Welding arc defect recognition method and system based on deep learning

Abstract

The present application relates to the technical field of welding, in particular to a welding arc defect identification method and system based on deep learning, the method comprising: acquiring basic data of a welding device, wherein the basic data comprises: device model, shielding gas, welding wire material, welding wire diameter, presence or absence of pulse, preset voltage and preset current; creating a waveform file for recording actual voltage and actual current in the welding process using the basic data; during the welding process of the welding device, collecting the actual voltage and actual current of the welding and recording them in the waveform file; converting the recorded waveform file into graph data; and inputting the graph data into a pre-trained welding defect detection model to obtain the welding defect type. The voltage and current data are less affected by external interference during the collection process, data acquisition is more convenient and accurate, and there is no need to increase detection equipment and detection procedures, thereby reducing detection cost and improving detection efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of welding, in particular to a welding arc defect identification method and system based on deep learning. BACKGROUND

[0002] At present, common welding quality detection methods include X-ray detection, ultrasonic detection, pressure detection, etc. The above detection methods need to purchase corresponding detection equipment and increase corresponding detection procedures to realize, which increases the cost of metal processing manufacturers and reduces the production capacity.

[0003] With the wide application of artificial intelligence algorithms represented by deep learning, the known welding quality detection methods currently include welding audio + big data analysis, weld photographing + artificial intelligence visual identification, etc. However, due to the complex processing site environment, the on-site noise pollution and light pollution have a great influence on the collection of audio data and image data, thereby affecting the identification accuracy. SUMMARY

[0004] In view of the above problems, the present application provides a welding arc defect identification method and system based on deep learning.

[0005] In one aspect of the present application, a welding arc defect identification method based on deep learning is provided, comprising:

[0006] Obtaining the basic data of the welding equipment, wherein the basic data includes: equipment model, protective gas, welding wire material, welding wire diameter, pulse presence or absence, preset voltage and preset current;

[0007] Creating a waveform file for recording the actual voltage and actual current in the welding process by using the basic data;

[0008] During the welding process of the welding equipment, the actual voltage and actual current of the welding are collected and recorded in the waveform file;

[0009] Converting the recorded waveform file into graph data;

[0010] Inputting the graph data into a pre-trained welding defect detection model to obtain the welding defect type.

[0011] Optionally, the process of converting the recorded waveform file into graph data comprises:

[0012] Reading the actual voltage and actual current recorded in the waveform file, and screening out the data in which the current is zero and the voltage is the no-load voltage;

[0013] For the screened data, a two-dimensional array of behavior voltage and column current is constructed;

[0014] The two-dimensional array constructed is converted into graph data by using a python graph processing method.

[0015] Optionally, the training process of the welding defect detection model comprises:

[0016] Sample data is acquired, the sample data comprising waveform files and actual corresponding welding defect types, the welding defect types comprising no defect, burn-through, arc breaking, arc drawing and weld snake shape;

[0017] The waveform files in the sample data are converted into graph data and input into an initial neural network model to obtain an estimated welding defect type;

[0018] The estimated welding defect type is compared with the corresponding welding defect type in the sample data, in the case that the estimated welding defect type does not meet the expectation, parameters in the initial neural network model are adjusted and the data quantity of the sample data is increased, the step of converting the waveform files in the sample data into graph data and inputting the graph data into the initial neural network model to obtain the estimated welding defect type is returned until the estimated welding defect type meets the expectation, and the current initial neural network model is taken as the welding defect detection model.

[0019] Optionally, the process of collecting the actual voltage and the actual current of welding comprises:

[0020] When the actual voltage and / or the actual current detected by the mutual inductor is greater than a predetermined threshold, the actual voltage and the actual current are recorded.

[0021] In still another aspect of the embodiment of the present application, a welding arc defect recognition system based on deep learning is also provided, which is characterized by comprising:

[0022] A data acquisition unit is configured to acquire basic data of a welding device, wherein the basic data comprises: device model, shielding gas, welding wire material, welding wire diameter, pulse presence or absence, preset voltage and preset current;

[0023] A file creation unit is configured to create a waveform file for recording the actual voltage and the actual current in a welding process by using the basic data;

[0024] A data collection unit is configured to collect the actual voltage and the actual current of welding and record them in the waveform file during the welding process of the welding device;

[0025] A data calculation unit is configured to convert the recorded waveform file into graph data;

[0026] A data transmission unit is configured to input the graph data into a pre-trained welding defect detection model to obtain a welding defect type.

