly complex signals, such as those with drastic fluctuations and short-term mutations.

[0064] Informer: Figure 5 This is a schematic diagram of the Informer architecture. The Informer is an improvement on the Transformer architecture, specifically designed for handling long-term dependencies and multi-scale time-series data. The core computation of self-attention is shown below: (25); Where Q represents the query term, K is used for matching, and V is the actual value. Informer adopts an Encoder-Decoder structure: the encoder extracts multi-scale features, and the decoder generates future predictions; by introducing a self-attention mechanism and a sparsity strategy, Informer significantly reduces the computational burden while capturing global information and long-range dependencies.

[0065] 3.6: AVOA Hyperparameter Optimization Process: like Figure 2 As shown, to reduce manual parameter tuning and improve model generalization ability, this invention uses the African Vulture Optimization Algorithm (AVOA) to globally optimize the key hyperparameters of BiLSTM and Informer. The hyperparameter search range table is shown in Table 3, and the specific process is as follows: a. Initialization phase: Set relevant parameters and initialize the vulture population within the problem search space. The search space is shown in Table 3 below, which lists the hyperparameter ranges.

[0066] b. Calculate fitness: Evaluate the fitness values ​​of individuals in the population and identify the best and second-best performing individuals.

[0067] c. Select the individual's movement direction: Based on the individual's status and the target optimization direction, decide whether the vulture will approach the best or second-best individual to simulate its dynamic foraging behavior.

[0068] (17); In the formula: R i t For the first t During the nth iteration, the population is at the th... i The vulture moves towards either the best or second-best vulture. Bestv 1 t , Bestv 2 t The first t The optimal and second-best individuals in the next iteration; L 1. L 2∈(0,1), where 2 is a parameter given before the search and the sum of the two parameters is 1; P i The selection probability of an individual in the current population (determined by roulette wheel). t This represents the current iteration number.

[0069] d. Calculate the current hunger level of individuals in the population: (18); In the formula: F i t The vulture's hunger level; r 1 t A random number within the interval (0,1); z t and h t A random number within the interval (-1, 1); t This represents the current iteration number; T The maximum number of iterations; wParameters defined by the user control the probability of the algorithm entering the exploration phase.

[0070] e. Based on hunger level F i t To determine the stage an individual has entered, when F i t When | ≥1, vultures are in a state of satiation and tend to fly farther to find food; when | F i t When |≤1, the AVOA algorithm enters the development stage. In this stage, if 0.5≤| F i t If |≤1, the algorithm enters development phase 1, employing an encirclement strategy or a rotational flight strategy to perform a fine-grained search of the target region; if | F i t If |≤0.5, then proceed to Development Phase 2, employing a centralized or aggressive acquisition strategy to accelerate the optimization process. Depending on the phase, individuals adopt corresponding position update strategies to ensure the algorithm's optimization effectiveness.

[0071] f. Iterative optimization. Repeat steps b through e until the maximum number of iterations is reached or the optimization termination condition is met. The final optimization results are shown in Tables 3 and 4 below.

[0072]

[0073]

[0074]

[0075] 3.7: Component Prediction, Point-by-Point Overlay Fusion and Output: For each reconstructed component, its corresponding model is used to make predictions, resulting in predictions at each time step. t component prediction values The final predicted value is obtained by point-by-point overlay and fusion: (19); in k To reconstruct the total number of components, point-by-point stacking is used in this embodiment to ensure the repeatability and simplicity of the method. Finally, the fusion result is back-normalized to the original temperature scale, and the evaluation index is calculated. The selected evaluation index is: (20); (twenty one); (twenty two); (twenty three); in The actual value; These are the model's predicted values; The mean of the true values, i.e. .

[0076] The prediction results of this invention are compared with those of traditional single prediction models and five coupled ablation models as follows: Figure 6 As shown in Table 5, the corresponding evaluation indicators are as follows.

