Water motor rotating wheel state data deep clustering method based on self-attention

Through the self-attention mechanism and KL divergence optimization clustering algorithm, the problem of subjectivity and inefficiency in the monitoring of the rotor status of the water motor is solved, and the deep understanding and accurate monitoring of the rotor status data of the water motor is achieved, which improves the Electric power engineering management efficiency and safety.

CN118965045APending Publication Date: 2024-11-15HUANENG LANCANG RIVER HYDROPOWER CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410972053.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-19
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The prior art has problems of subjectivity and inefficiency in monitoring the state of the water motor rotor of the water motor, and it is difficult to capture the implicit patterns and complex relationships in the data, resulting in insufficient in-depth understanding of the state of the water motor rotor of the water motor.

Method used

The deep clustering method of the state data of the water motor rotor based on self-attention is adopted. Low-dimensional features are extracted through convolutional neural networks, and the self-attention mechanism is introduced to deeply explore the intrinsic connections of the data. The KL divergence optimization clustering algorithm is used to achieve the Accurate monitoring and analysis of operating status data of water motor motor wheels.

Benefits of technology

It improves the accuracy of the analysis and the expressiveness of the model, optimizes the clustering effect, and enhances the guarantee of the safe operation of hydropower units and the efficiency of power engineering management.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118965045A_ABST
    Figure CN118965045A_ABST
Patent Text Reader

Abstract

The invention discloses a water motor rotating wheel state data deep clustering method based on self-attention, and the method comprises the steps: 1), extracting the low-dimensional feature representation of the water motor rotating wheel state data in a plurality of stable operation states, and achieving the preliminary dimension reduction and feature extraction of complex data; 2) excavating the internal relation between the state data of the water motor rotating wheel in different operation states, and effectively fusing the characteristics of the state data of the water motor rotating wheel; and 3) based on the fused features, optimizing clustering algorithm parameters by using KL divergence as a clustering objective function so as to realize effective clustering of the state data of the rotating wheel of the water motor. According to the method, the characterization and understanding capability of the state data of the water motor rotating wheel is improved, the distribution rule or potential distribution rule of the state data of the water motor rotating wheel can be excavated, and an effective data analysis tool is provided for operation monitoring and maintenance of the water motor.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a method for deep clustering of hydroelectric motor rotor state data, and in particular to a method for deep clustering of hydroelectric motor rotor state data based on self-attention, and belongs to the field of power engineering state monitoring. Background Art

[0002] With the rapid development of the power engineering industry, the demand for hydropower turbine runner condition monitoring is also increasing. As the core component of the hydropower unit, the operating status of the hydropower turbine runner directly affects the performance and stability of the entire unit. Therefore, in order to ensure the safe operation of the hydropower unit and improve the power generation efficiency, it is crucial to continuously monitor and analyze the status of the hydropower turbine runner. However, existing condition monitoring methods mainly rely on manual observation and experience-based judgment, which is inevitably affected by subjectivity and is inefficient and error-prone when processing large amounts of data. In addition, traditional methods have difficulty capturing implicit patterns and complex associations in the data, resulting in insufficient in-depth understanding of the status of the hydropower turbine runner. Summary of the invention

[0003] Purpose of the invention: The present invention aims to overcome the subjectivity and limitations of traditional manual monitoring methods, as well as the problem of low efficiency in processing large data sets. To this end, the present invention provides a deep clustering method for hydropower turbine runner status data based on self-attention. The method deeply mines and utilizes the hidden intrinsic connections in the data through the self-attention mechanism, effectively performs feature fusion, and thus realizes accurate monitoring and analysis of the hydropower turbine runner operating status data.

[0004] Technical solution: A deep clustering method for hydropower turbine runner status data based on self-attention, suitable for hydropower turbine runner status fault analysis and diagnosis, including the following steps:

[0005] (1) Extract low-dimensional feature representations of hydropower turbine runner status data under various stable operating conditions to achieve preliminary dimensionality reduction and feature extraction of complex data;

[0006] (2) To explore the intrinsic relationship between the status data of the hydropower turbine runner under different operating conditions and effectively integrate the characteristics of the status data of the hydropower turbine runner based on the correlation;

[0007] (3) Based on the fused features, the clustering algorithm parameters are optimized to achieve effective clustering of the hydropower turbine impeller status data.

[0008] Furthermore, in step (1), a convolutional neural network (CNN) encoder is used to extract low-dimensional feature representations of the hydroelectric motor rotor state data under various stable operating states, thereby achieving preliminary dimensionality reduction and feature extraction of complex data. The specific implementation steps are as follows:

[0009] In the CNN encoder layer, the data of the hydroelectric motor wheel under N stable operating states are processed, and the data under each stable operating state has multiple variables;

[0010] For each stable operating state, the corresponding hydroelectric motor rotor state data is processed by N independent CNN encoders; each CNN encoder establishes a nonlinear mapping for the input data; through this mapping, the original hydroelectric motor rotor state data features are converted into low-dimensional features of the latent layer.

[0011] Furthermore, in step (2), a self-attention mechanism is introduced to deeply explore the intrinsic connection between data under different operating states, and the features of the hydroelectric motor runner state data are effectively integrated based on the learned correlation, thereby enhancing the sensitivity and characterization ability of the method of the present invention to state changes. The specific implementation steps are as follows:

[0012] The obtained low-dimensional features are input into N self-attention mechanisms. The self-attention mechanism can learn the correlations between data in different states and fuse features based on these correlations. Through the processing of the self-attention mechanism, the features in each state are further enhanced and refined. The feature matrix processed by the self-attention mechanism is pooled by applying the pooling layer to reduce the dimension of the data and retain the most important information. The pooling layer helps to extract key features while reducing the number of parameters and preventing overfitting. The feature matrix after the pooling layer is input into the fully connected layer to obtain the final common feature representation.

[0013] Furthermore, in step (3), based on the fused features, KL divergence is used as the clustering objective function to optimize the clustering algorithm parameters to achieve effective clustering of the hydroelectric motor rotor status data. The specific steps are as follows:

[0014] Given a cluster cluster, the K-Means algorithm is used to cluster the fused features. Since the fused features contain cluster distribution structure, their actual cluster distribution lacks guidance, and it is difficult to form an ideal cluster distribution that is cluster-friendly between data. In order to obtain the most ideal clustering result, a distribution A is calculated to measure the similarity relationship between the obtained common feature representation points and the cluster center, that is, to calculate the probability of assigning a feature representation point to a cluster cluster. The probability can reflect the current cluster distribution structure after integrating the latent features of various stable operating status data of hydropower generators.

[0015] An ideal clustering distribution should be cluster-friendly, that is, similar data points are closer, different data points are farther, the intra-class distance is smaller, and the inter-class distance is larger; therefore, by calculating an ideal clustering cluster distribution P, we learn from high-confidence tasks to improve the matching degree between distributions A and P, thereby improving the clustering ability of the method and the similarity between the common feature representation points and the cluster center of the cluster to which they belong.

[0016] The ideal cluster representation is achieved by minimizing the KL divergence between two distributions.

[0017] In clustering, the ideal distribution P should be close to the true distribution A, so the KL divergence needs to be minimized; this can be achieved by optimizing the model parameters so that P is as close to A as possible; during the optimization process, on the one hand, the ideal distribution P can be used to help the method learn a representation that is more conducive to clustering, so that the obtained representation vector is closer to the cluster center of the cluster to which it belongs; on the other hand, the distribution P calculated from distribution A is also supervising the update of A.

[0018] A deep clustering system for hydroelectric turbine runner state data based on self-attention, including three modules: CNN encoder, self-attention decoder and KL divergence optimization;

[0019] The CNN encoder module uses CNN as an encoder to extract low-dimensional feature representations from the data of the hydroelectric motor runner under various stable operating states, thereby achieving preliminary dimensionality reduction and feature extraction of complex data;

[0020] The self-attention decoder module deeply explores the intrinsic connection between data under different operating states by introducing the self-attention mechanism; based on the learned correlation, the runner state data features under different operating states of the hydroelectric machine are effectively integrated, thereby improving the sensitivity and characterization ability of the method of the present invention to state changes;

[0021] The KL divergence optimization module uses the fused features, takes the KL divergence as the clustering objective function, and optimizes the clustering results by adjusting the clustering algorithm parameters.

[0022] The implementation method of the deep clustering model of the hydroelectric motor impeller status data based on self-attention is the same as the method and will not be repeated here.

[0023] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for deep clustering of hydroelectric motor wheel state data based on self-attention as described above is implemented.

[0024] A computer-readable storage medium stores a computer program for executing the self-attention-based deep clustering method for hydroelectric motor wheel state data as described above.

[0025] Beneficial effects: Compared with the prior art, the present invention provides an innovative method for analyzing and monitoring the status data of the hydropower turbine wheel by combining CNN, self-attention mechanism and KL divergence, and its beneficial effects are reflected in improving the accuracy of analysis, enhancing the expressiveness of the model and optimizing the final clustering effect. These effects are of great practical significance for ensuring the safe operation of hydropower units and improving the management efficiency of power engineering. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a framework diagram of a method for deep clustering of hydroelectric motor runner state data based on self-attention according to an embodiment of the present invention;

[0027] Figure 2 An example of a deep clustering result of the hydroelectric generator impeller status data according to an embodiment of the present invention; (a) is a visualization diagram of the distribution of hydroelectric generator impeller blade status data, and (b) is a diagram of the clustering result of the hydroelectric generator impeller blade status data. DETAILED DESCRIPTION

[0028] The present invention is further explained below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0029] Figure 1 The overall framework of the deep clustering method of the hydropower turbine rotor state data based on self-attention is divided into three parts of the implementation process, including the implementation of the CNN encoder, the self-attention decoder and the KL divergence optimization module. First, the original data is formed according to the monitoring data collected by the hydropower turbine rotor state monitoring sensor under N different stable operating states. In each stable operating state, data of M variables are collected. For the nth stable operating state, the operating state data of the hydropower generator runner can be expressed as where d n It is represented as the number of data values ​​collected in the nth stable operating state. Secondly, for each stable operating state, the corresponding data is processed by N independent CNN encoders. The original data The low-dimensional feature representations of the hydroelectric motor rotor state data under different stable operating states are respectively input into the CNN encoder to extract the low-dimensional feature representations of the hydroelectric motor rotor state data under different stable operating states, thus realizing the preliminary dimensionality reduction and feature extraction of complex data. Afterwards, the output representation of the CNN encoder is further input into the self-attention decoder to explore the intrinsic connection between the data under different operating states, and the features of the hydroelectric motor rotor state data are effectively fused according to the learned correlation. Finally, based on the common feature representation Z learned by the self-attention decoder, the KL divergence is used as the clustering objective function to optimize the clustering algorithm parameters to achieve effective data clustering.

[0030] Figure 2 This is an example of the deep clustering results of the hydroelectric turbine runner status data. Different colors represent different clusters. Specifically, Figure 2 (a) is the visualization of the status data distribution of the hydropower turbine runner blades. Figure 2 (b) is the deep clustering result of the corresponding hydro-turbine turbine runner blade status data. There are multiple sensors on the hydro-turbine turbine runner, which can collect data on multiple variables such as water pressure, static stress, and water flow velocity. According to the deep clustering method of hydro-turbine turbine runner status data based on self-attention, the multivariate hydro-turbine runner data under different stable operating states collected and calculated are clustered to obtain the data distribution characteristics under common stable operating states. Figure 2 (b) The obtained clustering results, we can clearly identify the data distribution of the hydroelectric motor runner during normal operation and diagnose and predict potential faults or abnormal conditions.

