Water turbine vibration data analysis method and system

Abstract

The invention relates to the technical field of rotary mechanical vibration analysis, in particular to a water turbine vibration data analysis method and system, which establishes a frequency and energy matching relation table through feature updating and synchronism screening, so that coupling analysis of a frequency domain and a time domain has pertinence and fineness, and energy features of key frequency responses among structures are highlighted; through feature similarity calculation and an updating mechanism of a low-similarity label, the dynamic adaptability of the frequency energy mode under the variable working condition clearly reflects a risk transmission link of frequency energy features between nodes, the dynamic adaptability of the frequency energy mode under the variable working condition is improved, and the reliability of the frequency energy mode under the variable working condition is improved. Through processing logic and organic connection of signal multi-scale feature extraction and key component frequency energy difference analysis, pertinence of frequency energy feature extraction, dynamic nature of working condition adaptation and fineness of risk transfer description are realized, and thus early abnormity capture precision and fault evolution path identification capability are improved.

Description

Technical Field

[0001] This invention relates to the field of rotating machinery vibration analysis technology, and in particular to a method and system for analyzing vibration data of a water turbine. Background Technology

[0002] The field of rotating machinery vibration analysis technology utilizes computer and signal processing technology to collect, store, process, and analyze vibration signals generated during the operation of rotating machinery. Its applications include rotating equipment such as water turbines, steam turbines, and wind turbine generators. It typically covers spectrum analysis, time-domain and frequency-domain feature extraction, pattern recognition, condition monitoring, and fault diagnosis. The aim is to identify the operating status of equipment by analyzing changes in vibration modes, prevent faults, extend service life, and improve operating efficiency.

[0003] A vibration data analysis method for hydro turbines is proposed to continuously acquire vibration signals during turbine operation and determine the operating status and potential anomalies through data analysis. The aim is to achieve early fault warning, operating status assessment, and maintenance decision support for hydro turbines, thereby reducing unexpected downtime, lowering maintenance costs, and enabling stable operation of hydro turbines under rated performance. This method is used to establish a technical system that enables automated operation, real-time monitoring, and accurate diagnosis, thereby improving the high availability and high safety of hydro turbines throughout their entire life cycle.

[0004] Existing technologies rely heavily on single-scale and fixed-band feature calculations in spectrum analysis, time-domain and frequency-domain feature extraction, and pattern recognition. This makes it difficult to capture the correlation patterns between structures in the dynamic changes of frequency and energy. Traditional methods are often disconnected from structural features and lack a mechanism for dynamically correcting frequency and energy characteristics based on operating conditions. This leads to deviations in analysis results when operating conditions fluctuate. They are mostly measured by overall vibration levels and lack node-level risk value distributions refined to frequency and energy attributes. This makes it difficult to accurately characterize the risk propagation chain. Existing technologies mostly rely on static correlations or empirical judgments and lack risk transmission paths built based on time series and connection relationships. This easily leads to the omission of the sequential relationship and evolution order of key nodes, increasing the probability of sudden equipment shutdowns under high-load operating conditions. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and to propose a method and system for analyzing vibration data of water turbines.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for analyzing vibration data of a water turbine, comprising the following steps: Step 1: Based on the vibration signal sequence of the turbine rotor, guide vanes, and main bearing, decompose the waveforms of different frequency bands and extract the amplitude and period curves. Use graph neural network to update node features and select high-synchronization segments to establish a frequency and energy matching table, and generate a multi-scale synchronous frequency and energy feature set. Step 2: Based on the multi-scale synchronous frequency and energy feature set, extract the amplitude difference and rate of change of the frequency points of the thrust bearing and the tailrace pipe, sort and screen the sample points that meet the spacing conditions, record the frequency position and energy ratio, and generate the mode center frequency and energy distribution table. Step 3: Based on the frequency and energy distribution table of the model center, extract the turbine operating condition parameters, map the model center to the guide vane clearance and rotor support matrix, use graph neural network to calculate node feature similarity and filter low similarity labels, update frequency position and energy ratio, and generate dynamic frequency and energy model matrix. Step 4: Based on the dynamic frequency energy mode matrix, assign node attributes to the main bearing and thrust bearing, establish node set and vibration propagation path, construct node risk value distribution table according to the adjacency matrix order, and generate node frequency energy risk matrix; Step 5: Based on the node frequency and energy risk matrix, sort the node risk values ​​and mark the tailrace pipe and rotor nodes that are higher than the threshold. Form risk transmission paths according to the connection order and arrange them by time to generate a cross-component risk transmission path set.

[0007] As a further embodiment of the present invention, the multi-scale synchronous frequency energy feature set includes a frequency location set, an energy density set, a phase difference parameter set, and an amplitude variance set; the mode center frequency energy distribution table includes a mode center frequency location column, a mode center energy ratio column, and an amplitude difference rate column; the dynamic frequency energy mode matrix includes frequency location parameters, energy ratio parameters, and a set of matching labels with guide vane clearance and rotor support; the node frequency energy risk matrix includes node number, frequency value, energy value, and risk value; and the cross-component risk transmission path set includes a high-risk node sequence, inter-node connection relationships, and time sequence information.

[0008] As a further aspect of the present invention, the specific steps for generating the multi-scale synchronous frequency energy feature set are as follows: Based on the vibration signal sequence of the turbine rotor, guide vanes, and main bearing, the waveforms of each frequency band in the signal are extracted, and the signal is clipped according to the set frequency band. The signal in each frequency band is divided into several time periods, and the amplitude and period in each time period are recorded and calculated to obtain the frequency band waveform segments. Based on the waveform segments of the frequency band, the amplitude and period characteristics of each waveform segment are compared, the variation trend of amplitude and period within each waveform segment is calculated, the part with high waveform synchronization is selected, and the waveforms are merged in time order to generate merged waveform data, thus obtaining the synchronized waveform segment. Based on the aforementioned synchronous waveform segments, each waveform segment is constructed as a node in a graph structure, and parameters such as mean frequency, energy density, phase difference, and amplitude variance are written into the node attributes. Weighted connecting edges are established based on temporal adjacency, frequency difference, and energy difference to form a complete graph structure. A graph neural network is used to perform information transmission and node feature updates on the graph structure. The feature similarity between the updated nodes is calculated and sorted from high to low values ​​to generate a multi-scale synchronous frequency and energy feature set.

[0009] As a further aspect of the present invention, the graph neural network specifically includes: First, for the graph structure constructed from synchronous waveform segments, each node is initialized as a multi-dimensional feature vector containing the mean frequency, energy density, phase difference, and amplitude variance. Based on the weighted edge relationship between nodes, the aggregated features of adjacent nodes and the features of the center node are updated in each iteration. The feature vectors of neighboring nodes are aggregated by averaging, and the aggregation weight is set according to the frequency difference and energy difference. The aggregation result and the current feature vector of the center node are transformed by a nonlinear transformation function to generate updated node features. The feature similarity between each pair of nodes in the graph is calculated by the cosine similarity function. The node pairs are sorted from high to low similarity, and node combinations that meet the preset synchronization threshold are selected. The corresponding frequency and energy attributes are extracted to form a multi-scale synchronous frequency and energy feature set.

[0010] As a further aspect of the present invention, the specific steps for generating the mode center frequency energy distribution table are as follows: Based on the multi-scale synchronous frequency energy feature set, each frequency point is sequentially located in the corresponding components of the thrust bearing and tailrace pipe, the amplitude is collected and the difference between the amplitudes of adjacent frequency points is calculated, the difference is compared point by point in frequency order and the rate of change is recorded, and the amplitude difference and the rate of change of the frequency points are combined into an ordered sequence to generate the frequency amplitude difference rate sequence. Based on the frequency amplitude difference rate sequence, the amplitude spacing between adjacent frequency points in the sequence is compared and it is determined whether it is greater than a set threshold. Frequency points that meet the conditions are marked as sample points. The frequency position and corresponding energy ratio of the sample points are extracted and stored in the table data. The parameters in the table are arranged and combined by row to generate the mode center frequency energy distribution table.

[0011] As a further aspect of the present invention, the specific steps for generating the dynamic frequency energy mode matrix are as follows: Based on the center frequency energy distribution table of the mode, the operating parameters of the turbine are extracted, including speed, load, etc., and the correlation positions of each operating parameter with the guide vane clearance and rotor support are matched one by one. The matching information between the operating parameters and the corresponding positions is recorded and organized into a corresponding mapping table to generate an operating parameter mapping table. Based on the aforementioned working condition parameter mapping table, the center frequency and energy information of each mode are compared with the guide vane clearance and rotor support parameters. The frequency position and energy ratio of each mode center are matched with the corresponding parameters. An input vector is constructed and input into the residual network to obtain the residual correction value. The matching degree of the corrected matching result is calculated. Labels with low matching degree are selected and marked as labels to be updated, generating a set of low matching degree labels. Based on the set of low-matching tags, the frequency position and energy ratio of each tag are adjusted, the updated frequency position and energy ratio are organized into the corresponding tags, the matching relationship between the tags and the operating parameters is updated, and the updated mode matrix is ​​generated to obtain the dynamic frequency-energy mode matrix.

[0012] As a further aspect of the present invention, the residual network specifically includes: first, combining the frequency position and energy ratio of each mode center with the guide vane clearance and rotor support parameters to form a fixed-length input vector; the input vector sequentially passes through a multi-layer linear transformation module, introducing a short-circuit connection structure between adjacent linear transformation layers, and directly adding the output of the intermediate layer with the output of the current layer to form a residual mapping, generating a residual correction value vector consistent with the original input dimension; then, performing element-wise addition between the correction value vector and the original input vector to obtain the frequency and energy parameter matching result after residual correction; finally, calculating the matching degree value based on the Euclidean distance between the corrected vector and the corresponding operating condition parameters, and filtering out low-matching-degree samples according to a preset matching degree threshold to generate a set of labels to be updated.

[0013] As a further aspect of the present invention, the specific steps for generating the node frequency-energy risk matrix are as follows: Based on the dynamic frequency and energy mode matrix, the frequency and energy values ​​of the main bearing and thrust bearing nodes are extracted sequentially. The frequency and energy values ​​are then filled into the attribute fields of the corresponding nodes. Within the attribute fields, similar parameters are categorized and sorted. Finally, the nodes and their attributes are combined into a complete list to generate a node attribute set. Based on the set of node attributes, a connection relationship is established between nodes according to the vibration propagation path. The position index of adjacent nodes is determined in sequence, and the risk calculation value of the node is filled into the risk distribution table using the index order. The risk distribution tables of the nodes are merged and sorted by node number to generate a node frequency-energy risk matrix.

[0014] As a further aspect of the present invention, the specific steps for generating the cross-component risk transmission path set are as follows: Based on the node frequency-energy risk matrix, the risk values ​​of each node are arranged in descending order. The risk values ​​of each node are compared with the set threshold. Nodes that are higher than the threshold are selected and marked as high-risk nodes. The frequency, energy and risk value of the nodes are recorded to generate a set of high-risk nodes. Based on the set of high-risk nodes, the transmission path between nodes is constructed one by one according to the connection relationship between the nodes, the connection order between each high-risk node is recorded, the transmission path is arranged in chronological order, and the path information is associated with the timestamp of the node to generate a set of cross-component risk transmission paths.

