Rotating machinery fault diagnosis method based on delay-weighted view and netpro2vec

Abstract

The application discloses a rotating machinery fault diagnosis method based on delay-weighted visual graph and Netpro2vec, and belongs to the technical field of mechanical state monitoring and fault diagnosis. The method comprises the following steps: collecting a rotating machinery vibration signal; converting the signal into a delay-weighted visual graph, wherein, when judging whether two data points are visual, adjacent nodes of the two data points are ignored, and a weight is assigned to a visual edge, and the weight is defined as the minimum vertical distance between the two points and all data points therebetween; based on the graph, a low-dimensional feature vector is extracted by using the Netpro2vec method, specifically, a word sequence is generated by calculating a node distance distribution and a transition matrix, and a document is formed, and then a vector representation is obtained by learning through a document embedding model; and finally, the feature vector is input into a classifier to obtain a diagnosis result. The application enhances the noise resistance through a delay mechanism, and finely depicts the signal structure through a weighting mechanism, so that the precision and robustness of fault diagnosis are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of mechanical condition monitoring and fault diagnosis technology, and in particular to a method for diagnosing rotating machinery faults based on delay-weighted visualization and Netpro2vec. Background Technology

[0002] Rotating machinery, such as gearboxes and bearings, is a critical piece of equipment in industrial production. Faults occurring during their operation not only affect production efficiency but can also lead to serious safety accidents. Fault diagnosis technology based on vibration signal analysis is currently the mainstream method, its core being the extraction of effective features characterizing the health status of equipment from complex vibration signals.

[0003] Visibility Graphs (VGs) are a method that maps one-dimensional time series data to complex networks, revealing hidden structural features and dynamic characteristics within the data. In recent years, they have been introduced into the field of rotating machinery fault diagnosis. For example, existing techniques (HVG-GAT) convert vibration signals into horizontal visibility graphs and utilize graph attention networks for feature learning and classification. Another technique (WHVG-GCN) constructs weighted horizontal visibility graphs, attempting to suppress noise through weights and combining them with graph convolutional networks for classification.

[0004] However, existing fault diagnosis methods based on visual data still have the following significant drawbacks:

[0005] (1) Weak noise resistance: For continuous vibration signals with high sampling frequency, the amplitudes of adjacent data points are very close. According to the definition of a visual graph, the judgment of the visual relationship between data points depends heavily on the amplitudes of their adjacent data points. Therefore, small noise disturbances in the data (manifested as small changes in the amplitude of individual data points) may cause drastic changes in the visual relationship, making the constructed graph structure unstable and the feature extraction robustness poor.

[0006] (2) Coarse information representation: Traditional visualization only establishes a binary relationship between nodes as "visible" or "invisible", ignoring the important continuous information of "visibility". This coarse binary representation may cause the loss of subtle but critical fault feature differences in the signal, thus limiting the further improvement of diagnostic accuracy. Summary of the Invention

[0007] The purpose of this invention is to provide a method for diagnosing rotating machinery faults based on delay-weighted visualization and Netpro2vec, which solves the problems of weak noise resistance and coarse information representation.

[0008] To achieve the above objectives, this invention provides a method for diagnosing rotating machinery faults based on delay-weighted visual graphs and Netpro2vec, comprising the following steps:

[0009] S1. Collect vibration signals from rotating machinery;

[0010] S2. Convert the vibration signal into a delay-weighted visual graph. ;

[0011] S3. Based on the aforementioned delay-weighted visual chart The image was extracted using the Netpro2vec method. The low-dimensional feature vector;

[0012] S4. Input the low-dimensional feature vector into the trained classifier and output the fault diagnosis result of the rotating machinery.

[0013] Preferably, the construction of the delayed weighted visual chart in step S2 is the core of the improvement of this invention, and its specific rules include two key innovations:

[0014] First, a delay visibility mechanism is introduced. For any two data points in the vibration signal... and When determining whether they are visible, all [items] are no longer considered. Instead of considering the midpoint, only consider all The midpoint (i.e., ignoring the two points that are adjacent to each other) (Data points), and requires that these intermediate points all satisfy the traditional visual inequality. This mechanism is equivalent to setting up a "buffer" in the visual judgment, making the visual relationship between nodes less susceptible to changes due to small noise fluctuations of their direct neighbor nodes, thereby significantly improving the noise resistance of the graph structure.

