CN114813838A - Method and system for detecting welding spot quality of vehicle body based on dynamic resistance signal - Google Patents

Abstract

The invention provides a method and a system for detecting the quality of welding spots of a vehicle body based on dynamic resistance signals, which comprises the following steps: step 1: the method comprises the steps that dynamic resistance signals of all stations of a welding workshop are collected and stored through networking of all welding machine equipment on an automobile welding production line, and a large database of welding spot quality information of an automobile body is built; step 2: on the basis of the consistent change trend of dynamic resistance signals under the same welding program number, a reference curve is constructed by adopting a low-rank sparse decomposition method, and an evaluation index of the stability of the welding process is established on the basis of the reference curve; and step 3: and developing a one-dimensional convolution network model based on a channel attention mechanism and residual connection to detect and classify the quality of the welding spots. The invention integrates the spot welding physical process information on the basis of the deep learning algorithm, improves the welding spot quality detection precision and reliability, and can realize accurate welding spot quality on-line detection.

Description

Method and system for detecting welding spot quality of vehicle body based on dynamic resistance signal

Technical Field

The invention relates to the technical field of vehicle body welding spot quality detection, in particular to a vehicle body welding spot quality detection method and system based on dynamic resistance signals.

Background

Resistance spot welding processes are widely used in the automotive industry because of their advantages of low cost, high degree of automation, and the like. On a large-batch and fast-paced white automobile body assembly line, even if the spot welding process parameters are well set, the random factor interference of a production field can easily cause quality problems of insufficient solder, solder penetration and the like. At present, the quality of welding spots is generally guaranteed by adopting a post-welding manual spot inspection method for domestic and foreign vehicle enterprises, but the quality inspection efficiency of the method is low, the production requirement of fast tempo cannot be met, and meanwhile, the welding spots of the whole vehicle cannot be comprehensively detected, so that the vehicle body faces a safety risk. The dynamic resistance signal can reflect the internal characteristic change of the welded plate under the action of thermal and force coupling in the resistance spot welding process, so the dynamic resistance signal is often used for online evaluation of the quality of the welding spot.

The existing welding spot quality detection based on the dynamic resistor still has a plurality of technical bottlenecks, and the detection precision and reliability of the welding spot quality detection cannot meet the actual production requirements. This detection technique has the following limitations:

(1) most of the welding spot quality detection research results are obtained by off-line experiments according to a plurality of plates and welding conditions under laboratory conditions, and training test samples are few, so that the methods are difficult to popularize in practical application.

(2) Most of welding spot quality detection algorithms based on dynamic resistance signals are related to feature engineering and machine learning models, the accuracy rate of the algorithms depends on the correlation between the features designed by experts and the quality of the welding spots, and the related features are difficult to accurately extract on a welding production line full of random interference.

Patent document CN201510091730.1 discloses a method for detecting quality of a resistance spot welding spot based on a dynamic resistance curve, which compares the dynamic resistance curve obtained by detection with a standard dynamic resistance curve to determine nucleation condition of the spot welding spot, calls a corresponding mathematical model in a database, inputs an average dynamic resistance obtained by detection, and a computer system calculates and outputs a diameter value and a maximum bearing force value of the spot welding spot, wherein if the diameter value and the maximum bearing force value are smaller than a corresponding set threshold value, the spot welding spot is determined to be unqualified.

Patent document CN202011329318.6 discloses a resistance spot welding quality prediction method based on ensemble learning, which collects welding process data of a welding spot according to process parameters measured by a sensor in the welding process; constructing a database; preprocessing the input data set by the characteristics; establishing an integrated learning model for the treatment and prediction of the welding spots; each classifier respectively outputs a quality prediction result of a sample to be detected; and integrating the quality prediction results of the different classifiers for the to-be-detected welding spot sample according to the output results of the different classifiers and the voting mode, and taking the majority judgment result as the final prediction output.

However, the above documents do not take the fusion of "physical process correspondence" and "data-driven classification" into consideration for the judgment of the quality of the welding spot, and the evaluation method is not comprehensive enough.

Disclosure of Invention

Aiming at the defects in the prior art, the invention aims to provide a method and a system for detecting the quality of a welding point of a car body based on a dynamic resistance signal.

The invention provides a method for detecting the quality of a welding spot of a vehicle body based on a dynamic resistance signal, which comprises the following steps:

step 1: the method comprises the steps that dynamic resistance signals of all stations of a welding workshop are collected and stored through networking of all welding machine equipment on an automobile welding production line, and a large database of welding spot quality information of an automobile body is built;

step 2: on the basis of the consistent change trend of dynamic resistance signals under the same welding program number, a reference curve is constructed by adopting a low-rank sparse decomposition method, and an evaluation index of the stability of the welding process is established on the basis of the reference curve;

and step 3: and developing a one-dimensional convolution network model based on a channel attention mechanism and residual connection to detect and classify the quality of the welding spots.