[0077]

[0078] The results of the examples show that the method of the present invention can achieve RMSE≈0.5161℃, MAE≈0.4181℃, MAPE≈1.0319%, and R²≈0.9975 on the test set. The error graph between the actual and predicted values ​​of the present invention is shown below. Figure 7 As shown, it also intuitively demonstrates its advantages in capturing multi-scale, nonlinear dynamic features, significantly improving prediction accuracy and robustness.

Claims

1. Based on SGMD decomposition and SE reconstruction combined with AVOA The BiLSTM / Informer thrust bearing bearing temperature prediction method is characterized by... Includes the following steps: Step 1: Collect raw temperature signals of the thrust bearing guide pads and data on temperature-related influencing factors; preprocess the raw data; Step 2: Apply symplectic geometric mode decomposition (SGMD) to the temperature sequence obtained in Step 1 to obtain several initial single-component sequences; Step 3: Calculate the sample entropy (SE) for each initial single component obtained in Step 2, and reconstruct similar single components by superimposing them based on a preset similarity threshold to obtain several reconstructed modal components; Step 4: Based on the sample entropy values ​​of each reconstructed modal component, divide them into high-complexity components and low-complexity components; Step 5: Input the high-complexity components into the Bidirectional Long Short-Term Memory (BiLSTM) network for training and prediction, and input the low-complexity components into the Informer model for training and prediction; Step 6: The key hyperparameters of the Bidirectional Long Short-Term Memory Network (BiLSTM) and the Informer model in Step 5 are optimized using the African Vulture Optimization Algorithm (AVOA). Step 7: Combine the predicted results of each component by point-by-point superposition to obtain the final predicted value of the guide tile temperature.

2. The method based on SGMD decomposition and SE reconstruction combined with AVOA as described in claim 1 The BiLSTM / Informer method for predicting thrust bearing bearing temperature is characterized by: In step 1, the collected influencing factors include: Thrust bearing oil level in the oil sump and thrust external circulation oil inlet pipe temperature; Unit speed, active power, and head; Vibration of the lower frame in the X, Y, and Z directions; Thrust external circulation main oil pipe pressure; cooling oil flow rate, Total outlet flow rate of technical water supply cooling water; pressure and thrust of technical water supply pump outlet pipe; inlet temperature of external circulating cooling water; The selection of influencing factors is based on the analysis of the heat generation and heat dissipation mechanism of the thrust bearing system. The mechanism analysis includes the relationship between oil film frictional heat generation and heat removal by lubricating oil / cooling water, and the influence of heat generation-related factors and heat dissipation-related factors on the temperature of the guide bearing is plotted at this point. The heat generation comes from two sources: first, the oil film formed between the thrust bearing and the sliding rotor, and the heat generated by the friction between the thrust bearing and the sliding rotor and the intermediate oil film during their relative movement. Q 1; Secondly, the oil film molecules generate heat through mutual friction in a laminar flow state. Q 2; Under laminar flow conditions and with an approximately linear velocity profile, the two values ​​are the same, and their sum yields the frictional heat generated by the bearing per unit time. P 0: （1）； Under empirical correction, it is written as follows: （2）； In formula (2): λ This refers to the viscosity coefficient of the lubricating oil. v N The linear velocity of the sliding rotor. P For the unit pressure of the guide tile, C These are the unit's factory parameters; This shows that heat source loss P 0 and λ , v N , P There is a positive correlation; when the lubricating oil temperature increases, the viscosity coefficient decreases, and the generation of oil film frictional heat changes accordingly. Therefore, the oil level in the thrust bearing oil groove and the temperature of the thrust external circulation oil inlet pipe are selected to reflect the changes in lubricating oil temperature and viscosity. The drastic fluctuations in speed and load during unit start-up and shutdown directly affect v N Due to friction conditions, the unit speed, active power, and head are included in the model. In addition, the vibration of the lower frame in the X, Y, and Z directions will change the unit pressure on the friction surface. P Therefore, triaxial vibration is taken as the influencing factor of vibration characteristics; The above eight parameters together characterize the generation mechanism of oil film frictional heat.

