Bearing life prediction method based on knowledge graph under cold start

Abstract

The invention discloses a bearing life prediction method based on a knowledge graph under cold start. The method comprises the following steps: S1, building a vector-based equipment life knowledge graph according to equipment text data and equipment vibration information of a public data set in the field of bearing life; s2, collecting bearing operation vibration data of the current environment; s3, carrying out variational mode decomposition on the newly collected bearing operation vibration data; s4, according to the feature information of the new nodes, using a TransH life prediction method based on cosine similarity to select life type nodes most possibly associated with new life vibration signal nodes, and combining with a deep learning model to assist in prediction; and S5, adding rule knowledge such as the new life vibration feature nodes, the similarity relationship between the new nodes and the basic map nodes, the life prediction result and the like into the map.

Description

Technical Field

[0001] The present invention relates to the field of industrial bearing life prediction and knowledge graphs, and in particular to a bearing life prediction method based on a knowledge graph under cold start. Background Art

[0002] With the rapid development of modern mechanical equipment, the complexity and running accuracy of bearings have increased significantly, and their life and reliability have a great impact on the safe operation of the entire mechanical equipment. Therefore, bearing health status monitoring and life prediction have become a hot topic.

[0003] Bearing life prediction methods can be divided into three categories: physical model-based methods, statistical methods, and data-driven methods. Physical model-based methods are usually limited to specific equipment types and have poor portability; statistical methods require fewer parameter types and larger amounts of data to build a "black box" model, producing opaque predictions; data-driven prediction methods have a wider range of uses in life, and data-driven methods combined with artificial intelligence have become the most widely used life prediction methods. Article [Yupeng Wei1CA1; Dazhong Wu2. Conditional variational transformer for bearing remaining useful life prediction [J]. Advanced Engineering Informatics, 2024, Vol. 59: 102247.] In order to solve the problem that the self-attention mechanism in Transformer does not work to select highly correlated features, a new conditional variational Transformer architecture is proposed. In order to solve the problem that life prediction methods are overly dependent on prior knowledge and lack adaptability, the article [Wen Jinghui, Wu Rongsen, Li Shuaiyong, et al. Bearing Remaining Useful Life Prediction Method Based on DRSN and Optimized BiLSTM [J]. Computer Integrated Manufacturing Systems, 2024, 30(05): 1877-1888.] proposes a bearing remaining life prediction method based on deep residual shrinkage network (DRSN) and bidirectional long short-term memory network (BiLSTM) with adaptive feature extraction. In order to solve the problem of poor interpretability of deep learning models, the article [Xiaoyu Yang1; Ying Zheng1; Yong Zhang2; David Shan-Hill Wong3; Weidong Yang1. Bearing Remaining Useful Life Prediction Based on Regression Shapalet and Graph Neural Network [J]. IEEE Transactions on Instrumentation and Measurement, 2022, Vol. 71: 1-12.] adopts a graph evolution method to provide an explanation for the bearing degradation process by detecting the changes of the graph over time. In summary, the prediction method based on data-driven solves the problem that the bearing life prediction mechanism is complex and difficult to establish a mathematical model, and has received widespread attention. However, in the actual application of industrial production, many equipment do not have complete operating history data due to the lack of data acquisition, and the accuracy of the life prediction model based on data-driven is difficult to improve. Therefore, how to predict the bearing life under cold start has become a research focus.

[0004] The cold start problem has been widely studied in machine learning and recommendation systems. Due to the lack of sufficient historical data, it is difficult for the system to make accurate predictions or recommendations on user preferences or item features. To address this challenge, researchers have developed a variety of solutions, which can be divided into two categories: data-driven recommendation algorithms and method-driven recommendation algorithms. Data-driven recommendation algorithms include solutions based on demographic data, solutions based on social networks, and cross-domain solutions to alleviate the cold start problem. Among them, cross-domain recommendation uses useful information from the original domain to migrate to the target domain, effectively solving the problems of data scarcity and cold start. The article [WANG X, PENG Z, WANG S, et al. CDLFM: cross-domain recommendation for cold-start users via latent feature mapping. [J]. Knowledge & Information Systems, 2020, Vol. 62 (5): 1723-1750.] proposes a cross-domain latent feature mapping model based on learning the feature mapping function of neighboring users. In order to further explore the natural relationship between the two domains, the article [Zhao, Guoshuaia; Zhang, Xiaolonga; Tang, Haoa; Shen, Jialieb; Qian, Xueminga. Domain-Oriented Knowledge Transfer for Cross-Domain Recommendation[J]. IEEE Transactions on Multimedia, 2024,: 1-12.] proposed a domain-oriented knowledge transfer method for cross-domain recommendation. The above methods extract user potential information through the association of auxiliary information and cross-domain data. However, for the cold start problem when the device is just put into use, cross-domain recommendation requires the consistency of data in the source domain and the target domain. If there is a deviation in the data, the cross-domain recommendation system cannot guarantee the accuracy of the prediction. Method-driven cold start problem solution recommendation algorithms include meta-learning-based recommendation algorithms, context information and conversation strategy-based solution algorithms, graph random walk-based solution algorithms, adversarial mechanism-based solution algorithms, and knowledge graph embedding task-based solution algorithms. Among them, the knowledge graph can integrate data from different sources and types by building a knowledge base of structured device information to form a unified knowledge representation framework, thereby discovering deeper connections between nodes and effectively solving the cold start problem of new projects.The article [Yuntao Du1; Xinjun Zhu1; Lu Chen1; Ziquan Fang1; YunjunGao1.MetaKG: Meta-Learning on Knowledge Graph for Cold-Start Recommendation[J].IEEE Transactions on Knowledge and Data Engineering,2023,Vol.35(10):9850-9863.] proposes a new collaborative-aware cold-start recommendation framework that adaptively selects information tasks for learning. In order to preserve the key co-occurrence patterns between node features, the article [Hui Pan11; Senlin Luo12; Xinshuai Li13; LiminPan1CA4; Zhouting Wu15.Meta-learning on dynamic node clustering knowledgegraph for cold-start recommendation[J].Neurocomputing,2024,Vol.602:128192.] proposes a recommendation method based on dynamic node clustering knowledge graph to promote connections between users based on attribute relationships and achieve information balance across paths.