[0015] Secondly, a weighted visibility mechanism is introduced. For two nodes determined to be visible... and The edges connected to it are assigned a weight. The weight is defined as the connection node. and The straight line, intersecting all intermediate data points (including those ignored). The minimum vertical distance between neighboring points. This weight value quantifies the "clarity" or "intensity" of the visibility between two points; a larger weight indicates less occlusion of the visible path and a stronger relationship. This allows the constructed graph to contain not only topological connectivity information but also rich intensity information, enabling a more refined characterization of differences in signal waveform structures.

[0016] Preferably, the delay parameter The value can be set according to the characteristics of the actual vibration signal, and the range is usually from 1 to 5. Experiments have verified that, for the gearbox vibration signal in the described embodiment, It can achieve a good balance between noise resistance and feature retention.

[0017] Preferably, step S3 uses the Netpro2vec method to extract graph features, specifically including:

[0018] S31. Calculation Graph The distance distribution vector of each node in the graph is used to capture the global topological features of the graph.

[0019] S32. Calculation diagram The transition matrix of each node is used to capture local neighborhood features based on random walks.

[0020] S33. Convert the two features of each node into a sequence of “words” respectively.

[0021] S34. Integrate the words from all nodes to form a representation of the entire graph. The "document".

[0022] S35. Use a document embedding model (such as PV-DBOW) to learn the distributed representation of the document and obtain a low-dimensional, dense feature vector of the graph.

[0023] Preferably, global feature documents and local feature documents can be generated separately, and the corresponding feature vectors can be learned and then fused to make comprehensive use of global and local information.

[0024] Preferably, to further explore the structural information of the graph, after step S2, the process may further include constructing the delayed-weighted visual graph. The steps of the first-order weighted edge graph (WSGN). The nodes of the WSGN correspond to the original graph. The edges are defined, reflecting the relationships between edges in the original graph, and their weights are inherited from the corresponding edges in the original graph. Subsequently, the original graph is analyzed... Features are extracted and fused with the WSGN graph. This step enhances the expressive power of the features from the perspective of edge association.

[0025] Preferably, the classifier may employ a neural network model containing one or more fully connected layers and using a Softmax function as the output layer.

[0026] This invention effectively solves the problems of poor noise resistance and coarse information representation in traditional methods by using a visualization construction method that combines "delay" and "weighting". Furthermore, by combining Netpro2vec's powerful graph structure feature learning capabilities, it can stably and precisely extract highly discriminative fault features from vibration signals in a noisy background, ultimately achieving high-precision and robust fault diagnosis of rotating machinery through a classifier.

[0027] Therefore, the rotating machinery fault diagnosis method based on delay-weighted visualization and Netpro2vec using the above structure of the present invention has the following beneficial effects:

[0028] (1) The delayed visualization mechanism of the present invention effectively isolates the interference of local noise on the graph structure, enabling the method of the present invention to construct a stable feature map under strong noise background, which significantly improves the robustness of the diagnostic model.

[0029] (2) The weighted visualization mechanism of the present invention introduces continuous weight information, which enables the visualization to distinguish the subtle differences in the signal waveform structure under different fault states more finely, providing the classifier with richer features.

[0030] (3) The present invention combines the feature learning strategy of delayed weighted visual image with Netpro2vec. Experiments on public datasets show that the key indicators such as diagnostic accuracy and recall rate of the present invention are superior to traditional visual image methods and other mainstream diagnostic models, especially under low signal-to-noise ratio conditions.

[0031] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0032] Figure 1 This is a flowchart of the overall process for fault diagnosis of rotating machinery based on delay-weighted visualization in this invention.

[0033] Figure 2 This is a schematic diagram illustrating the definition of the edge weights in a visible view according to an embodiment of the present invention;

[0034] Figure 3 A schematic diagram comparing traditional visual representations of noise-free signals and noise-contaminated signals;

[0035] Figure 4 This is a schematic diagram illustrating the definition of a delay visualization.

[0036] Figure 5 A schematic diagram comparing the delay of noise-free signals and noise-contaminated signals;

[0037] Figure 6 Construct a schematic diagram for the weighted delay visualization;

[0038] Figure 7 A schematic diagram illustrating the method for constructing WSGN from a visual image;

[0039] Figure 8 This is a schematic diagram of the SEU gearbox dataset acquisition platform;

[0040] Figure 9 This is the feature distribution map extracted by the method of the present invention after dimensionality reduction using t-SNE. Detailed Implementation

[0041] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0042] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0043] Example

[0044] like Figure 1 As shown, this invention provides a method for diagnosing rotating machinery faults based on a delayed weighted visual graph and Netpro2vec. The overall diagnostic process of this invention includes: acquiring vibration signals, constructing a delayed weighted visual graph, constructing a weighted edge graph (WSGN), extracting feature vectors using Netpro2vec, and outputting diagnostic results through a classifier.

