CN105741306A - Penetration state determination method based on small hole characteristic on back side - Google Patents

Abstract

The invention discloses a penetration state determination method based on a small hole characteristic on a back side, comprising steps of dividing a variable-polarity plasma arc-welding into P penetration states, choosing the back side small hole image as data resource, collecting the back side small hole image data under various penetration states, performing characteristic extraction on the small hole image data, determining a plurality of characteristic variables, choosing sample data under various penetration states, wherein each group of the sample data comprises a characteristic variable and a penetration state, randomly dividing the sample data into training sample data and test sample data, determining the quantity of nodes of an input layer, a hidden layer and a output layer of a limit learning machine model to be established, choosing an excitation function of a limit learning machine, performing learning and establishing a limit learning machine model with the training sample data as the input of the limit learning machine, and adopting the test sample data to verify the prediction accuracy of the limit learning machine model. The invention establishes the limit learning machine model to perform determination on the penetration state of the small hole, the prediction accuracy and the convergence rate is quick.

Description

A kind of penetration signal decision method based on the little hole characteristic in the back side

Technical field

The present invention relates to welding technology field, especially a kind of penetration signal decision method based on the little hole characteristic in the back side.

Background technology

Plasma arc welding with adjustable polarity parameters (Variablepolarityplasmaarcwelding, VPPAW) owing to having energy density height, penetration power is strong, can realize the advantages such as cut deal aluminum alloy materials one side welding with back formation, is therefore widely used in space flight and aviation welding technology field.For VPPAW joint form, weigh the penetration signal that the most important and the most basic index of welding forming quality is weld seam.But in actual welding process, owing to affecting the many factors of penetration signal, as

Workpiece radiating condition, thermal deformation, gap clearance etc. all can affect the stability of welding process and final penetration signal, are difficult to ensure that only by fixing welding conditions and obtain uniform and stable appropriate penetration signal.In view of the aperture dynamic behaviour of plasma arc welding with adjustable polarity parameters is the key factor directly affecting appearance of weld stability and joint quality, therefore study the useful signal that can characterize molten bath aperture characteristic behavior and be to maintain the main path that weld joint stable shapes.

In the sensing technology of current existing sign aperture characteristic signal, mainly carry out the research of acoustical signal, arc light spectrum, wake flame and plasma cloud voltage sensor aspect, but these sensing technologies all can only reflect the presence or absence of aperture indirectly.Although image vision sensing can directly reflect the dynamic behaviour of aperture, it is also carried out the research of welding procedure and aperture characteristic relation, but do not analyse in depth the relation of little hole characteristic and actual welding penetration signal, thus it is difficult to quantification ground to analyze appearance of weld stability and joint quality.

Summary of the invention

It is an object of the invention to provide a kind of penetration signal decision method based on the little hole characteristic in the back side, with solve prior art cannot the problem of relation of the little hole characteristic of quantitative analysis and actual welding penetration signal.

In order to achieve the above object, the invention provides a kind of penetration signal decision method based on the little hole characteristic in the back side, including:

Step 1: plasma arc welding with adjustable polarity parameters is divided into P penetration signal, P is the integer more than or equal to 2, and is demarcated as the 1st label, the 2nd label P label respectively, chooses the back side aperture image of plasma arc welding with adjustable polarity parameters as data source；

Step 2: gather the back side aperture view data under each penetration signal, aperture view data is carried out feature extraction, it is determined that multiple characteristic variables；

Step 3: choose plasma arc welding with adjustable polarity parameters sample data under each penetration signal, often group sample data all includes characteristic variable and penetration signal；

Step 4: sample data is randomly divided into training sample data and test specimens notebook data, training sample data are used for learning extreme learning machine model, test specimens notebook data is used for verifying extreme learning machine model, determine the input layer of extreme learning machine model to be set up, hidden layer and output layer nodes, choose the excitation function of extreme learning machine；

Step 5: using training sample data as the input of extreme learning machine, carry out the study of extreme learning machine penetration signal decision model, and then set up extreme learning machine model, judged the penetration signal of aperture according to the aperture characteristic variable of input by described extreme learning machine model；

Step 6: adopt test specimens notebook data that the extreme learning machine model set up is verified, the prediction accuracy of this extreme learning machine model that checking is set up.

