CN-121982374-A - Method and system for detecting welding quality of steel structural part

CN121982374ACN 121982374 ACN121982374 ACN 121982374ACN-121982374-A

Abstract

The invention relates to the technical field of welding quality detection, in particular to a method and a system for detecting the welding quality of a steel structure. The method comprises the steps of obtaining perception data of a welding workpiece, extracting defect feature vectors, converting the defect feature vectors into semantic descriptions, storing the semantic descriptions in a defect case library, constructing a weld defect field ontology model, organizing and indexing cases in the defect case library according to the ontology model, constructing a machine learning model, carrying out reasoning prediction on defects to be detected, outputting a prediction result and prediction confidence, dynamically selecting a reasoning path according to the prediction confidence, triggering case retrieval when the prediction confidence is lower than a preset threshold, carrying out similarity matching in the defect case library by adopting a heuristic optimization algorithm in case retrieval, obtaining an optimal matching case through fitness evaluation, and determining the defect type according to the optimal matching case. The invention realizes high accuracy and quick adaptability of welding defect detection through the organic fusion of the neural network quick reasoning and the case reasoning accurate matching.

Inventors

HU XIAODONG
DAN BINBIN
XIONG LING
HU TAO

Assignees

武汉汉光钢品建设工程有限公司
武汉科技大学

Dates

Publication Date: 20260505
Application Date: 20251229

Claims (10)

1. The method for detecting the welding quality of the steel structural member is characterized by comprising the following steps of: obtaining the perceived data of a welding workpiece, extracting a defect feature vector, converting the defect feature vector into semantic description, and storing the semantic description into a defect case library; Constructing a body model in the welding defect field, wherein the body model comprises a defect type layer, a defect attribute layer and a defect instance layer, and cases in the defect case library are organized and indexed according to the body model; constructing a machine learning model taking the perception data and the defect feature vector as input, carrying out reasoning prediction on the defect to be detected, and outputting a defect type prediction result and a prediction confidence; Dynamically selecting an inference path according to the prediction confidence coefficient, namely directly outputting the defect type prediction result when the prediction confidence coefficient is higher than or equal to a preset threshold value, and triggering case retrieval when the prediction confidence coefficient is lower than the preset threshold value; In the case retrieval, similarity matching is carried out in the defect case library by adopting a heuristic optimization algorithm, an optimal matching case is obtained through fitness evaluation, and the defect type is determined according to the optimal matching case.
2. The method for detecting the welding quality of the steel structural member according to claim 1, wherein the sensing data is multi-mode image data acquired through a visual sensing device.
3. The method for detecting the welding quality of the steel structural part according to claim 2 is characterized in that the visual sensing device comprises a three-dimensional visual camera and a multispectral visual sensor array, wherein the three-dimensional visual camera acquires three-dimensional shape data of a welding seam by adopting a structured light projection triangulation principle, the sampling precision reaches the level of 0.01mm, and the multispectral visual sensor array comprises a visible light sensor, a near infrared sensor and an ultraviolet sensor which are respectively used for acquiring texture characteristic, heat distribution characteristic and fluorescence characteristic information of the welding seam.
4. The method for detecting the welding quality of the steel structural member according to claim 1, wherein the conversion of the semantically described is realized by adopting an ontology instance mapping mode, semantic association of the defect feature vector and concepts in the ontology model is realized through a query language, and the mapping process comprises three stages of feature value normalization processing, threshold value classification and semantic annotation.
5. The method of claim 1, wherein the defect type layer defines eight types of welding defects and sub-type hierarchies thereof, including blow holes, slag inclusions, cracks, unfused, incomplete penetration, undercut, flash, and crater.
6. The method of claim 1, wherein the defect attribute layer defines defect location attributes, defect size attributes, defect morphology attributes, and defect severity attributes, and spatial, causal, and concurrency relationships between defects.
7. The method for detecting welding quality of steel structural members according to claim 1, wherein the heuristic optimization algorithm is a genetic algorithm, and the fitness evaluation adopts a fitness function F (x): F(x)=α·Sim(x,C)+β·(1/D(x)) Wherein x is a feature vector of a defect to be detected, C is a history case set in the defect case library, sim (x, C) is a comprehensive similarity of x and C, D (x) is a calculation complexity of case matching, alpha and beta are weight coefficients, and alpha+beta=1, wherein the comprehensive similarity is calculated by multi-attribute weighting, sim (x, C) =Σ (wi· simi (xi, ci))/Σwi, wherein wi is a weight of an ith attribute, simi (xi, ci) is a local similarity of the ith attribute, and the attribute weight is determined by a hierarchical analysis method.
8. The method for detecting welding quality of steel structural members according to claim 1, wherein the machine learning model is a deep neural network model, a multi-layer perceptron structure is adopted, the machine learning model comprises an input layer, a plurality of hidden layers and an output layer, a hidden layer activation function adopts a ReLU function, the output layer adopts a Softmax function to output probability distribution of each defect type, a probability maximum value is used as a defect type prediction result, and the probability value is used as the prediction confidence.
9. The method for detecting the welding quality of the steel structural member according to claim 8, wherein the number of neurons of the hidden layer is 256, 128 and 64 respectively, an Adam optimizer is adopted in training of the deep neural network model, an initial value of learning rate is set to 0.001, a cosine annealing strategy is adopted for dynamic adjustment, batch size is set to 32, the maximum training round is 200, and early shutdown is adopted for preventing overfitting.
10. A steel structure welding quality detection system, comprising: the data acquisition and storage module is used for acquiring the perceived data of the welded workpiece, extracting the defect feature vector, converting the defect feature vector into semantic description and storing the semantic description into a defect case library; The body model construction module is used for constructing a body model in the welding defect field, the body model comprises a defect type layer, a defect attribute layer and a defect instance layer, and cases in the defect case library are organized and indexed according to the body model; The machine learning reasoning module is used for constructing a machine learning model taking the perception data and the defect characteristic vector as input, carrying out reasoning prediction on the defect to be detected, and outputting a defect type prediction result and a prediction confidence; The confidence level routing module is used for dynamically selecting an inference path according to the prediction confidence level, wherein the inference path is used for directly outputting the defect type prediction result when the prediction confidence level is higher than or equal to a preset threshold value, and triggering case retrieval when the prediction confidence level is lower than the preset threshold value; And the case retrieval module is used for carrying out similarity matching in the defect case library by adopting a heuristic optimization algorithm in the case retrieval, obtaining an optimal matching case through fitness evaluation, and determining the defect type according to the optimal matching case.