[0015] A turbine vibration data analysis system is provided, the system being used to execute the aforementioned turbine vibration data analysis method, the system comprising: Frequency and energy feature construction module: Based on the vibration signal sequence of rotor, guide vane and main bearing, frequency band division and amplitude and period extraction are performed, waveform segments are constructed into graph structure and node features are updated, and frequency and energy matching table is established by combining synchronous segments to generate multi-scale synchronous frequency and energy feature set; Model distribution analysis module: Based on the multi-scale synchronous frequency and energy feature set, extract the amplitude difference and rate of change of the frequency points of the thrust bearing and the tailrace pipe, screen sample points that meet the spacing conditions, record the frequency position and energy ratio, and generate the model center frequency and energy distribution table. Operating condition matching and correction module: Based on the mode center frequency energy distribution table, obtain operating condition parameters and establish a mapping with guide vane clearance and rotor support, match frequency position and energy ratio, adjust low similarity label parameters, and generate dynamic frequency energy mode matrix; Node risk modeling module: Based on the dynamic frequency energy mode matrix, assign node attributes to the main bearing and thrust bearing, determine the node connection relationship and generate a risk value distribution table to form a node frequency energy risk matrix; Cross-component risk tracking module: Based on the node frequency and energy risk matrix, high-risk tailrace pipes and rotor nodes are screened, risk transmission paths are constructed and sorted, and a cross-component risk transmission path set is generated.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: 1. In this invention, a frequency and energy matching relationship table is established through feature updating and synchronization screening, which makes the frequency domain and time domain coupling analysis more targeted and precise. By using the joint sorting and spacing conditions of the frequency point amplitude difference and change rate of the thrust bearing and tailpipe, the energy characteristics of the key frequency response between structures are highlighted. 2. In this invention, the dynamic adaptability of frequency energy mode under changing operating conditions is improved by using feature similarity calculation and low similarity label update mechanism, which clearly reflects the risk transmission link of frequency energy characteristics between nodes and improves the dynamic adaptability of frequency energy mode under changing operating conditions. 3. In this invention, by organically linking signal multi-scale feature extraction and frequency energy difference analysis of key components through processing logic, the targeted nature of frequency energy feature extraction, the dynamic nature of operating condition adaptation, and the fineness of risk transmission characterization are achieved, thereby improving the accuracy of early anomaly detection and the ability to identify fault evolution paths. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the workflow of the present invention; Figure 2 This is a system flowchart of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] Example 1 Please see Figure 1 This invention provides a technical solution: a method for analyzing vibration data of a water turbine, comprising the following steps: Step 1: Based on the vibration signal sequence of the turbine rotor, guide vanes, and main bearing, decompose the waveforms of different frequency bands and extract the amplitude and period curves. Use graph neural network to update node features and select high-synchronization segments to establish a frequency and energy matching table, and generate a multi-scale synchronous frequency and energy feature set. Step 2: Based on the multi-scale synchronous frequency and energy feature set, extract the amplitude difference and rate of change of the frequency points of the thrust bearing and the tailrace pipe, sort and screen the sample points that meet the spacing conditions, record the frequency position and energy ratio, and generate the model center frequency and energy distribution table. Step 3: Based on the frequency and energy distribution table of the model center, extract the turbine operating condition parameters, map the model center to the guide vane clearance and rotor support matrix, use graph neural network to calculate node feature similarity and filter low similarity labels, update frequency position and energy ratio, and generate dynamic frequency and energy model matrix. Step 4: Based on the dynamic frequency energy mode matrix, assign node attributes to the main bearing and thrust bearing, establish node set and vibration propagation path, construct node risk value distribution table according to the adjacency matrix order, and generate node frequency energy risk matrix; Step 5: Based on the node frequency and energy risk matrix, sort the node risk values ​​and mark the tailrace pipe and rotor nodes that are higher than the threshold. Form risk transmission paths according to the connection order and arrange them by time to generate a cross-component risk transmission path set.

[0021] The multi-scale synchronous frequency and energy feature set includes a set of frequency locations, a set of energy densities, a set of phase difference parameters, and a set of amplitude variances. The mode center frequency and energy distribution table includes a column of mode center frequency locations, a column of mode center energy proportions, and a column of amplitude difference rates. The dynamic frequency and energy mode matrix includes frequency location parameters, energy proportion parameters, and a set of matching labels with guide vane clearance and rotor support. The node frequency and energy risk matrix includes node number, frequency value, energy value, and risk value. The cross-component risk transmission path set includes a sequence of high-risk nodes, inter-node connection relationships, and time sequence information.

