CN116561302A - Fault diagnosis method, device and storage medium based on mixed knowledge graph reasoning - Google Patents

Fault diagnosis method, device and storage medium based on mixed knowledge graph reasoning Download PDF

Info

Publication number
CN116561302A
CN116561302A CN202310072544.8A CN202310072544A CN116561302A CN 116561302 A CN116561302 A CN 116561302A CN 202310072544 A CN202310072544 A CN 202310072544A CN 116561302 A CN116561302 A CN 116561302A
Authority
CN
China
Prior art keywords
entity
rule
knowledge graph
representing
knowledge
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202310072544.8A
Other languages
Chinese (zh)
Inventor
韩慧慧
王坚
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Tongji University
Original Assignee
Tongji University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Tongji University filed Critical Tongji University
Priority to CN202310072544.8A priority Critical patent/CN116561302A/en
Publication of CN116561302A publication Critical patent/CN116561302A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/35—Clustering; Classification
    • G06F16/353—Clustering; Classification into predefined classes
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33—Querying
    • G06F16/3331—Query processing
    • G06F16/334—Query execution
    • G06F16/3346—Query execution using probabilistic model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
    • G06F16/367—Ontology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/901—Indexing; Data structures therefor; Storage structures
    • G06F16/9024—Graphs; Linked lists
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/02—Knowledge representation; Symbolic representation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/04—Inference or reasoning models
    • G06N5/042—Backward inferencing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/04—Manufacturing
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/30—Computing systems specially adapted for manufacturing

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Business, Economics & Management (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Probability & Statistics with Applications (AREA)
  • Manufacturing & Machinery (AREA)
  • Animal Behavior & Ethology (AREA)
  • Health & Medical Sciences (AREA)
  • Economics (AREA)
  • General Health & Medical Sciences (AREA)
  • Human Resources & Organizations (AREA)
  • Marketing (AREA)
  • Primary Health Care (AREA)
  • Strategic Management (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention relates to a fault diagnosis method, a device and a storage medium based on mixed knowledge graph reasoning, wherein the method comprises the following steps: acquiring data related to faults and equipment of a steel production line, and constructing a fault knowledge graph; constructing a mixed knowledge graph reasoning model, wherein the mixed knowledge graph reasoning model comprises a knowledge graph embedding model constructed based on a graph attention mechanism and transpark and a knowledge graph reasoning model constructed based on logic rules and reinforcement learning; training a mixed knowledge graph reasoning model; and performing fault diagnosis on the steel production line equipment by using the trained mixed knowledge graph reasoning model. Compared with the prior art, the invention has the advantages of high accuracy, stable training and the like.