[0027] Optionally, the data calculation unit is specifically configured to:

[0028] read the actual voltage and the actual current recorded in the waveform file, and screen out data in which the current is zero and the voltage is the no-load voltage from the read actual voltage and actual current;

[0029] For the screened data, a two-dimensional array of behavior voltage and column current is constructed.

[0030] The constructed two-dimensional array is converted into graph data using a python graph processing method.

[0031] Optionally, it further comprises a model training unit for

[0032] Obtain sample data, which includes waveform files and actual corresponding welding defect types, and the welding defect types include no defect, burn-through, arc break, arc draw and weld seam snake;

[0033] After the waveform files in the sample data are converted into graph data, input the initial neural network model to obtain an estimated welding defect type.

[0034] Compare the estimated welding defect type with the corresponding welding defect type in the sample data, and in the case that the estimated welding defect type does not meet the expectation, adjust the parameters in the initial neural network model and increase the data amount of the sample data, return to the step of converting the waveform files in the sample data into graph data and inputting the initial neural network model to obtain the estimated welding defect type, until the obtained estimated welding defect type meets the expectation, and the current initial neural network model is taken as a welding defect detection model.

[0035] Optionally, the data acquisition unit is specifically configured to start recording the actual voltage and the actual current when the actual voltage and / or the actual current detected by the mutual inductor is greater than a predetermined threshold.

[0036] Compared with the prior art, the voltage and current data are less affected by external interference during the acquisition process, the data acquisition is more convenient and accurate, and the detection cost is reduced without increasing the detection equipment and detection process, thereby improving the detection efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0037] The accompanying drawings, which are included to provide a further understanding of the application and constitute a part of this application, illustrate embodiments of the application and together with the description serve to explain the application. In the drawings:

[0038] Figure 1 A flowchart of a welding arc defect recognition method based on deep learning provided for the embodiment of the application;

[0039] Figure 2 A graph data conversion process diagram provided for the embodiment of the application;

[0040] Figure 3A model training process schematic diagram provided for an embodiment of the present application;

[0041] Figure 4 A structure schematic diagram of a welding arc defect recognition system based on deep learning provided for an embodiment of the present application;

[0042] Figure 5 A recognition system architecture schematic diagram in a specific implementation of the present application;

[0043] Figure 6 A structure schematic diagram of a waveform data acquisition and calculation device in a specific implementation of the present application. DETAILED DESCRIPTION

[0044] To make the objectives, technical solutions and advantages of the present application clearer, further detailed description will be made to the present application with reference to the embodiments and drawings. Herein, the illustrative embodiments of the present application and the description thereof are used to explain the present application, but not as a limitation to the present application.

[0045] Referring to Figure 1 The embodiment of the present application provides a welding arc defect recognition method based on deep learning, which comprises:

[0046] S100, acquiring basic data of a welding equipment.

[0047] The basic data comprises: equipment model, shielding gas, welding wire material, welding wire diameter, pulse presence or absence, preset voltage and preset current.

[0048] S110, creating a waveform file for recording actual voltage and actual current in a welding process by using the basic data.

[0049] In the implementation, the waveform file can be a file with an extension of.csv, and the name of the waveform file is named by using the basic data, and the naming format is: model-shielding gas-welding wire material-welding wire diameter-pulse presence or absence-pre-set current-pre-set voltage.csv.

[0050] S120, during the welding process of the welding equipment, collecting actual voltage and actual current of welding and recording in the waveform file.

[0051] In the implementation, a mutual inductor is commonly used to collect voltage and current in the current technical background; specifically, when the actual voltage and / or actual current detected by the mutual inductor is greater than a predetermined threshold, the actual voltage and actual current are recorded.

[0052] S130, converting the recorded waveform file into graph data.

[0053] In implementation, different welding dry elongations cause different durations of no-load voltage after welding starts, which leads to difficulties in model calculation, so the waveform file needs to be processed, such as Figure 2 as shown specifically comprising:

[0054] S131, read the actual voltage and actual current recorded in the waveform file, and filter out the data with zero current and no-load voltage in the read actual voltage and actual current;

[0055] S132, for the filtered data, construct a two-dimensional array of behavior voltage and column current;

[0056] S133, convert the constructed two-dimensional array into graph data using a python graph processing method, that is, convert the voltage and current waveform file into a visual picture.

[0057] S140, input the graph data into the pre-trained welding defect detection model to obtain the welding defect type.

[0058] In implementation, as shown in Figure 3 the training process of the welding defect detection model comprises:

[0059] S141, obtain sample data, the sample data comprising a waveform file and an actual corresponding welding defect type, the welding defect type comprising no defect (normal), burn-through, arc break, arc draw and weld snake;

[0060] S142, convert the waveform file in the sample data into graph data and input it into the initial neural network model to obtain the estimated welding defect type;

[0061] S143, compare the estimated welding defect type with the corresponding welding defect type in the sample data, in the case that the estimated welding defect type does not meet the expectation, execute S144 to adjust the parameters in the initial neural network model and increase the data amount of the sample data, and return to S142, until the estimated welding defect type meets the expectation, execute S145 to take the current initial neural network model as the welding defect detection model.

[0062] In implementation, the waveform file can be named in combination with the welding defect type during the model training process, and the naming format is: defect type-model type-protection gas-welding wire material-welding wire diameter-pulse presence-pre-set current-pre-set voltage.csv, for example, normal-500GS6-MAG-carbon steel-1.2-pulse has-200-28.csv.

[0063] During the training process, the actual voltage and the actual current of different welding equipment (welding machine / welding robot) under different conditions of protective gas, welding wire material, welding wire diameter, pulse, preset current, and preset voltage can be collected to obtain sample data and record the corresponding actual welding defect type. One recording form is shown in Table 1.

[0064]

[0065]

[0066] Table 1 sample data example

[0067] During the training process, the sample data can be divided into a training set and a test set. The training set is used for model training, and the test set is used to evaluate the performance of the model. If the evaluation is not up to standard, the training set needs to be further increased for training until the final welding defect detection model is obtained.

[0068] Referring to Figure 4 In another aspect of the embodiment of the present application, a structure diagram of a welding arc defect recognition system based on deep learning is also provided, which comprises:

[0069] The data acquisition unit 400 is configured to acquire basic data of the welding equipment, wherein the basic data comprises: equipment model, protective gas, welding wire material, welding wire diameter, pulse, preset voltage, and preset current.

[0070] The file creation unit 410 is configured to create a waveform file for recording the actual voltage and the actual current during the welding process by using the basic data.

[0071] The data acquisition unit 420 is configured to acquire the actual voltage and the actual current during the welding process of the welding equipment and record them in the waveform file.

[0072] The data calculation unit 430 is configured to convert the recorded waveform file into graph data.

[0073] The data transmission unit 440 is configured to input the graph data into the pre-trained welding defect detection model to obtain the welding defect type.

[0074] In the implementation, the data calculation unit 430 is specifically configured to:

[0075] read the actual voltage and the actual current recorded in the waveform file, and filter out the data with zero current and empty voltage in the read actual voltage and actual current;

[0076] construct a two-dimensional array of behavior voltage and current column for the filtered data.

[0077] The two-dimensional array constructed is converted into graph data by using a python graph processing method.

[0078] In the implementation, the method further includes a model training unit configured to

[0079] Obtaining sample data, the sample data including waveform files and actual corresponding welding defect types, the welding defect types including no defect, burn-through, arc breaking, arc drawing, and weld snake shape;

[0080] Converting the waveform files in the sample data into graph data and inputting the graph data into an initial neural network model to obtain estimated welding defect types;

[0081] Comparing the estimated welding defect types with the corresponding welding defect types in the sample data, in a case where the estimated welding defect types do not meet expectations, adjusting parameters in the initial neural network model and increasing a data amount of the sample data, returning to the step of converting the waveform files in the sample data into graph data and inputting the graph data into the initial neural network model to obtain estimated welding defect types, until the estimated welding defect types meet expectations, and taking the current initial neural network model as a welding defect detection model.

[0082] In the implementation, the data acquisition unit 420 is specifically configured to start recording the actual voltage and the actual current after detecting that the actual voltage and / or the actual current are greater than a predetermined threshold value through the mutual inductor.

[0083] In a specific implementation, referring to Figure 5 and Figure 6 Each functional unit can be integrated in the waveform data acquisition and calculation device. In this implementation, an operator can input basic data of a welding power source through the UI interaction unit. After receiving the basic data, the waveform data calculation unit creates a waveform file and stores the waveform file in the storage. After starting welding, the actual voltage and the actual current are acquired through the waveform data acquisition unit, and the acquired data is written into the waveform file through the waveform data calculation unit. After the data is written, the waveform data calculation unit converts the waveform file into graph data and uploads the graph data to the server through the data transmission unit. The server pre-trains a welding defect detection model and saves the trained welding defect detection model. After receiving the graph data, the server inputs the graph data into the welding defect detection model to obtain corresponding welding defect types and sends the welding defect types to the data transmission unit. After receiving the detection result, the waveform data calculation unit can display the detection result through the UI interaction unit.

[0084] In another implementation, the server can also send the trained welding defect detection model to the memory in the waveform data acquisition and calculation device through the data transmission unit for storage. After the waveform data calculation unit converts the waveform file into graph data, the welding defect detection model in the memory can be directly called for welding defect detection, and the detection result is displayed through the UI interaction unit.

[0085] The above merely provides the preferred embodiments of the application, but not for limiting the protective scope of the application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the application shall be included in the protective scope of the application.

Claims

1. A method for identifying a welding arc defect based on deep learning, characterized by, The method comprises the following steps: obtaining basic data of a welding device, wherein the basic data comprises: device model, shielding gas, welding wire material, welding wire diameter, pulse presence or absence, preset voltage and preset current; creating a waveform file for recording actual voltage and actual current in a welding process by using the basic data; acquiring actual voltage and actual current in the welding process and recording the actual voltage and the actual current in the waveform file; converting the recorded waveform file into graph data; inputting the graph data into a pre-trained welding defect detection model to obtain a welding defect type; the process of converting the recorded waveform file into graph data comprises: reading actual voltage and actual current recorded in the waveform file, and screening out data in which current is zero and voltage is no-load voltage from the read actual voltage and actual current; constructing a two-dimensional array with behavior voltage as rows and current as columns based on the screened data; converting the constructed two-dimensional array into graph data by using a python graph processing method; the training process of the welding defect detection model comprises: obtaining sample data, wherein the sample data comprises waveform files and actual corresponding welding defect types, and the welding defect types comprise no defect, burn-through, arc break, arc draw and weld snake shape; inputting the waveform files in the sample data into an initial neural network model after converting the waveform files into graph data to obtain an estimated welding defect type; comparing the estimated welding defect type with the corresponding welding defect type in the sample data, adjusting parameters in the initial neural network model and increasing the data amount of the sample data in the case that the estimated welding defect type does not meet the expectation, returning to the step of inputting the waveform files in the sample data into the initial neural network model after converting the waveform files into graph data to obtain the estimated welding defect type, and taking the current initial neural network model as the welding defect detection model when the obtained estimated welding defect type meets the expectation.

2. The deep learning-based welding arc defect recognition method of claim 1, wherein, the process of acquiring actual voltage and actual current in the welding process comprises: starting to record actual voltage and actual current when the actual voltage and / or the actual current detected by a mutual inductor is greater than a predetermined threshold.

3. A deep learning-based welding arc defect recognition system, characterized by, The method comprises the following steps: a data acquisition unit is configured to obtain basic data of a welding device, wherein the basic data comprises: device model, shielding gas, welding wire material, welding wire diameter, pulse presence or absence, preset voltage and preset current; a file creation unit is configured to create a waveform file for recording actual voltage and actual current in a welding process by using the basic data; a data acquisition unit is configured to acquire actual voltage and actual current in the welding process and record the actual voltage and the actual current in the waveform file; a data calculation unit is configured to convert the recorded waveform file into graph data; a data transmission unit is configured to input the graph data into a pre-trained welding defect detection model to obtain a welding defect type; the data calculation unit is specifically configured to: read actual voltage and actual current recorded in the waveform file, and screen out data in which current is zero and voltage is no-load voltage from the read actual voltage and actual current; construct a two-dimensional array with behavior voltage as rows and current as columns based on the screened data; convert the constructed two-dimensional array into graph data by using a python graph processing method; a model training unit is further configured to Obtaining sample data, the sample data including waveform files and actual corresponding welding defect types, the welding defect types including no defect, burn-through, arc breaking, arc drawing and weld snake shape; After converting the waveform files in the sample data into graph data, inputting the graph data into an initial neural network model to obtain an estimated welding defect type; Comparing the estimated welding defect type with the corresponding welding defect type in the sample data, in the case that the estimated welding defect type does not meet the expectation, adjusting parameters in the initial neural network model and increasing the data quantity of the sample data, returning to the step of converting the waveform files in the sample data into graph data and inputting the graph data into the initial neural network model to obtain the estimated welding defect type, until the obtained estimated welding defect type meets the expectation, taking the current initial neural network model as a welding defect detection model.

4. The deep learning-based welding arc defect recognition system of claim 3, wherein, The data acquisition unit is specifically configured to start recording the actual voltage and the actual current when the actual voltage and / or the actual current detected by the mutual inductor is greater than a predetermined threshold.

Images