[0005] From the above research, it can be seen that the current data-driven bearing life prediction method requires a large number of training samples, and it is difficult to complete the bearing life prediction in the absence of historical operating data. The traditional data-driven recommendation algorithm cannot solve the cold start problem of new projects in an industrial context. Summary of the invention

[0006] The purpose of the present invention is to provide a bearing life prediction method based on knowledge graph under cold start, which includes two modules: vector-based equipment life knowledge graph construction and cosine similarity-based TransH life prediction. Among them, the vector-based equipment life knowledge graph construction realizes the life extraction of equipment text data and the feature extraction of equipment vibration information, obtains the description of equipment life information and time-frequency domain feature information, combines the life information and time-frequency domain feature extraction to extract degradation rules, and realizes the construction of knowledge graph. The TransH life prediction based on cosine similarity obtains the similarity between the new node and the graph node through similarity analysis, establishes the similar_to correspondence, and uses the TransH algorithm to project the entity onto the hyperplane defined by the relationship, and learns the low-dimensional vector representation of the entity and the relationship. By calculating the scores to link the potential associations between the predicted entities and combining them with the prediction results of the deep learning model, the most likely connections are finally identified and recommended in the knowledge graph to realize the remaining life prediction of the vibration feature nodes in the new environment. New life vibration feature nodes and their associated rule knowledge are continuously added to the graph, and the basic knowledge graph is iteratively updated to realize the personalized growth of the graph, thereby realizing accurate prediction of bearing life under cold start conditions.

[0007] The technical solution adopted by the present invention is:

[0008] A bearing life prediction method based on knowledge graph under cold start includes the following steps:

[0009] S1: Build a vector-based equipment life knowledge graph based on equipment text data and equipment vibration information from public data sets in the field of bearing life;

[0010] S2: Collect the bearing operation vibration data of the current environment and pre-process the data, including removing null values ​​and abnormal values, and cutting the data to make it consistent with the data set content size of the basic map, and generate a data set that meets the requirements of subsequent models;

[0011] S3: performing variational mode decomposition on the newly collected bearing operation vibration data;

[0012] S4: According to the characteristic information of the new node, the TransH life prediction method based on cosine similarity is used to select the life type node that is most likely to be associated with the new life vibration signal node, and the deep learning model is combined to assist the prediction;

[0013] S5: Add new life vibration feature nodes, similarity relationships between new nodes and basic graph nodes, life prediction results and other rule knowledge into the graph.

[0014] Furthermore, in S1, building a vector-based equipment life knowledge graph based on equipment text data and equipment vibration information of a public data set in the field of bearing life includes the following steps:

[0015] 1-1) Obtain a public data set in the field of bearing life, and define the knowledge graph rule layer and the data layer to represent the knowledge graph life rule graph and the actual vibration feature data respectively. The rule layer contains the life rule knowledge graph, and the data layer stores the actual operation feature data, which corresponds one-to-one with the nodes in the rule layer;

[0016] 1-2) Extract the equipment text data of the data set to obtain the bearing life information and data source information; extract the equipment vibration information to obtain the data vibration characteristics, and define the labels of the life vibration characteristic node, life type node and data source node corresponding to the node;

[0017] 1-3) According to the characteristics of its own data, select the appropriate Δd as the life value difference, and the relationship between adjacent life type nodes is X n =X n+1 +Δd, where X n ,X n+1 are the values ​​corresponding to adjacent life type nodes, and there is a difference in life values ​​of Δd;

[0018] 1-4) Establishment <X n , loss_Δd_rul,X n+1 > triple relationship, and link adjacent life type nodes;

[0019] 1-5) define a life vibration characteristic node corresponding to each vibration signal, and customize the corresponding number to ensure that the number is unique and corresponds to a life vibration characteristic node;

[0020] 1-6) If there is X rul , represents the actual life corresponding to a certain life vibration characteristic node, if X rul =X n , then establish the triple relationship of <life vibration feature node, belong_to, life type node>, that is, <life vibration feature node, belong_to, X n >; if X n <X rul <X n+1 , then establish the triple relationship of <life vibration feature node, belong_to*value1, life type node>, <life vibration feature node, belong_to*value2, life type node>, that is, <life vibration feature node, belong_to*value1, X n>, <Life Vibration Feature Node, belong_to*value2, X n+1 >, the calculation process of value1 and value2 is as follows:

[0021]

[0022] Among them, the k value is the modulus selected according to the operating environment;

[0023] 1-7) Perform variational mode decomposition (VMD) on the equipment time series vibration data to extract features. The corresponding calculation process is as follows:

[0024]

[0025]