Description

Fault diagnosis method, device and storage medium based on mixed knowledge graph reasoning
Technical Field
The invention relates to the technical field of knowledge discovery, in particular to a fault diagnosis method, device and storage medium based on mixed knowledge graph reasoning.
Background
The purpose of knowledge-graph reasoning is to infer unknown facts or relationships from the facts or relationships that are already in the knowledge graph. A large number of knowledge-graph inference methods are proposed, such as an embedded learning-based knowledge-graph inference method, a rule-based knowledge-graph inference method, a reinforcement learning-based knowledge-graph inference method. In order to make the model obtain an interpretable reasoning result, a multi-hop reasoning method based on reinforcement learning, such as a DeepPath, DIVA, MINERVA, M-Walk, multiHop classical knowledge graph model, is proposed. However, a significant challenge faced by these reinforcement learning-based knowledge-graph inference models in processing sparse knowledge graphs is that, due to insufficient information in the sparse knowledge graphs, there are insufficient paths between some entity pairs as inference evidence, and agents cannot accurately select the correct search direction and perform the inference process. In addition, in order to improve the interpretation of knowledge graph reasoning, several models based on symbol rules are proposed, including classical knowledge graph models such as NTP, neuralLP, ruleN, DRUM and RNNLogic. The rule-based knowledge graph reasoning model improves the interpretability by learning logic rules, and also enables the model to have generalization capability for similar tasks. However, the rule-based knowledge-graph inference method is difficult to extend to large-scale knowledge graphs, and the rule-based knowledge-graph inference model only focuses on the relationships of rule components, without considering the properties of related entities.
The steel production line mainly comprises manufacturing processes of steelmaking, continuous casting, hot rolling and the like, and comprises a large number of complex mechanical equipment such as a stepping furnace, a finishing mill, a coiling machine, a motor and the like. This is a typical complex manufacturing environment, with complex mechanisms, multiple parameters, frequent changes in various factors, and significant challenges in equipment failure maintenance. The knowledge graph has important significance in many applications, and is one of the latest research directions in the field of fault diagnosis. However, the traditional text searching method is low in efficiency in fault diagnosis due to too scattered knowledge, and the rule-based knowledge graph reasoning method cannot be expanded to a large-scale knowledge graph, and the properties of related entities are not considered.
Disclosure of Invention
The invention aims to provide a fault diagnosis method, a fault diagnosis device and a fault diagnosis storage medium based on mixed knowledge graph reasoning, which can be expanded into a large-scale knowledge graph by considering the properties of related entities, so as to improve the fault diagnosis accuracy.
The aim of the invention can be achieved by the following technical scheme:
a fault diagnosis method based on mixed knowledge graph reasoning comprises the following steps:
s1, acquiring data related to faults and equipment of a steel production line, and constructing a fault knowledge graph;
s2, constructing a mixed knowledge graph reasoning model, wherein the mixed knowledge graph reasoning model comprises a knowledge graph embedding model and a knowledge graph reasoning model;
the step S2 includes:
s21, constructing a knowledge graph embedding model based on a graph attention mechanism and transpark:
s211, obtaining entity embedding in the fault knowledge graph based on the multi-level graph attention mechanism module,
s212, embedding relations in the coding fault knowledge graph based on the transpark model;
s22, combining the knowledge graph embedding model, and constructing a knowledge graph reasoning model based on logic rules and reinforcement learning;
s3, training a mixed knowledge graph reasoning model;
s4, performing fault diagnosis on the steel production line equipment by using the trained mixed knowledge graph reasoning model.
The step S211 includes the steps of:
s2111, the constructed fault knowledge graph is expressed as G= { V, R, F }, wherein V represents an entity set, R represents a relation set, and F= { (e) 0 ，r，e t )|e 0 ，e t E, V, R E, R } represent fact triples, namely knowledge graphs are composed of a large number of triples;
s2112, using a relationship level attention mechanism to order importance of the relationship related to the current entity:
v hR ＝W R [h||R]
wherein the head entity vector h and the relation vector R are combined into a vector, and all neighborhood relation vectors related to the head entity h are evaluated, W R and wR Representing a learnable parameter, N h A neighborhood relation vector representing a head entity h;
s2113, under the current entity h and various relations of different importance degrees obtained in the previous step, evaluating entity groups corresponding to different relations;
v hRr ＝W T [v hR ||r]
wherein ,NhR Representing the intersection of the sub-relationship of R with the adjacent sub-relationship of h;
the relationships in the triples are converted into corresponding sub-relationships in advance, and the attention of the two levels is defined as:
v hr ＝W 1 [h||r]
γ hr ＝α hR ·β hRr
wherein ,W1 and w1 Representing a learnable parameter;
s2114, evaluating the importance degree of each neighborhood entity for the current entity of the given neighborhood relation based on the attention of the entity level:
v hrt ＝W 2 [v hr ||t]
wherein ,W2 and w2 Representing a learnable parameter, N hr Representing a relationship r lower head entity h neighborhood entity;
s2115, defining the attention degree of the head entity h to the fact triplet (h, r, t) as:
att hrt ＝γ hr ·η hrt
wherein ,γhr Attention, η, representing type level and relationship level hrt Representing entity-level attention;
s2116, defining neighborhood information of the entity:
r′＝W 3 r
wherein ,W3 Representing a learnable parameter;
s2117, obtaining an entity coding mode in the knowledge graph, wherein the entity coding mode is as follows:
h＝LeakyReLU(W 4 (h+h N ))
wherein ,W4 Representing a learnable parameter;
the encoded entity has rich neighborhood information, relationship type information and entity type information.
The step S212 specifically includes:
the TranSparse model is used for embedding the relation of the coding knowledge graph, and the target loss function of the relation embedding stage is defined as follows:
wherein I 2 Representing an L2 distance function, T representing a positive case aligned entity pair set, T' representing a negative case entity pair set resulting from negative sampling, gamma 2 Is a boundary super parameter；
Defining the objective function of the entity phase as:
wherein ,γ1 Is a boundary superparameter;
defining an objective function of the knowledge graph embedding stage as:
L Embed ＝L Ent +L Rel
in the model training process, the objective function L is optimized by continuously adjusting parameters Embed And obtaining better knowledge graph embedding.
The step S22 includes the steps of:
s221, combining the knowledge graph embedding model to construct a knowledge graph reasoning model based on reinforcement learning;
s232, storing paths of K before ranking of a knowledge graph reasoning model in each time step based on an improved beam searching algorithm, wherein K is the size of a beam;
s233, a high-level policy network based on logic rules is adopted to select the most probable path, and the path is expanded at each time step until the target entity is reached.
The step S221 specifically includes:
for knowledge graph reasoning tasks, the objective is to base on the problem q= (e) q ，r q What is? ) Deducing correct a from knowledge graph G nsw, wherein ,eq Representing an interrogating entity, r q Representing query relationships; the core of the knowledge graph reasoning model based on reinforcement learning is that the reinforcement learning network trains interaction between agent learning and knowledge graph environment;
reinforcement learning is essentially a Markov decision process, defined by a four-tuple of bitsI.e., status, action, transition and rewards, wherein,
the state is: at the position ofStatus of time->Is defined as +.> wherein /> and />Respectively represents historical track embedding, comprehensive entity embedding and node level graph attention embedding, wherein,