[0031] The deep clustering method of hydropower turbine rotor state data based on self-attention is used to conduct experiments on hydropower turbine rotor state data under different stable operating states, including the following steps:

[0032] (1) A convolutional neural network (CNN) encoder is used to extract low-dimensional feature representations of the hydropower turbine runner state data under various stable operating conditions. The specific steps for achieving preliminary dimensionality reduction and feature extraction of complex data are as follows:

[0033] In the CNN encoder layer, the hydroelectric motor rotor state data under N stable operating states are processed. Each stable operating state data has M variables, so the original data set can be expressed as in, The original characteristic matrix representing the state data of the hydroelectric turbine rotor under the nth stable operating state, d n It is expressed as the number of data values ​​collected in the nth stable operating state;

[0034] For each stable operating state, the corresponding data is processed by N independent CNN encoders Each CNN encoder creates a nonlinear mapping f for the input data. n (X n θ n )→Z n , where f n represents the nonlinear mapping function of the nth CNN encoder, θ n (1≤n≤M) represents the parameters of the nth CNN encoder; through this mapping, the original data features are converted into low-dimensional features of the latent layer Among them, l n is the output dimension of the CNN encoder, that is, the dimension of the low-dimensional features.

[0035] (2) The self-attention mechanism is introduced to deeply explore the intrinsic connection between data under different operating states, and the features of the hydropower turbine runner state data are effectively integrated based on the learned correlation to enhance the model's sensitivity and representation ability to state changes. The specific steps are as follows:

[0036] The obtained low-dimensional feature Z n Input into N self-attention mechanisms, the self-attention mechanism can learn the correlation between data in different states and fuse features based on these correlations; through the processing of the self-attention mechanism, the features in each state are further enhanced and refined; the calculation formula of the self-attention mechanism is:

[0037]

[0038] Among them, Q, K, and V represent the n The query, key and value matrices obtained by linear transformation, d k is the dimension of each column feature in the key matrix, and the softmax function calculates the attention weight; the feature matrix after self-attention mechanism processing is Z′ n , apply the pooling layer to get the pooled feature matrix Z″ n , to reduce the dimension of the data and retain the most important information; the pooling layer helps to extract key features while reducing the number of parameters and preventing overfitting; the feature matrix after the pooling layer is input into the fully connected layer to obtain the final common feature representation Z; the mapping formula of the fully connected layer is:

[0039] FC(Z″ n )=Z

[0040] Among them, FC represents the fully connected layer operation, which maps the pooled features to the required output dimension and generates a common feature representation Z={z1,z2,…,z M}.

[0041] (3) Based on the fused features, KL divergence is used as the clustering objective function to optimize the clustering algorithm parameters to achieve effective data clustering. The specific steps are as follows:

[0042] Given k clusters, i j Refers to the cluster center of the jth cluster, and uses the K-Means algorithm to cluster the fused feature Z. Since the fused feature contains a cluster distribution structure, its actual cluster distribution lacks guidance, and it is difficult to form a cluster-friendly ideal cluster distribution between data. In order to obtain the most ideal clustering result, a distribution A is calculated to measure the obtained common feature representation point z. i and cluster heart j The similarity relationship between them is calculated as follows:

[0043]

[0044] Where a is the degree of freedom of the distribution A, q ij It represents the probability of assigning the i-th feature representation point to the j-th cluster, which can reflect the cluster distribution structure after integrating the latent features of various stable operation status data of hydropower generators.

[0045] The ideal cluster distribution should be cluster-friendly, that is, similar data points are closer, different data points are farther, the intra-class distance is smaller, and the inter-class distance is larger; therefore, by calculating an ideal cluster distribution P, we can learn from high-confidence tasks to improve the matching degree between distribution A and P, thereby improving the clustering ability of the method and the similarity between the common feature representation points and the cluster centers of the clusters to which they belong; the calculation formula for the ideal cluster distribution is as follows:

[0046]

[0047] where p ij It represents the ideal probability of assigning the i-th feature representation point to the j-th cluster, which can reflect the ideal cluster distribution structure after integrating the latent features of various stable operating state data of hydropower generators. In order to obtain the optimal clustering result, the KL divergence between distribution P and distribution A is minimized to achieve the ideal cluster representation. The specific loss function is as follows:

[0048]

[0049] In clustering, the ideal distribution P should be close to the true distribution A, so the KL divergence needs to be minimized; this can be achieved by optimizing the model parameters so that P is as close to A as possible; during the optimization process, on the one hand, the ideal distribution P can be used to help the method learn a representation that is more conducive to clustering, so that the obtained representation vector is closer to the cluster center of the cluster to which it belongs; on the other hand, the distribution P calculated from distribution A is also supervising the update of A.

[0050] The deep clustering system of hydroelectric turbine rotor state data based on self-attention includes three modules: CNN encoder, self-attention decoder and KL divergence optimization;

[0051] CNN encoder module, which uses CNN as an encoder to extract low-dimensional feature representations from the data of the hydroelectric turbine runner under various stable operating states, thereby achieving preliminary dimensionality reduction and feature extraction of complex data;

[0052] The self-attention decoder module deeply explores the intrinsic connection between data under different operating states by introducing the self-attention mechanism; based on the learned correlation, the data features of the runner state under different operating states of the hydroelectric generator are effectively integrated, thereby improving the sensitivity and characterization ability of the method of the present invention to state changes;

[0053] The KL divergence optimization module uses the fused features and takes the KL divergence as the clustering objective function to optimize the clustering results by adjusting the clustering algorithm parameters.

[0054] Obviously, those skilled in the art should understand that the various steps of the method for deep clustering of water motor wheel state data based on self-attention or the various modules of the deep clustering model for water motor wheel state data based on self-attention of the above-mentioned embodiment of the present invention can be implemented by a general computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, optionally, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, and in some cases, the steps shown or described can be executed in a different order than that here, or they can be made into individual integrated circuit modules respectively, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. In this way, the embodiments of the present invention are not limited to any specific combination of hardware and software.

[0055] The parameter settings and experimental evaluation indicators are as follows:

[0056] AParameter Settings

[0057] A model is obtained by implementing the method of the present invention. The following are the parameters that affect the performance of the model: the convolutional neural network structure of the CNN encoder, including the convolution kernel size and step size, the output dimension l of the CNN encoder n Settings, Adam optimizer learning rate, iteration rounds and model batch size and degrees of freedom a of A distribution.

[0058] Each CNN encoder in the model contains the same 3-layer convolutional neural network architecture; the convolution kernel size of the first convolutional network is 4×4×10, and the convolution step is 2; the convolution kernel size of the second convolutional network is 3×3×20, and the convolution step is 1; the convolution kernel size of the third convolutional network is 4×4×30, and the convolution step is 2; the output dimension of the CNN encoder is l n The initial learning rate of the Adam optimizer is 0.001; the entire model is trained using a batch size of 64 and 300 epochs; the degree of freedom a of the A distribution is designed to be 1.

[0059] B Experimental evaluation index

[0060] The following are specific evaluation indicators of the embodiments of the present invention: In this experiment, two common clustering evaluation indicators are used to evaluate the performance of the method, including clustering accuracy (ACC) and normalized mutual information (NMI). The higher the value, the better the clustering effect.

[0061] According to the above embodiments, the method of the present invention first uses a CNN encoder to extract low-dimensional feature representations of the hydroelectric motor wheel under a variety of stable operating states, thereby achieving preliminary dimensionality reduction and feature extraction of complex data. Next, the self-attention mechanism is introduced to deeply explore the intrinsic connections between data under different operating states, and the features of the hydroelectric motor wheel state data are effectively fused based on the learned correlations, thereby enhancing the model's sensitivity and characterization capabilities to state changes. Finally, based on the fused features, the KL divergence is used as the clustering objective function to optimize the clustering algorithm parameters to achieve effective clustering of data. The method adopted by the present invention utilizes deep learning and self-attention mechanisms to accurately identify and analyze complex nonlinear relationships in the hydroelectric motor wheel state data, significantly improving the model's perception and description capabilities for operating state changes. This method provides comprehensive and reliable diagnosis and fault prediction services for the hydroelectric motor state by performing efficient feature extraction and detailed clustering analysis on the data.

Claims

1. A deep clustering method for hydroelectric turbine runner state data based on self-attention, characterized in that: The steps include: (1) Extract low-dimensional feature representations of hydropower turbine runner state data under various stable operating conditions; (2) Mining the intrinsic relationship between the status data of the hydropower turbine runner under different operating conditions and fusing the characteristics of the status data of the hydropower turbine runner according to the correlation; (3) Based on the fused features, the clustering algorithm parameters are optimized to achieve effective clustering of the hydropower turbine impeller status data.

2. The method for deep clustering of hydroelectric motor runner state data based on self-attention according to claim 1 is characterized in that: In the step (1), a convolutional neural network encoder is used to extract low-dimensional feature representations of the hydroelectric motor rotor state data under various stable operating states, thereby achieving preliminary dimensionality reduction and feature extraction of complex data. The specific implementation steps are as follows: In the CNN encoder layer, the data of the hydroelectric motor wheel under N stable operating states are processed, and the data under each stable operating state has multiple variables; For each stable operating state, the corresponding hydroelectric motor rotor state data is processed by N independent CNN encoders; each CNN encoder establishes a nonlinear mapping for the input data; through this mapping, the original hydroelectric motor rotor state data features are converted into low-dimensional features of the latent layer.

3. The method for deep clustering of hydroelectric motor runner state data based on self-attention according to claim 1 is characterized in that: In step (2), a self-attention mechanism is introduced to deeply explore the intrinsic relationship between the status data of the hydroelectric motor wheel under different operating states, and the features of the status data of the hydroelectric motor wheel are fused according to the learned correlation. The specific implementation steps are as follows: The obtained low-dimensional features are input into N self-attention mechanisms. The self-attention mechanism can learn the correlations between data of different states and fuse features based on these correlations. The feature matrix processed by the self-attention mechanism is then applied to the pooling layer to obtain the pooled feature matrix; the feature matrix after the pooling layer is input into the fully connected layer to obtain the final common feature representation.

4. The method for deep clustering of hydroelectric motor runner state data based on self-attention according to claim 1 is characterized in that: In step (3), based on the fused features, the KL divergence is used as the clustering objective function to optimize the model parameters to achieve effective data clustering. The specific steps are as follows: Given a cluster, the K-Means algorithm is used to cluster the fused features. The similarity between the common feature representation points and the cluster centers is measured by calculating a distribution A, that is, the probability of assigning a feature representation point to a cluster is calculated. The probability reflects the current cluster distribution structure after the latent features of various stable operating state data of the hydropower generator are integrated. By calculating an ideal cluster distribution P, we learn to improve the match between distributions A and P from high-confidence tasks, and achieve an ideal cluster representation by minimizing the KL divergence between the two distributions.

5. A deep clustering system for hydroelectric turbine runner state data based on self-attention, characterized in that: include: Three modules: CNN encoder, self-attention decoder and KL divergence optimization; The CNN encoder module uses CNN as an encoder to extract low-dimensional feature representations from the data of the hydroelectric turbine runner under various stable operating states; The self-attention decoder module deeply explores the intrinsic connection between data in different operating states by introducing the self-attention mechanism; Based on the learned correlation, the data features of the runner status under different operating conditions of the hydropower generator are effectively integrated; The KL divergence optimization module uses the fused features, takes the KL divergence as the clustering objective function, and optimizes the clustering results by adjusting the clustering algorithm parameters.

6. A computer device, characterized in that: The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the deep clustering method for water motor wheel status data based on self-attention as described in any one of claims 1 to 4 is implemented.

7. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the self-attention-based deep clustering method for hydroelectric motor wheel state data as described in any one of claims 1 to 4 are implemented.