Graph-driven modal decomposition method, system and device for extracting rotating machinery fault features

Abstract

The application belongs to the technical field of mechanical equipment vibration signal processing and fault diagnosis. A kind of graph-driven modal decomposition method, system and equipment for rotating machinery fault feature extraction are provided, the original vibration signal is discretized and the initial finite impulse response filter bank is constructed, the initial modal set is obtained after filtering processing, and the current modal set is obtained according to the reweighted kurtosis iterative update filter coefficient;Extract multi-dimensional modal relationship features, determine adaptive weights and construct fully connected weighted undirected modal relationship graph according to information distribution, obtain modal community by modularity optimization community discovery, carry out group constraint filter joint optimization to obtain updated modal set with community as unit. Loop execution mapping, community division and joint optimization until meeting termination condition, output final modal set. The application realizes global correlation representation of modal, merges similar fault information, and improves the processing effect of periodic fault modal aggregation.

Description

Technical Field

[0001] This invention relates to the field of vibration signal processing and fault diagnosis technology for mechanical equipment, specifically to a graph-driven modal decomposition method, system, and device for extracting fault features of rotating machinery. Background Technology

[0002] Rotating machinery transmission components in industrial settings are subjected to long-term load fluctuations and multi-source excitation coupling. Fault information generated when components experience early damage is often mixed within complex vibration signals. Modal decomposition (MDD) is an important signal processing technique in rotating machinery fault diagnosis. It decomposes the original vibration signal into several modal components with different physical meanings, facilitating the extraction of periodic impact characteristics hidden within the signal. Various existing MDD methods can be used for vibration signal analysis. Some methods, combined with impact evaluation indices to drive filter iteration, are adapted to the extraction requirements of fault impact characteristics and have already been practically used in the condition analysis of components such as bearings and gearboxes, providing signal-level support for equipment condition assessment.

[0003] Most existing modal processing methods employ pairwise comparisons, relying solely on a single criterion to measure the correlation between modes, thus only capturing localized relationships. This approach lacks a method for global correlation modeling of all modes. The mode selection and merging process is easily affected by the processing order, and the iterative process generates unnecessary computational steps, making it difficult to effectively merge similar fault information scattered across different modes. When dealing with measured vibration signals containing multiple coupled fault components, the shortcomings of this approach become even more apparent, exhibiting limitations in the aggregation and processing of periodic fault modes within the vibration signal. Summary of the Invention

[0004] To address the limitations of existing modal processing methods, which can only capture local modal relationships, struggle to perform global correlation modeling for all modes, are susceptible to interference from the order of modal processing, and are difficult to merge similar fault information scattered across different modes, this invention provides a graph-driven modal decomposition method, system, and device for extracting fault features from rotating machinery. By constructing a modal relationship graph, it achieves global correlation representation of modes and conducts group constraint joint optimization at the modal community level, improving the aggregation processing effect of periodic fault modes and realizing stable extraction of fault features from rotating machinery vibration signals.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a graph-driven mode decomposition method for extracting fault features of rotating machinery.

[0006] A graph-driven mode decomposition method for extracting fault features of rotating machinery includes the following steps: The original vibration signal is discretized, the corresponding frequency band is divided, and an initial finite impulse response filter bank is constructed. The initial finite impulse response filter bank is applied to the original vibration signal to obtain the initial mode set. Based on the reweighted kurtosis corresponding to each mode in the initial mode set, the coefficients of each finite impulse response filter are iteratively updated to obtain the current mode set; Calculate the multidimensional modal relationship features between any two modes in the current modality set, determine the corresponding adaptive weights based on the information distribution of each multidimensional modal relationship feature, and construct a fully connected weighted undirected modal relationship graph based on the multidimensional modal relationship features and the corresponding adaptive weights; Community discovery based on modularity optimization is performed on the modality graph to obtain multiple modal communities. Group constraint filters are jointly optimized for each modal community to obtain the group representative filter and group representative mode corresponding to each modal community, forming an updated mode set. The updated modality set will be re-executed with multidimensional modal relationship feature calculation, modal relationship graph construction, community detection and joint optimization of group constraint filter until the preset termination condition is met, and the final modality set will be output.

[0007] In one implementation of the first aspect of the present invention, the coefficients of each finite impulse response filter are iteratively updated based on the reweighted kurtosis corresponding to each mode in the initial mode set, including: The corresponding mode is divided into signal segments of equal length under multiple candidate segment numbers, and the kurtosis of each signal segment under each candidate segment number is calculated; For each candidate segment number, the kurtosis of the corresponding signal segment is sorted in ascending order, and the corresponding normalized weights are sorted in descending order. The kurtosis and the corresponding normalized weights are weighted and summed to obtain the reweighted kurtosis for the corresponding candidate segment number. The number of candidate segments with the largest reweighted kurtosis is selected as the optimal number of segments for the corresponding mode.

[0008] As a further limitation of the first aspect of the present invention, for each signal segment under the optimal number of segments, a corresponding sliding observation matrix is ​​constructed, and the local filter coefficients corresponding to each signal segment are solved. The corresponding fusion weights are determined based on the kurtosis of each signal segment, and the local filter coefficients of each signal segment are weighted and fused to obtain the updated filter coefficients. The updated filter coefficients are normalized to obtain the finite impulse response filter to be used in the next iteration.

