A rotor slip tracing and regulation method based on heterogeneous knowledge graph reasoning

Abstract

The present invention discloses a method for rotor slip traceability and regulation based on heterogeneous knowledge graph reasoning. First, text knowledge and data knowledge related to rotor slip are collected. Based on the text knowledge and slip level division rules, text ontology modeling is carried out. Based on the data knowledge and adaptive distributed meta-learning, data ontology modeling is carried out. Among them, the text ontology modeling results and the adaptive communication strategy are used to construct a distributed network, and a slip knowledge graph is constructed through heterogeneous node mapping. Secondly, based on the meta-graph convolution-relation meta-learner, slip knowledge graph reasoning is carried out, and slip traceability is performed for the input text or data knowledge under unknown working conditions. Then, the prior knowledge of slip sensitivity is obtained by analyzing the influence degrees of working condition parameters, assembly parameters, structural parameters, and rotor characteristic parameters on the slip behavior. Finally, based on the slip traceability results and the prior knowledge of slip sensitivity, slip suppression measures are formulated and the control system is adjusted, thereby realizing rotor slip regulation.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing slip diagnosis, and mainly relates to a rotor slip tracing and regulation method based on heterogeneous knowledge graph reasoning. Background Technique

[0002] The phenomenon of rotor slip has become the main factor restricting the development of aeroengines towards ultra-high speed, low friction and light weight. The rotor slip behavior will cause equipment vibration, fatigue wear, motion instability and reduce the service life of the equipment, thereby significantly increasing the operation and maintenance costs. Rotor slip tracing helps to analyze the slip mechanism, and the regulation of rotor slip helps to inhibit the slip rate and reduce damage. Therefore, it is very necessary to study rotor slip tracing and regulation.

[0003] Research shows that the rotor slip behavior is closely related to working condition parameters, structural parameters, assembly parameters and rotor characteristic parameters. Therefore, it is difficult to meet the slip tracing requirements under actual complex and changeable working conditions only by using the classification algorithm based on deep learning. With the continuous enrichment of multi-source heterogeneous knowledge describing rotor slip behavior and the continuous development of knowledge graph research, the knowledge graph has great advantages in mining the potential relationship between text knowledge and data knowledge, and can realize the accurate tracing and mechanism analysis of slip behavior affected by multi-factor coupling. However, due to the complexity of the semantic relationship of slip text knowledge, the high cost of knowledge collection and the lack of global information capture, the constructed slip knowledge graph is incomplete. Therefore, it is necessary to apply knowledge reasoning to solve the problems of incomplete entity types and serious lack of entity relationships in the slip knowledge graph. Summary of the Invention

[0004] Object of the Invention: Based on the problems existing in the above background technique, the present invention provides a rotor slip tracing and regulation method based on heterogeneous knowledge graph reasoning, which can realize the accurate tracing and regulation of rotor slip behavior based on multi-source heterogeneous knowledge and knowledge graph reasoning.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A rotor slip tracing and regulation method based on heterogeneous knowledge graph reasoning, comprising the following steps:

[0007] Step S1, collect slip text knowledge and slip data knowledge related to rotor slip behavior under different working state parameters;

[0008] Step S2: Ontology modeling is carried out based on the slip text knowledge collected in Step S1; different working state parameters are used as slip condition nodes and divided into three levels according to the influence degree of different working state parameters on the slip rate. According to the slip level division rule, the slip rates in different ranges are divided into five levels and used as slip level nodes. Further, the association relationship between the slip condition nodes and the slip level nodes is established according to the slip level division rule;

[0009] Step S3: An adaptive distributed meta-learning network is constructed based on the slip text knowledge ontology modeling and the adaptive communication strategy obtained in Step S2, and the features of the slip data knowledge collected in Step S1 are extracted based on the constructed adaptive distributed meta-learning network; the working state parameters corresponding to different slip features are numbered to make different slip features correspond to a unique ID, thereby realizing the ontology modeling of the data knowledge;

[0010] Step S4: Based on the text ontology modeling result obtained in Step S2, the text data is mapped into symbolic slip condition nodes, slip level nodes and their corresponding relationships and stored in the knowledge graph; based on the features of the slip data knowledge extracted in Step S3, according to the unique ID corresponding to the working condition information of different slip features and mapping the ID into the fault time series nodes in the knowledge graph, a slip heterogeneous knowledge graph is constructed through heterogeneous node mapping and the storage and visualization of the slip heterogeneous knowledge are carried out using Neo4j software;

[0011] Step S5: Reasoning of the slip knowledge graph constructed in Step S4 is carried out based on the relational meta-learner - meta-graph convolutional network, and slip traceability is carried out for the input text knowledge or data knowledge under unknown working conditions, thereby determining the cause of the slip;

[0012] Step S6: Analyze the influence degree of the working condition parameters, assembly parameters, structural parameters and rotor characteristic parameters on the slip behavior and calculate and obtain the prior knowledge of slip sensitivity;

[0013] Step S7: Formulate slip suppression measures based on the slip traceability result obtained in Step S5 and the prior knowledge of slip sensitivity obtained in Step S6, and adjust the working state parameters through the control system to realize the regulation of rotor slip.

[0014] Preferably, the implementation process of Step S1 is as follows:

[0015] Step S1.1. The working state parameters related to the slipping behavior include working condition parameter A, assembly parameter B, structural parameter C, and rotor characteristic parameter D; collect the working condition parameter A, including lubricating oil flow rate A1, lubricating oil viscosity A2, lubricating oil temperature A3, axial load A4, radial load A5, acceleration A6, and speed A7; assembly parameter B, including assembly preload B1 and assembly interference B2; structural parameter C, including clearance ratio C1, guiding mode C2, number of rollers C3, radial clearance C4, surface roughness C5, and axial clearance C6; rotor characteristic parameter D, including resonance frequency D1, support stiffness D2, and the numerical values of the slipping rate caused by different working state parameters. Organize the working state parameters related to the slipping behavior and the numerical values of the slipping rate caused by different working state parameters into slipping text knowledge in text form;

[0016] Step S1.2. Determine the slipping test working state according to the range of the working state parameters covered by the slipping text knowledge collected in step S1.1. Collect vibration acceleration data under different working state parameters and organize them into slipping data knowledge in data form;

[0017] Preferably, the implementation process of step S2 is as follows:

[0018] Step S2.1. Take different working state parameters as slipping working condition nodes and divide the working state parameters into three levels according to the influence degree of different working state parameters on the slipping rate, which are respectively represented as severe level Ls = {A Ls , B Ls , C Ls , D Ls}, medium level Lm = {A Lm , B Lm , C Lm , D Lm}, and mild level Ll = {A Ll , B Ll , C Ll , D Ll}. Then assign evaluation scores of S Ls = 7, S Lm = 1, and S Ll = 0 to the three different levels:

[0019]

[0020]

[0021]

[0022] For any working state of the rotor, use the total evaluation score S all to characterize as follows:

[0023]

[0024] Among them, where S represents the evaluation score, x represents the working state parameter, and y represents the level of the influence degree of the corresponding working state parameter on the slip rate;

[0025] Step S2.2: Different slip rate ranges are divided into five slip level nodes according to the following rules: If the total evaluation score S corresponding to the rotor working state all ≤3, then the slip level node is K1: 0-2%; if the total evaluation score S corresponding to the rotor working state all =4, then the slip level node is K2: 2%-10%; if the total evaluation score S corresponding to the rotor working state all =5, then the slip level node is K3: 10%-20%; if the total evaluation score S corresponding to the rotor working state all =6, then the slip level node is K4: 20%-50%; if the total evaluation score S corresponding to the rotor working state all ≥7, then the slip level node is K5: >50%;

[0026] Step S2.3: Establish the association relationship between the slip working condition node and the slip level node, so that the slip working condition node obtained in step S2.1 corresponds to the slip level node obtained in step S2.2, and then obtain the slip text knowledge ontology modeling result.

