Bearing fault knowledge graph construction method driven by multi-source vibration information

Abstract

The invention discloses a bearing fault knowledge graph construction method driven by multi-source vibration information, and relates to the technical field of bearing fault diagnosis methods. The method comprises the following steps that bearing fault data are preprocessed, the preprocessed data are standardized, and the length of the data is limited according to documents; a data set needs to be divided for construction of the knowledge graph; f-value scores of 18 features of each divided data set are calculated, importance analysis is carried out on the features, and the first ten important features are reserved; and constructing a bearing fault knowledge graph based on the first ten important features. According to the method, information such as the sampling frequency, the bearing position, the fault diameter and the fault type is organized into a hierarchical structure, and the availability of the knowledge graph is enhanced through interactive visualization.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault diagnosis methods, and in particular to a method for constructing a bearing fault knowledge graph driven by multi-source vibration information. Background Art

[0002] With the rapid development of Industry 4.0 and intelligent manufacturing, fault diagnosis and health management of mechanical equipment have become critical to ensuring the reliability of production systems. Bearings, as core components of rotating machinery, have a crucial role in determining the operational stability and service life of the entire equipment. Bearing fault diagnosis methods based on vibration signal analysis, with their real-time and non-invasive nature, have become a research hotspot in the industrial monitoring field.

[0003] Traditional bearing fault diagnosis methods primarily rely on expert experience and signal processing techniques, including time-domain feature extraction, spectrum analysis, and joint time-frequency analysis. These methods face three major challenges in practical application: feature engineering relies heavily on domain expertise, making automation difficult; the effective fusion of heterogeneous data from multiple sources creates information silos; and the lack of transparency and interpretability in the diagnostic process limits their adoption in critical applications. While recent advances in deep learning have enabled some progress in neural network-based fault diagnosis methods, they still face challenges such as strong data dependency and poor model interpretability.

[0004] Traditional fault diagnosis methods, such as support vector machines (SVMs) and multilayer perceptrons (MLPs), are limited by their shallow feature extraction capabilities and are no longer able to meet the high-precision diagnostic requirements of today's industrial development. Convolutional neural networks (CNNs), with their powerful feature extraction capabilities, have gradually become a research hotspot in the field of fault diagnosis.

[0005] To address the inaccurate feature localization issues inherent in traditional multi-scale convolutional neural network models, Zhang Li et al. converted the one-dimensional vibration signal of a rolling bearing into a two-dimensional grayscale image as input to enrich the feature information. Chen Qi et al. proposed a multi-channel fusion multi-scale dynamic adaptive residual network (MC-MSDARL). This network dynamically adjusts the weights of convolution kernels of different scales to adaptively extract local and global feature information from single-channel data. This allows for weighted fusion of multi-scale features from multiple channels, resulting in a more comprehensive data representation.

[0006] While these methods enrich feature information, they rely heavily on large amounts of labeled data. In real-world industrial environments, acquiring sufficient fault samples, especially those for rare faults, is extremely difficult and costly. Furthermore, existing methods lack effective integration of domain knowledge and are unable to fully leverage expert experience and existing fault diagnosis knowledge.

[0007] As a structured knowledge representation method, knowledge graphs can effectively organize and leverage domain expert knowledge, providing new insights into solving these problems. Knowledge graphs can explicitly represent the concepts, attributes, and relationships related to bearing faults, forming a complete knowledge network. By combining deep learning with knowledge graphs, knowledge-driven intelligent fault diagnosis can be achieved, maintaining the high accuracy of deep learning models while introducing knowledge-based interpretability mechanisms.

[0008] As a semantic network structure capable of representing entities and their complex relationships, knowledge graphs have demonstrated powerful knowledge representation and reasoning capabilities in fields such as healthcare and finance. They can structuredly store dispersed domain knowledge in the form of entity-relationship-entity triples, providing a new paradigm for modeling complex systems. Key components of industrial equipment have diverse failure modes, complex causes, and are difficult to troubleshoot. With their powerful knowledge network construction and associative display capabilities, knowledge graphs can provide a new way to acquire, store, organize, and manage fault knowledge. They also support fault knowledge application and reasoning models that are more in line with human cognitive habits, significantly improving the efficiency and accuracy of fault diagnosis.