Further, in step 1, being divided into by actual penetration signal: partial penetration, complete penetration and three kinds of penetration signal of cutting, partial penetration being set as, the 1st label, complete penetration are set as the 2nd label, cutting is set as the 3rd label.

Further, in step 1, using the back side aperture image that obtains through visual sensing system as data source.

Further, in step 2, characteristic variable is the aperture characteristic parameter that aperture image obtains through image processing method and camera calibration, and this characteristic variable includes aperture width and aperture area.

Further, in step 3, often group sample data all includes current time and the characteristic variable of first three historical juncture acquisition, and also includes the penetration signal of current time.

Further, in step 4, described input layer number is the dimension Q of described characteristic variable, this dimension Q is multiplied by the quantity gathering the moment equal to the quantity of characteristic variable, often group sample data all includes the characteristic variable of Q dimension and 1 penetration signal of current time, described output layer nodes is the quantity of described penetration signal, and node in hidden layer is empirical value, and excitation function g (x) selects sigmoidal function: g (x)=1/ (1+exp (-x)).

Further, randomly select an initial node in hidden layer, set up extreme learning machine model and adopt test specimens notebook data to calculate prediction accuracy, choose a node in hidden layer at interval of the interval set and calculate prediction accuracy, choose prediction accuracy the highest time corresponding node in hidden layer be node in hidden layer used in step 4.

Further, described step 5 specifically includes:

Step 51: determine hidden layer neuron number, the random connection weight w setting input layer and implicit interlayer_iBiasing with hidden layer neuronFor node in hidden layer；

Step 52: calculate the implicit output matrix H of corresponding training sample data_tr；Wherein,

$H_{tr}\left(w_1,...w_{\tilde{N}},b_1,...,b_{\tilde{N}},x_1^{tr},...x_N^{tr}\right)={\left[\begin{array}{ccc}g\left(w_1{x}_1^{tr}+b_1\right)& ...& g\left(w_{\tilde{N}}{x}_1^{tr}+b_{\tilde{N}}\right)\\ \begin{array}{c}.\\ .\\ .\end{array}& \begin{array}{c}\\ ...\\ \end{array}& \begin{array}{c}.\\ .\\ .\end{array}\\ g\left(w_1{x}_N^{tr}+b_1\right)& ...& g\left(w_{\tilde{N}}{x}_N^{tr}+b_{\tilde{N}}\right)\end{array}\right]}_{N\×\tilde{N}}$

For the training sample data of input, N is the group number of training sample data；

Step 53: adopt method of least square to solve hidden layer output matrix H_trMoore-Penrose generalized inverse

Step 54: calculate output weightWhereinT_tr=[t₁,…,t_N]^T, and T_trFor training sample output matrix.

Further, step 6 specifically includes:

Step 61: calculate the implicit output matrix H of corresponding test specimens notebook data_te；Wherein,

$H_{te}\left(w_1,...w_{\tilde{N}},b_1,...,b_{\tilde{N}},x_1^{te},...x_M^{te}\right)={\left[\begin{array}{ccc}g\left(w_1{x}_1^{te}+b_1\right)& ...& g\left(w_{\tilde{N}}{x}_1^{te}+b_{\tilde{N}}\right)\\ \begin{array}{c}.\\ .\\ .\end{array}& \begin{array}{c}\\ ...\\ \end{array}& \begin{array}{c}.\\ .\\ .\end{array}\\ g\left(w_1{x}_M^{te}+b_1\right)& ...& g\left(w_{\tilde{N}}{x}_M^{te}+b_{\tilde{N}}\right)\end{array}\right]}_{M\×\tilde{N}}$

For Q dimensional feature variable data in test specimens notebook data, M is the group number of test specimens notebook data；

Step 62: the output T of calculating limit learning machine_te=H_teβ；

Step 63: the penetration signal type that element value is corresponding in the output row of extreme learning machine is as judged result, and this element value is positive integer, and each element value is in 1,2 P, and corresponding corresponding 1st label, the 2nd label P label；

Step 64: the judged result of extreme learning machine and actual penetration signal are contrasted, calculates classification accuracy, extreme learning machine model is verified, and actual penetration signal is in test specimens notebook data except Q dimensional feature variable dataPenetration signal data in addition.

The present invention proposes a kind of penetration signal decision method based on the little hole characteristic in the back side in welding technology field, obtaining, back side aperture image basis extracts aperture characteristic parameter through image procossing, and choose penetration signal sample and carry out extreme learning machine model learning, by the extreme learning machine model set up, the penetration signal of aperture is judged.Find that this extreme learning machine model prediction accuracy rate is high through checking, more than 90% can be reached, and the study of extreme learning machine model is without iteration, fast convergence rate, has good Generalization Capability.

Accompanying drawing explanation

Fig. 1 is the flow chart of the penetration signal decision method based on the little hole characteristic in the back side provided by the invention；

Aperture sequence image corresponding under three kinds of typical penetration signal that Fig. 2 provides for the embodiment of the present invention；

Fig. 3 is extreme learning machine structural model；

The node in hidden layer that Fig. 4 provides for the embodiment of the present invention impact on extreme learning machine performance.

Detailed description of the invention

Below in conjunction with schematic diagram, the specific embodiment of the present invention is described in more detail.According to description below and claims, advantages and features of the invention will be apparent from.It should be noted that, accompanying drawing all adopts the form simplified very much and all uses non-ratio accurately, only in order to convenience, the purpose aiding in illustrating the embodiment of the present invention lucidly.

Refer to Fig. 1, present embodiments provide a kind of penetration signal decision method based on the little hole characteristic in the back side, including:

Step 1: plasma arc welding with adjustable polarity parameters is divided into P penetration signal, P is the integer more than or equal to 2, and is demarcated as the 1st label, the 2nd label P label respectively, chooses the back side aperture image of plasma arc welding with adjustable polarity parameters as data source；

Step 2: gather the back side aperture view data under each penetration signal, aperture view data is carried out feature extraction, it is determined that multiple characteristic variables；

Step 3: choose plasma arc welding with adjustable polarity parameters sample data under each penetration signal, often group sample data all includes characteristic variable and penetration signal；

Step 4: sample data is randomly divided into training sample data and test specimens notebook data, training sample data are used for learning extreme learning machine model, test specimens notebook data is used for verifying extreme learning machine model, refer to Fig. 3, determine the input layer of extreme learning machine model to be set up, hidden layer and output layer nodes, choose the excitation function of extreme learning machine；

Step 5: using training sample data as the input of extreme learning machine, carry out the study of extreme learning machine penetration signal decision model, and then set up extreme learning machine model, judged the penetration signal of aperture according to the aperture characteristic variable of input by described extreme learning machine model；

Step 6: adopt test specimens notebook data that the extreme learning machine model set up is verified, the prediction accuracy of this extreme learning machine model that checking is set up.

In the present embodiment, in step 1, as in figure 2 it is shown, actual penetration signal is divided into: partial penetration, complete penetration and three kinds of penetration signal of cutting, partial penetration is set as, and the 1st label, complete penetration are set as the 2nd label, cutting is set as the 3rd label.

Further, in step 1, using the back side aperture image that obtains through visual sensing system as data source, visual sensing system includes ccd video camera, composite filter-dimming system, sensor stand and industrial computer.As a nonrestrictive example, it is 55 width that ccd video camera camera is adopted as maximum number of pictures per second, time of exposure 1 microsecond, resolution is 1280 × 1084, the narrow-band-filter eyeglass of wavelength 660nm centered by filter glass, dim light eyeglass is the combined lens of 5%, 10%, 20% percent of pass, and the diameter of dim light eyeglass and filter glass is 25mm.

In the present embodiment, in step 2, characteristic variable is the aperture characteristic parameter that aperture image obtains through image processing method and camera calibration, and this characteristic variable includes aperture width and aperture area.

Extreme learning machine model for VPPAW, consider that welding process exists thermal inertia, namely there is certain relation between current aperture characteristic parameter and some historical juncture values, in addition to prevent model excessively complicated, therefore only take currency and above several historical juncture values as the input of extreme learning machine model.

Preferably, in step 3, often group sample data all includes current time and the characteristic variable of first three historical juncture acquisition, and also includes the penetration signal of current time.

In step 4, described input layer number is the dimension Q of described characteristic variable, this dimension Q is multiplied by the quantity gathering the moment equal to the quantity of characteristic variable, often group sample data all includes the characteristic variable of Q dimension and 1 penetration signal of current time, described output layer nodes is the quantity of described penetration signal, node in hidden layer is empirical value, and excitation function g (x) selects sigmoidal function: g (x)=1/ (1+exp (-x)).

So, owing to aperture characteristic parameter is aperture width and area, each 4 inputs, therefore input layer number is 8.Output layer nodes is the status number of the penetration signal divided, and here output layer nodes is 3.And in the present embodiment, each sample data comprises aperture width, aperture area and first three historical juncture value totally eight input vectors and a penetration signal output vector, amount to 2100 sample datas.

In step 4, randomly draw front 1500 samples of all sample datas as training, all the other 600 samples are as test, the extreme learning machine of test sample is predicted the outcome and makes comparisons with actual result, calculate the prediction accuracy Accuarcy (see formula 1) of test set, wherein K is the sample number consistent with actual result that predict the outcome, and R is for always testing sample number.Consider that node in hidden layer needs artificial setting, therefore to obtain maximum prediction accuracy, then analyze the impact on accuracy of the different node in hidden layer.

Preferably, randomly select an initial node in hidden layer, set up extreme learning machine model and adopt test specimens notebook data to calculate prediction accuracy, choose a node in hidden layer at interval of the interval set and calculate prediction accuracy, choose prediction accuracy the highest time corresponding node in hidden layer be node in hidden layer used in step 4.It is, looking for an optimum hidden layer node numerical value by test of many times is the node in hidden layer selected.

Fig. 4 show the node in hidden layer impact on extreme learning machine performance, it can be seen that when node in hidden layer is gradually increased, the prediction accuracy of test set is in the trend being gradually reduced, therefore the result according to Fig. 4, finally determines that the node in hidden layer in the present embodiment is 200.

Accuracy=K/R*100% (1)

Further, described step 5 specifically includes:

Step 51: determine hidden layer neuron number, the random connection weight w setting input layer and implicit interlayer_iBiasing with hidden layer neuronFor node in hidden layer；

Step 52: calculate the implicit output matrix H of corresponding training sample data_tr；Wherein,

$H_{tr}\left(w_1,...w_{\tilde{N}},b_1,...,b_{\tilde{N}},x_1^{tr},...x_N^{tr}\right)={\left[\begin{array}{ccc}g\left(w_1{x}_1^{tr}+b_1\right)& ...& g\left(w_{\tilde{N}}{x}_1^{tr}+b_{\tilde{N}}\right)\\ \begin{array}{c}.\\ .\\ .\end{array}& \begin{array}{c}\\ ...\\ \end{array}& \begin{array}{c}.\\ .\\ .\end{array}\\ g\left(w_1{x}_N^{tr}+b_1\right)& ...& g\left(w_{\tilde{N}}{x}_N^{tr}+b_{\tilde{N}}\right)\end{array}\right]}_{N\×\tilde{N}}$

For the training sample data of input, N is the group number of training sample data；

Step 53: adopt method of least square to solve hidden layer output matrix H_trMoore-Penrose generalized inverse

Step 54: calculate output weightWhereinT_tr=[t₁,…,t_N]^T, and T_trFor training sample output matrix.

Further, step 6 specifically includes:

Step 61: calculate the implicit output matrix H of corresponding test specimens notebook data_te；Wherein,

$H_{te}\left(w_1,...w_{\tilde{N}},b_1,...,b_{\tilde{N}},x_1^{te},...x_M^{te}\right)={\left[\begin{array}{ccc}g\left(w_1{x}_1^{te}+b_1\right)& ...& g\left(w_{\tilde{N}}{x}_1^{te}+b_{\tilde{N}}\right)\\ \begin{array}{c}.\\ .\\ .\end{array}& \begin{array}{c}\\ ...\\ \end{array}& \begin{array}{c}.\\ .\\ .\end{array}\\ g\left(w_1{x}_M^{te}+b_1\right)& ...& g\left(w_{\tilde{N}}{x}_M^{te}+b_{\tilde{N}}\right)\end{array}\right]}_{M\×\tilde{N}}$

For Q dimensional feature variable data in test specimens notebook data, M is the group number of test specimens notebook data, M and N's and the n that is in Fig. 3；

Step 62: the output T of calculating limit learning machine_te=H_teβ；

Step 63: the penetration signal type that element value is corresponding in the output row of extreme learning machine is as judged result, and this element value is positive integer, and each element value is in 1,2 P, and corresponding corresponding 1st label, the 2nd label P label；

Step 64: the judged result of extreme learning machine and actual penetration signal are contrasted, calculates classification accuracy, extreme learning machine model is verified, and actual penetration signal is in test specimens notebook data except Q dimensional feature variable dataPenetration signal data in addition.

Owing to training set and test set randomly generate, in order to verify the quality of extreme learning machine performance better, then test specimens originally being calculated ten times respectively, obtain final prediction accuracy, wherein node in hidden layer is fixed as 200.

Table 1 extreme learning machine the result

Table 1 shows the classification accuracy of ten times and final Average Accuracy, it can be seen that the classification accuracy of extreme learning machine model is substantially all more than 90%.And after being averaged calculating, more than 90%, it is possible to reach very accurate prediction effect.

The present invention proposes a kind of penetration signal decision method based on the little hole characteristic in the back side in welding technology field, obtaining, back side aperture image basis extracts aperture characteristic parameter through image procossing, and choose penetration signal sample and carry out extreme learning machine model learning, by the extreme learning machine model set up, the penetration signal of aperture is judged.Find that this extreme learning machine model prediction accuracy rate is high through checking, more than 90% can be reached, and the study of extreme learning machine model is without iteration, fast convergence rate, has good Generalization Capability.

Above are only the preferred embodiments of the present invention, the present invention is not played any restriction effect.Any person of ordinary skill in the field; without departing from the scope of technical scheme; the technical scheme that the invention discloses and technology contents are made the variations such as any type of equivalent replacement or amendment; all belong to the content without departing from technical scheme, still fall within protection scope of the present invention.

Claims (9)

1. the penetration signal decision method based on the little hole characteristic in the back side, it is characterised in that including:

Step 1: plasma arc welding with adjustable polarity parameters is divided into P penetration signal, P is the integer more than or equal to 2, and is demarcated as the 1st label, the 2nd label P label respectively, chooses the back side aperture image of plasma arc welding with adjustable polarity parameters as data source；

Step 2: gather the back side aperture view data under each penetration signal, aperture view data is carried out feature extraction, it is determined that multiple characteristic variables；

Step 3: choose plasma arc welding with adjustable polarity parameters sample data under each penetration signal, often group sample data all includes characteristic variable and penetration signal；

Step 4: sample data is randomly divided into training sample data and test specimens notebook data, training sample data are used for learning extreme learning machine model, test specimens notebook data is used for verifying extreme learning machine model, determine the input layer of extreme learning machine model to be set up, hidden layer and output layer nodes, choose the excitation function of extreme learning machine；

Step 5: using training sample data as the input of extreme learning machine, carry out the study of extreme learning machine penetration signal decision model, and then set up extreme learning machine model, judged the penetration signal of aperture according to the aperture characteristic variable of input by described extreme learning machine model；

Step 6: adopt test specimens notebook data that the extreme learning machine model set up is verified, the prediction accuracy of this extreme learning machine model that checking is set up.

2. the penetration signal decision method based on the little hole characteristic in the back side as claimed in claim 1, it is characterized in that, in step 1, actual penetration signal is divided into: partial penetration, complete penetration and three kinds of penetration signal of cutting, partial penetration being set as, the 1st label, complete penetration are set as the 2nd label, and cutting is set as the 3rd label.

3. the penetration signal decision method based on the little hole characteristic in the back side as claimed in claim 1, it is characterised in that in step 1, using the back side aperture image that obtains through visual sensing system as data source.

4. the penetration signal decision method based on the little hole characteristic in the back side as claimed in claim 1, it is characterized in that, in step 2, characteristic variable is the aperture characteristic parameter that aperture image obtains through image processing method and camera calibration, and this characteristic variable includes aperture width and aperture area.

5. the penetration signal decision method based on the little hole characteristic in the back side as claimed in claim 4, it is characterised in that in step 3, often group sample data all includes current time and the characteristic variable of first three historical juncture acquisition, and also includes the penetration signal of current time.

6. the penetration signal decision method based on the little hole characteristic in the back side as claimed in claim 1, it is characterized in that, in step 4, described input layer number is the dimension Q of described characteristic variable, this dimension Q is multiplied by the quantity gathering the moment equal to the quantity of characteristic variable, often group sample data all includes the characteristic variable of Q dimension and 1 penetration signal of current time, described output layer nodes is the quantity of described penetration signal, node in hidden layer is empirical value, and excitation function g (x) selects sigmoidal function: g (x)=1/ (1+exp (-x)).

7. the penetration signal decision method based on the little hole characteristic in the back side as claimed in claim 6, it is characterized in that, randomly select an initial node in hidden layer, set up extreme learning machine model and adopt test specimens notebook data to calculate prediction accuracy, choose a node in hidden layer at interval of the interval set and calculate prediction accuracy, choose prediction accuracy the highest time corresponding node in hidden layer be node in hidden layer used in step 4.

8. the penetration signal decision method based on the little hole characteristic in the back side as claimed in claim 6, it is characterised in that described step 5 specifically includes:

Step 51: determine hidden layer neuron number, the random biasing b connecting weights wi and hidden layer neuron setting input layer and implicit interlayer_i, For node in hidden layer；

Step 52: calculate the implicit output matrix H of corresponding training sample data_tr；Wherein,

$H_{tr}(w_1,\mathrm{...},w_{\tilde{N}},b_1,\mathrm{...},b_{\tilde{N}},x_1^{tr},\mathrm{...}{x}_N^{tr})={\left[\begin{array}{ccc}g\left(w_1{x}_1^{tr}+b_1\right)& ...& g\left(w_{\tilde{N}}{x}_1^{tr}+b_{\tilde{N}}\right)\\ .& & .\\ .& ...& .\\ .& & .\\ g\left(w_1{x}_N^{tr}+b_1\right)& ...& g\left(w_{\tilde{N}}{x}_N^{tr}+b_{\tilde{N}}\right)\end{array}\right]}_{N\×\tilde{N}}$

For the training sample data of input, N is the group number of training sample data；

Step 53: adopt method of least square to solve hidden layer output matrix H_trMoore-Penrose generalized inverse

Step 54: calculate output weightWhereinT_tr=[t₁,…,t_N]^T, and T_trFor training sample output matrix.

9. the penetration signal decision method based on the little hole characteristic in the back side as claimed in claim 8, it is characterised in that step 6 specifically includes:

Step 61: calculate the implicit output matrix H of corresponding test specimens notebook data_te；Wherein,

$H_{te}(w_1,\mathrm{...},w_{\tilde{N}},b_1,\mathrm{...},b_{\tilde{N}},x_1^{te},\mathrm{...}{x}_M^{te})={\left[\begin{array}{ccc}g\left(w_1{x}_1^{te}+b_1\right)& ...& g\left(w_{\tilde{N}}{x}_1^{te}+b_{\tilde{N}}\right)\\ .& & .\\ .& ...& .\\ .& & .\\ g\left(w_1{x}_M^{te}+b_1\right)& ...& g\left(w_{\tilde{N}}{x}_M^{te}+b_{\tilde{N}}\right)\end{array}\right]}_{M\×\tilde{N}}$

For Q dimensional feature variable data in test specimens notebook data, M is the group number of test specimens notebook data；

Step 62: the output T of calculating limit learning machine_te=H_teβ；

Step 63: the penetration signal type that element value is corresponding in the output row of extreme learning machine is as judged result, and this element value is positive integer, and each element value is in 1,2 P, and corresponding corresponding 1st label, the 2nd label P label；

Step 64: the judged result of extreme learning machine and actual penetration signal are contrasted, calculates classification accuracy, extreme learning machine model is verified, and actual penetration signal is in test specimens notebook data except Q dimensional feature variable dataPenetration signal data in addition.