[0027] Preferably, the implementation process of step S3 is as follows:

[0028] Step S3.1: Based on the slip text knowledge ontology modeling result obtained in step S2 and the adaptive communication strategy, construct an adaptive distributed meta-learning network, and extract the characteristics of the slip data knowledge collected in step S1.2 based on the constructed adaptive distributed meta-learning network, specifically including:

[0029] First, group the vibration acceleration data corresponding to different working state parameters collected in step S1.2 according to the slip level node division rule shown in step S2.2. Then, the vibration acceleration data set corresponding to the slip level node K1 is Among them, represents the N1th group of vibration acceleration data corresponding to the slip level node K1, and N1 is the number of groups of vibration acceleration data contained in the vibration acceleration data set The vibration acceleration data set corresponding to the slip level node K2 is Among them, represents the N2th group of vibration acceleration data corresponding to the slip level node K2, and N2 is the vibration acceleration data set The number of groups of vibration acceleration data included; the vibration acceleration data set corresponding to the slip level node K3 is Among them, represents the N3th group of vibration acceleration data corresponding to the slip level node K3, where N3 is the number of groups of vibration acceleration data included; the vibration acceleration data set corresponding to the slip level node K4 is Among them, represents the N4th group of vibration acceleration data corresponding to the slip level node K4, where N4 is the number of groups of vibration acceleration data included; the vibration acceleration data set corresponding to the slip level node K5 is Among them, represents the N5th group of vibration acceleration data corresponding to the slip level node K5, where N5 is the number of groups of vibration acceleration data included;

[0030] Secondly, according to the vibration acceleration data sets under different slip level nodes corresponding to different working state parameters train the adaptive distributed meta-learning network sub-model Among them, m = {1, 2, 3, 4, 5} is the slip level; is the parameter of the sub-model ; τ m is the number of parameters; is the sample included in the vibration acceleration data set; is the sample corresponding label; l m is the number of samples corresponding to the slip level m; N m is the number of sub-models corresponding to the slip level m; the sub-model parameters are trained according to the following rules:

[0031]

[0032]

[0033] Among them, θ is the learning rate; α is the balance parameter; is the parameter of the sub-model at the (t + 1)th iteration step; is the local cross-entropy loss function; is the loss function gradient; is the sub-model corresponding input x mi average output;

[0034] Then, based on the vibration acceleration data set The trained sub - model is integrated into a local model where is the parameter of the local model ; the sub - model trained based on the vibration acceleration dataset The trained sub - model is integrated into a local model where is the parameter of the local model ; the sub - model trained based on the vibration acceleration dataset The trained sub - model is integrated into a local model where is the parameter of the local model ; the sub - model trained based on the vibration acceleration dataset The trained sub - model is integrated into a local model where is the parameter of the local model ; the sub - model trained based on the vibration acceleration dataset The trained sub - model is integrated into a local model where is the parameter of the local model ; the parameters of the local model are integrated according to the following rules:

[0035]

[0036] where, is the integration weight of the local model at the t - th iteration step, which is updated according to the following rules:

[0037]

[0038] where, is the change amount of the loss function ; the parameter integration frequency of the local model is designed to be The integration frequency is adaptively adjusted according to the corresponding change amount of the loss function at different iteration steps {t, t + 1, …, t+ξ}; ξ is the total number of iteration steps;

[0039] Then, the parameters of the local models obtained under different slip levels are integrated into a global model According to the following rules:

[0040]

[0041] Among them, is the global model M global the integration weight at the t-th iteration step;

[0042] Finally, the global model parameters obtained at the t-th iteration step are used as the starting point of the sub-model training parameters for the (t + 1)-th iteration step, expressed as follows:

[0043]

[0044] Step S3.3: Number them according to the working state parameters corresponding to different slip characteristics, so that different slip characteristics correspond to unique IDs, specifically including:

[0045] Respectively, for the vibration acceleration data set and Use the adaptive distributed meta-learning network constructed in step S3.2 to extract features and assign unique IDs to the corresponding features as follows:

[0046]

[0047]

[0048]

[0049]

[0050]

[0051] Preferably, the implementation process of step S4 is as follows:

[0052] Step S4.1: Take the working condition parameter A, assembly parameter B, structural parameter C, and rotor characteristic parameter D as slip working condition nodes, take the slip levels {K1, K2, K3, K4, K5} as slip level nodes, and store the association relationship between the slip working condition nodes and slip level nodes obtained in step S2.3 into the knowledge graph;

[0053] Step S4.2: Based on the slip data features extracted in step S3, determine the corresponding according to the working state information corresponding to different slip data features, and map the IDs corresponding to different slip data features to the slip time sequence nodes in the knowledge graph respectively;

[0054] Step S4.4: Map the slip condition nodes established in Step S4.1 and the slip timing nodes established in Step S4.2 according to the working state parameters corresponding to the slip timing nodes to obtain a slip heterogeneous knowledge graph G = {E, R, T}, and use Neo4j software to store and visualize the slip heterogeneous knowledge, where E is the entity set of the knowledge graph; R is the relationship set of the knowledge graph; T = {(h, r, t) ∈ E × R × E} is the triple set; h and t are the head entity and tail entity corresponding to the relationship r.