Preferably, the step 1 comprises:

collecting electrode voltage and welding current signals of each station in real time, calculating to obtain dynamic resistance signals through a preset program, and storing process signal data and corresponding spot welding information in a MySQL and MongoDB database;

the raw signal is pre-processed to eliminate the spike portion of the signal based on the pulse time of the current while reducing the high frequency noise of the main weld with a low pass filter.

Preferably, the step 2 comprises:

for welding spots with the same welding procedure, acquiring dynamic resistance signals of the welding spots meeting preset requirements through an offline experiment, and decomposing a matrix X consisting of the dynamic resistance signals to obtain low-rank prior information of a reference curve, wherein the expression is as follows:

X＝L+S

wherein: s is sparse noise, and L represents a low-rank matrix formed by reference curve information;

and then carrying out low-rank sparse decomposition according to a Semi-Soft GoDec algorithm to obtain a reference curve.

Preferably, since the lengths of the spot welding dynamic resistance signals under the same welding procedure are the same, the spot welding stability factor SF is defined as the deviation of the measured dynamic resistance signal from the reference curve based on the euclidean distance using the following equation:

wherein: x and r represent the measured signal and the corresponding reference curve, n is the length of the signal; the smaller the SF value is, the closer the acquired dynamic resistance signal is to the reference curve, the more stable the actual spot welding process is, and by setting a threshold value, stable and unstable welding spots are distinguished and a warning is given out.

Preferably, the step 3 comprises: dividing the collected dynamic resistance signal into different channels according to the pulse time for analysis, constructing a depth network model, and automatically learning the weights of different characteristic graphs based on a channel attention mechanism to improve the classification prediction precision of the model;

the deep network model comprises a squeezing excitation module, the weight of each feature map is automatically learned in each layer of the network model through the squeezing excitation module, an input signal initially enters a convolution layer, and a series of feature maps of X-X are output ₁ ,x ₂ ,…,x _c ]Feature map x _i Is W × 1, the number of channels is C, global spatial information is obtained by generating a channel-level compression value through a global average pooling layer, and the GAP layer compression operation is expressed as:

wherein: z is a radical of formula _c A compression value representing the c-th channel of the input feature map; x _c (i, j) represents the value of the c channel at positions j and k;

after the compression operation, the shape of the output feature map becomes C × 1 × 1, and then two fully-connected layers are used to learn the weight of each channel, the first fully-connected layer reduces the dimensionality of the compressed feature map, using the ReLU activation function:

s _e1 ＝max(W ₁ ×z+b ₁ ,0)

wherein: z represents a feature mapping value after a compression operation, W ₁ And b ₁ Respectively representing the weight and the deviation;

the second fully-connected layer learns the weight of each channel and maps its value to 0-1 by a sigmoid function, which is mathematically represented as:

wherein: w ₂ And b ₂ Respectively representing the weight and the offset of the FC layer;

and then adding the calculated weight to the original feature map, wherein the mathematical expression is as follows:

wherein: f _scale (x, s) is a channel multiplication method between the scalar s and the feature map x, and channel feature recalibration is carried out; the output of the channel attention mechanism in conjunction with residual concatenation is expressed as:

wherein: x is the input, H (x) is the output of the residual join;

based on welding spot labeling data, representing different types of important features through model learning, overcoming data imbalance through a weighted cross entropy loss function, wherein an expression is as follows:

wherein: t is t _j And y _j Respectively representing the actual probability and the prediction probability of the observed value belonging to the j-th class; w is a _j Is the weight coefficient for each class; n is a radical of _c Represents the number of classes;

after the cross entropy error is calculated, a random gradient descent algorithm is applied to train model parameters, and the quality type of the welding spot is directly output through the trained network model.

The invention provides a system for detecting the quality of welding spots of a vehicle body based on dynamic resistance signals, which comprises:

module M1: the method comprises the steps that dynamic resistance signals of all stations of a welding workshop are collected and stored through networking of all welding machine equipment on an automobile welding production line, and a large database of welding spot quality information of an automobile body is built;

module M2: on the basis of the consistent change trend of dynamic resistance signals under the same welding program number, a reference curve is constructed by adopting a low-rank sparse decomposition method, and an evaluation index of the stability of the welding process is established on the basis of the reference curve;

module M3: and developing a one-dimensional convolution network model based on a channel attention mechanism and residual connection to detect and classify the quality of the welding spots.

Preferably, the module M1 includes:

collecting electrode voltage and welding current signals of each station in real time, calculating by a preset program to obtain a dynamic resistance signal, and storing process signal data and corresponding spot welding information in a MySQL and MongoDB database;

the raw signal is pre-processed to eliminate the spike portion of the signal based on the pulse time of the current while reducing the high frequency noise of the main weld with a low pass filter.

Preferably, the module M2 includes:

for welding spots with the same welding procedure, acquiring dynamic resistance signals of the welding spots meeting preset requirements through an offline experiment, and decomposing a matrix X consisting of the dynamic resistance signals to obtain low-rank prior information of a reference curve, wherein the expression is as follows:

X＝L+S

wherein: s is sparse noise, and L represents a low-rank matrix formed by reference curve information;

and then carrying out low-rank sparse decomposition according to a Semi-Soft GoDec algorithm to obtain a reference curve.

Preferably, since the lengths of the spot welding dynamic resistance signals under the same welding procedure are the same, the spot welding stability factor SF is defined as the deviation of the measured dynamic resistance signal from the reference curve based on the euclidean distance using the following equation:

wherein: x and r represent the measured signal and the corresponding reference curve, n is the length of the signal; the smaller the SF value is, the closer the acquired dynamic resistance signal is to the reference curve, the more stable the actual spot welding process is, and by setting a threshold value, stable and unstable welding spots are distinguished and a warning is given out.

Preferably, the module M3 includes: dividing the collected dynamic resistance signal into different channels according to the pulse time for analysis, constructing a depth network model, and automatically learning the weights of different characteristic graphs based on a channel attention mechanism to improve the classification prediction precision of the model;

the deep network model comprises a squeezing excitation module, the weight of each feature map is automatically learned in each layer of the network model through the squeezing excitation module, an input signal initially enters a convolutional layer, and a series of features are outputThe figure is X ═ X ₁ ,x ₂ ,…,x _c ]Feature map x _i Is W × 1, the number of channels is C, global spatial information is obtained by generating a channel-level compression value through a global average pooling layer, and the GAP layer compression operation is expressed as:

wherein: z is a radical of _c A compression value representing the c-th channel of the input feature map; x _c (i, j) represents the value of the c channel at positions j and k;

after the compression operation, the shape of the output feature map becomes C × 1 × 1, and then two fully-connected layers are used to learn the weight of each channel, the first fully-connected layer reduces the dimensionality of the compressed feature map, using the ReLU activation function:

s _e1 ＝max(W ₁ ×z+b ₁ ,0)

wherein: z represents a feature mapping value after a compression operation, W ₁ And b ₁ Respectively representing the weight and the deviation;

the second fully-connected layer learns the weight of each channel and maps its value to 0-1 by a sigmoid function, which is mathematically represented as:

wherein: w ₂ And b ₂ Respectively representing the weight and the offset of the FC layer;

and then adding the calculated weight to the original feature map, wherein the mathematical expression is as follows:

wherein: f _scale (x, s) is a channel multiplication method between the scalar s and the feature map x, and channel feature recalibration is performed; output representation of channel attention mechanism in conjunction with residual concatenationComprises the following steps:

wherein: x is the input, H (x) is the output of the residual join;

based on welding spot labeling data, representing different types of important features through model learning, overcoming data imbalance through a weighted cross entropy loss function, wherein an expression is as follows:

wherein: t is t _j And y _j Respectively representing the actual probability and the prediction probability of the observed value belonging to the j-th class; w is a _j Is the weight coefficient for each class; n is a radical of _c Represents the number of classes;

after the cross entropy error is calculated, a random gradient descent algorithm is applied to train model parameters, and the quality type of the welding spot is directly output through the trained network model.

Compared with the prior art, the invention has the following beneficial effects:

the invention integrates the spot welding physical process information on the basis of the deep learning algorithm, greatly improves the welding spot quality detection precision and reliability, can realize accurate welding spot quality on-line detection, and hardly needs signal processing and professional knowledge in the aspect of spot welding.

Drawings

Other features, objects and advantages of the invention will become more apparent upon reading of the detailed description of non-limiting embodiments with reference to the following drawings:

FIG. 1 is a deep network model structure, FIG. 1a is an SE module, and FIG. 1b is a network model overall architecture;

FIG. 2 is a schematic diagram of dynamic resistance and welding current signals collected during a welding procedure at a production site; note that RL020_ L3_201 is the 201 welding program number of the on-site RL020 station L3 welding robot;

FIG. 3 is a schematic diagram of the raw and pre-processed signals of the spot weld dynamic resistance at RL020_ L3_ 201;

FIG. 4 is a schematic diagram of a reference curve and signal decomposition obtained by a Semi-Soft GoDec algorithm at RL020_ L3_201, and FIG. 4a is an acquired dynamic resistance signal; FIG. 4b shows the low rank component resulting from signal decomposition; FIG. 4c shows the noise component resulting from the signal decomposition;

FIG. 5 is a signal after an initial determination of weld process stability at RL020_ L3_201, FIG. 5a is a signal requiring further determination, and FIG. 5b is good;

FIG. 6 is a graph of the prediction result of RL020_ L3_201 based on a deep network model, and classification performance is characterized by a confusion matrix, wherein the higher the value of a diagonal element is, the better the performance is;

FIG. 7 is a flow chart of the method of the present invention.

Detailed Description

The present invention will be described in detail with reference to specific examples. The following examples will assist those skilled in the art in further understanding the invention, but are not intended to limit the invention in any way. It should be noted that it would be obvious to those skilled in the art that various changes and modifications can be made without departing from the spirit of the invention. All falling within the scope of the present invention.

Example (b):

the invention provides a white car body welding spot quality evaluation method fusing spot welding process stability and a deep learning model, which utilizes dynamic resistance closely related to growth formed by a welding spot nugget as a signal source for quality judgment, is suitable for on-line quality evaluation of welding spots on an automobile welding production line, and fuses spot welding physical process information on the basis of a data driving algorithm, thereby greatly improving the welding spot quality detection precision and reliability, as shown in figure 7, the method specifically comprises the following steps:

(1) the collection and storage of dynamic resistance signals of all stations in a welding workshop are realized through the networking of all welding machine equipment on an automobile welding production line, and a large database of welding spot quality information of an automobile body is established.

(2) Aiming at the fact that dynamic resistance signals under the same welding program number have consistent variation trends, a reference curve is constructed by adopting a low-rank sparse decomposition method, and an evaluation index of welding process stability is established based on the reference curve.

(3) A one-dimensional convolution network model based on a channel attention mechanism and residual connection is developed to achieve weld spot quality classification, the channel attention mechanism considers weights of different feature maps to enhance the classification performance of the model, and the residual connection structure can promote feature reuse and reduce redundancy.

The step (1) is specifically as follows:

the data acquisition system on the automobile welding production line acquires electrode voltage and welding current signals of each station in real time, dynamic resistance signals are obtained through internal program calculation, and the process signal data and corresponding spot welding information are stored in a MySQL and MongoDB database. Due to the high strength of the car body material, a multi-pulse current welding machine with preheating current is often adopted on the production line. The basic spot welding process consists of several constant current pulses of a certain pulse duration and cooling between the pulses. Due to the removal of the cooling part, the acquired dynamic resistance signal has many discontinuities, and in order to improve the accuracy of classification, the original signal needs to be preprocessed. The spike portion of the signal is eliminated according to the pulse time of the current while the high frequency noise of the main welding section is reduced with a low pass filter.

Collecting defective weld data on an actual production line is very expensive and time consuming, and therefore poor weld condition experiments are conducted in order to generate typical weld defects. Various welding abnormal conditions which may occur in the spot welding process of the white automobile body, such as current shunting, plate gap, distortion and other working conditions, are adopted in the experiment. And (3) performing a welding experiment in a laboratory by adopting a welding program and a plate stacking combination which are the same as those of the production line, refining expert experience knowledge, acquiring energy characteristics of welding defects, recording types of the welding spot defects and storing the types of the welding spot defects into a welding spot quality information database.

The step (2) is specifically as follows:

through analysis of on-site vehicle body welding spot signal data, the vehicle body welding spots adopting the same welding procedure have similar welding working conditions including plate stacking combination and the like, and the acquired dynamic resistance signals have an approximate uniform change mode. Thus, the difference between the dynamic resistance signal acquired in real time and the uniformly varying pattern can be used as an indicator of the stability of the welding process, the result of which can be used as a priori condition for quality detection based on deep learning. The key to the stability metric of the welding process is the construction of the reference curve and the selection of the similarity metric.

The construction of the reference curve specifically comprises the following steps: for the welding spots of the same welding procedure, dynamic resistance signals of a series of qualified welding spots are obtained through off-line experiments. The matrix of these signal data can be decomposed into physically significant features, and the low rank prior information of the reference curve is retained in the signal matrix. By mathematical description, the measured signal data may be formed into a matrix

X ═ L + S, where S is sparse noise and L denotes a low rank matrix composed of reference curve information. In the low-rank sparse decomposition algorithm, the Semi-Soft GoDec algorithm with high speed and strong rank parameter selection capability is selected.

The Semi-Soft GoDec algorithm specifically comprises the following steps: the algorithm can decompose the matrix X into three components defined in equation 1 by the minimization problem given in equation 2:

X＝L+S+Rrank(L)≤r,card(S)≤k#(1)

wherein: r and k are each

Maximum rank sum of

The maximum value of the number of the medium elements,

is the approximation error. I | · | purple wind _L1 Represents the norm L1, is defined as | | | S | | non-volatile wind _L1 ＝∑ _i,j |s(i,j)|，||.|| _F Represents the Frobenius norm and is defined as

The regularization parameter λ is a trade-off between the error term and the sparsity S.

The Semi-Soft GoDec algorithm adopts low-rank approximation based on bilateral random projection BRP to replace a Singular Value Decomposition (SVD) method, so that the time cost is reduced. From random projections at iteration t

And

low rank component L of medium estimate X _t As shown in equation 3:

wherein:

and

is a random matrix. Sparse component (S) of iteration _t ) Can be estimated by a soft threshold operator as shown in equation 4:

S _t ＝Soft_threshold(X-L _t ,λ)#(4)

the soft threshold operator for an arbitrary matrix is defined as:

Soft_threshold(Q,λ)＝max(|q|-λ,0)sgn(q)#(5)

wherein: q represents each element of the matrix Q, sgn () being a sign function.

The flow chart of Semi-Soft GoDec is as follows:

inputting: measuring a dynamic resistance signal matrix X and hyper-parameters r, lambda, p and epsilon;

and (3) outputting: a low rank matrix L and a sparse matrix S;

1. initialization: l is ₀ ＝X，S ₀ ＝0，t＝0；

2. When in use

Executing;

3.t＝t+1；

4.

5.

A ₂ ＝Y ₁ ；

6.

7. if the rank is

Then

Turning to the step 3; finishing;

8.

9.S _t ＝Soft_threshold(X-L _t ,λ)；

10. the loop is ended.

The selection of the similarity measure is specifically as follows: in order to measure the stability of the welding process for each weld spot, a reasonable evaluation criterion needs to be established. The length of the spot welding dynamic resistance signals under the same welding procedure is the same, and calculation by utilizing the Euclidean distance is simpler and quicker. Based on the euclidean distance, the spot weld Stability Factor (SF) is defined as the deviation of the measured dynamic resistance signal from the reference curve using equation 6:

wherein: x and r represent the measured signal and the corresponding reference curve, and n is the length of the signal.

In a mass production process, SF can be calculated in real time for on-line quantitative assessment of spot welding process stability. The smaller the SF value is, the closer the acquired dynamic resistance signal is to the reference curve, and the more stable the actual spot welding process is. In the present invention, the welding point highly consistent with the reference curve is considered to have good quality, otherwise, it means that an unstable situation occurs in the actual welding process, and there may be a quality problem. By setting the threshold, stable and unstable weld spots can be distinguished and a warning issued. Thus, by calculating the percentage of suspicious welds at each location online, the locations of welds with poor overall stability can be identified. The judgment of the stability of the welding process has clear physical significance, is beneficial to identifying unstable factors on site, and is suitable for the vehicle body welding production line with the characteristic of batch production.

The step (3) is specifically as follows:

in resistance spot welding processes, the purpose of preheating is to melt the sheet coating, while the main weld segment pulses are used to form the weld nugget, so that nugget quality information is mainly embedded in the time sequence between pulses. The acquired dynamic resistance signals are divided into different channels according to the pulse time for analysis, and the influence of each channel on final welding spot quality prediction is different, so that the proposed depth network model automatically learns the weights of different feature maps based on a channel attention mechanism to improve the classification prediction accuracy of the model. The deep network model structure is shown in fig. 1. The network has some of the same basic components as a traditional convolutional network (CNN) model, including convolutional layers, a rectifying linear unit (ReLU) activation function, a Batch Normalization (BN) layer, and a Global Average Pooling (GAP) layer.

The convolution layer is used for performing convolution operation with shared weight on input data to extract features and generate a series of feature graphs; the ReLU activation function is used for increasing the nonlinear relation among layers of the depth network and generating sparsity to relieve the problem of model overfitting; the BN layer is used for improving the model training speed and relieving the over-fitting problem; the GAP is used to calculate the average value of the channel level for a series of feature maps.

The basic components of the network model of the present invention are the Squeeze Excitation (SE) modules, as shown in fig. 1 (a). The SE module can explicitly model intra-module channel correlation and selectively enhance important feature maps and suppress less useful feature maps. The residual join mechanism may facilitate feature reuse and reduce redundancy. The weight of each feature map can be automatically learned at each layer of the network model by the SE module, and the SE module can be stacked in the entire network model to enhance the discriminability of the features.

Specifically, in the SE module, an input signal initially enters the convolutional layer, and a series of characteristic maps X ═ X are output ₁ ,x ₂ ,…,x _c ]Feature map x _i The size of (2) is W × 1, and the number of channels is C. The global spatial information is obtained by generating a channel-level compression value through the GAP layer, which may be expressed as:

wherein: z is a radical of _c Representing the compression value of the c-th channel of the input profile. X _c (i, j) represents the value of the c-th channel at positions j and k.

After the compression operation, the shape of the output feature map becomes C × 1 × 1. Then, two fully-connected (FC) layers are employed to learn the weight of each channel. The first FC layer may reduce the dimensionality of the compressed feature map, using the ReLU activation function:

s _e1 ＝max(W ₁ ×z+b ₁ ,0)#(8)

wherein: z represents a feature mapping value after a compression operation, W ₁ And b ₁ Respectively representing the weight and the deviation.

The second FC layer may learn the weight of each channel and map its value to 0-1 by a sigmoid function, which may be mathematically expressed as:

wherein: w ₂ And b ₂ The weight and offset of the FC layer are respectively represented.

And then adding the calculated weight to the original feature map, wherein the mathematical expression is as follows:

wherein: f _scale (x, s) is the channel multiplication between the scalar s and the feature map x. This operation may perform channel feature recalibration.

Furthermore, using residual concatenation is more advantageous than conventional CNN. Thus, the output of the channel attention mechanism in conjunction with residual concatenation can be expressed as:

wherein: x is the input and h (x) is the output of the residual concatenation.

Weld quality prediction can be translated into a classification problem. In the multi-classification task, cross-entropy errors are often used as the objective function for minimization. By providing weld point annotation data, the model is able to learn important features that represent different classes. Due to the imbalance of data collected by the actual production line, a weighted cross entropy loss function is selected for the prediction problem, and is defined as:

wherein: t is t _j And y _j Respectively representing the actual probability and the predicted probability of the observed value belonging to the j-th class. w is a _j Is a weight coefficient, N, for each class _c Representing the number of classes. After cross entropy errors are calculated, a stochastic gradient descent algorithm may be applied to trainAnd (4) training the model parameters, wherein the trained network model can directly output the quality type of the welding spot.

Further, when the quality of the welding spot is evaluated on line, the quality of the welding spot is preliminarily judged through welding stability, the constructed reference curve is used for preliminarily judging, if SF of the measurement signal is smaller than a set threshold value, the quality of the welding spot is good, and if not, the next deep network model is used for carrying out quality judgment. And predicting the quality type of the welding point by loading the trained network weight.

The invention provides a white car body welding spot quality evaluation method based on dynamic resistance signals. Compared with the conventional mainstream prediction method, the method integrates the stability judgment and the deep learning model in the spot welding process, and has advantages in prediction precision and reliability. The invention provides a new method for predicting the quality of the welding spot based on the dynamic resistance signal, and has good application prospect in the modern automobile manufacturing industry.

The essence of the invention is further illustrated in connection with the specific embodiment RL020_ L3_201 (dynamic resistance signal collected from the 201 welding program number of the production site RL020 station L3 welding robot). Please refer to fig. 2 to fig. 6.

First, the dynamic resistance and welding current signals collected on the automobile welding production line are shown in fig. 2. The collected dynamic resistance signal can be spiked according to the current pulse time and a low pass filter is used to remove high frequency noise of the main welding section. Fig. 3 shows the raw and preprocessed signals of the dynamic resistance signal.

Then, the judgment of the stability of the welding process is realized by constructing a reference curve. The invention adopts a Semi-Soft GoDec algorithm to extract and measure the low-rank component of the dynamic resistance signal matrix so as to construct a reference curve. FIG. 4 shows a diagram of signal decomposition with a Semi-Soft GoDec and the corresponding reference curve. The low rank component reflects the general trend of the dynamic resistance signal change of normal solder joints and can be used to construct a reference curve. The reference curve is used for carrying out preliminary evaluation on the dynamic resistance signal acquired in real time, spot welding of the dynamic resistance signal which is obviously deviated from the reference curve is considered to be abnormal and an alarm is output, and spot welding consistent with the reference curve is considered to be good in quality. Fig. 5 shows the signal after the preliminary determination of the stability of the welding process.

And finally, predicting the quality type of the welding spot by using a one-dimensional convolution network based on a channel attention mechanism and a residual error structure, and realizing online evaluation of the quality of the welding spot by the input signal through feedforward calculation of a trained depth network model. FIG. 6 illustrates a confusion matrix based on model prediction.

The result shows that the method for detecting the welding spot quality of the car body, which integrates the stability of the welding process and the deep learning model, has reliable result and low detection cost, is particularly suitable for the on-line quality detection of the welding production field, has advantages in prediction precision compared with the existing quality prediction method based on machine learning, and has higher effectiveness and popularization.

Those skilled in the art will appreciate that, in addition to implementing the systems, apparatus, and various modules thereof provided by the present invention in purely computer readable program code, the same procedures can be implemented entirely by logically programming method steps such that the systems, apparatus, and various modules thereof are provided in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers, embedded microcontrollers and the like. Therefore, the system, the device and the modules thereof provided by the present invention can be considered as a hardware component, and the modules included in the system, the device and the modules thereof for implementing various programs can also be considered as structures in the hardware component; modules for performing various functions may also be considered to be both software programs for performing the methods and structures within hardware components.

The foregoing description has described specific embodiments of the present invention. It is to be understood that the present invention is not limited to the specific embodiments described above, and that various changes or modifications may be made by one skilled in the art within the scope of the appended claims without departing from the spirit of the invention. The embodiments and features of the embodiments of the present application may be combined with each other arbitrarily without conflict.

Claims (10)

1. A method for detecting the quality of a welding spot of a vehicle body based on a dynamic resistance signal is characterized by comprising the following steps:

step 1: the method comprises the steps that dynamic resistance signals of all stations of a welding workshop are collected and stored through networking of all welding machine equipment on an automobile welding production line, and a large database of welding spot quality information of an automobile body is built;

step 2: on the basis of the consistent change trend of dynamic resistance signals under the same welding program number, a reference curve is constructed by adopting a low-rank sparse decomposition method, and an evaluation index of the stability of the welding process is established on the basis of the reference curve;

and step 3: and developing a one-dimensional convolution network model based on a channel attention mechanism and residual connection to detect and classify the quality of the welding spots.

2. The method for detecting the quality of the welding spot of the car body based on the dynamic resistance signal as claimed in claim 1, wherein the step 1 comprises the following steps:

collecting electrode voltage and welding current signals of each station in real time, calculating by a preset program to obtain a dynamic resistance signal, and storing process signal data and corresponding spot welding information in a MySQL and MongoDB database;

the raw signal is pre-processed to eliminate the spike portion of the signal based on the pulse time of the current while reducing the high frequency noise of the main weld with a low pass filter.

3. The method for detecting the quality of the welding spot of the car body based on the dynamic resistance signal as claimed in claim 1, wherein the step 2 comprises the following steps:

for welding spots with the same welding procedure, acquiring dynamic resistance signals of the welding spots meeting preset requirements through an offline experiment, and decomposing a matrix X consisting of the dynamic resistance signals to obtain low-rank prior information of a reference curve, wherein the expression is as follows:

X＝L+S

wherein: s is sparse noise, and L represents a low-rank matrix formed by reference curve information;

and then carrying out low-rank sparse decomposition according to a Semi-Soft GoDec algorithm to obtain a reference curve.

4. The method for detecting the quality of the welding spot of the car body based on the dynamic resistance signal as claimed in claim 3, wherein the welding stability factor SF is defined as the deviation of the measured dynamic resistance signal from the reference curve based on the Euclidean distance by using the following formula because the lengths of the welding dynamic resistance signals under the same welding procedure are the same:

wherein: x and r represent the measured signal and the corresponding reference curve, n is the length of the signal; the smaller the SF value is, the closer the acquired dynamic resistance signal is to the reference curve, the more stable the actual spot welding process is, and by setting a threshold value, stable and unstable welding spots are distinguished and a warning is given out.

5. The method for detecting the quality of the welding spot of the car body based on the dynamic resistance signal as claimed in claim 1, wherein the step 3 comprises the following steps: dividing the collected dynamic resistance signal into different channels according to the pulse time for analysis, constructing a depth network model, and automatically learning the weights of different characteristic graphs based on a channel attention mechanism to improve the classification prediction precision of the model;

the deep network model comprises a squeezing excitation module, the weight of each feature map is automatically learned in each layer of the network model through the squeezing excitation module, an input signal initially enters a convolution layer, and a series of feature maps of X-X are output ₁ ，x ₂ ，...，x _c ]Feature map x _i Is W × 1, the number of channels is C, global spatial information is obtained by generating a channel-level compression value through a global average pooling layer, and the GAP layer compression operation is expressed as:

wherein: z is a radical of _c Representing input featuresCompression value of the c-th channel of the graph; x _c (i, j) represents the value of the c channel at positions j and k;

after the compression operation, the shape of the output feature map becomes C × 1 × 1, and then two fully-connected layers are used to learn the weight of each channel, the first fully-connected layer reduces the dimensionality of the compressed feature map, using the ReLU activation function:

s _e1 ＝max(W ₁ ×z+b ₁ ，0)

wherein: z represents a feature mapping value after a compression operation, W ₁ And b ₁ Respectively representing the weight and the deviation;

the second fully-connected layer learns the weight of each channel and maps its value to 0-1 by a sigmoid function, which is mathematically represented as:

wherein: w ₂ And b ₂ Respectively representing the weight and the offset of the FC layer;

and then adding the calculated weight to the original feature map, wherein the mathematical expression is as follows:

wherein: f _scale (x, s) is a channel multiplication method between the scalar s and the feature map x, and channel feature recalibration is performed;

the output of the channel attention mechanism in conjunction with residual concatenation is expressed as:

wherein: x is the input, H (x) is the output of the residual join;

based on welding spot labeling data, representing different types of important features through model learning, overcoming data imbalance through a weighted cross entropy loss function, wherein an expression is as follows:

wherein: t is t _j And y _j Respectively representing the actual probability and the prediction probability of the observed value belonging to the j-th class; w is a _j Is the weight coefficient for each class; n is a radical of _c Represents the number of classes;

after the cross entropy error is calculated, a random gradient descent algorithm is applied to train model parameters, and the quality type of the welding spot is directly output through the trained network model.

6. The utility model provides a car body solder joint quality detection system based on dynamic resistance signal which characterized in that includes:

module M1: the method comprises the steps that dynamic resistance signals of all stations of a welding workshop are collected and stored through networking of all welding machine equipment on an automobile welding production line, and a large database of welding spot quality information of an automobile body is built;

module M2: on the basis of the consistent change trend of dynamic resistance signals under the same welding program number, a reference curve is constructed by adopting a low-rank sparse decomposition method, and an evaluation index of the stability of the welding process is established on the basis of the reference curve;

module M3: and developing a one-dimensional convolution network model based on a channel attention mechanism and residual connection to detect and classify the quality of the welding spots.

7. The system for detecting the quality of the welding spot on the vehicle body based on the dynamic resistance signal as claimed in claim 6, wherein the module M1 comprises:

collecting electrode voltage and welding current signals of each station in real time, calculating by a preset program to obtain a dynamic resistance signal, and storing process signal data and corresponding spot welding information in a MySQL and MongoDB database;

the raw signal is pre-processed to eliminate the spike portion of the signal based on the pulse time of the current while reducing the high frequency noise of the main weld with a low pass filter.

8. The system for detecting the quality of the welding spot on the vehicle body based on the dynamic resistance signal as claimed in claim 6, wherein the module M2 comprises:

for welding spots with the same welding procedure, acquiring dynamic resistance signals of the welding spots meeting preset requirements through an offline experiment, and decomposing a matrix X consisting of the dynamic resistance signals to obtain low-rank prior information of a reference curve, wherein the expression is as follows:

X＝L+S

wherein: s is sparse noise, and L represents a low-rank matrix formed by reference curve information;

and then carrying out low-rank sparse decomposition according to a Semi-Soft GoDec algorithm to obtain a reference curve.

9. The system of claim 8, wherein the weld stability factor SF is defined as a deviation of the measured dynamic resistance signal from a reference curve based on euclidean distance using the following equation, since the lengths of the spot welding dynamic resistance signals under the same welding procedure are the same:

wherein: x and r represent the measured signal and the corresponding reference curve, n is the length of the signal; the smaller the SF value is, the closer the acquired dynamic resistance signal is to the reference curve, the more stable the actual spot welding process is, and by setting a threshold value, stable and unstable welding spots are distinguished and a warning is given out.

10. The system for detecting the quality of the welding spot on the vehicle body based on the dynamic resistance signal as claimed in claim 6, wherein the module M3 comprises: dividing the collected dynamic resistance signal into different channels according to the pulse time for analysis, constructing a depth network model, and automatically learning the weights of different characteristic graphs based on a channel attention mechanism to improve the classification prediction precision of the model;

the deep network model comprises a squeezing excitation module, the weight of each feature map is automatically learned in each layer of the network model through the squeezing excitation module, an input signal initially enters a convolution layer, and a series of feature maps of X-X are output ₁ ，x ₂ ，...，x _c ]Feature map x _i Is W × 1, the number of channels is C, global spatial information is obtained by generating a channel-level compression value through a global average pooling layer, and the GAP layer compression operation is expressed as:

wherein: z is a radical of _c A compression value representing the c-th channel of the input feature map; x _c (i, j) represents the value of the c channel at positions j and k;

after the compression operation, the shape of the output feature map becomes C × 1 × 1, and then two fully-connected layers are used to learn the weight of each channel, the first fully-connected layer reduces the dimensionality of the compressed feature map, using the ReLU activation function:

s _e1 ＝max(W ₁ ×z+b ₁ ，0)

wherein: z represents a feature mapping value after a compression operation, W ₁ And b ₁ Respectively representing the weight and the deviation;

the second fully-connected layer learns the weight of each channel and maps its value to 0-1 by a sigmoid function, which is mathematically represented as:

wherein: w ₂ And b ₂ Respectively representing the weight and the offset of the FC layer;

and then adding the calculated weight to the original feature map, wherein the mathematical expression is as follows:

wherein: f _scale (x, s) is a channel multiplication method between the scalar s and the feature map x, and channel feature recalibration is performed;

the output of the channel attention mechanism in conjunction with residual concatenation is expressed as:

wherein: x is the input, H (x) is the output of the residual join;

based on welding spot labeling data, different types of important features are represented through model learning, imbalance of the data is overcome through a weighted cross entropy loss function, and the expression is as follows:

wherein: t is t _j And y _j Respectively representing the actual probability and the prediction probability of the observed value belonging to the j-th class; w is a _j Is the weight coefficient for each class; n is a radical of _c Represents the number of classes;

after the cross entropy error is calculated, a random gradient descent algorithm is applied to train model parameters, and the quality type of the welding spot is directly output through the trained network model.

Images