[0022] The specific steps for generating a multi-scale synchronous frequency energy feature set are as follows: Based on the vibration signal sequence of the turbine rotor, guide vanes, and main bearing, the waveforms of each frequency band in the signal are extracted, and the signal is clipped according to the set frequency band. The signal in each frequency band is divided into several time periods, and the amplitude and period in each time period are recorded and calculated to obtain the frequency band waveform segments. Based on frequency band waveform segments, the amplitude and period characteristics of each waveform segment are compared, the variation trend of amplitude and period within each waveform segment is calculated, the part with high waveform synchronization is selected, and the waveforms are merged in time order to generate merged waveform data, thus obtaining synchronized waveform segments. Based on synchronous waveform segments, each waveform segment is constructed as a node of a graph structure, and the parameters of mean frequency, energy density, phase difference and amplitude variance are written into the node attributes. Weighted connecting edges are established according to time adjacency and frequency and energy differences to form a complete graph structure. A graph neural network is used to perform information transmission and node feature update on the graph structure. The feature similarity between updated nodes is calculated and sorted from high to low values ​​to generate a multi-scale synchronous frequency and energy feature set. Based on the vibration signal sequences of the turbine rotor, guide vanes, and main bearing, spectral analysis was performed on each vibration signal. The sampling frequency was set to 2560Hz, the signal length was 2048 points, and the output was a complex spectrum array containing real and imaginary parts. The frequency band was filtered according to the preset frequency range of 0.5Hz to 500Hz. The filtered signal was divided into segments of 256 points each. The peak amplitude and period difference of the signal in each time period were calculated and recorded in the result table to generate frequency band waveform segments. Based on frequency band waveform segments, the amplitude and period characteristics of each segment are compared pairwise. The amplitude correlation threshold is set to 0.85 and the period correlation threshold is set to 0.80. The changing trends of the amplitude sequence and period sequence of each segment are calculated. Segments with both amplitude correlation and period correlation greater than the threshold are marked as high synchronization segments. The high synchronization segments are merged in time order to generate synchronized waveform segments. Based on synchronous waveform segments, a graph neural network algorithm is used to construct graph structure nodes and perform feature updates. The features of each node are defined as mean frequency, energy density, phase difference, and amplitude variance. The mean frequency is the average frequency value within the segment, the energy density is the sum of the segment's energy per unit time, the phase difference is the difference between phase values ​​in the same frequency band, and the amplitude variance is the variance of the segment's amplitude change. The weights of the weighted connection edges are calculated by the weighted average of the frequency difference and energy difference, with weight parameters of 0.6 for the frequency difference and 0.4 for the energy difference. Iterative updates are performed using a graph convolutional network structure with ReLU as the activation function, Adam as the optimizer, a learning rate of 0.001, and 200 iterations. After completing the node feature updates, the nodes are sorted from high to low feature similarity to generate a multi-scale synchronous frequency and energy feature set.

[0023] The graph neural network specifically includes the following steps: First, for the graph structure constructed from synchronous waveform segments, each node is initialized as a multi-dimensional feature vector containing the mean frequency, energy density, phase difference, and amplitude variance. Based on the weighted edge relationship between nodes, the aggregated features of neighboring nodes and the features of the center node are updated in each iteration. The feature vectors of neighboring nodes are aggregated by averaging, and the aggregation weight is set according to the frequency difference and energy difference. The aggregation result and the current feature vector of the center node are transformed by a nonlinear transformation function to generate updated node features. The feature similarity between each pair of nodes in the graph is calculated by the cosine similarity function. The node pairs are sorted from high to low similarity, and node combinations that meet the preset synchronization threshold are selected. The corresponding frequency and energy attributes are extracted to form a multi-scale synchronization frequency and energy feature set. Graph neural networks, according to the formula: ; in: Indicates the first Layer Time Node Update the feature vector. This indicates the number of the central node currently being updated. Represents nodes The neighboring node numbers, Indicates the layer index of the network. Represents a node The set of neighboring nodes, Represents a node With nodes The initial weighted connection coefficients, This represents the frequency difference weighting coefficient. Represents a node With nodes The frequency difference is the node frequency mean minus nodes The frequency mean, Indicates the frequency difference attenuation coefficient. Represents a node With nodes Energy difference, Indicates the energy difference attenuation coefficient. This indicates that the position in the formula represents the base of the exponentiation operation, taking the natural logarithm constant. , Represents a node The normalization factor, Represents the feature transformation weight matrix. Indicates the first Layer-time neighbor nodes eigenvectors, Represents the nonlinear activation function ReLU; Execution process: First, determine the current node. The set of neighboring nodes Read each neighbor node in turn. eigenvectors Get nodes With nodes initial connectivity coefficients Calculate the frequency difference between the two. And substitute in the frequency difference attenuation coefficient attenuation coefficient The calculation method is to take the average of the frequency differences at each node, then take the reciprocal and multiply it by the adjustment coefficient of 1.2 to calculate the energy difference. And substitute the energy difference attenuation coefficient Energy attenuation coefficient The calculation method is to take the reciprocal of the mean of the energy differences at each node and multiply it by an adjustment coefficient of 0.8 to calculate the frequency difference weighting coefficient. The value is the square root of the ratio of the variance of the frequency difference to the variance of the energy difference at each node, and the initial connection coefficients are... and Multiply by, then multiply by each of them. and The values ​​obtained from the two exponentially decaying terms are used as temporary weights for neighboring nodes. The normalization factor is obtained by summing the temporary weights of the neighboring nodes. Then divide the temporary weights of each neighbor by Obtain the normalized weights, and then combine the normalized weights with the feature vectors of the neighboring nodes. Multiply by, and then by the feature transformation weight matrix Multiplying the results and summing the neighbor calculations yields an aggregated vector. Finally, the aggregated vector is input into the ReLU activation function to obtain the node's result. In the Updating the feature vector of the layer .

[0024] The specific steps for generating the center frequency energy distribution table of the model are as follows: Based on the multi-scale synchronous frequency energy feature set, each frequency point is located sequentially in the corresponding components of the thrust bearing and tailrace pipe, the amplitude is collected and the difference between the amplitudes of adjacent frequency points is calculated, the difference is compared point by point in frequency order and the rate of change is recorded, and the amplitude difference and the rate of change of the frequency points are combined into an ordered sequence to generate the frequency amplitude difference rate sequence. Based on the frequency amplitude difference rate sequence, the amplitude spacing between adjacent frequency points in the sequence is compared and it is determined whether it is greater than a set threshold. Frequency points that meet the conditions are marked as sample points. The frequency position and corresponding energy ratio of the sample points are extracted and stored in the table data. The parameters in the table are arranged and combined by row to generate the mode center frequency energy distribution table. Based on the multi-scale synchronous frequency energy feature set, frequency points are sequentially read in the corresponding components of the thrust bearing and tailrace pipe. The reading step size is set to 0.1Hz, the amplitude acquisition accuracy is 0.001 units, the amplitude values ​​of each frequency point are recorded, and the amplitudes of two adjacent frequency points are differentially calculated according to the frequency order. The differential result is retained to three decimal places. The difference sequence is arranged in order of frequency from low to high. The amplitude change rate calculation process is called. The sampling interval in the rate calculation is 0.5 seconds, and the rate value is retained to two decimal places. The amplitude difference and change rate of each frequency point are combined into an ordered data sequence according to the frequency order to generate the frequency amplitude difference rate sequence. Based on the frequency amplitude difference rate sequence, the amplitude spacing between adjacent frequency points is compared item by item in the ordered sequence. The preset amplitude spacing threshold is 0.05 units. It is determined whether the spacing between each frequency point is greater than the threshold. Frequency points that meet the condition are marked as sample points. The frequency position and corresponding energy ratio of the sample points are extracted and retained to three decimal places. The frequency position and energy ratio are stored in two columns of a table. The table is sorted in ascending order according to the frequency position. The frequency position column and energy ratio column are combined and arranged by row to generate the center frequency energy distribution table of the model.

[0025] The specific steps for generating the dynamic frequency energy mode matrix are as follows: Based on the center frequency energy distribution table of the model, the operating parameters of the turbine are extracted, including speed and load. Each operating parameter is matched with the associated position of the guide vane clearance and rotor support. The matching information between the operating parameters and the corresponding positions is recorded and organized into a corresponding mapping table to generate an operating parameter mapping table. Based on the working condition parameter mapping table, the center frequency and energy information of each mode are compared with the guide vane clearance and rotor support parameters. The frequency position and energy ratio of each mode center are matched with the corresponding parameters. An input vector is constructed and input into the residual network to obtain the residual correction value. The matching degree of the corrected matching result is calculated. Labels with low matching degree are selected and marked as labels to be updated, generating a set of low matching degree labels. Based on the low-matching tag set, the frequency position and energy ratio of each tag are adjusted, the updated frequency position and energy ratio are organized into the corresponding tags, the matching relationship between the tags and the operating parameters is updated, and the updated mode matrix is ​​generated to obtain the dynamic frequency-energy mode matrix. Based on the center frequency energy distribution table of the model, the operating parameters of the turbine are extracted. In the data acquisition module, the speed sampling interval is set to 1 second, the speed data accuracy is 0.1 revolutions per minute, the load sampling interval is 5 seconds, and the load data accuracy is 0.01 megawatts. The collected speed and load data are matched item by item to the association position table of guide vane clearance and rotor support. The guide vane clearance is recorded in millimeters with an accuracy of 0.01 millimeters, and the rotor support position is stored in coordinate form with an accuracy of 0.1 millimeters. Each set of operating parameters and position data is formed into a two-dimensional association pair, sorted by parameter name and stored in a mapping table to generate an operating parameter mapping table. Based on the operating condition parameter mapping table, a residual correction matching method is used to match the center frequency and energy information of the mode to the guide vane clearance and rotor support parameters item by item. The frequency position and energy ratio of each mode center are retained to three decimal places. The frequency position, energy ratio and guide vane clearance and rotor support parameters are combined to form a fixed-length input vector with a length of 4 dimensions. The input vector is input to the residual network algorithm model for calculation. The residual network contains a three-layer linear mapping structure, with each layer having a mapping dimension of 128. The weight initialization method is uniform distribution, the activation function type is ReLU, and short-circuit connection structure is added between every two adjacent layers. The learning rate is set to 0.001, and the number of iterations is 300 rounds. The output correction value is added element by element to the original input vector to obtain the corrected matching result. The Euclidean distance between the corrected result and the operating condition parameters is calculated and retained to two decimal places. Labels with a matching degree lower than the preset threshold of 0.75 are added to the list to generate a low matching degree label set. Based on a low-matching tag set, the frequency position and energy ratio of each tag in the set are modified item by item. The frequency position is adjusted by 0.01 to 0.05 Hz in the direction of difference from the original value, and the energy ratio is adjusted by 0.001 to 0.005 in the direction of difference from the original value. All updated values ​​are retained to three decimal places. The updated frequency position and energy ratio are written into the corresponding tag record, and the tags are rematched with the operating parameters. The generated new matching results are sorted by tag number and stored as a two-dimensional matrix file to generate a dynamic frequency-energy pattern matrix.

[0026] The residual network specifically includes the following steps: First, the frequency position and energy ratio of each mode center are combined with the guide vane clearance and rotor support parameters to form a fixed-length input vector. The input vector is then passed through a multi-layer linear transformation module. A short-circuit connection structure is introduced between adjacent linear transformation layers. The output of the intermediate layer is directly added to the output of the current layer to form a residual mapping, generating a residual correction value vector with the same dimension as the original input. Then, the correction value vector is added to the original input vector element-wise to obtain the frequency and energy parameter matching results after residual correction. Finally, the matching degree value is calculated based on the Euclidean distance between the corrected vector and the corresponding operating parameters, and low matching degree samples are selected according to the preset matching degree threshold to generate a set of labels to be updated. Residual networks, according to the formula: ; in: This represents the vector of corrected frequency and energy parameter matching results. Indicates the layer index of the network. This represents the total number of layers in the residual network. Indicates the first Layer weight coefficients, ReLU represents the nonlinear activation function. Indicates the first The weight matrix of the layer, This represents a fixed-length input vector of length 4 dimensions, which includes, in sequence, the mode center frequency position, the mode center energy ratio, the guide vane clearance parameter, and the rotor support parameter. Indicates the first The layer's bias vector, This represents the frequency difference weighting coefficient. Indicates the first The frequency difference obtained from layer calculation, This represents the energy difference weighting coefficient. Indicates the first The energy difference obtained from layer calculation, This represents the overall scaling factor for the input vector residual information; Execution process: First, the mode center frequency position, mode center energy ratio, guide vane clearance parameters, and rotor support parameters are combined in sequence to form a 4-dimensional input vector. The input vector is then fed into the residual network for layer-by-layer computation, at the th... Within the layer, the input vector First with the weight matrix Multiply and add the bias vector The difference between the center frequency of the calculation mode and the corresponding operating frequency is obtained. And multiplied by the frequency difference weighting factor Then, the difference between the energy ratio at the center of the mode and the energy ratio under the corresponding operating condition is calculated to obtain the result. And multiplied by the energy difference weighting factor The linear transformation output is compared with the frequency difference correction term. and energy difference correction term After element-wise addition, input a non-linear activation function. Generate the feature output of this layer, and then multiply the feature output by the first... Layer weight coefficient The weighted output of the layer is summed, and the sum vector is compared with the original input vector. Multiply element by element and multiply by the overall scaling factor. Output the corrected frequency and energy parameter matching result vector. .

[0027] The specific steps for generating the node frequency-energy risk matrix are as follows: Based on the dynamic frequency and energy pattern matrix, the frequency and energy values ​​of the main bearing and thrust bearing nodes are extracted sequentially. The frequency and energy values ​​are then filled into the attribute fields of the corresponding nodes. Within the attribute fields, parameters of the same type are categorized and sorted. Finally, the nodes and their attributes are combined into a complete list to generate a set of node attributes. Based on the node attribute set, a connection relationship is established between nodes according to the vibration propagation path. The position index of adjacent nodes is determined in sequence, and the risk calculation value of the node is filled into the risk distribution table using the index order. The node risk distribution tables are merged and sorted by node number to generate a node frequency-energy risk matrix. Based on the dynamic frequency-energy pattern matrix, the frequency and energy values ​​of the main bearing and thrust bearing nodes are read sequentially. The frequency value is accurate to 0.001 Hz, and the energy value is accurate to 0.0001 units. The read frequency value is filled into the frequency column of the node attribute field, and the energy value is filled into the energy column of the node attribute field. The frequency column is sorted in ascending order in the attribute field. The sorting results of the frequency column and energy column are combined into node attribute entries. Each entry contains the node number, frequency value, and energy value. The node attribute entries are arranged in ascending order by node number and stored in list format to generate a node attribute set. Based on the node attribute set, connections are established between nodes according to the vibration propagation path. First, node pairs are read sequentially from the path table, with the preceding node number as the source index and the following node number as the target index, to create a set of adjacency index pairs. The index numbers are all kept in integer form. The positional order of each set of adjacency indices is determined sequentially, and the risk calculation values ​​are filled in according to the order of the adjacency indices. The risk value is taken with a precision of 0.0001 units. The risk values ​​corresponding to each adjacency index are filled into the specified columns of the risk distribution table. The generated risk distribution table is sorted in ascending order by node number. The sorted risk distribution tables are merged into a complete two-dimensional matrix file to generate the node frequency-energy risk matrix.

[0028] The specific steps for generating a cross-component risk transmission path set are as follows: Based on the node frequency-energy risk matrix, the risk values ​​of each node are arranged in descending order. The risk values ​​of each node are compared with the set threshold. Nodes that are higher than the threshold are selected and marked as high-risk nodes. The frequency, energy and risk value of the nodes are recorded to generate a set of high-risk nodes. Based on the set of high-risk nodes, the transmission path between nodes is constructed one by one according to the connection relationship between the nodes, the connection order between each high-risk node is recorded, the transmission path is arranged in chronological order, and the path information is associated with the timestamp of the node to generate a set of cross-component risk transmission paths. Based on the node frequency-energy risk matrix, the risk values ​​of each node are sorted in descending order. The sorting method is set with a comparison precision of 0.0001 units. The node number and corresponding risk value in the sorting result are read sequentially. The threshold is set to 0.75 units. The risk value of each node is compared with the threshold one by one. The node number with the risk value greater than the threshold is added to the high-risk marker list. A high-risk identification code is generated for each node in the marker list. The identification code is 8 bits long and is generated by combining the node number with a fixed prefix. The frequency value and energy of the node are read. The frequency value is taken with a precision of 0.001 Hz and the energy value is taken with a precision of 0.0001 units. At the same time, the risk value is retained to four decimal places and stored in the record table along with the frequency value and energy value. The column order of the table is node number, frequency value, energy value and risk value, generating a set of high-risk nodes. Based on the set of high-risk nodes, the transmission paths between high-risk nodes are constructed one by one according to the node connection relationship table. The starting node is selected from the set, and the next node directly connected to it is found according to the connection relationship. The two nodes form a path segment. The path segment is recorded in the path table in the form of node number pairs. The path segments are combined into a complete path according to the connection order. During the path combination process, the timestamp information of each node is recorded. The timestamp precision is in seconds. The path is globally sorted according to the order of the timestamps. The sequence number of the path segment and the node number are stored together in the path file. The path file columns are in the order of sequence number, starting node number, target node number, starting timestamp, and target timestamp. The path file is renumbered in ascending order of node number to generate a set of cross-component risk transmission paths.