embedding history tracks into a memory module by adopting a long-short-term memory network LSTMThe definition is as follows:
wherein ,/>Representing the last entity +.>And the current entity->A relationship between;
comprehensive entity embeddingConsists of two parts:
wherein ,representing the current entity->And target entity e targ Is a distance of (2);
node level graph meaning embeddingFor helping agent to pay attention to current entity +.>Neighbor information of (c):
wherein ,W5 Represents a linear transformation matrix, N s Representation ofIs a neighborhood entity in the knowledge graph of (a)>Is->The attention weight between the individual entity and the j-th entity is calculated by a single-layer self-attention neural network:
wherein ,a learnable weight vector representing a share of all entities;
the action is a forward relationship path selected by the agent: proxy slave source entity e s Initially, a higher-level policy network based on logic rules is applied to select the most probable path, which is then extended at each step until the target entity e is reached targ The method comprises the steps of carrying out a first treatment on the surface of the Status ofIs->Is the current entity->A set of direction edges extending in the knowledge graph:
the transition is represented as a transition matrixThe effect is to ensure the probability distribution of the next state, defined as a mapping function +.>The strategy network encodes the current state and then outputs probability distribution wherein ,/>The goal of the transition strategy is +.>Selecting the action with the highest probability;
the reward R a Feedback to the agent is based on whether the behavior is valid and whether a series of behaviors can help the agent reach the target entity; given a pair ofAnd knowledge-graph, if the agent arrives at the target entity, i.e. +.>The agent gets a positive prize; if the agent arrives at the wrong entity, the agent will get a negative incentive; if the agent arrives at the entity without answer, i.e. +.>The agent will get a neutral reward; the ternary rewards structure is as follows:
the improved beam search algorithm specifically comprises the following steps:
will be at timePath set obtained here->The definition is as follows:
selecting candidate pathsThe principle of (1) is that the relation sequence can be matched with the related rule from left to right, i.e. the relation sequence in the previous path can be selected to generate +.>
The selection of candidate actions is divided into three phases:
a) And (3) random sampling: to eliminate the terrible influence of the pseudo path, the knowledge graph is based onRandom slaveLambda selection 1 K P Candidate paths other than K P The candidate paths with the highest scores are selected;
b) Action matching rules: selecting lambda meeting the correlation rule according to the action score 2 K P Action, and lambda 2 +λ 1 ≤1；
c) Action with higher score: selecting lambda from the remaining unmatched rule paths based on its score 3 K P A plurality of actions as supplement, wherein, when lambda 2 +λ 1 When=1, there is no stage c).
The high-level policy network based on the logic rule specifically comprises the following steps:
knowledge-graph reasoning is formalized in a probabilistic manner, where a set of logical rules z are treated as potential variables, the target distribution p (a nsw |g, q) is modeled jointly by a rule generator and an inference predictor:
rule generator p θ It is intended to generate a set of potential logic rules z for reasoning about the knowledge graph G to answer the query question q= (e) q ，r q What is? ) Formally, a query q= (e) q ，r q What is? ) By considering only the query relationship r q Irrespective of the querying entity e q Generating a combinational logic rule that allows the defined rule to be generalized across entities;
for short called r q ←r 1 ∧…∧r l Is regarded as a relation sequence r q ，r 1 ，r 2 …r l ，r END], wherein rq In order to query the relationship or rule header,is a rule main body, r END To represent a special relationship of the end of the relationship sequence;
introducing an LSTM network to parameterize a rule generator:
given a query relationship r q ，LSTM θ Generating each relation in the rule main body in turn until reaching the end relation r END In the process, the probability of generating the rule is calculated simultaneously;
then the distribution over a set of rules z is defined as a multiple-term distribution:
p θ (z|q)＝Mu(z|N，LSTM θ (·|r q ))
where Mu represents a polynomial distribution, N represents a hyper-parameter of the size of the set z, LSTM θ (·|r q ) Defining a distribution of combination rules, the rule head is r q ；
Inference predictor p w The objective of (a) is to infer a candidate answer e on G based on a given query q and a set of rules z cand Wherein, for each candidate answer e cand Scalar score of (a) w (e cand ) The calculation is as follows:
wherein ,ecand ∈A cand ，A cand Is a candidate answer set identifiable by any logical rule in the set z, P (e q ，rule，e cand ) Representing the rule being followed from e q Start to e cand The set of real paths that end up, and />Scalar weights representing each rule and path, respectively; each candidate answer e cand Is the sum of the scores contributed by each rule, score w (e cand The rule) is obtained by summing each real path found in the knowledge graph;
answer a to query q nsw Is entity e cand Is defined by applying a softmax function:
the step S3 includes the steps of:
s31 rule generator p θ First, a set of rules is generatedThe score H for each rule is calculated as follows:
wherein ,Acand Representation ofA set of all candidate answers learned by the rules in the rule score w (e|rule) represents the contribution rate of each rule to entity e, RNN θ (rule|r) represents the prior probability of the rule calculated by the generator;
s32, updating the reasoning predictor p according to the rule w ；
S33, drawing a sample for query qEach training example (G, q, a nsw ) Is approximated as follows:
for rulesLog p of a subset of (a) θ，w (z I |G，q，a nsw ) The approximation is:
wherein the method comprises the steps ofConst represents a single element independent of z I Constant term of (c), gamma (z) I ) Representing the set z I Rule indicates that each rule is at z I The sampling probability for each rule is calculated as follows:
s34, taking H (rule) of each rule as an evaluation of rule quality: k rules of highest H (rule) are selected, for each data instance (G, q, a nsw ) Forming a set of high quality logic rulesThe objective function is expressed as:
s35, updating the rule generator to be consistent with the selected high-quality rule, and applying the parameter theta in the high-quality rule updating rule generator to update the data of each data instance (G, q, a nsw ) High quality logic rules of (a)Is considered as part of the training data by maximizing +.>Updating and optimizing rule generator p by log likelihood θ 。
A fault diagnosis device based on mixed knowledge graph reasoning comprises a memory, a processor and a program stored in the memory, wherein the processor realizes the method when executing the program.
A storage medium having stored thereon a program which when executed performs a method as described above.
Compared with the prior art, the invention has the following beneficial effects:
(1) The knowledge graph embedding model constructed based on the graph attention mechanism and the transpark is particularly suitable for solving the problem of embedding complex heterogeneous knowledge graphs, can capture entity neighborhood information, relationship type information and entity type information in the knowledge graph, enables the model to embed richer semantic information, converts discrete graph search into vector space calculation, and greatly reduces search space.
(2) The invention provides a knowledge graph reasoning model based on logic rules and reinforcement learning, which greatly improves the interpretability of knowledge graph reasoning, effectively solves the problem of lack of paths of sparse knowledge graphs, and is not only suitable for small-sized knowledge graphs, but also can be expanded into large-sized knowledge graphs.
(3) The invention introduces an optimized beam searching algorithm, is beneficial to the opportunity of searching the discarded low-branch paths in the standard beam, improves the accuracy of the agent to select the correct action, and ensures that the reasoning result is more accurate.
Drawings
FIG. 1 is a flow chart of the method of the present invention;
FIG. 2 is a schematic diagram of knowledge graph construction;
FIG. 3 is a schematic diagram of a mixed knowledge graph inference model;
fig. 4 is a schematic diagram of a high-level policy network based on logic rules.
Detailed Description
The invention will now be described in detail with reference to the drawings and specific examples. The present embodiment is implemented on the premise of the technical scheme of the present invention, and a detailed implementation manner and a specific operation process are given, but the protection scope of the present invention is not limited to the following examples.
The embodiment provides a fault diagnosis method based on mixed knowledge graph reasoning, as shown in fig. 1, comprising the following steps:
s1, acquiring data related to faults and equipment of a steel production line, and constructing a fault knowledge graph, as shown in FIG. 2.
Specifically, the acquired data is divided into structured data, semi-structured data, and unstructured data.
The construction of the fault knowledge graph comprises the following steps:
s11, acquiring a relation triplet;
s12, knowledge fusion; in this embodiment, knowledge fusion is achieved through reference resolution, entity disambiguation, and entity linking.
S13, constructing and obtaining a fault knowledge graph.
S2, constructing a mixed knowledge graph reasoning model, as shown in figure 3.
S21, constructing a knowledge graph embedding model based on a graph attention mechanism and transpark.
S211, obtaining entity embedding in the fault knowledge graph based on the multi-level graph attention mechanism module.
S2111, the constructed fault knowledge graph is expressed as G= { V, R, F }, wherein V represents an entity set, R represents a relation set, and F= { (e) 0 ，r，e t )|e 0 ，e t E, V, R E, R } represent fact triples, namely knowledge graphs are composed of a large number of triples;
s2112, using a relationship level attention mechanism to order importance of the relationship related to the current entity:
v hR ＝W R [h||R]
wherein the head entity vector h and the relation vector R are combined into a vector, and all neighborhood relation vectors related to the head entity h are evaluated, W R and wR Representing a learnable parameter, N h A neighborhood relation vector representing a head entity h;
s2113, under the current entity h and various relations of different importance degrees obtained in the previous step, evaluating entity groups corresponding to different relations;
v hRr ＝W T [v hR ||r]
wherein ,NhR Representing the intersection of the sub-relationship of R with the adjacent sub-relationship of h;
the relationships in the triples are converted into corresponding sub-relationships in advance, and the attention of the two levels is defined as:
v hr ＝W 1 [h||r]
γ hr ＝α hR ·β hRr
wherein ,W1 and w1 Representing a learnable parameter;
s2114, evaluating the importance degree of each neighborhood entity for the current entity of the given neighborhood relation based on the attention of the entity level:
v hrt ＝W 2 [v hr ||t]
wherein ,W2 and w2 Representing a learnable parameter, N hr Representing a relationship r lower head entity h neighborhood entity;
s2115, defining the attention degree of the head entity h to the fact triplet (h, r, t) as:
att hrt ＝γ hr ·η hrt
wherein ,γhr Attention, η, representing type level and relationship level hrt Representing entity-level attention;
s2116, defining neighborhood information of the entity:
r′＝W 3 r
wherein ,W3 Representing a learnable parameter;
s2117, obtaining an entity coding mode in the knowledge graph, wherein the entity coding mode is as follows:
h＝LeakyReLU(W 4 (h+h N ))
wherein ,W4 Representing a learnable parameter;
the encoded entity has rich neighborhood information, relationship type information and entity type information.
S212, embedding relations in the coding fault knowledge graph based on the TranSparse model.
Because the constructed fault knowledge graph of the steel production line equipment has complex relationship and serious heterogeneity and unbalance, and the optimized multistage graph convolution attention network model can not encode the relationship embedding, the TranSparse model is used for the relationship embedding of the encoding knowledge graph, and the target loss function of the relationship embedding stage is defined as:
wherein I 2 Representing an L2 distance function, T representing a positive case aligned entity pair set, T' representing a negative case entity pair set resulting from negative sampling, gamma 2 Is a boundary superparameter.
Defining the objective function of the entity phase as:
wherein ,γ1 Is a boundary superparameter.
Defining an objective function of the knowledge graph embedding stage as:
L Embed ＝L Ent +L Rel
in the model training process, the objective function L is optimized by continuously adjusting parameters Embed And obtaining better knowledge graph embedding.
S22, combining the knowledge graph embedding model, and constructing a knowledge graph reasoning model based on logic rules and reinforcement learning.
S221, combining the knowledge graph embedding model to construct a knowledge graph reasoning model based on reinforcement learning.
For knowledge graph reasoning tasks, the objective is to base on the problem q= (e) q ，r q What is? ) Deducing correct a from knowledge graph G nsw, wherein ,eq Representing an interrogating entity, r q Representing query relationships.
The core of the knowledge graph reasoning model based on reinforcement learning is that the reinforcement learning network trains interaction between agent learning and knowledge graph environment; reinforcement learning is essentially a Markov decision process, defined by a four-tuple of bitsI.e., status, action, transition and rewards, in particular,
a) Status of
At the position ofStatus of time->Is defined as +.> wherein /> and />The historical track embedding, the comprehensive entity embedding and the node level graph meaning embedding are respectively represented.
Introducing historical track embeddingThe reason for (a) is that the agent needs to make sequence decisions, history trace embedded +.>The historical information can be recorded, so that the agent is effectively guided to effectively walk on the knowledge graph.
In order to better help the agent record and learn the history path, the long-short-term memory network LSTM is adopted as a memory component, so that the dependence of the knowledge graph on pre-training can be greatly reduced.
Embedding historical tracksThe definition is as follows:
wherein ,rt Representing the last entity e t-1 And current entity e t Relationship between them.
wherein ,representing the last entity +.>And the current entity->A relationship between;
comprehensive entity embeddingConsists of two parts:
wherein ,representing the current entity->And target entity e targ Is a distance of (2);
node level drawingItalian embeddingFor helping agent to pay attention to current entity +.>Neighbor information of (c):
wherein ,W5 Represents a linear transformation matrix, N s Representation ofIs a neighborhood entity in the knowledge graph of (a)>Is->The attention weight between the individual entity and the j-th entity is calculated by a single-layer self-attention neural network:
wherein ,/>A learnable weight vector representing a share of all entities;
b) Action
The action is a forward relationship path selected by the agent. Proxy slave source entity e s Initially, a higher-level policy network based on logic rules is applied to select the most probable path, which is then extended at each step until the target entity e is reached targ 。
Status ofIs->Is whenFront entity->A set of direction edges extending in the knowledge graph:
c) Transition
Transition matrixThe function of (a) is to ensure that the probability distribution of the next state, defined as a mapping functionThe strategy network encodes the current state and then outputs probability distribution +.> wherein ,the goal of the transition strategy is +.>The action with the highest probability is selected.
D) Rewards
Rewards R a Feedback to the agent is based on whether the action is valid and whether a series of actions can help the agent reach the target entity.
To solve the problem of inability to reach the answering entity within a limited number of steps performed by the agent, the present embodiment adds an additional "no answer" e NOANSWER In operation, therefore, a ternary rewards structure is presented. Given a pair ofAnd knowledge-graph, if the agent arrives at the target entity, i.e. +.>The agent gets a positive prize; such asThe fruit proxy arrives at the wrong entity, and the proxy obtains negative rewards; if the agent arrives at the entity without answer, i.e. +.>The agent will get a neutral reward; the ternary rewards structure is as follows:
s232, storing paths of K before ranking in each time step by a knowledge graph reasoning model based on an improved beam searching algorithm, wherein K is the size of a beam.