[0009] In one implementation of the first aspect of the present invention, in the step of calculating the multidimensional modal relationship features between any two modes in the current modal set, the multidimensional modal relationship features include at least one of periodic consistency, envelope spectrum consistency, reweighted kurtosis similarity, filter band overlap rate, center frequency similarity, bandwidth similarity, and energy similarity. The corresponding adaptive weights are determined based on the information distribution of each multidimensional modal relationship feature, including: Statistically analyze the similarity distribution of all modal pairs across all dimensions, and calculate the information entropy of the corresponding similarity distribution across each dimension; The discriminant degree is calculated based on the information entropy of each dimension. The discriminant degree of each dimension is then normalized to obtain the adaptive weights corresponding to the multidimensional modal relationship features.

[0010] In one implementation of the first aspect of the present invention, a fully connected weighted undirected modal relationship graph is constructed based on multidimensional modal relationship features and corresponding adaptive weights, including: Each mode in the current mode set is used as a node in the mode relationship graph, and a connection edge is established between any two nodes; Based on the similarity of the corresponding modalities of the two nodes in each dimension and the corresponding adaptive weights, the comprehensive relationship strength between the two nodes is calculated and used as the edge weight of the corresponding connection edge. The fully connected weighted undirected modal graph is composed of all nodes and the edge weights of all connected edges.

[0011] In one implementation of the first aspect of the present invention, community discovery based on modularity optimization is performed on the modal relationship graph to obtain multiple modal communities, including: Calculate the total edge weight and the weighted degree of each node in the modal graph, and calculate the modularity of the modal graph based on the total edge weight and the weighted degree of each node; To maximize modularity, each node is divided into communities, resulting in multiple modal communities.

[0012] In one implementation of the first aspect of the present invention, joint optimization of group constraint filters is performed for each modal community to obtain the group representative filter and group representative mode corresponding to each modal community, including: The filter with the largest reweighted kurtosis within the modal community is used as the initial value for the group representative filter; The optimization objective is to maximize the reweighted kurtosis of the mode corresponding to the group representative filter, while introducing the intra-group filter distance constraint to construct the group-constrained optimization objective function. Solve the objective function of group constraint optimization to obtain the optimized group representative filter. Apply the group representative filter to the original vibration signal to obtain the corresponding group representative mode. The regularization parameter corresponding to the distance constraint of the intra-group filter is adaptively linked to the degree of dispersion of the filter coefficients within the modal community.

[0013] In one implementation of the first aspect of the present invention, the preset termination condition includes any one of the following: The current modality set contains fewer than or equal to the preset target modality number; the community partitioning results obtained from two consecutive community detections are consistent; and the number of iterations reaches the preset maximum number of iterations.

[0014] Secondly, the present invention provides a graph-driven mode decomposition system for extracting fault features of rotating machinery.

[0015] A graph-driven mode decomposition system for extracting fault features of rotating machinery includes: The initialization unit is configured to: discretize the original vibration signal, divide the corresponding frequency band range and construct an initial finite impulse response filter bank, apply the initial finite impulse response filter bank to the original vibration signal, and obtain an initial mode set; The filter iteration unit is configured to iteratively update the coefficients of each finite impulse response filter based on the reweighted kurtosis corresponding to each mode in the initial mode set, so as to obtain the current mode set. The graph building unit is configured to: calculate the multidimensional modal relationship features between any two modes in the current modality set, determine the corresponding adaptive weights based on the information distribution of each multidimensional modal relationship feature, and construct a fully connected weighted undirected modal relationship graph based on the multidimensional modal relationship features and the corresponding adaptive weights; The community optimization unit is configured to: perform community discovery based on modularity optimization on the modality graph to obtain multiple modal communities; perform joint optimization of group constraint filters for each modal community to obtain the group representative filter and group representative mode corresponding to each modal community, forming an updated modality set; The iteration termination unit is configured to re-execute the multidimensional modal relationship feature calculation, modal relationship graph construction, community detection and group constraint filter joint optimization on the updated modal set until the preset termination condition is met, and output the final modal set.

[0016] Thirdly, the present invention provides a computer device, comprising: a processor and a computer-readable storage medium; A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the graph-driven mode decomposition method for extracting rotating machinery fault features according to the first aspect of the present invention.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention utilizes reweighted kurtosis to iterate filters and obtain the current mode set. It then constructs a mode relationship graph by combining multi-dimensional mode relationship features and adaptively determined feature weights, thus representing the global correlation between modes. This differs from processing methods that rely solely on a single criterion for pairwise comparisons, reducing interference from the order of mode processing. The invention groups modes through community discovery, performs joint optimization of group-constrained filters using mode communities as units, and iteratively executes graph construction, community partitioning, and joint optimization until a termination condition is triggered. This allows for the merging of similar fault information scattered across different modes. The combined use of multiple types of mode relationship features with adaptive weights adapts to changes in mode features during iteration, ensuring that the mode correlation measurement results closely match the actual state of the modes output in each iteration, helping to reduce invalid processing steps during iteration. In the scenario of fault feature extraction from rotating machinery vibration signals, this invention improves the aggregation processing effect of periodic fault modes, facilitating the acquisition of relatively stable mode components from vibration signals with mixed noise and multi-source interference, providing a reliable signal analysis foundation for subsequent rotating machinery fault identification.