[0029] Please see Figure 2 A hydraulic turbine vibration data analysis system, the system comprising: Frequency and energy feature construction module: Based on the vibration signal sequence of rotor, guide vane and main bearing, frequency band division and amplitude and period extraction are performed, waveform segments are constructed into graph structure and node features are updated, and frequency and energy matching table is established by combining synchronous segments to generate multi-scale synchronous frequency and energy feature set; Model distribution analysis module: Based on the multi-scale synchronous frequency and energy feature set, extract the amplitude difference and rate of change of the frequency points of the thrust bearing and tailrace pipe, screen sample points that meet the spacing conditions, record the frequency position and energy ratio, and generate the model center frequency and energy distribution table. Operating condition matching and correction module: Based on the mode center frequency energy distribution table, it obtains operating condition parameters and establishes a mapping with guide vane clearance and rotor support, matches frequency position and energy ratio, adjusts low similarity label parameters, and generates dynamic frequency energy mode matrix. Node risk modeling module: Based on the dynamic frequency energy pattern matrix, assign values ​​to the node attributes of the main bearing and thrust bearing, determine the node connection relationship, generate a risk value distribution table, and form a node frequency energy risk matrix; Cross-component risk tracking module: Based on the node frequency and energy risk matrix, it filters high-risk tailrace pipes and rotor nodes, constructs and sorts risk transmission paths, and generates a set of cross-component risk transmission paths.

[0030] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for analyzing vibration data of a water turbine, characterized in that, Includes the following steps: Step 1: Based on the vibration signal sequence of the turbine rotor, guide vanes, and main bearing, decompose the waveforms of different frequency bands and extract the amplitude and period curves. Use graph neural network to update node features and select high-synchronization segments to establish a frequency and energy matching table, and generate a multi-scale synchronous frequency and energy feature set. Step 2: Based on the multi-scale synchronous frequency and energy feature set, extract the amplitude difference and rate of change of the frequency points of the thrust bearing and the tailrace pipe, sort and screen the sample points that meet the spacing conditions, record the frequency position and energy ratio, and generate the mode center frequency and energy distribution table. Step 3: Based on the frequency and energy distribution table of the model center, extract the turbine operating condition parameters, map the model center to the guide vane clearance and rotor support matrix, use graph neural network to calculate node feature similarity and filter low similarity labels, update frequency position and energy ratio, and generate dynamic frequency and energy model matrix. Step 4: Based on the dynamic frequency energy mode matrix, assign node attributes to the main bearing and thrust bearing, establish node set and vibration propagation path, construct node risk value distribution table according to the adjacency matrix order, and generate node frequency energy risk matrix; Step 5: Based on the node frequency and energy risk matrix, sort the node risk values ​​and mark the tailrace pipe and rotor nodes that are higher than the threshold. Form risk transmission paths according to the connection order and arrange them by time to generate a cross-component risk transmission path set.

2. The method for analyzing turbine vibration data according to claim 1, characterized in that, The multi-scale synchronous frequency and energy feature set includes a frequency location set, an energy density set, a phase difference parameter set, and an amplitude variance set. The mode center frequency and energy distribution table includes a mode center frequency location column, a mode center energy ratio column, and an amplitude difference rate column. The dynamic frequency and energy mode matrix includes frequency location parameters, energy ratio parameters, and a set of matching labels with guide vane clearance and rotor support. The node frequency and energy risk matrix includes node number, frequency value, energy value, and risk value. The cross-component risk transmission path set includes a high-risk node sequence, inter-node connection relationships, and time sequence information.

3. The method for analyzing turbine vibration data according to claim 1, characterized in that, The specific steps for generating the multi-scale synchronous frequency energy feature set are as follows: Based on the vibration signal sequence of the turbine rotor, guide vanes, and main bearing, the waveforms of each frequency band in the signal are extracted, and the signal is clipped according to the set frequency band. The signal in each frequency band is divided into several time periods, and the amplitude and period in each time period are recorded and calculated to obtain the frequency band waveform segments. Based on the waveform segments of the frequency band, the amplitude and period characteristics of each waveform segment are compared, the variation trend of amplitude and period within each waveform segment is calculated, the part with high waveform synchronization is selected, and the waveforms are merged in time order to generate merged waveform data, thus obtaining the synchronized waveform segment. Based on the aforementioned synchronous waveform segments, each waveform segment is constructed as a node in a graph structure, and parameters such as mean frequency, energy density, phase difference, and amplitude variance are written into the node attributes. Weighted connecting edges are established based on temporal adjacency, frequency difference, and energy difference to form a complete graph structure. A graph neural network is used to perform information transmission and node feature updates on the graph structure. The feature similarity between the updated nodes is calculated and sorted from high to low values ​​to generate a multi-scale synchronous frequency and energy feature set.

4. The method for analyzing turbine vibration data according to claim 1, characterized in that, The graph neural network specifically includes the following steps: First, for the graph structure constructed from synchronous waveform segments, each node is initialized as a multi-dimensional feature vector containing the mean frequency, energy density, phase difference, and amplitude variance. Based on the weighted edge relationship between nodes, the aggregated features of adjacent nodes and the features of the center node are updated in each iteration. The feature vectors of neighboring nodes are aggregated by averaging, and the aggregation weight is set according to the frequency difference and energy difference. The aggregation result and the current feature vector of the center node are transformed by a nonlinear transformation function to generate updated node features. The feature similarity between each pair of nodes in the graph is calculated by the cosine similarity function. The node pairs are sorted from high to low similarity, and node combinations that meet the preset synchronization threshold are selected. The corresponding frequency and energy attributes are extracted to form a multi-scale synchronous frequency and energy feature set.