[0045] Step S1: Data acquisition and experimental setup.

[0046] This embodiment uses gearbox vibration signals from the SEU mechanical dataset. For example... Figure 8As shown, this dataset was collected from a Dynamic Drive System Simulator (DDS). The platform consists of a motor, brake, control module, and two gearboxes under test. Data was acquired via an accelerometer at a sampling frequency of 5120 Hz and a sampling length of 800. The dataset covers five health states: normal, rolling element fault, outer race fault, inner race fault, and combined fault (combination of inner and outer race faults), and was collected under two operating conditions (20 Hz / 0 V and 30 Hz / 2 V). The dataset was randomly divided into training and test sets in a 7:3 ratio. To simulate a high-noise environment, Gaussian noise was introduced into the original vibration signal. Four noise levels were considered in the test, with signal-to-noise ratios of -9, -6, -3, and 0 dB.

[0047] Step S2: Construct a delayed weighted view.

[0048] For a single vibration signal sample Convert it into a delay-weighted visual chart. This is the core step of the present invention, and its specific implementation is as follows:

[0049] Weighted Visibility View Definition: Traditional visibility views only determine whether two data points are visible. To capture detailed structural information, this invention proposes a weighted visibility view. For example... Figure 2 As shown, for sampling points and If they are visible, then define their edge weights. This is the minimum distance between the line connecting two points and all data points between them, i.e.:

[0050] ;

[0051] Figure 2 China and Israel and An example is provided for illustration. The introduction of weights allows the visualization not only to reflect whether something is visible, but also to determine the degree of visibility.

[0052] The necessity of delayed visibility mechanisms: High sampling frequency of vibration signals and small amplitude differences between adjacent points make visibility sensitive to adjacent nodes and result in poor noise resistance. For example... Figure 3 As shown, using the noise-free signal above as an example, the visual relationship (connections) between the 29th data point and other points is illustrated. When the amplitude of the 30th data point is slightly increased due to noise (the signal below), the visual relationship of the 29th point changes significantly. This demonstrates that traditional visualization charts are sensitive to noise.

[0053] Delay Visibility Definition: To improve noise immunity, this invention proposes a delay visibility diagram. Its visibility relationship is defined as follows: when all intermediate points... When all conditions are met, the node and Only then is it visible. This definition ignores nodes. and Each adjacent Each node. For example... Figure 4 As shown, with For example, when constructing the visibility relationship between nodes 8 and 29, it ignores their two adjacent nodes (represented in blue). Nodes 8 and 29 are visible, while nodes 29 and 52 are invisible due to occlusion in between.

[0054] Noise robustness verification of delay visibility: such as Figure 5 As shown, for and Figure 3 The same time series application delay visualization Comparing the delay visualizations of the noise-free signal (left) and the noise-contaminated signal (right), it can be observed that the visibility of the 29th data point remains unchanged before and after noise contamination. This verifies that the delay visualization has good noise resistance.

[0055] Integration of Weighted Delay Visualization: Finally, by combining the weighting and delay mechanisms, the final weighted delay visualization is constructed. For example... Figure 6 As shown, when constructing the visual relationship between node 7 and node 31, their two neighboring nodes are ignored. The minimum difference between the amplitude of the line connecting these two nodes and the amplitude of the intermediate node is used as the weight. The thickness of the edges in the diagram visually reflects the magnitude of the weights. This results in a visualization that accurately captures detailed data structure information while also exhibiting good noise resistance. .

[0056] Step S3: Construct a weighted edge graph (WSGN) for feature enhancement.

[0057] To further explore deep structural features, a weighted edge graph (WSGN) is used to enhance the weighted delay visualization G.

[0058] For weighted delay visualization Construct its first-order WSGN, denoted as WSGN(1). Figure 7 As shown, its construction rule is: each edge in G is mapped to a node in WSGN(1). If two edges in G share the same endpoint, their corresponding nodes are connected in WSGN(1). The edge weight in WSGN(1) is defined as the average of the weights of two edges in G that connect to the same node. For example, Figure 7In the original graph G on the left, edges (1,6) and (6,5) both connect node 6 with relatively large weights. Therefore, in the WSGN(1) on the right, the edge between nodes (1,6) and (6,5) has a relatively large weight. Similarly, the edges (2,3) and (3,4) in the original graph have small weights, so the edge weights between the corresponding nodes in WSGN(1) are small. WSGN reflects the relationships between edges in the original graph and inherits the weight information.

