A hard-facing welding process parameter recommendation method, system and medium

Abstract

The application provides a hard-facing process parameter recommendation method, system and medium, comprising building a hard-facing process knowledge graph, offline pre-storing the metallurgical compatibility basic score of the base material and the hard-facing material under the reference parameter; pruning the original knowledge graph according to the query semantics to generate a reasoning subgraph facing a specified task; embedding the subgraph into a low dimension by means of a graph attention network and carrying out multi-hop path searching, and deleting or raising the path cost by the compatibility score below the threshold; the attention coefficient is jointly regulated by the path aggregation feature and the uncertainty; a parameterized fuzzy interval model is built for the numerical constraint nodes in the path, and the membership interval boundary is fine-tuned by the upstream process vector; after the candidate process path is encoded, it is sent into a path evaluation network to output a performance compliance confidence score, and a parameter stability score is obtained through a Monte Carlo parameter disturbance evaluation unit, and the two scores are combined for weighted sorting to optimize the optimal hard-facing process parameter.

Description

Technical Field

[0001] This application belongs to the field of parameter recommendation, and in particular relates to a method, system and medium for recommending hardface welding process parameters. Background Technology

[0002] Hardfacing, as a surface modification and repair technology, is applied in aerospace, petrochemical, metallurgical manufacturing, and heavy machinery industries. This technology deposits alloy materials with specific properties onto the surface of the workpiece base material, improving the material's wear resistance, corrosion resistance, and high-temperature resistance, thereby extending the service life of core components. Hardfacing is a multidisciplinary physicochemical process; the forming quality and performance depend on the matching relationship between the base material and the welding material, as well as the coordinated control of multiple process parameters such as welding current, voltage, welding speed, and preheating temperature. These parameters exhibit interrelationships. In traditional industrial production, process parameters are determined based on the personal experience of senior engineers or require extensive, time-consuming, and costly trial-and-error experiments. This traditional experience-guided model is no longer sufficient to provide optimal process parameter solutions for new material combinations or specific processing conditions. Therefore, applying artificial intelligence and knowledge engineering to process planning in the manufacturing field has become a feasible direction to improve the efficiency of process parameter recommendation. However, conventional data-driven models often lack interpretability and cannot incorporate underlying metallurgical mechanisms as constraints into the reasoning process, leading to recommended parameters that easily cause porosity or crack defects in practical applications. Faced with knowledge networks containing material properties and process specifications, existing recommendation methods mostly employ static global retrieval, lacking the ability to reconstruct and focus based on specific contextual semantics. Furthermore, they cannot adequately handle the ambiguity and fluctuation ranges commonly found in parameter numerical constraints in actual processes, failing to adapt to ever-changing task requirements. Industrial environments are complex, and even minor deviations in process parameters can cause drastic quality changes. Existing technologies only focus on the matching degree of theoretical predictions, neglecting the process's resilience to interference during actual execution, and failing to balance the reliability of achieving expected performance with the stability of the parameters themselves. Therefore, there is an urgent need for a recommendation system that can integrate physical mechanism constraints, achieve intelligent knowledge reasoning, and comprehensively weigh prediction confidence and process stability to overcome application limitations. Summary of the Invention

[0003] This invention proposes a method for recommending hardfacing welding process parameters to address the problem that existing technologies lack a recommendation system capable of integrating physical mechanism constraints, realizing knowledge-based intelligent reasoning, and comprehensively balancing prediction confidence and process stability. The method includes: Construct a knowledge graph of hardfacing welding process, and calculate offline the basic metallurgical compatibility score between the cached workpiece base material and the candidate welding material under the benchmark process parameters. For the reasoning request of hardfacing welding process parameters, combine the semantic information of the query context to prune and reconstruct the knowledge graph of hardfacing welding process to generate a reasoning subgraph that serves a specific task. A graph attention network is used to embed the inference subgraph into a low-dimensional vector space for multi-hop inference path search. During the multi-hop inference path search, the basic metallurgical compatibility score below a preset threshold is used as a constraint to prune the inference path or increase its cost. During the inference path expansion process, the attention coefficient is modulated by the aggregated information and uncertainty measure of the current inference path. A parameterized fuzzy interval model is established for the numerical constraint nodes in the inference path. The membership function boundary of the numerical constraint interval is adjusted using the upstream process parameter vector of the current inference path as input. The candidate process paths, including the adjusted numerical constraint interval boundaries, are encoded as a sequence vector. The sequence vector is then input into a multi-task path evaluation network, which outputs the confidence scores of the candidate process paths in achieving the expected performance indicators in parallel. At the same time, the stability scores of the process parameter combinations corresponding to the candidate process paths are obtained. Based on the confidence scores and the stability scores, the optimal process parameter combinations are selected as the recommended results for hardfacing welding process parameters.

[0004] Furthermore, the present invention also relates to a hardfacing welding process parameter recommendation system, comprising: The reconstruction module is used to construct a knowledge graph of hardfacing welding process, and to calculate offline the basic metallurgical compatibility score between the cached workpiece base material and the candidate welding material under the benchmark process parameters. For the reasoning request of hardfacing welding process parameters, the module combines the semantic information of the query context to prune and reconstruct the knowledge graph of hardfacing welding process to generate a reasoning subgraph that serves a specific task. The adjustment module is used to embed the inference subgraph into a low-dimensional vector space using a graph attention network to perform multi-hop inference path search. During the multi-hop inference path search, the basic metallurgical compatibility score below a preset threshold is used as a constraint to prune the inference path or increase its cost. During the inference path expansion process, the attention coefficient is modulated by the current inference path aggregation information and uncertainty measure. A parameterized fuzzy interval model is established for the numerical constraint nodes in the inference path. The membership function boundary of the numerical constraint interval is adjusted using the upstream process parameter vector of the current inference path as input. The filtering module is used to encode the candidate process paths containing the adjusted numerical constraint interval boundaries into a sequence vector, input the sequence vector into a multi-task path evaluation network, and output the confidence scores of the candidate process paths in achieving the expected performance indicators in parallel. At the same time, it obtains the stability scores of the process parameter combinations corresponding to the candidate process paths, and filters out the optimal process parameter combinations based on the weighted sorting of the confidence scores and the stability scores as the recommended results of hardfacing welding process parameters.

