Electric arc welding process defect identification method based on multi-sensor information

Abstract

The invention belongs to the technical field of welding, discloses an electric arc welding process defect identification method based on multi-sensor information, and provides an electric arc welding process defect monitoring system based on welding process multi-sensor information. According to the system, current signals, temperature signals and visual image signals in the welding process are collected, the information theory and machine learning are combined, and real-time recognition of welding defects is achieved. The system aims at solving the problems of massive high-dimensional data volume generated by multi-sensor information fusion, rapid solving of appropriate model parameters and the like in a traditional welding defect detection method, and provides a welding defect identification method combining a mutual information method and a particle swarm optimization algorithm-least square support vector machine. Feature selection is carried out on multi-source sensing high-dimensional data through a mutual information method, optimization training is carried out on a particle swarm optimization algorithm-least square support vector machine model by utilizing a simplified data set, and a classification model established by the method has the characteristics of low complexity, short training and prediction time, relatively strong generalization ability and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of arc welding, and in particular to a method for identifying defects in the arc welding process based on multi-sensor information. Background Art

[0002] Welding technology is one of the key technologies in modern industrial manufacturing. With the gradual popularization of robotic welding in the manufacturing industry, welding quality monitoring has become one of the core links in automated production. Although the design and optimization of pre-welding process parameters play a leading role, since welding is a highly non-linear, multi-variable coupled, and complex process with uncertain factors at the same time, it is difficult to achieve "zero defects" in the surface and internal quality of welds. At present, the detection of welding quality mainly focuses on after welding, with low detection efficiency, high cost, and high requirements for operators. It can only be used for spot checks of small-batch products. On-line welding quality detection can discover various welding defects in a timely manner based on the effective quality information obtained during the welding process, and then adjust the process conditions or parameters to control the welding quality and process in real time, so as to meet the needs of large-scale automated welding production.

[0003] Sensing technology is the key factor to realize the automation and intelligence of the welding process. Aiming at the deficiencies in the comprehensiveness and reliability of information acquisition by a single sensor during the welding process, multi-sensor information fusion utilizes the complementarity of different types of source information, and can describe the welding process and its quality characteristics from multiple angles and aspects, which is conducive to improving the reliability and accuracy of welding defect identification and prediction. Chinese invention patent CN119246531A trains a deep learning model to identify and classify defects in the fused feature representation to obtain defect identification results. However, at the same time, the increase in the number of sensors inevitably causes a sharp increase in the dimensionality of the feature space, generating "big data" in the welding process. Therefore, how to eliminate redundant information and noise data from large-scale raw data and establish the connection between the effective information in the welding process and the defect categories is of great value for improving and perfecting on-line welding quality monitoring technology.

[0004] Chinese invention patent CN114140669A discloses a method, device and computer terminal for training a welding defect recognition model, which uses a support vector machine model to improve the recognition accuracy of the welding defect recognition model. As a machine learning algorithm based on statistical learning theory, the support vector machine shows unique advantages in solving small-sample and non-linear classification scenarios, and at the same time has strong generalization ability and robustness. The least squares support vector machine is an improvement and development of the traditional support vector machine. Although the least squares support vector machine is not overly sensitive to complex situations in the feature space, it cannot completely solve the problem of feature simplification. In addition, for the selection of the parameters - the penalty factor c and the kernel parameter g, which have a great influence on the classification performance of the least squares support vector machine, it is obviously difficult to achieve the expected effect by relying on experience or trial and error. How to quickly obtain the appropriate values of c and g is of great significance for improving the classification ability and learning efficiency of the least squares support vector machine model.

[0005] Through the above analysis, the problems and defects existing in the prior art are as follows:

[0006] 1. At present, the detection of welding quality mainly focuses on after welding, with low detection efficiency, high cost, high requirements for operators, and can only be used for spot checks of small-batch products.

[0007] 2. The increase in the number of sensors will inevitably cause a sharp increase in the dimensionality of the feature space, generating "big data" in the welding process. It is necessary to eliminate redundant information and noise data from the large-scale raw data and establish the connection between the effective information in the welding process and the defect categories.

[0008] 3. The least squares support vector machine has unique advantages but lacks means of feature simplification, and it is difficult to achieve the expected effect by relying on experience and trial and error for parameter selection. Quickly obtaining appropriate parameter values is of great significance for its classification and learning efficiency. Summary of the Invention

[0009] Aiming at the problems existing in the prior art, the present invention provides a method for identifying defects in the arc welding process based on multi-sensor information.

[0010] A method for identifying defects in the arc welding process based on multi-sensor information includes the following steps:

[0011] Step 1: Conduct multiple welding experiments through a multi-sensor information acquisition test platform, and collect and obtain the welding process current signal, temperature signal, and visual image signal.

[0012] Step 2: Obtain and record the weld profile features of all weld surfaces in the welding test, and classify the welded workpieces according to the weld morphology features after welding. They are classified into six weld types, and the category results of the forming quality of each weld surface are obtained.

[0013] Step 3: Use statistical methods to compress and statistically process the welding current signal data to extract features;

[0014] Use statistical methods to compress and statistically process the temperature signal data to extract features;

[0015] Use a convolutional neural network to process the visual image signal and extract visual image features;

[0016] Step 4: Based on the characteristic of a large amount of high-dimensional data generated by multi-information fusion in the welding process, use the mutual information method to perform feature selection and data compression on the data, and generate a reduced data set containing key features;

[0017] Step 5: According to the morphology type of each weld after welding, classify and label the welding process data of the preprocessed weld, and divide the data set into a training set, a validation set, and a test set;

[0018] Among them, the division ratio of the training set and the validation set is 1:1. At the same time, the test set uses the weld process data of welds different from those of the training set and the validation set to verify the generalization ability of the model;

[0019] Step 6: Use the least squares support vector machine model to train, validate, and test the current signal, temperature signal, and visual image signal obtained from the welding test. At the same time, use the particle swarm optimization algorithm to select the key parameters of the least squares support vector machine model to obtain the classification result of the weld surface forming quality.

[0020] Furthermore, use the least squares support vector machine algorithm to perform feature-level information fusion on the current signal, temperature signal, and visual image signal collected and obtained in Step 1.

[0021] Furthermore, in the six types of welds in Step 2, the six weld types include: good weld (Ⅰ), welding deviation (Ⅱ), incomplete fusion (Ⅲ), porosity (Ⅳ), lack of penetration (Ⅴ), burn-through (Ⅵ).

[0022] Furthermore, the welding current signal in Step 3 includes the following characteristic values:

[0023] Standard deviation, mean, root mean square, peak-to-peak value, peak factor, waveform factor, skewness, kurtosis.

[0024] Furthermore, the temperature signal in Step 3 includes the following characteristic values:

[0025] Maximum value, minimum value, peak-to-peak value, variance, standard deviation, mean, skewness, kurtosis, slope.

[0026] Furthermore, the visual image in Step 3 is processed using a convolutional neural network, including the following steps:

[0027] First, the input features are convolved through a convolutional layer to extract local features;

[0028] Then, the ReLU activation function is used to perform a non-linear transformation on the output of the convolutional layer;

[0029] Finally, the extracted features are sent to the average pooling layer to extract different feature information; specifically:

[0030] The convolution operation can be expressed as

[0031]

[0032] where Y i,j,d is the value of the output feature map Y at position (i, j) and channel d; X i+m,j+n,c is the value of the input data X at position (i + m, j + n) and channel c; K m,n,c,d is the value of the convolutional kernel K at position (m, n) and channel c corresponding to the output channel d; F is the size of the convolutional kernel; C is the number of channels;

[0033] The ReLU activation function can be expressed as

[0034] f(x) = max(0, x)

[0035] where x is the input value; f(x) is the output value;

[0036] The average pooling operation can be expressed as

[0037]

[0038] where Y i,j,c is the value of the output feature map at position (i, j) and channel c; X iS+m,jS+n,c is the value of the input feature map at position (iS + m, jS + n) and channel c; S is the stride; F is the size of the pooling window.