[0018] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0019] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0020] Figure 1 A schematic flowchart of a graph-driven modal decomposition method for extracting fault features of rotating machinery, provided as an exemplary embodiment of the present invention; Figure 2 A schematic diagram of measured vibration signals of rotating machinery provided as an exemplary embodiment of the present invention, wherein, Figure 2 (a) in the diagram is the time-domain plot of the original vibration signal. Figure 2 (b) in the diagram is the original vibration signal spectrum. Figure 3 A Graph-RMD decomposition signal component graph (iterated twice) is provided as an exemplary embodiment of the present invention, wherein, Figure 3 (a) in the diagram is the time-domain waveform of the first component. Figure 3 (b) in the diagram is the spectrum of the first component. Figure 3 (c) in the diagram is the time-domain waveform of the second component. Figure 3 (d) in the diagram represents the spectrum of the second component; Figure 4The RMD decomposition signal component map (iterated twice) provided as an exemplary embodiment of the present invention, wherein, Figure 4 (a) in the diagram is the time-domain waveform of the first component. Figure 4 (b) in the diagram is the spectrum of the first component. Figure 4 (c) in the diagram is the time-domain waveform of the second component. Figure 4 (d) in the diagram represents the spectrum of the second component. Figure 4 (e) in the diagram is the time-domain waveform of the third component. Figure 4 In the diagram, (f) represents the spectrum of the third component. Figure 4 (g) in the diagram represents the time-domain waveform of the fourth component. Figure 4 (h) in the diagram represents the spectrum of the fourth component. Figure 4 In the diagram, (i) represents the time-domain waveform of the fifth component. Figure 4 In the diagram, (j) represents the spectrum of the fifth component. Figure 4 In the diagram, (k) represents the time-domain waveform of the sixth component. Figure 4 (l) in the diagram represents the spectrum of the sixth component. Figure 4 In the diagram, (m) represents the time-domain waveform of the seventh component. Figure 4 In this context, (n) represents the spectrum of the seventh component. Figure 4 (o) in the diagram represents the time-domain waveform of the eighth component. Figure 4 (p) in the diagram represents the spectrum of the eighth component; Figure 5 The RMD decomposition signal component map (iterated 10 times) provided as an exemplary embodiment of the present invention, wherein, Figure 5 (a) in the diagram is the time-domain waveform of the first component. Figure 5 (b) in the diagram is the spectrum of the first component. Figure 5 (c) in the diagram is the time-domain waveform of the second component. Figure 5 (d) in the diagram represents the spectrum of the second component; Figure 6 A schematic diagram of the framework of a graph-driven mode decomposition system for extracting fault features of rotating machinery, provided as an exemplary embodiment of the present invention; Figure 7 A schematic diagram of a computer device provided for an exemplary embodiment of the present invention. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0023] During operation, rotating machinery generates vibration signals from its drive and transmission components, exhibiting significant periodicity while being affected by random noise and non-target components. Modal decomposition (MDD) of these vibration signals can extract periodic features, providing a foundation for equipment condition analysis and fault identification. However, different modes exhibit feature correlations and redundancy. Accurately determining the degree of modal correlation and quickly obtaining representative periodic modes is crucial for improving the effectiveness of vibration signal decomposition. Existing methods such as EEMD improve modal aliasing by adding auxiliary noise and performing ensemble averaging, but this is computationally intensive and susceptible to noise and ensemble number. VMD obtains finite-bandwidth modes through variational optimization, but relies on a preset number of modes and penalty parameters. SVD extracts signal features using singular value decomposition, but lacks specificity for frequency structure and periodic characteristics. FMD achieves mode extraction through filter bank optimization, but it does not adequately utilize the relationships between multiple modes. RMD uses an FIR filter bank and reweighted kurtosis (RK) for adaptive mode extraction and processes modes pairwise through correlation coefficients. However, a single correlation index is difficult to fully describe the comprehensive correlation between modes, and pairwise fusion also has processing order dependence, which can easily lead to invalid fusion.

[0024] To address the aforementioned issues, this invention proposes a graph-driven modal decomposition (Graph-RMD) method for extracting fault features in rotating machinery. Based on the RK-driven filter iteration mechanism, modes are constructed as graph nodes. A weighted modal relationship graph is built using multidimensional modal relationship features and adaptive entropy weights. Community self-organization grouping is performed according to the graph structure, and joint optimization of group-constrained filters is carried out on a community-by-community basis. The optimized group representative filter re-participates in subsequent RK iterations, and the relationship graph is reconstructed based on the updated modes, achieving dynamic evolution of modal relationships. This method utilizes multidimensional, global modal association information to replace the traditional local pairwise judgment based on a single correlation coefficient, reducing the impact of fusion order on the decomposition results and minimizing invalid fusion and iteration processes. Testing with actual rotating machinery vibration signal data and comparison with traditional RMD demonstrates that Graph-RMD can reduce invalid iterations and computational time during the decomposition process, and improve the extraction accuracy and decomposition stability of periodic target modes.

[0025] This invention decomposes the original vibration signal into a filter bank and iteratively optimizes each filter using reweighted kurtosis to obtain an initial mode set with strong impact characteristics and periodic information. Furthermore, it extracts multidimensional modal relationship features from the modal signals and their corresponding filters, constructs a modal relationship graph, and uses the graph structure to self-organize and group the modes. Finally, it performs joint optimization of the filters on a mode group basis, thereby reducing redundant information between modes and obtaining a mode set with strong characterization capabilities. This achieves self-organized reduction and mode fusion of the number of modes.

[0026] More specifically, such as Figure 1As shown, it includes the following four steps: Step S1: Initialize the original vibration signal and filter. Discretize the original vibration signal and construct the initial filter bank.

[0027] Step S2: Iterative optimization of the filter based on reweighted kurtosis to obtain the current mode set.

[0028] Step S3: Construction of Multidimensional Modal Relationship Graph Model. Calculate the multidimensional similarity relationships between modes and construct a fully connected weighted undirected modal relationship graph.

[0029] Step S4: Joint optimization of group-constrained filters based on graph structure. Self-organizing grouping and group-constrained filters are jointly optimized based on the modal graph, the modal graph is updated, and a termination decision is made.

[0030] Step S1 also includes: Step S101: Definition of the original vibration signal. The vibration signal of the rotating machinery to be processed... Represented as a discrete-time series and converted into a column vector: , The signal length is and the sampling frequency is . .

[0031] Step S102: Divide the frequency band corresponding to the original vibration signal and construct an initial FIR filter bank. Divide the normalized frequency range of the signal into... The frequency sub-band, the first The frequency band range of each sub-band Represented as: When using an equal-interval frequency band division method, the first... The lower cutoff frequency of each frequency sub-band and upper cutoff frequency They are respectively: (1); For each frequency subband, the construction length is... A finite impulse response bandpass filter, defining the first... The coefficient vector of each filter is FIR filters can be obtained using the window function method, frequency sampling method, or other methods that can construct bandpass filters based on given upper and lower cutoff frequencies. Preferably, a window function is used to truncate the ideal bandpass filter to obtain filter coefficients of finite length.

[0032] It should be further noted that step S2 also includes: Step S201: Apply each filter to the original vibration signal to obtain the initial mode: (2); in, For the first In the first iteration The modal signals corresponding to each filter, when step S2 is executed for the first time. =0, For the h-th iteration The first FIR filter One coefficient, For delay The original vibration signal after sampling points. This yields the initial mode set corresponding to the current filter bank. : , Indicates the first The number of finite impulse response filters in the filter bank during round iteration, i.e. the number of modes contained in the current mode set.

[0033] Step S202: Calculate the reweighted kurtosis of each initial mode. (The text abruptly ends here, likely due to an incomplete sentence or a formatting error.) The modes are divided into Let the nth signal segment be of equal length. Each signal segment is , Given the number of sampling points for a single signal segment, calculate the kurtosis of each signal segment. : (3); in, For the h-th iteration The modality of the first Kuness of a signal segment This represents the mean of the corresponding signal segment. Reweighting coefficients are generated based on the relative magnitudes of the kurtosis among the signal segments. The kurtosis is then sorted in ascending order, and the corresponding normalized weights are rearranged in descending order. The number of segments in the current segment is then calculated. The weighted kurtosis below: (4); in, For the first The modality in the number of segments The weighted kurtosis under the following conditions This represents the kurtosis of the segments arranged in ascending order. The normalized weights, which correspond to and are arranged in descending order, are obtained by normalizing the kurtosis of each signal segment: (5); Step S203: Determine the optimal number of segments for each mode. The loop index for summing fragments, given the set of candidate segment counts. , This represents the minimum number of candidate segments. This represents the maximum number of candidate segments. This indicates the number of candidate segments currently selected. The segment with the largest reweighted kurtosis is selected as the optimal number of segments for the current mode. : (6); in, For the first The optimal number of segments corresponding to each mode. This represents the value of the variable that maximizes the objective function.

[0034] Step S204: Update the coefficients of each filter according to the reweighted kurtosis. For the first... There are modes, with a construction length of . Sliding observation matrix: (7); Represents the original vibration signal The There are 1 sampling points, among which . For the signal length. Regarding the optimal number of segments. For each corresponding signal segment, the local filter is solved based on the reweighted kurtosis target. For the ... The local filter coefficients of a signal segment are expressed as follows: (8); In the formula, For the first In the first iteration The modality of the first Local filter update results corresponding to each signal segment; Indicates the first In the first iteration The first mode The sliding observation matrix corresponding to each signal segment is constructed from the signal segment as shown in equation (7); Indicates the first The first mode The number of sampling points contained in a signal segment; This represents the signal amplitude at sampling point n of the p-th segment in the k-th mode. Since each signal segment is divided into equal lengths, the number of sampling points in each segment is the same. Further, the local filters are weighted and fused based on the kurtosis of each signal segment to obtain the updated filter. : (9); in, For the first The weighted update result of each filter, weights From the kurtosis of each segment The calculation yielded: (10); The updated results are normalized to obtain the filter to be used in the next iteration: (11); in, No. The filter coefficient vector used in the next filter iteration. ε represents the L2 norm of the vector, and ε represents a predefined positive decimal to avoid zero denominators during calculation.

[0035] Step S205: Iterate the filter update process and generate the current mode set. Re-extract the modes using the updated filter and repeat steps S202 to S204 to gradually bring the filter closer to a state with better reweighted kurtosis for the corresponding modes, thus obtaining the mode set corresponding to the current filter bank. : .

[0036] It should be further noted that step S3 also includes: Step S301: Calculate the multidimensional modal relationship features between modes. For any two modes in the current mode set... and Select at least one feature dimension that can reflect the relationship between modalities, calculate the similarity between two modalities on that feature dimension, and form a multidimensional relationship feature of the modality pair: (12); in, Indicates the first The first mode and the first The mode in the th ... Similarity across relational feature dimensions This represents the relation feature dimension, where D represents the number of relation feature dimensions currently used.