[0026] Among them, K is the number of decomposed modes, {v k} is the IMFs component obtained by segment decomposition of time series vibration data, {ω k} is the center frequency of the component, and δ(t) is the Dirac function. The Lagrange multiplication operator τ is introduced to transform it into an unconstrained variational problem:

[0027]

[0028] Among them, α is the second-order penalty factor, which ensures the accuracy of reconstruction of time series vibration data fragments in Gaussian noise environment. The alternating direction multiplier method (ADMM) is used to update the components and their central frequencies, search for the saddle point of the Lagrangian function, and optimize the v after iteration. k ,ω k and τ is:

[0029]

[0030] Among them, γ is the noise tolerance, and They are v i (t), g(t) and τ n+1 The Fourier transform of (t), n is the number of iterative optimizations, and the iteration stops when the following equation is satisfied:

[0031]

[0032] Where ε is the given judgment accuracy;

[0033] 1-8) Establish a triple relationship of <life vibration feature node, from, data source node>, where from indicates that the life vibration feature node comes from the data source node.

[0034] Furthermore, in S4, according to the characteristic information of the new node, the TransH life prediction method based on cosine similarity is used to select the life type node that is most likely to be associated with the new life vibration signal node, and combined with the deep learning model to assist in prediction, including the following steps:

[0035] 2-1) Use short-time Fourier transform (STFT) on the new vibration signal. The steps are as follows:

[0036] Step 1: Move the window to the beginning of the signal and perform windowing on the life vibration data. Where f(t) is the original signal, w(t) is the window function, and the Hanning window is used to reduce spectrum leakage. The corresponding formula is as follows:

[0037] g(t)=f(t)·w(t)-τ0)

[0038] Step 2: Perform Fourier transform on g(t), as shown below:

[0039]

[0040] The spectrum distribution G(ω) of the first segment of the life vibration data is obtained. ω is a frequency variable, corresponding to the frequency component in Fourier transform. Define the function S(ω,τ), which represents the spectrum result G(ω) after the original function f(t) is transformed when the center of the window function is τ. The process is:

[0041]

[0042] Step 3: Move the window to τ1 and perform the next operation. Repeat the sliding of the corresponding window to finally obtain the spectrum results of all segments;

[0043] 2-2) Calculate the power spectrum of the obtained spectrum results, and the process is as follows:

[0044] P(ω,τ)=|S(ω,τ)| 2

[0045] 2-3) Calculate the cosine similarity between the vibration data of the new life vibration feature node and the vibration data of the spectrum life vibration feature node. The calculation formula is as follows:

[0046]

[0047] Determine whether the similarity between the vibration signal of the new life vibration feature node and the vibration signal of the spectrum life vibration feature node is greater than ε. If the similarity is greater than ε, establish a triple relationship of <new life vibration feature node, similar_to, spectrum life vibration feature node>;

[0048] 2-4) Use the TransH algorithm to project the entities onto the hyperplane defined by the relations and learn the low-dimensional vector representation of the entities and relations. The specific steps are as follows:

[0049] Step 1: Project the original vector h onto the relational plane W r , the unit normal vector h obtained by its projection w The process is:

[0050] h w =w T hw

[0051] Where w is the relationship plane W r The unit normal vector of the hyperplane W r For each relation r, a hyperplane is given on which the relation d is defined. r , map the original head entity h and tail entity t to the hyperplane and name it h ⊥ ,t ⊥ , and h ⊥ , t ⊥ Satisfy the following formula:

[0052] h ⊥ +d r =t ⊥

[0053] Step 2: According to h w Obtain the mapping vector h ⊥ , the calculation process is:

[0054] h ⊥ =hh w =hw T hw

[0055] Step 3: Find the overall mapping vector. The calculation process is:

[0056]

[0057]

[0058] Among them, h′ ⊥ , h′, etc. are generated negative examples;

[0059] Step 4: Calculate the loss function and reduce the loss value during the iteration process. The corresponding loss calculation formula is:

[0060]

[0061] Among them, G represents the set of correct triples, and G′ is the set of negative examples;

[0062] 2-5) Use the Euclidean distance formula to calculate the matching degree of entity and relationship pairs, and obtain the life type node that is most likely related to the new life vibration feature node. The corresponding Euclidean distance calculation formula is:

[0063]

[0064] The negative sign is used to convert the distance into a score. The smaller the distance, the higher the score, indicating that the matching degree between the entity pair and the relationship pair is higher. i and t i Represent the embedding vectors of the head entity and the tail entity respectively;

[0065] 2-6) Use multiple deep learning models to predict bearing life, and combine the knowledge graph prediction results to comprehensively predict bearing life.

[0066] The beneficial effects of adopting the above technical solution are:

[0067] The purpose of the present invention is to provide a bearing life prediction method based on knowledge graph under cold start, which includes two modules: vector-based equipment life knowledge graph construction and cosine similarity-based TransH life prediction. Among them, the vector-based equipment life knowledge graph construction realizes the life extraction of equipment text data and the feature extraction of equipment vibration information, obtains the description of equipment life information and time-frequency domain feature information, combines the life information and time-frequency domain feature extraction to realize the construction of knowledge graph. TransH life prediction based on cosine similarity obtains the similarity between new nodes and graph nodes through similarity analysis, establishes a similar_to correspondence, and uses the TransH algorithm to project entities onto a hyperplane defined by relationships, and learns low-dimensional vector representations of entities and relationships. By calculating scores to link and predict potential associations between entities, and combining the prediction results of deep learning models, the most likely connections are finally identified and recommended in the knowledge graph, and the remaining life prediction of vibration feature nodes in the new environment is realized. The new life vibration feature nodes and their associated rule knowledge are continuously added to the graph, and the basic knowledge graph is iteratively updated to realize the personalized growth of the graph.

[0068] The life prediction method proposed in this paper is applied to the PHM2012 bearing dataset and the XJTU bearing dataset. The PHM2012 bearing dataset is used in the IEEE PHM 2012 competition and contains the vibration and temperature data of the bearings, which is used to predict the remaining service life of the bearings. The XJTU bearing dataset is provided by Xi'an Jiaotong University and contains the vibration data of the bearings under different working conditions, which is used for bearing fault diagnosis and life prediction research. Among them, the PHM2012 dataset is used to build the basic graph, and the XJTU dataset is used for experimental verification. Through experimental analysis, on the XJTU dataset, the mean absolute error and root mean square error are reduced by at least about 16 and at most about 111 compared with the traditional data-driven life prediction method, which are better than the traditional data-driven life prediction method, verifying the life prediction ability of the bearing life prediction method based on the knowledge graph under cold start, and realizing the accurate prediction of the bearing life under cold start conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 This is a framework diagram of the bearing life prediction method based on knowledge graph under cold start;

[0070] Figure 2 Build graphs for vector-based equipment life knowledge graphs;

[0071] Figure 3 Extract lifespan rule relationship graph for text;

[0072] Figure 4 is the hyperplane mapping graph of the relationship vector;

[0073] Figure 5 This is the content diagram of some equipment life knowledge graphs;

[0074] Figure 6 Comparison chart between LSTM, TCN-LSTM-A, LSTM-SA and the method in this paper;

[0075] Figure 7 This is a comparison chart of the prediction results of NSC, CS and the KG-CS method in this paper; DETAILED DESCRIPTION

[0076] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0077] Based on the background of industrial bearing life prediction, the present invention uses the knowledge graph architecture as a carrier and deep learning as an aid to propose a bearing life prediction method based on knowledge graph in a cold start environment. Figure 1 As shown, the following steps are included:

[0078] S1: Build a vector-based equipment life knowledge graph based on the equipment text data and equipment vibration information of the public data set in the field of bearing life, including the following steps:

[0079] 1-1) Obtain a public data set in the field of bearing life, and define the knowledge graph rule layer and the data layer to represent the knowledge graph life rule graph and the actual vibration feature data respectively. The rule layer contains the life rule knowledge graph, and the data layer stores the actual operation feature data, which corresponds one-to-one with the nodes in the rule layer;

[0080] 1-2) Extract the equipment text data of the data set to obtain the bearing life information and data source information; extract the equipment vibration information to obtain the data vibration characteristics, and define the labels of the life vibration characteristic node, life type node and data source node corresponding to the node;

[0081] 1-3) According to the characteristics of its own data, select the appropriate Δd as the life value difference, and the relationship between adjacent life type nodes is X n =X n+1 +Δd, where X n ,X n+1 are the values ​​corresponding to adjacent life type nodes, and there is a difference in life values ​​of Δd;

[0082] 1-4) Establishment <X n , loss_Δd_rul,X n+1 > triple relationship, and link adjacent life type nodes;

[0083] 1-5) define a life vibration characteristic node corresponding to each vibration signal, and customize the corresponding number to ensure that the number is unique and corresponds to a life vibration characteristic node;

[0084] 1-6) If there is X rul , represents the actual life corresponding to a certain life vibration characteristic node, if X rul =X n , then establish the triple relationship of <life vibration feature node, belong_to, life type node>, that is, <life vibration feature node, belong_to, X n >; if X n <X rul <X n+1 , then establish the triple relationship of <life vibration feature node, belong_to*value1, life type node>, <life vibration feature node, belong_to*value2, life type node>, that is, <life vibration feature node, belong_to*value1, X n >, <Life Vibration Feature Node, belong_to*value2, Xn+1 >, the calculation process of value1 and value2 is as follows:

[0085]

[0086] Among them, the k value is the modulus selected according to the operating environment;

[0087] 1-7) Perform variational mode decomposition (VMD) on the equipment time series vibration data to extract features. The corresponding calculation process is as follows:

[0088]

[0089]

[0090] Among them, K is the number of decomposed modes, {v k} is the IMFs component obtained by segment decomposition of time series vibration data, {ω k} is the center frequency of the component, and δ(t) is the Dirac function. The Lagrange multiplication operator τ is introduced to transform it into an unconstrained variational problem:

[0091]

[0092] Among them, α is the second-order penalty factor, which ensures the accuracy of reconstruction of time series vibration data fragments in Gaussian noise environment. The alternating direction multiplier method (ADMM) is used to update the components and their central frequencies, search for the saddle point of the Lagrangian function, and optimize the v after iteration. k ,ω k and τ is:

[0093]

[0094] Among them, γ is the noise tolerance, and They are v i (t), g(t) and τ n+1 The Fourier transform of (t), n is the number of iterative optimizations, and the iteration stops when the following equation is satisfied:

[0095]

[0096] Where ε is the given judgment accuracy;

[0097] 1-8) Establish a triple relationship of <life vibration feature node, from, data source node>, where from indicates that the life vibration feature node comes from the data source node.

[0098] S2: Collect the bearing operation vibration data of the current environment and pre-process the data, including removing null values ​​and abnormal values, and cutting the data to make it consistent with the data set content size of the basic map, and generate a data set that meets the requirements of subsequent models;

[0099] S3: Perform VMD decomposition on the newly collected bearing operation vibration data;

[0100] S4: According to the feature information of the new node, the TransH life prediction method based on cosine similarity is used to select the life type node that is most likely to be associated with the new life vibration signal node, and the deep learning model is combined to assist the prediction, including the following steps:

[0101] 2-1) Use short-time Fourier transform (STFT) on the new vibration signal. The steps are as follows:

[0102] Step 1: Move the window to the beginning of the signal and perform windowing on the life vibration data. Where f(t) is the original signal, w(t) is the window function, and the Hanning window is used to reduce spectrum leakage. The corresponding formula is as follows:

[0103] g(t)=f(t)·w(t-τ0)

[0104] Step 2: Perform Fourier transform on g(t), as shown below:

[0105]

[0106] The spectrum distribution G(ω) of the first segment of the life vibration data is obtained. ω is a frequency variable, corresponding to the frequency component in Fourier transform. Define the function S(ω,τ), which represents the spectrum result G(ω) after the original function f(t) is transformed when the center of the window function is τ. The process is:

[0107]

[0108] Step 3: Move the window to τ1 and perform the next operation. Repeat the sliding of the corresponding window to finally obtain the spectrum results of all segments;

[0109] 2-2) Calculate the power spectrum of the obtained spectrum results, and the process is as follows:

[0110] P(ω,τ)=|S(ω,τ)| 2

[0111] 2-3) Calculate the cosine similarity between the vibration data of the new life vibration feature node and the vibration data of the spectrum life vibration feature node. The calculation formula is as follows:

[0112]

[0113] Determine whether the similarity between the vibration signal of the new life vibration feature node and the vibration signal of the spectrum life vibration feature node is greater than ε. If the similarity is greater than ε, establish a triple relationship of <new life vibration feature node, similar_to, spectrum life vibration feature node>;

[0114] 2-4) Use the TransH algorithm to project the entities onto the hyperplane defined by the relations and learn the low-dimensional vector representation of the entities and relations. The specific steps are as follows:

[0115] Step 1: Project the original vector h onto the relational plane W r , the unit normal vector process obtained by projection is:

[0116] h w =w T hw

[0117] Where w is the relationship plane W r The unit normal vector of the hyperplane W r For each relation r, a hyperplane is given on which the relation d is defined. r , map the original head entity h and tail entity t to the hyperplane and name it h ⊥ ,t ⊥ , and h ⊥ , t ⊥ Satisfy the following formula:

[0118] h ⊥ +d r =t ⊥

[0119] Step 2: According to h w Obtain the mapping vector h ⊥ , the calculation process is:

[0120] h ⊥ =hh w =hw T hw

[0121] Step 3: Find the overall mapping vector. The calculation process is:

[0122]

[0123]

[0124] Among them, h′ ⊥ , h′, etc. are generated negative examples;

[0125] Step 4: Calculate the loss function and reduce the loss value during the iteration process. The corresponding loss calculation formula is:

[0126]

[0127] Among them, G represents the set of correct triples, and G′ is the set of negative examples;

[0128] 2-5) Use the Euclidean distance formula to calculate the matching degree of entity and relationship pairs, and obtain the life type node that is most likely related to the new life vibration feature node. The corresponding Euclidean distance calculation formula is:

[0129]

[0130] The negative sign is used to convert the distance into a score. The smaller the distance, the higher the score, indicating that the matching degree between the entity pair and the relationship pair is higher. i and t i Represent the embedding vectors of the head entity and the tail entity respectively;

[0131] 2-6) Use multiple deep learning models (LSTM, TCN-LSTM-A, LSTM-SA) to predict bearing life, and combine the knowledge graph prediction results to comprehensively predict bearing life. Among them, the LSTM model (Long Short-Term Memory) is a special recurrent neural network (RNN) designed to solve the gradient vanishing or gradient exploding problem of standard RNN when processing long sequence data; TCN-LSTM-A is a deep learning model that combines temporal convolutional network (TCN) and long short-term memory network (LSTM) and attention mechanism (Attention) to process sequence data; LSTM-SA model is a neural network model that combines long short-term memory network (LSTM) and self-attention mechanism (Self-Attention).

[0132] S5: Add new life vibration feature nodes, similarity relationships between new nodes and basic graph nodes, life prediction results and other rule knowledge into the graph.

[0133] Based on the above steps, the present invention effectively solves the problem that the bearing lacks historical operating data and cannot be accurately predicted under the cold start environment, and proposes a bearing life prediction method based on knowledge graph under cold start. The method includes two modules: vector-based equipment life knowledge graph construction and cosine similarity-based TransH life prediction. Among them, the vector-based equipment life knowledge graph construction realizes the life extraction of equipment text data and the feature extraction of equipment vibration information, obtains the description of equipment life information and time-frequency domain feature information, combines the life information and time-frequency domain feature extraction to extract degradation rules, and realizes the construction of knowledge graph. The TransH life prediction based on cosine similarity obtains the similarity between the new node and the graph node through similarity analysis, establishes a similar_to correspondence, and uses the TransH algorithm to project the entity onto the hyperplane defined by the relationship, and learns the low-dimensional vector representation of the entity and the relationship. By calculating the scores to link the potential associations between predicted entities and combining them with the prediction results of the deep learning model, the most likely connections are finally identified and recommended in the knowledge graph to realize the remaining life prediction of vibration feature nodes in the new environment. New life vibration feature nodes and their associated rule knowledge are continuously added to the graph, and the basic knowledge graph is iteratively updated to realize personalized growth of the graph, thereby realizing accurate prediction of bearing life in a cold start environment.

[0134] Experimental verification of the bearing life prediction based on knowledge graph under cold start of the present invention:

[0135] 1. Test environment

[0136] Windows 11 system, Intel(R) Core(TM) i5-8250U CPU@1.60GHz, development environment is Python3.8.19, deep learning framework is PyTorch 2.3.0, and Neo4j is Neo4j Desktop.

[0137] 2. Test verification

[0138] Experimental results and analysis are performed on the public datasets PHM2012 and XJTU.

[0139] (1) Dataset description

[0140] The experiment uses the PHM2012 challenge dataset to build the equipment life knowledge graph, and the XJTU rolling bearing dataset is used as the equipment life data collected in the new environment. The sampling frequency of the bearing vibration signal in the PHM2012 dataset is 25.6kHz, and it is sampled once every 10s, and 2560 vibration data points are collected each time. The PHM2012 dataset contains three different operating environment conditions, involving 13 bearing samples, and 6 of them are selected as the starting data layer data of the graph. The specific graph construction bearing data is shown in Table 1.

[0141] Table 1 Bearing data for graph construction

[0142]

[0143] The XJTU rolling bearing dataset is used as the equipment life data collected in the new environment. It contains three working conditions, each of which contains five sets of bearing vibration data. The sampling frequency is 25.6kHz, and the data is collected once every 1 minute. The sampling time is 1.28s. The specific atlas validation set data is shown in Table 2.

[0144] Table 2 Bearing data of the atlas validation set

[0145]

[0146] (2) Experimental details and evaluation indicators

[0147] The model code is implemented based on Pytorch, and the knowledge graph is built through Neo4j. Adam optimizer is selected, where Adam is an adaptive learning rate optimization algorithm used to effectively and efficiently train models in deep learning. As the optimization model parameter. The data is normalized before model training and testing, based on part of the data of the PHM2012 challenge dataset. This paper uses 1min as the life rule chain Δd, and uses A X represents the source conditions of different vibration data in the PHM2012 challenge dataset, N represents different equipment life vibration information datasets from the PHM2012 challenge dataset, and the vectorization results of the atlas content are shown in Table 3.

[0148] Table 3. Partial vectorization results of equipment life knowledge graph

[0149]

[0150] In the TCN-LSTM-A model for decision support, the model configuration includes a convolution kernel size of 3 for the TCN part, 64 hidden units in the LSTM layer, 2 layers, and a dropout of 0.4. The training parameters are the Adam optimizer, learning rate 0.001, 100 rounds of training, and batch size 2; the LSTM model includes a bidirectional LSTM layer with 128 hidden units, 2 layers stacked, and a dropout of 0.4. The training parameters are also the Adam optimizer, learning rate 0.001, 100 rounds of training, and batch size 2; the LSTM-SA model includes an LSTM layer with 128 hidden units, 2 layers stacked, and an attention layer. The training parameters are the Adam optimizer, learning rate 0.001, and 100 rounds of training.

[0151] In order to evaluate the performance of the model, mean absolute error (MAE) and root mean square error (RMSE) are used as evaluation indicators. The calculation process is shown in the following formula:

[0152]

[0153]

[0154] In the formula, y i represents the true value of the remaining equipment life at the i-th moment, It represents the prediction result of equipment life at the i-th moment, and n represents the number of samples in the test set.

[0155] (3) Analysis of comparative experimental results

[0156] In order to verify the prediction effect of the bearing life prediction method based on knowledge graph under cold start, predictions were made in 1_1, 1_2, 1_3, and 2_2 of the XJTU dataset respectively. The models selected for the comparative experiment were LSTM, TCN-LSTM-A, and LSTM-SA algorithms. These three models can capture key information in the bearing degradation process from different angles, and can further verify the generalization ability and prediction performance of the model under different working conditions. For the convenience of comparison, the learning rates of the three models are set to 0.001, and the training rounds are all 100 times. In the cold start environment, the experimental comparison results of the three models and the method in this paper are shown in Table 4.

[0157] Table 4 Comparison of prediction results of LSTM, TCN-LSTM, LSTM-SA and this method

[0158]

[0159]

[0160] According to the experimental results, in the cold start environment, the overall prediction accuracy of the three models, LSTM, TCN-LSTM-A, and LSTM-SA, is low, and the MAE and RMSE values ​​of the models are generally high. The MAE of LSTM on the Bearing1_1 dataset is as high as 112.8231, and the RMSE reaches 117.1915, while the model in this paper only shows a MAE of 1.3333 and a RMSE of 1.8697 on the same dataset. This is because the LSTM model and its variants usually rely on long-term dependencies in the data for prediction. In the absence of historical training data in the new environment, the three models cannot establish effective time dependencies, which in turn affects the stability of the prediction results. Based on the deep learning prediction results as auxiliary decision-making, the method in this paper introduces knowledge graphs, so that the model can more accurately capture the features related to the prediction task while considering the temporal and spatial relationships between samples, and find the life type nodes that are most likely to be associated with the new life vibration data. It has more reliable prediction results and prediction accuracy, which reflects the superiority of the method in this paper in the life prediction problem in the cold start environment.

[0161] (4) Similarity analysis ablation experiment

[0162] To verify the effectiveness of the similarity analysis in this paper, the similarity analysis method of this paper is compared with the non-similarity analysis method, the basic cosine similarity analysis method (Cosine Similarity) and the similarity analysis method in 1_4, 1_5, 2_4, and 3_5 of the XJTU dataset. The comparison results are shown in the figure. Figure 7 shown.

[0163] Table 5 Comparison of prediction results between NSC, CS and KG-CS method in this paper

[0164]

[0165]

[0166] In the figure, NSC, CS, and KG-CS represent the partial running results of the non-similarity analysis method, the basic cosine similarity analysis method, and the similarity analysis method of this paper. From the results in the table, it can be seen that the loss value of the similarity analysis method of this paper is lower. This is because the method of this paper can analyze each component of the signal in detail, rather than treating the signal as a whole, thereby improving the frequency resolution and accuracy of the analysis, identifying the key frequency components in the signal, and reflecting the energy distribution of the signal, thereby providing more accurate signal analysis results. According to the experiment, the accuracy of the similarity analysis method of this paper in finding similar vibration signal nodes can be verified, which can effectively assist the spectrum life reasoning.

[0167] In summary, the bearing life prediction method based on knowledge graph under cold start proposed in this paper has more advantages than LSTM and its variant deep learning algorithms in the cold start industrial environment. This method provides an effective solution for bearing life prediction under cold start environment by establishing a bearing life knowledge graph and effectively using bearing operation data to construct a life rule knowledge graph.

[0168] 3. Conclusion

[0169] In view of the problem that it is difficult to complete the bearing life prediction by building a deep learning model under cold start, this paper proposes a bearing life prediction method based on knowledge graph under cold start. The method consists of two parts. The first part is the construction of equipment life knowledge graph based on vector, that is, the life rule chain of vibration signal is extracted through the text data and vibration information of bearing life vibration data in the same field to realize the construction of equipment life knowledge graph; the second part is TransH life prediction based on cosine similarity, that is, by calculating the power spectrum and cosine similarity of the decomposed signal, the graph vibration signal similar to the new vibration signal is obtained, and the entity is embedded into the hyperplane defined by the relationship through the TransH algorithm, and the link prediction of the new life vibration feature node is completed by combining the prediction results of the deep learning model. During the operation process, the new life vibration feature nodes and their associated rule knowledge are continuously added to the graph to realize the personalized growth of the graph. The experimental results show that the proposed method effectively improves the accuracy of bearing life prediction under cold start environment, and can effectively use the bearing operation data to construct the life rule knowledge graph, which is convenient for on-site staff to understand, improves its applicability and reliability in actual industrial applications, and provides an effective solution for bearing life prediction under cold start environment.

Claims

1. A bearing life prediction method based on knowledge graph under cold start, characterized in that: The steps include: S1: Build a vector-based equipment life knowledge graph based on equipment text data and equipment vibration information from public data sets in the field of bearing life; S2: Collect the bearing operation vibration data of the current environment and pre-process the data, including removing null values ​​and abnormal values, and cutting the data to make it consistent with the data set content size of the basic map, and generate a data set that meets the requirements of subsequent models; S3: performing variational mode decomposition on the newly collected bearing operation vibration data; S4: According to the characteristic information of the new node, the TransH life prediction method based on cosine similarity is used to select the life type node that is most likely to be associated with the new life vibration signal node, and the deep learning model is combined to assist the prediction; S5: Add new life vibration feature nodes, similarity relationships between new nodes and basic graph nodes, life prediction results and other rule knowledge into the graph.

2. According to the method for predicting bearing life under cold start based on knowledge graph in claim 1, it is characterized by: In step S1, a vector-based equipment life knowledge graph is constructed based on equipment text data and equipment vibration information of a public data set in the field of bearing life, including the following steps: 1-1) Obtain a public data set in the field of bearing life, and define the knowledge graph rule layer and the data layer to represent the knowledge graph life rule graph and the actual vibration feature data respectively. The rule layer contains the life rule knowledge graph, and the data layer stores the actual operation feature data, which corresponds one-to-one with the nodes in the rule layer; 1-2) Extract the equipment text data of the data set to obtain the bearing life information and data source information; extract the equipment vibration information to obtain the data vibration characteristics, and define the labels of the life vibration characteristic node, life type node and data source node corresponding to the node; 1-3) According to the characteristics of its own data, select the appropriate Δd as the life value difference, and the relationship between adjacent life type nodes is X n =X n+1 +Δd, where X n , X n+1 are the values ​​corresponding to adjacent life type nodes, and there is a difference in life values ​​of Δd; 1-4) Establish <X n , loss_Δd_rul,X n+1 > triple relationship, and link adjacent life type nodes, where loss_Δd_rul represents the loss from X n To X n+1 The life span value of Δd is lost; 1-5) define a life vibration characteristic node corresponding to each vibration signal, and customize the corresponding number to ensure that the number is unique and corresponds to a life vibration characteristic node; 1-6) If there is X rul , represents the actual life corresponding to a certain life vibration characteristic node, if X rul =X n , then establish the triple relationship of <life vibration feature node, belong_to, life type node>, that is, <life vibration feature node, belong_to, X n >; if X n <X rul <X n+1 , then establish the triple relationship of <Life vibration feature node, belong_to*value1, life type node>, <Life vibration feature node, belong_to*value2, life type node>, that is, <Life vibration feature node, belong_to*value1, X n >, <Life Vibration Feature Node, belong_to*value2, X n+1 >, the calculation process of value1 and value2 is as follows: Among them, the k value is the modulus selected according to the operating environment, belong_to indicates the relationship between a certain life vibration feature node and a certain life type node, and value1 and value2 indicate the weight of the life vibration feature node belonging to the life type node; 1-7) Perform variational modal decomposition on the equipment time series vibration data to extract features. The corresponding calculation process is as follows: Among them, K is the number of decomposed modes, {v k } is the IMFs component obtained by segment decomposition of time series vibration data, {ω k } is the center frequency of the component, δ(t) is the Dirac function, j is the imaginary unit, and j satisfies 2 =-1, is the time derivative operator, which is used to calculate the time derivative of the mode function. Indicated by ω k is the complex exponential function of the center frequency, g(t) is the original signal, t is the time variable, st represents that, that is, the constraint condition that the above formula needs to satisfy, min represents the minimum value, ||.||2 represents the L2 norm, that is, the square root of the sum of the squares of the vector components, and the Lagrange multiplication operator τ is introduced to transform it into an unconstrained variational problem: Among them, α is the second-order penalty factor, which ensures the accuracy of reconstruction of time series vibration data fragments under Gaussian noise environment, τ(t) represents the Lagrange multiplier, which is used to incorporate constraints into the optimization objective function, <·, ·> represents the inner product, calculates the dot product of two vectors, and uses the alternating direction multiplier method to update the components and their central frequencies, search for the saddle point of the Lagrangian function, and optimize the v after iteration. k ,ω k and τ is: Among them, γ is the noise tolerance, and They are v i (t), g(t) and τ n+1 The Fourier transform of (t), n is the number of iterative optimizations, and the iteration stops when the following equation is satisfied: Where ε is the given judgment accuracy; 1-8) Establish a triple relationship of <life vibration feature node, from, data source node>, where from indicates that the life vibration feature node comes from the data source node.

3. The method for predicting bearing life under cold start based on knowledge graph according to claim 1 is characterized in that: In step S4, according to the characteristic information of the new node, the TransH life prediction method based on cosine similarity is used to select the life type node that is most likely to be associated with the new life vibration signal node, and combined with the deep learning model to assist in prediction, including the following steps: 2-1) Use short-time Fourier transform on the new vibration signal. The steps are as follows: Step 1: Move the window to the beginning of the signal and perform windowing on the life vibration data, where f(t) is the original signal, w(t) is the window function, and the Hanning window is used to reduce spectrum leakage. τ0 represents the starting point of the window function. The corresponding formula is as follows: g(t)=f(t)·w(t-τ0) Step 2: Perform Fourier transform on g(t), as shown below: The spectrum distribution G(ω) of the first segment of the life vibration data is obtained. ω is the frequency variable, corresponding to the frequency component in Fourier transform. The function S(ω, τ) is defined to represent the spectrum result G(ω) after the original function f(t) is transformed when the center of the window function is τ. The process is as follows: Step 3: Move the window to τ1 and perform the next operation. Repeat sliding the corresponding window and finally obtain the spectrum results of all segments. 2-2) Calculate the power spectrum of the obtained spectrum results, and the process is as follows: P(ω,τ)=|S(ω,τ)| 2 2-3) Calculate the cosine similarity between the vibration data of the new life vibration feature node and the vibration data of the spectrum life vibration feature node. The calculation formula is as follows: Determine whether the similarity between the vibration signal of the new life vibration feature node and the vibration signal of the spectrum life vibration feature node is greater than ε. If the similarity is greater than ε, establish a triple relationship of <new life vibration feature node, similar_to, spectrum life vibration feature node>, where similar_to indicates that there is a similar relationship between the two life feature nodes; 2-4) Use the TransH algorithm to project entities onto the hyperplane defined by the relationship and learn low-dimensional vector representations of entities and relationships. TransH is a knowledge graph representation learning model that projects entities onto the hyperplane and vectorizes relationships. The specific steps are as follows: Step 1: Project the original vector h onto the relational plane W r , the unit normal vector h obtained by its projection w The process is: h w =w T hw Where w is the relationship plane W r The unit normal vector of the hyperplane W r For each relation r, a hyperplane is given on which the relation d is defined. r , map the original head entity h and tail entity t to the hyperplane and name it h ⊥ ,t ⊥ , and h ⊥ , t ⊥ Satisfy the following formula: h ⊥ +d r =t ⊥ Step 2: According to h w Obtain the mapping vector h ⊥ , the calculation process is: h ⊥ =h-h w =h_w T hw Step 3: Find the overall mapping vector. The calculation process is: Among them, h′ ⊥ , h′, etc. are generated negative examples; Step 4: Calculate the loss function and reduce the loss value during the iteration process. The corresponding loss calculation formula is: Among them, G represents the set of correct triples, and G′ is the set of negative examples; 2-5) Use the Euclidean distance formula to calculate the matching degree of entity and relationship pairs, and obtain the life type node that is most likely related to the new life vibration feature node. The corresponding Euclidean distance calculation formula is: The negative sign is used to convert the distance into a score. The smaller the distance, the higher the score, indicating that the matching degree between the entity pair and the relationship pair is higher. i and t i Represent the embedding vectors of the head entity and the tail entity respectively; 2-6) Use multiple deep learning models to predict bearing life, and combine the knowledge graph prediction results to comprehensively predict bearing life.

Images