CN113894390A - Pulse tungsten electrode argon arc welding penetration state detection side system, terminal and medium - Google Patents
Pulse tungsten electrode argon arc welding penetration state detection side system, terminal and medium Download PDFInfo
- Publication number
- CN113894390A CN113894390A CN202111201556.3A CN202111201556A CN113894390A CN 113894390 A CN113894390 A CN 113894390A CN 202111201556 A CN202111201556 A CN 202111201556A CN 113894390 A CN113894390 A CN 113894390A
- Authority
- CN
- China
- Prior art keywords
- penetration
- welding
- laser
- model
- training
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B23—MACHINE TOOLS; METAL-WORKING NOT OTHERWISE PROVIDED FOR
- B23K—SOLDERING OR UNSOLDERING; WELDING; CLADDING OR PLATING BY SOLDERING OR WELDING; CUTTING BY APPLYING HEAT LOCALLY, e.g. FLAME CUTTING; WORKING BY LASER BEAM
- B23K9/00—Arc welding or cutting
- B23K9/16—Arc welding or cutting making use of shielding gas
- B23K9/167—Arc welding or cutting making use of shielding gas and of a non-consumable electrode
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B23—MACHINE TOOLS; METAL-WORKING NOT OTHERWISE PROVIDED FOR
- B23K—SOLDERING OR UNSOLDERING; WELDING; CLADDING OR PLATING BY SOLDERING OR WELDING; CUTTING BY APPLYING HEAT LOCALLY, e.g. FLAME CUTTING; WORKING BY LASER BEAM
- B23K9/00—Arc welding or cutting
- B23K9/095—Monitoring or automatic control of welding parameters
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Mechanical Engineering (AREA)
- Plasma & Fusion (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computational Linguistics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Laser Beam Processing (AREA)
Abstract
The invention belongs to the technical field of image processing in welding technology, and discloses a pulse tungsten electrode argon arc welding penetration state detection side system, a terminal and a medium, wherein a laser reflection stripe image in a multi-welding process is obtained, and the obtained image is divided into training data, verification data and test data; constructing a deep neural network, optimizing the deep neural network, and defining model parameters; and training, verifying, testing and evaluating the constructed deep neural network model respectively by using the training data, the verifying data and the testing data, and carrying out welding penetration detection on the trained model network. The invention carries out detection classification based on the improved LeNet-5 deep neural network framework recognition learning training on the stripe images related to the penetration obtained by the laser vision method, the penetration recognition accuracy of the test images can reach 98.6 percent, the real-time regulation and control of the welding state can be realized, the manual participation degree is reduced, and the welding quality is improved.
Description
Technical Field
The invention belongs to the technical field of image processing in a welding technology, and particularly relates to a pulse tungsten electrode argon arc welding penetration state detection system, a terminal and a medium.
Background
At present, pulsed argon tungsten arc welding (P-GTAW) has a function of replacing in a material processing technology, and is one of the most widely applied welding methods in the manufacturing industry at present. The heat input is accurately controlled by using the current with periodic pulse change, the base current maintains the stable combustion of the electric arc during the condensation and melting of the molten pool, and the peak current melts the base material. Compared with other welding methods, the method has the advantages of high arc stability, accurate and controllable heat input, easiness in automation realization and the like, and is widely applied to backing welding of precise structural members such as aerospace, pressure vessels and the like.
However, the existing penetration state detection methods, including an arc voltage method, an arc sound method and a visual method, generally have the problems of low identification precision, poor detection effect, poor robustness, difficulty in realizing identification from a molten pool characteristic end to a penetration state end and the like; the generalization capability of the trained model is poor, and the practical value is lacked.
Through the above analysis, the problems and defects of the prior art are as follows: the arc voltage method has the problems that the identification precision is low, and the method cannot be applied to the identification of welding states under high-frequency pulses and alternating-current power supplies; the arc sound method has the problems that the arc sound method is easily interfered by the outside, and the correlation between the arc sound and the penetration is low; the visual method has a problem of lack of means for effectively extracting the characteristics of the molten pool.
The difficulty in solving the above problems and defects is: the three methods have the problem of poor adaptability, namely: each method is only suitable for one or a few welding methods, has high popularization difficulty, and cannot realize the identification from a characteristic end of a molten pool to a penetration state end.
The significance of solving the problems and the defects is as follows: the system disclosed by the patent uses a laser vision method to amplify the surface characteristics of the molten pool, so that the system is well adapted; and the deep neural network is used for feature recognition on the basis of the laser vision method, the response speed is high, the deployment difficulty and the cost are low, the automation degree of a welding device system is promoted, the manpower can be liberated, and the method has great significance in the aspects of improving the working environment quality of welders, improving the welding quality, saving energy and protecting environment.
Disclosure of Invention
Aiming at the problems in the prior art, the invention provides a pulse argon tungsten-arc welding penetration state detection system, a terminal and a medium, and particularly relates to a pulse argon tungsten-arc welding penetration state detection method based on a deep neural network.
The invention is realized in this way, a pulse tungsten argon arc welding penetration state detection method based on a deep neural network, which comprises the following steps:
acquiring laser reflection stripe images in multiple welding processes, and dividing the acquired images into training data, verification data and test data; constructing a deep neural network, optimizing the deep neural network, and defining model parameters; and training, verifying, testing and evaluating the constructed deep neural network model respectively by using the training data, the verifying data and the testing data, and carrying out welding penetration detection on the trained model network.
Further, the pulse argon tungsten-arc welding penetration state detection method based on the deep neural network comprises the following steps:
carrying out a plurality of welding experiments by utilizing a built molten pool oscillation frequency laser visual platform to obtain molten pool oscillation laser stripe images under different penetration states (non-penetration, critical penetration and full penetration) of non-consumable electrode gas shielded welding;
adding labels in three penetration states of non-penetration, critical penetration and full penetration into the obtained oscillation laser fringe image of the non-consumable electrode gas shielded welding molten pool, classifying the laser fringe images in the three different penetration states according to the added labels, and establishing a laser fringe image database in the different penetration states of the molten pool;
selecting reasonable LeNet-5 network parameters and training according to the constructed laser stripe image database of the molten pool in different penetration states, establishing a molten pool penetration state classification model, and obtaining labels and accuracy rates corresponding to the images after testing;
and step four, carrying out model deployment on the trained model, transmitting the model into a laser reflection stripe image in the welding process in real time through connection of high-speed camera shooting, and calculating real-time feedback fusion penetration information.
Further, the laser vision platform for sensing the molten pool oscillation reflection laser stripe image comprises a welding device and a laser vision detection device;
the welding device comprises a tungsten electrode argon arc welding gun, a welding power supply, a protective gas cylinder and a weldment;
the laser visual detection device comprises a laser, a laser fixing frame, an imaging screen, a high-speed camera and an image processing workstation.
Further, the laser visual inspection device includes:
the laser emits light beams, the five-line stripes and the single-line stripes are adjustable, and the included angle between the laser and the plane is 30-50 degrees;
the imaging screen and the high-speed camera are arranged on the other side of the laser and are used for extracting a light and shade change image of the reflected laser stripe; the high-speed camera is used for collecting the oscillation laser reflection stripes of the pulsed tungsten argon arc welding molten pool and feeding the oscillation laser reflection stripes back to the image processing workstation;
the image processing workstation comprises: the model deployment module is used for deploying the models; the model training module is used for substituting the divided training data sets into the initial model for training; and the model deployment module is used for performing trained model deployment and performing real-time calculation and feedback fusion penetration state on welding fusion penetration by inputting data flow in real time through high-speed camera shooting.
Further, the second step comprises:
dividing laser reflection stripe images shot by high-speed shooting into non-penetration, full-penetration and critical penetration according to actual penetration states, and adding corresponding labels to obtain a penetration information database; and (4) cutting all the images by using a conversion program, then disorganizing all the images, and packaging according to the proportion of 0.42, 0.28 and 0.30 to generate a training data set, a verification data set and a test data set.
Further, in the third step, according to the constructed molten pool penetration state laser stripe image database, selecting reasonable LeNet-5 network parameters and training, establishing a molten pool penetration state classification model, and obtaining labels and accuracy corresponding to the images after testing comprises:
(1) defining a parent folder path, a model parameter storage path, a test set picture path, a classified error picture storage path and three data set file paths;
(2) defining a LeNet-5 network structure, and optimizing and improving the LeNet-5 network structure; defining the size of data flow and calling parameters for each calling;
(3) defining a classified error picture storage rule; defining a loss function SoftmaxClossEntrol; defining training parameters adopted by the model and initializing the model;
(4) defining a training process, using a training data set and a verification data set to perform model network learning and verify network classification accuracy, using a test data set to perform test evaluation on a learning result, and storing trained model network parameters after training is finished;
(5) defining a processing method of classified error pictures, counting the test set images after the evaluation of the model network, independently classifying the classified error pictures, and determining the common points of the classified error pictures.
Further, in the step (2), the optimizing and improving the LeNet-5 network structure includes:
reducing the number of channels of the two convolution layers; reducing the number of fully connected layers to two while slightly increasing the output of the first fully connected layer and thereafter adding a discard layer; in addition to the output layer, each layer of activation function selects a ReLU function, and a Sigmoid activation function is added behind the output layer; and adding a batch normalization layer after each layer for normalization treatment.
Further, in the step (3), the model training hyper-parameters comprise a learning rate, a model training frequency and a data quantity used each time;
the per-use data quantity includes: the number of pictures used by the training set, the number of pictures used by the verification set and the number of pictures used by the test set.
Further, the evaluating learning results using the test dataset includes:
calculating the identification success rate of the fusion penetration state image classification model; the calculation formula is as follows: accuracy is the number of predicted results in the test set that are the same as the actual tags per test set size.
Further, in the fourth step, the trained model is deployed, and is transmitted into the laser reflection stripe image in the welding process in real time by connecting high-speed camera shooting, and the feedback penetration information is calculated, and the regulation and control of welding parameters include:
defining a model network parameter storage path, and storing the trained model network parameters;
when the molten pool oscillation frequency laser visual platform carries out a new round of experiment, calling the stored model network parameters, and connecting with a high-speed camera to carry out data stream acquisition; meanwhile, the input image is cut and preprocessed, and the size of the obtained picture is the same as that of the picture in the training data set;
substituting the data flow sheets into a model network for calculation, judging the penetration state of the data flow sheets, and adjusting the welding current, the welding voltage, the welding speed and other welding parameters in real time according to the penetration state of the data flow sheets.
The invention also aims to provide a pulse argon tungsten-arc welding penetration state detection system based on the deep neural network, which implements the pulse argon tungsten-arc welding penetration state detection method based on the deep neural network.
Another object of the present invention is to provide an information processing terminal, which includes a memory and a processor, wherein the memory stores a computer program, and the computer program, when executed by the processor, causes the processor to execute the method for detecting the penetration state of pulsed tungsten argon arc welding based on the deep neural network.
Another object of the present invention is to provide a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, the processor executes the method for detecting the penetration state of pulsed tungsten argon arc welding based on the deep neural network.
By combining all the technical schemes, the invention has the advantages and positive effects that:
the invention provides a method for identifying and classifying non-consumable electrode gas shielded welding molten pool surface reflection laser stripe images obtained by a laser vision method based on an improved LeNet-5 depth neural network, which is used for detecting and classifying the stripe images obtained by the laser vision method and related to penetration based on an improved LeNet-5 depth neural network framework after identification and learning training, wherein the penetration identification accuracy of a test image can reach 98.6%, and an input image is classified after a model is deployed, and the image identification accuracy is 98.2%. The system can realize real-time regulation and control of the welding state, reduce the manual participation and improve the welding quality.
Drawings
FIG. 1 is a schematic diagram of a pulse argon tungsten-arc welding penetration state detection method based on a deep neural network according to an embodiment of the present invention.
Fig. 2 is a flowchart of a pulse argon tungsten-arc welding penetration state detection method based on a deep neural network according to an embodiment of the present invention.
FIG. 3 is a schematic diagram of a laser visual platform of a molten pool oscillation frequency provided by an embodiment of the invention.
Fig. 4 is a schematic structural diagram of an image workstation according to an embodiment of the present invention.
FIG. 5(a) is a schematic illustration of laser reflection fringes on a surface of a non-gas metal arc welding in an unmelted condition according to an embodiment of the present invention.
Fig. 5(b) is a schematic diagram of laser reflection stripes on a surface of a full penetration condition non-gas metal arc welding according to an embodiment of the present invention.
FIG. 5(c) is a schematic diagram of laser reflection stripes on the surface of a Critical penetration condition non-gas metal arc welding according to an embodiment of the present invention.
Fig. 6 is a schematic structural diagram of a deep neural network model provided in an embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail with reference to the following embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
Aiming at the problems in the prior art, the invention provides a pulse argon tungsten-arc welding penetration state detection method based on a deep neural network, and the invention is described in detail below with reference to the accompanying drawings.
As shown in fig. 1, the method for detecting a penetration state of pulsed argon tungsten-arc welding based on a deep neural network according to an embodiment of the present invention includes:
acquiring laser reflection stripe images in multiple welding processes, and dividing the acquired images into training data, verification data and test data; constructing a deep neural network, optimizing the deep neural network, and defining model parameters; and training, verifying, testing and evaluating the constructed deep neural network model respectively by using the training data, the verifying data and the testing data, and carrying out welding penetration detection on the trained model network.
As shown in fig. 2, the method for detecting a penetration state of pulsed argon tungsten-arc welding based on a deep neural network according to an embodiment of the present invention includes the following steps:
s101, carrying out a plurality of welding experiments by utilizing the built molten pool oscillation frequency laser visual platform to obtain molten pool oscillation laser stripe images in different penetration states (non-penetration, critical penetration and full penetration) of non-consumable electrode gas shielded welding;
s102, adding labels in three penetration states of non-penetration, critical penetration and full penetration to the obtained oscillation laser fringe image of the non-consumable electrode gas shielded welding molten pool, classifying the laser fringe images in three different penetration states according to the added labels, and establishing a laser fringe image database in different penetration states of the molten pool;
s103, selecting reasonable LeNet-5 network parameters and training according to the constructed laser stripe image database of the molten pool in different penetration states, establishing a molten pool penetration state classification model, and obtaining labels and accuracy rates corresponding to the images after testing;
and S104, carrying out model deployment on the trained model, transmitting the model into a laser reflection stripe image in the welding process in real time through connection of high-speed camera shooting, and calculating real-time feedback fusion penetration information.
As shown in fig. 3, the laser vision platform for sensing the image of the molten pool oscillation reflection laser stripe provided by the embodiment of the present invention includes a welding device and a laser vision detection device;
the welding device comprises a tungsten electrode argon arc welding gun, a welding power supply, a protective gas cylinder and a weldment;
the laser visual detection device comprises a laser, a laser fixing frame, an imaging screen, a high-speed camera and an image processing workstation.
The laser visual detection device provided by the embodiment of the invention comprises:
the laser emits light beams, the five-line stripes and the single-line stripes can be adjusted, and the included angle between the laser and the plane is 30-50 degrees;
the imaging screen and the high-speed camera are arranged on the other side of the laser and used for extracting a light and shade change image of the reflected laser stripe. The high-speed camera is used for collecting the oscillation laser reflection stripes of the pulsed tungsten argon arc welding molten pool and feeding the oscillation laser reflection stripes back to the image processing workstation;
the image processing workstation according to the embodiment of the present invention shown in fig. 4 includes: the model deployment module is used for deploying the models; the model training module is used for substituting the divided training data sets into the initial model for training; and the model deployment module is used for performing trained model deployment and performing real-time calculation and feedback fusion penetration state on welding fusion penetration by inputting data flow in real time through high-speed camera shooting.
Step S102 provided in the embodiment of the present invention includes:
dividing laser reflection stripe images shot by high-speed shooting into non-penetration, full-penetration and critical penetration according to actual penetration states, and adding corresponding labels to obtain a penetration information database; and (4) cutting all the images by using a conversion program, then disorganizing all the images, and packaging according to the proportion of 0.42, 0.28 and 0.30 to generate a training data set, a verification data set and a test data set.
In step S103, selecting and training reasonable LeNet-5 network parameters according to the constructed molten pool penetration state laser fringe image database, and establishing a molten pool penetration state classification model, where the labels and the accuracy corresponding to the obtained images after the test include:
(1) defining a parent folder path, a model parameter storage path, a test set picture path, a classified error picture storage path and three data set file paths;
(2) defining a LeNet-5 network structure, and optimizing and improving the LeNet-5 network structure; defining the size of data flow and calling parameters for each calling;
(3) defining a classified error picture storage rule; defining a loss function SoftmaxClossEntrol; defining training parameters adopted by the model and initializing the model;
(4) defining a training process, using a training data set and a verification data set to perform model network learning and verify network classification accuracy, using a test data set to perform test evaluation on a learning result, and storing trained model network parameters after training is finished;
(5) defining a processing method of classified error pictures, counting the test set images after the evaluation of the model network, independently classifying the classified error pictures, and determining the common points of the classified error pictures.
In step (2), the optimization and improvement of the LeNet-5 network structure provided by the embodiment of the invention comprises:
reducing the number of channels of the two convolution layers; reducing the number of fully connected layers to two while slightly increasing the output of the first fully connected layer and thereafter adding a discard layer; in addition to the output layer, each layer of activation function selects a ReLU function, and a Sigmoid activation function is added behind the output layer; and adding a batch normalization layer after each layer for normalization treatment.
In the step (3), the model training hyper-parameters provided by the embodiment of the invention comprise learning rate, model training times and data quantity used each time;
the per-use data quantity includes: the number of pictures used by the training set, the number of pictures used by the verification set and the number of pictures used by the test set.
The evaluation of the learning result by using the test data set provided by the embodiment of the invention comprises the following steps:
calculating the identification success rate of the fusion penetration state image classification model; the calculation formula is as follows: accuracy is the number of predicted results in the test set that are the same as the actual tags per test set size.
The method for deploying the trained model, provided by the embodiment of the invention, transmits the laser reflection stripe image in the welding process in real time by connecting high-speed camera shooting, calculates feedback penetration information, and regulates and controls welding parameters, and comprises the following steps:
defining a model network parameter storage path, and storing the trained model network parameters;
when the molten pool oscillation frequency laser visual platform carries out a new round of experiment, calling the stored model network parameters, and connecting with a high-speed camera to carry out data stream acquisition; meanwhile, the input image is cut and preprocessed, and the size of the obtained picture is the same as that of the picture in the training data set;
substituting the data flow sheets into a model network for calculation, judging the penetration state of the data flow sheets, and adjusting the welding current, the welding voltage, the welding speed and other welding parameters in real time according to the penetration state of the data flow sheets.
The technical solution of the present invention is further described with reference to the following specific embodiments.
Example 1:
a non-consumable electrode gas shielded welding penetration state detection method based on a deep neural network comprises the following steps:
step S1: building a molten pool oscillation frequency laser visual platform; the experimental platform comprises a welding system and a laser visual detection system, wherein the welding system comprises a tungsten electrode argon arc welding gun, a welding power supply, a protective gas cylinder and a weldment; the laser visual detection system comprises a laser, a laser fixing frame, an imaging screen, a high-speed camera and an image processing workstation, wherein the high-speed camera is used for collecting the oscillation laser reflection stripes of the pulsed tungsten argon arc welding molten pool and feeding the oscillation laser reflection stripes back to the image processing workstation;
step S2: carrying out a plurality of welding experiments to obtain an oscillation laser stripe image of a non-consumable electrode gas shielded welding molten pool; all the obtained image data are transmitted to a model training module in the image processing workstation to go to the subsequent steps S3, S4 and S5;
step S3: adding a penetration state label, classifying the image according to the label, and establishing a molten pool penetration state laser stripe image database;
step S4: and (5) selecting reasonable network parameters for the LeNet-5 network according to the database in the step S3, training, establishing a molten pool fusion penetration state classification model, and testing to obtain a label and accuracy corresponding to the image.
Step S5: and deploying the trained model in the S4, transmitting the model into a laser reflection stripe image in the welding process in real time by connecting high-speed camera shooting, calculating feedback fusion penetration information, and regulating and controlling welding parameters.
The non-consumable electrode gas shielded welding adopts pulse current, peak current provides external excitation force to promote molten pool oscillation, molten pool oscillation is violent in initial stage of base current moment, and molten pool oscillation frequency can be collected.
The axes of the laser and the welding gun are positioned on the same plane, the laser emits light beams with adjustable five-line stripes and single-line stripes, and the included angle between the laser and the plane is 30-50 degrees.
The imaging screen and the high-speed camera are arranged on the other side of the laser and used for extracting light and shade change images of the reflected laser stripes.
The image processing workstation in step S2 is characterized by a model training module and a model deployment module, wherein the model training module is used to train the created data set into the initial model, and the model deployment module is used to deploy the trained model, and the weld penetration is calculated and fed back in real time by inputting data flow in real time through high-speed camera shooting.
In step S3, labeling the laser reflection stripe image captured by high-speed imaging, and summarizing the images in a folder form in three types of non-penetration/full-penetration/critical-penetration to serve as a penetration information database; the folders are named by Critical (label is 0), Partial (label is 1) and Full (label is 2). Classifying all pictures into three folders according to the actual penetration state reflected by the pictures, wherein each picture only has one label, then properly cutting all the pictures by using a conversion program, then disordering all the pictures, packaging the pictures according to the proportion of 0.42, 0.28 and 0.30 to generate and store a training data set, a verification data set and a test data set.
The laser reflection stripe on the surface of the non-consumable electrode gas shielded welding is characterized in that: under the condition of no penetration, at the pulse base value stage, the electric arc force is weakened, the molten pool rebounds to generate oscillation, and the surface oscillation is mainly longitudinal waves due to the lifting effect of solid metal at the bottom of the molten pool and the shallow fusion depth, so that the amplitude is small and regular, and the laser reflection stripe patterns are concentrated, high in definition and symmetrical; under the condition of full penetration, the back of a welding seam is completely penetrated, so that the volume of a molten pool is larger, more liquid metal is contained, the bottom solid metal support is lost, the oscillation form is transverse wave, the amplitude is larger, the amplitude is reflected in a laser reflection stripe pattern, the reflection stripe distribution is wider, the pattern is more dispersed, and the definition of a reflection image is low; under the critical penetration condition, the molten pool at the back of the welding seam is smaller, so that the molten pool oscillation in the critical penetration state has the characteristics of non-penetration and full penetration at the same time, the oscillation mode has two frequencies of one higher frequency and one lower frequency of longitudinal wave and transverse wave at the same time, and the oscillation mode is characterized in that a stripe image forms a similar 8-shaped graph in a laser reflection stripe image on the surface of the molten pool. As shown in fig. 5.
Wherein, the step S4 specifically includes the following steps:
step S4.1: defining a data path, a model parameter storage path, a test set picture path, a classified error picture storage path and three data set file paths;
step S4.2: defining a LeNet-5 network structure, and optimizing and improving a framework of the LeNet-5 network structure;
step S4.3: defining the size of data flow and calling parameters for each calling;
step S4.4: defining a classified error picture storage rule;
step S4.5: defining a loss function SoftmaxClossEntrol;
step S4.6: defining training parameters adopted by the model and initializing the model;
step S4.7: defining a training process, performing network learning and testing by using a training data set and a verification data set, finally evaluating a learning result by using a testing data set, and storing trained network parameters after training is finished;
step S4.8: defining a processing method of classified error pictures, counting the test set pictures after the network is evaluated, and individually classifying the classified error pictures to evaluate the rationality of the network structure.
Wherein, the step S5 specifically includes the following steps:
step S5.1: defining a model network parameter storage path, and storing the trained model network parameters;
step S5.2: when a new round of experiment is carried out on the experiment table, the stored model network parameters are called and connected with the high-speed camera to carry out data stream acquisition;
step S5.3: substituting the data flow sheets into a model network for calculation, judging the penetration state of the data flow sheets, and adjusting welding parameters in real time according to the penetration state of the data flow sheets.
In step S4.2, in order to adapt to the characteristics of the identified object and prevent overfitting, the LeNet-5 network is improved by reducing the number of channels of two convolutional layers; reducing the number of fully connected layers from three to two but slightly increasing the output of the first fully connected layer and thereafter adding a discard layer to reduce overfitting; except for the output layer, each layer of activation function is changed into a ReLU function from a Sigmoid function, and a Sigmoid activation function is added to the output layer; and adding batch normalization layer normalization after each layer.
In step S4.6, the model training parameters include a Learning rate (Learning _ rate ═ 0.3), a model training number of times (Epoch ═ 10), a number of pieces of usage data per time (training set usage picture number Batch _ size _ train ═ 1260, verification set usage picture number Batch _ size _ value ═ 840, and test set usage picture number Batch _ size _ test ═ 900).
In step S4.7, the identification success rate of the penetration state image classification model is calculated, and the calculation formula is that accuacy is the number of prediction results in the test set that is the same as the number of actual tags/the size of the test set.
In the step 4.8, the obtained images are evaluated, and common points of the images are searched, so that a direction is provided for improvement of the structure of the network model in the next step.
In step S5.2, the input image needs to be cut and preprocessed, and the size of the obtained picture is the same as the size of the picture in the training data set.
In the step S5.3, welding parameters including welding current, welding voltage and welding speed are adjusted for the obtained real-time penetration state of the welding.
The positive effects of the present invention are further described below in conjunction with specific experimental data.
In the experiment, the deep neural network trained by the training data set and the verification data set is verified by using the pictures of the test data set, wherein the test data set comprises 900 pictures in total, 12 pictures with wrong classification are obtained after the test, and the accuracy rate of the neural network reaches 98.6%. And then deploying the model, inputting 1800 pictures in total, classifying by using the trained neural network, and manually counting the classified error pictures. Wherein, the critical penetration state pictures are divided into 8 pieces of non-penetration groups, the critical penetration state pictures are divided into 10 pieces of full-penetration groups, the other two state pictures are divided into 15 pieces of critical penetration groups, and the image identification accuracy rate is 98.2%.
It should be noted that the embodiments of the present invention can be realized by hardware, software, or a combination of software and hardware. The hardware portion may be implemented using dedicated logic; the software portions may be stored in a memory and executed by a suitable instruction execution system, such as a microprocessor or specially designed hardware. Those skilled in the art will appreciate that the apparatus and methods described above may be implemented using computer executable instructions and/or embodied in processor control code, such code being provided on a carrier medium such as a disk, CD-or DVD-ROM, programmable memory such as read only memory (firmware), or a data carrier such as an optical or electronic signal carrier, for example. The apparatus and its modules of the present invention may be implemented by hardware circuits such as very large scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of hardware circuits and software, e.g., firmware.
The above description is only for the purpose of illustrating the present invention and the appended claims are not to be construed as limiting the scope of the invention, which is intended to cover all modifications, equivalents and improvements that are within the spirit and scope of the invention as defined by the appended claims.
Claims (10)
1. A pulse argon tungsten-arc welding penetration state detection method based on a deep neural network is characterized by comprising the following steps of:
acquiring laser reflection stripe images in multiple welding processes, and dividing the acquired images into training data, verification data and test data; constructing a deep neural network, optimizing the deep neural network, and defining model parameters; and training, verifying, testing and evaluating the constructed deep neural network model respectively by using the training data, the verifying data and the testing data, and carrying out welding penetration detection on the trained model network.
2. The method for detecting the penetration state of the pulsed argon tungsten-arc welding based on the deep neural network as claimed in claim 1, wherein the method for detecting the penetration state of the pulsed argon tungsten-arc welding based on the deep neural network comprises the following steps:
carrying out a plurality of welding experiments by utilizing a built molten pool oscillation frequency laser visual platform to obtain molten pool oscillation laser stripe images in different penetration states of non-consumable electrode gas shielded welding;
adding labels in three penetration states of non-penetration, critical penetration and full penetration into the obtained oscillation laser fringe image of the non-consumable electrode gas shielded welding molten pool, classifying the laser fringe images in the three different penetration states according to the added labels, and establishing a laser fringe image database in the different penetration states of the molten pool;
selecting reasonable LeNet-5 network parameters and training according to the constructed laser stripe image database of the molten pool in different penetration states, establishing a molten pool penetration state classification model, and obtaining labels and accuracy rates corresponding to the images after testing;
and step four, carrying out model deployment on the trained model, transmitting the model into a laser reflection stripe image in the welding process in real time through connection of high-speed camera shooting, and calculating real-time feedback fusion penetration information.
3. The deep neural network-based pulsed argon tungsten-arc welding penetration state detection method as claimed in claim 2, wherein the laser vision platform for sensing the image of the molten pool oscillation reflection laser stripe comprises a welding device and a laser vision detection device;
the welding device comprises a tungsten electrode argon arc welding gun, a welding power supply, a protective gas cylinder and a weldment;
the laser visual detection device comprises a laser, a laser fixing frame, an imaging screen, a high-speed camera and an image processing workstation.
4. The deep neural network-based pulsed argon tungsten-arc welding penetration state detection method according to claim 3, wherein the laser visual detection device comprises:
the laser emits light beams with adjustable five-line stripes, and the included angle between the laser and the plane is 30-50 degrees;
the imaging screen and the high-speed camera are arranged on the other side of the laser and are used for extracting a light and shade change image of the reflected laser stripe; the high-speed camera is used for collecting the oscillation laser reflection stripes of the pulsed tungsten argon arc welding molten pool and feeding the oscillation laser reflection stripes back to the image processing workstation;
the image processing workstation comprises: the model deployment module is used for deploying the models; the model training module is used for substituting the divided training data sets into the initial model for training; and the model deployment module is used for performing trained model deployment and performing real-time calculation and feedback fusion penetration state on welding fusion penetration by inputting data flow in real time through high-speed camera shooting.
5. The deep neural network-based pulsed argon tungsten-arc welding penetration state detection method as claimed in claim 2, wherein the second step comprises:
dividing laser reflection stripe images shot by high-speed shooting into non-penetration, full-penetration and critical penetration according to actual penetration states, and adding corresponding labels to obtain a penetration information database; and (4) cutting all the images by using a conversion program, then disorganizing all the images, and packaging according to the proportion of 0.42, 0.28 and 0.30 to generate a training data set, a verification data set and a test data set.
6. The deep neural network-based pulsed argon tungsten-arc welding penetration state detection method according to claim 2, wherein in step three, according to a constructed molten pool penetration state laser fringe image database, reasonable LeNet-5 network parameters are selected and trained, a molten pool penetration state classification model is established, and obtaining labels and accuracy corresponding to images after testing comprises:
(1) defining a parent folder path, a model parameter storage path, a test set picture path, a classified error picture storage path and three data set file paths;
(2) defining a LeNet-5 network structure, and optimizing and improving the LeNet-5 network structure; defining the size of data flow and calling parameters for each calling;
(3) defining a classified error picture storage rule; defining a loss function SoftmaxClossEntrol; defining training parameters adopted by the model and initializing the model;
(4) defining a training process, using a training data set and a verification data set to perform model network learning and verify network classification accuracy, using a test data set to perform test evaluation on a learning result, and storing trained model network parameters after training is finished;
(5) defining a processing method of classified error pictures, counting the test set images after the evaluation of the model network, independently classifying the classified error images, and determining common points of the classified error pictures;
the method for detecting the penetration state of the pulsed argon tungsten-arc welding based on the deep neural network is characterized in that in the step (2), the optimization and improvement of the LeNet-5 network structure comprises the following steps:
reducing the number of channels of the two convolution layers; reducing the number of fully connected layers to two while slightly increasing the output of the first fully connected layer and thereafter adding a discard layer; in addition to the output layer, each layer of activation function selects a ReLU function, and a Sigmoid activation function is added behind the output layer; adding a batch normalization layer behind each layer for normalization treatment;
in the step (3), the model training hyper-parameters comprise learning rate, model training times and data quantity used each time;
the per-use data quantity includes: the number of pictures used by a training set, the number of pictures used by a verification set and the number of pictures used by a test set;
the evaluating learning results using the test dataset comprises:
calculating the identification success rate of the fusion penetration state image classification model; the calculation formula is as follows: accuracy is the number of predicted results in the test set that are the same as the actual tags per test set size.
7. The deep neural network-based pulsed argon tungsten-arc welding penetration state detection method according to claim 2, wherein in the fourth step, the trained model is deployed, and feedback penetration information is calculated by connecting a high-speed camera and transmitting a laser reflection fringe image in a welding process in real time, and the adjusting and controlling of the welding parameters comprises:
defining a model network parameter storage path, and storing the trained model network parameters;
when the molten pool oscillation frequency laser visual platform carries out a new round of experiment, calling the stored model network parameters, and connecting with a high-speed camera to carry out data stream acquisition; meanwhile, the input image is cut and preprocessed, and the size of the obtained picture is the same as that of the picture in the training data set;
substituting the data flow sheets into a model network for calculation, judging the penetration state of the data flow sheets, and adjusting the welding current, the welding voltage, the welding speed and other welding parameters in real time according to the penetration state of the data flow sheets.
8. A pulse argon tungsten-arc welding penetration state detection system based on a deep neural network is used for implementing the pulse argon tungsten-arc welding penetration state detection method based on the deep neural network according to any one of claims 1-6.
9. An information processing terminal, characterized in that the information processing terminal comprises a memory and a processor, the memory stores a computer program, and the computer program is executed by the processor, so that the processor executes the deep neural network based pulsed tungsten argon arc welding penetration state detection method according to any one of claims 1 to 6.
10. A computer-readable storage medium storing a computer program, which when executed by a processor, causes the processor to execute the method for detecting the penetration state of pulsed argon tungsten arc welding based on the deep neural network according to any one of claims 1 to 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111201556.3A CN113894390B (en) | 2021-10-15 | 2021-10-15 | Pulse tungsten electrode argon arc welding penetration state detection side system, terminal and medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111201556.3A CN113894390B (en) | 2021-10-15 | 2021-10-15 | Pulse tungsten electrode argon arc welding penetration state detection side system, terminal and medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113894390A true CN113894390A (en) | 2022-01-07 |
| CN113894390B CN113894390B (en) | 2023-03-24 |
Family
ID=79192170
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111201556.3A Active CN113894390B (en) | 2021-10-15 | 2021-10-15 | Pulse tungsten electrode argon arc welding penetration state detection side system, terminal and medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113894390B (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115121895A (en) * | 2022-04-28 | 2022-09-30 | 广东省威汇智能科技有限公司 | Selective wave-soldering early warning method and device based on deep learning and storage medium |
| CN116038077A (en) * | 2022-12-30 | 2023-05-02 | 深圳市麦格米特焊接技术有限公司 | Gas shielded welding system, control method thereof, controller and storage medium |
| CN117464182A (en) * | 2023-12-21 | 2024-01-30 | 哈尔滨工业大学（威海） | Laser welding state discrimination method based on convolutional neural network |
| CN118371866A (en) * | 2024-06-27 | 2024-07-23 | 苏州松德激光科技有限公司 | Intelligent correction method and system for welding position |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH11309576A (en) * | 1998-04-28 | 1999-11-09 | Hitachi Ltd | Automatic welding equipment |
| CN108856978A (en) * | 2018-08-20 | 2018-11-23 | 南京理工大学 | Corner joint penetration control method based near infrared binocular visual identity |
| CN110935983A (en) * | 2018-09-21 | 2020-03-31 | 天津大学 | Method for controlling welding penetration by utilizing reflected laser stripe image |
| CN112215834A (en) * | 2020-10-22 | 2021-01-12 | 深圳大学 | Power battery welding quality detection method and system based on convolutional neural network |
| CN112894101A (en) * | 2019-12-03 | 2021-06-04 | 山东大学 | Perforated plasma arc welding molten pool state prediction method based on deep learning algorithm |
| CN112975071A (en) * | 2021-03-02 | 2021-06-18 | 兰州理工大学 | Device and method for detecting oscillation frequency of molten pool aiming at GMAW |
-
2021
- 2021-10-15 CN CN202111201556.3A patent/CN113894390B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH11309576A (en) * | 1998-04-28 | 1999-11-09 | Hitachi Ltd | Automatic welding equipment |
| CN108856978A (en) * | 2018-08-20 | 2018-11-23 | 南京理工大学 | Corner joint penetration control method based near infrared binocular visual identity |
| CN110935983A (en) * | 2018-09-21 | 2020-03-31 | 天津大学 | Method for controlling welding penetration by utilizing reflected laser stripe image |
| CN112894101A (en) * | 2019-12-03 | 2021-06-04 | 山东大学 | Perforated plasma arc welding molten pool state prediction method based on deep learning algorithm |
| CN112215834A (en) * | 2020-10-22 | 2021-01-12 | 深圳大学 | Power battery welding quality detection method and system based on convolutional neural network |
| CN112975071A (en) * | 2021-03-02 | 2021-06-18 | 兰州理工大学 | Device and method for detecting oscillation frequency of molten pool aiming at GMAW |
Non-Patent Citations (2)
| Title |
|---|
| 林俊 等: "电弧焊熔池表征与熔透状态映射研究", 《焊接》 * |
| 石玗 等: "基于激光视觉的脉冲GTAW熔池振荡检测与分析", 《机械工程学报》 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115121895A (en) * | 2022-04-28 | 2022-09-30 | 广东省威汇智能科技有限公司 | Selective wave-soldering early warning method and device based on deep learning and storage medium |
| CN116038077A (en) * | 2022-12-30 | 2023-05-02 | 深圳市麦格米特焊接技术有限公司 | Gas shielded welding system, control method thereof, controller and storage medium |
| CN116038077B (en) * | 2022-12-30 | 2024-03-12 | 深圳市麦格米特焊接技术有限公司 | Gas shielded welding system, control method thereof, controller and storage medium |
| CN117464182A (en) * | 2023-12-21 | 2024-01-30 | 哈尔滨工业大学（威海） | Laser welding state discrimination method based on convolutional neural network |
| CN118371866A (en) * | 2024-06-27 | 2024-07-23 | 苏州松德激光科技有限公司 | Intelligent correction method and system for welding position |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113894390B (en) | 2023-03-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113894390B (en) | Pulse tungsten electrode argon arc welding penetration state detection side system, terminal and medium | |
| Wang et al. | Intelligent welding system technologies: State-of-the-art review and perspectives | |
| CN111014879B (en) | Automatic welding method for corrugated plate of robot based on laser weld seam tracking | |
| Gu et al. | Autonomous seam acquisition and tracking system for multi-pass welding based on vision sensor | |
| He et al. | Autonomous detection of weld seam profiles via a model of saliency-based visual attention for robotic arc welding | |
| CN110102886A (en) | A kind of Intelligent welding system applied to metal structure | |
| Chen et al. | Prediction of weld bead geometry of MAG welding based on XGBoost algorithm | |
| US20230191540A1 (en) | Method, device and computer program for determining the performance of a welding method via digital processing of an image of the welded workpiece | |
| Zhang et al. | Adaptive intelligent welding manufacturing | |
| Aviles-Viñas et al. | On-line learning of welding bead geometry in industrial robots | |
| CN109940305A (en) | Laser Deep Penetration Welding intelligent monitor system and method based on plasma electric signal | |
| CN103218665A (en) | Real-time detection and weld quality prediction in vibration welding process | |
| CN110441329B (en) | Laser welding defect identification method, device and equipment based on deep learning | |
| CN111738369A (en) | Weld penetration state and penetration depth real-time prediction method based on visual characteristics of molten pool | |
| CA3158524A1 (en) | Systems and methods for analyzing weld quality | |
| CN115464263A (en) | Automatic tracking method, detection method and device for laser welding seam | |
| CN206567687U (en) | Detect the penetration control device of frequency of oscillation in a kind of pulse laser exciting TIG molten baths | |
| CN114842311A (en) | Welding penetration state monitoring method and device based on multi-dimensional fusion sensing | |
| CN115255731A (en) | Welding quality on-line detection welding seam marking device and method | |
| CN114581368A (en) | Bar welding method and device based on binocular vision | |
| Schmitz et al. | Enabling rewards for reinforcement learning in laser beam welding processes through deep learning | |
| CN115630512A (en) | Swing laser welding equivalent heat source modeling method and swing laser welding simulation method | |
| US12005532B2 (en) | Systems, methods, and apparatuses, of an arc welding (AW) process and quality monitoring | |
| Li et al. | Multi-layer and Multi-channel Dynamic Routing Planning and Initial Point Positioning of Weld Seam Based on Machine Vision | |
| CN112809130A (en) | Intelligent welding seam detection and trajectory planning method and system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |