A method for predicting the bearing runout trend of a hydro-generator unit

Abstract

This invention provides a method for predicting the swing trend of a hydro-generator bearing, comprising the following steps: S1, collecting swing data, operating condition parameters, and environmental variables of the hydro-generator, performing denoising and normalization operations to form a standardized feature dataset; S2, performing feature filtering, generating comprehensive features, constructing a three-dimensional input tensor, and transmitting it to a dynamic nested model; S3, capturing the long-term trend of swing data through a first-layer nested network, and capturing the short-term fluctuation characteristics of swing data through a second-layer nested network; S4, optimizing the weight combination of the two nested networks using a genetic algorithm to generate a swing trend prediction result; S5, combining with operating condition parameters, calculating a swing safety threshold based on dynamic correlation analysis; S6, comparing the safety threshold with the swing data to predict the current state of the bearing. This invention achieves accurate swing trend prediction by combining a dynamic nested model with multi-source feature extraction and weight optimization.

Description

Technical Field

[0001] This invention belongs to the field of bearing runout prediction technology, specifically relating to a method for predicting the runout trend of a hydro-generator bearing. Background Technology

[0002] With the increasing demand for electricity in modern times, hydro-generator units have become crucial equipment in hydropower generation, and their operational reliability directly affects the stability of the power system. However, due to the complex environment and various operating conditions during operation, abnormal changes in bearing runout can lead to equipment damage, reduced operating efficiency, or even shutdown accidents. Therefore, accurate prediction of bearing runout trends has become one of the key technologies to ensure the safe operation of hydro-generator units. Currently, existing runout monitoring technologies mostly rely on single-variable analysis or simple statistical models, making it difficult to capture the potential multidimensional nonlinear dynamic characteristics in runout data. This results in insufficient prediction accuracy and stability, especially under complex operating conditions, where prediction errors increase significantly.

[0003] Traditional methods often lack the ability to co-model the long-term trends and short-term fluctuations of swing data, and cannot effectively integrate the correlation between environmental variables, operating conditions, and swing data. In addition, existing technologies rely heavily on fixed parameters or single models to optimize prediction results, lacking the ability for dynamic adjustment and multi-level optimization. Summary of the Invention

[0004] The main objective of this invention is to provide a method for predicting the bearing swing trend of a hydro-generator set, addressing the aforementioned problems.

[0005] Therefore, the above-mentioned objective of the present invention is achieved through the following technical solution:

[0006] A method for predicting the bearing runout trend of a hydro-generator unit includes the following steps:

[0007] S1. Collect swing data, operating parameters and environmental variables of the hydro-generator unit, and perform noise reduction and normalization operations to form a standardized feature dataset.

[0008] S2. Perform feature filtering on the standardized feature dataset, generate comprehensive features, construct a three-dimensional input tensor, and transmit it to the dynamic nested model;

[0009] S3. In the dynamic nested model, the long-term trend of swing data is captured by the first layer of network nesting, and the short-term fluctuation characteristics of swing data are captured by the second layer of network nesting.

[0010] S4. Optimize the weight combination of the two nested networks using a genetic algorithm to generate swing trend prediction results;

[0011] S5. Combine the operating condition parameters and swing trend prediction results, use fuzzy logic reasoning for fuzzification processing, and calculate the swing safety threshold based on dynamic correlation analysis.

[0012] S6. Compare the safety threshold with the swing data to predict the trend of the bearing's current state.

[0013] While adopting the above technical solutions, the present invention may also adopt or combine the following technical solutions:

[0014] As a preferred technical solution of the present invention: in step S1, the swing data includes swing peak value, swing amplitude and swing frequency data.

[0015] As a preferred technical solution of the present invention: in step S1, the operating parameters include speed, active power and cooling water pressure.

[0016] As a preferred technical solution of the present invention: in step S1, the environmental variables include ambient temperature and water head height.

[0017] As a preferred embodiment of the present invention, step S1 further includes the following sub-steps:

[0018] S11. Use the swing data, operating condition parameters, and environmental variables as the original dataset;

[0019] S12. Convert the original dataset into a time series format, fill in the missing values ​​through time series interpolation, and form time series data;

[0020] S13. After wavelet denoising, the time series data is normalized and smoothed by sliding window, and the feature matrix is ​​generated by aligning the timestamps.

[0021] S14. Perform outlier removal and integrity checks on the standardized feature matrix, and output the formatted standardized feature dataset.

[0022] As a preferred embodiment of the present invention, step S2 further includes the following sub-steps:

[0023] S21. Use random forest to rank the variables in the standardized feature dataset by importance, and select the top contributors based on the average reduction in recursive splits. One candidate feature;

[0024] S22. Set a threshold for removing highly correlated features, filter candidate features based on Pearson coefficient and random forest importance ranking, and generate comprehensive features through weighted fusion and attention mechanism;

[0025] S23. Generate time series segments based on comprehensive features, and use multi-scale convolutional neural networks to construct three-dimensional input tensors and transmit them to a dynamic nested model.

[0026] As a preferred embodiment of the present invention, step S3 further includes the following sub-steps:

[0027] S31. Use the first-level network nesting and the second-level network nesting as the basic framework of the dynamic nesting model;

[0028] S32. In the first layer of nested network, the three-dimensional input tensor is received and decomposed into time-segment decomposition. Each node corresponds to a feature vector of a time segment.

[0029] S33. Calculate the spatial correlation between nodes based on the cosine similarity between feature vectors;

[0030] S34. Calculate the spatial dynamic offset of each node by weighted averaging of the spatial correlation of each node, represent each time segment as multiple feature vectors, and construct a graph structure.

[0031] S35. Traverse the nodes in the graph structure sequentially, calculate the edge weights between the nodes in the graph structure for each edge, update the edge weights to the graph structure, and generate a weighted graph structure.

[0032] S36. Apply a hierarchical topological attention mechanism to the weighted graph structure, dynamically adjust the feature weights of nodes through the edge weights between nodes, and generate the long-term trend feature vector for each time segment.

[0033] S37. Use the long-term trend feature vector of each time segment as the output of long-term trend modeling;

[0034] S38. In the second layer of the nested network, the fast Fourier transform is used to extract short-term characteristic variables, and the output of the long-term trend modeling is dynamically combined with the short-term characteristic variables to construct a joint feature matrix.

[0035] S39. Use multi-branch convolution to extract the high-frequency and low-frequency characteristics of short-term characteristic variables, and perform nonlinear characteristic transformation. Use the short-term characteristics after nonlinear characteristic transformation as the characteristic output of short-term fluctuations.

[0036] As a preferred embodiment of the present invention, step S4 further includes the following sub-steps:

[0037] S41. Use the time dimension dynamic weights, spatial dynamic offsets, and edge weights as the dynamic weights of the first layer of the network.

[0038] S42, Low-frequency characteristics Amplitude weight High frequency characteristics Amplitude weight Frequency parameters of low-frequency characteristics Phase parameters of low-frequency characteristics As multi-branch convolutional weights nested in the second layer of the network;

[0039] S43. Randomly generate an initial population. Each individual in the initial population is represented by a vector in the form of a weight combination of dynamic weights and multi-branch convolution weights in the two-layer network.

[0040] S44. Use a genetic algorithm to perform selection, crossover, and mutation operations;

[0041] S45. Repeated genetic algorithm: evaluate the fitness value of individuals in the population, retain the optimal solution until the iteration limit is reached, and apply the optimized genetic algorithm optimal solution to a two-layer nested network.

[0042] S46. The first layer adjusts the dynamic weights for long-term trend modeling, and the second layer optimizes the convolution weights for short-term volatility characteristics to generate swing trend prediction results.

[0043] As a preferred embodiment of the present invention, step S5 further includes the following sub-steps:

[0044] S51. Through fuzzy logic reasoning, the operating condition parameters and swing trend prediction results are divided into three categories of fuzzy linguistic variables, each corresponding to a corresponding membership function.

[0045] S52. Analyze the correlation between historical operating condition parameters and swing trend prediction results, and establish a fuzzy rule base;

[0046] S53. Fuzzyen the numerical values ​​in the comprehensive features by using membership functions to generate a fuzzy membership matrix;

[0047] S54. Take the fuzzy membership matrix as input, match and output fuzzy inference;

[0048] S55. Defuzzify using the centroid method and generate an initial safety threshold;

[0049] S56. Based on the operating condition parameters and swing trend prediction results, construct a dynamic correlation graph, evaluate the correlation between the features of the correlation graph through the Gaussian kernel function, and extract time-series correlation features;

[0050] S57. The temporal correlation features are fused with the initial safety threshold, and the swing safety threshold is generated by weighted combination.

[0051] As a preferred embodiment of the present invention, step S6 further includes the following sub-steps:

[0052] S61. Align the safety threshold and swing data by timestamp, calculate the degree of deviation of the swing data from the safety threshold, and generate a state deviation index.

[0053] S62. By integrating the deviation index of the fusion state with historical operating condition parameters, the bearing runout trend is predicted through dynamic correlation analysis, and intuitive risk warnings are generated.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] This invention employs a dynamic nested model to collaboratively model the long-term trend and short-term fluctuation characteristics of swing amplitude data, enhancing the ability to capture complex dynamic features. On one hand, through multimodal feature extraction, it fully explores the correlation between swing amplitude data, operating condition parameters, and environmental variables, forming a standardized feature dataset and overcoming the limitations of single-variable analysis methods. On the other hand, it utilizes a genetic algorithm to optimize the weight combination of the dynamic nested model, significantly improving its adaptability and prediction accuracy. Furthermore, through fuzzy logic reasoning and dynamic correlation analysis, this invention can dynamically calculate the swing amplitude safety threshold, achieving accurate prediction of the bearing condition of hydro-generator units and effectively reducing equipment risks caused by abnormal swing amplitude. Attached Figure Description

[0056] Figure 1 The flowchart shows the method for predicting the bearing runout trend of a hydro-generator set provided by the present invention.

[0057] Figure 2 This is a flowchart of step S1. Detailed Implementation

[0058] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0059] like Figure 1 As shown, a method for predicting the bearing runout trend of a hydro-generator unit specifically includes the following steps:

[0060] S1. Collect swing data, operating parameters, and environmental variables of the hydro-generator unit, perform noise reduction and normalization operations to form a standardized feature dataset. The specific process is as follows: Figure 2 As shown;

[0061] Swing data includes peak sway, sway amplitude, and sway frequency data.

[0062] Operating parameters include speed, active power, and cooling water pressure.

[0063] Environmental variables include ambient temperature and water head.

[0064] S11. Use the swing data, operating condition parameters, and environmental variables as the original dataset;

[0065] High-precision vibration sensors are installed at the generator set bearings to collect swing data, including peak value, amplitude, and frequency information. These sensors must possess high sensitivity and anti-interference capabilities to ensure data accuracy. Secondly, speed sensors, power meters, and pressure sensors are installed on the rotor, stator, cooling water pipes, and related devices of the hydro-generator. These sensors monitor the real-time rotor speed, stator active power output, and cooling water pressure changes, respectively. The data acquisition accuracy and stability of each sensor must meet the generator set's operational status monitoring requirements. For environmental variables, ambient temperature sensors and head height measurement devices are deployed in the generator set's operating area to record changes in ambient temperature and head height. All data acquisition devices must maintain a sampling frequency consistent with the generator set's operating parameter acquisition. The collected data is transmitted to edge computing devices via industrial communication protocols (such as Modbus or CAN bus) and stored as raw datasets in a unified timestamp format, laying the foundation for subsequent data preprocessing and analysis.

[0066] S12. Convert the original dataset into a time series format, fill in the missing values ​​through time series interpolation, and form time series data;

[0067] S13. After wavelet denoising, the time series data is normalized and smoothed by sliding window, and the feature matrix is ​​generated by aligning the timestamps.

[0068] Wavelet decomposition is performed on the time series of each time series variable in the time series data to output high-frequency and low-frequency components. The low-frequency components are retained and the high-frequency noise is suppressed by the soft thresholding method to reconstruct a denoised signal.

[0069] High-frequency components refer to the rapidly changing parts separated from the time-series signal of a time-series variable during wavelet decomposition. These rapid changes typically correspond to noise, sudden events, or short-term high-frequency dynamic characteristics in the original acquired data. High-frequency components mainly consist of fluctuations or changes within short time intervals in time-series signals such as swing data, operating condition parameters (e.g., speed, power), and environmental variables (e.g., temperature, head). Low-frequency components refer to the slowly changing parts retained from the time-series signal of a time-series variable after wavelet decomposition. These slowly changing changes represent the long-term trend or macroscopic characteristics of the data. Low-frequency components consist of low-frequency vibrations in signals such as swing data, operating condition parameters, and environmental variables, and typically reflect the main trends in swing characteristics, stable changes in operating conditions, and fundamental fluctuations in environmental conditions.

[0070] Decomposing a time series into high-frequency and low-frequency components helps distinguish between the main trend and short-term fluctuations in the signal. This method is advantageous in denoising, and is particularly suitable for processing complex industrial data. The low-frequency component consists of low-frequency vibrations in the time series signal, which typically represent the main trend or fundamental pattern of the data.

[0071] The denoised signal is uniformly scaled to a value using the Min-Max normalization method. The range is determined, and a sliding window smoothing operation is used to reduce the impact of extreme values ​​on the normalization results.

[0072] Scaling the signal to the range of [0,1] facilitates unified processing of variables with different dimensions. At the same time, the sliding window smoothing operation reduces the impact of extreme values ​​on the results and enhances the stability of subsequent calculation processes.

[0073] S14. Perform outlier removal and integrity checks on the standardized feature matrix, and output the formatted standardized feature dataset.

[0074] The presence of outliers can significantly affect the statistical properties of data, and removing outliers helps improve the accuracy of the model; integrity checks ensure the quality of data preprocessing.

[0075] S2. Perform feature filtering on the standardized feature dataset, generate comprehensive features, construct a three-dimensional input tensor, and transmit it to the dynamic nested model;

[0076] S21. Use random forest to rank the variables in the standardized feature dataset by importance, and select the top contributors based on the average reduction in recursive splits. One candidate feature;

[0077] This represents the number of features selected from all feature variables and ranked highest according to importance, used to determine the size of the core feature set for subsequent modeling.

[0078] Taking the bearing runout data of a hydro-generator unit as an example, assume the original dataset contains multiple variables such as rotational speed, active power, cooling water pressure, runout amplitude, runout frequency, ambient temperature, and head height. After denoising and normalizing the data, a random forest model is used to calculate the importance of these variables. It is found that runout amplitude and runout frequency contribute the most to the predicted runout trend, followed by cooling water pressure and active power, while ambient temperature and head height contribute less. Based on importance, the top-ranked variables are retained. (For example, the first four) variables, namely swing amplitude, swing frequency, cooling water pressure and active power, are used as candidate features for subsequent modeling.

[0079] Random forests use the average reduction in recursive splits as a metric to evaluate feature importance. This method is nonparametric and robust, effectively handling high-dimensional data and the interactions between complex features, providing data support for subsequent selection of candidate features.

[0080] S22. Set a threshold for removing highly correlated features, filter candidate features based on Pearson coefficient and random forest importance ranking, and generate comprehensive features through weighted fusion and attention mechanism;

[0081] The distribution of Pearson correlation coefficients between pairs of variables in the statistically standardized feature dataset is analyzed, and the central tendency of the correlation coefficients is observed using histograms or kernel density estimation methods. Then, based on the requirements of the feature selection task and the principle of redundancy control, a preliminary critical value is determined as a candidate threshold. Next, through cross-validation experiments, the changing trend of model performance under different candidate thresholds is evaluated, and the threshold that achieves the best balance between model accuracy and feature redundancy control is selected. Finally, the threshold for removing highly correlated features is fixed as the final criterion for removing highly correlated features, and is used as the basis for judgment in the subsequent feature selection process.

[0082] Before calculation The Pearson correlation coefficient between candidate features is used to determine whether the candidate features are retained based on the order of importance in the random forest. If the Pearson correlation coefficient between two candidate features is greater than the threshold for removing highly correlated features, the candidate features are retained. Otherwise, the candidate features are removed.

[0083] Traditional feature selection methods may not be able to simultaneously consider the nonlinear relationships and interactions of features. Random forests, by reducing the amount of recursive splitting, capture the importance ranking of features, thus solving the technical problem of quantifying feature importance.

[0084] When the Pearson correlation coefficient is less than or equal to the threshold for removing highly correlated features, all candidate features are retained.

[0085] The retained candidate features are weighted and summed using swing data, operating condition parameters and environmental variables respectively, and the weights are dynamically adjusted through an attention mechanism to generate comprehensive features.

[0086] Based on the importance ranking results of the random forest, initial weights are assigned to the retained candidate features. Then, according to the distribution characteristics of each feature in different scenarios, the importance scores of the features are dynamically calculated, and these importance scores are combined with the initial weights to obtain preliminary weighted results. Next, an attention mechanism is adopted to establish a correlation model of the weights of different features in the time series through global analysis of historical data, and the weight distribution of features is dynamically adjusted so that the weights can reflect the changing trend of the current data state. Finally, the adjusted weights are applied to the candidate features, and a weighted summation operation is performed item by item to generate comprehensive features as input for subsequent processing.

[0087] S23. Generate time series segments based on comprehensive features, and use multi-scale convolutional neural networks to construct three-dimensional input tensors and transmit them to a dynamic nested model.

[0088] The time-series data is segmented according to a set window length and step size using a sliding window method. Each data segment consists of feature vectors extracted during the sliding window's movement, ensuring that the time-series segments contain complete short-term and long-term information. Subsequently, the generated time-series segments are input into a multi-scale convolutional neural network. Multiple convolutional kernels with different receptive field sizes extract features from the time-series segments. Different scale kernels capture rapid changes in short periods and slow trends over long periods, generating multi-scale feature representations. Next, max pooling is used to reduce the dimensionality of the convolutional features, reducing redundant information while preserving important dynamic information about the time series. Finally, the extracted multi-scale features are uniformly normalized to eliminate the influence of feature scale differences, outputting a structured feature tensor that provides multi-dimensional, fine-grained time-series feature support for subsequent modeling.

[0089] The comprehensive features and multi-scale features are normalized, and the three-dimensional input tensor is constructed and transmitted to the dynamic nested model by aligning the timestamps.

[0090] S3. In the dynamic nested model, the long-term trend of swing data is captured by the first layer of network nesting, and the short-term fluctuation characteristics of swing data are captured by the second layer of network nesting.

[0091] S31. Use the first-level network nesting and the second-level network nesting as the basic framework of the dynamic nesting model;

[0092] S32. In the first layer of nested network, the three-dimensional input tensor is received and decomposed into time-segment decomposition. Each node corresponds to a feature vector of a time segment.

[0093] A sliding window is used to analyze the distribution of feature vectors in the time dimension and to calculate the dynamic weights in the time dimension.

[0094] By segmenting feature vectors using a sliding window, time series data is sliced ​​into fixed-length windows, each containing a certain number of continuous time segments. Statistical properties such as mean, variance, and rate of change of trend are calculated for the feature vector set within each window, thereby analyzing the distribution pattern of feature vectors in the time dimension. Based on this, dynamic weights related to time segments are assigned according to the degree of change of statistical properties within each window. For example, windows with drastic changes are assigned higher weights to highlight their impact on the overall dynamic characteristics, while windows with stable changes are assigned lower weights. This process achieves the calculation of dynamic weights in the time dimension.

[0095] S33. Calculate the spatial correlation between nodes based on the cosine similarity between feature vectors;

[0096] S34. Calculate the spatial dynamic offset of each node by weighted averaging of the spatial correlation of each node, represent each time segment as multiple feature vectors, and construct a graph structure.

[0097] The spatial correlation between nodes is calculated using cosine similarity, which reflects the degree of similarity between the feature vectors of different nodes. Next, the geometric distance between nodes is used as a weighting factor, and the influence weight of nearby nodes is enhanced by normalizing the distance using the inverse of the distance, while the contribution of distant nodes is weakened. Finally, the spatial correlation between nodes is multiplied by the corresponding weighting factor and summed using a weighted average method. After normalization, the spatial dynamic offset of the node is obtained, thus fully expressing the dynamic change characteristics of the spatial relationship between the node and its surrounding nodes.

[0098] S35. Traverse the nodes in the graph structure sequentially, calculate the edge weights between the nodes in the graph structure for each edge, update the edge weights to the graph structure, and generate a weighted graph structure.

[0099] The nodes in the graph structure are traversed sequentially, and the edge weights between nodes are calculated one by one based on the node's feature vector, time-dimensional dynamic weights, and spatial dynamic offsets. The expression is as follows:

[0100] ;

[0101] In the formula, Represents a node and nodes The edge weights between nodes and nodes They are two different nodes. and These correspond to index variables of two different nodes. Representing nodes in a graph structure eigenvectors, Representing nodes in a graph structure eigenvectors, This represents the smoothing factor in the Gaussian kernel function. Represents a node and nodes The Euclidean distance between the eigenvectors Represents nodes Relevant dynamic weights over the time dimension Represents nodes The relevant spatial dimension dynamic offset, Indicates the interval Integrate the sine function within the range. Representing the integral variable The differential;

[0102] Dynamic weights capture the correlation between time segments, improving the adaptability to long-term trends; the introduction of spatial dynamic offsets enhances the ability to represent the spatiotemporal relationships between nodes, providing a more detailed description for long-term trend modeling. Specifically, traditional methods only process static data and cannot dynamically adjust node weights, while this invention significantly improves the modeling accuracy of long-term trends by combining spatiotemporal characteristics.

[0103] S36. Apply a hierarchical topological attention mechanism to the weighted graph structure, dynamically adjust the feature weights of nodes through the edge weights between nodes, and generate the long-term trend feature vector for each time segment.

[0104] S37. Use the long-term trend feature vector of each time segment as the output of long-term trend modeling;

[0105] In the dynamic nested model, the primary goal of the first layer of nested networks is to capture the long-term trend of swing data. It receives time-series data, extracts stability features across time segments, and reveals the core dynamic patterns of the long-term trend. The long-term trend feature vector is the concrete output of this process, comprehensively representing the regularity of long-term changes across each time segment.

[0106] The long-term trend feature vector is a mathematical expression of the long-term changing trend extracted from the first layer of the nested network. It quantifies long-term behavioral patterns through dynamic weights and node correlations, and serves as a key input for subsequent steps. It dynamically combines with short-term characteristics, laying the foundation for global and local collaborative modeling. This vectorized expression not only preserves the stability of the long-term trend but also resolves the problem of trend ambiguity, facilitating a comprehensive analysis of the dynamic behavior of swings. Therefore, the long-term trend feature vector is a concrete expression of the long-term changing trend; the two are causally related and functionally connected, jointly ensuring the integrity and accuracy of the dynamically nested model.

[0107] S38. In the second layer of the nested network, the fast Fourier transform is used to perform frequency domain transformation on the time segment, capture the dynamic fluctuations in a short period of time, extract short-term characteristic variables, and dynamically combine the output of the long-term trend model with the short-term characteristic variables to construct a joint feature matrix.

[0108] The dynamic combination strategy enables collaborative modeling of short-term and long-term characteristics, compensating for the interaction relationships that may be overlooked when modeling a single characteristic. The multi-branch convolution design improves the accuracy of feature extraction across multiple time scales. Specifically, traditional methods typically model long-term or short-term characteristics separately, ignoring the dynamic relationships between them. This invention, through the joint feature matrix and multi-branch convolution, makes the modeling more comprehensive and the prediction results more accurate.

[0109] S39. Using multi-branch convolution to extract the high-frequency and low-frequency characteristics of short-term characteristic variables by corresponding different time scales to different branches in the joint feature matrix, and performing nonlinear characteristic transformation, the short-term characteristics after nonlinear characteristic transformation are used as the characteristic output of short-term fluctuations.

[0110] The nonlinear characteristic transformation of the high-frequency and low-frequency characteristics is expressed as follows:

[0111] ;

[0112] ;

[0113] ;

[0114] In the formula, Indicates time segment The output of short-term volatility characteristics, Index variables representing time segments, This represents the modified linear unit function. Represents the high-frequency characteristic function. Represents the low-frequency characteristic function. Indicates high frequency characteristics Amplitude weights, Indicates high frequency characteristics Smoothing factor, Represents the decay function. Indicates low-frequency characteristics Amplitude weights, Frequency parameters representing low-frequency characteristics Phase parameters representing low-frequency characteristics, Represents the sine function;

[0115] Frequency domain transformation decomposes the dynamic characteristics of time-series data into frequency components, enabling precise capture of short-term fluctuations. The design of nonlinear characteristic transformation further enhances the ability to represent nonlinear relationships, allowing for more accurate capture of complex dynamic changes during the fusion of high-frequency and low-frequency characteristics in time segments. Specifically, existing technologies largely rely on time-domain analysis, making it difficult to separate and model short-term dynamic characteristics. This invention solves the accuracy problem in short-term fluctuation modeling through frequency domain analysis and nonlinear transformation.

[0116] The primary objective of the second-layer nested network is to capture the short-term fluctuation characteristics of swing data, complementing the long-term trend modeling of the first-layer nested network. Through frequency domain transformation, multi-branch convolution, and nonlinear characteristic transformation, the second-layer nested network deeply mines the dynamic characteristics of short-term fluctuations, making the system more accurate in capturing both rapid changes and slow trends. Simultaneously, its combination with the first-layer nested network ensures a comprehensive integration of global and local dynamic characteristics, laying a solid foundation for accurate prediction of swing trends.

[0117] S4. Optimize the weight combination of the two nested networks using a genetic algorithm to generate swing trend prediction results;

[0118] S41. Use the time dimension dynamic weights, spatial dynamic offsets, and edge weights as the dynamic weights of the first layer of the network.

[0119] S42, Low-frequency characteristics Amplitude weight High frequency characteristics Amplitude weight Frequency parameters of low-frequency characteristics Phase parameters of low-frequency characteristics As multi-branch convolutional weights nested in the second layer of the network;

[0120] By using dynamic weights in the time dimension, spatial dynamic offsets, and edge weights as dynamic weights in the first-layer network, the correlation between time segments and the dynamic nature of spatial feature changes can be effectively captured. The optimization of the dynamic weights in the first layer, by integrating multi-dimensional feature information, provides fundamental support for accurate modeling of long-term trends. Simultaneously, low-frequency characteristics... Amplitude weight High frequency characteristics Amplitude weight Frequency parameters of low-frequency characteristics Phase parameters of low-frequency characteristics The multi-branch convolutional weights, nested in the second layer of the network, further refine the modeling of signal characteristics at different time scales. This weight division balances the importance of low-frequency smoothing characteristics and high-frequency fluctuation characteristics in the modeling, thereby improving the accuracy of swing trend prediction. This multi-branch convolutional weight design compensates for the problem of insufficient modeling of time series signal characteristics in traditional methods, enabling the model to more clearly identify the combined characteristics of long-term trends and short-term fluctuations in complex dynamic environments.

[0121] S43. Randomly generate an initial population. Each individual in the initial population is represented by a vector in the form of a weight combination of dynamic weights and multi-branch convolution weights in the two-layer network.

[0122] Traditional optimization methods struggle to simultaneously handle the complexity of high-dimensional dynamic weights and the nonlinear characteristics of multi-branch convolutions. Genetic algorithms, offering global search capabilities, help address this issue. Specifically, by generating the initial population using genetic algorithms, the randomness of weight initialization is avoided from impacting model performance. Optimization based on the fitness function further ensures the evaluation of the initial population's merits, making weight allocation more closely reflective of the actual data.

[0123] The fitness function based on mean squared error evaluates the quality of weight combinations and minimizes the error.

[0124] S44. Use a genetic algorithm to perform selection, crossover, and mutation operations based on the fitness value in the fitness function.

[0125] The fitness value of each individual's weight combination is calculated based on the fitness function to reflect its contribution to the objective optimization. Then, a selection operation is performed, choosing individuals with high fitness values ​​to have a higher probability of entering the next generation of the population. Next, a crossover operation is used to randomly swap the weight combinations of the selected individuals, generating new individuals to increase population diversity. Finally, a mutation operation is used to make small random adjustments to some of the individuals' weight values, further introducing randomness to avoid the algorithm getting trapped in local optima.

[0126] S45. Repeated genetic algorithm: evaluate the fitness value of individuals in the population, retain the optimal solution until the iteration limit is reached, and apply the optimized genetic algorithm optimal solution to a two-layer nested network.

[0127] In each iteration, the fitness value of each individual in the current population is recalculated, and they are sorted based on their fitness values. The individual with the highest fitness is retained as the current optimal solution. Simultaneously, selection, crossover, and mutation operations are performed to generate the next generation of the population, gradually updating the individual weight combinations to approximate the global optimum. This process continues until a set iteration limit is reached or the population fitness value converges, ultimately outputting the optimal solution for the weight combinations.

[0128] The optimal weight combination output by the genetic algorithm is distributed to the first and second layers of the network, respectively. In the first layer, the dynamic weights are adjusted to the values ​​corresponding to the optimal solution to optimize the modeling ability of long-term trends. In the second layer, the amplitude weights, frequency parameters, and phase parameters in the convolution weights are also updated to their optimal values ​​to further refine the extraction of short-term fluctuation characteristics. In this way, the weight combination optimized by the genetic algorithm is used to adjust the core parameters of both layers, thereby comprehensively improving the performance of both layers in swing trend prediction.

[0129] S46. The first layer adjusts the dynamic weights for long-term trend modeling, and the second layer optimizes the convolution weights for short-term volatility characteristics to generate swing trend prediction results.

[0130] Traditional modeling methods often struggle to achieve a dynamic balance when dealing with long-term and short-term characteristics. This method, however, employs layered optimization, making targeted adjustments for both long-term trends and short-term fluctuations, significantly improving prediction accuracy and adaptability. The first layer, through dynamic weight adjustments, more accurately captures the changing characteristics of long-term trends, ensuring a thorough understanding of long-term time series behavior. The second layer, by optimizing the convolutional weights for short-term fluctuations, becomes more sensitive to abnormal fluctuations within a short period, effectively enhancing the refined representation of short-term characteristics.

[0131] S5. Combine the operating condition parameters and swing trend prediction results, use fuzzy logic reasoning for fuzzification processing, and calculate the swing safety threshold based on dynamic correlation analysis.

[0132] S51. Through fuzzy logic reasoning, the operating condition parameters and swing trend prediction results are divided into three categories of fuzzy linguistic variables, each corresponding to a corresponding membership function.

[0133] The membership function employs a combination of triangular and Gaussian functions; fuzzy linguistic variables (such as high, medium, and low) are core concepts in fuzzy logic reasoning, used to map continuous numerical values ​​to semantically defined intervals, facilitating the expression of uncertainty and fuzziness in complex systems. The combination of triangular and Gaussian functions can respectively reflect the central tendency and smooth change characteristics of data. For example, triangular functions have definite sharp boundaries, suitable for expressing linear changes, while Gaussian functions provide a gentle transition, suitable for handling noise interference.

[0134] S52. Analyze the correlation between historical operating condition parameters and swing trend prediction results, and establish a fuzzy rule base;

[0135] The linear and nonlinear correlation between operating condition parameters and swing trend prediction results is assessed using Pearson correlation coefficient or mutual information in statistics, and highly correlated variable pairs are screened. Then, based on the selected variables, fuzzy clustering algorithms (such as C-means clustering) are applied to divide historical data into different fuzzy classes, and representative rules are extracted from the distribution characteristics of each class. Finally, fuzzy rule extraction algorithms (such as Apriori algorithm or fuzzy association rule mining) are used to generate a rule base, thereby constructing a fuzzy rule base covering a wide range of operating conditions to support the subsequent reasoning process.

[0136] S53. Fuzzyen the numerical values ​​in the comprehensive features by using membership functions to generate a fuzzy membership matrix;

[0137] By using fuzzy linguistic variables and membership functions to fuzzify the operating condition parameters and swing prediction results, complex numerical relationships are effectively transformed into easily processed fuzzy relationships, overcoming the limitations of traditional numerical computation in expressing fuzzy information.

[0138] S54. Take the fuzzy membership matrix as input, match and output fuzzy inference;

[0139] The input variables in the fuzzy membership matrix are mapped to the antecedents (conditional parts) in the fuzzy rule base. Then, the matching degree between the input variables and the antecedents of each rule in the rule base is calculated one by one using the max-min synthesis method to determine the optimal matching rule or multiple partial matching rules. Finally, based on the consequent (conclusion part) of the matching rule and its matching degree, multiple rules are fused using a weighted method to output the fuzzy inference result, thereby completing the mapping from the input fuzzy membership degree to the output fuzzy conclusion.

[0140] S55. Calculate fuzzy inference using the maximum-minimum synthesis method and defuzzify using the centroid method to generate an initial safety threshold.

[0141] S56. Based on the operating condition parameters and swing trend prediction results, construct a dynamic correlation graph, evaluate the correlation between the features of the correlation graph through the Gaussian kernel function, and extract time-series correlation features;

[0142] Structured graphs representing the relationship between operating parameters and swing trend predictions help capture the nonlinear and dynamic characteristics of complex systems.

[0143] S57. The temporal correlation features are fused with the initial safety threshold, and the swing safety threshold is generated by weighted combination.

[0144] Initial weights are assigned to time-series correlation features and initial safety thresholds. Based on the statistical characteristics of historical operating data and the current swing trend prediction results, the weight allocation is optimized through a dynamic adjustment model of weighting factors (such as gradient descent or adaptive weighting algorithm), so that the weight of high correlation features increases and the weight of low correlation features decreases. Then, the optimized weights are multiplied by the corresponding feature values ​​to calculate the fusion result of the weighted combination, generating a dynamically adjusted swing safety threshold to ensure that it can adapt to the complex changing trends under different operating conditions in real time.

[0145] S6. Compare the safety threshold with the swing data to predict the trend of the bearing's current state.

[0146] S61. Align the safety threshold and swing data by timestamp, calculate the degree of deviation of the swing data from the safety threshold, and generate a state deviation index.

[0147] This deviation-based indicator extraction avoids relying solely on traditional fixed threshold methods and can dynamically adapt to changing operating conditions.

[0148] S62. By integrating the deviation index of the fusion state with historical operating condition parameters, the bearing runout trend is predicted through dynamic correlation analysis, and intuitive risk warnings are generated.

[0149] By integrating the state deviation index with historical operating condition parameters, a dynamic correlation diagram is used to assess the changing trend of the current state.

[0150] The deviation indicators and historical operating parameters are standardized and aligned by timestamps to ensure data consistency across time. Then, an attention mechanism is used to dynamically allocate weights to different variables, generating a joint feature vector as input for subsequent analysis. When evaluating state change trends using a dynamic correlation graph, each time step is treated as a node. The edge weights between nodes are determined by calculating feature similarity (e.g., Euclidean distance) or temporal correlation (e.g., Pearson correlation coefficient), and smoothed using a Gaussian kernel function. Furthermore, a topological attention mechanism dynamically adjusts the weights of node features to extract long-term trends and short-term fluctuations, ultimately accurately reflecting the current state's changing trend and providing a reliable basis for subsequent trend prediction.

[0151] By combining the current state change trend with a deep learning model, the current bearing state is divided into states, and the bearing swing trend is predicted.

[0152] Long-term trend features and short-term fluctuation characteristics extracted through dynamic correlation graphs are encoded into a time-series input deep learning model. The deep learning model can employ a combination of a multilayer perceptron and a temporal convolutional network; the former is responsible for modeling nonlinear relationships, while the latter captures temporal features. State classification maps feature vectors to predefined state categories (e.g., normal, warning, and danger) by adding a classification head. Trend prediction outputs the swing value sequence and uncertainty range for future time steps through a regression head. Through deep learning model training, the loss function is optimized to simultaneously minimize classification and regression errors, thereby achieving high-precision state classification and trend prediction, ensuring adaptability to complex changes and robustness of results.

[0153] Based on the bearing swing trend, the probability distribution of the bearing's future state is output, and intuitive risk warnings are generated.

[0154] Generating intuitive risk alerts can present complex probability distribution results in a visual or categorized form, making it easier for operations and maintenance personnel to quickly understand and make decisions, thereby effectively improving the efficiency and accuracy of risk management.

[0155] To further verify the technical solution of the present invention, experimental simulation data of the method for predicting the bearing swing trend of a hydro-generator unit are provided.

[0156] This experiment was conducted on Unit 3 of a large hydroelectric power plant. The unit has a rated speed of 375 r / min, a rated active power of 180 MW, and an average head height that fluctuates between 50 m and 65 m.

[0157] To accurately collect swing-related data, several high-precision vibration sensors are deployed near the unit bearings to capture raw vibration signals such as bearing swing peak value, swing amplitude, and swing frequency in real time. Simultaneously, an operating condition parameter acquisition module is deployed to record speed, active power, and cooling water pressure, and a temperature and humidity acquisition module is configured to measure ambient temperature, as well as to obtain environmental variables such as water head height through the power plant's water condition monitoring system.

[0158] All data from high-precision vibration sensors, operating condition parameter acquisition modules (including speed, active power, and cooling water pressure), and environmental variable acquisition modules (including ambient temperature and humidity, and water head height) are tagged with a unified timestamp to form the initial raw dataset.

[0159] In the data processing stage, time-series interpolation was first performed on the original dataset to compensate for missing values ​​caused by synchronization deviations in the acquisition system or instantaneous sensor malfunctions. Subsequently, wavelet decomposition was used to separate the time-series data into high-frequency and low-frequency components, retaining the low-frequency components. After suppressing high-frequency noise using a soft thresholding method, reconstruction was completed, resulting in a relatively stable swirl-denoised signal. To avoid the interference of numerical magnitude differences on model training and prediction accuracy, the Min-Max normalization method was used to uniformly scale all denoised signals to a scaled value. The data is then processed within a range, followed by a sliding window smoothing operation to reduce the impact of extreme outliers on the overall prediction. Afterwards, outlier removal and integrity checks are performed on the standardized feature matrix to ensure that the feature dataset used in subsequent stages has good representational quality and stability.

[0160] After data preprocessing, the importance of variables is ranked using a random forest method. The top-ranking features based on average recursive split reduction are selected as candidate features, and their correlation is tested. When the Pearson correlation coefficient between any two candidate features exceeds a set threshold, the candidate features are retained or removed according to their importance ranking in the random forest. This process allows subsequent models to focus on the variables most valuable for swing prediction. Next, based on three categories of information—swing, operating parameters, and environmental variables—the retained candidate features are weighted and summed, with dynamic weight adjustment achieved through an attention mechanism, resulting in a comprehensive feature output. Subsequently, multi-scale time series segments are generated using a sliding window, and features at different scales are extracted within a multi-scale convolutional neural network structure. The comprehensive feature and multi-scale features are then normalized again and aligned by timestamps to construct a three-dimensional input tensor suitable for input into a dynamically nested model.

[0161] The first layer of the dynamic nested model focuses on capturing the long-term evolution trend of the swing data. This layer analyzes each time segment in a graph structure and calculates the edge weights between nodes node by node by combining dynamic weights in the time dimension and dynamic spatial offsets. Through a hierarchical topology attention mechanism, nodes with higher weights in long-term trend identification can be given stronger influence, thereby outputting the long-term trend feature vector of the unit swing.

[0162] Subsequently, in the second layer of the dynamic nested model, a fast Fourier transform is used to project time segments into the frequency domain, extracting short-term high-frequency and low-frequency features. A multi-branch convolutional structure is then applied to process different frequency bands, and a nonlinear characteristic transformation method is used to enhance the capture of transient changes, resulting in a short-term fluctuation characteristic output. The long-term trend features and short-term fluctuation characteristics are finally combined with weights under optimization by a genetic algorithm to minimize the prediction error and output the swing trend prediction result. This prediction result, along with operating condition parameters, is input into a fuzzy logic inference module. Combined with dynamic correlation analysis, the safety threshold of the unit bearing under different operating conditions and environments can be calculated. Finally, this safety threshold is compared with the real-time swing value to determine the current bearing health status and future trend.

[0163] This experiment involved nearly two months of continuous data collection, and compared the proposed method with existing ARIMA models, BP neural networks, and grey prediction models in stages. The experimental data comparison ultimately demonstrated that the proposed method has significant advantages in swing trend prediction and safety threshold assessment. Details are shown in Table 1.

[0164] As can be seen from the performance comparison table, the dynamic nested model of this invention is significantly lower than traditional methods in terms of both mean absolute error (MAE) and peak error (PE). Specifically, the ARIMA model has a MAE of 0.061 mm and a PE of 0.088 mm; while the BP neural network reduces the PE to some extent, it still has limitations in handling real-world swing data fluctuations; the grey prediction model performs poorly in adapting to complex nonlinear systems, as evidenced by its PE and deviation from the safety threshold setting, indicating insufficient accuracy in judging extreme values. In contrast, the proposed method captures long-term trends through a first-layer network and short-term fluctuations through a second-layer network, and combines these with a genetic algorithm for optimal weight combination, resulting in a MAE of only 0.022 mm and a PE of only 0.039 mm, significantly outperforming the other three existing technologies.

[0165] Table 1. Performance Comparison of Methods for Predicting Bearing Runout Trend of Hydro-generator Units

[0166]

[0167] It should be noted that although the ARIMA model has a shorter computation time, it relies heavily on the assumption of data stationarity, making it difficult to accurately capture short-term fluctuations when dealing with non-stationary swing data influenced by multiple environmental variables. While BP neural networks possess some nonlinear fitting capabilities, they are prone to getting trapped in local optima in scenarios with significant noise and multi-scale features, and lack rapid response to sudden fluctuation anomalies. Grey prediction models can make predictions using relatively little historical data, but their adaptability to real-time changing environmental variables and operating parameters is insufficient, resulting in a 6.8% deviation in the safety threshold setting. This deviation can lead to overly conservative or overly optimistic assessments of bearing conditions during actual operation. In contrast, the dynamic nested model of this invention, by establishing a graph structure in the first layer network and combining dynamic weights of the time and spatial dimensions, effectively characterizes the correlation of nodes under long-term trends. The second layer network performs branch convolution and nonlinear characteristic transformation on short-term frequency domain features, which can meticulously distinguish transient changes within different frequency ranges. This allows for the synergistic analysis of long-term trends and short-term fluctuations, improving the overall accuracy and stability of predictions.

[0168] Furthermore, the setting of the safety threshold is also an important indicator for evaluating the merits of a prediction method. As can be seen from the table, the safety threshold setting deviation of the dynamic nested model of this invention is only 2.1%, indicating that the method can organically integrate operating condition parameters and swing trend prediction results, and effectively eliminate the influence of random noise through fuzzy logic reasoning and dynamic correlation analysis. Since the cooling water pressure, speed, and ambient temperature of the hydro-generator unit may vary significantly under different operating conditions, this invention performs multi-scale decomposition and dynamic weight allocation of each influencing factor in the time-series dimension, enabling the safety threshold to be adjusted in real time according to changes in the environment and operating status, thereby reasonably assessing the bearing health status.

[0169] Therefore, compared with the ARIMA model, BP neural network and grey prediction model, the dynamic nested model of this invention exhibits higher accuracy, stability and adaptability in bearing swing trend prediction, short-term fluctuation capture and safety threshold setting, and has obvious technical advantages, and can effectively improve the prediction accuracy and safety in engineering applications.

[0170] The technical solution of the present invention has been described in conjunction with the specific experimental procedures shown in the accompanying drawings. However, the scope of protection of the present invention is not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions resulting from such changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for predicting the trend of bearing runout of a hydroelectric generating unit, characterized in that, The method comprises the following steps: S1, collecting the runout data, operating condition parameters and environmental variables of the hydro-generator unit, performing denoising processing and normalization operation to form a standardized feature data set; S2, performing feature screening on the standardized feature data set, generating a comprehensive feature, and then constructing a three-dimensional input tensor and transmitting it to a dynamic nested model; S3, in the dynamic nested model, the long-term change trend of the runout data is captured through the first layer network nesting, and the short-term fluctuation characteristics of the runout data are captured through the second layer network nesting; S4, the weight combination of the two-layer network nesting is optimized by a genetic algorithm to generate a runout trend prediction result; S5, combining the operating condition parameters and the runout trend prediction result, performing fuzzy processing by fuzzy logic reasoning, and calculating the runout safety threshold according to dynamic correlation analysis; S6, comparing the safety threshold with the runout data, and predicting the current state of the bearing trend; Step S3 further comprises the following sub-steps: S31, taking the first layer network nesting and the second layer network nesting as the basic framework of the dynamic nested model; S32, in the first layer network nesting, receiving the three-dimensional input tensor and performing time slice-by-time slice decomposition, each node corresponding to a feature vector of a time slice; S33, based on the cosine similarity between the feature vectors, the spatial correlation between the nodes is calculated; S34, the spatial correlation of each node is weighted and averaged to calculate the spatial dynamic offset of each node, each time slice is represented as a plurality of feature vectors, and a graph structure is constructed; S35, sequentially traversing the nodes in the graph structure, calculating the edge weight value between the nodes in the graph structure edge by edge, updating the edge weight value to the graph structure, and generating a weighted graph structure; S36, applying a hierarchical topological attention mechanism to the weighted graph structure, dynamically adjusting the feature weight of the node through the edge weight value between the nodes, and generating a long-term trend feature vector of each time slice; S37, taking the long-term trend feature vector of each time slice as the output of the long-term trend modeling; S38, in the second layer network nesting, the fast Fourier transform is used to extract the short-term characteristic variable, the output of the long-term trend modeling is dynamically combined with the short-term characteristic variable, and a joint feature matrix is constructed; S39, multi-branch convolution is used to extract the high-frequency characteristics and low-frequency characteristics of the short-term characteristic variable, and nonlinear characteristic transformation is performed, and the short-term characteristics after nonlinear characteristic transformation are taken as the characteristic output of the short-term fluctuation.

2. The method of claim 1, wherein: In step S1, the runout data includes runout peak value, runout amplitude and runout frequency data.

3. The method of claim 1, wherein: In step S1, the operating condition parameters include rotating speed, active power and cooling water pressure.

4. The method of claim 1, wherein: In step S1, the environmental variables include environmental temperature and water head height.

5. The method of claim 1, wherein: Step S1 further comprises the following sub-steps: S11, taking the runout data, operating condition parameters and environmental variables as the original data set; S12, converting the original data set into time series format, filling in the missing values by time series interpolation, and forming time series data; S13, after wavelet denoising of the time series data, normalization and sliding window smoothing are performed, and a feature matrix is generated according to the time stamp alignment; S14, outlier elimination and integrity check are performed on the standardized feature matrix, and a formatted standardized feature dataset is output.

6. The method of claim 1, wherein: Step S2 further includes the following sub-steps: S21, using random forest to sort the importance of variables in the standardized feature dataset, taking the average recursive split reduction amount as the measurement standard, and screening out the top k candidate features according to the contribution value; S22, set a high correlation feature elimination threshold, screen candidate features based on Pearson coefficient and random forest importance ranking, and generate comprehensive features through weighted fusion and attention mechanism; S23, generate time series segments based on comprehensive features, and use multi-scale convolutional neural network to construct three-dimensional input tensor and transmit to dynamic nested model.

7. The method of claim 1, wherein: Step S4 further includes the following sub-steps: S41, taking the time dimension dynamic weight, spatial dynamic offset and edge weight as the dynamic weight of the first layer network nesting; S42, the amplitude weight a of the low-frequency characteristic l l , the amplitude weight a of the high-frequency characteristic s s , the frequency parameter ω of the low-frequency characteristic and the phase parameter φ of the low-frequency characteristic as the multi-branch convolution weight of the second layer network nesting S43, randomly generate an initial population, and each individual in the initial population is represented in vector form as a combination of dynamic weights and multi-branch convolution weights in the two-layer network; S44, adopt genetic algorithm to perform selection operation, crossover operation and mutation operation; S45, repeat the genetic algorithm, evaluate the fitness value of the population individuals, retain the optimal solution, and apply the optimized genetic algorithm solution to the two-layer network nesting until the iteration limit is reached; S46, the first layer adjusts the dynamic weight of the long-term trend modeling, and the second layer optimizes the convolution weight of the short-term fluctuation characteristics to generate the swing trend prediction result.

8. The method of claim 1, wherein: Step S5 further includes the following sub-steps: S51, divide the operating condition parameters and swing trend prediction results into three fuzzy language variables through fuzzy logic reasoning, and correspond to the respective membership functions; S52, analyze the relevance of historical operating condition parameters and swing trend prediction results, and establish a fuzzy rule base; S53, fuzz the values in the comprehensive features through the membership function to generate a fuzzy membership matrix; S54, take the fuzzy membership matrix as input, match and output fuzzy reasoning; S55, solve the fuzziness by the barycentric method to generate an initial safety threshold; S56, based on the operating condition parameters and swing trend prediction results, construct a dynamic correlation graph, evaluate the correlation between the features of the correlation graph through a Gaussian kernel function, and extract time series correlation features; S57, fuse the time series correlation features with the initial safety threshold, and optimize the swing safety threshold through weighted combination.

9. The method of claim 1, wherein: Step S6 further includes the following sub-steps: S61, align the safety threshold with the swing data according to the time stamp, calculate the deviation of the swing data relative to the safety threshold, and generate a state deviation index; S62, fuse the state deviation index and the historical operating condition parameters, predict the bearing swing trend through dynamic correlation analysis, and generate an intuitive risk prompt.

Images