A helicopter transmission system fault information extraction method based on RL-GPNet

Abstract

The present invention discloses a helicopter transmission system fault information extraction method based on RL-GPNet, which relates to the technical field of helicopter transmission system faults. The method comprises entity recognition based on a global pointer network, and effectively captures long-distance dependencies and complex semantic information in fault description texts through global normalization and relative position encoding mechanisms. Simultaneously, the PPO algorithm is combined to model triple generation as a multi-step decision task, and entity recognition and relationship classification strategies are collaboratively trained to alleviate error accumulation caused by task conflicts. The method exhibits good generalization ability when processing helicopter transmission system faults involving overlapping entities in long texts and a large number of professional terms, providing effective technical support for intelligent fault diagnosis of helicopter transmission systems. The method has important theoretical significance and application value for improving the intelligent level of helicopter equipment maintenance and support.

Description

Technical Field

[0001] The present invention relates to the technical field of helicopter transmission system failure, and in particular to a method for extracting helicopter transmission system failure information based on RL-GPNet. Background Art

[0002] As a core component of mechanical equipment, helicopter transmission systems are becoming increasingly important for fault diagnosis and predictive maintenance. Failures in helicopter transmission systems can not only cause abnormal flight conditions but can also lead to major safety incidents. Therefore, constructing an accurate and complete knowledge graph for transmission system failures is crucial for intelligent fault diagnosis and improving equipment reliability. As a structured semantic knowledge base, a knowledge graph can represent and store entities, attributes, and relationships within a domain in the form of a graph, providing knowledge support for intelligent applications. In the field of helicopter transmission system fault diagnosis, knowledge graphs can integrate multi-dimensional information such as fault phenomena, causes, diagnostic methods, and maintenance strategies to form a systematic knowledge system.

[0003] However, there are some technical problems in current knowledge extraction, including: (1) There are a large number of professional terms and complex technical expressions in the field of helicopter transmission system failures, and traditional named entity recognition methods are difficult to accurately identify domain-specific entities. Rule-based methods require a large number of manually constructed rules, which are costly and have limited coverage; although statistical learning-based methods can automatically learn features, they perform poorly when faced with complex contextual dependencies and long-distance semantic associations. (2) The relationships between entities in the field of helicopter transmission system failures are complex and diverse, including causal relationships, inclusion relationships, influence relationships, and other types. Traditional relationship extraction methods mainly rely on predefined templates or manually annotated training data, and have problems with poor generalization ability and poor adaptability to new fields, especially when dealing with long-range dependencies in long texts. (3) Most existing knowledge graph construction methods use independent entity recognition and relationship extraction processes, and lack an effective coordination mechanism between the two tasks. Errors in entity recognition will directly affect the effect of relationship extraction, and relationship information has not been effectively fed back to the entity recognition process, resulting in poor overall construction results. Summary of the Invention

[0004] The purpose of the present invention is to provide a helicopter transmission system fault information extraction method based on RL-GPNet to solve the problems raised in the background technology.

[0005] To solve the above technical problems, the present invention adopts the following technical solution: a method for extracting fault information of a helicopter transmission system based on RL-GPNet, comprising:

[0006] Obtain fault data and construct datasets for entity recognition and relationship extraction tasks; preprocess the original text and entity types to obtain word vector sequences and entity type embedding vectors;

[0007] Encoding based on BERT: Obtain the hidden state of the text and perform feature fusion with the entity type embedding vector;

[0008] Decoding based on a global pointer network: using pointer pairs to predict whether a text segment is an entity, and calculating the entity recognition probability of each span belonging to each entity type; for the identified entity pairs, using pointer pairs to predict the relationship type between the entity pairs;

[0009] An optimization layer based on reinforcement learning is introduced to automatically optimize and update the decoding dynamically through trial and error learning; and the PPO algorithm is combined to limit the update amplitude.

[0010] Preferably, the helicopter transmission system component fault data is obtained from structured and semi-structured data, wherein the entities include fault system, fault phenomenon, fault cause, fault location and solution; the relationships between the entities include composition, occurrence, solution, cause and unknown relationships.