[0009] Currently, research on fault knowledge graphs has been conducted in various fields. Jiang Da constructed a high-speed rail equipment fault knowledge graph using text mining techniques; Xu et al. combined BERT with a BiLSTM-CRF framework to extract engine fault entities; Zhao et al. proposed a top-down and bottom-up approach to construct aerospace software knowledge graph; and Hu Jie et al. optimized the automobile fault knowledge graph by introducing text pre-classification and entity reorganization techniques based on the traditional construction process.

[0010] Existing research focuses on extracting knowledge from text and tabular data, ignoring the valuable fault information contained in the measured data generated by system operation. In addition, although bearing fault knowledge comes from a wide range of sources, its forms are diverse and limited, making it difficult to apply conventional methods of pure data mining. In response to multi-source heterogeneous bearing fault information, especially when it contains time series data, Tang et al. [7] proposed a knowledge graph construction method of "ontology-entity bidirectional constraint": the initial ontology is constructed from top to bottom based on expert knowledge, while the ontology structure is optimized from bottom to top through fault data mining entities, and the knowledge fusion of heterogeneous information is achieved through the bidirectional constraint of the two. Summary of the Invention

[0011] The technical problem to be solved by the present invention is how to provide a bearing fault knowledge graph construction method driven by multi-source vibration information that can effectively enhance the usability of the knowledge graph.

[0012] To solve the above technical problems, the technical solution adopted by the present invention is: a method for constructing a bearing fault knowledge graph driven by multi-source vibration information, comprising the following steps:

[0013] Preprocess the bearing fault data and standardize the preprocessed data. The data should be within the document length limit.

[0014] Building a knowledge graph requires partitioning of the dataset;

[0015] Calculate the F-value scores of the 18 features of each divided data set, perform feature importance analysis, and retain the top ten important features;

[0016] Based on the first ten important features, a bearing fault knowledge graph is constructed.

[0017] A further technical solution is that preprocessing the bearing fault data includes the following steps:

[0018] Data is hierarchically classified by fault diameter, fault type, and motor load. Fault types include inner ring fault, steel ball fault, and outer ring fault. For outer ring faults, the fault locations are further subdivided into center position, orthogonal position, and opposite position. By parsing data naming rules, a complete data mapping relationship is established.

[0019] A further technical solution is that the data standardization process includes the following steps:

[0020] All signals are uniformly cut into set sampling points, and the vibration signal data is normalized using MinMaxScaler to map the data range to the [0,1] interval.

[0021] A further technical solution is that the construction of the bearing fault knowledge graph includes:

[0022] The bearing fault knowledge graph (BF-KG) is defined as the set G = {E, A, V, R, T}, where:

[0023] E (entity set): includes bearing fault entity and vibration signal entity E Signal , that is, E=E Fault ∪E Signal ;

[0024] A (attribute set): describes the set of entity attributes;

[0025] V (attribute value set): the specific value of the attribute, such as text value or numeric value;

[0026] R (relationship set): relationship type between entities;

[0027] T (triplet set): consists of attribute triples TA and relation triples TR, that is, T = TA ∪ TR;

[0028] TA=T Fault A∪T SignalA

[0029] T Fault A=E Fault ×A Fault ×V Fault (attribute triplet of bearing fault entity)

[0030] T Signal A=E Signal ×A Signal ×V Signal (Attribute triplet of vibration signal entity)

[0031] TR=E Fault ×R×E Signal (Relationship triplet between bearing fault entity and vibration signal entity).

[0032] A further technical solution is that the bearing fault knowledge graph construction framework adopts a three-stage design approach to achieve a complete conversion from raw vibration data to a structured knowledge graph with a multi-layer fault node hierarchy:

[0033] Zero-level node: set to CWRU, serving as the central node of the knowledge graph;

[0034] Level 1 node: sampling frequency, including 12k and 48k;

[0035] Secondary node: bearing position, including Drive End and Fan End;

[0036] Level 3 node: fault diameter, including 4 categories from 0.007 inches to 0.028 inches;

[0037] Level 4 node: fault type, including inner ring, outer ring and steel ball, a total of 3 categories;

[0038] Level 5 node: fault location, including three types: central location, orthogonal location, and opposite location;

[0039] Entity node: specific fault instance, a total of 60 entities;

[0040] Feature nodes: There are two types: time domain features and frequency domain features, with a total of 120 nodes.