[0037] In one optional implementation, the modal relationship features include: periodic consistency, envelope spectrum consistency, reweighted kurtosis similarity, filter band overlap rate, center frequency similarity, bandwidth similarity, and energy similarity. These can be selected based on the actual signal type, analysis objective, and computational conditions; it is not required that all feature dimensions be used simultaneously.

[0038] 1) Periodic consistency, calculated based on the closeness between the periodic characteristics corresponding to the two modes: (13); in, and The first The and the first The characteristic period of each mode, ε represents a preset positive decimal to avoid the denominator being zero during the calculation process.

[0039] 2) Envelope spectrum consistency, calculated based on the Pearson correlation coefficient between the normalized envelope spectra of the two modes: (14); in, and The first The and the first The normalized envelope spectral vector of each mode, Describing covariance, It represents the standard deviation.

[0040] 3) Reweighted kurtosis similarity can be expressed as: (15); in, and The first The and the first The reweighted kurtosis corresponding to each mode.

[0041] 4) The filter bandwidth overlap rate can be expressed as: (16); in, and The first The and the first Each mode corresponds to the frequency band range of the filter. Indicates the bandwidth length.

[0042] 5) Center frequency similarity, calculated based on the closeness of the center frequencies of the two filters: (17); in, and The first The and the first The center frequency of each filter.

[0043] 6) Bandwidth similarity, calculated based on the similarity of the bandwidths of the two filters: (18); in, and The first The and the first The bandwidth of each filter, where ε represents a preset positive decimal to avoid zero denominators during calculation.

[0044] 7) Energy similarity, based on the correlation coefficient of the energy change trends of the two filters during the iterative optimization process. calculate: (19); in, and The first The and the first A time series vector consisting of the energy values ​​of each mode in the previous iterations. represent and covariance, represent standard deviation represent The standard deviation.

[0045] Step S302: Determine the adaptive feature weights based on the discriminative power of each modality relationship feature. For the first... For each relational feature dimension, the similarity distribution of all modality pairs along that dimension is statistically analyzed, and the information entropy is calculated based on the information distribution. (20); in, For the first Information entropy of each relation feature dimension Indicates the first The number of mode pairs participating in the statistics during each iteration, when the current number of modes is And when any two different modes form a mode pair and are not counted repeatedly, . For the first The modality pair in the th ... Normalized eigenvalues ​​across each feature dimension. Further, the weights are determined based on the information entropy of each feature dimension: (twenty one); in, No. Adaptive weights for each relation feature dimension. This represents the number of relation feature dimensions currently used. Feature weights are recalculated based on the current modality set each time a modal relation graph is constructed, so that the contribution of each relation feature to the graph structure can be adjusted accordingly when the modal relations change in different iteration stages.

[0046] Step S303: Construct a fully connected weighted modality graph based on the modal similarity relationships and adaptive weights of each feature dimension. All modes in the modality set are treated as nodes in the modality graph. With mode correspond: Any two modal nodes and Connections are established between all nodes, and the overall relationship strength between two nodes is calculated based on the similarity relationships and corresponding weights across multiple feature dimensions. (twenty two); in, For the modal relationship diagram, the first... The and the first Edge weights between nodes For the first Adaptive weights for each relation feature dimension. For two modes in the first Similarity across the feature dimensions of each relationship.

[0047] The modal relationship graph is composed of all modal nodes and their edge weights. , This is a modal relationship diagram. For a set of modal nodes, It is a set of connections between nodes. Let be the edge weight matrix.

[0048] It should be further noted that step S4 also includes: Step S401: Perform community discovery based on modularity optimization using the modal relationship graph. Based on the modal relationship graph obtained in step S303, divide the modal nodes into communities, grouping closely related modes into the same community. Define the total edge weight of the modal relationship graph: (twenty three); Define nodes weighting : (twenty four); Regarding node community affiliation The modularity of a modal graph is expressed as: (25); in, For the first Modularity of the wheel modal diagram and The first The and the first The weighted degree of each node; and These are the communities to which the two nodes belong; This is a community affiliation discrimination function; it takes a value of 1 when two nodes belong to the same community, and a value of 0 otherwise. This is achieved by optimizing the modularity. This yields the community partitioning results corresponding to the current modal relationship graph, and is represented as follows: , For the current community collection, For the first A community This represents the current number of communities.

[0049] In one alternative implementation, the Louvain algorithm is used to implement the above-mentioned modularity optimization process. However, the present invention is not limited to using the Louvain algorithm, and other modularity optimization methods that can perform community partitioning based on graph structure can also be used.

[0050] The number of communities discovered by the community When the number of modalities exceeds the target number, further community merging is performed based on the overall graph relationships between different communities. For any two communities... and The strength of inter-community relations is defined as: (26); in, For the community With the community The overall strength of the relationship between them and These represent the number of modalities contained in the two communities, respectively.

[0051] Step S402: Perform joint optimization of group-constrained filters for each mode group. For the first... Modal group The filters corresponding to all modes within the group are taken as the joint optimization object, and the reweighted kurtosis of the mode extracted by the group representative filter is taken as the main optimization objective. At the same time, the distance constraint of the filters within the group is introduced to construct the joint optimization objective function of the group-constrained filter: (27); in, For the first The group representative filter obtained by modal group optimization Represented by the filter coefficient vector The group-constrained filter constructed for the optimization variables jointly optimizes the objective function. The group representative mode is extracted from the original vibration signal by the group representative filter. Group representative mode The reweighted kurtosis, Let represent the coefficient vector of the filter to be optimized in the group-represented filter, and represent the optimization variables in the joint optimization process of the group-constrained filter. For mode group Inner The filter corresponding to each mode For the first Constraint regularization parameters for each mode group.

[0052] Regularization parameters The value of is adaptively linked to the dispersion of the filter coefficients within the group, and the calculation formula is: (28); in, This is a global scaling factor used to adjust the regularization strength in the joint optimization of group-constrained filters. For the first The degree of dispersion of filter coefficients within each mode group For the first The mean vector of all filters within a mode group. When the differences between filters within the group are large, the group constraint is strengthened to limit the deviation of the group representative filter from the original mode group; when the differences between filters within the group are small, the group constraint is weakened, making the optimization process focus more on improving the reweighted kurtosis. This regularization parameter adaptively links with the differences between filters within the group and is an important component of the group-level joint optimization in this invention.

[0053] When performing joint optimization of group-constrained filters, the filter with the largest reweighted kurtosis within the current mode group is used as the initial value for the group representative filter: (29); in, Let the initial values ​​of the group represent the filter. This is the index corresponding to the mode with the largest reweighted kurtosis within the mode group. For the corresponding filter, Mode group Inner The modal signal corresponding to each mode.

[0054] Step S403: Generate group representative modes, update the relational graph, and perform a termination judgment. Apply the group representative filter obtained in step S402 back to the original vibration signal to obtain the corresponding group representative modes: (30); in, For the first Each modal group corresponds to a group representing a mode. This is the corresponding group representative filter. The updated mode set is composed of the group representative modes corresponding to all mode groups: , For the first The modal set of round iterations, This represents the current number of mode groups.

[0055] After obtaining the updated modality set, the updated modalities are re-input into step S3 to recalculate the multidimensional relationship features between modalities, redetermine the adaptive weights of each feature dimension, and reconstruct the modality relationship graph: , This is the updated modal relationship diagram. This represents the graph structure obtained by re-performing multidimensional relation feature calculation, feature weight determination, and graph construction based on the current modality set.

[0056] This allows the modal changes generated in the previous round of group-level joint optimization to be fed back into the next round of modal relationship modeling, enabling the modal relationship graph to dynamically evolve as the modal set changes. After updating the modal relationship graph, a termination judgment is made on the current modal set and its community structure. Iteration stops when any of the following conditions are met: First, the current number of modes is less than or equal to the target number of modes. : ; Second, the community division results obtained from two consecutive rounds of community discovery remain unchanged, meaning that the community affiliation of all modal nodes has not changed: (31); in, The set of all modal community partitioning results in the h-th iteration. This is the set of all modal community partitioning results for the (h+1)th iteration.

[0057] Third, the current iteration count has reached the preset maximum iteration count. : .

[0058] Repeat steps S2 to S4 until the preset termination condition is met, ultimately obtaining the final mode set composed of the group representative modes corresponding to each mode group. Through the above process, the mode relationship graph is dynamically updated with the mode extraction and group-level joint optimization results. Community division and mode group boundaries are automatically determined based on the current mode relationships. Mode fusion no longer depends on a pre-specified fixed pairwise fusion order, but determines the fusion objects and fusion relationships based on the overall structure of the current mode relationship graph, thereby achieving self-organizing mode fusion.

[0059] To verify the effectiveness of the method of this invention in the analysis of real vibration signals of rotating machinery, the Graph-RMD method of this invention was compared with the traditional RMD method under the same experimental conditions and parameter configuration. Graph-RMD uses fewer outer iterations (2 times) for decomposition, while the results of traditional RMD under the same number of iterations (2 times) and a larger number of iterations (10 times) were compared to analyze the differences between the two methods in terms of modal aggregation speed and modal simplification.

[0060] Figure 2 This is a schematic diagram of the measured vibration signal of rotating machinery. Figure 3 This is the modal component graph after decomposition using the Graph-RMD method of this invention. Figure 4 The modal component maps are obtained by traditional RMD with the same number of iterations. Figure 5 This is the modal component map obtained by traditional RMD with a large number of iterations.

[0061] The decomposition results from Graph-RMD show that, with fewer iterations (2 iterations), the correlations between multiple initial modes can be uniformly described by the modal relationship graph, and modal grouping can be automatically completed based on the community structure. After joint optimization with group constraint filters, modes related to the main periodic vibration features gradually concentrate into representative modes, while the remaining modes mainly contain random noise and non-target vibration components. Therefore, Graph-RMD can obtain a relatively simple modal set with fewer iterations.

[0062] In contrast, traditional RMD still retains more modal components at the same number of iterations. Some modes with similar periodic structures exist in different components, and there is a certain degree of modal splitting for the same periodic vibration feature. As the number of RMD iterations increases, some of the dispersed periodic information gradually concentrates into a few modes, but multiple rounds of pairwise modal comparisons and processing are still required to obtain a more concise result.

[0063] Therefore, the Graph-RMD of this invention describes the global correlation between multiple modes using a modal relationship graph, determines the self-organized grouping of modes through community structure, and then performs joint optimization of filters on a mode group basis. Compared with the traditional RMD approach of processing modes pairwise, this invention can reduce the impact of fusion order on the decomposition results, enabling related modes to aggregate in fewer iterations, thereby improving the simplicity and computational efficiency of mode decomposition.

[0064] Figure 6 A graph-driven mode decomposition system for extracting fault features in rotating machinery is shown, comprising: The initialization unit 601 is configured to: discretize the original vibration signal, divide the corresponding frequency band range and construct an initial finite impulse response filter bank, apply the initial finite impulse response filter bank to the original vibration signal, and obtain an initial mode set; The filter iteration unit 602 is configured to iteratively update the coefficients of each finite impulse response filter based on the reweighted kurtosis corresponding to each mode in the initial mode set to obtain the current mode set; The graph building unit 603 is configured to: calculate the multidimensional modal relationship features between any two modes in the current modality set, determine the corresponding adaptive weights based on the information distribution of each multidimensional modal relationship feature, and construct a fully connected weighted undirected modal relationship graph based on the multidimensional modal relationship features and the corresponding adaptive weights; The community optimization unit 604 is configured to: perform community discovery based on modularity optimization on the modal relationship graph to obtain multiple modal communities; perform joint optimization of group constraint filters for each modal community to obtain the group representative filter and group representative mode corresponding to each modal community, forming an updated modal set; The iteration termination unit 605 is configured to: re-execute the multidimensional modal relationship feature calculation, modal relationship graph construction, community detection and group constraint filter joint optimization on the updated modal set until the preset termination condition is met, and output the final modal set.

[0065] It is understood that the aforementioned units can be individually or entirely merged into one or more other units, or some of the units can be further divided into multiple functionally smaller units. This achieves the same operation without affecting the technical effects of the embodiments of the present invention. The aforementioned units are based on logical functional division. In practical applications, the function of one unit can be implemented by multiple units, or the function of multiple units can be implemented by one unit. In other embodiments of the present invention, the system may also include other units. In practical applications, these functions can also be implemented with the assistance of other units, and can be implemented collaboratively by multiple units.

[0066] According to another embodiment of the present invention, the system of this embodiment can be constructed by running a computer program (including program code) capable of performing the steps involved in the corresponding method of the present invention on a general-purpose computing device, such as a computer, which includes processing elements and storage elements such as a central processing unit (CPU), random access memory (RAM), and read-only memory (ROM). The computer program can be recorded on, for example, a computer-readable recording medium, loaded into the aforementioned computing device through the computer-readable recording medium, and run therein.

[0067] Figure 7 A computer device is shown, which includes a processor 701, a communication interface 702, and a computer-readable storage medium 703. The processor 701, communication interface 702, and computer-readable storage medium 703 can be connected via a bus or other means.

[0068] The communication interface 702 is used to receive and send data. The computer-readable storage medium 703 can be stored in the memory of the electronic device. The computer-readable storage medium 703 is used to store computer programs, which include program instructions. The processor 701 is used to execute the program instructions stored in the computer-readable storage medium 703.

[0069] The processor 701 is the computing and control core of an electronic device. It is suitable for implementing one or more instructions, specifically for loading and executing one or more instructions to achieve the corresponding method flow or corresponding function.

[0070] Processor 701 is configured to perform the following procedure: The original vibration signal is discretized, the corresponding frequency band is divided, and an initial finite impulse response filter bank is constructed. The initial finite impulse response filter bank is applied to the original vibration signal to obtain the initial mode set. Based on the reweighted kurtosis corresponding to each mode in the initial mode set, the coefficients of each finite impulse response filter are iteratively updated to obtain the current mode set; Calculate the multidimensional modal relationship features between any two modes in the current modality set, determine the corresponding adaptive weights based on the information distribution of each multidimensional modal relationship feature, and construct a fully connected weighted undirected modal relationship graph based on the multidimensional modal relationship features and the corresponding adaptive weights; Community discovery based on modularity optimization is performed on the modality graph to obtain multiple modal communities. Group constraint filters are jointly optimized for each modal community to obtain the group representative filter and group representative mode corresponding to each modal community, forming an updated mode set. The updated modality set will be re-executed with multidimensional modal relationship feature calculation, modal relationship graph construction, community detection and joint optimization of group constraint filter until the preset termination condition is met, and the final modality set will be output.

[0071] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can implement the described functions using different methods for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0072] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic cable, digital cable) or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data processing device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive), etc.

[0073] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A graph-driven mode decomposition method for extracting fault features in rotating machinery, characterized in that, The process includes the following: The original vibration signal is discretized, the corresponding frequency band is divided, and an initial finite impulse response filter bank is constructed. The initial finite impulse response filter bank is applied to the original vibration signal to obtain the initial mode set. Based on the reweighted kurtosis corresponding to each mode in the initial mode set, the coefficients of each finite impulse response filter are iteratively updated to obtain the current mode set; Calculate the multidimensional modal relationship features between any two modes in the current modality set, determine the corresponding adaptive weights based on the information distribution of each multidimensional modal relationship feature, and construct a fully connected weighted undirected modal relationship graph based on the multidimensional modal relationship features and the corresponding adaptive weights; Community discovery based on modularity optimization is performed on the modality graph to obtain multiple modal communities. Group constraint filters are jointly optimized for each modal community to obtain the group representative filter and group representative mode corresponding to each modal community, forming an updated mode set. The updated modality set will be re-executed with multidimensional modal relationship feature calculation, modal relationship graph construction, community detection and joint optimization of group constraint filter until the preset termination condition is met, and the final modality set will be output.

2. The graph-driven mode decomposition method for extracting fault features of rotating machinery as described in claim 1, characterized in that, Based on the reweighted kurtosis corresponding to each mode in the initial mode set, the coefficients of each finite impulse response filter are iteratively updated, including: The corresponding mode is divided into signal segments of equal length under multiple candidate segment numbers, and the kurtosis of each signal segment under each candidate segment number is calculated; For each candidate segment number, the kurtosis of the corresponding signal segment is sorted in ascending order, and the corresponding normalized weights are sorted in descending order. The kurtosis and the corresponding normalized weights are weighted and summed to obtain the reweighted kurtosis for the corresponding candidate segment number. The number of candidate segments with the largest reweighted kurtosis is selected as the optimal number of segments for the corresponding mode.

3. The graph-driven mode decomposition method for extracting fault features of rotating machinery as described in claim 2, characterized in that, For each signal segment with the optimal number of segments, construct the corresponding sliding observation matrix and solve for the local filter coefficients corresponding to each signal segment. The corresponding fusion weights are determined based on the kurtosis of each signal segment, and the local filter coefficients of each signal segment are weighted and fused to obtain the updated filter coefficients. The updated filter coefficients are normalized to obtain the finite impulse response filter to be used in the next iteration.

4. The graph-driven mode decomposition method for extracting fault features of rotating machinery as described in claim 1, characterized in that, In the step of calculating the multidimensional modal relationship features between any two modes in the current modal set, the multidimensional modal relationship features include at least one of the following: periodic consistency, envelope spectrum consistency, reweighted kurtosis similarity, filter band overlap rate, center frequency similarity, bandwidth similarity, and energy similarity. The corresponding adaptive weights are determined based on the information distribution of each multidimensional modal relationship feature, including: Statistically analyze the similarity distribution of all modal pairs across all dimensions, and calculate the information entropy of the corresponding similarity distribution across each dimension; The discriminant degree is calculated based on the information entropy of each dimension. The discriminant degree of each dimension is then normalized to obtain the adaptive weights corresponding to the multidimensional modal relationship features.

5. The graph-driven mode decomposition method for extracting fault features of rotating machinery as described in claim 1, characterized in that, A fully connected weighted undirected modal relationship graph is constructed based on multidimensional modal relationship features and corresponding adaptive weights, including: Each mode in the current mode set is used as a node in the mode relationship graph, and a connection edge is established between any two nodes; Based on the similarity of the corresponding modalities of the two nodes in each dimension and the corresponding adaptive weights, the comprehensive relationship strength between the two nodes is calculated and used as the edge weight of the corresponding connection edge. The fully connected weighted undirected modal graph is composed of all nodes and the edge weights of all connected edges.

6. The graph-driven mode decomposition method for extracting fault features of rotating machinery as described in claim 1, characterized in that, Community discovery based on modularity optimization is performed on the modality graph, resulting in multiple modal communities, including: Calculate the total edge weight and the weighted degree of each node in the modal graph, and calculate the modularity of the modal graph based on the total edge weight and the weighted degree of each node; To maximize modularity, each node is divided into communities, resulting in multiple modal communities.

7. The graph-driven mode decomposition method for extracting fault features of rotating machinery as described in claim 1, characterized in that, Joint optimization of group constraint filters is performed for each modal community to obtain the group representative filter and group representative mode corresponding to each modal community, including: The filter with the largest reweighted kurtosis within the modal community is used as the initial value for the group representative filter; The optimization objective is to maximize the reweighted kurtosis of the mode corresponding to the group representative filter, while introducing the intra-group filter distance constraint to construct the group-constrained optimization objective function. Solve the objective function of group constraint optimization to obtain the optimized group representative filter. Apply the group representative filter to the original vibration signal to obtain the corresponding group representative mode. The regularization parameter corresponding to the distance constraint of the intra-group filter is adaptively linked to the degree of dispersion of the filter coefficients within the modal community.

8. The graph-driven mode decomposition method for extracting fault features of rotating machinery as described in claim 1, characterized in that, The preset termination conditions include any one of the following: The current modality set contains fewer than or equal to the preset target modality number; the community partitioning results obtained from two consecutive community detections are consistent; and the number of iterations reaches the preset maximum number of iterations.

9. A graph-driven mode decomposition system for extracting fault features of rotating machinery, characterized in that, include: The initialization unit is configured to: discretize the original vibration signal, divide the corresponding frequency band range and construct an initial finite impulse response filter bank, apply the initial finite impulse response filter bank to the original vibration signal, and obtain an initial mode set; The filter iteration unit is configured to iteratively update the coefficients of each finite impulse response filter based on the reweighted kurtosis corresponding to each mode in the initial mode set, so as to obtain the current mode set. The graph building unit is configured to: calculate the multidimensional modal relationship features between any two modes in the current modality set, determine the corresponding adaptive weights based on the information distribution of each multidimensional modal relationship feature, and construct a fully connected weighted undirected modal relationship graph based on the multidimensional modal relationship features and the corresponding adaptive weights; The community optimization unit is configured to: perform community discovery based on modularity optimization on the modality graph to obtain multiple modal communities; perform joint optimization of group constraint filters for each modal community to obtain the group representative filter and group representative mode corresponding to each modal community, forming an updated modality set; The iteration termination unit is configured to re-execute the multidimensional modal relationship feature calculation, modal relationship graph construction, community detection and group constraint filter joint optimization on the updated modal set until the preset termination condition is met, and output the final modal set.

10. A computer device, characterized in that, include: Processor and computer-readable storage media; A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program that, when executed by the processor, implements the graph-driven mode decomposition method for extracting fault features of rotating machinery as described in any one of claims 1 to 8.

Images