[0011] Preferably, the preprocessing is specifically:

[0012] The original text is segmented and converted into subword units that can be processed by BERT;

[0013] Use the pre-trained BERT model to deeply encode the text and generate context-related word vectors;

[0014] Load a predefined set of entity type labels and convert them into dense vectors.

[0015] Preferably, during the encoding process, the hidden state H of the text is obtained through BERT; feature fusion is performed based on the hidden state of the text:

[0016] F = Fusion(H,E)= H⊕Repeat(E,n);

[0017] Among them, F represents the fused features; Fusion represents a feature fusion function with inputs H and E; ⊕ represents the splicing operation, Repeat(E,n) repeats the entity type embedding n times to match the sequence length; E is the entity type embedding vector.

[0018] Preferably, an attention mechanism is constructed to enhance the interaction between entity type information and text representation and context-aware enhanced text representation matrix:

[0019] ;

[0020] Among them, MultiHead represents the multi-head attention function; Q, K, V are attention weights, W Q , W K , W V , W O is the learning parameter matrix, h is the number of attention heads, is a context-aware representation, n Indicates that the input sequence token the number of dc Represents the feature dimension of the final output of multi-head attention, Indicates that C is a matrix with n rows and dc columns; Concat represents a concatenation operation, which concatenates the outputs of multiple attention heads according to the feature dimension; Indicates the h The output of an attention head.

[0021] Preferably, during the entity prediction process:

[0022] Directly predict span-level entities through pointer pairs (i, j), where i≤j represents the text segment from position i to position j, and calculate the entity recognition pointer score :

[0023] ;

[0024] Among them, C i and C j is the context representation vector of the i-th and j-th token in the sequence; Feature weights for learning to identify the starting position of an entity; Feature weights for learning to identify the end position of an entity; is a bias term that adjusts the overall score baseline;

[0025] Calculate entity recognition probability based on entity recognition pointer score:

[0026] ;

[0027] Where entity represents the recognized entity; k is the entity type; and softmax represents the activation function, which is used to convert a real number vector into a probability distribution.

[0028] Preferably, in the process of predicting the relationship type between entity pairs:

[0029] Calculate the relationship extraction pointer score :

[0030] ;

[0031] Among them, (m,n) and (k,l) are the starting and ending positions of the head and tail entities respectively; C m and C n are the boundary information of the head entity, which is used to capture the complete semantics of the head entity; C k and C l It is the boundary information of the tail entity, which is used to capture the complete semantics of the tail entity;

[0032] Relationship extraction pointer score Relationship extraction pointer score Relationship recognition probability :

[0033] ;

[0034] In the formula, relation represents the extracted relationship; r represents the relationship type.

[0035] Preferably, the dynamic optimization update includes:

[0036] Construct a vector containing multi-dimensional information, including the context of the input text, the set of entities currently recognized, and the set of relations currently extracted;

[0037] Build a policy network to learn the strategy for selecting the optimal action in a given state, and design a value network to evaluate the value of the current state; and

[0038] Update the policy network based on the CLIP loss function in the PPO algorithm;

[0039] Update the value network by minimizing the value function loss;

[0040] Design a joint loss function and update the global pointer network based on the joint loss function. The joint loss function is:

[0041] ;

[0042] in, , is the weight coefficient; is the entity recognition loss, P represents the probability of entity recognition; is the true label. If span (i,j) is indeed an entity of type k, then =1, otherwise 0; , is the relationship extraction loss, is the true relation label, if the entity pair ( , ) is 1 if a relation of type r does exist.

[0043] Preferably, entities are determined by maximizing entity boundary prediction probabilities, and based on the identified entity pairs, relationships between entities are determined by maximizing relationship prediction probabilities;

[0044] Construct a joint decoding strategy to obtain the optimal entity-relationship joint extraction result and calculate the total loss function. The joint decoding strategy is:

[0045] ;

[0046] Where: E * represents the optimal entity set; R * represents the optimal relation set; x represents the input text sequence; Represents the conditional probability of entity extraction; Represents the probability of relation extraction under a given entity set; E represents a candidate entity set identified by the model in the input text sequence x; R represents a candidate relation set identified by the model based on the identified entity set E; Represents the product of all elements, which is used to calculate the joint probability obtained by multiplying the conditional probability and the relation extraction probability; arg max represents the parameters E and R when the joint probability is maximized.

[0047] Preferably, the joint loss function and reinforcement learning loss function Linear combination forms the total loss function :

[0048] ;

[0049] in: and are the weight coefficients of the two losses, which are used to balance the contribution of supervised learning and reinforcement learning;

[0050] Joint update strategy: Based on standard gradient descent, after the kth iteration, L total right The gradient of L is contributed by the joint extraction loss part and the reinforcement learning strategy and entropy loss part; total right The gradient comes only from the loss part of the value function in reinforcement learning;

[0051] Policy network parameter update:

[0052]

[0053] in: represents the value network parameter of the k-th step, is the learning rate of the value network, is the total loss of the model, Represents the total loss function gradient.

[0054] This collaborative training mechanism enables the model to use labeled data to ensure the lower limit of extraction accuracy, while using the exploration ability of reinforcement learning to break through the bottleneck of supervised learning, thereby achieving higher robustness and performance ceiling in complex extraction tasks.

[0055] Beneficial effects: The global pointer network architecture designed by the present invention has strong global optimization capabilities, can capture long-range semantic dependencies in text, achieve joint optimization of entity recognition and relationship extraction, greatly alleviate the error propagation problem, and by combining the reinforcement learning mechanism, the model can automatically optimize the extraction strategy through trial and error learning, reduce dependence on manually labeled data, and improve the automation level of knowledge graph construction;

[0056] In addition, the PPO algorithm designed in the present invention avoids drastic changes in strategy parameters by limiting the strategy update amplitude. In the joint extraction task, it ensures the smooth optimization of entity recognition and relationship classification strategies and avoids oscillations during the training process. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0058] In the attached figure:

[0059] Figure 1 This is a fault information relationship category diagram of the present invention;

[0060] Figure 2 The overall flow chart of the RL-GPNet model provided by the present invention;

[0061] Figure 3 The global pointer network decoding flow chart provided by the present invention;

[0062] Figure 4 Flowchart of the PPO optimization algorithm provided by the present invention;

[0063] Figure 5 Provides a comparison chart of evaluation indicators of different entity relationship joint extraction models for the present invention;

[0064] Figure 6 This is a partial annotation diagram of the gear structure fault information of the entity dataset of the present invention. DETAILED DESCRIPTION

[0065] To make the objects and advantages of the present invention more clearly understood, the present invention is described in detail below with reference to embodiments. It should be understood that the following text is merely intended to describe a helicopter transmission system fault information extraction method based on RL-GPNet or several specific implementations of the present invention, and does not strictly limit the scope of protection specifically claimed in the present invention.

[0066] like Figures 1-4 As shown in the figure, a helicopter transmission system fault information extraction method based on RL-GPNet is proposed. An entity recognition module based on the Global Pointer network is constructed. Through global normalization and relative position encoding mechanisms, the long-distance dependencies and complex semantic information in the fault description text are effectively captured. At the same time, a reinforcement learning framework based on the Proximal Policy Optimization (PPO) algorithm is designed to model triple generation as a multi-step decision task. The entity recognition and relationship classification strategies are trained collaboratively to alleviate the error accumulation caused by task conflicts. The specific steps are as follows:

[0067] Step 1: Data acquisition and preprocessing:

[0068] 1.1: Obtain helicopter transmission system component failure data from structured and semi-structured data to construct a helicopter transmission system failure information dataset;

[0069] The entities in the data are divided into five categories: fault system, fault phenomenon, fault cause, fault location and solution; and BMES is used to annotate them, as shown in Table 1:

[0070] Table 1 Data annotation description of entity recognition task

[0071]

[0072] Finally, the processed annotated corpus is formed into a dataset consisting of more than 10,000 rows.

[0073] Taking the helicopter transmission system fault entity as an example, the relationship between each entity category is established as follows: Figure 1 As shown;

[0074] The relationships between entities are defined into three categories: composition, occurrence, resolution, cause, and unknown. A composition relationship indicates that the head entity can be composed of the tail entity, an occurrence relationship indicates that the head entity can occur in the tail entity, a resolution relationship indicates that the head entity can be resolved by the tail entity, a cause relationship indicates that the head entity is caused by the tail entity, and an unknown relationship indicates that the relationship between two entities is unclear and needs to be discovered. The three relationships and their descriptions are shown in Table 2:

[0075] Table 2 Relationship annotation

[0076]

[0077] 1.2: Tokenize the original text and convert it into subword units that can be processed by BERT. Use the pre-trained BERT model to deeply encode the text after word segmentation and generate context-related word vector representations:

[0078] ;

[0079] Where: T is the original input text; X is the token sequence after word segmentation; n is the sequence length.

[0080] 1.3: Load a predefined set of entity type labels and convert them into dense vector representations for easy fusion with text features.

[0081] Ttypes={t1,t2,...,t m};

[0082] Where m is the number of entity types;

[0083] Step 2: Reference Figure 3 As shown, based on BERT encoding:

[0084] 2.1: Obtaining hidden state representation of text through BERT:

[0085] H=BERT(X)={h1,h2,...,h n};

[0086] in: is the BERT hidden state matrix; d h is the BERT hidden layer dimension (usually 768 or 1024);

[0087] 2.2: Convert entity types to embedded vector representations:

[0088] E=EntityEmbedding(Ttypes)={e1,e2, ...,e m};

[0089] in: is the entity type embedding matrix; d e is the entity type embedding dimension;

[0090] 2.3: Feature fusion: Fusing BERT-encoded text information with entity type embedding information:

[0091] F=Fusion(H,E) =H⊕Repeat(E,n);

[0092] Where: ⊕ represents the concatenation operation; Repeat(E,n) repeats the entity type embedding n times to match the sequence length; ;

[0093] 2.4: Cross-Attention Mechanism: The attention mechanism is used to enhance the interaction and context-aware representation between entity type information and text representation, generating enhanced text representation that incorporates entity type information, providing richer features for subsequent extraction;

[0094] 2.41: Attention weight calculation:

[0095] ; ; ;

[0096] Where: F is the feature representation of the input; W Q is the query weight matrix; W K is the bond weight matrix; W V is the value weight matrix; n is the sequence length; d k Dimensions of query and key; d v is the dimension of the value.

[0097] 2.42: Scaled Dot Product Attention:

[0098] ;

[0099] Where: d k is the dimension of the key vector, used as a scaling factor to prevent gradient vanishing.

[0100] 2.43: Cross-Attention Output:

[0101] ;

[0102] Where: W Q , W K , W V , W O is the learning parameter matrix; h is the number of attention heads; is a context-aware representation, n Indicates that the input sequence token the number of dc Represents the feature dimension of the final output of multi-head attention, Indicates the h The output of an attention head;

[0103] Step 3: Decoding based on global pointer network:

[0104] 3.1: Entity Recognition, including:

[0105] 3.11: Entity Recognition Pointer Score: Directly predict span-level entities through pointer pairs (i, j), where i≤j represents the text segment from position i to position j:

[0106] ;

[0107] Where: C i and Cj is the context representation vector of the i-th and j-th token in the sequence; Learn feature weights to identify the starting position of an entity; Learn feature weights to identify where entities end;

[0108] The semantic information of the start and end positions is fused through linear transformation. Bent is a bias term that adjusts the overall score baseline.

[0109] 3.12: Entity Recognition Probability:

[0110] ;

[0111] Where k represents the entity type; for each possible span (i, j), the probability of it belonging to each entity type is calculated. k traverses all predefined entity types, including a "non-entity" type to filter invalid spans.

[0112] 3.2: Relationship extraction, including:

[0113] 3.21: Relationship Extraction Pointer Score:

[0114] ;

[0115] Where (m,n) and (k,l) are the start and end positions of the head and tail entities respectively; the context vectors of the four positions are concatenated: [C m ;C n ;C k ;C l ], and process this high-dimensional concatenation vector through a multi-layer perceptron (MLP). m and C n is the boundary information of the head entity, which is used to capture the complete semantics of the head entity; C k and C l It is the boundary information of the tail entity, which is used to capture the complete semantics of the tail entity. The combination of the four vectors can model the positional relationship and semantic interaction between entity pairs;

[0116] 3.22: Relationship recognition probability:

[0117] ;

[0118] Where r represents the relationship type; for each entity pair ( , ) predict the relationship type between them, r traverses all predefined relationship types, which also include the "no relationship" type;

[0119] 3.3: Joint loss function, design multi-task learning loss function, including:

[0120] 3.31 Entity Recognition Loss:

[0121] ;

[0122] Where: is the true label. If span (i,j) is indeed an entity of type k, then =1, otherwise 0, is the probability predicted by the model, only when the true label is 1, corresponding The term will contribute to the loss, and the negative sign makes the higher the prediction probability, the smaller the loss. This loss function punishes the model for incorrect predictions of real entities and encourages the model to give a high probability to the correct entity span.

[0123] 3.32: Relation extraction loss:

[0124] ;

[0125] is the true relation label, if the entity pair ( , ) If there is indeed a relationship of type r, then it is 1. The relationship label is meaningful only when both spans are real entities, and some entity pairs are in a "no relationship" state.

[0126] 3.33: Joint extraction loss:

[0127] ;

[0128] Where: α and β are weight coefficients;

[0129] Step 4: Reference Figure 4 As shown, the global pointer network decoding combined with reinforcement learning optimization:

[0130] 4.1: Reinforcement Learning Framework. To improve the sequential decision-making capabilities of knowledge graph construction, an optimization layer based on reinforcement learning is introduced. This layer dynamically adjusts entity recognition and relationship extraction strategies through interactive learning between the agent and the environment, thereby achieving more accurate knowledge extraction in complex text scenarios.

[0131] 4.11: State Space Design,In the reinforcement learning framework, state representation is a key element,of the decision-making process.

[0132] Contextual representation of the input text:

[0133] ;

[0134] The current set of recognized entities:

[0135] ;

[0136] The currently extracted relationship set:

[0137] ;

[0138] Define the state at time t as a vector containing multi-dimensional information;

[0139] ;

[0140] Where: Indicates the text position of the current extraction operation, which is used to guide subsequent sequence decisions

[0141] 4.12: Policy Network:

[0142] ;

[0143] Where: The network parameter set is defined as θ={W π , W s , b s , b π} is the policy network parameter, W s and b s Construct a state encoding layer to map the input state to the hidden representation space, W π and b π It forms the strategy output layer and generates the logit value of each action. The ReLU activation function enhances the nonlinear expression ability of the network. The Softmax function ensures that the output is a valid probability distribution.

[0144] 4.13: Value Network Design:

[0145] ;

[0146] Where: The value network parameter set is φ={W v ,W s ,b s ,b v}, the policy network and the value network share the state encoding layer parameters, W s and b s ;

[0147] 4.2: PPO optimization algorithm, including:

[0148] 4.21: Importance sampling, in order to effectively utilize historical experience data, the importance sampling technique is used to calculate the strategy update ratio:

[0149] ;

[0150] This ratio reflects the probability change of the new strategy relative to the old strategy for a specific state-action pair and is a key quantity for controlling the update step size in the PPO algorithm.

[0151] 4.22: Advantage function estimation (GAE), using the generalized advantage estimation (GAE) method to calculate the advantage function:

[0152] ;

[0153] The timing differential error is defined as:

[0154] ;

[0155] Where: Parameter is the discount factor, is a GAE parameter used to balance bias and variance.

[0156] 4.23: PPO objective function construction: The core objective function of the PPO algorithm adopts truncated importance sampling:

[0157] ;

[0158] Where: To truncate parameters and prevent the policy update from being too aggressive.

[0159] The value function loss is defined as: ;

[0160] Where: is the target value, usually estimated by GAE.

[0161] In order to promote exploration, an entropy regularization term is introduced: ;

[0162] Where: Indicates that the policy is in state The entropy value below.

[0163] Combining the above, the loss function is expressed as:

[0164] ;

[0165] Step 5: Joint training optimization:

[0166] 5.1: Gradient update strategy. During the joint training process, a multi-network collaborative update method is adopted to optimize the parameters of the policy network, value network, and global pointer network respectively. The specific gradient update rules are as follows:

[0167] Policy network update: The parameter update of the policy network adopts the CLIP loss function in the proximal policy optimization (PPO) algorithm. The update formula is: ;

[0168] Where: represents the policy network parameters of step t; is the learning rate of the policy network; is the CLIP loss function.

[0169] Value network update: The value network updates its parameters by minimizing the value function loss:

[0170] ;

[0171] Where: represents the value network parameter of step t; is the learning rate of the value network; is the value function loss.

[0172] Global pointer network update: The global pointer network uses a joint loss function to update parameters:

[0173] ;

[0174] Where: W t Represents the global pointer network parameters of step t; is the learning rate of the global pointer network; is the joint loss function.

[0175] 5.2: Reward function design, including:

[0176] 5.21: Instant reward mechanism. In order to effectively guide the learning process of the model, an instant reward function is designed that includes entity recognition reward, relation extraction reward and penalty term:

[0177] ;

[0178] In the formula, the specific definitions of each item are as follows:

[0179] Entity Recognition Rewards: ;

[0180] Relationship extraction rewards: ;

[0181] Prediction error penalty: ;

[0182] 5.22: Cumulative reward calculation, taking into account the impact of long-term benefits on decision-making, adopts a discounted cumulative reward mechanism:

[0183] ;

[0184] Where: is the discount factor; T is the total length of the sequence.

[0185] 5.3: Reasoning stage, including

[0186] 5.31: Entity Extraction,In the reasoning stage, entities are determined by maximizing the predicted probability of,entity boundaries.

[0187] Entity Boundary Prediction:

[0188] ;

[0189] Where i and j represent the starting and ending positions of the entity respectively; k represents the entity type.

[0190] 5.32: Relation extraction, based on the identified entity pairs, determines the relationship between entities by maximizing the relationship prediction probability.

[0191] Relationship Prediction: ;

[0192] Where: head and tail are the recognized entity pairs.

[0193] 5.4: Joint decoding strategy, including:

[0194] 5.41 Global Optimal Solution,In order to obtain the global optimal entity-relationship joint extraction result, the joint decoding strategy is:

[0195] ;

[0196] Where: E * represents the optimal entity set; R * represents the optimal relation set; x represents the input text sequence; Represents the conditional probability of entity extraction; It represents the probability of relation extraction under a given entity set, E represents a candidate entity set identified by the model in the input text sequence x; R represents a candidate relation set identified by the model based on the identified entity set E.

[0197] The total loss function of training is composed of the joint loss function and reinforcement learning loss function The linear combination forms the total loss function: .

[0198] Where: and These are weight coefficients for the two losses, respectively, used to balance the contributions of supervised learning and reinforcement learning. By adjusting these two hyperparameters, we can effectively control the balance between model accuracy and robustness. Ljoint is the global pointer extraction loss, and LRL is the reinforcement learning loss.

[0199] Using standard gradient descent as the criterion, after the kth iteration, L total right The gradient of L is contributed by the joint extraction loss part and the reinforcement learning strategy and entropy loss part; total right The gradient comes only from the loss part of the value function in reinforcement learning;

[0200] Policy network parameter update:

[0201]

[0202] in: represents the value network parameter of the k-th step, is the learning rate of the value network, is the total loss of the model, Represents the total loss function gradient;

[0203] This collaborative training mechanism enables the model to use labeled data to ensure the lower limit of extraction accuracy, while using the exploration ability of reinforcement learning to break through the bottleneck of supervised learning, thereby achieving higher robustness and performance ceiling in complex extraction tasks.

[0204] In a specific embodiment, the helicopter transmission system fault information extraction method based on RL-GPNet extracts the fault knowledge of a helicopter transmission system; taking "the gear of the reducer is glued due to insufficient heat dissipation in the plateau environment, and the transmission efficiency is improved by 15% after adding a forced air cooling system" as an example, sequence annotation is performed as follows: Figure 6 As shown;

[0205] The data format of the annotation examples is shown in Table 3, including the head entity, tail entity, the relationship between entities, and the sentence text.

[0206] Table 3 Partial annotation of relational dataset

[0207]

[0208] Precision, recall, and F1 are used as evaluation indicators, and the calculation formulas are as follows:

[0209] ;

[0210] ;

[0211] ;

[0212] In the formula, TP represents the number of positive examples (correct entities or relations) correctly predicted by the model, that is, the number of entities or relations that actually exist that are correctly identified by the model; FP represents the number of positive examples incorrectly predicted by the model, that is, the number of cases where non-existent entities or relations are incorrectly identified as existing; FN represents the number of negative examples (unidentified entities or relations) incorrectly predicted by the model, that is, the number of entities or relations that actually exist that are not correctly identified by the model.

[0213] Table 4 Results of the present invention and other baseline models on the dataset

[0214]

[0215] Table 4 above lists the results of the present invention and other baseline models on the dataset. Figure 5 It can be observed that the proposed model outperforms the comparison models in almost all evaluation metrics, and even outperforms the recent TPLinker baseline model. Furthermore, the proposed method demonstrates good generalization ability when dealing with helicopter transmission system faults involving long texts with overlapping entities and extensive technical terminology, providing effective technical support for intelligent fault diagnosis of helicopter transmission systems. This research has important theoretical significance and application value for improving the intelligent level of helicopter equipment maintenance and support.

[0216] The above describes the embodiments of the present invention in detail with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. After knowing the contents described in the present invention, ordinary technicians in this technical field can make several equivalent changes and substitutions without departing from the principles of the present invention. These equivalent changes and substitutions should also be regarded as falling within the scope of protection of the present invention.

Claims

1. A helicopter transmission system fault information extraction method based on RL-GPNet, characterized in that: include: Obtain fault data and construct datasets for entity recognition and relationship extraction tasks; Preprocess the original text and entity types to obtain word vector sequences and entity type embedding vectors; Encoding based on BERT: Obtain the hidden state of the text and perform feature fusion with the entity type embedding vector; Decoding based on a global pointer network: using pointer pairs to predict whether a text segment is an entity, and calculating the entity recognition probability of each span belonging to each entity type; for the identified entity pairs, using pointer pairs to predict the relationship type between the entity pairs; Introducing a reinforcement learning-based optimization layer to automatically and dynamically optimize and update the decoding through trial-and-error learning; And combined with the PPO algorithm to limit the update range; Dynamic Optimization updates include: Construct a vector containing multi-dimensional information, including the context of the input text, the set of entities currently recognized, and the set of relations currently extracted; Build a policy network to learn the strategy for selecting the optimal action in a given state, and design a value network to evaluate the value of the current state; and Update the policy network based on the CLIP loss function in the PPO algorithm; Update the value network by minimizing the value function loss; Design a joint loss function and update the global pointer network based on the joint loss function. The joint loss function is: ； in, , is the weight coefficient; is the entity recognition loss, P represents the probability of entity recognition, entity represents the recognized entity; k is the entity type; is the true label. If span(i,j) is indeed an entity of type k, then =1, otherwise 0; , is the relationship extraction loss, is the true relation label, if the entity pair ( , ) If a relation of type r does exist, then it is 1, relation represents the extracted relation, and r represents the relation type.

2. The method for extracting helicopter transmission system fault information based on RL-GPNet according to claim 1, characterized in that: Fault data of helicopter transmission system components are obtained from structured and semi-structured data. The entities include fault system, fault phenomenon, fault cause, fault location and solution. The relationships between the entities include composition, occurrence, solution, cause and unknown relationships.

3. The method for extracting helicopter transmission system fault information based on RL-GPNet according to claim 1 or 2, characterized in that: The pre-processing is specifically as follows: The original text is segmented and converted into subword units that BERT can process; Use the pre-trained BERT model to deeply encode the text and generate context-related word vectors; Load a predefined set of entity type labels and convert them into dense vectors.

4. The method for extracting helicopter transmission system fault information based on RL-GPNet according to claim 3 is characterized in that: During the encoding process, BERT is used to obtain the hidden state H of the text; feature fusion is performed based on the hidden state of the text: F = Fusion(H,E)= H⊕Repeat(E,n); Among them, F represents the fused features; Fusion represents a feature fusion function with inputs H and E; ⊕ represents the splicing operation, Repeat(E,n) repeats the entity type embedding n times to match the sequence length; E is the entity type embedding vector.

5. The method for extracting helicopter transmission system fault information based on RL-GPNet according to claim 4, characterized in that: Construct an attention mechanism to enhance the interaction between entity type information and text representation and context-aware enhanced text representation matrix: ； Among them, MultiHead represents the multi-head attention function; Q, K, V are attention weights, W Q , W K , W V , W O is the learning parameter matrix, h is the number of attention heads, is a context-aware representation, n Indicates that the input sequence token the number of dc Represents the feature dimension of the final output of multi-head attention, Indicates that C is a matrix with n rows and dc columns; Concat represents a concatenation operation, which concatenates the outputs of multiple attention heads according to the feature dimension; Indicates the h The output of an attention head.

6. The method for extracting helicopter transmission system fault information based on RL-GPNet according to claim 1, characterized in that: During entity prediction: Directly predict span-level entities through pointer pairs (i, j), where i≤j represents the text segment from position i to position j, and calculate the entity recognition pointer score : ； Among them, C i and C j is the context representation vector of the i-th and j-th token in the sequence; Feature weights for learning to identify the starting position of an entity; Feature weights for learning to identify the end position of an entity; is a bias term that adjusts the overall score baseline; Calculate entity recognition probability based on entity recognition pointer score : ； Where entity represents the recognized entity; k is the entity type; and softmax represents the activation function, which is used to convert a real number vector into a probability distribution.

7. The method for extracting helicopter transmission system fault information based on RL-GPNet according to claim 6, characterized in that: In the process of predicting the relationship type between entity pairs: Calculate the relationship extraction pointer score : ； Among them, (m,n) and (k,l) are the starting and ending positions of the head and tail entities respectively, C m and C n are the boundary information of the head entity, which is used to capture the complete semantics of the head entity; C k and C l It is the boundary information of the tail entity, which is used to capture the complete semantics of the tail entity; Relationship extraction pointer score Relationship extraction pointer score Relationship recognition probability : ； In the formula, relation represents the extracted relationship; r represents the relationship type.

8. A helicopter transmission system fault information extraction method based on RL-GPNet according to any one of claims 4 to 7, characterized in that: Entities are determined by maximizing the predicted probability of entity boundaries, and based on the identified entity pairs, the relationships between entities are determined by maximizing the predicted probability of relationships; Construct a joint decoding strategy to obtain the optimal entity-relationship joint extraction result and calculate the total loss function. The joint decoding strategy is: ； Where: E * represents the optimal entity set; R * represents the optimal relation set; x represents the input text sequence; Represents the conditional probability of entity extraction; represents the probability of relation extraction under a given entity set; E represents a candidate entity set identified by the model in the input text sequence x; R represents a candidate relation set identified by the model based on the identified entity set E; ∏ represents the product of all elements, which is used to calculate the joint probability obtained by multiplying the conditional probability and the relation extraction probability; arg max represents the parameters E and R that maximize the joint probability.

9. The method for extracting helicopter transmission system fault information based on RL-GPNet according to claim 8, characterized in that: By the joint loss function and reinforcement learning loss function Linear combination forms the total loss function : ； in: and are the weight coefficients of the two losses, which are used to balance the contribution of supervised learning and reinforcement learning; Joint update strategy: Based on standard gradient descent, k After iterations, L total right The gradient of is contributed by the joint extraction loss part and the reinforcement learning strategy and entropy loss part; L total right The gradient comes only from the loss part of the value function in reinforcement learning; Policy network parameter update: ； in: Indicates the k step value network parameters, is the learning rate of the value network, is the total loss of the model, Represents the total loss function gradient.

Images