[0055] Preferably, the implementation process of Step S5 is as follows:

[0056] Step S5.1: Decompose the slip heterogeneous knowledge graph G = {E, R, T} into five local neighborhoods according to the slip levels {K1, K2, K3, K4, K5} obtained in Step S2.2, and further divide the local neighborhoods into Z meta-graphs {μ1, μ2,..., μ Z} according to the number of nodes contained in each local neighborhood. Each meta-graph μ = {E μ , R μ , T μ} contains several entity types (h i , t i ) and relationship types r i , satisfying and where E μ is the entity set contained in the meta-graph μ; R μ is the relationship set contained in the meta-graph μ; T μ is the triple set contained in the meta-graph μ;

[0057] Step S5.2: Perform slip knowledge graph reasoning based on the relation meta-learner - meta-graph convolutional network, specifically including:

[0058] First, based on the meta-graphs {μ1, μ2,..., μ Z} obtained in Step S5.1, use the relation meta-learner - meta-graph convolutional network to extract relation metas respectively. The calculation process is as follows:

[0059] In the training task, there is a support set Ξ r ={(h i , r i ) ∈ E × E|(h i , r i , t i ) ∈ T}, and perform the following meta-graph convolution operation on all entities contained in the meta-graph μ = {μ1, μ2,…, μ χ}:

[0060]

[0061] Among them, R μ (h i , t i ) is the relational element corresponding to the entity pair (h i , t i ) in the meta-graph μ; normal(·) is a normalization function; x i and xj are the features of entities h i and t i respectively; σ(·) is an activation function; λ is the number of entities contained in the meta-graph μ; ω0 and ω η are weight coefficients; is the adjacency tensor of the meta-graph μ, which satisfies:

[0062]

[0063] Among them, I(·) is an indicator function; F μ (·) is an entity mapping function; η is the type corresponding to entity h i ;

[0064] Then, the embedding learner is used to score and rank the extracted relational elements and the generated triples, and the training loss of the relational element learner - meta-graph convolutional network is calculated based on the scoring function to update the model parameters. The calculation process is as follows:

[0065]

[0066] Among them, Γ(h i , t i ) is the scoring function corresponding to the training task Task r ; is the relational element corresponding to the training task Task r , which satisfies α zi is the weighted weight corresponding to the z-th meta-graph containing the entities (h i , t i ), which is obtained by applying the attention mechanism; Q is the number of entity pairs contained in the meta-graph μ; is the L2 norm; for the support set Ξ r , the designed loss function is as follows:

[0067]

[0068] Among them, [·] + is the positive operation; γ is the margin hyperparameter; then the meta-relation is updated as follows:

[0069]

[0070] Among them, β is the step size of the gradient element; the updated relational element The scores of the evaluation query tasks are as follows:

[0071]

[0072] Among them, Ω r ={(h j ,r j )∈E×E|(h j ,r j ,t j )∈T} is the query set; the proposed relational meta-learner-meta graph convolutional network will be updated based on the following loss function:

[0073]

[0074] Among them, T tra For all tasks of reasoning on heterogeneous knowledge graphs;

[0075] Finally, the ranking results of the relation elements and the generated triples are obtained, and the relation elements and triples with the highest scores are selected as the reasoning results of the sliding knowledge graph;

[0076] Step S5.3: Tracing the source of the slippage based on the input text knowledge or data knowledge under the unknown working condition to determine the cause of the slippage, specifically including:

[0077] First, for the text knowledge under the unknown working condition, the Cypher language provided by neo4j is used to realize the retrieval of the knowledge graph; for the data knowledge under the unknown working condition, the adaptive distributed meta-learning network constructed in step S3.1 is first used to extract the slip features of the data and match them with the slip data features corresponding to different IDs extracted in step S4.2, so as to obtain the corresponding slip timing nodes; for the slip condition nodes missing in the query process of the constructed slip heterogeneous knowledge graph G = {E, R, T}, the slip knowledge graph reasoning is performed using the relational meta-learner-metagraph convolutional network designed in step S5;

[0078] Then, for the corresponding slipping working state parameter information queried from the input text knowledge or data knowledge under the unknown working condition, the evaluation scores corresponding to different working state parameters are further queried based on the slipping state parameter level classification result in step S2.1. If a certain working state parameter is of severity level Ls={A Ls ,B Ls ,C Ls ,D Ls} or medium level Lm={A Lm ,B Lm ,C Lm ,D Lm}, then it is the cause of the rotor slip, and finally forms the slip cause set That is, the slip traceability result, where is the number of working state parameters that cause the rotor to slip, is for the th working state parameter corresponding to the slip cause;

[0079] Preferably, the implementation process of step S6 is as follows:

[0080] Analyze the influence degrees of different working condition parameters A, assembly parameters B, structural parameters C, and rotor characteristic parameters D on the slip behavior, and calculate and obtain the prior knowledge of slip sensitivity. The calculation process of the slip sensitivity of different working state parameters is as follows:

[0081]

[0082]

[0083]

[0084] where P i ini is the initial value of the i-th working state parameter; P i end is the final value of the i-th working state parameter; max(·) is to take the maximum value of the variable; is the change rate of the i-th working state parameter; is the initial value of the slip rate corresponding to the i-th working state parameter; is the final value of the slip rate corresponding to the i-th working state parameter; is the change rate of the slip rate corresponding to the i-th working state parameter; κ i is the slip sensitivity corresponding to the i-th working state parameter.

[0085] Preferably, the implementation process of step S7 is as follows:

[0086] Step S7.1, formulate slip suppression measures based on the slip traceability result obtained in step S5 and the prior knowledge of slip sensitivity corresponding to different working state parameters obtained in step S6:

[0087] According to the slip traceability result obtained in step S5 Search in the slip sensitivities corresponding to different working state parameters obtained in step S6. According to the slip traceability result C cau sort the slip sensitivities corresponding to different slip state parameters in C from large to small, and adjust the slip traceability result C in order from large to small according to the slip sensitivity cauThe slip state parameters therein; further query the evaluation scores corresponding to different working state parameters based on the slip state parameter level classification result in step S2.1, and sequentially set the adjustment target values of the slip working state parameters ;

[0088] Step S7.2, adjust the working state parameters through the control system according to the slip suppression measures determined in step S7.1, so as to realize the regulation of rotor slip.

[0089] Beneficial effects:

[0090] 1) The present invention constructs a slip heterogeneous knowledge graph to fully mine the correlation information between slip text knowledge and vibration data, and proposes a knowledge graph reasoning algorithm based on a relational meta-learner - meta-graph convolutional network, realizing accurate slip traceability and effective slip regulation; the proposed algorithm integrates rich theoretical, engineering and experimental knowledge, and has extremely high engineering application value;

[0091] 2) Formulated classification rules for working condition parameters, assembly parameters, structural parameters and rotor characteristic parameters closely related to the slip behavior according to the influence degree of different working state parameters on the slip behavior; and realized the embedding of slip time series nodes by assigning a unique ID to the slip data knowledge;

[0092] 3) Designed an adaptive distributed meta-learning network based on the specified slip level classification rules to extract data features under different working state parameters, and the network communication frequency is adaptively designed based on the change rate of the loss function;

[0093] 4) Realized slip knowledge graph reasoning based on the proposed relational meta-learner - meta-graph convolutional network, in which meta-graph partitioning and meta-graph convolutional operations are introduced to process the slip knowledge graph, significantly improving the accuracy of slip traceability. Description of the drawings

[0094] Figure 1 is a flowchart of a rotor slip traceability and regulation method based on heterogeneous knowledge graph reasoning provided by the present invention;

[0095] Figure 2 is a structural diagram of the adaptive distributed meta-learning network provided by the present invention;

[0096] Figure 3 is the adaptive communication frequency provided by the present invention;

[0097] Figure 4 is a classification confusion matrix diagram of vibration acceleration data under different slip levels obtained based on the adaptive distributed meta-learning network in the embodiment of the present invention;