Will be at timePath set obtained here->The definition is as follows:
selecting candidate pathsThe principle of (a) is that the relationship sequence can match the correlation rule from left to right. For example, given rule r q ←r 1 ∧r 2 And two candidate paths (e s ，r 1 ，e 1 ，r 2 ，e 2 )，(e s ，r 2 ，e 3 ，r 3 ，e 4 ) Only if the relation sequence in the previous path can successfully match the rule will be selected to generate +.>
The selection of candidate actions is divided into three phases:
a) And (3) random sampling: to eliminate the terrible influence of the pseudo path, the knowledge graph is based onRandom slaveLambda selection 1 K P Candidate paths other than K P The candidate paths with the highest scores are selected;
b) Action matching rules: selecting lambda meeting the correlation rule according to the action score 2 K P Action, and lambda 2 +λ 1 ≤1；
c) Action with higher score: selecting lambda from the remaining unmatched rule paths based on its score 3 K P A plurality of actions as supplement, wherein, when lambda 2 +λ 1 When=1, there is no stage c).
With the optimized three-phase beam search algorithm described above, there is also an opportunity to explore the low-path that is discarded in the standard beam. Furthermore, the motion random sampling scheme may avoid that the model only selects the rewards path.
S233, a high-level policy network based on logic rules is adopted to select the most probable path, and the path is expanded at each time step until the target entity is reached.
The selection of actions in the markov decision process described above faces mainly two challenges, 1) random initialization of initial parameters and an increase in path length lead to difficulty in the model in selecting the correct actions to reach the target entity; 2) Because of the complexity of knowledge maps, the action space can be very large, and reinforcement learning-based methods often require extensive experimentation from scratch to find a reliable evidence path to obtain a non-zero reward. Since rules can accurately describe the mapping from query relationships to semantic composition paths, this embodiment designs a logic rule-based high-level policy network, as shown in FIG. 4, modeling reinforcement learning agents in sequential space. The rule is used as prior information of the action, so that the probability of obtaining rewards by the path can be improved, and effective exploration is facilitated.
The present embodiment uses five types of horns rules to mine on the knowledge graph: inverse rules, symmetric rules, pass rules, combining rules, and closed path rules.
Knowledge-graph reasoning is formalized in a probabilistic manner, where a set of logical rules z are treated as potential variables, the target distribution p (a nsw |g, q) is modeled jointly by a rule generator and an inference predictor:
rule generator p θ It is intended to generate a set of potential logic rules z for reasoning about the knowledge graph G to answer the query question q= (e) q ，r q What is? ) Formally, a query q= (e) q ，r q What is? ) By considering only the query relationship r q Irrespective of the querying entity e q Generating the combinational logic rules allows the defined rules to be generalized across entities.
For short called r q ←r 1 ∧…∧r l Is regarded as a relation sequence r q ，r 1 ，r 2 …r l ，r END], wherein rq In order to query the relationship or rule header,is a rule main body, r END To a special relationship that represents the end of a sequence of relationships.
Introducing an LSTM network to parameterize a rule generator:
given a query relationship r q ，LSTM θ Generating each relation in the rule main body in turn until reaching the end relation r END In this process, the probability of generating a rule is calculated at the same time.
Then the distribution over a set of rules z is defined as a multiple-term distribution:
p θ (z|q)＝Mu(z|N，LSTM θ (·|r q ))
where Mu represents a polynomial distribution, N represents a hyper-parameter of the size of the set z, LSTM θ (·|r q ) Define aDistribution of combination rules, rule head r q 。
Inference predictor p w The objective of (a) is to infer a candidate answer e on G based on a given query q and a set of rules z cand Wherein, for each candidate answer e cand Scalar score of (a) w (e cand ) The calculation is as follows:
wherein ,ecand ∈A cand ，A cand Is a candidate answer set identifiable by any logical rule in the set z, P (e q ，rule，e cand ) Representing the rule being followed from e q Start to e cand The set of real paths that end up, and />Scalar weights representing each rule and path, respectively; each candidate answer e cand Is the sum of the scores contributed by each rule, score w (e cand I rule) is obtained by summing each real path found in the knowledge-graph.
Answer a to query q nsw Is entity e cand Is defined by applying a softmax function:
s3, training a mixed knowledge graph reasoning model.
S31 rule generator p θ First, a set of rules is generatedThe score H for each rule is calculated as follows:
wherein ,Acand Representation ofA set of all candidate answers learned by the rules in the rule score w (e|rule) represents the contribution rate of each rule to entity e, RNN θ (rule|r) represents the prior probability of the rule calculated by the generator.
S32, updating the reasoning predictor p according to the rule w 。
S33, drawing a sample for query qEach training example (G, q, a nsw ) Is approximated as follows:
for rulesLog p of a subset of (a) θ，w (z I |G，q，a nsw ) The approximation is:
wherein const represents a single element independent of z I Constant term of (c), gamma (z) I ) Representing the set z I Rule indicates that each rule is at z I The sampling probability for each rule is calculated as follows:
s34, taking H (rule) of each rule as an evaluation of rule quality: k rules of highest H (rule) are selected, for each data instance (G, q, a nsw ) Forming a set of high quality logic rulesThe objective function is expressed as:
s35, updating the rule generator to be consistent with the selected high-quality rule, and applying the parameter theta in the high-quality rule updating rule generator to update the data of each data instance (G, q, a nsw ) High quality logic rules of (a)Is considered as part of the training data by maximizing +.>Updating and optimizing rule generator p by log likelihood θ 。
Therefore, the rule generator can learn to generate high-quality rules for the inference predictor to explore, so that the search space is reduced, and better inference results are generated.
S4, performing fault diagnosis on the steel production line equipment by using the trained mixed knowledge graph reasoning model.
The above functions, if implemented in the form of software functional units and sold or used as a stand-alone product, may be stored in a computer readable storage medium. Based on this understanding, the technical solution of the present invention may be embodied essentially or in a part contributing to the prior art or in a part of the technical solution, in the form of a software product stored in a storage medium, comprising several instructions for causing a computer device (which may be a personal computer, a server, a network device, etc.) to perform all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk, or an optical disk, or other various media capable of storing program codes.
In this embodiment, the data set is from a steel plant, 19 types of relationships are selected by sorting statistics, 2136 entities, the training set includes 32500 triplets, the test set includes 2080 triplets, the verification set includes 7600 triplets, and the comparison experiment of link prediction is performed between our model and other classical models, and MRR and Hit@1 are selected as evaluation indexes. The experimental results are shown in table 1. Experimental results show that compared with other methods, the mixed knowledge graph reasoning model provided by the invention has higher accuracy.
Table 1 experimental results
Model MRR hits@1
TransE 70.1 63.1
TransH 73.4 64.8
TransD 74.2 67.9
PRA 68.5 72.3
DeepPath 73.3 65.8
Multi-hop 74.1 66.7
The invention is that 83.5 78.1
The foregoing describes in detail preferred embodiments of the present invention. It should be understood that numerous modifications and variations can be made in accordance with the concepts of the invention by one of ordinary skill in the art without undue burden. Therefore, all technical solutions which can be obtained by logic analysis, reasoning or limited experiments based on the prior art by a person skilled in the art according to the inventive concept shall be within the scope of protection defined by the claims.

