CN-122021933-A - Self-adaptive control method and system for steel pipe weld scraping

CN122021933ACN 122021933 ACN122021933 ACN 122021933ACN-122021933-A

Abstract

The invention provides a steel pipe weld scraping self-adaptive control method and a system, which relate to the technical field of deep learning and comprise the steps of obtaining a training data set formed by weld morphology time sequence observation, tool intervention parameters and morphology change labeling, constructing a causal prediction neural network comprising a time sequence coding module, a differentiable causal graph module and a counterfactual reasoning module which are fused with periodic constraints, wherein the time sequence coding module considers weld annular continuity coding morphology samples, the differentiable causal graph module represents causal relations between tool parameters and residual heights through parameterized adjacency matrixes, the counterfactual reasoning module predicts morphology changes under the condition that intervention parameters are not implemented, dividing task subset training according to steel pipe specifications by adopting a meta learning frame, optimizing adjacency matrixes through causal sparsity constraint, identifying causal effect intensity, and optimizing learning task-crossing initialization parameters in a double-layer mode.

Inventors

DAI XIANGYANG
ZHANG ZUOQUAN
SUN HAICHENG
GAO BEN
GAO SHIGUI
JIN BOJUN

Assignees

陕西友发钢管有限公司

Dates

Publication Date: 20260512
Application Date: 20260319

Claims (9)

1. The self-adaptive control method for the scraping of the welding line of the steel pipe is characterized by comprising the following steps: Acquiring a training data set from a steel pipe weld scraping process, wherein the training data set comprises a weld morphology time sequence observation sample, a tool intervention parameter sample and a morphology change labeling sample; The method comprises the steps of constructing a causal prediction neural network, wherein the causal prediction neural network comprises a time sequence coding module, a differentiable causal graph module and a counterfactual reasoning module, the time sequence coding module is used for coding a weld joint morphology time sequence observation sample in consideration of the annular continuity of a weld joint, the differentiable causal graph module is used for representing the causal relation between tool intervention parameters and the residual height of the weld joint through a parameterized adjacency matrix, and the counterfactual reasoning module is used for predicting morphology change under the tool intervention parameters which are not actually applied based on a causal graph topological structure; The causal prediction neural network is trained by adopting a meta-learning training framework, a training data set is divided into a plurality of task subsets according to the specification of a steel pipe, a sample supervision network is marked on each task subset through morphological change to output, causal structure learning is carried out on the differentiable causal graph module, a loss function optimization adjacent matrix parameter containing causal sparsity constraint is constructed to identify causal effect intensity, and a trained network model is obtained through double-layer optimization learning of cross-task network initialization parameters.
2. The method of claim 1, wherein the timing encoding module encodes the weld topography timing observation samples taking into account the annular continuity of the weld, comprising: Mapping the circumferential initial measurement position and the circumferential end measurement position of the weld joint into continuous connected boundary nodes on a topological structure, constructing a coding network in cyclic connection, and outputting the coding state of the end measurement position as the input of an initial measurement position coding layer to form a closed information flow path; Position identification is carried out on each measuring position in a weld morphology time sequence observation sample by adopting a periodic index mapping strategy in the coding network which is circularly connected, and a continuously-changing periodic index value is given to the measuring positions at adjacent periodic boundaries; and the circularly connected coding network takes the periodically indexed mapped measurement position features as input to carry out forward propagation, and outputs a coding vector sequence for generating the weld joint morphology baseline features by the inverse fact reasoning module.
3. The method of claim 2, wherein the step of constructing a cyclically connected coding network comprises: Constructing a bidirectional cyclic convolution coding network, sequentially extracting features from a starting measurement position to a final measurement position along the circumferential direction of a welding seam in a forward propagation path, reversely extracting features from the final measurement position to the starting measurement position in a reverse propagation path, establishing a cross-boundary connection channel between the starting position and the final position, taking a hidden state vector of the final position as an additional input for extracting the features of the starting position, and taking a hidden state vector of the starting position as an additional input for extracting the features of the final position to form a closed-loop information flow topology; Generating position coding vectors for each measurement position in the circumferential direction of the weld joint according to the angular coordinates thereof through sine and cosine transformation of which the period is the circumferential angle of the complete weld joint; And splicing the periodic position coding vector with the weld morphology observation numerical value characteristic to input the bidirectional cyclic convolution coding network, and outputting a characteristic vector sequence after the bidirectional cyclic convolution coding network transmits hidden state vectors of the starting boundary position and the ending boundary position in a crossing manner through a cross-boundary connecting channel.
4. The method of claim 1, wherein the differentiable causal graph module represents causal relationships between tool intervention parameters and weld residual height by parameterized adjacency matrices, and wherein the anti-facts inference module predicts topographical changes under tool intervention parameters not actually applied based on causal graph topology, comprising: Initializing an adjacent matrix, wherein a row index corresponds to each dimension of the tool intervention parameter, a column index corresponds to the numerical value of the residual height of the welding seam at each measurement position in the circumferential direction, the numerical value of matrix elements represents causal influence intensity, and the adjacent matrix parameters can be learned and optimized in the training process through gradient back propagation; The inverse facts inference module receives the topology of the adjacency matrix, and for a given hypothetical tool intervention parameter, obtains a weld residual height distribution prediction under inverse facts by cutting off causal paths in the adjacency matrix that are associated with actual observed tool parameters and activating causal paths that correspond to the hypothetical tool parameters, and performing forward propagation calculations along the activation paths.
5. The method of claim 4, wherein the step of the counterfactual inference module receiving topology of the adjacency matrix to obtain a weld residual height distribution prediction for a given hypothetical tool intervention parameter by cutting off a causal path in the adjacency matrix associated with the actual observed tool parameter and activating a causal path corresponding to the hypothetical tool parameter for forward propagation calculations along the activation path comprises: constructing a causal path mask tensor, wherein a row dimension corresponds to a tool intervention parameter dimension of an adjacency matrix, a column dimension corresponds to a circumferential measurement position of a residual height of a weld, a mask value is set to zero for a row corresponding to an actually observed tool intervention parameter to cut off causal propagation of the parameter, and a mask value is set to one for a row corresponding to an assumed tool intervention parameter to activate causal propagation of the parameter; Performing element-by-element multiplication operation on the causal path mask tensor and an adjacent matrix to obtain a causal graph structure of the dry prognosis, reserving the connection weight from the hypothetical tool parameter to the residual height position of the welding line, and simultaneously blocking the influence transmission of the actual observation tool parameter; Matrix multiplication is carried out on the numerical vector of the assumed tool intervention parameter and a causal graph structure of the dry prognosis, a causal effect accumulation value received by each weld residual height measurement position is calculated through an aggregation operation summed along a row, and a residual height distribution prediction under the counter-facts condition is generated by combining the weld appearance baseline characteristic provided by the time sequence coding module.
6. The method of claim 1, wherein performing causal structure learning on the differentiable causal graph module, constructing a loss function optimization adjacency matrix parameter containing causal sparsity constraints to identify causal effect strengths, comprises: Constructing a total loss function, wherein the total loss function comprises a prediction loss term and a sparsity regular term, the prediction loss term measures the deviation between the predicted morphology change and the morphology change labeling sample output by the neural network for predicting morphology change, and the sparsity regular term penalizes the number of non-zero elements of the adjacent matrix; Simultaneously calculating gradients of the total loss function on network parameters and adjacent matrix parameters of the time sequence coding module and the inverse fact reasoning module in each gradient updating iteration, and synchronously updating the network parameters and the adjacent matrix parameters through gradient descent; After training, a significance threshold value is adaptively determined according to the numerical distribution of adjacent matrix elements, elements with absolute values exceeding the significance threshold value in the adjacent matrix are screened, the connection between the corresponding tool intervention parameters and the residual height position of the welding seam is determined as a causal effect relation pair, and the absolute values of the elements represent the causal effect intensity.
7. The method of claim 1, wherein the use of the trained network model in adaptive control of steel pipe weld skiving comprises: And acquiring real-time morphology observation data of the weld joint of the steel pipe to be processed, inputting the network model after training, predicting the morphology evolution trend of the weld joint under the combination of a plurality of candidate tool intervention parameters by using the inverse fact reasoning module, and selecting the tool intervention parameters which enable the residual height uniformity index of the weld joint to be optimal as control output to realize the self-adaptive accurate control of the scraping of the weld joint of the steel pipe.
8. A steel pipe weld skiving adaptive control system for implementing the method of any of the preceding claims 1-7, comprising: The first unit is used for acquiring a training data set from the steel pipe weld scraping process, wherein the training data set comprises a weld morphology time sequence observation sample, a tool intervention parameter sample and a morphology change labeling sample; The second unit is used for constructing a causal prediction neural network and comprises a time sequence coding module, a differentiable causal graph module and a counterfactual reasoning module which are fused with periodic constraints; the time sequence coding module codes a weld seam morphology time sequence observation sample in consideration of the annular continuity of a weld seam, the differentiable causal graph module represents the causal relationship between tool intervention parameters and the residual height of the weld seam through a parameterized adjacency matrix, and the inverse fact reasoning module predicts morphology changes under tool intervention parameters which are not actually applied based on a causal graph topological structure; The third unit is used for training the causal prediction neural network by adopting a meta-learning training framework, dividing a training data set into a plurality of task subsets according to the specification of the steel pipe, and labeling a sample supervision network output on each task subset through morphology change; performing causal structure learning on the differentiable causal graph module, and constructing a loss function optimization adjacency matrix parameter containing causal sparsity constraint to identify causal effect intensity; and obtaining a trained network model through double-layer optimization learning of cross-task network initialization parameters.
9. A computer readable storage medium having stored thereon computer program instructions, which when executed by a processor, implement the method of any of claims 1 to 7.

Description

Self-adaptive control method and system for steel pipe weld scraping Technical Field The invention relates to a deep learning technology, in particular to a steel pipe weld scraping self-adaptive control method and system. Background The scraping of the welding seam of the steel pipe is a key post-treatment procedure in the production of the steel pipe, and the residual metal protruding from the surface of the welding seam is removed through a mechanical tool, so that the flatness of the outer surface of the steel pipe is ensured to meet the technological requirements. However, the existing adaptive control method mainly adopts an empirical formula or a traditional machine learning model to establish a mapping relation between tool parameters and morphology changes, and is difficult to accurately describe a complex nonlinear coupling mechanism. Although the deep learning method can fit a high-dimensional nonlinear relation, the method lacks of interpretability, and the topological characteristics of the annular continuity of the welding line are not fully considered, so that characteristic extraction deviation is easy to generate at the circumferential start-stop boundary. More importantly, the existing method can only predict the morphology change under the current tool parameters, cannot support the counterfactual reasoning, namely, the effect of the intervention parameters which are not actually applied is predicted, and limits the optimization space of the control strategy. Therefore, development of a novel modeling method integrating causal reasoning and deep learning is needed to realize intelligent control of precisely interpretable steel pipe weld scraping. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a self-adaptive control method and a self-adaptive control system for the scraping of a steel pipe welding line, which can solve the problems in the prior art. In a first aspect of the embodiment of the present invention, a method for adaptively controlling scraping of a weld joint of a steel pipe is provided, including: Acquiring a training data set from a steel pipe weld scraping process, wherein the training data set comprises a weld morphology time sequence observation sample, a tool intervention parameter sample and a morphology change labeling sample; The method comprises the steps of constructing a causal prediction neural network, wherein the causal prediction neural network comprises a time sequence coding module, a differentiable causal graph module and a counterfactual reasoning module, the time sequence coding module is used for coding a weld joint morphology time sequence observation sample in consideration of the annular continuity of a weld joint, the differentiable causal graph module is used for representing the causal relation between tool intervention parameters and the residual height of the weld joint through a parameterized adjacency matrix, and the counterfactual reasoning module is used for predicting morphology change under the tool intervention parameters which are not actually applied based on a causal graph topological structure; The causal prediction neural network is trained by adopting a meta-learning training framework, a training data set is divided into a plurality of task subsets according to the specification of a steel pipe, a sample supervision network is marked on each task subset through morphological change to output, causal structure learning is carried out on the differentiable causal graph module, a loss function optimization adjacent matrix parameter containing causal sparsity constraint is constructed to identify causal effect intensity, and a trained network model is obtained through double-layer optimization learning of cross-task network initialization parameters. Optionally, the time sequence encoding module encodes the weld morphology time sequence observation sample in consideration of the annular continuity of the weld, including: Mapping the circumferential initial measurement position and the circumferential end measurement position of the weld joint into continuous connected boundary nodes on a topological structure, constructing a coding network in cyclic connection, and outputting the coding state of the end measurement position as the input of an initial measurement position coding layer to form a closed information flow path; Position identification is carried out on each measuring position in a weld morphology time sequence observation sample by adopting a periodic index mapping strategy in the coding network which is circularly connected, and a continuously-changing periodic index value is given to the measuring positions at adjacent periodic boundaries; and the circularly connected coding network takes the periodically indexed mapped measurement position features as input to carry out forward propagation, and outputs a coding vector sequence for generating the weld joint morphology baseline features