[0098] Figure 5The slipping heterogeneous knowledge graph in the embodiments of the present invention;

[0099] Figure 6 The flowchart of the meta-learner - meta-graph convolutional network provided by the present invention;

[0100] Figure 7 The diagram of the division result of the inference data of the slipping knowledge graph in the embodiments of the present invention;

[0101] Figure 8 The result of the first-round slipping traceability in the embodiments of the present invention;

[0102] Figure 9 The result of the second-round slipping traceability in the embodiments of the present invention;

[0103] Figure 10 The similarity of candidate entities in the embodiments of the present invention;

[0104] Figure 11 The bar chart of the slipping sensitivity of different working state parameters in the embodiments of the present invention;

[0105] Figure 12 The topology diagram of the slipping sensitivity of different working state parameters in the embodiments of the present invention;

[0106] Figure 13 The circuit diagram of the rotor slipping regulation system in the embodiments of the present invention;

[0107] Figure 14 The slipping regulation result in the embodiments of the present invention. Detailed implementation manners

[0108] The present invention will be further described in detail below with reference to the accompanying drawings.

[0109] Refer to Figures 1-14 , this embodiment provides a method for rotor slipping traceability and regulation based on heterogeneous knowledge graph inference, and the flow of this method is as Figure 1 shown. The specific strategy includes the following steps:

[0110] Step S1, collect the text knowledge and data knowledge related to rotor slipping under different working state parameters {working condition parameter A, assembly parameter B, structural parameter C, and rotor characteristic parameter D};

[0111] Specifically, in this embodiment, this step S1 includes:

[0112] Step S1.1: Collect the slipping text knowledge corresponding to different working state parameters {operating condition parameter A, assembly parameter B, structural parameter C, and rotor characteristic parameter D}, and record the operating condition parameter A {lubricating oil flow rate A1, lubricating oil viscosity A2, lubricating oil temperature A3, axial load A4, radial load A5, acceleration A6, speed A7}, assembly parameter B {assembly pre-tightening force B1, assembly interference B2}, structural parameter C {clearance ratio C1, guiding method C2, number of rollers C3, radial clearance C4, surface roughness C5, axial clearance C6}, and rotor characteristic parameter D {resonance frequency D1, support stiffness D2} and the corresponding numerical values of the slipping rate caused by them.

[0113] Step S1.2: Determine the slipping test working state according to the range of working state parameters covered by the text knowledge collected in Step S1.1, and collect the corresponding vibration acceleration data under different operating condition parameters A {lubricating oil flow rate A1, lubricating oil viscosity A2, lubricating oil temperature A3, axial load A4, radial load A5, acceleration A6, speed A7}, assembly parameter B {assembly pre-tightening force B1, assembly interference B2}, structural parameter C {clearance ratio C1, guiding method C2, number of rollers C3, radial clearance C4, surface roughness C5, axial clearance C6}, and rotor characteristic parameter D {resonance frequency D1, support stiffness D2}.

[0114] Step S2: Conduct ontology modeling based on the text knowledge collected in Step S1; use different working state parameters as slipping condition nodes and divide them into three levels according to their influence on the slipping rate. According to the slipping level division rule, divide the slipping rates in different ranges into five levels and use them as slipping level nodes, and further establish the association relationship between the slipping condition nodes and the slipping level nodes according to the slipping level division rule.

[0115] Specifically, in this embodiment, this Step S2 includes:

[0116] Step S2.1: Use different working state parameters {operating condition parameter A, assembly parameter B, structural parameter C, rotor characteristic parameter D} as slipping condition nodes and divide them into three levels according to their influence on the slipping rate, which are respectively represented as severe level Ls = {A Ls , B Ls , C Ls , D Ls}, medium level Lm = {A Lm , B Lm , C Lm , D Lm}, and slight level Ll = {A Ll , B Ll , C Ll , D Ll}, and further assign evaluation scores of S to different levels respectively.Ls = 7, S Lm = 1 and S Ll = 0; Then the evaluation scores of the corresponding different working state parameters can be expressed as follows:

[0117]

[0118]

[0119]

[0120] For any working state of the rotor, the total evaluation score S can be used all to characterize as follows:

[0121]

[0122] Step S2.2, The slip rates in different ranges are divided into five slip levels according to the following rules: If the total evaluation score S corresponding to the rotor working state all ≤ 3, then the slip level is K1: 0 - 2%; If the total evaluation score S corresponding to the rotor working state all = 4, then the slip level is K2: 2% - 10%; If the total evaluation score S corresponding to the rotor working state all = 5, then the slip level is K3: 10% - 20%; If the total evaluation score S corresponding to the rotor working state all = 6, then the slip level is K4: 20% - 50%; If the total evaluation score S corresponding to the rotor working state all ≥ 7, then the slip level is K5: > 50%;

[0123] Step S2.3, Based on Equation (4) and the slip level division rules shown in Step S2.2, establish the association relationship between the slip working condition nodes and the slip level nodes, so that the corresponding slip working condition nodes correspond to the corresponding slip level nodes, and then obtain the slip text knowledge ontology modeling result;

[0124] Step S3, Based on the text knowledge ontology modeling result and the adaptive communication strategy obtained in Step S2, construct an adaptive distributed meta-learning network, and extract the slip characteristics of the data knowledge collected in Step S1 based on the constructed adaptive distributed meta-learning network; Number them according to the working state parameters corresponding to different slip characteristics, so that they correspond to a unique ID, and then realize the ontology modeling of the data knowledge;

[0125] Specifically, in this embodiment, this Step S3 includes:

[0126] Step S3.1, Based on the text knowledge ontology modeling result and the adaptive communication strategy obtained in Step S2, construct an adaptive distributed meta-learning network, such asFigure 2 As shown, and based on the constructed adaptive distributed meta-learning network, extract the characteristics of the slip data collected in step S1.2, specifically including:

[0127] First, group the vibration acceleration data corresponding to different working state parameters collected in step S1.1 according to the slip level division rule shown in step S2.2. Then, the vibration acceleration data set corresponding to slip level K1 is The vibration acceleration data set corresponding to slip level K2 is The vibration acceleration data set corresponding to slip level K3 is The vibration acceleration data set corresponding to slip level K4 is The vibration acceleration data set corresponding to slip level K5 is

[0128] Secondly, according to the vibration acceleration data sets under different slip levels corresponding to different working state parameters Train the sub-models of the adaptive distributed meta-learning network where m = {1, 2, 3, 4, 5} is the slip level; is the parameter of the sub-model τ m is the number of parameters; is the vibration acceleration data sample; is the sample corresponding label; l m is the number of samples corresponding to slip level m; N m is the number of sub-models corresponding to slip level m; the parameters of the sub-model are trained according to the following rules:

[0129]

[0130]

[0131] where θ is the learning rate; α is the balance parameter; is the parameter of the sub-model at the (t + 1)-th iteration step; is the local cross-entropy loss function; is the loss function gradient; is the sub-model corresponding input x mi average output;

[0132] Then, integrate the sub-models trained based on the vibration acceleration data sets into the local model where For the local model parameters; the sub-models trained based on the vibration acceleration dataset are integrated into the local model where are the parameters of the local model ; the sub-models trained based on the vibration acceleration dataset are integrated into the local model where are the parameters of the local model ; the sub-models trained based on the vibration acceleration dataset are integrated into the local model where are the parameters of the local model ; the sub-models trained based on the vibration acceleration dataset are integrated into the local model where are the parameters of the local model ; the parameters of the local model are integrated based on the following rules:

[0133]

[0134] wherein, is the integration weight of the local model at the t-th iteration step, which will be updated according to the following rules:

[0135]

[0136] wherein, is the change amount of the loss function ; the parameter integration frequency is adaptively designed as according to the change amounts of the loss functions corresponding to different iteration steps {t, t + 1,..., t + ξ} ξ is the total number of iteration steps; the results of adapting the communication frequency in the constructed adaptive distributed meta-learning network are as Figure 3 shown.