5. The method for analyzing turbine vibration data according to claim 1, characterized in that, The specific steps for generating the center frequency energy distribution table of the mode are as follows: Based on the multi-scale synchronous frequency energy feature set, each frequency point is sequentially located in the corresponding components of the thrust bearing and tailrace pipe, the amplitude is collected and the difference between the amplitudes of adjacent frequency points is calculated, the difference is compared point by point in frequency order and the rate of change is recorded, and the amplitude difference and the rate of change of the frequency points are combined into an ordered sequence to generate the frequency amplitude difference rate sequence. Based on the frequency amplitude difference rate sequence, the amplitude spacing between adjacent frequency points in the sequence is compared and it is determined whether it is greater than a set threshold. Frequency points that meet the conditions are marked as sample points. The frequency position and corresponding energy ratio of the sample points are extracted and stored in the table data. The parameters in the table are arranged and combined by row to generate the mode center frequency energy distribution table.

6. The method for analyzing turbine vibration data according to claim 1, characterized in that, The specific steps for generating the dynamic frequency energy mode matrix are as follows: Based on the center frequency energy distribution table of the mode, the operating parameters of the turbine are extracted, including speed, load, etc., and the correlation positions of each operating parameter with the guide vane clearance and rotor support are matched one by one. The matching information between the operating parameters and the corresponding positions is recorded and organized into a corresponding mapping table to generate an operating parameter mapping table. Based on the aforementioned working condition parameter mapping table, the center frequency and energy information of each mode are compared with the guide vane clearance and rotor support parameters. The frequency position and energy ratio of each mode center are matched with the corresponding parameters. An input vector is constructed and input into the residual network to obtain the residual correction value. The matching degree of the corrected matching result is calculated. Labels with low matching degree are selected and marked as labels to be updated, generating a set of low matching degree labels. Based on the set of low-matching tags, the frequency position and energy ratio of each tag are adjusted, the updated frequency position and energy ratio are organized into the corresponding tags, the matching relationship between the tags and the operating parameters is updated, and the updated mode matrix is ​​generated to obtain the dynamic frequency-energy mode matrix.

7. The method for analyzing turbine vibration data according to claim 1, characterized in that, The residual network specifically includes the following steps: First, the frequency position and energy ratio of each mode center are combined with the guide vane clearance and rotor support parameters to form a fixed-length input vector. The input vector is then passed through a multi-layer linear transformation module, with a short-circuit connection structure introduced between adjacent linear transformation layers. The output of the intermediate layer is directly added to the output of the current layer to form a residual mapping, generating a residual correction value vector with the same dimension as the original input. Then, the correction value vector is added element-wise to the original input vector to obtain the frequency and energy parameter matching result after residual correction. Finally, the matching degree value is calculated based on the Euclidean distance between the corrected vector and the corresponding operating parameters, and low-matching-degree samples are selected according to a preset matching degree threshold to generate a set of labels to be updated.

8. The method for analyzing turbine vibration data according to claim 1, characterized in that, The specific steps for generating the node frequency-energy risk matrix are as follows: Based on the dynamic frequency and energy mode matrix, the frequency and energy values ​​of the main bearing and thrust bearing nodes are extracted sequentially. The frequency and energy values ​​are then filled into the attribute fields of the corresponding nodes. Within the attribute fields, similar parameters are categorized and sorted. Finally, the nodes and their attributes are combined into a complete list to generate a node attribute set. Based on the set of node attributes, a connection relationship is established between nodes according to the vibration propagation path. The position index of adjacent nodes is determined in sequence, and the risk calculation value of the node is filled into the risk distribution table using the index order. The risk distribution tables of the nodes are merged and sorted by node number to generate a node frequency-energy risk matrix.

9. The method for analyzing turbine vibration data according to claim 1, characterized in that, The specific steps for generating the cross-component risk transmission path set are as follows: Based on the node frequency-energy risk matrix, the risk values ​​of each node are arranged in descending order. The risk values ​​of each node are compared with the set threshold. Nodes that are higher than the threshold are selected and marked as high-risk nodes. The frequency, energy and risk value of the nodes are recorded to generate a set of high-risk nodes. Based on the set of high-risk nodes, the transmission path between nodes is constructed one by one according to the connection relationship between the nodes, the connection order between each high-risk node is recorded, the transmission path is arranged in chronological order, and the path information is associated with the timestamp of the node to generate a set of cross-component risk transmission paths.

10. A hydraulic turbine vibration data analysis system, characterized in that, The method for analyzing turbine vibration data according to any one of claims 1-9, wherein the system comprises: Frequency and energy feature construction module: Based on the vibration signal sequence of rotor, guide vane and main bearing, frequency band division and amplitude and period extraction are performed, waveform segments are constructed into graph structure and node features are updated, and frequency and energy matching table is established by combining synchronous segments to generate multi-scale synchronous frequency and energy feature set; Model distribution analysis module: Based on the multi-scale synchronous frequency and energy feature set, extract the amplitude difference and rate of change of the frequency points of the thrust bearing and the tailrace pipe, screen sample points that meet the spacing conditions, record the frequency position and energy ratio, and generate the model center frequency and energy distribution table. Operating condition matching and correction module: Based on the mode center frequency energy distribution table, obtain operating condition parameters and establish a mapping with guide vane clearance and rotor support, match frequency position and energy ratio, adjust low similarity label parameters, and generate dynamic frequency energy mode matrix; Node risk modeling module: Based on the dynamic frequency energy mode matrix, assign node attributes to the main bearing and thrust bearing, determine the node connection relationship and generate a risk value distribution table to form a node frequency energy risk matrix; Cross-component risk tracking module: Based on the node frequency and energy risk matrix, high-risk tailrace pipes and rotor nodes are screened, risk transmission paths are constructed and sorted, and a cross-component risk transmission path set is generated.

Images