CN-121980510-A - Multilayer multi-channel welding defect detection method based on multi-source sensor feature fusion

CN121980510ACN 121980510 ACN121980510 ACN 121980510ACN-121980510-A

Abstract

The invention provides a multi-layer multi-channel welding defect detection method based on multi-source sensor feature fusion, which is characterized in that multi-mode signals in the multi-layer multi-channel welding process are collected through deploying a plurality of types of sensors, feature screening is carried out on high-dimensional feature vectors by using a Fisher discriminant method, iterative optimization is carried out on a PSO-LightGBM model to obtain a highest-accuracy model and an optimal feature subset, the feature subset is built by evaluating the degree of distinction between the categories based on the Fisher discriminant method, the high-accuracy features are gradually reserved as input sets of the model according to the gradual accumulation of the score, the optimal feature subset and the highest accuracy can be obtained through model training, the redundant feature interference model performance is effectively avoided, and a LightGBM model adopts a histogram algorithm and a leaf-phase growth strategy, can support incremental training and parallel calculation, has strong adaptability to large-scale data and high-dimensional features, and has low memory consumption and strong anti-interference performance.

Inventors

ZHANG RUNFENG
MA XIAOYANG
WANG XINGHUA
CHENG BIN

Assignees

洛阳船舶材料研究所(中国船舶集团有限公司第七二五研究所)

Dates

Publication Date: 20260505
Application Date: 20260126

Claims (10)

1. The multi-layer multi-channel welding defect detection method based on multi-source sensor feature fusion is characterized by comprising the following steps of: s1, constructing a multi-source signal acquisition platform, deploying a multi-type sensor, and synchronously acquiring multi-mode signals in a multi-layer multi-channel welding process; s2, designing a welding test scheme, preparing a welding defect sample, and classifying the welding defect sample according to backing welding and filling welding; S3, dividing a training set, a verification set and a test set, determining the time window size of a feature extraction unit according to the sampling rates of the space feature signals and the time sequence feature signals, and preprocessing the multi-mode signals; S4, extracting characteristics of the preprocessed multi-mode signals to obtain multi-source characteristics; s5, performing Z-Score standardization processing on the multi-source features, evaluating the degree of distinction between the categories of the features based on a Fisher discriminant method, accumulating and constructing feature subsets in a gradual way according to the Score level and the dimension by dimension, and gradually reserving high-Score features; And S6, gradually accumulating the 1-to-n-dimensional features to be used as input values of the PSO-LightGBM model for training, iterating and optimizing the model through the training set and the verification set, and evaluating the performance of the model by the test set to obtain the model with the highest accuracy and the optimal feature subset.
2. The multi-layer, multi-channel welding defect detection method based on multi-source sensor feature fusion of claim 1, wherein in S1 the multi-mode signal comprises a puddle image signal, an infrared image signal, a current signal, a voltage signal, and a sound signal.
3. The multi-layer multi-channel welding defect detection method based on multi-source sensor feature fusion according to claim 2, wherein in S1, the multi-type sensor includes a puddle camera for acquiring puddle images, a pre-polarized free field microphone for acquiring sound signals at the time of welding, a thermal infrared imager for acquiring temperature signals, a closed-loop hall current sensor for acquiring welding current, and a closed-loop hall voltage sensor for acquiring welding voltage.
4. The multi-layer, multi-pass weld defect detection method based on multi-source sensor feature fusion of claim 1, wherein in S2, the weld state of the foundation weld comprises no defects, leaks, lack of penetration, weld bias, voids, and misalignment, and the weld state of the filler weld comprises no defects, lack of fusion, voids, and slag inclusions.
5. The multi-layer and multi-channel welding defect detection method based on multi-source sensor feature fusion according to claim 1, wherein in S3, the data volume ratio of the training set, the verification set and the test set is 7:2:1, the time window size is determined according to the sampling rates of the space feature signal and the time sequence feature signal in the multi-mode signal, and a time stamp is added to the multi-mode signal, and synchronous alignment of signals is achieved based on the time stamp.
6. The multi-layer and multi-channel welding defect detection method based on multi-source sensor feature fusion according to claim 2, wherein in S3, the preprocessing includes performing image normalization processing on the molten pool image and clipping the size of the molten pool image, clipping a region of interest on the infrared image, and denoising the sound signal, the current signal, and the voltage signal.
7. The multi-layer and multi-channel welding defect detection method based on multi-source sensor feature fusion according to claim 6, wherein in S4, a lightweight convolutional neural network MobileNetV is adopted to perform feature extraction on the molten pool image, a statistical feature extraction method is adopted to perform feature extraction on the infrared image, and a time domain feature extraction method is adopted to perform feature extraction on the sound signal, the current signal and the voltage signal.
8. The multi-layer, multi-channel welding defect detection method based on multi-source sensor feature fusion of claim 7, wherein the statistical features extracted from the infrared image include mean, peak and root mean square, the statistical features extracted from the current and voltage signals include mean, root mean square, peak factor, kurtosis, skewness and variance, and the statistical features extracted from the sound signals include short-time zero-crossing rate, root mean square, peak-to-peak, peak factor, kurtosis and skewness.
9. The multi-layer, multi-channel weld defect detection method based on multi-source sensor feature fusion of claim 1, wherein the Fisher discriminant comprises the steps of: A1, calculating Fisher score of each feature, and calculating the ratio of the inter-class divergence to the intra-class divergence to obtain the score of each feature, wherein the higher the score is, the more effective the feature can distinguish different defect categories; A2, gradually selecting features, starting from the features with highest scores, gradually increasing feature dimensions, starting from the first feature, sequentially reserving the first two features, the first three features and the n-dimensional features until all the features are reserved, and recalculating the accuracy of the model after adding one feature each time; a3, selecting an optimal dimension, and selecting an optimal feature subset capable of maximizing the model accuracy by comparing the classification accuracy under different feature dimensions.
10. The multi-layer multi-channel welding defect detection method based on multi-source sensor feature fusion of claim 1, wherein the LightGBM model is a gradient lifting decision tree model.

Description

Multilayer multi-channel welding defect detection method based on multi-source sensor feature fusion Technical Field The invention relates to the technical field of welding quality on-line monitoring, in particular to a multi-layer multi-channel welding defect detection method based on multi-source sensor feature fusion. Background With the increase of technical progress and industry demands, the application of thick plates in the fields of bridges, petrochemical equipment, ships, buildings and the like is gradually increased, and the multilayer multi-channel welding technology is widely applied to welding of medium and thick plates due to the advantages of low single-channel heat input, interlayer local heat treatment and the like. However, due to the fact that the thickness of the welding seam is large, the number of passes is large, the problems of improper heat input control, irregular interlayer treatment and the like easily occur in multi-layer multi-pass welding, and stability and reliability of the welding process cannot be guaranteed. The quality monitoring in the welding process is used as a key technology of automatic and intelligent welding, is an important means for realizing the driving of the welding manufacture from experience driving to signal driving, and is very important for constructing a defect on-line monitoring model by collecting signals such as acousto-optic and the like in the welding process. The invention patent with publication number of CN120355645A provides a welding defect detection method based on deep learning, a welding defect image is processed through a deep learning algorithm, and a multi-scale convolution attention module used by the welding defect detection method can effectively capture the characteristics of strip-shaped welding defects while effectively reducing the calculated amount, so that the multi-scale defect detection performance of the model under a complex background is improved. The method has the defects that only a single sensor is adopted for signal acquisition, the detection dimension is single, the anti-interference capability is weak, and complex defects are difficult to comprehensively and accurately identify. The invention patent with publication number of CN120177482A provides a defect identification method in an arc welding process based on multi-sensor signals, which combines a mutual signal method with a particle swarm optimization algorithm-least square support vector machine to perform characteristic selection on multi-source sensing high-dimensional signals by collecting current signals, temperature signals and visual image signals in the welding process and performing characteristic extraction, and has the defects that a simple signal set obtained by performing characteristic selection and signal compression on the signals by the mutual signal method cannot be ensured to be an optimal signal set, and on the other hand, the calculation complexity of the least square support vector machine is obviously increased along with the increase of the sample quantity, and the efficiency is lower when processing large-scale signals. Disclosure of Invention In view of the above, the present invention aims to provide a multi-layer multi-channel welding defect detection method based on multi-source sensor feature fusion, so as to solve the problem that an optimal feature subset cannot be screened and a large-scale high-dimensional signal cannot be processed rapidly in welding defect detection. In order to achieve the above purpose, the technical scheme of the invention is realized as follows: A multi-layer multi-channel welding defect detection method based on multi-source sensor feature fusion comprises the following steps: s1, constructing a multi-source signal acquisition platform, deploying a multi-type sensor, and synchronously acquiring multi-mode signals in a multi-layer multi-channel welding process; s2, designing a welding test scheme, preparing a welding defect sample, and classifying the welding defect sample according to backing welding and filling welding; S3, dividing a training set, a verification set and a test set, determining the time window size of a feature extraction unit according to the sampling rates of the space feature signals and the time sequence feature signals, and preprocessing the multi-mode signals; S4, extracting characteristics of the preprocessed multi-mode signals to obtain multi-source characteristics; s5, performing Z-Score standardization processing on the multi-source features, evaluating the degree of distinction between the categories of the features based on a Fisher discriminant method, accumulating and constructing feature subsets in a gradual way according to the Score level and the dimension by dimension, and gradually reserving high-Score features; And S6, gradually accumulating the 1-to-n-dimensional features to be used as input values of the PSO-LightGBM model for training, iterating and optimizing the