[0041] The beneficial effect of adopting the above technical solution is that: the method proposes a hierarchical knowledge graph construction method, organizes information such as sampling frequency, bearing position, fault diameter, fault type, etc. into a hierarchical structure, and enhances the usability of the knowledge graph through interactive visualization. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0043] Figure 1 is a flow chart of the method according to an embodiment of the present invention;

[0044] Figure 2 This is a fault type characteristic distribution diagram in an embodiment of the present invention;

[0045] Figure 3 This is a characteristic distribution diagram of fault diameters in an embodiment of the present invention;

[0046] Figure 4 This is a characteristic distribution diagram of bearing positions in an embodiment of the present invention;

[0047] Figure 5 is a fault type feature importance analysis diagram according to an embodiment of the present invention;

[0048] Figure 6 is a diagram analyzing the importance of fault diameter characteristics according to an embodiment of the present invention;

[0049] Figure 7 is a diagram analyzing the importance of bearing position features in an embodiment of the present invention;

[0050] Figure 8 is a visualization diagram of fault type characteristics in an embodiment of the present invention;

[0051] Figure 9 is a visualization diagram of the fault diameter characteristics in an embodiment of the present invention;

[0052] Figure 10 is a visualization diagram of bearing position characteristics in an embodiment of the present invention;

[0053] Figure 11 This is a partial knowledge graph in an embodiment of the present invention;

[0054] Figure 12 This is outer ring failure example 1 in the embodiment of the present invention. DETAILED DESCRIPTION

[0055] The following is a clear and complete description of the technical solutions in the embodiments of the present invention, in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts are within the scope of protection of the present invention.

[0056] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0057] like Figure 1 As shown, an embodiment of the present invention discloses a method for constructing a bearing fault knowledge graph driven by multi-source vibration information, comprising the following steps:

[0058] Preprocess the bearing fault data and standardize the preprocessed data. The data should be within the document length limit.

[0059] Building a knowledge graph requires partitioning of the dataset;

[0060] Calculate the F-value scores of the 18 features of each divided data set, perform feature importance analysis, and retain the top ten important features;

[0061] Based on the first ten important features, a bearing fault knowledge graph is constructed.

[0062] The following describes the above steps in combination with specific content:

[0063] 1) Bearing failure data

[0064] The CWRU bearing dataset contains bearing fault data under various operating conditions, covering different combinations of fault types (inner race faults, outer race faults, and rolling element faults), fault diameters (0.007 inch, 0.014 inch, 0.021 inch, and 0.028 inch), and motor loads (0, 1, 2, and 3). Outer race faults are further categorized by fault location into central (at 6 o'clock), orthogonal (at 3 o'clock), and reverse (at 12 o'clock).

[0065] Each working condition in the data set corresponds to a different motor speed: load 0 corresponds to 1797 rpm, load 1 corresponds to 1772 rpm, load 2 corresponds to 1750 rpm, and load 3 corresponds to 1730 rpm.

[0066] 2) Data collection and preprocessing

[0067] Data is hierarchically categorized by fault diameter, fault type (inner race, steel ball, outer race), and motor load. Outer race faults are particularly categorized by fault location (center, orthogonal, and reverse).

[0068] The data files use specific naming conventions. For example, "IR007_0" indicates an inner race (IR) fault with a diameter of 0.007 inches and a motor load of 0. "OR014@6_2" indicates an outer race (OR) fault with a diameter of 0.014 inches, a fault location at 6 o'clock, and a motor load of 2. By parsing these naming conventions, a complete data mapping relationship can be established.

[0069] There is some irregularity in the variable naming in the CWRU dataset. The regular expression pattern = re.compile(r'X\d+_DE_time') is used to match the Drive End time series variables.

[0070] 3) Data standardization

[0071] 12k Drive End data was uniformly truncated into 119,808 sampling points to ensure data dimensionality consistency, facilitating subsequent feature extraction and model training. The vibration signal data was normalized using the MinMaxScaler function, mapping the data range to the [0, 1] interval. This normalization process eliminated dimensional differences in vibration signal amplitudes under different operating conditions, effectively improving the accuracy and stability of feature extraction.

[0072] Appropriate data structures are used for different datasets. The 12k Drive End data uses a DataFrame structure, while the 12k Fan End and 48k Drive End data use a dictionary structure to adapt to the characteristics and processing requirements of different datasets.

[0073] Through these preprocessing steps, the raw vibration signal data is converted into a structured and standardized data format, providing high-quality input data for the subsequent feature extraction algorithm and ensuring the reliability and consistency of the feature extraction results.

[0074] Feature Engineering and Enhancement

[0075] To further improve the discriminative power of features, the basic features were engineered to create additional composite features:

[0076] Kurtosis-to-mean ratio: The ratio of kurtosis to mean, calculated as:

[0077]

[0078] where ε is a small constant that prevents division by zero.

[0079] Frequency Domain Peak-to-Root Mean Square Ratio: The ratio of the frequency domain peak value to the root mean square value. The calculation formula is:

[0080]

[0081] Pulse margin ratio: The ratio of pulse factor to margin factor, calculated as:

[0082]

[0083] Kurtosis-peak product: The product of kurtosis and crest factor, calculated as:

[0084] K F =K·Crest Factor

[0085] Standard Deviation RMS: The ratio of the standard deviation to the RMS, calculated as:

[0086]

[0087] Frequency domain energy ratio: the ratio of frequency domain energy to time domain energy, calculated as:

[0088]

[0089] Through feature engineering, a feature set containing 18 features was constructed. These features describe the characteristics of vibration signals from different perspectives and provide a rich information basis for fault diagnosis.

[0090] Feature analysis and screening:

[0091] In order to determine which features are most valuable for fault diagnosis, the extracted features are analyzed and screened in detail.

[0092] Feature distribution analysis:

[0093] like Figure 2 The figure shows the distribution of each feature in the fault type classification task. From the figure, we can observe that:

[0094] The skewness feature is symmetrically distributed and close to zero under normal conditions, but is dispersed during faults, indicating that faults increase signal asymmetry. The root mean square (RMS) feature clearly distinguishes between normal and faulty conditions, with values ranging from 0.44 to 0.56. The RMS value is generally higher during faults, reflecting an increase in signal energy caused by the fault. The frequency domain energy feature exhibits a bimodal distribution in faulty samples, with peak values of approximately 3.6×109 and 4.2×109, which may be related to the energy levels of different fault types. The crest factor is lower in faulty samples than in normal samples, with a median of approximately 2.0, indicating a systematic change in the peak characteristics of the fault signal. The pulse factor has a median of approximately 0.48, showing a relatively concentrated distribution, with faulty samples tending towards higher values.

[0095] Figure 3 The characteristic distributions under different fault diameter conditions are presented. The analysis results show that the skewness increases with increasing fault diameter, and the median shifts positively. The frequency domain energy is distributed hierarchically, with energy values increasing with larger fault diameters. The root mean square (RMS) value between 0.44 and 0.56 shows the most obvious stratification effect, indicating strong fault size measurement capabilities. The crest factor and impulse factor vary little across different fault diameters, making them more suitable for fault type identification rather than fault severity assessment.

[0096] Figure 4The distribution of bearing position-related features is shown. It is worth noting that the kurtosis-to-mean ratio, kurtosis-to-peak product, and kurtosis show significant differences between the drive-end and fan-end bearings (in the range of 0-90), with the drive-end values generally higher. The frequency domain energy also shows good discrimination ability, with the drive-end concentrated at 3.6×10 9 -4.2×10 9 The fan end is mainly in the range of 3.0×10 9 -3.6×10 9 interval; in the fault state, skewness and kurtosis reflect the enhanced nonlinearity of the signal, the frequency domain energy and root mean square related vibration energy increase, and the crest factor and impulse factor are related to the impact characteristics; the difference between the two position characteristics is due to the different workloads and structures. The greater force on the drive end leads to differences in the statistical characteristics of the vibration signal, which is especially obvious in the kurtosis-related characteristics, providing a reliable basis for automatic identification of bearing positions.

[0097] In summary, feature distribution analysis not only validates the effectiveness of the extracted features but also provides intuitive insights into signal characteristics under different fault conditions, laying an empirical foundation for subsequent feature selection and classifier design. The diversity and discriminability of feature distributions are key factors in achieving high-precision fault diagnosis and directly demonstrate the effectiveness of this study's feature engineering approach.

[0098] Feature Importance Analysis:

[0099] like Figure 5 As shown in the figure, in fault type identification, skewness (0.112), standard mean square root deviation (0.096), and kurtosis peak product (0.094) are the three most discriminative features, contributing about 30.2% of the importance. This shows that time domain statistical features and our proposed combined features play a key role in fault type identification.