[0137] Then, the parameters of the local models obtained under different slip levels are integrated into the global model according to the following rules:

[0138]

[0139] wherein, is the global model M globalThe integration weight at the t-th iteration step;

[0140] Finally, the global model parameters obtained at the t-th iteration step are used as the starting point of the sub-model training parameters for the (t + 1)-th iteration step, which is expressed as follows:

[0141]

[0142] The classification confusion matrix results of the vibration acceleration data under different slip levels obtained based on the adaptive distributed meta-learning network are as Figure 4 shown.

[0143] Step S3.3: Number the working state parameters corresponding to different slip characteristics to make them correspond to a unique ID, specifically including:

[0144] Extract features from the vibration acceleration data set and using the adaptive distributed meta-learning network constructed in step S3.2, and assign a unique ID to the corresponding features as follows:

[0145]

[0146]

[0147]

[0148]

[0149]

[0150] Step S4: Map the text data into symbolic slip working condition nodes, slip level nodes and their corresponding relationships based on the text ontology modeling result obtained in step S2, and store them in the knowledge graph; based on the characteristics of the slip data knowledge extracted in step S3, map it into the fault time series nodes in the knowledge graph according to the unique ID corresponding to its working condition information, construct a slip heterogeneous knowledge graph through heterogeneous node mapping, and use Neo4j software to store and visualize the slip heterogeneous knowledge; Figure 5 The slip heterogeneous knowledge graph constructed for the present invention.

[0151] Specifically, in this embodiment, this step S4 includes:

[0152] Step S4.1: Map the text data into symbolic slip working condition nodes, slip level nodes and their corresponding relationships based on the text ontology modeling result obtained in step S2, and store them in the knowledge graph, specifically including:

[0153] Take the working condition parameters A {lubricating oil flow rate A1, lubricating oil viscosity A2, lubricating oil temperature A3, axial load A4, radial load A5, acceleration A6, speed A7}, assembly parameters B {assembly pre-tightening force B1, assembly interference B2}, structural parameters C {clearance ratio C1, guiding method C2, number of rollers C3, radial clearance C4, surface roughness C5, axial clearance C6} and rotor characteristic parameters D {resonance frequency D1, support stiffness D2} as the slip working condition nodes, take the slip levels {K1, K2, K3, K4, K5} as the slip level nodes, and store the association relationship between the slip working condition nodes and slip level nodes obtained in step S2.3 into the knowledge graph;

[0154] Step S4.2, based on the slip data features extracted in step S3, determine the corresponding and map them respectively to the slip time series nodes in the knowledge graph;

[0155] Step S4.4, construct a slip heterogeneous knowledge graph through heterogeneous node mapping and use Neo4j software for storage and visualization of slip heterogeneous knowledge, specifically including:

[0156] Map the slip working condition nodes established in step S4.1 and the slip time series nodes established in step S4.2 according to their corresponding working state parameters to obtain a slip heterogeneous knowledge graph G = {E, R, T}, and use Neo4j software for storage and visualization of slip heterogeneous knowledge, where E is the entity set of the knowledge graph; R is the relationship set of the knowledge graph; T = {(h, r, t) ∈ E × R × E} is the triple set; h and t are the head entity and tail entity corresponding to the relationship r;

[0157] Step S5, perform slip knowledge graph reasoning based on the relational meta-learner - meta graph convolutional network, and conduct slip traceability for the input text knowledge or data knowledge under unknown working conditions, and then determine the cause of slip;

[0158] Specifically, in this embodiment, this step S5 includes:

[0159] Step S5.1, decompose the slip heterogeneous knowledge graph G = {E, R, T} into five local neighborhoods according to the slip levels {K1, K2, K3, K4, K5} obtained in step S2.2, and further divide them into Z meta-graphs {μ1, μ2,..., μ Z} according to the number of nodes contained in each local neighborhood, and each meta-graph μ = {E μ , R μ , T μ} contains several entity types (h i , t i ) and relationship type r i, satisfying and

[0160] Step S5.2: Perform skidding knowledge graph reasoning based on the relational meta-learner - meta-graph convolutional network. The reasoning algorithm process is as Figure 6 shown, specifically including:

[0161] First, based on the meta-graph {μ1, μ2,..., μ Z} obtained in Step S5.1, use the relational meta-learner - meta-graph convolutional network to extract relational metas. The skidding knowledge graph reasoning data partitioning result is as Figure 7 shown, and the calculation process is as follows:

[0162] In the training task, there is a support set Ξ r ={(h i , r i ) ∈ E × E|(h i , r i , t i ) ∈ T}. Perform the following meta-graph convolutional operation on all entities included in the meta-graph μ = {μ1, μ2,..., μ χ}:

[0163]

[0164] where R μ (h i , t i ) is the relational meta corresponding to the entity pair (h i , t i ) in the meta-graph μ; normal(·) is the normalization function; x i and x j are the features of entities h i and t i respectively; σ(·) is the activation function; λ is the number of entities contained in the meta-graph μ; ω0 and ω η are weight coefficients; is the adjacency tensor of the meta-graph μ, which satisfies:

[0165]

[0166] where I(·) is the indicator function; F μ (·) is the entity mapping function; η is the type corresponding to entity h i ;

[0167] Then, use the embedding learner to score and rank the extracted relational metas and the generated triples, and calculate the training loss of the relational meta-learner - meta-graph convolutional network based on the scoring function to update the model parameters. The calculation process is as follows:

[0168]

[0169] Among them, Γ(h i ,t i ) is the scoring function corresponding to the training task Task r ; is the relational element corresponding to the training task Task r , satisfying α zi is the weighted weight corresponding to the z-th meta-graph containing the entity (h i ,t i ), which is obtained by applying the attention mechanism; Q is the number of entity pairs contained in the meta-graph μ; is the L2 norm; for the support set Ξ r , the designed loss function is as follows:

[0170]

[0171] Among them, [·] + is the positive operation; γ is the margin hyperparameter; then the meta-relation can be updated as follows:

[0172]

[0173] Among them, β is the step size of the gradient element; the updated relational element is used to evaluate the score of the query task as follows:

[0174]

[0175] Among them, Ω r = {(h j ,r j ) ∈ E×E|(h j ,r j ,t j ) ∈ T} is the query set; the proposed relational element learner - meta-graph convolutional network will be updated based on the following loss function:

[0176]

[0177] Among them, T tra is all tasks for heterogeneous knowledge graph reasoning;

[0178] Finally, the sorting results of the relational elements and the generated triples are obtained, and the relational element and triple with the highest score are selected as the skidding knowledge graph reasoning result;

[0179] Step S5.3, perform skidding traceability on the input text knowledge or data knowledge under unknown working conditions to determine the cause of skidding, specifically including:

[0180] First, the Cypher language provided by neo4j is used to retrieve the knowledge graph for the input text knowledge under unknown working conditions; for the input data knowledge under unknown working conditions, the adaptive distributed meta-learning network constructed in step S3.1 is first used to extract the slip features of the data and match them with the slip data features corresponding to different IDs extracted in step S4.2, so as to obtain the corresponding slip timing nodes; for the slip condition nodes missing in the query process of the constructed slip heterogeneous knowledge graph G = {E, R, T}, the slip knowledge graph reasoning is performed using the relational meta-learner-metagraph convolutional network designed in step S5;

[0181] Then, for the corresponding slip working state parameter information queried from the text knowledge or data knowledge under the unknown working condition, the evaluation scores corresponding to different working state parameters {working condition parameter A, assembly parameter B, structural parameter C, rotor characteristic parameter D} are further queried based on the slip state parameter level classification result in step S2.1. If a working state parameter is of severity level Ls={A Ls ,B Ls ,C Ls ,D Ls} or medium level Lm={A Lm ,B Lm ,C Lm ,D Lm}, then it is the cause of the rotor slip, and finally forms the slip cause set That is, the slip tracing result, where is the number of working state parameters that cause rotor slip, To correspond to Working status parameters of the reasons for slipping;

[0182] For example, the results of tracing the slip condition parameters of an unknown working condition vibration data are shown in Table 1. For the inference results output by the model, we hope to obtain a more accurate prediction of the slip level, so we can use the knowledge retrieval function of the knowledge graph to assist in decision-making. If the corresponding specific fault sequence node can be found, the slip level of the fault sequence node is used as the result. If not, the neighboring similar samples can be used as a reference or as new knowledge to update the graph in real time. The search results for the two-wheel slip condition parameters using the slip heterogeneous knowledge graph are shown in Table 1. Figure 8 and Figure 9 The first round of retrieval narrowed the search scope to 78 operating conditions, and the second round of retrieval narrowed the search scope to 6 candidate operating conditions.

[0183] Table 1 Tracing results of slip condition parameters

[0184]

[0185]

[0186] In the second-round retrieval, the slip rate levels for all possible working conditions are "Level 2: 2% - 10%". Comparing with the actual results, it is found that the prediction is correct and there is no need to correct the predicted slip level. Calculate the similarity between the current data and the candidate working conditions, and the results are as Figure 10 shown.

[0187] Step S6: Analyze the influence degrees of working condition parameters, assembly parameters, structural parameters, and rotor characteristic parameters on the slip behavior and calculate to obtain the prior knowledge of slip sensitivity;

[0188] Specifically, in this embodiment, this step S6 includes:

[0189] Analyze the influence degrees of different working condition parameters A {lubricating oil flow rate A1, lubricating oil viscosity A2, lubricating oil temperature A3, axial load A4, radial load A5, acceleration A6, speed A7}, assembly parameters B {assembly pre-tightening force B1, assembly interference B2}, structural parameters C {clearance ratio C1, guiding method C2, number of rollers C3, radial clearance C4, surface roughness C5, axial clearance C6}, and rotor characteristic parameters D {resonance frequency D1, support stiffness D2} on the slip behavior and calculate to obtain the prior knowledge of slip sensitivity. The calculation processes of slip sensitivity for different working state parameters are as follows:

[0190]

[0191]

[0192]

[0193] Among them, P i ini is the initial value of the i-th working state parameter; P i end is the final value of the i-th working state parameter; max(·) is to take the maximum value of the variable; is the change rate of the i-th working state parameter; is the initial value of the slip rate corresponding to the i-th working state parameter; is the final value of the slip rate corresponding to the i-th working state parameter; is the change rate of the slip rate corresponding to the i-th working state parameter; κ i is the slip sensitivity corresponding to the i-th working state parameter; Figure 11 and Figure 12 respectively show the bar chart and topological graph of the slip sensitivity corresponding to different working state parameters.

[0194] Step S7: Develop a slip suppression measure based on the slip traceability result obtained in Step S5 and the prior knowledge of slip sensitivity obtained in Step S6, and adjust the working state parameters through the control system to achieve rotor slip regulation;

[0195] Specifically, in this embodiment, Step S7 includes:

[0196] Step S7.1: Develop a slip suppression measure based on the slip traceability result obtained in Step S5 and the prior knowledge of slip sensitivity corresponding to different working state parameters obtained in Step S6, which specifically includes:

[0197] According to the slip traceability result obtained in Step S5 Search in the slip sensitivities corresponding to different working state parameters obtained in Step S6, and preferentially adjust the working state parameters that cause slip Among them, the working state parameters with large slip sensitivities; further query the evaluation scores corresponding to different working state parameter values {operating condition parameter A, assembly parameter B, structural parameter C, rotor characteristic parameter D} based on the slip state parameter level division result in Step S2.1, and sequentially set the adjustment target values of the slip working state parameters ;

[0198] Step S7.2: Adjust the working state parameters through the control system according to the slip suppression measure determined in Step S7.1, thereby achieving rotor slip regulation. Figure 13 The circuit diagram of the rotor slip regulation system is shown. The working state parameters such as rotor speed, acceleration, lubricating oil flow rate, lubricating oil temperature, axial load, and radial load can be adjusted respectively through the motorized spindle, lubricating oil pump, lubricating oil heating rod, and loading equipment. The slip test results under different working state parameters are as Figure 14 shown. By formulating accurate suppression measures based on the slip traceability result and using the control system to adjust the corresponding working state parameters, the slip rate will be significantly reduced.