[0039] Furthermore, step 4 uses the mutual information method to perform feature selection and data compression on the data, including the following steps:

[0040] First, calculate the mutual information value I(X; Y) between each feature and the target variable;

[0041] Then, sort the features according to the calculated mutual information values. The higher the mutual information value, the stronger the correlation between the feature and the target variable;

[0042] Finally, according to the mutual information values, retain the features with higher mutual information values and eliminate the features with lower mutual information values; the calculation process of the mutual information value I(X; Y) is specifically:

[0043]

[0044] where P XY (x, y) is the joint probability distribution; P X (x) is the marginal probability distribution of variable X; P Y (y) is the marginal probability distribution of variable Y.

[0045] Furthermore, step 6 constructs a least squares support vector machine model based on the particle swarm optimization algorithm to improve the model classification performance; it includes the following steps:

[0046] (1) Randomly initialize the particle swarm according to the upper and lower bounds of the parameter combination (c, g);

[0047] (2) Use the generated (c, g) to perform cross-validation on the least squares support vector machine to obtain the fitness of each particle;

[0048] (3) Update the individual optimal position and the global optimal position of each particle;

[0049] (4) Update the velocity and position of each particle according to the historical optimal position of the particle and the global optimal position of the group;

[0050] (5) Calculate the fitness of the updated particle, and update the individual optimal position and the global optimal position according to the new fitness;

[0051] (6) Determine whether the algorithm meets the termination criterion. If not, return to step (3). Otherwise, output the current global optimal position as the optimal parameter combination (c, g) and establish a least squares support vector machine model based on the particle swarm optimization algorithm; specifically:

[0052] The specific least squares support vector machine is as follows:

[0053] The training data set can be expressed as

[0054]

[0055] where x i is the data vector; y i is the data category corresponding to x i ;

[0056] The optimal hyperplane equation can be expressed as

[0057] w·x + b = 0

[0058] where w is the plane normal vector; b is the plane offset;

[0059] The optimization problem of the least squares support vector machine can be expressed as

[0060]

[0061] where w is the weight vector; c is the penalty factor; e i is the error vector; is a function for mapping the input space to a high-dimensional space; b is the bias vector; n is the number of data points;

[0062] Using the Lagrange multiplier method, the least squares support vector machine model can be expressed as

[0063]

[0064] where α is the Lagrange multiplier;

[0065] Using a non-linear kernel function that satisfies the Mercer condition to calculate the inner product of high-dimensional vectors, the classification decision function can be expressed as

[0066]

[0067] where α * is the optimal solution; w * is the optimal weight; b * is the optimal bias; K(x i , x) is the non-linear kernel function;

[0068] The said K(x i , x) can be expressed as

[0069] K(xi, x) = exp(-g||xi - x|| 2 )

[0070] where g is the kernel parameter.

[0071] Another object of the present invention is to provide a defect monitoring system for the arc welding process based on multi-sensor information, including:

[0072] A welding work platform, an arc welding torch, a wire feeder, a shielding gas, a multi-information acquisition system;

[0073] The multi-information acquisition includes a current signal acquisition system, a temperature signal acquisition system, and an image acquisition system;

[0074] The current signal acquisition system includes a Hall current sensor, a data acquisition card, and a current signal detection system;

[0075] The temperature signal acquisition system includes an infrared temperature sensor, a data acquisition card, and a temperature signal detection system;

[0076] The image acquisition system includes a CCD camera, a data acquisition card, an auxiliary light source, and an image monitoring system;

[0077] Among them, the infrared temperature sensor and the CCD camera are installed at the end welding torch of the robot.

[0078] Combined with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by the present invention are as follows:

[0079] 1. An arc welding process defect monitoring system based on multi-sensor information in the welding process proposed by the present invention. This system is based on collecting current signals, temperature signals, and visual image signals during the welding process, and combines information theory and machine learning to achieve real-time identification of welding defects.

[0080] This system aims to solve problems such as the massive high-dimensional data volume generated by multi-sensor information fusion and the rapid calculation of appropriate model parameters in traditional welding defect detection methods. A welding defect identification method combining the mutual information method and the particle swarm optimization algorithm-least squares support vector machine is proposed. The multi-source sensing high-dimensional data is feature-selected by the mutual information method, and the particle swarm optimization algorithm-least squares support vector machine model is optimized and trained using the reduced data set. The classification model established by this method has the characteristics of low complexity, short training and prediction time, and strong generalization ability.

[0081] 2. Compared with the welding defect identification model established using a single sensor, the model based on current signals, temperature signals, and visual image signals has stronger identification and prediction capabilities. After performing feature-level fusion processing on these multi-source information, their complementarity and redundancy contribute to further improving the accuracy and reliability of the classification model prediction.

[0082] 3. Aiming at the welding big data problem brought by multi-sensor information fusion in the welding process, a feature reduction method based on mutual information value is introduced. The advantage of using the mutual information method is that it can effectively remove features with low correlation with the target variable while keeping the classification ability of the model unchanged, improving the performance and efficiency of the model.

[0083] 4. Selecting appropriate hyperparameters is a key issue in establishing a least squares support vector machine classification model. For this reason, a particle swarm optimization algorithm is proposed to optimize and train the least squares support vector machine model. Compared with the least squares support vector machine model established by the conventional grid search method, this algorithm can more effectively improve the classification performance of the model.

[0084] 5. A welding defect identification method combining the mutual information method and the particle swarm optimization algorithm-least squares support vector machine is proposed. The multi-source sensing high-dimensional data is feature-selected by the mutual information method, and the particle swarm optimization algorithm-least squares support vector machine model is optimized and trained using the reduced data set. The classification model established by this method has the characteristics of low complexity, short training and prediction time, and strong generalization ability. Description of the Drawings

[0085] Figure 1 It shows a schematic diagram of the basic process of an embodiment of the present invention;

[0086] Figure 2 It shows a detailed flowchart of step S104 in the embodiment of the present invention;

[0087] Figure 3 It shows a detailed flowchart of step S106 in the embodiment of the present invention. Specific embodiments

[0088] In order to make the objectives, technical solutions and advantages of the present invention more clear and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0089] For the specific embodiments of the arc welding process defect recognition method based on multi-sensor information, it can be described based on the above features and working principles. The following are the embodiments:

[0090] As Figure 1 shown, a kind of arc welding process defect monitoring system based on multi-sensor information provided by the embodiment of the present invention has the following monitoring steps:

[0091] S101: Conduct multiple welding experiments through a multi-sensor information acquisition test platform, and collect and obtain the welding process current signal, temperature signal, and visual image signal.

[0092] S102: Obtain and record the weld profile features of all weld surfaces in the welding test, and classify the welded workpieces according to the weld morphology features after welding. They are classified into six weld types, and the category results of the forming quality of each weld surface are obtained.

[0093] S103: Use statistical methods to compress and statistically process the welding current signal data to extract features;

[0094] Use statistical methods to compress and statistically process the temperature signal data to extract features;

[0095] Use a convolutional neural network to process the visual image signal and extract visual image features.

[0096] S104: Based on the characteristic of a large amount of high-dimensional data generated by multi-information fusion in the welding process, use the mutual information method to perform feature selection and data compression on the data, and generate a reduced data set containing key features.

[0097] S105: Classify and label the welding process data of each welded seam after preprocessing according to the morphology type of each welded seam after welding, and divide the data set into a training set, a validation set, and a test set;

[0098] Among them, the division ratio of the training set and the validation set is 1:1. At the same time, the test set uses the weld process data of welds different from those of the training set and the validation set to verify the generalization ability of the model.

[0099] S106: Use the least squares support vector machine model to train, validate, and test the current signal, temperature signal, and visual image signal obtained from the welding experiment. At the same time, use the particle swarm optimization algorithm to select the key parameters of the least squares support vector machine model to obtain the classification result of the weld surface forming quality.

[0100] The present invention provides, in an embodiment, a method for performing feature-level information fusion on the current signal, temperature signal, and visual image signal collected and obtained in S101 by using the least squares support vector machine algorithm.

[0101] The six types of welded seams provided in S102 of the present invention include: good weld (Ⅰ), weld deviation (Ⅱ), incomplete fusion (Ⅲ), porosity (Ⅳ), lack of penetration (Ⅴ), burn-through (Ⅵ). At the same time, multiple welding experiments with different process parameters are carried out for each type of welded seam.

[0102] The welding current signal provided in S103 of the present invention includes the following characteristic values:

[0103] Standard deviation, mean value, root mean square, peak-to-peak value, peak factor, waveform factor, skewness, kurtosis;

[0104] The basis for extracting the above characteristic values is that there is a correlation between the obtained welding current signal characteristics and the welding quality.

[0105] The temperature signal provided in S103 of the present invention includes the following characteristic values:

[0106] Maximum value, minimum value, peak-to-peak value, variance, standard deviation, mean value, skewness, kurtosis, slope;

[0107] The basis for extracting the above characteristic values is that there is a correlation between the workpiece temperature information and the welding quality.

[0108] The visual image provided in S103 of the present invention is processed by a convolutional neural network, including the following steps:

[0109] First, perform a convolution operation on the input features through a convolutional layer to extract local features;

[0110] Then, use the ReLU activation function to perform a non-linear transformation on the output of the convolutional layer;

[0111] Finally, the extracted features are sent to the average pooling layer to extract different feature information; specifically:

[0112] The convolution operation can be expressed as

[0113]

[0114] where Y i,j,d is the value of the output feature map Y at position (i, j) and channel d; X i+m,j+n,c is the value of the input data X at position (i + m, j + n) and channel c; K m,n,c,d is the value of the convolution kernel K at position (m, n) and channel c corresponding to the output channel d; F is the size of the convolution kernel; C is the number of channels;

[0115] The ReLU activation function can be expressed as

[0116] f(x) = max(0, x)

[0117] where x is the input value; f(x) is the output value;

[0118] The average pooling operation can be expressed as

[0119]

[0120] where Y i,j,c is the value of the output feature map at position (i, j) and channel c; X iS+m,jS+n,c is the value of the input feature map at position (iS + m, jS + n) and channel c; S is the stride; F is the size of the pooling window;

[0121] The main parameter settings of each layer of the convolutional neural network are as follows:

[0122] (1) Input layer: The size of the input image is 32×32×3, and the data is normalized with zero mean.

[0123] (2) First convolutional block: Convolution layer 1 uses 64 3×3 convolutional kernels, the stride is [1, 1], the padding method is same, the activation function is ReLU1, and the max pooling layer uses a 2×2 pooling kernel with a stride of [2, 2];

[0124] (3) Second convolutional block: Convolution layer 2 uses 64 3×3 convolutional kernels, the stride is [1, 1], the padding method is same, the activation function is ReLU2, and the max pooling layer uses a 2×2 pooling kernel with a stride of [2, 2];

[0125] (4) Third Convolution Block: Convolution Layer 3 uses 128 3×3 convolutional kernels, with a stride of [1,1], a padding method of same, and an activation function of ReLU3;

[0126] (5) Feature Convolution Block: Convolution Layer 4 uses 128 1×1 convolutional kernels, with a stride of [1,1], a padding method of same, and an activation function of ReLU4;

[0127] The advantages of using a convolutional neural network are

[0128] 1. A convolutional neural network can automatically learn features from raw data without the need for manual feature design, reducing the need for manual intervention and feature engineering;

[0129] 2. A convolutional neural network automatically extracts the spatial features of an image through convolutional layers, enabling it to extract a series of features from low-level to high-level features.

[0130] S104 provided by the embodiment of the present invention uses the mutual information method for feature selection and data compression of data, including the following steps:

[0131] First, calculate the mutual information value I(X;Y) between each feature and the target variable;

[0132] Then, sort the features according to the calculated mutual information value. The higher the mutual information value, the stronger the correlation between the feature and the target variable;

[0133] Finally, according to the mutual information value, retain the features with higher mutual information values and eliminate the features with lower mutual information values; The specific calculation process of the mutual information value I(X;Y) is:

[0134]

[0135] where P XY (x,y) is the joint probability distribution; P X (x) is the marginal probability distribution of variable X; P Y (y) is the marginal probability distribution of variable Y;

[0136] The advantage of using the mutual information method is that it can effectively remove features with low correlation with the target variable while maintaining the classification ability of the model unchanged, improving the performance and efficiency of the model.

[0137] S106 provided by the embodiment of the present invention constructs a least squares support vector machine model based on the particle swarm optimization algorithm to improve the model classification performance; including the following steps:

[0138] (1) Randomly initialize the particle swarm according to the upper and lower bounds of the parameter combination (c,g);

[0139] (2) Use the generated (c, g) to perform least squares support vector machine cross-validation to obtain the fitness of each particle;

[0140] (3) Update the individual optimal position and global optimal position of each particle;

[0141] (4) Update the velocity and position of each particle according to the historical optimal position of the particle and the global optimal position of the population;

[0142] (5) Calculate the fitness of the updated particle, and update the individual optimal position and global optimal position according to the new fitness;

[0143] (6) Determine whether the algorithm meets the termination criterion. If not, return to step (3). Otherwise, output the current global optimal position as the optimal parameter combination (c, g) and establish a least squares support vector machine model based on the particle swarm optimization algorithm; specifically:

[0144] The least squares support vector machine mentioned above is specifically:

[0145] The training dataset can be expressed as

[0146]

[0147] where x i is the data vector; y i is the data category corresponding to x i ;

[0148] The optimal hyperplane equation can be expressed as

[0149] w·x + b = 0

[0150] where w is the plane normal vector; b is the plane offset;

[0151] The optimization problem of the least squares support vector machine can be expressed as

[0152]

[0153] where w is the weight vector; c is the penalty factor; e i is the error vector; is the function used to map the input space to a high-dimensional space; b is the bias vector; n is the number of data points;

[0154] Using the Lagrange multiplier method, the least squares support vector machine model can be expressed as

[0155]

[0156] where α is the Lagrange multiplier;

[0157] The inner product of high-dimensional vectors is calculated using a non-linear kernel function that satisfies the Mercer condition, and the classification decision function can be expressed as

[0158]

[0159] where α * is the optimal solution; w * is the optimal weight; b * is the optimal bias; K(x i , x) is the non-linear kernel function;

[0160] The aforementioned K(x i , x) can be expressed as

[0161] K(xi, x) = exp(-g||xi - x|| 2 )

[0162] where g is the kernel parameter;

[0163] The advantage of using the particle swarm optimization algorithm is that it has characteristics such as a simple mechanism, high stability, strong optimization ability, and fast convergence speed;

[0164] The advantage of using the least squares support vector machine model is that it shows unique advantages in solving small-sample and non-linear classification scenarios, has strong generalization ability and robustness, and high computational efficiency;

[0165] The advantage of constructing a least squares support vector machine model based on the particle swarm optimization algorithm is that it can select the values of the penalty factor c and the kernel parameter g, which have a greater impact on the classification performance of the least squares support vector machine. The particle swarm optimization algorithm can quickly obtain appropriate values of c and g, thereby improving the classification ability and learning efficiency of the least squares support vector machine model.

[0166] An arc welding process defect monitoring system for an arc welding process defect recognition method based on multi-sensor information provided by an embodiment of the present invention, the arc welding process defect monitoring system based on multi-sensor information includes:

[0167] A welding work platform, an arc welding torch, a wire feeder, a shielding gas, and a multi-information acquisition system;

[0168] Multi-information acquisition includes a current signal acquisition system, a temperature signal acquisition system, and an image acquisition system;

[0169] The current signal acquisition system includes a Hall current sensor, a data acquisition card, and a current signal detection system;

[0170] The temperature signal acquisition system includes an infrared temperature sensor, a data acquisition card, and a temperature signal detection system;

[0171] The image acquisition system includes a CCD camera, a data acquisition card, an auxiliary light source, and an image monitoring system;

[0172] Among them, the infrared temperature sensor and the CCD camera are installed at the end of the robot welding torch;

[0173] Through the above-mentioned devices, the automatic acquisition, storage, and display of the molten pool current signal, temperature signal, and visual image signal during the welding process are realized.

[0174] The above are the specific implementation manners of the present invention, but the protection scope of the present invention should not be limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope defined by the claims.

Claims

1. A method for identifying defects in arc welding process based on multi-sensor information, characterized in that: The following steps are involved: Step 1: Conduct multiple welding experiments through a multi-sensor information acquisition test platform to collect and obtain welding process current signals, temperature signals, and visual image signals; Step 2: Obtain and record the weld profile features of all weld surfaces in the welding test, and classify the welded workpieces into six weld types based on the weld morphology features after welding, and obtain the category results of the forming quality of each weld surface; Step 3: Use statistical methods to compress and statistically process the welding current signal data to extract features; Statistical methods are used to compress temperature signal data and perform statistical processing to extract features; Use convolutional neural network to process visual image signals and extract visual image features; Step 4: Based on the large amount of high-dimensional data generated by multi-information fusion in the welding process, the mutual information method is used to perform feature selection and data compression on the data to generate a simplified data set containing key features; Step 5: According to the morphology type of each weld after welding, the welding process data of the preprocessed weld is classified and marked, and the data set is divided into a training set, a validation set, and a test set; The ratio of the training set to the validation set is 1:1, and the test set uses welding process data of welds different from those in the training set and validation set to verify the generalization ability of the model; Step 6: The least squares support vector machine model is used to train, verify and test the current signal, temperature signal and visual image signal obtained from the welding test. At the same time, the particle swarm optimization algorithm is used to select the key parameters of the least squares support vector machine model to obtain the category result to which the weld surface forming quality belongs.

2. The arc welding process defect identification method based on multi-sensor information according to claim 1 is characterized in that: The least squares support vector machine algorithm is used to perform feature-level information fusion on the current signal, temperature signal and visual image signal collected and obtained in step 1.

3. The arc welding process defect identification method based on multi-sensor information according to claim 1 is characterized in that: The six types of welds in step 2 include: good weld (I), off-center weld (II), incomplete weld (III), pores (IV), incomplete weld (V), and burn-through (VI).

4. The arc welding process defect identification method based on multi-sensor information according to claim 1 is characterized in that: The welding current signal in step 3 includes the following characteristic values: Standard deviation, mean, RMS, peak-to-peak value, crest factor, crest factor, skewness, kurtosis.

5. The arc welding process defect identification method based on multi-sensor information according to claim 1 is characterized in that: The temperature signal in step 3 includes the following characteristic values: Maximum, minimum, peak-to-peak, variance, standard deviation, mean, skewness, kurtosis, slope.

6. The arc welding process defect identification method based on multi-sensor information according to claim 1 is characterized in that: In step 3, the visual image is processed using a convolutional neural network.

7. The arc welding process defect identification method based on multi-sensor information according to claim 1 is characterized in that: The step 4 uses the mutual information method to perform feature selection and data compression on the data.

8. The arc welding process defect identification method based on multi-sensor information according to claim 1 is characterized in that: The step 6 constructs a least squares support vector machine model based on a particle swarm optimization algorithm to improve the model classification performance.

9. An arc welding process defect monitoring system implementing the arc welding process defect identification method based on multi-sensor information as claimed in any one of claims 1 to 8, characterized in that: The arc welding process defect monitoring system based on multi-sensor information includes: Welding work platform, arc welding gun, wire feeder, shielding gas, multi-information collection system; Multi-information acquisition includes current signal acquisition system, temperature signal acquisition system, and image acquisition system; The current signal acquisition system includes a Hall current sensor, a data acquisition card and a current signal detection system; The temperature signal acquisition system includes an infrared temperature sensor, a data acquisition card and a temperature signal detection system; The image acquisition system includes a CCD camera, a data acquisition card, an auxiliary light source and an image monitoring system; The infrared temperature sensor and CCD camera are installed on the welding gun at the end of the robot.

Images