[0100] like Figure 6 As shown in the figure, skewness (0.104), frequency domain peak-to-root mean square ratio (0.097), and standard deviation root mean square to kurtosis (both 0.085) are the most discriminative features in fault diameter classification, contributing approximately 28.7% of the importance. This finding emphasizes the importance of combining frequency and time domain indicators, especially for identifying different fault diameters.

[0101] like Figure 7 As shown in the figure, in the bearing position classification, skewness (0.124), kurtosis-to-mean ratio (0.107), and kurtosis-to-peak product (0.106) are the three most discriminative features, contributing a total of approximately 33.7% importance. This shows that our proposed combined feature can effectively capture bearing position information and provide reliable feature support for bearing position identification.

[0102] Feature Selection:

[0103] To improve the model's computational efficiency, reduce the risk of overfitting, and enhance generalization, a systematic feature selection method was used to screen the most relevant feature subset. In bearing fault diagnosis, effective feature selection not only reduces computational complexity but also improves classification accuracy and model robustness.

[0104] ANOVA F-value-based filtering method: This method evaluates the discriminative power of features by calculating the F statistic between each feature and the target variable. This method is computationally efficient and can quickly screen out features that are significantly correlated with the target variable.

[0105] Embedded method based on random forest: Utilizes the feature importance scoring mechanism of the random forest algorithm to obtain feature contributions from the model training process. This method considers the interaction between features and can capture nonlinear relationships.

[0106] The feature selection process is performed separately for three different fault diagnosis tasks (fault type classification, fault diameter classification, and bearing position classification) to obtain the optimal feature subset for each task.

[0107] Based on feature importance analysis and the F-statistic score of the SelectKBest algorithm, the optimal feature subsets were determined for different classification tasks. Table 1 shows the features selected for each task and their corresponding F-scores.

[0108] Table 1 - Optimal feature subsets and their F-value scores for different classification tasks

[0109]

[0110] Feature visualization:

[0111] In order to more intuitively understand the relationship between features and faults, a visual analysis of the features is performed:

[0112] t-SNE dimensionality reduction visualization Figures 8-10 ,The results show that based on the extracted features, different types of bearing faults ,form clusters in the feature space, which proves the effectiveness of the ,extracted features.

[0113] Through the above feature extraction, analysis, and screening process, a high-quality feature set was constructed, providing a solid foundation for bearing fault diagnosis. These features include not only traditional time-domain and frequency-domain statistical features, but also our proposed engineering features, which can more comprehensively describe the characteristics of bearing faults.

[0114] Bearing Fault Knowledge Graph (BF-KG) Construction Method

[0115] 1) Knowledge Graph Definition

[0116] The bearing fault knowledge graph (BF-KG) is defined as the set G = {E, A, V, R, T}, where:

[0117] E (entity set): includes bearing fault entity and vibration signal entity E Signal , that is, E=E Fault ∪E Signal

[0118] A (attribute set): describes the set of entity attributes.

[0119] V (attribute value set): The specific value of the attribute, such as text value or numeric value.

[0120] R (relationship set): the relationship type between entities.

[0121] T (triplet set): consists of attribute triples TA and relation triples TR, that is, T = TA∪TR.

[0122] TA=T Fault A∪T Signal A

[0123] T Fault A=E Fault ×A Fault ×V Fault (attribute triplet of bearing fault entity)

[0124] T Signal A=E Signal ×A Signal ×V Signal (Attribute triplet of vibration signal entity)

[0125] TR=E Fault ×R×E Signal (Relationship triplet between bearing fault entity and vibration signal entity)

[0126] Example:

[0127] Attribute triplet: (inner ring fault, fault frequency, 162Hz)∈T Fault A indicates that the attribute "Fault Frequency" of the bearing fault entity "Inner Race Fault" has a value of "162 Hz".

[0128] Relation triple: (Signal 001 ,Signal Of , inner ring fault)∈TR represents the vibration signal entity "Signal 001 "With the bearing fault entity "Inner Ring Fault" through the relationship "Signal Of "Relationship.

[0129] 2) Knowledge graph construction framework

[0130] The Bearing Fault Knowledge Graph (BF-KG) construction framework proposed in this paper adopts a three-stage design approach to achieve a complete conversion from raw vibration data to a structured knowledge graph. It has a multi-layer fault node hierarchy:

[0131] Zero-level node: set to CWRU, serving as the central node of the knowledge graph;

[0132] Level 1 node: sampling frequency, including 12k and 48k;

[0133] Secondary node: bearing position, including Drive End and Fan End;

[0134] Level 3 node: fault diameter, including 4 categories from 0.007 inches to 0.028 inches;

[0135] Level 4 node: fault type, including inner ring, outer ring and steel ball, a total of 3 categories;

[0136] Level 5 node: fault location (only for outer ring faults), including three types: center location, orthogonal location, and opposite location;

[0137] Entity node: specific fault instance, a total of 60 entities;

[0138] Feature nodes: There are two types: time domain features and frequency domain features, with a total of 120 nodes.

[0139] The node attributes entered into the knowledge graph include node label, node type, node description and feature value.

[0140] 3) Entity construction

[0141] A Python program was used to construct a knowledge graph by combining entity nodes and relationship nodes. This study used the NetworkX library to create a directed graph and Pyvis for visualization. The graph detailed the entity features and clearly presented the relationships between nodes.

[0142] like Figure 11 As shown in the figure, the core node of the knowledge graph is "CWRU", which is first classified by sampling frequency ("48k" and "12k"), and then hierarchically divided by bearing position ("Drive End" and "Fan End"), fault diameter, and fault type (such as "outer ring", etc.).

[0143] like Figure 12For outer ring faults, the fault location information is further identified. Feature information is divided into two categories: time domain features (such as "mean", "standard deviation", "skewness", "kurtosis", etc.) and frequency domain features ("FFT peak", "FFT energy"). Each entity node is connected to its corresponding feature node.

Claims

1. A method for constructing a bearing fault knowledge graph driven by multi-source vibration information, characterized in that The steps include: Preprocess the bearing fault data and standardize the preprocessed data. The data should be within the document length limit. Building a knowledge graph requires partitioning of the dataset; Calculate the F-value scores of the 18 features of each divided data set, perform feature importance analysis, and retain the top ten important features; Based on the first ten important features, a bearing fault knowledge graph is constructed.

2. The method for constructing a bearing fault knowledge graph driven by multi-source vibration information according to claim 1, characterized in that: The preprocessing of bearing fault data includes the following steps: Data is hierarchically classified by fault diameter, fault type, and motor load. Fault types include inner ring fault, steel ball fault, and outer ring fault. For outer ring faults, the fault locations are further subdivided into center position, orthogonal position, and opposite position. By parsing data naming rules, a complete data mapping relationship is established.

3. The method for constructing a bearing fault knowledge graph driven by multi-source vibration information according to claim 1, characterized in that: The data standardization process includes the following steps: All signals are uniformly cut into set sampling points, and the vibration signal data is normalized using MinMaxScaler to map the data range to the [0,1] interval.

4. The method for constructing a bearing fault knowledge graph driven by multi-source vibration information according to claim 1, characterized in that: The construction of bearing fault knowledge graph includes: The bearing fault knowledge graph (BF-KG) is defined as the set G = {E, A, V, R, T}, where: E (entity set): includes bearing fault entity and vibration signal entity E Signal , that is, E=E Fault ∪E Signal ; A (attribute set): describes the set of entity attributes; V (attribute value set): the specific value of the attribute, such as text value or numeric value; R (relationship set): relationship type between entities; T (triplet set): consists of attribute triples TA and relation triples TR, that is, T = TA ∪ TR; TA=T Fault A∪T Signal A T Fault A=E Fault ×A Fault ×V Fault (attribute triplet of bearing fault entity) T Signal A=E Signal ×A Signal ×V Signal (Attribute triplet of vibration signal entity) TR=E Fault ×R×E Signal (Relationship triplet between bearing fault entity and vibration signal entity).

5. The method for constructing a bearing fault knowledge graph driven by multi-source vibration information according to claim 1, characterized in that: The bearing fault knowledge graph construction framework adopts a three-stage design approach to achieve a complete conversion from raw vibration data to a structured knowledge graph with a multi-layer fault node hierarchy: Zero-level node: set to CWRU, serving as the central node of the knowledge graph; Level 1 node: sampling frequency, including 12k and 48k; Secondary node: bearing position, including Drive End and Fan End; Level 3 node: fault diameter, including 4 categories from 0.007 inches to 0.028 inches; Level 4 node: fault type, including inner ring, outer ring and steel ball, a total of 3 categories; Level 5 node: fault location, including three types: central location, orthogonal location, and opposite location; Entity node: specific fault instance, a total of 60 entities; Feature nodes: There are two types: time domain features and frequency domain features, with a total of 120 nodes.

Images