[0199] The above is only the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A rotor slip tracing and control method based on heterogeneous knowledge graph reasoning, characterized in that: The following steps are involved: Step S1, collecting slip text knowledge and slip data knowledge related to rotor slip behavior under different working state parameters; Step S2, performing ontology modeling based on the slip text knowledge collected in step S1; taking different working state parameters as slip condition nodes and dividing them into three levels according to the influence of different working state parameters on the slip rate, dividing the slip rates in different ranges into five levels according to the slip level division rule and taking them as slip level nodes, and further establishing the association relationship between the slip condition node and the slip level node according to the slip level division rule; Step S3: construct an adaptive distributed meta-learning network based on the slip text knowledge ontology modeling and adaptive communication strategy obtained in step S2, and extract the features of the slip data knowledge collected in step S1 based on the constructed adaptive distributed meta-learning network; number different slip features according to the working state parameters corresponding to them, so that different slip features correspond to unique IDs, thereby realizing the ontology modeling of data knowledge; Step S4: Based on the text ontology modeling result obtained in step S2, the text data is mapped into symbolic slip condition nodes and slip level nodes and corresponding relationships and stored in the knowledge graph; based on the characteristics of the slip data knowledge extracted in step S3, the unique ID corresponding to the condition information of different slip characteristics is mapped into the fault sequence node in the knowledge graph, and the slip heterogeneous knowledge graph is constructed through heterogeneous node mapping, and the Neo4j software is used to store and visualize the slip heterogeneous knowledge; Step S5: performing reasoning on the slip knowledge graph constructed in step S4 based on the relational meta-learner-meta-graph convolutional network, and tracing the slip source for the input text knowledge or data knowledge under the unknown working condition, thereby determining the cause of the slip; Step S6, analyzing the influence of operating parameters, assembly parameters, structural parameters and rotor characteristic parameters on the slip behavior and calculating and obtaining prior knowledge of slip sensitivity; Step S7: based on the slip tracing result obtained in step S5 and the prior knowledge of slip sensitivity obtained in step S6, formulate slip suppression measures and adjust the working state parameters through the control system to achieve rotor slip control.

2. According to claim 1, a rotor slip tracing and control method based on heterogeneous knowledge graph reasoning is characterized in that: The implementation process of step S1 is: Step S1.1: Working state parameters related to the slip behavior include working condition parameters Assembly parameters Structural parameters and rotor characteristic parameters Collecting working condition parameters Including lubricating oil flow Lubricating oil viscosity Lubricating oil temperature Axial load Radial load Acceleration speed Assembly parameters Including assembly preload Assembly interference Structural parameters Including gap ratio Boot method Number of rollers Radial clearance Surface roughness Axial clearance Rotor characteristic parameters Including resonant frequency Support stiffness and the values ​​of slip rates caused by different working state parameters, and organizing the working state parameters related to the slip behavior and the values ​​of slip rates caused by different working state parameters into slip text knowledge in the form of text; Step S1.2, determine the working state of the slip test according to the working state parameter range covered by the slip text knowledge collected in step S1.1, collect vibration acceleration data under different working state parameters, and organize them into slip data knowledge in the form of data.

3. The rotor slip tracing and control method based on heterogeneous knowledge graph reasoning according to claim 2 is characterized in that: The implementation process of step S2 is: Step S2.1, taking different working state parameters as slip condition nodes and dividing the working state parameters into three levels according to the influence of different working state parameters on the slip rate, which are respectively expressed as severity level Medium and minor grade Then the three different levels are assigned evaluation scores: and Use the total evaluation score for any working state of the rotor The characterization is as follows: in, middle represents the evaluation score, x represents the working state parameter, and y represents the level of influence of the corresponding working state parameter on the slip rate; Step S2.2: Different slip rate ranges are divided into five slip level nodes according to the following rules: If the total evaluation score corresponding to the rotor working state is The slip level node is K1: 0-2%; if the total evaluation score corresponding to the rotor working state is The slip level node is K2: 2%-10%; if the total evaluation score corresponding to the rotor working state is The slip level node is K3: 10%-20%; if the total evaluation score corresponding to the rotor working state The slip level node is K4: 20%-50%; if the total evaluation score corresponding to the rotor working state Then the slip level node is K5:>50%; Step S2.3, establish an association relationship between the slip condition node and the slip level node, so that the slip condition node obtained in step S2.1 corresponds to the slip level node obtained in step S2.2, and then obtain the slip text knowledge ontology modeling result.

4. The rotor slip tracing and control method based on heterogeneous knowledge graph reasoning according to claim 3 is characterized in that: The implementation process of step S3 is: Step S3.1: construct an adaptive distributed meta-learning network based on the skidding text knowledge ontology modeling results and adaptive communication strategy obtained in step S2, and extract the features of the skidding data knowledge collected in step S1.2 based on the constructed adaptive distributed meta-learning network, specifically including: First, the vibration acceleration data corresponding to the different working state parameters collected in step S1.2 are grouped according to the slip level node division rule shown in step S2.2, and the vibration acceleration data set corresponding to the slip level node K1 is in, Indicates the N1th group of vibration acceleration data corresponding to the slip level node K1, N1 is the vibration acceleration data set The number of vibration acceleration data sets included; the vibration acceleration data set corresponding to the slip level node K2 is in, Indicates the N2th group of vibration acceleration data corresponding to the slip level node K2, N2 is the vibration acceleration data set The number of vibration acceleration data sets included; the vibration acceleration data set corresponding to the slip level node K3 is in, It represents the N3th group of vibration acceleration data corresponding to the slip level node K3, where N3 is the vibration acceleration data set θ K3 The number of vibration acceleration data sets included; the vibration acceleration data set corresponding to the slip level node K4 is in, Indicates the N4th group of vibration acceleration data corresponding to the slip level node K4, N4 is the vibration acceleration data set The number of vibration acceleration data sets included; the vibration acceleration data set corresponding to the slip level node K5 is in, Indicates the N5th group of vibration acceleration data corresponding to the slip level node K5, N5 is the vibration acceleration data set The number of vibration acceleration data sets included; Secondly, according to the vibration acceleration data set under different slip level nodes corresponding to different working state parameters Training Adaptive Distributed Meta-Learning Network Sub-Models Among them, m = {1, 2, 3, 4, 5} is the slip level; For sub-model Parameters; τ m is the number of parameters; is the sample contained in the vibration acceleration data set; For sample Corresponding label; l m N is the number of samples corresponding to the slip level m; m is the number of sub-models corresponding to the slip level m; The parameters of are trained according to the following rules: Among them, θ is the learning rate; α is the balance parameter; It is a sub-model Parameters at the (t+1)th iteration step; is the local cross entropy loss function; is the loss function The gradient of For sub-model Corresponding input x mi The average output of Then, based on the vibration acceleration data set The trained sub-model Integrate into local models in For local models Parameters; will be based on the vibration acceleration data set The trained sub-model Integrate into local models in For local models Parameters; will be based on the vibration acceleration data set The trained sub-model Integrate into local models in For local models Parameters; will be based on the vibration acceleration data set The trained sub-model Integrate into local models in For local models Parameters; will be based on the vibration acceleration data set The trained sub-model Integrate into local models in For local models Parameters of local model The parameters of are integrated based on the following rules: in, For local models The integrated weight at the tth iteration step is updated according to the following rules: in, is the loss function The amount of change; local model The parameter integration frequency is designed to be The integration frequency is adaptively adjusted according to the corresponding loss function change At different iteration steps {t, t+1, ..., t+ξ}; ξ is the total number of iteration steps; Then, the local models obtained at different slip levels are Parameter integration into a global model According to the following rules: in, is the global model M global The integration weight at the tth iteration step; Finally, the global model parameters obtained at the tth iteration step are As the starting point of the (t+1)th iteration step model training parameters, it is expressed as follows: Step S3.3, numbering different slip characteristics according to their corresponding working state parameters, so that different slip characteristics correspond to unique IDs, specifically including: The vibration acceleration data set and The adaptive distributed meta-learning network constructed in step S3.2 is used to extract features and assign unique IDs to the corresponding features as follows:

5. The rotor slip tracing and control method based on heterogeneous knowledge graph reasoning according to claim 4 is characterized in that: The implementation process of step S4 is: Step S4.1: Set the working condition parameters Assembly parameters Structural parameters and rotor characteristic parameters As the slip condition node, the slip levels {K1, K2, K3, K4, K5} are used as slip level nodes, and the association relationship between the slip condition node and the slip level node obtained in step S2.3 is stored in the knowledge graph; Step S4.2: Based on the slip data features extracted in step S3, determine the corresponding working state information corresponding to different slip data features. The IDs corresponding to different slip data features are mapped to slip time series nodes in the knowledge graph; Step S4.4: Map the slip condition node established in step S4.1 and the slip timing node established in step S4.2 according to the working state parameters corresponding to the slip timing node to obtain a slip heterogeneous knowledge graph Neo4j software is used to store and visualize heterogeneous knowledge. is the entity set of the knowledge graph; is the relationship set of the knowledge graph; is a set of triples; h and t are the head entity and tail entity corresponding to the relation r.

6. The rotor slip tracing and control method based on heterogeneous knowledge graph reasoning according to claim 5 is characterized in that: The implementation process of step S5 is: Step S5.1: According to the slip levels {K1, K2, K3, K4, K5} obtained in step S2.2, the slip heterogeneous knowledge graph It is decomposed into five local neighborhoods, and the local neighborhood is further divided into Z meta-graphs {μ1,μ2,...,μ Z }, each meta-graph Contains several entity types (h i ,t i ) and relationship type r i ,satisfy and in, is the entity set contained in the meta-graph μ; is the set of relations contained in the metagraph μ; is the set of triples contained in the metagraph μ; Step S5.2, performing skidding knowledge graph reasoning based on the relational meta-learner-meta-graph convolutional network, specifically includes: First, based on the meta-graph {μ1,μ2,...,μ Z }Relational meta-learner-meta graph convolutional network is used to extract relational meta-data, and the calculation process is as follows: There is a support set in the training task The meta-graph μ={μ1,μ2,...,μ χ All entities contained in} perform the following meta-graph convolution operation: in, is the entity pair (h i ,t i ) is the relational element; normal(·) is the normalization function; x i and x j The entity h i and t i ; σ(·) is the activation function; λ is the number of entities contained in the meta-graph μ; ω0 and ω η is the weight coefficient; is the adjacency tensor of the metagraph μ, which satisfies: Where I(·) is the indicator function; F μ (·) is the entity mapping function; η is the entity h i The corresponding type; Then, the extracted relational meta-tuples and generated triples are scored and ranked using the embedding learner, and the training loss of the relational meta-learner-meta graph convolutional network is calculated based on the scoring function to update the model parameters. The calculation process is as follows: Among them, Γ(h i ,t i ) is the training task Task r The corresponding scoring function; Task for training task r The corresponding relation element satisfies α zi To contain the entity (h i ,t i ) is the weighted weight corresponding to the zth meta-graph, which is obtained by applying the attention mechanism; Q is the number of entity pairs contained in the meta-graph μ; is the L2 norm; for the support set Ξ r The designed loss function is as follows: in,[·] + is the positive operation; γ is the marginal hyperparameter; then the element relationship Is updated as follows: Among them, β is the step size of the gradient element; the updated relationship element The scores of the evaluation query tasks are as follows: in, is the query set; the proposed relational meta-learner-meta graph convolutional network will be updated based on the following loss function: Among them, T tra For all tasks of reasoning on heterogeneous knowledge graphs; Finally, the ranking results of the relation elements and the generated triples are obtained, and the relation elements and triples with the highest scores are selected as the reasoning results of the sliding knowledge graph; Step S5.3: Tracing the source of the slippage based on the input text knowledge or data knowledge under the unknown working condition to determine the cause of the slippage, specifically including: First, for the text knowledge under the unknown working condition, the Cypher language provided by neo4j is used to realize the retrieval of the knowledge graph; for the data knowledge under the unknown working condition, the adaptive distributed meta-learning network constructed in step S3.1 is first used to extract the slip features of the data and match them with the slip data features corresponding to the different IDs extracted in step S4.2, so as to obtain the corresponding slip timing nodes; for the constructed slip heterogeneous knowledge graph If there is a missing node of the slip condition in the query process, the slip knowledge graph reasoning is performed using the relational meta-learner-meta-graph convolutional network designed in step S5; Then, for the corresponding slip working state parameter information queried from the text knowledge or data knowledge under the unknown working condition, the evaluation scores corresponding to different working state parameters are further queried based on the slip state parameter level classification result in step S2.

1. If a working state parameter is of the severe level or medium level Then it is the cause of the rotor slip, and finally forms a slip cause cluster. That is, the slip tracing result, where is the number of working state parameters that cause rotor slip, To correspond to Working status parameters of the reasons for slipping.

7. The rotor slip tracing and control method based on heterogeneous knowledge graph reasoning according to claim 6 is characterized in that: The implementation process of step S6 is: Analyze different operating parameters Assembly parameters Structural parameters and rotor characteristic parameters The influence degree on the slip behavior is calculated and the prior knowledge of slip sensitivity is obtained. The slip sensitivity calculation process of different working state parameters is as follows: Among them, P i ini is the initial value of the i-th working state parameter; P i end is the final value of the i-th working state parameter; max(·) is the maximum value of the variable; is the rate of change of the i-th working state parameter; is the initial value of the slip rate corresponding to the i-th working state parameter; is the final value of the slip rate corresponding to the i-th working state parameter; is the rate of change of the slip rate corresponding to the i-th working state parameter; κ i is the slip sensitivity corresponding to the i-th working state parameter.

8. The rotor slip tracing and control method based on heterogeneous knowledge graph reasoning according to claim 7 is characterized in that: The implementation process of step S7 is: Step S7.1: Formulate skidding suppression measures based on the skidding tracing result obtained in step S5 and the prior knowledge of skidding sensitivity corresponding to different working state parameters obtained in step S6: According to the skidding tracing result obtained in step S5 Search the slip sensitivity corresponding to the different working state parameters obtained in step S6, and according to the slip tracing result The slip sensitivity corresponding to different slip state parameters is sorted from large to small, and the slip tracing result is adjusted in the order of slip sensitivity from large to small. The slip state parameters in step S2.1 are further used to query the evaluation scores corresponding to different working state parameters based on the results of the slip state parameter classification, and the slip working state parameters are set in turn. The adjustment target value; Step S7.2: According to the slip suppression measures determined in step S7.1, the operating state parameters are adjusted through the control system to achieve rotor slip control.

Images