Description

Method and system for detecting welding quality of steel structural part Technical Field The invention relates to the technical field of welding quality detection, in particular to a method and a system for detecting the welding quality of a steel structure. Background The existing welding quality detection technology is mostly dependent on a traditional image processing algorithm or a single deep learning model. The former has the problems of incomplete feature extraction and low classification accuracy when facing complex and changeable welding defect forms; the latter has the disadvantages of poor interpretability and high requirement for training data amount, although breaking through in image recognition, and is difficult to meet the requirement for traceability of detection results in industrial application. The case reasoning is used as an intelligent method for simulating human experience reasoning, has good interpretability and knowledge accumulation capability, but the traditional case reasoning method has obvious defects in terms of case retrieval efficiency and matching precision, and a single strategy is difficult to cover diversified defect detection scenes, so that the robustness in a dynamic complex welding environment is insufficient, and the detection precision and efficiency in various scenes can not be ensured. Disclosure of Invention Aiming at the problems that in the prior art, feature extraction is incomplete, retrieval efficiency is low, single strategy robustness is insufficient, detection accuracy and efficiency in various scenes cannot be guaranteed, and the like, the invention provides a method and a system for detecting welding quality of a steel structural member. The technical scheme for solving the technical problems is as follows: in a first aspect, the present invention provides a method for detecting welding quality of a steel structural member, including: obtaining the perceived data of a welding workpiece, extracting a defect feature vector, converting the defect feature vector into semantic description, and storing the semantic description into a defect case library; Constructing a body model in the welding defect field, wherein the body model comprises a defect type layer, a defect attribute layer and a defect instance layer, and cases in the defect case library are organized and indexed according to the body model; constructing a machine learning model taking the perception data and the defect feature vector as input, carrying out reasoning prediction on the defect to be detected, and outputting a defect type prediction result and a prediction confidence; Dynamically selecting an inference path according to the prediction confidence coefficient, namely directly outputting the defect type prediction result when the prediction confidence coefficient is higher than or equal to a preset threshold value, and triggering case retrieval when the prediction confidence coefficient is lower than the preset threshold value; In the case retrieval, similarity matching is carried out in the defect case library by adopting a heuristic optimization algorithm, an optimal matching case is obtained through fitness evaluation, and the defect type is determined according to the optimal matching case. Optionally, the sensing data is multi-mode image data, and is acquired through a vision sensing device. Optionally, the visual sensing device comprises a three-dimensional visual camera and a multispectral visual sensor array, wherein the three-dimensional visual camera acquires three-dimensional shape data of the welding seam by adopting a structured light projection triangulation principle, the sampling precision reaches the level of 0.01mm, and the multispectral visual sensor array comprises a visible light sensor, a near infrared sensor and an ultraviolet sensor which are respectively used for acquiring texture characteristics, heat distribution characteristics and fluorescence characteristic information of the surface of the welding seam. Optionally, the conversion of the semantically described is realized by adopting an ontology instance mapping mode, semantic association of the defect feature vector and concepts in the ontology model is realized through a query language, and the mapping process comprises three stages of feature value normalization processing, threshold value classification and semantic annotation. Optionally, the defect type layer defines eight types of welding defects and their sub-type hierarchies of porosity, slag inclusion, cracking, lack of fusion, lack of penetration, undercut, flash, and crater. Optionally, the defect attribute layer defines defect location attributes, defect size attributes, defect morphology attributes, and defect severity attributes, as well as spatial, causal, and concurrency relationships between defects. Optionally, the heuristic optimization algorithm is a genetic algorithm, and the fitness evaluation adopts a fitness function F (x): F(x)=