Method for establishing resistance spot welding quality monitoring system based on integrated algorithm

Abstract

The present invention relates to a method for establishing a resistance spot welding quality monitoring system based on an integrated algorithm, and belongs to the field of quality monitoring. By collecting welding parameters and ultrasonic detection information of welds in the actual welding production process, a long short-term memory network algorithm and a random forest algorithm are introduced to establish a basic learner and an auxiliary learner, and cross-validation and weighted output are completed through an integrated algorithm to realize resistance spot welding quality monitoring. The present invention extracts process information in actual welding production to the maximum extent, and also significantly reduces the impact caused by the difference between test piece welding and actual production welded parts. Therefore, the established model shows excellent adaptability and effectiveness. The introduction of integrated learning and reinforcement learning technology further improves the accuracy of the model, making the prediction results closer to the true value, which not only enhances the prediction ability of the model, but also provides a more reliable basis for the evaluation of welding quality.

Description

Technical Field

[0001] The present invention relates to the field of quality monitoring, and in particular to a resistance spot welding quality monitoring method based on an integrated algorithm and the collection of actual welding production process information to establish a quality assessment model, and more particularly to a method for establishing a resistance spot welding quality monitoring system based on the integrated algorithm. Background Art

[0002] Resistance spot welding technology, with its significant advantages such as high efficiency, low cost, ease of operation, and high degree of automation, has occupied a pivotal position in traditional manufacturing, particularly in the automotive and rail vehicle manufacturing sectors. However, the welding process of rail vehicle bodies is often accompanied by numerous interfering factors such as electrode wear, assembly gaps, and work surface contamination. These factors not only increase the instability of the welding process but can also lead to a series of welding quality issues such as cold welds and spattering, threatening the safety performance and service life of the vehicle body. To ensure the quality of welds, real-time monitoring of weld quality has become an indispensable measure. Through real-time monitoring, problems in the welding process can be promptly identified and resolved, thereby ensuring the stability and reliability of welding quality and improving the overall performance and service life of the vehicle body.

[0003] In today's industrial landscape, welding quality monitoring technology primarily relies on the collection of parameters from test pieces during welding and comprehensive testing of their post-weld mechanical properties. This data is then used to construct quality assessment models to monitor welding quality. However, this technology faces significant challenges in practical application. The inherent differences between test piece welding and actual vehicle body welding limit its effectiveness in real-world production environments. Furthermore, the limited number of test pieces further impairs the technology's monitoring accuracy and adaptability to different welding machines. Summary of the Invention

[0004] The present invention aims to provide a method for establishing a resistance spot welding quality monitoring system based on an integrated algorithm. This method addresses the existing issues of low monitoring accuracy and poor adaptability, while also eliminating the reliance on large amounts of welding test piece data. By collecting key parameters during the actual welding process, such as current and voltage, in real time, and performing nondestructive monitoring after welding, particularly ultrasonic monitoring signals, the present invention establishes a quality assessment model based on this detailed information about the actual welding process, thereby constructing an accurate resistance spot welding quality monitoring system.

[0005] The above-mentioned purpose of the present invention is achieved through the following technical solutions:

[0006] This method, based on an integrated algorithm, establishes a resistance spot welding quality monitoring system. This system uses welding parameter information and ultrasonic testing information from the actual production process as quality assessment parameters, extracting dynamic resistance curves and ultrasonic testing images. A basic learner and an auxiliary learner are constructed, and their outputs are fed into the integrated algorithm unit. Appropriate weights are determined using a retained validation set or cross-validation to balance the contributions of different models, thus constructing a resistance spot welding quality assessment model. The model predicts and assesses weld quality, then determines whether a weld is qualified based on industry quality requirements, thereby enabling quality monitoring of resistance spot welding.

[0007] Establish a quality assessment model. During the modeling process, the dynamic resistance eigenvalues ​​are input, and a long short-term memory network algorithm is introduced to generate multi-task regression predictions for nugget size and tensile shear strength. Through multiple iterative training, a base learner is constructed. The ultrasonic nugget size eigenvalues ​​are input, and a random forest algorithm is introduced to generate regression predictions for tensile shear strength. Through multiple iterative training, an auxiliary learner is constructed. The outputs of the base learner and auxiliary learner are fed into the integrated algorithm unit as input. Appropriate weights are determined using a retained validation set or cross-validation to balance the contributions of different models, completing the quality assessment model.

[0008] The quality assessment model established based on the integrated algorithm has the function of reinforcement learning. The learning method is: the prediction results of the basic learner and the auxiliary learner are weighted averaged to output the tensile shear strength prediction result. At the same time, the output result and the characteristic value of the weld nugget size are transmitted back to the basic learner, the basic learner data is updated, and the basic learner is subjected to reinforcement learning.

[0009] The method for obtaining the dynamic resistance characteristic value is as follows: collecting the welding current and welding voltage in the actual welding production process, filtering the mean value and calculating the effective value of each 1ms data, extracting the dynamic resistance characteristic value in chronological order, and extracting the slope and mean value of the dynamic resistance curve from characteristic values ​​1 to 9 in sequence.

[0010] The integrated algorithm described in the present invention is:

[0011]

[0012] in, r 真实 In order to verify the true value of concentrated tensile shear strength, r 输出 is the final prediction output of the tensile and shear strength after model integration, h t(1) 、 h t(2) They are respectively the predicted values ​​of tensile shear strength and nugget size of the basic learner, y (n) 、l (n) are the input value of the nugget size and the predicted value of tensile shear strength of the auxiliary learner, respectively. is the integration parameter, and λ is the integration weight, which will be continuously modified according to each new data input.

[0013] The method for updating the basic learner is:

[0014] .

[0015] The beneficial effects of the present invention are:

[0016] 1. This method not only maximizes the extraction of process information from actual welding production, but also significantly reduces the impact of differences between test piece welding and actual production welds. As a result, the model established by this method demonstrates excellent adaptability and effectiveness.

[0017] 2. This model overcomes the limitation of the number of test pieces. By collecting signal characteristics from actual welding production in real time, it theoretically has a nearly unlimited number of learning samples. This innovative approach greatly enriches the model's data foundation and provides a broader learning environment.

[0018] 3. By introducing ensemble learning and reinforcement learning techniques, the model’s accuracy is further improved, making the prediction results closer to the true value. This integration of technologies not only enhances the model’s predictive capabilities but also provides a more reliable basis for welding quality assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The drawings described herein are used to provide further understanding of the present invention and constitute a part of this application. The illustrative examples of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0020] Figure 1 A flow chart of establishing a monitoring system of the present invention;

[0021] Figure 2 This is a dynamic resistance characteristic value extraction diagram of the present invention;

[0022] Figure 3 This is a characteristic value extraction diagram of the ultrasonic monitoring nugget size of the present invention;

[0023] Figure 4 This is a diagram of the long short-term memory network algorithm of the present invention;

[0024] Figure 5 It is the random forest algorithm diagram of the present invention;

[0025] Figure 6 It is the prediction effect diagram of the present invention. DETAILED DESCRIPTION

[0026] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0027] See also Figures 1 to 6 As shown, the present invention presents a method for establishing a resistance spot welding quality monitoring system based on an integrated algorithm. This method collects welding parameters and ultrasonic detection information from actual welding production processes, introduces a long short-term memory network algorithm and a random forest algorithm to establish a basic learner and an auxiliary learner, and uses the integrated algorithm to perform cross-validation and weighted output to achieve resistance spot welding quality monitoring. The monitoring system includes six components: an input unit, a data processing unit, a basic learning unit, an auxiliary learner, an integrated algorithm unit, and an output unit.

[0028] Input unit: collects welding parameters during the actual welding process, uses twisted pair wires to collect welding voltage, uses current sensors to collect welding current, and inputs welding current and voltage data; performs ultrasonic testing on welding points and inputs ultrasonic testing information.

[0029] Data processing unit: obtain the dynamic resistance effective value curve through filtering, effective value calculation, etc.; extract the dynamic resistance characteristic value, extract the dynamic resistance characteristic value in chronological order, and extract the slope and mean of the dynamic resistance curve from characteristic value 1 to characteristic value 9 in sequence; process the ultrasonic detection information, extract the weld nugget size characteristics, extract the weld nugget diameter, weld nugget area, and weld nugget perimeter characteristic values, take the average value of the distance between the two farthest points of a line in the four directions in the image as the weld nugget diameter, and the weld nugget perimeter and area are calculated according to the circumference and area formula of a circle to extract the characteristic values.

[0030] During the basic learning phase, the Long Short-Term Memory (LSTM) network algorithm is introduced. Dynamic resistance eigenvalues ​​are input and multi-task regression predictions are made for nugget size and tensile-shear strength. Through multiple iterations of training, a basic learner is constructed. The LSTM network algorithm consists of three units: a forget gate, an input gate, and an output gate. These units work together to ensure the network can effectively process and learn complex sequence data.

[0031] Forget gate: used to keep the information in the memory unit discarded or retained

[0032]

[0033] inf t is the output of the forget gate, x t is the input feature data, h t-1 is the output of the previous unit, W if 、 W hf As parameters, b if 、 b hf is the weight;

[0034] Input gate: determines the information to update the memory unit, including sigmoid and tanh functions, both of which include the input at the current moment x t Hidden state at the previous moment h t-1 :

[0035]

[0036] in i t and g t They are the outputs of the sigmoid and tanh functions respectively. c t is the new state output, W ii 、 W hi 、 W ig 、 W hg As parameters, b ii 、 b hi 、 b ig 、 b hg is the weight;

[0037] Output gate: Its function is to read the neural network state that has just been updated and output the memory unit:

[0038]

[0039] in, o t is the output gate output, h t is the predicted output of the unit and the oldest state is passed to the next unit. If the unit is the final unit, then h tis the final prediction output, W io 、 W ho As parameters, b io 、 b ho is the weight.

[0040] Finally, the predicted value of tensile shear strength is obtained h t(1) and nugget size h t(2) .

[0041] Auxiliary learner: Introducing the random forest algorithm, inputting the characteristic value of the weld nugget size to make a regression prediction of the tensile shear strength, and constructing an auxiliary learner through multiple iterative training. y n Input the characteristic value of the nugget size, l n Output for tensile shear strength prediction.

[0042] Integration algorithm unit: The output of the basic learning period and the auxiliary learner are passed into the integration algorithm unit as input. The appropriate weights are determined in the retained validation set or by using cross-validation to balance the contributions of different models. The prediction results of the long short-term memory network algorithm and the random forest algorithm are weighted averaged to output the tensile shear strength prediction result. The integration algorithm is as follows:

[0043]

[0044] in, r 真实 In order to verify the true value of concentrated tensile shear strength, r 输出 is the final prediction output of the tensile and shear strength after model integration, h t(1) 、 h t(2) They are respectively the predicted values ​​of tensile shear strength and nugget size of the basic learner, y (n) 、 l (n) are the input value of the nugget size and the predicted value of tensile shear strength of the auxiliary learner, respectively. is the integration parameter, and λ is the integration weight, which will be continuously modified according to each new data input;

[0045] At the same time, the output results and the characteristic value of the weld nugget size are transmitted back to the basic learner, the basic learner data is updated, and reinforcement learning is performed on the basic learner;

[0046]

[0047] Output unit: The integrated tensile shear strength results are output to the model, and at the same time, compared with the industry's regulations on the tensile shear strength of welds, whether the welds are qualified is determined, and real-time monitoring of spot welding quality is completed.

[0048] See also Figures 1 to 6 As shown, the present invention provides a method for establishing a resistance spot welding quality monitoring system based on an integrated algorithm and the collection of information from the actual welding production process. The system collects current and voltage signals from the actual welding process, filters them, and calculates their effective values ​​to extract dynamic resistance eigenvalues. Subsequently, a long short-term memory network algorithm is introduced to construct a basic learner using these eigenvalues ​​to achieve multi-task regression prediction of tensile shear strength and weld nugget size. Simultaneously, the system also collects ultrasonic testing information from the weld spot, extracts weld nugget size information, and uses a random forest algorithm to construct an auxiliary learner for regression prediction of tensile shear strength. Ultimately, a multi-model integrated algorithm is used to achieve model training, enhancement updates, and accurate prediction of weld quality.

[0049] Step 1: Data collection and processing:

[0050] See also Figure 2 As shown in the figure, the system first collects welding parameters during the actual welding process. The welding voltage is collected using a twisted pair cable, and the welding current is collected using a current sensor. The collected signals are filtered and the effective value is calculated. The sampling frequency for both current and voltage is 10 kHz, and the number of sampling points per 1 ms is 10. Therefore, the effective values ​​of resistance, voltage, and current are calculated once every 1 ms as follows:

[0051]

[0052] Step 2: Perform ultrasonic testing or other non-destructive testing on the weld to extract the nugget size information and nugget size characteristics, including the nugget diameter, circumference and area. Figure 3 As shown in the figure, the shape of the weld nugget is approximately circular. When measuring the diameter of the weld nugget, the average value of the distance between the two farthest points of a line in the four directions in the image is taken as the weld nugget diameter. The circumference and area of ​​the weld nugget are calculated according to the circumference and area formula of a circle to extract the characteristic values.

[0053] Step 3: Introduce the long short-term memory network algorithm. As a special recurrent neural network, the long short-term memory network algorithm has shown excellent performance in processing sequence data, especially in capturing long-term dependencies in time series. The dynamic resistance curve of resistance spot welding, as a typical time series data, records the change of resistance value over time during the welding process. The characteristic of this data is that the data points follow a strict time sequence, and each data point is closely associated with a specific time point or time period. Therefore, the characteristic value of dynamic resistance is extracted in time sequence, see Figure 2 , extract the mean and slope of the dynamic resistance curve from eigenvalue 1 to eigenvalue 9 in sequence. Input the eigenvalues ​​into the long short-term memory network algorithm, and build a basic learner through multiple training iterations to achieve the prediction output of tensile shear strength and weld nugget size. Figure 4 The long short-term memory network algorithm consists of three units: forget gate, input gate, and output gate. These units work together to ensure that the network can effectively process and learn complex sequence data.

[0054] Forget gate: used to discard or retain the information in the memory unit;

[0055]

[0056] in f t is the output of the forget gate, x t is the input feature data, h t-1 is the output of the previous unit, W if 、 W hf As parameters, b if 、 b hf is the weight;

[0057] Input gate: determines the information to update the memory unit, including sigmoid and tanh functions, both of which include the input at the current moment x t Hidden state at the previous moment h t-1 .

[0058]

[0059] in i t and g t They are the outputs of the sigmoid and tanh functions respectively. c tis the new state output, W ii 、 W hi 、 W ig 、 W hg As parameters, b ii 、 b hi 、 b ig 、 b hg is the weight;

[0060] Output gate: Its function is to read the neural network state that has just been updated and output the memory unit

[0061]

[0062] in, o t is the output gate output, h t is the predicted output of the unit and the oldest state is passed to the next unit. If the unit is the final unit, then h t is the final prediction output, W io 、 W ho As parameters, b io 、 b ho is the weight.

[0063] Finally, the predicted value of tensile shear strength is obtained h t(1) and nugget size h t(2) .

[0064] Step 4: Introduce the random forest algorithm, input the nugget size feature value to build an auxiliary learner, and make a regression prediction of the tensile shear strength; see Figure 5 , y n Input the characteristic value of the nugget size, l n Output for tensile shear strength prediction.

[0065] Step 5: Using an ensemble learning algorithm, the base learner and auxiliary learner are combined to achieve more accurate predictions. During this process, the performance of each model is evaluated using a retained validation set or cross-validation techniques. Appropriate weights are set to balance the contributions of different models. Ultimately, the prediction results of these models are combined and output through a weighted average to achieve the optimal prediction effect. The ensemble algorithm is as follows:

[0066]

[0067] in, r 真实 In order to verify the true value of concentrated tensile shear strength, r 输出 is the final prediction output of the tensile and shear strength after model integration, h t(1) 、 h t(2) They are respectively the predicted values ​​of tensile shear strength and nugget size of the basic learner, y (n) 、 l (n) are the input value of the nugget size and the predicted value of tensile shear strength of the auxiliary learner, respectively. is the integration parameter, and λ is the integration weight, which will be continuously modified according to each new data input;

[0068] Step 6: r 真实 and y (n) Input back to the basic learner to update the data and strengthen the basic learner;

[0069]

[0070] Step 7: The quality assessment model is constructed through the above steps. In the actual welding production process, the welding parameters of the welding points are collected in real time. The welding current and voltage are filtered and the effective value is calculated through the data processing unit. The dynamic resistance characteristic value is input and the result is evaluated by the model. r 真实 Output, compared with the industry's regulations on weld shear strength and nugget size, to determine whether the weld is qualified. This completes real-time monitoring of weld quality. Example

[0071] This embodiment provides a quality monitoring method based on actual welding process information to evaluate the quality of double-pulse DC welding of SUS301L, 2+0.6mm stainless steel plates. The spot welding process parameters are a primary pulse current of 5kA, secondary pulse currents of 7kA, 7.5kA, 8kA, 8.5kA, and 9kA, and an electrode pressure of 6000N. Five groups of tests were conducted, with eight welds in each group. The dynamic resistance characteristic value was extracted by collecting welding parameters, see Figure 2 As shown; collect ultrasonic detection information and extract the characteristic value of the weld nugget size, see Figure 3 shown.

[0072] The dynamic resistance characteristic value and the nugget size characteristic value are input into the model to build the basic learning period and auxiliary learning device. After verification by the integrated algorithm, the tensile shear strength prediction value is output after weighting, and the prediction accuracy can reach 97.47%. Figure 6 shown.

[0073] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that the present invention is susceptible to various modifications and variations. Any modifications, equivalent substitutions, or improvements to the present invention are intended to fall within the scope of protection of the present invention.

Claims

1. A method for establishing a resistance spot welding quality monitoring system based on an integrated algorithm, characterized by: The welding parameter information and ultrasonic detection information from the actual production process are used as quality assessment parameters to extract dynamic resistance curves and ultrasonic detection images. A basic learner and an auxiliary learner are constructed, and the outputs of the basic learner and the auxiliary learner are input into the integrated algorithm unit. The contribution of different models is balanced by weighting in a retained validation set or through cross-validation to construct a resistance spot welding quality assessment model. Establish a quality assessment model. During the modeling process, the dynamic resistance characteristic value obtained from the dynamic resistance curve is input, and a long short-term memory network algorithm is introduced to make a multi-task regression prediction output for the weld nugget size and tensile shear strength. Through multiple iterative training, a basic learner is constructed. The ultrasonic detection weld nugget size characteristic value obtained from the ultrasonic detection image is input, and a random forest algorithm is introduced to make a regression prediction for the tensile shear strength. Through multiple iterative training, an auxiliary learner is constructed. The outputs of the basic learner and the auxiliary learner are passed as input to the integrated algorithm unit. The weights are determined in the retained validation set or through cross-validation to balance the contributions of different models, completing the establishment of the quality assessment model. The quality assessment model established based on the integrated algorithm has a reinforcement learning function. The learning method is as follows: the prediction results of the basic learner and the auxiliary learner are weighted averaged to output the tensile shear strength prediction result. At the same time, the output result and the characteristic value of the weld nugget size are transmitted back to the basic learner to update the basic learner data and perform reinforcement learning on the basic learner. The dynamic resistance characteristic value acquisition method is as follows: the welding current and welding voltage are collected during the actual welding production process, the dynamic resistance characteristic values ​​are extracted in chronological order after mean filtering and effective value calculation of each 1ms data, and the slope and mean of the dynamic resistance curve are extracted in sequence from characteristic values ​​1 to 9; The integration algorithm is: ； in, r 真实 In order to verify the true value of concentrated tensile shear strength, r 输出 is the final prediction output of the tensile and shear strength after model integration, h t(1) 、 h t(2) They are respectively the predicted values ​​of tensile shear strength and nugget size of the basic learner, y (n) 、 l (n) are the input value of the nugget size and the predicted value of tensile shear strength of the auxiliary learner, respectively. is the integration parameter, and λ is the integration weight, which will be continuously modified according to each new data input.

2. The method for establishing a resistance spot welding quality monitoring system based on an integrated algorithm according to claim 1, characterized in that: The method for updating the basic learner is: 。

Images