[0005] This invention mitigates material mismatch risks by constructing a knowledge graph of hardfacing welding processes and calculating basic metallurgical compatibility scores offline. For specific inference requests, the knowledge graph is pruned and reconstructed based on contextual semantics to generate task-specific inference subgraphs, reducing the search space. During multi-hop inference in the graph attention network, compatibility scores are used as constraints for pruning, and attention coefficients are modulated using uncertainty metrics. Parameterized fuzzy interval models are established for numerically constrained nodes, and boundaries are adjusted based on upstream parameters. A multi-task network is used to comprehensively evaluate the confidence level of expected performance indicators and the stability score of process parameter combinations, ensuring that the selected recommendations meet performance requirements while also considering engineering fault tolerance and operational stability, thus improving the intelligence level and production consistency of hardfacing welding process settings. Attached Figure Description

[0006] Figure 1 A flowchart for recommending process parameters for hardfacing welding; Figure 2 This is a schematic diagram illustrating the basic metallurgical compatibility scores between the parent material and the candidate material. Figure 3 This diagram illustrates the relationship between node correlation scoring and pruning effects. Detailed Implementation

[0007] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0008] Example 1 In Embodiment 1 of the present invention, as Figure 1 As shown, a method for recommending hardfacing welding process parameters includes: S1. Construct a knowledge graph of hardfacing welding process and calculate offline the basic metallurgical compatibility score between the cached workpiece base material and the candidate welding materials under the baseline process parameters.

[0009] Using a named entity recognition model with a bidirectional long short-term memory network (BSSN) and a conditional random field (CRF), entities representing workpiece base material, weld overlay material, welding current, welding voltage, shielding gas, and hardness are extracted from welding process specification text. A relation extraction model based on a converter-driven bidirectional encoder representation technique is then used to extract process combination relationships and performance influence relationships between entities. The extracted entities and relationships are persistently stored in the Neo4j graph database to construct an initial knowledge graph.

[0010] By calling the Thermo-Calc application interface of the phase diagram calculation library to read basic material property data, the dilution rate is first calculated based on the weld cross-sectional geometry parameters or a pre-calibrated heat input-deposition mapping model. This dilution rate is then used to correct the equivalent chemical composition of the workpiece base material and candidate weld overlay material in the interface region. Subsequently, based on the Shergulier solidification model, the solidification path, the proportion of favorable wear-resistant phases, and the risk of brittle intermetallic compound precipitation under baseline heat input conditions are predicted. The Shergulier solidification model is a model used to simulate the non-equilibrium solidification process of alloys, assuming that solutes do not diffuse in the solid phase and are completely mixed in the liquid phase, thereby predicting the solidification path, phase precipitation sequence, and compositional segregation trend. For variables such as dilution rate and the risk of brittle intermetallic compound precipitation, which are not suitable as direct positive indicators, they are first converted into dilution rate fit and brittle phase inhibition index, and then normalized and weighted summed with positive indicators such as the proportion of favorable wear-resistant phases and bonding strength assessment value to obtain the basic metallurgical compatibility score. The metallurgical compatibility scores of each base material and candidate weld overlay material pair are cached offline using the key-value in-memory database Redis.

[0011] Specifically, if weld overlay cross-section inspection data can be obtained, then based on the molten cross-sectional area of ​​the base material... and the cross-sectional area of ​​the welded metal Calculate dilution rate If cross-sectional inspection data is not yet available during the online recommendation phase, a pre-calibrated dilution rate estimation model is constructed based on welding current, arc voltage, welding speed, powder feed rate, or wire feed speed, and the estimated dilution rate is output. The equivalent composition of the interface is calculated based on the dilution rate D. The equivalent composition of this interface is then input into the phase diagram calculation model or the metallurgical phase prediction model. Thus, the Shergulier solidification model is used to predict solidification paths and phase precipitation trends, rather than directly calculating the dilution rate itself. This addresses the risk of brittle intermetallic compound precipitation. The brittle phase suppression index was obtained by using reverse normalization. This results in a higher risk of brittle phase precipitation and a lower B value; for dilution rate D, it is based on the preset target dilution rate. and allowable deviation Build dilution ratio fit The greater the deviation of the dilution rate from the target window, the lower the fit. The basic metallurgical compatibility score can be expressed as... ,in This is the normalized value for the proportion of the favorable phase. The normalized value for bonding strength is given, and B is the brittle phase suppression index. For dilution rate fit, the sum of all weights is 1.

[0012] In some embodiments, the offline calculation of the basic metallurgical compatibility score between the cached workpiece base material and the candidate weld overlay material under baseline process parameters includes: Extract the chemical element mass fraction list of the workpiece base material and the candidate weld overlay material, input the chemical element mass fraction list and the benchmark process parameters into a pre-trained metallurgical phase prediction model to predict the proportion of beneficial phases and the bonding strength evaluation value at the interface of the weld overlay layer, normalize the proportion of beneficial phases and the bonding strength evaluation value and calculate the basic metallurgical compatibility score by weighted summation, and store the basic metallurgical compatibility score in a static mapping table for online query and retrieval.

[0013] The pre-trained metallurgical phase prediction model employs a deep feedforward neural network. The network structure includes an input layer, multiple fully connected hidden layers, and an output layer. The input consists of a list of chemical element mass fractions of the workpiece base material and candidate weld overlay materials, as well as a comprehensive input tensor composed of baseline process parameters. The output consists of the predicted volume percentage of the favorable wear-resistant phase and the evaluated tensile strength of the interfacial metallurgical bond.

[0014] For specific workpiece base materials such as 45 steel and candidate surfacing materials such as high-chromium cast iron alloy powder, a list of chemical element mass fractions with preset dimensions is extracted, constructing a content vector of 20 core elements, such as carbon content 0.45%, chromium content 15.0%, and silicon content 1.2%. Simultaneously, a baseline process parameter vector containing physical values ​​is configured, for example, setting the welding current to 220A, the arc voltage to 26V, and the surfacing travel speed to 250mm / min. The chemical composition characteristics and the baseline process parameters are then combined to form a 43-dimensional comprehensive input tensor, which is fed into the metallurgical phase prediction model.

[0015] The forward inference process of this model can be represented as follows: Where X is a 43-dimensional integrated input tensor, Y is the output predicted value vector, W and b are the weight matrix and bias term of each hidden layer, respectively, and f is the activation function. The model predicts the state parameters in the micro-region of the interface based on multidimensional fitting. Specifically, this includes the predicted volume percentage of primary carbides or martensite, which is set between 10% and 65%, with an example output of 45%, and the evaluated tensile strength of the interfacial metallurgical bond, which is set between 300 MPa and 800 MPa, with an example output of 620 MPa.

[0016] The proportion of beneficial phases, bonding strength assessment values, and negative risk assessment values ​​are uniformly converted into scores ranging from 0 to 1 using pre-defined scoring rules. Specifically, for the proportion of beneficial phases and bonding strength assessment values, higher values ​​correspond to higher scores; for negative assessment values ​​such as brittle phase risk, crack sensitivity, or excessive dilution risk, higher risks correspond to lower risk suppression scores. After the score conversion, preset weights are assigned according to the importance of different indicators, and the scores are weighted and synthesized to obtain the basic metallurgical compatibility score. For example, the proportion of beneficial phases, bonding strength assessment values, and brittle phase risk can be converted into phase scores, strength scores, and risk suppression scores, respectively, and then synthesized according to preset weights to obtain the basic metallurgical compatibility score ranging from 0 to 1. This ensures that the more abundant the beneficial metallurgical factors, the higher the compatibility score; and the more obvious the negative risk, the lower the compatibility score. Quantitative scoring results for multiple sets of base material-weld alloy pairings are shown below. Figure 2 As shown.

[0017] The basic metallurgical compatibility scores formed by different combinations of parent materials and candidate materials are used as key-value pairs and persistently stored in a static mapping table of a high-speed in-memory database, thereby enabling low-latency state query calls with a time of less than 5ms during the online multi-hop inference stage.

[0018] S2, for the reasoning request of hardfacing welding process parameters, the hardfacing welding process knowledge graph is pruned and reconstructed in combination with the semantic information of the query context to generate a reasoning subgraph that serves a specific task.

[0019] The system receives a reasoning request text containing the target base material and desired wear resistance performance. It then uses the Sentence-BERT model to map the reasoning request text into a 768-dimensional query semantic vector. Next, it traverses the entity nodes and relation edges in the knowledge graph, calculating the cosine similarity between the query semantic vector and the text embedding vectors of each entity node. Nodes with a cosine similarity below 0.5 and their connected edges are marked as unrelated graph structures. Finally, it uses the subgraph extraction function provided by the NetworkX graph processing library to filter out nodes and edges marked as unrelated graph structures.

[0020] Meanwhile, the Personalized PageRank algorithm is used, with the source node in the inference request as the seed node, to calculate the restart random walk probability of all nodes in the graph, and retain the top 200 nodes with the highest access probability and their edges, thereby completing the pruning and reconstruction of the graph to generate an inference subgraph that serves a specific task.

[0021] In some embodiments, the step of pruning and reconstructing the hardface welding process knowledge graph by combining query context semantic information to generate a reasoning subgraph serving a specific task includes: The current query task expression is extracted into an intent sequence vector using a natural language processing model. The feature representations of entities in the hardface welding process knowledge graph are obtained. The intent sequence vector is converted into a single vector through pooling aggregation and projected to the same dimensional space as the feature representation. The inner product similarity is calculated with the feature representations of each entity to obtain the relevance score between each entity and the query task. Nodes and edges with relevance scores lower than a preset threshold are filtered out, and the remaining nodes and relationships are retained to form the reasoning subgraph.

[0022] The natural language processing model comprises a multi-layer bidirectional Transformer encoder, an average pooling layer, and a fully connected linear mapping layer. The input is a sequence of text describing the current query task, and the output is a projected global intent sequence vector.

[0023] Using a pre-trained BERT or RoBERTa model as a natural language processing model, the input current query task description, such as improving the impact and wear resistance of Q345 base material surface and requiring a hardness of not less than HRC55, is mapped into a hidden layer state sequence with a dimension of 768. The average pooling strategy is applied to aggregate and compress this sequence into a single global intent sequence vector with a dimension of 1×768.

[0024] A fully connected linear mapping layer with a weight matrix size of 768×256 is used to project the global intent sequence vector onto a 256-dimensional shared low-dimensional feature space that matches the entity embedding vectors of the hardfacing welding process knowledge graph. The calculation formula for this projection process is as follows: ,in Let W be the projected global intent vector, and W be the weight matrix of the fully connected layer with a size of 768×256. is a single global intent sequence vector after pooling with a size of 1×768, and b is a bias term with a size of 1×256. Therefore, the calculated result is... The size is 1×256, enabling similarity calculations in the same dimension as entity embedding vectors. If implemented using column vectors, it would be written as... The only difference between the two is the direction of matrix representation.

[0025] Within this 256-dimensional feature space, dot product operations are performed on the projected global intent vector and the embedded feature representations of each node in the knowledge graph, covering entity categories such as base material type, welding process method, and alloy powder composition, to calculate the inner product similarity. The calculated normalized inner product similarity is then mapped to the interval between 0 and 1 through a linear transformation and used as the relevance score. Before calculating the inner product similarity, the inner product similarity is... and L2 normalization was performed to reduce the impact of vector magnitude differences on the correlation score. The correlation score was calculated using the following formula: ,in Let be the embedding feature of the i-th entity. Thus, the cosine similarity is linearly mapped from [-1,1] to the [0,1] interval, which can be directly compared with the subsequent association threshold of 0.65.

[0026] A pre-defined threshold, preferably 0.65, is set between 0.60 and 0.75. The scoring list is traversed, and irrelevant nodes with a correlation score below 0.65 and their associated semantic edges are removed. For example, interfering nodes such as submerged arc welding and copper-based alloy powder with a score of only 0.15 are filtered out. Only the topology of highly correlated nodes that conform to the specific welding intention is retained, and the computationally reduced inference subgraph is output. The graph simplification effect after filtering nodes according to the correlation threshold is as follows: Figure 3 As shown. It should be noted that the cosine similarity threshold of 0.5 is used for coarse screening of semantically irrelevant nodes, the personalized PageRank is used to retain nodes with high topological proximity to the source node, and the projected relevance score threshold of 0.65 is used for fine screening of nodes after coarse screening; the above three screening conditions are performed in the order of semantic coarse screening, topological preservation, and relevance fine screening.

[0027] S3, a graph attention network is used to embed the inference subgraph into a low-dimensional vector space for multi-hop inference path search.

[0028] The graph deep learning library DGL is used to construct a multi-layer graph attention network (GAT) architecture. The node feature matrix and adjacency matrix of the inference subgraph are input into the graph attention network. The attention weights of neighboring nodes are calculated by linear transformation of the LeakyReLU activation function with a leaky linear rectified unit. The features of neighboring nodes are aggregated through a multi-head attention mechanism to generate a low-dimensional dense node embedding vector.

[0029] A reinforcement learning Markov decision process environment is constructed based on the generated node embedding vectors. The current state is initialized as the source node of the inference request, and the path search agent is trained using the Proximity Policy Optimization (PPO) algorithm. For multi-hop inference path search, a beam-search algorithm is employed, with a beam width of 5. In each inference step, the action value function score of the current node's transition to all valid neighbor nodes is calculated, and the top 5 branches with the highest scores are retained as candidate paths for the next hop to continue extending downwards. The beam-search algorithm is a heuristic path search algorithm that retains only the top few candidate branches with the highest scores at each step to approximately find a globally optimal path while reducing computational complexity.

[0030] In some embodiments, the step of embedding the inference subgraph into a low-dimensional vector space using a graph attention network for multi-hop inference path search includes: The initial features of nodes in the inference subgraph are mapped to hidden layer representations using a multilayer perceptron. During message passing between nodes, attention weights are calculated for the features of neighboring nodes and the current path state features. The weighted features of neighboring nodes are then fused with the features of the current node. The node vector representation is updated using an activation function and used as the current state input for multi-hop inference to the next-hop search network.

[0031] The graph attention network structure includes an attention computation head composed of a multilayer perceptron, a gated recurrent unit, and a feedforward neural network. The input is the initial features of each entity in the inference subgraph and the current exploration path, and the output is a new node vector representation after contextual semantic enhancement.

[0032] Extract initial statistical or categorical One-Hot features of 128 dimensions for each entity in the inference subgraph, and input them into a multilayer perceptron with a first layer output dimension of 256 and a second layer output dimension of 128. Then, perform feature processing using a LeakyReLU function with a slope of 0.2 to generate hidden layer feature representations for each node. The formula for calculating these hidden layer features is as follows: ,in Initial features of the nodes and This is the weight matrix. and For bias.

[0033] When performing multi-hop message passing in a graph attention network, a gated recurrent unit encodes the currently formed exploration path to obtain the current path state feature. This current path state feature is then concatenated with the node features of each candidate neighbor node to form a combined feature vector for the corresponding candidate neighbor node, which is input into the attention calculation head with shared parameters. The attention calculation head can be implemented using a single-layer or multi-layer feedforward neural network to output the raw matching score of each candidate neighbor node relative to the current path state. Subsequently, the raw matching scores of all candidate neighbor nodes under the same current node are normalized to obtain the attention weights corresponding to each candidate neighbor node. These attention weights are in the range of 0 to 1, and the sum of the attention weights of all candidate neighbor nodes under the same current node is 1. When a candidate neighbor node has a high matching degree with the current task semantics and the formed path, its attention weight increases accordingly; when the matching degree is low, its attention weight decreases accordingly. The network performs weighted aggregation of neighbor node features based on the attention weights, thereby updating the vector representation of the current node and using it as the state input for the next-hop inference search.

[0034] Based on the attention weight coefficients calculated above, the graph attention network performs a neighborhood feature aggregation operation. Specifically, the network weights and aggregates the feature information of neighboring nodes according to the attention weights corresponding to each first-order neighbor node, so that neighboring nodes more relevant to the current path and task semantics contribute more to the aggregation result. Subsequently, the aggregated neighborhood features are fused with the hidden layer representation of the current node itself. The fusion method can be vector concatenation, weighted addition, or feedforward neural network mapping. The fused features are then smoothly transformed by a non-linear activation function to generate a new node vector representation enhanced with contextual semantics. This new node vector representation is used to update the current inference state and serves as the input for subsequent multi-hop path search.

[0035] The enhanced vector will synchronously update the state memory of the hidden layer of the gated recurrent unit and be passed into the search network of the next hop as the input state variable of the multi-hop inference in step t+1. The maximum number of inference hops is set to 3 to 5 hops, preferably 4 hops.

[0036] S4, during the multi-hop inference path search, the basic metallurgical compatibility score below a preset threshold is used as a constraint to prune the inference path or increase its cost.

[0037] When the path search agent selects the next-hop candidate weld overlay material node, it simultaneously retrieves the basic metallurgical compatibility score between the workpiece parent material node and the candidate weld overlay material node in the current inference state from the key-value pair in-memory database Redis. A pre-set threshold is set to 0.6. When the retrieved basic metallurgical compatibility score is greater than or equal to this threshold, the transition probability of this action remains unchanged. When the score is lower than 0.6, a rule-based soft pruning mechanism is triggered.

[0038] A negative penalty term proportional to the difference between the threshold and the score is added to the agent's reward function as a cost penalty. This means that the priority of the path is reduced by subtracting a preset penalty coefficient from the original action value and multiplying it by the difference between the threshold and the score. This process can be represented as follows: ,in For the original action value, The corrected value of the action, where λ is the penalty coefficient. Let S be the compatibility threshold, and S be the basic metallurgical compatibility score. Therefore, the lower S is and the greater the difference from the threshold, the greater the penalty; when S is not lower than the threshold, the penalty is zero. If the score is extremely low and approaches zero, the candidate node is masked out of the action space of the cluster search to achieve pruning, preventing the path from extending to infeasible metallurgical combinations.

[0039] S5, during the inference path expansion process, the attention coefficient is modulated by the aggregated information of the current inference path and the uncertainty measure.

[0040] A gated recurrent unit (GRU) is used to encode and aggregate the sequence vectors of traversed path nodes step-by-step, outputting the hidden layer state vector as the aggregation information representation of the current inference path. Simultaneously, Monte Carlo Dropout is employed to perform ten forward propagation calculations during the network inference phase, and the variance of the output features from multiple forward propagations is used as a measure of cognitive uncertainty. Monte Carlo Dropout involves repeatedly enabling the Dropout random deactivation mechanism during the model inference phase, estimating model prediction uncertainty by assessing the volatility of multiple forward prediction results.

[0041] When calculating the attention coefficient of the next candidate node, multiple forward propagations are performed for the j-th candidate neighbor node, the variance of the transition score corresponding to the candidate neighbor node is calculated, and then normalized to obtain the node-level uncertainty. ∈[0,1]. Further calculate the uncertainty suppression coefficient. , where β is the preset modulation intensity and satisfies 0 ≤ β ≤ 1. The adjusted correlation score is Therefore, candidate neighbor nodes with higher uncertainty have lower adjusted relevance scores, thus reducing their attention coefficients after Softmax normalization. This method suppresses the attention coefficients of candidate nodes with high uncertainty or those that deviate significantly from the current historical path aggregation information, thereby modulating the expansion direction.

[0042] S6. Establish a parameterized fuzzy interval model for the numerical constraint nodes in the inference path, and adjust the membership function boundary of the numerical constraint interval using the upstream process parameter vector of the current inference path as input.

[0043] Numerical constraint nodes representing continuous variables such as welding current, welding speed, preheating temperature, or interpass temperature are identified in the inference subgraph, and a parameterized fuzzy interval model is constructed using a generalized bell-shaped membership function. The upstream determined feature vectors of the workpiece base material, the weld overlay material, and upstream process parameters such as interpass temperature are concatenated into a one-dimensional feature input fully connected feedforward neural network.

[0044] This feedforward neural network comprises a three-layer hidden layer structure and uses hyperbolic tangent Tanh as the activation function. The network output predicts and outputs three control parameters of the generalized bell-shaped membership function: the interval center value, the width parameter, and the slope parameter. By updating these three control parameters, the shape and left and right boundary points of the membership function are changed, thus narrowing or widening the allowable variation range of the welding process parameters according to different material combinations.

[0045] In some embodiments, adjusting the membership function boundary of the numerical constraint interval includes: Obtain the upstream process parameter feature latent vector that has been determined in the current inference path, input the upstream process parameter feature latent vector into the boundary bias prediction network to calculate the boundary offset and width scaling factor of the fuzzy interval membership function, and modify the center position and limit width of the pre-set initial membership function based on the boundary offset and the width scaling factor to generate the numerical constraint interval boundary.

[0046] The boundary offset prediction network consists of a three-layer fully connected feedforward neural network, with the terminal branches activated by the Tanh and Sigmoid functions, respectively. The input of the network is the total feature tensor formed by concatenating the feature latent vectors of the determined upstream process parameters through channels, and the output is a continuous real value, namely the boundary offset and the width scaling factor.

[0047] During the process of extending the multi-hop inference sequence to deeper nodes, the upstream process node features with determined values ​​are extracted from the context memory of the current inference search state. For example, when inferring to the weld overlay thickness node, the powder feeding rate and welding current have been confirmed in the previous sequence. The 64-dimensional local embedding features corresponding to the two nodes are extracted and channel spliced ​​to construct an upstream process parameter feature latent vector with a total dimension of 128.

[0048] The latent feature vector is input into the boundary offset prediction network, and the regression outputs two continuous real values, where the boundary offset prediction formula is: Where k represents the k-th type of numerical constraint parameter, This refers to the maximum permissible offset scale corresponding to this type of parameter. Specifically, it refers to the welding current. It can be set to 20A, corresponding to the arc voltage. It can be set to 3V, corresponding to the welding speed. The speed can be set to 50 mm / min, corresponding to the preheating temperature. It can be set to 30℃, corresponding to the thickness of the weld overlay. It can be set to 1mm. Therefore, the dimensionless translation trend output by the network can be converted into an actual boundary offset with units consistent with the target process parameters. The formula for predicting the width scaling factor is: ,in =0.5, =1.5, therefore The value ranges from 0.5 to 1.5. This coefficient represents the tolerance width adjustment ratio. The smaller the value, the narrower the corrected range and the stricter the process tolerance; The larger the value, the wider the corrected range and the broader the process tolerance. Features of the hidden layer of the network and This is the output layer weight matrix. and This is a bias term.

[0049] The geometric parameters of the pre-defined generalized bell-shaped membership function are corrected using algebraic operations. The formula for correcting the center position is as follows: The width correction formula is: If the original initial membership function sets the center position reference value for the weld overlay thickness parameter to 5.0 mm and the allowable total control width to 2.0 mm, and predicts the boundary offset of this thickness parameter to be -0.5 mm and the width scaling factor to be 0.8, then the new center position is corrected to 4.5 mm and the new allowable total control width to be 1.6 mm. Based on this, the legal range for this thickness parameter is defined as 3.7 mm to 5.3 mm.

[0050] S7. The candidate process path containing the adjusted numerical constraint interval boundary is encoded as a sequence vector. The sequence vector is input into the multi-task path evaluation network, and the confidence score of the candidate process path achieving the expected performance index is output in parallel. At the same time, the stability score of the process parameter combination corresponding to the candidate process path is obtained.

[0051] The classic Transformer-Encoder architecture in natural language processing is adopted to map the long sequence data consisting of the base material, welding material and the center values ​​of each process numerical constraint interval after boundary adjustment into a fixed-dimensional global sequence representation vector.

[0052] A path evaluation network consisting of a low-level parameter sharing layer and a multilayer perceptron branch is constructed using the PyTorch deep learning framework. The global sequence representation vector is input into the low-level parameter sharing layer for high-dimensional feature extraction. The confidence branch predicts the probability that candidate process paths will achieve target constraints such as desired hardness, dilution rate, or wear loss rate, outputting a continuous floating-point number between 0 and 1 as the confidence score. The stability score for the process parameter combination is calculated by a subsequent Monte Carlo parameter perturbation evaluation unit based on the performance variance under perturbation samples.

[0053] In some embodiments, obtaining the stability score of the process parameter combination corresponding to the candidate process path includes: Extract the process parameter combination corresponding to the candidate process path, input the process parameter combination into a pre-trained Monte Carlo parameter perturbation evaluation unit, inject random noise within the set tolerance range of the process parameter combination to simulate operating condition fluctuations, calculate the variance of the predicted value of the weld overlay performance index after injecting random noise, construct an inverse bounded mapping based on the variance, and calculate the stability score of the process parameter combination corresponding to the candidate process path.

[0054] The Monte Carlo parameter perturbation evaluation unit embeds a pre-trained surrogate response network model, which includes a multilayer perceptron regressor. The input is a sample set of perturbation process parameters containing random noise, and the output is the predicted value of performance indicators such as Rockwell hardness or wear loss rate of the weld overlay.

[0055] For a complete candidate process path decoded by a multi-hop network, its specific parameter set is analyzed, such as extracting multiple continuous real-type control variable sets including preheating temperature 200℃, welding current 250A, and wire feeding speed 5m / min. This combination of process parameters serves as the basic input variable matrix for a preset Monte Carlo parameter disturbance evaluation unit. The evaluation unit, combined with the actual grid power fluctuations and material feeding machine vibration patterns, sets floating tolerance boundaries of, for example, ±5% for the aforementioned core control variables.

[0056] Within this range, a pseudo-random number generator is used to inject truncated Gaussian noise or limited Gaussian noise, and a sample group of micro-perturbation operating condition parameters with a scale of up to 1000 sets is generated through iterative resampling. The perturbation generation formula is as follows: , where ε follows a Gaussian distribution with a mean of 0, r is the tolerance ratio, preferably r=5%, and clip means that the perturbed parameter value is limited to the preset tolerance range.

[0057] A batch of 1000 perturbed sample sets containing random noise is fed into a pre-trained surrogate response network model. High-throughput performance metric prediction is performed through the model's forward mapping. The statistical variance is calculated based on the data distribution sequence consisting of the 1000 generated discrete predictions, using the formula: Where N is the sample size, For a single predicted value, The mean is used, for example, the variance of 2.5 is calculated for hardness fluctuations, which is used to represent the degree of dispersion of the current process path against environmental noise fluctuations.

[0058] Extract the variance value and perform a bounded inverse mapping to generate a stability score. The stability score is calculated using the following formula: Where Var is the variance of the predicted performance index of the weld overlay corresponding to the disturbed sample, and κ is the preset sensitivity coefficient, preferably ranging from 0.1 to 1.0. Therefore, the smaller the performance variance, the closer the stability score is to 1, indicating that the combination of process parameters is less sensitive to field disturbances; the larger the performance variance, the lower the stability score, indicating that the stability of the combination of process parameters is worse. For example, when Var = 2.5 and κ = 0.1, the stability score R = 1 / (1 + 0.1 × 2.5) = 0.80. This stability score and the aforementioned confidence score are both within the range of 0 to 1, and can directly proceed to the subsequent weighted sorting step.

[0059] S8. Based on the confidence score and the stability score, a weighted sort is performed to select the optimal combination of process parameters as the recommended result of hardfacing welding process parameters.

[0060] An objective weighting method based on information entropy, namely the entropy weighting method, is used to pre-calculate the information entropy distribution of confidence scores and stability scores in the sample set, and then allocate confidence weights and stability weights accordingly. The comprehensive evaluation score is obtained by multiplying the confidence score of each candidate process path by its confidence weight, and then adding the stability score multiplied by its stability weight.

[0061] The Timsort algorithm, a built-in list sorting algorithm in Python, is used to sort all candidate process paths generated through cluster search in descending order based on their comprehensive evaluation scores. The candidate process path corresponding to the highest comprehensive evaluation score is then extracted. The underlying entity attributes carried in this path are parsed, and the corresponding base material type, recommended welding wire grade, matching welding current range, arc voltage range, and preheating temperature parameter combination are output as the recommended hardfacing welding process parameters.

[0062] In some embodiments, the step of selecting the optimal combination of process parameters as the recommended result for hardfacing welding process parameters based on a weighted sorting of the confidence score and the stability score includes: Assign a first weighting coefficient and a second weighting coefficient to the confidence score and the stability score respectively. Multiply the confidence score by the first weighting coefficient and the stability score by the second weighting coefficient, and sum them to obtain the comprehensive evaluation total score of the candidate process path. Sort all candidate process paths in descending order according to the comprehensive evaluation total score. Select the process parameter combination corresponding to the candidate process path ranked first as the optimal recommendation result output. Alternatively, after passing unit consistency verification, parameter upper and lower limit verification, equipment interlock status verification and manual confirmation verification, write it into the loading parameter area of ​​the welding equipment.

[0063] The preceding evaluation network receives confidence scores and stability scores for all compliant paths falling within the 0-1 standard range, representing the probability of that path achieving a specific surface strengthening target. Based on application requirements, an allocation strategy is developed to meet the normalization requirement of 1.0. For example, in heavy-duty mining machinery repair applications, a first weighting coefficient of 0.65 and a second weighting coefficient of 0.35 are assigned.

[0064] The first weighting coefficient is multiplied by the confidence score of each path using basic arithmetic multiplication. The stability score is processed similarly, and the products of the two paths are combined using linear superposition to obtain the overall evaluation score. For example, if the confidence score of a candidate process path is 0.90 and the stability score is 0.80, the calculated overall evaluation score is 0.865.

[0065] After calculating all candidate paths in the complete candidate pool, which contains up to 500 alternative paths, a sorting key list is constructed from all the comprehensive evaluation scores. A sorting algorithm with a time complexity of O(NlogN) is then called to sort all candidate process paths in descending order based on their comprehensive evaluation scores.

[0066] After the sorting results are determined, the optimal candidate path ranked first in the global leaderboard is extracted based on the first index of the array. The optimal combination of process parameters, including preheating parameters, wire feeding parameters, and AC setting parameters implicit in this topology, is then extracted. This generates a human-machine readable parameter recommendation sheet containing parameter names, units, recommended ranges, recommended center values, and applicable material combinations, and stores it in the pending confirmation queue of the process management system. If the field equipment has a pre-configured communication interface, the recommended parameters can be mapped to equipment-recognizable data items according to the OPC UA, Modbus-TCP, or vendor-open SDK interface fields supported by the equipment. Before distribution, unit consistency checks, parameter upper and lower limit checks, equipment interlock status checks, and manual confirmation checks are performed. If any check fails, only the recommendation result is retained, and no control command is issued to the robotic welding machine. If all checks pass, the parameters are written to the equipment's pending loading parameter area. Therefore, in this embodiment, "execution" is preferably understood as the confirmation, loading, or pending execution configuration of the recommendation result, without directly overriding the equipment's existing safety control procedures.

[0067] Example 2 Embodiment 2 of the present invention proposes a hardfacing welding process parameter recommendation system, which includes the following modules: The reconstruction module is used to construct a knowledge graph of hardfacing welding process, and to calculate offline the basic metallurgical compatibility score between the cached workpiece base material and the candidate welding material under the benchmark process parameters. For the reasoning request of hardfacing welding process parameters, the module combines the semantic information of the query context to prune and reconstruct the knowledge graph of hardfacing welding process to generate a reasoning subgraph that serves a specific task. The adjustment module is used to embed the inference subgraph into a low-dimensional vector space using a graph attention network to perform multi-hop inference path search. During the multi-hop inference path search, the basic metallurgical compatibility score below a preset threshold is used as a constraint to prune the inference path or increase its cost. During the inference path expansion process, the attention coefficient is modulated by the current inference path aggregation information and uncertainty measure. A parameterized fuzzy interval model is established for the numerical constraint nodes in the inference path. The membership function boundary of the numerical constraint interval is adjusted using the upstream process parameter vector of the current inference path as input. The filtering module is used to encode the candidate process paths containing the adjusted numerical constraint interval boundaries into a sequence vector, input the sequence vector into a multi-task path evaluation network, and output the confidence scores of the candidate process paths in achieving the expected performance indicators in parallel. At the same time, it obtains the stability scores of the process parameter combinations corresponding to the candidate process paths, and filters out the optimal process parameter combinations based on the weighted sorting of the confidence scores and the stability scores as the recommended results of hardfacing welding process parameters.

[0068] In some embodiments, the step of pruning and reconstructing the hardface welding process knowledge graph by combining query context semantic information to generate a reasoning subgraph serving a specific task includes: The current query task expression is extracted into an intent sequence vector using a natural language processing model. The feature representations of entities in the hardface welding process knowledge graph are obtained. The intent sequence vector is converted into a single vector through pooling aggregation and projected to the same dimensional space as the feature representation. The inner product similarity is calculated with the feature representations of each entity to obtain the relevance score between each entity and the query task. Nodes and edges with relevance scores lower than a preset threshold are filtered out, and the remaining nodes and relationships are retained to form the reasoning subgraph.

[0069] In some embodiments, the step of embedding the inference subgraph into a low-dimensional vector space using a graph attention network for multi-hop inference path search includes: The initial features of nodes in the inference subgraph are mapped to hidden layer representations using a multilayer perceptron. During message passing between nodes, attention weights are calculated for the features of neighboring nodes and the current path state features. The weighted features of neighboring nodes are then fused with the features of the current node. The node vector representation is updated using an activation function and used as the current state input for multi-hop inference to the next-hop search network.

[0070] In some embodiments, the offline calculation of the basic metallurgical compatibility score between the cached workpiece base material and the candidate weld overlay material under baseline process parameters includes: Extract the chemical element mass fraction list of the workpiece base material and the candidate weld overlay material, input the chemical element mass fraction list and the benchmark process parameters into a pre-trained metallurgical phase prediction model to predict the proportion of beneficial phases and the bonding strength evaluation value at the interface of the weld overlay layer, normalize the proportion of beneficial phases and the bonding strength evaluation value and calculate the basic metallurgical compatibility score by weighted summation, and store the basic metallurgical compatibility score in a static mapping table for online query and retrieval.

[0071] In some embodiments, adjusting the membership function boundary of the numerical constraint interval includes: Obtain the upstream process parameter feature latent vector that has been determined in the current inference path, input the upstream process parameter feature latent vector into the boundary bias prediction network to calculate the boundary offset and width scaling factor of the fuzzy interval membership function, and modify the center position and limit width of the pre-set initial membership function based on the boundary offset and the width scaling factor to generate the numerical constraint interval boundary.

[0072] In some embodiments, obtaining the stability score of the process parameter combination corresponding to the candidate process path includes: Extract the process parameter combination corresponding to the candidate process path, input the process parameter combination into a pre-trained Monte Carlo parameter perturbation evaluation unit, inject random noise within the set tolerance range of the process parameter combination to simulate operating condition fluctuations, calculate the variance of the predicted value of the weld overlay performance index after injecting random noise, construct an inverse bounded mapping based on the variance, and calculate the stability score of the process parameter combination corresponding to the candidate process path.

[0073] In some embodiments, the step of selecting the optimal combination of process parameters as the recommended result for hardfacing welding process parameters based on a weighted sorting of the confidence score and the stability score includes: Assign a first weighting coefficient and a second weighting coefficient to the confidence score and the stability score respectively. Multiply the confidence score by the first weighting coefficient and the stability score by the second weighting coefficient, and sum them to obtain the comprehensive evaluation total score of the candidate process path. Sort all candidate process paths in descending order according to the comprehensive evaluation total score. Select the process parameter combination corresponding to the candidate process path ranked first as the optimal recommendation result output. Alternatively, after passing unit consistency verification, parameter upper and lower limit verification, equipment interlock status verification and manual confirmation verification, write it into the loading parameter area of ​​the welding equipment.

[0074] In this specification, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise limited, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. In this document, "a," "an," "the," "the," and "its" may also include plural forms unless the context clearly indicates otherwise. "Multiple" refers to at least two, such as 2, 3, 5, or 8, etc. "And / or" includes any and all combinations of the associated listed items.

[0075] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0076] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for recommending hardfacing welding process parameters, characterized in that, include: Construct a knowledge graph of hardfacing welding process, and calculate offline the basic metallurgical compatibility score between the cached workpiece base material and the candidate welding material under the benchmark process parameters. For the reasoning request of hardfacing welding process parameters, combine the semantic information of the query context to prune and reconstruct the knowledge graph of hardfacing welding process to generate a reasoning subgraph that serves a specific task. A graph attention network is used to embed the inference subgraph into a low-dimensional vector space for multi-hop inference path search. During the multi-hop inference path search, the basic metallurgical compatibility score below a preset threshold is used as a constraint to prune the inference path or increase its cost. During the inference path expansion process, the attention coefficient is modulated by the aggregated information and uncertainty measure of the current inference path. A parameterized fuzzy interval model is established for the numerical constraint nodes in the inference path. The membership function boundary of the numerical constraint interval is adjusted using the upstream process parameter vector of the current inference path as input. The candidate process paths, including the adjusted numerical constraint interval boundaries, are encoded as a sequence vector. The sequence vector is then input into a multi-task path evaluation network, which outputs the confidence scores of the candidate process paths in achieving the expected performance indicators in parallel. At the same time, the stability scores of the process parameter combinations corresponding to the candidate process paths are obtained. Based on the confidence scores and the stability scores, the optimal process parameter combinations are selected as the recommended results for hardfacing welding process parameters.

2. The method according to claim 1, characterized in that, The step of pruning and reconstructing the hardface welding process knowledge graph by combining query context semantic information to generate a reasoning subgraph serving a specific task includes: The current query task expression is extracted into an intent sequence vector using a natural language processing model. The feature representations of entities in the hardface welding process knowledge graph are obtained. The intent sequence vector is converted into a single vector through pooling aggregation and projected to the same dimensional space as the feature representation. The inner product similarity is calculated with the feature representations of each entity to obtain the relevance score between each entity and the query task. Nodes and edges with relevance scores lower than a preset threshold are filtered out, and the remaining nodes and relationships are retained to form the reasoning subgraph.

3. The method according to claim 1, characterized in that, The step of embedding the inference subgraph into a low-dimensional vector space using a graph attention network for multi-hop inference path search includes: The initial features of nodes in the inference subgraph are mapped to hidden layer representations using a multilayer perceptron. During message passing between nodes, attention weights are calculated for the features of neighboring nodes and the current path state features. The weighted features of neighboring nodes are then fused with the features of the current node. The node vector representation is updated using an activation function and used as the current state input for multi-hop inference to the next-hop search network.

4. The method according to claim 3, characterized in that, The offline calculation of the basic metallurgical compatibility score between the cached workpiece base material and the candidate weld overlay material under the baseline process parameters includes: Extract the chemical element mass fraction list of the workpiece base material and the candidate weld overlay material, input the chemical element mass fraction list and the benchmark process parameters into a pre-trained metallurgical phase prediction model to predict the proportion of beneficial phases and the bonding strength evaluation value at the interface of the weld overlay layer, normalize the proportion of beneficial phases and the bonding strength evaluation value and calculate the basic metallurgical compatibility score by weighted summation, and store the basic metallurgical compatibility score in a static mapping table for online query and retrieval.

5. The method according to claim 1, characterized in that, The adjustment of the membership function boundary of the numerical constraint interval includes: Obtain the upstream process parameter feature latent vector that has been determined in the current inference path, input the upstream process parameter feature latent vector into the boundary bias prediction network to calculate the boundary offset and width scaling factor of the fuzzy interval membership function, and modify the center position and limit width of the pre-set initial membership function based on the boundary offset and the width scaling factor to generate the numerical constraint interval boundary.

6. The method according to claim 1 or 2, characterized in that, The process of obtaining the stability score of the process parameter combination corresponding to the candidate process path includes: Extract the process parameter combination corresponding to the candidate process path, input the process parameter combination into a pre-trained Monte Carlo parameter perturbation evaluation unit, inject random noise within the set tolerance range of the process parameter combination to simulate operating condition fluctuations, calculate the variance of the predicted value of the weld overlay performance index after injecting random noise, construct an inverse bounded mapping based on the variance, and calculate the stability score of the process parameter combination corresponding to the candidate process path.

7. The method according to claim 1, characterized in that, The process of selecting the optimal combination of process parameters based on the weighted ranking of the confidence score and the stability score as the recommended result of the hardfacing welding process parameters includes: Assign a first weighting coefficient and a second weighting coefficient to the confidence score and the stability score respectively. Multiply the confidence score by the first weighting coefficient and the stability score by the second weighting coefficient, and sum them to obtain the comprehensive evaluation total score of the candidate process path. Sort all candidate process paths in descending order according to the comprehensive evaluation total score. Select the process parameter combination corresponding to the candidate process path ranked first as the optimal recommendation result output. Alternatively, after passing unit consistency verification, parameter upper and lower limit verification, equipment interlock status verification and manual confirmation verification, write it into the loading parameter area of ​​the welding equipment.

8. A hardfacing welding process parameter recommendation system, characterized in that, include: The reconstruction module is used to construct a knowledge graph of hardfacing welding process, and to calculate offline the basic metallurgical compatibility score between the cached workpiece base material and the candidate welding material under the benchmark process parameters. For the reasoning request of hardfacing welding process parameters, the module combines the semantic information of the query context to prune and reconstruct the knowledge graph of hardfacing welding process to generate a reasoning subgraph that serves a specific task. The adjustment module is used to embed the inference subgraph into a low-dimensional vector space using a graph attention network to perform multi-hop inference path search. During the multi-hop inference path search, the basic metallurgical compatibility score below a preset threshold is used as a constraint to prune the inference path or increase its cost. During the inference path expansion process, the attention coefficient is modulated by the current inference path aggregation information and uncertainty measure. A parameterized fuzzy interval model is established for the numerical constraint nodes in the inference path. The membership function boundary of the numerical constraint interval is adjusted using the upstream process parameter vector of the current inference path as input. The filtering module is used to encode the candidate process paths containing the adjusted numerical constraint interval boundaries into a sequence vector, input the sequence vector into a multi-task path evaluation network, and output the confidence scores of the candidate process paths in achieving the expected performance indicators in parallel. At the same time, it obtains the stability scores of the process parameter combinations corresponding to the candidate process paths, and filters out the optimal process parameter combinations based on the weighted sorting of the confidence scores and the stability scores as the recommended results of hardfacing welding process parameters.

9. The system according to claim 8, characterized in that, The step of pruning and reconstructing the hardface welding process knowledge graph by combining query context semantic information to generate a reasoning subgraph serving a specific task includes: The current query task expression is extracted into an intent sequence vector using a natural language processing model. The feature representations of entities in the hardface welding process knowledge graph are obtained. The intent sequence vector is converted into a single vector through pooling aggregation and projected to the same dimensional space as the feature representation. The inner product similarity is calculated with the feature representations of each entity to obtain the relevance score between each entity and the query task. Nodes and edges with relevance scores lower than a preset threshold are filtered out, and the remaining nodes and relationships are retained to form the reasoning subgraph.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program, when executed by a processor, implements the method as described in any one of claims 1-7.

Images