[0059] Step S4: Extract graph feature vectors using Netpro2vec.

[0060] Low-dimensional feature vectors are extracted from the delayed weighted visual image G and its first-order WSGN(1) using the Netpro2vec method.

[0061] Calculate the global feature, i.e., the node distance distribution: for each node in the graph Calculate its distance distribution vector It describes the global topology of the nodes in the graph:

[0062] ;

[0063] in For nodes arrive The shortest path distance, This is the starting value of the distance interval. The interval width, This represents the total number of nodes. Indicates from node Departure, distance within range The proportion of nodes within the set. The distance distribution set of all nodes. Together, they characterize the global topological features of the graph.

[0064] Calculate the local features, i.e., the transition matrix (TM): The transition matrix describes the local connectivity information in the graph based on random walks. Consider each node. Step transition matrix Its elements Defined as a node Departure, Passing through After a random walk, the node is reached. The probability of TM. TM captures the local connectivity and proximity relationships between nodes in the graph.

[0065] Generate word sequences:

[0066] For nodes , its distance distribution vector Arranged in descending order of probability values, the sequence is obtained. ,in Represents a node The interval number with the largest distance to other nodes.

[0067] For the transition matrix Sort the nodes in descending order of probability value to obtain the index sequence of the target nodes. ,in Indicates the process The sequence number of the node most likely to be reached after the step.

[0068] The generated sequence number represents the "word" of each node.

[0069] Forming a document and learning embeddings: Integrating the words from all nodes to form a document representing the visual representation. Using... These represent weighted delay visualizations. (or WSGN) global and local documents. Then, a PV-DBOW model is used to learn the distribution patterns of words in the documents, and the weights of the hidden layers are used as low-dimensional vector representations of the visual document. and In order to comprehensively capture fault characteristics, data from... The eigenvectors of WSGN(1) are fused together, for example... .

[0070] Step S5: Model training and fault diagnosis.

[0071] The fused feature vector The input is fed into a fully connected layer, followed by a softmax layer to complete the classification task. The classification network is trained using the training set data, and its performance is evaluated using the test set.

[0072] Experimental verification and result analysis

[0073] To verify the effectiveness of the method of the present invention, the following experiments were conducted:

[0074] Diagnostic results for each fault type: As shown in Table 1, the method of this invention demonstrates superior performance in identifying most fault types. The lowest accuracy and recall were for rolling element faults (30 Hz / 2 V), at 0.963 and 0.975, respectively; the lowest specificity was for combined faults (30 Hz / 2 V), at 0.993. This indicates that the method has a high ability to identify various types of faults. Figure 9 The t-SNE plot of the feature distribution also intuitively demonstrates the good separability of features of different categories.

[0075] Table 1. Fault diagnosis results of rotating machinery based on delay-weighted visual charts

[0076]

[0077] Ablation experiments: As shown in Table 2, the effectiveness of each module was verified by comparing different benchmark models. Model A (traditional visual graph, without WSGN) had an average accuracy of 0.786; Model B (traditional visual graph, with WSGN) had an accuracy of 0.823; Model C (delay-weighted visual graph, without WSGN) improved to 0.912; while the complete method of this invention (delay-weighted visual graph, with WSGN) achieved the highest accuracy of 0.993. The results show that both delay-weighted visual graph and WSGN can effectively improve performance, and the combination of the two has the best effect.

[0078] Table 2 Comparison of model structure and diagnostic accuracy

[0079]

[0080] Comparison with existing methods: To comprehensively evaluate the performance advantages of this invention, we compared our method with five mainstream fault diagnosis models (FCTNN, HVG-GAT, WHVG-GCN, CEEMD-GCN, and ACNN-BILSTM) under different signal-to-noise ratios (-9dB, -6dB, -3dB, 0dB). Evaluation metrics included accuracy, recall, and specificity to comprehensively measure the model's diagnostic capability from different dimensions. The comparison results are shown in Table 3.

[0081] Table 3 Performance comparison with existing methods at different signal-to-noise ratios

[0082]

[0083] As shown in Table 3, in high-noise environments with low signal-to-noise ratios (-9dB, -6dB), the accuracy, recall, and specificity of the method proposed in this invention are significantly better than all the comparative methods. For example, at -9dB, the accuracy of the method proposed in this invention (0.967) is about 3.1 percentage points higher than the second-best FCTNN (0.936). Even under relatively ideal conditions with a signal-to-noise ratio of 0dB, the method proposed in this invention (accuracy 0.992) still maintains top-tier performance comparable to the best method (FCTNN, 0.993), while showing better performance in recall and specificity. This fully demonstrates that the feature learning strategy proposed in this invention, which combines delayed weighted visual images with Netpro2vec, can stably and accurately extract discriminative features from high-noise backgrounds, significantly improving the accuracy and robustness of fault diagnosis models, and is particularly suitable for complex noise environments in industrial settings.

[0084] In summary, the fault diagnosis method based on delay-weighted visual graphs and Netpro2vec, as detailed in this invention, effectively solves the problems pointed out in the background art through innovative graph construction methods and feature learning strategies, and achieves high-precision, noise-resistant fault diagnosis of rotating machinery.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for diagnosing rotating machinery faults based on delay-weighted visual graphs and Netpro2vec, characterized in that, Includes the following steps: S1. Collect vibration signals from rotating machinery; S2. Convert the vibration signal into a delay-weighted visual graph. ,in, For a set of nodes, For the set of connected edges, For the set of weights of the connected edges; the construction rule of the delayed weighted visualization is: for any two data points in the vibration signal and If all data points All have Then at node and Establish an edge between them and assign a weight to the edge. ,in, It is a positive integer; S3. Based on the aforementioned delay-weighted visualization The image was extracted using the Netpro2vec method. The low-dimensional feature vector; S4. Input the low-dimensional feature vector into the trained classifier and output the fault diagnosis result of the rotating machinery.

2. The rotating machinery fault diagnosis method based on delay-weighted visualization and Netpro2vec according to claim 1, characterized in that, In step S1, the health status corresponding to the vibration signal includes at least two of the following: normal status, rolling element fault, outer ring fault, inner ring fault, and combined fault.

3. The rotating machinery fault diagnosis method based on delay-weighted visualization and Netpro2vec according to claim 1, characterized in that, In step S2, the The value range is from 1 to 5.

4. The rotating machinery fault diagnosis method based on delay-weighted visualization and Netpro2vec according to claim 1, characterized in that, After step S2 and before step S3, the method further includes: based on the delay-weighted visual diagram. Construct its first-order weighted edge graph WSGN(1); in step S3, extract the low-dimensional feature vectors of the delayed weighted visual graph G and the first-order weighted edge graph WSGN(1) and fuse them.

5. The rotating machinery fault diagnosis method based on delay-weighted visualization and Netpro2vec according to claim 4, characterized in that, The construction rule of the first-order weighted edge graph WSGN(1) is as follows: The original graph... Each edge in Mapped to a node in WSGN(1); if the original graph Two connecting edges in and If two nodes share a common endpoint, an edge is established between them in WSGN(1), and the weight of this edge is equal to the weight of the original graph. Middle of the border and The average of the weights.

6. The rotating machinery fault diagnosis method based on delay-weighted visualization and Netpro2vec according to claim 1, characterized in that, Step S3 specifically includes: S31. Calculation diagram Each node Distance distribution vector ,in, This is the starting value of the distance interval. The interval width, Indicates from node Starting point, shortest path distance in interval The proportion of nodes within a given area to the total number of nodes in the graph; S32. Calculation diagram Transition matrix of each node Its elements Indicates from node Departure process After a random walk, the node is reached. The probability of; S33. For each node, determine its distance distribution vector. and transition matrix row vector Sort by probability value in descending order to obtain the corresponding distance distribution word sequence. and transition probability word sequence ; S34. Integrate the word sequences of all nodes to form a representative graph. Documents ; S35. Learn the document using a document embedding model. , to obtain the diagram The low-dimensional feature vectors.

7. The rotating machinery fault diagnosis method based on delay-weighted visualization and Netpro2vec according to claim 6, characterized in that, The document formed in step S34 This includes a global document consisting of word sequences whose distances are distributed across all nodes. And a local document consisting of word sequences representing the transition probabilities of all nodes. .

8. The method for diagnosing rotating machinery faults based on delay-weighted visualization and Netpro2vec according to claim 7, characterized in that, In step S35, the global document is learned respectively. and the local document The corresponding global feature vector is obtained. and local feature vectors and will and The features are then spliced ​​and merged to form the final low-dimensional feature vector.

9. A method for diagnosing rotating machinery faults based on delay-weighted visual graphs and Netpro2vec according to claim 6, characterized in that, In step S35, the document embedding model is a paragraph vector distributed bag-of-words model.

10. The method for diagnosing rotating machinery faults based on delay-weighted visualization and Netpro2vec according to claim 1, characterized in that, In step S4, the classifier is a neural network classifier containing a fully connected layer and a Softmax layer.

Images