3. The method based on SGMD decomposition and SE reconstruction combined with AVOA as described in claim 2 The BiLSTM / Informer method for predicting thrust bearing bearing temperature is characterized by: Heat dissipation occurs through two pathways: one is the heat carried away by the lubricating oil through oil circulation. Q 油 ; Q 油 Heat dissipation through pipes Q 管 Oil tank heat dissipation Q 槽 and thrust external circulation cooling water heat dissipation Q 水 ; Secondly, the heat dissipated by the sliding rotor and guide bearings. Q 滑 , Q 导 ; Q 管 , Q 滑 , Q 导 , Q 槽 Heat is dissipated through radiation to the surroundings. 空 and heat dissipation through soaking lubricating oil Q 油 ; According to thermodynamic formulas, the heat carried away by a fluid per unit time... P With traffic Q P The following relationship exists: （3）； In formula (3): C P Δ is the specific heat capacity of the fluid; T For temperature rise; γ Let be the fluid density; it can be seen from equation (3) that, Q 油 and Q 水 All of them are positively correlated with their fluid flow rate; Given that the lubricating oil circulation intensity is affected by pipeline pressure, the thrust external circulation total oil pipe pressure and cooling oil flow rate are used to evaluate the oil cooling heat dissipation capacity. Meanwhile, in order to fully characterize the heat exchange performance of the water cooling system, the total outlet flow rate of the technical water supply cooling water, the outlet pipe pressure of the technical water supply pump, and the inlet temperature of the thrust external circulation cooling water were selected. These parameters together reflect the influence of cooling water flow rate, pressure, and inlet temperature on the oil heat exchange efficiency. The above five factors fully cover the two-stage heat dissipation process of lubricating oil cooling and cooling water cooling.

4. The method based on SGMD decomposition and SE reconstruction combined with AVOA as described in claim 3 The BiLSTM / Informer method for predicting thrust bearing bearing temperature is characterized by: In step 1, in order to ensure the input quality of subsequent SGMD decomposition and model training, the original monitoring data must be systematically preprocessed. First, outlier detection and removal are performed on the original sequence using the interquartile range (ICM) method: ICM calculation. If the original observed values x satisfy or Then mark it as an exception; in IQR It is four parts of the positional distance; Q 1 、Q 3 represents the first and third quartiles, respectively; Marked outliers can be replaced with nearby valid values ​​or by interpolation to maintain sequence continuity; for missing values, linear interpolation is used for short-term consecutive missing values. （4）； In equation (4): This is the estimated value after interpolation or filling. t、 1 t 2 represents the interpolation endpoint time; For long-term or periodic deficiencies, the daily average value under the same working conditions is used to fill the gaps. （5）； In formula (5): Representing history M At the same time τ The observed values; to adapt to different modeling granularities, the high-frequency raw data were resampled and aggregated, and the aggregated values ​​after resampling were... : （6）； In formula (6): t i To fall into the resampling window T The original moment, n This represents the number of samples within the window. Finally, all numerical features are normalized to a uniform scale to facilitate neural network training, using Min-Max normalization: （7）； In equation (7): where Normalized values; The minimum and maximum values ​​used for normalization.

5. The method based on SGMD decomposition and SE reconstruction combined with AVOA as described in claim 4 The BiLSTM / Informer method for predicting thrust bearing bearing temperature is characterized by: In step 2, several initial single-component sequences, i.e., initial SGC components, are obtained through phase space reconstruction, symplectic geometric matrix transformation, and eigenvalue decomposition. The specific implementation steps include: S2.1: Phase Space Reconstruction: For a given original time series input signal x ={ x 1, x 2,… x n }, where n represents the length of the signal data; the trajectory matrix is ​​constructed according to the Takens embedding theorem. X for: （8）； In equation (8): d For the embedding dimension; τ For delay time; m = n -( d -1) τ This indicates that the time series is mapped to... m dimensional vector; S2.2: Symplectic geometric matrix transformation: For the trajectory matrix X, perform autocorrelation analysis to obtain the covariance matrix. A = X T X ,use A Construct the Hamilton matrix M : （9）； In equation (9): A T This is the transpose of matrix A; make N = M 2 According to the definition, N Also known as the Hamiltonian matrix, construct a symplectic orthogonal matrix. Q : （10）； In formula (10): B It is an upper triangular matrix. R It is a symmetric matrix; At this time the matrix A eigenvalues σ i for: （11）； In equation (11): λ i It is a matrix B The eigenvalues ​​are denoted as d, where d is the embedding dimension. Will σ i Sort in descending order to obtain the corresponding feature vectors. Q i Then reconstruct the trajectory matrix Z A series Z i Constructed, that is: （12）； In equation (12): Z i = Q i S i Let be the initial single-component matrix, 1≤ i ≤ d ; S i = Q i T X T This is the matrix of change coefficients; Q i T , X T They represent Q i and X Transpose of; S2.3: Diagonal averaging: Initial single component Z i (1≤ i ≤ d Use the diagonal averaging method to transform it into a set of lengths. n The time series, therefore d Group length is n The sum of the time series is the original time series. x The expression for the diagonal mean transformation method is: （13）； In equation (13): It is a one-dimensional time series; Embedding dimension d Conjugate; Time series are mapped to m-dimensional conjugates; The summation index in the diagonal transformation matrix; Z i The elements in are defined as z ij ,1≤ i ≤ m, 1≤ j ≤ d ; d = min ( m , d ); m = max ( m , d ), m < d hour, z ij = z ij ,otherwise z ij = z ji , z ij express z ij Conjugate; z ji Representation matrix Z i The transposed element; This gives us a new set of lengths. n The time series, i.e., the reconstruction matrix Z The reconstructed independent components were obtained by diagonal averaging. d Each reconstructed independent component is shown in equation (14) below: （14）； In equation (14): Represents the initial symplectic geometric components; Y i Represents the initial single signal component, 1≤ i ≤ d .

6. The method based on SGMD decomposition and SE reconstruction combined with AVOA as described in claim 5 The BiLSTM / Informer method for predicting thrust bearing bearing temperature is characterized by: In step 3, each initial single component obtained in step 2 is denoted as SGC. i (0) ,1≤ i ≤ n , length is N Calculate the sample entropy (SE) to measure component complexity; Based on a preset sample entropy similarity threshold, similar components are superimposed and reconstructed to obtain several reconstructed modal components, specifically including: Sample entropy is defined as: （15）； In equation (15): r For similarity tolerance; B m ( r ), B m+1 ( r ) are SGC in r Next Match m points and m +1 point probability; If the difference or similarity of their sample entropy (SE) meets the preset threshold, the two components are superimposed and reconstructed into one component; after reconstruction, the sample entropy (SE) is calculated again, and the component is divided into two categories, high complexity and low complexity, according to the classification threshold.

7. The method based on SGMD decomposition and SE reconstruction combined with AVOA as described in claim 6 The BiLSTM / Informer method for predicting thrust bearing bearing temperature is characterized by: For any two components SGC a (0) SGC b (0) Calculate its sample entropy SE a SE b Define the difference Given a similarity threshold SE threshold ,when At that time, the two components are added together point by point according to the time interval to form a new component. ; The final reconstruction K Each component Calculate the entropy SE of each sample according to equation (15). i Given a classification threshold SE class_threshold Using a discriminant: （16）； And can be pressed SE Sort by size from largest to smallest and then renumber them, so that the high-complexity components are... The low-complexity component is .

8. The method based on SGMD decomposition and SE reconstruction combined with AVOA as described in claim 7 The BiLSTM / Informer method for predicting thrust bearing bearing temperature is characterized by: In step 5, the high-complexity components are input into a bidirectional long short-term memory network (BiLSTM) for training and prediction in order to capture local dynamics and short-term mutations. The hyperparameters of the Bidirectional Long Short-Term Memory (BiLSTM) network include: number of hidden layer neurons, number of network layers, number of fully connected layer neurons, random inactivation rate, and learning rate. Low-complexity components are input into the Informer model for training and prediction to efficiently capture long-range dependencies and overall trends. The hyperparameters of the Informer model include: embedding dimension, number of attention heads, number of network layers, number of neurons in fully connected layers, random inactivation rate, and learning rate.

9. The method based on SGMD decomposition and SE reconstruction combined with AVOA as described in claim 8 The BiLSTM / Informer method for predicting thrust bearing bearing temperature is characterized by: In step 6, the African Vulture Optimization Algorithm (AVOA) is used to globally optimize the key hyperparameters of BiLSTM and Informer in step 5. The African Vulture Optimization Algorithm (AVOA) iteratively updates the population within a preset search space and selects the optimal parameter combination through fitness evaluation to improve the generalization ability and prediction accuracy of the model. The hyperparameter optimization process of the African Vulture Optimization Algorithm (AVOA) includes the following steps: a. Initialization phase: Set relevant parameters and initialize the vulture population within the problem search space; b. Calculate fitness: Evaluate the fitness values ​​of individuals in the population and identify the best and second-best performing individuals; c. Selecting the individual's movement direction: Based on the individual's state and the target optimization direction, decide whether the vulture will approach the best or second-best individual to simulate its dynamic foraging behavior: （17）； In equation (17): R i t For the first t During the nth iteration, the population is at the th... i The vulture moves towards either the best or second-best vulture; Bestv 1 t , Bestv 2 t The first t The optimal and second-best individuals in the next iteration; L 1. L 2∈(0,1), where 2 is a parameter given before the search and the sum of the two parameters is 1; P i The selection probability of an individual in the current population; t This represents the current iteration number; d. Calculate the current hunger level of individuals in the population: （18）； In equation (18): F i t The vulture's hunger level; r 1 t A random number within the interval (0,1); z t and h t A random number within the interval (-1, 1); t This represents the current iteration number; T This represents the maximum number of iterations. w Users can define parameters to control the probability of the algorithm entering the exploration phase; e. Based on hunger level F i t To determine the stage an individual has entered: when F i t When the temperature is ≥1, vultures are in a state of satiation and tend to fly to greater distances in search of food; When | F i t When |≤1, the AVOA algorithm enters the development stage. In this stage, if 0.5≤| F i t If |≤1, the algorithm enters development phase 1, employing an encirclement strategy or a rotational flight strategy to perform a fine-grained search of the target region; if | F i t If |≤0.5, then proceed to development phase 2, adopting a centralized strategy or an aggressive acquisition strategy to accelerate the optimization process; depending on the different phases, individuals adopt corresponding position update strategies to ensure the algorithm optimization effect; f. Iterative optimization: Repeat steps b to e until the maximum number of iterations is reached or the optimization termination condition is met, and output the final optimization result.

10. The method based on SGMD decomposition and SE reconstruction combined with AVOA as described in claim 9 The BiLSTM / Informer method for predicting thrust bearing bearing temperature is characterized by: In step 7, each component from step 5 is predicted separately to obtain the predicted value of each component at each prediction time. At each time point, the predicted values ​​of each component are summed point by point to obtain the final predicted value of the conductor tile temperature. The details are as follows: For each reconstructed component, its corresponding model is used to make predictions, resulting in predictions at each time step. t component prediction values Final predicted value The result is obtained by point-by-point superposition and fusion: （19）； In equation (19): k The total number of reconstructed components; Finally, the fusion results are inversely normalized back to the original temperature scale, and evaluation indicators are calculated. The selected evaluation indicators are: （20）； （21）； （22）； （23）； in: The number of samples; The actual value; These are the model's predicted values; The mean of the true values, i.e. .

Images