Claims (10)

1. The fault diagnosis method based on the mixed knowledge graph reasoning is characterized by comprising the following steps of:
s1, acquiring data related to faults and equipment of a steel production line, and constructing a fault knowledge graph;
s2, constructing a mixed knowledge graph reasoning model, wherein the mixed knowledge graph reasoning model comprises a knowledge graph embedding model and a knowledge graph reasoning model;
the step S2 includes:
s21, constructing a knowledge graph embedding model based on a graph attention mechanism and transpark:
s211, obtaining entity embedding in the fault knowledge graph based on the multi-level graph attention mechanism module,
s212, embedding relations in the coding fault knowledge graph based on the transpark model;
s22, combining the knowledge graph embedding model, and constructing a knowledge graph reasoning model based on logic rules and reinforcement learning;
s3, training a mixed knowledge graph reasoning model;
s4, performing fault diagnosis on the steel production line equipment by using the trained mixed knowledge graph reasoning model.
2. The fault diagnosis method based on mixed knowledge-graph inference of claim 1, wherein the step S211 comprises the steps of:
s2111, the constructed fault knowledge graph is expressed as G= { V, R, F }, wherein V represents an entity set, R represents a relation set, and F= { (e) 0 ,r,e t )|e 0 ,e t E, V, R E, R } represent fact triples, namely knowledge graphs are composed of a large number of triples;
s2112, using a relationship level attention mechanism to order importance of the relationship related to the current entity:
v hR ＝W R [h||R]
wherein the head entity vector h and the relation vector R are combined into a vector, and all neighborhood relation vectors related to the head entity h are evaluated, W R and wR Representing a learnable parameter, N h A neighborhood relation vector representing a head entity h;
s2113, under the current entity h and various relations of different importance degrees obtained in the previous step, evaluating entity groups corresponding to different relations;
v hRr ＝W T [v hR ||r]
wherein ,NhR Representing the intersection of the sub-relationship of R with the adjacent sub-relationship of h;
the relationships in the triples are converted into corresponding sub-relationships in advance, and the attention of the two levels is defined as:
v hr ＝W 1 [h||r]
γ hr ＝α hR ·β hRr
wherein ,W1 and W1 Representing a learnable parameter;
s2114, evaluating the importance degree of each neighborhood entity for the current entity of the given neighborhood relation based on the attention of the entity level:
v hrt ＝W 2 [v hr ||t]
wherein ,W2 and w2 Representing a learnable parameter, N hr Representing a relationship r lower head entity h neighborhood entity;
s2115, defining the attention degree of the head entity h to the fact triplet (h, r, t) as:
att hrt ＝γ hr ·η hrt
wherein ,γhr Attention, η, representing type level and relationship level hrt Representing entity-level attention;
s2116, defining neighborhood information of the entity:
r′＝W 3 r
wherein ,W3 Representing a learnable parameter;
s2117, obtaining an entity coding mode in the knowledge graph, wherein the entity coding mode is as follows:
h＝LeakyReLU(W 4 (h+h N ))
wherein ,W4 Representing a learnable parameter;
the encoded entity has rich neighborhood information, relationship type information and entity type information.
3. The fault diagnosis method based on mixed knowledge-graph reasoning according to claim 2, wherein the step S212 is specifically:
the TranSparse model is used for embedding the relation of the coding knowledge graph, and the target loss function of the relation embedding stage is defined as follows:
wherein I 2 Representing an L2 distance function, T represents a set of positive case-aligned entity pairs, T ′ Representing a set of negative instance entity pairs resulting from negative sampling, gamma 2 Is a boundary superparameter;
defining the objective function of the entity phase as:
wherein ,γ1 Is a boundary superparameter;
defining an objective function of the knowledge graph embedding stage as:
L Embed ＝L Ent +L Rel
in the model training process, the objective function L is optimized by continuously adjusting parameters Embed And obtaining better knowledge graph embedding.
4. A fault diagnosis method based on mixed knowledge-graph reasoning according to claim 3, wherein said step S22 comprises the steps of:
s221, combining the knowledge graph embedding model to construct a knowledge graph reasoning model based on reinforcement learning;
s232, storing paths of K before ranking of a knowledge graph reasoning model in each time step based on an improved beam searching algorithm, wherein K is the size of a beam;
s233, a high-level policy network based on logic rules is adopted to select the most probable path, and the path is expanded at each time step until the target entity is reached.
5. The fault diagnosis method based on mixed knowledge-graph reasoning according to claim 4, wherein the step S221 is specifically:
for knowledge graph reasoning tasks, the objective is to base on the problem q= (e) q ,r q What is? ) Deducing correct a from knowledge graph G nsw, wherein ,eq Representing an interrogating entity, r q Representing query relationships; the core of the knowledge graph reasoning model based on reinforcement learning is that the reinforcement learning network trains interaction between agent learning and knowledge graph environment;
reinforcement learning is essentially a Markov decision process, defined by a four-tuple of bitsI.e., status, action, transition and rewards, wherein,
the state is: at the position ofStatus of time->Is defined as +.> wherein /> and />Respectively represents historical track embedding, comprehensive entity embedding and node level graph attention embedding, wherein,
embedding history tracks into a memory module by adopting a long-short-term memory network LSTMThe definition is as follows:
wherein ,representing the last entity +.>And the current entity->A relationship between;
comprehensive entity embeddingConsists of two parts:
wherein ,representing the current entity->And target entity e targ Is a distance of (2);
node level graph meaning embeddingFor helping agent to pay attention to current entity +.>Neighbor information of (c):
wherein ,W5 Represents a linear transformation matrix, N s Representation ofIs a neighborhood entity in the knowledge graph of (a)>Is->The attention weight between the individual entity and the j-th entity is calculated by a single-layer self-attention neural network:
wherein ,a learnable weight vector representing a share of all entities;
the action is a forward relationship path selected by the agent: proxy slave source entity e s Initially, a higher-level policy network based on logic rules is applied to select the most probable path, which is then extended at each step until the target entity e is reached targ The method comprises the steps of carrying out a first treatment on the surface of the Status ofIs->Is the current entity->A set of direction edges extending in the knowledge graph:
the transition is represented as a transition matrixThe effect is to ensure the probability distribution of the next state, defined as a mapping function +.>Sxa→s; the strategy network encodes the current state and then outputs probability distribution +.> wherein ,/>The goal of the transition strategy is +.>Selecting the action with the highest probability;
the reward R a Feedback to the agent is based on whether the behavior is valid and whether a series of behaviors can help the agent reach the target entity; given a pair ofAnd knowledge-graph, if the agent arrives at the target entity, i.e. +.>Then take the place ofObtaining forward rewards; if the agent arrives at the wrong entity, the agent will get a negative incentive; if the agent arrives at the entity without answer, i.e. +.>The agent will get a neutral reward; the ternary rewards structure is as follows:
6. the fault diagnosis method based on mixed knowledge-graph reasoning according to claim 5, wherein the improved beam search algorithm specifically comprises:
will be at timePath set obtained here->The definition is as follows:
selecting candidate pathsThe principle of (1) is that the relation sequence can be matched with the related rule from left to right, i.e. the relation sequence in the previous path can be selected to generate +.>
The selection of candidate actions is divided into three phases:
a) And (3) random sampling: to eliminate the terrible influence of the pseudo path, the knowledge graph is based onRandom from->Lambda selection 1 K P Candidate paths other than K P The candidate paths with the highest scores are selected;
b) Action matching rules: selecting lambda meeting the correlation rule according to the action score 2 K P Action, and lambda 2 +λ 1 ≤1；
c) Action with higher score: selecting lambda from the remaining unmatched rule paths based on its score 3 K P A plurality of actions as supplement, wherein, when lambda 2 +λ 1 When=1, there is no stage c).
7. The fault diagnosis method based on mixed knowledge graph reasoning of claim 6, wherein the high-level policy network based on logic rules specifically comprises:
mining on the knowledge graph by using five types of horns rules, namely an inverse rule, a symmetry rule, a transfer rule, a combination rule and a closed path rule;
knowledge-graph reasoning is formalized in a probabilistic manner, where a set of logical rules z are treated as potential variables, the target distribution p (a nsw |g, q) is modeled jointly by a rule generator and an inference predictor:
rule generator p θ It is intended to generate a set of potential logic rules z for reasoning about the knowledge graph G to answer the query question q= (e) q ,r q What is? ) Formally, a query q= (e) q ,r q What is? ) By considering only the query relationship r q Irrespective of the querying entity e q Generating a combinational logic rule that allows the defined rule to be generalized across entities;
for short called r q ←r 1 ∧…∧r l Is regarded as a relation sequence r q ,r 1 ,r 2 …r l ,r END], wherein rq In order to query the relationship or rule header,is a rule main body, r END To represent a special relationship of the end of the relationship sequence;
introducing an LSTM network to parameterize a rule generator:
given a query relationship r q ，LSTM θ Generating each relation in the rule main body in turn until reaching the end relation r END In the process, the probability of generating the rule is calculated simultaneously;
then the distribution over a set of rules z is defined as a multiple-term distribution:
p θ (z|q)＝Mu(z|N,LSTM θ (·|r q ))
where Mu represents a polynomial distribution, N represents a hyper-parameter of the size of the set z, LSTM θ (·|r q ) Defining a distribution of combination rules, the rule head is r q ；
Inference predictor p w The objective of (a) is to infer a candidate answer e on G based on a given query q and a set of rules z cand Wherein, for each candidate answer e cand Scalar score of (a) w (e cand ) The calculation is as follows:
wherein ,ecand ∈A cand ,A cand Is a candidate answer set identifiable by any logical rule in the set z, P (e q ,rule,e cand ) Representing the rule being followed from e q Start to e cand The set of real paths that end up,andscalar weights representing each rule and path, respectively; each candidate answer e cand Is the sum of the scores contributed by each rule, score w (e cand The rule) is obtained by summing each real path found in the knowledge graph;
answer a to query q nsw Is entity e cand Is defined by applying a softmax function:
8. the fault diagnosis method based on mixed knowledge-graph inference of claim 7, wherein the step S3 comprises the steps of:
s31 rule generator p θ First, a set of rules is generatedThe score H for each rule is calculated as follows:
wherein ,Acand Representation ofA set of all candidate answers learned by the rules in the rule score w (e|rule) represents the contribution rate of each rule to entity e, RNN θ (rule|r) represents the prior probability of the rule calculated by the generator;
s32, updating the reasoning predictor p according to the rule w ；
S33, drawing a sample for query qEach training example (G, q, a nsw ) Is approximated as follows:
for rulesIs a subset of the logp θ,w (z I |G,q,a nsw ) The approximation is:
wherein const represents a single element independent of z I Constant term of (c), gamma (z) I ) Representing the set z I Rule indicates that each rule is at z I The sampling probability for each rule is calculated as follows:
s34, taking H (rule) of each rule as an evaluation of rule quality: k rules of highest H (rule) are selected, for each data instance (G, q, a nsw ) Forming a set of high quality logic rulesThe objective function is expressed as:
s35, updating the rule generator to be consistent with the selected high-quality rule, and applying the high-quality ruleThe parameter θ in the new rule generator will be the parameter θ for each data instance (G, q, a nsw ) High quality logic rules of (a)Is considered as part of the training data by maximizing +.>Updating and optimizing rule generator p by log likelihood θ 。
9. A fault diagnosis device based on mixed knowledge graph reasoning, comprising a memory, a processor and a program stored in the memory, characterized in that the processor implements the method according to any of claims 1-8 when executing the program.
10. A storage medium having a program stored thereon, wherein the program, when executed, implements the method of any of claims 1-8.
CN202310072544.8A 2023-01-13 2023-01-13 Fault diagnosis method, device and storage medium based on mixed knowledge graph reasoning Pending CN116561302A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310072544.8A CN116561302A (en) 2023-01-13 2023-01-13 Fault diagnosis method, device and storage medium based on mixed knowledge graph reasoning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310072544.8A CN116561302A (en) 2023-01-13 2023-01-13 Fault diagnosis method, device and storage medium based on mixed knowledge graph reasoning

Publications (1)

Publication Number Publication Date
CN116561302A true CN116561302A (en) 2023-08-08

Family

ID=87497194

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310072544.8A Pending CN116561302A (en) 2023-01-13 2023-01-13 Fault diagnosis method, device and storage medium based on mixed knowledge graph reasoning

Country Status (1)

Country Link
CN (1) CN116561302A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117076653A (en) * 2023-10-17 2023-11-17 安徽农业大学 Knowledge base question-answering method based on thinking chain and visual lifting context learning
CN118193756A (en) * 2024-05-16 2024-06-14 南京邮电大学 Knowledge graph rule learning method and system based on graph structure

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117076653A (en) * 2023-10-17 2023-11-17 安徽农业大学 Knowledge base question-answering method based on thinking chain and visual lifting context learning
CN117076653B (en) * 2023-10-17 2024-01-02 安徽农业大学 Knowledge base question-answering method based on thinking chain and visual lifting context learning
CN118193756A (en) * 2024-05-16 2024-06-14 南京邮电大学 Knowledge graph rule learning method and system based on graph structure

Similar Documents

Publication Publication Date Title
Li et al. Improved binary particle swarm optimization for feature selection with new initialization and search space reduction strategies
Qu et al. Rnnlogic: Learning logic rules for reasoning on knowledge graphs
Huang et al. Addressing the loss-metric mismatch with adaptive loss alignment
Zeng et al. Enhancing attention-based LSTM with position context for aspect-level sentiment classification
Risi et al. How novelty search escapes the deceptive trap of learning to learn
CN116561302A (en) Fault diagnosis method, device and storage medium based on mixed knowledge graph reasoning
Yu et al. Evolutionary fuzzy neural networks for hybrid financial prediction
Xiong Fuzzy rule-based similarity model enables learning from small case bases
Chen et al. Explain the explainer: Interpreting model-agnostic counterfactual explanations of a deep reinforcement learning agent
Wei et al. MoCo4SRec: A momentum contrastive learning framework for sequential recommendation
Bai et al. CLR-DRNets: Curriculum learning with restarts to solve visual combinatorial games
Kumar et al. Software Fault Prediction using Wrapper based Feature Selection Approach employing Genetic Algorithm
Cheddadi et al. Improving equity and access to higher education using artificial intelligence
Yang et al. Designing novel cognitive diagnosis models via evolutionary multi-objective neural architecture search
Lu et al. Counting crowd by weighing counts: A sequential decision-making perspective
Wang et al. Automated design of action advising trigger conditions for multiagent reinforcement learning: a genetic programming-based approach
Shao et al. Average User-Side Counterfactual Fairness for Collaborative Filtering
Newaz et al. A hybrid classification technique using belief rule based semi-supervised learning
Chen et al. Generating quantum feature maps for SVM classifier
Huang et al. In-Context Decision Transformer: Reinforcement Learning via Hierarchical Chain-of-Thought
Duan MCEDA: A novel many-objective optimization approach based on model and clustering
Ebrahimian et al. A CNN-based hybrid model and architecture for shilling attack detection
Zou et al. A multi-hop path query answering model for knowledge graph based on neighborhood aggregation and transformer
Farzi The design of self-organizing evolved polynomial neural networks based on learnable evolution model 3.
Xue et al. Neural architecture search via similarity adaptive guidance

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination