A Welding Feature Quantity Evaluation Method Based on Monotonic Relevance and Permutation Importance

Abstract

The present invention relates to a method for evaluating welding feature quantities based on monotonic correlation degree and permutation importance, which includes extracting an n-dimensional welding signal feature set from the original welding signal, analyzing the monotonic correlation of the n-dimensional welding signal feature set, eliminating feature variables with a relatively high monotonic correlation degree according to the monotonic correlation distance in the feature space, then using the feature set after eliminating redundancy for model training, calculating the permutation importance of each feature variable based on the trained model, and finally realizing the evaluation of feature variables. The monotonic correlation degree analysis can greatly suppress the redundancy in the feature space, and the permutation importance analysis can screen out the main feature variables, greatly reducing the dimension of the feature space. The feature analysis combining the monotonic correlation degree and the permutation importance and realizing the evaluation of feature variables can not only reduce the redundancy of the feature space and the complexity of the analysis process, but also improve the accuracy of the analysis, and is applicable to the field of welding information processing.

Description

Technical Field

[0001] The present invention belongs to the technical field of welding signal processing, and particularly relates to a method for evaluating welding feature quantities based on monotonic correlation degree and permutation importance. Background Art

[0002] The welding dynamic process is a complex physico-chemical process involving a high degree of coupling of physical fields such as sound, light, heat, force, magnetism, and electricity. For various physical information of the welding dynamic state, corresponding physical signals can be collected through corresponding microphones, cameras, infrared sensors, Hall sensors, etc., providing information such as arc sound, arc spectrum, arc image, weld pool image, and arc voltage of the welding dynamic process. Different information sources require different sensing means to obtain information, and the forms and contents of the obtained information are quite different.

[0003] Extracting multi-dimensional (n-dimensional) physical features from the original physical signals is essentially equivalent to a dimensionality reduction of the original high-dimensional data at one time. However, there are still two questions in the obtained n-dimensional feature space. One is whether the features are redundant, and the other is the contribution of the extracted features to the model. The process of solving these two questions is the further analysis and evaluation of the extracted features, and another dimensionality reduction of the feature space. Redundancy in the feature space is often caused by a strong correlation degree between features. Principal Component Analysis (PCA) is a commonly used data dimensionality reduction method. PCA transforms a set of possibly correlated features into a set of linearly uncorrelated features through an orthogonal transformation. Specifically, it deletes the repeated and redundant features (features with close relationships), retains as few features as possible, so that the retained features are pairwise uncorrelated, and these retained features can reflect the information of the system as much as possible while maintaining the original information. PCA can perform dimensionality reduction processing on a multi-dimensional feature system and convert it into a low-dimensional feature system, but there are still the following deficiencies:

[0004] (1) When the correlation between features in the original multi-dimensional feature system is weak, PCA will still retain more features after dimensionality reduction processing, which is not conducive to system analysis, and PCA has not pointed out the direction to solve this problem;

[0005] (2) For the strongly correlated features that are deleted, PCA only realizes the screening function and does not perform classification processing, that is, it does not clarify what role the deleted features play in the system, or whether there is an association between the deleted features;

[0006] (3) When the retained features are difficult to implement in experimental operations, whether they can be replaced by features with convenient operations. Regarding this problem, PCA is obviously powerless. Summary of the Invention

[0007] Aiming at the technical problems existing in the prior art, the object of the present invention is to provide a method for evaluating welding feature quantities based on monotonic correlation degree and permutation importance, which can greatly suppress the redundancy of the feature space and the monotonic correlation between features, is beneficial to selecting the main feature variables with high permutation importance and having an important impact on the welding process, and improves the accuracy of analysis.

[0008] The object of the present invention is achieved by the following technical solutions:

[0009] A method for evaluating welding feature quantities based on monotonic correlation degree and permutation importance extracts an n-dimensional welding signal feature set from the original welding signal, analyzes the monotonic correlation of the n-dimensional welding signal feature set, eliminates the feature variables with higher monotonic correlation degree according to the monotonic correlation distance of the feature space, then uses the feature set after eliminating redundancy for model training, calculates the permutation importance of each feature variable based on the trained model, and finally realizes the evaluation of the feature variables.

[0010] Further, it includes the following steps:

[0011] Step S1, extract an n-dimensional welding signal feature set Ω from the original welding signal and number the features, perform Spearman monotonic correlation analysis on the feature set Ω to obtain the monotonic correlation degree of each feature variable in the feature set Ω, and the calculation formula is:

[0012]

[0013] In the formula, ρ ij is the Spearman monotonic correlation degree between the feature variables x i and x j , x i , x j (i, j = 1, 2, 3, ···, n) are any two feature variables in the feature set Ω respectively, each feature variable is an N×1 vector, n is the feature dimension, N is the number of samples, X i and X j are the position averages after descending order arrangement of the feature variables x i and x j respectively, X i,p is the p-th position number after descending order arrangement of the feature variable x i , X j,p is the p-th position number after descending order arrangement of the feature variable x j ;

[0014] Step S2, after calculating the Spearman monotonic correlation degree in step S1, screen out the feature set Ω with higher monotonic correlation degree a(a = 1, 2,..., and a ≤ n / 2), calculate each feature set Ω a the monotonic correlation distance between the feature variables in and the entire feature space, and retain each feature set Ω a the one feature variable with the largest monotonic correlation distance in, discard the other feature variables in the feature set Ω a the other feature variables in, and at the same time the feature set Ω also discards the corresponding feature variables, so as to obtain the feature set Ω1 after eliminating redundancy;

[0015] Step S3, use the feature set Ω1 and combine it with K-fold cross-validation and the network search method to train the SVM model, solve each set of hyperparameter combinations of the model, and then determine the SVM model;

[0016] Step S4, evaluate the relative importance of each feature variable in the feature set Ω1 to the SVM model based on the permutation importance method, sort each feature variable in descending order of importance, determine and retain the feature variables ranked in the front and the sum of whose importance exceeds 95%, discard the other feature variables, and form a new feature set Ω′. Each feature variable in the new feature set Ω′ is the main feature variable describing the welding dynamic process.

[0017] Furthermore, when the monotonic correlation degree |ρ ij | ≥ 0.8 in step S1, the feature variables x i and x j have a high monotonic correlation degree and are marked as attribute A1; when 0.3 < |ρ ij | < 0.8, the feature variables x i and x j have a general monotonic correlation degree and are marked as attribute A2; when |ρ ij | ≤ 0.3, the feature variables x i and x j have a low monotonic correlation degree and are marked as attribute A3; when |ρ ij | = 0, the feature variables x i and x j are completely uncorrelated and are marked as attribute A4.

[0018] Furthermore, for the n-dimensional feature space, the number k of feature variables in the feature set Ω with attribute A1 in step S2 a satisfies 2 ≤ k ≤ n - 1.

[0019] Furthermore, expand ρ ij into an n×n feature space, then ρ ij measures the monotonic correlation distance between two feature variables;

[0020] When the number of feature variables k = 2, calculate the feature set Ω aThe monotonic correlation distance between two feature variables. Retain the feature variable with the largest monotonic correlation distance. The calculation formulas for the monotonic correlation distances between two feature variables and the entire feature space are

[0021]

[0022] When 3 ≤ k ≤ n - 1, for the feature set Ω a there is a strong monotonic correlation degree between each pair of feature variables. For each feature set Ω a select the feature variable with the largest monotonic correlation distance as the representative of the feature set Ω a and the calculation formula is

[0023]

[0024] Furthermore, the permutation importance method described in step S4 includes the following process.

[0025] Step S41: Input the feature set Ω1 into the SVM model for prediction, and calculate the accuracy s of the SVM model in the feature set Ω1.

[0026] Step S42: Calculate the permutation importance of each feature variable in the feature set Ω1. The calculation process is as follows. In step S421, randomly permute the order of the feature variable w to be evaluated in the feature set Ω1 so that its corresponding relationship with the sample label y w is disrupted, and obtain the permuted and disordered feature set Ω t,w , where t is the mark of the number of disruptions in each loop. Input the disordered feature set Ω t,w back into the SVM model, and then obtain the accuracy rate s t,w of the SVM model in the disordered feature set Ω t,w ;

[0027] Step S422: Repeat step S421 for T times, and then obtain T accuracy rates s t,w (t = 1, 2,..., T) of the feature variable w. The calculation formula for the permutation importance PI w of the feature variable w is

[0028]

[0029] Step S43: Repeat step S42 until the permutation importance of each feature variable in the feature set Ω1 is traversed and calculated.

[0030] Furthermore, the number of repetitions T of step S42 is equal to the number of feature variables in the feature set Ω1, T = number(Ω1).

[0031] Furthermore, when the feature variable w is relatively important, the accuracy rate s t,wThe value is small, and its permutation importance PI w is high and closer to the original accuracy rate s. Conversely, for the accuracy rate s t,w The value is large, and its permutation importance PI w is low.

[0032] Furthermore, the sample label y w ∈ {1, 2, 3}, where 1 represents the incomplete penetration state, 2 represents the full penetration state, and 3 represents the over-penetration state.

[0033] Furthermore, when the correlation degree among multiple features is high, when calculating the permutation importance of one feature, the remaining features with strong correlation can still play an important role in the model. The permutation importance of multiple feature variables with high correlation degrees will all be low. At this time, repeat steps S1 and S2 to eliminate multiple feature variables with high monotonic correlation degrees and reduce the redundancy of the feature space.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] (1) Based on the analysis of feature variables using Spearman monotonic correlation degree, it can greatly suppress the redundancy of the feature space and the monotonic correlation among features, avoid considering feature variables that have the same effect on the welding dynamic process and increase the complexity of the analysis process, and achieve dimensionality reduction of the feature space once again.

[0036] (2) Based on the analysis of feature variables using Spearman monotonic correlation degree, it provides a potential idea for analyzing the relationship among feature variables in a feature set with high Spearman monotonic correlation degree. When there are deficiencies in analyzing the influence of a certain feature variable on the welding process, feature variables with high monotonic correlation degree with this feature variable and beneficial to the analysis of the welding process can be selected by permutation.

[0037] (3) Based on the analysis of the permutation importance of feature variables, it is beneficial to select the main feature variables with high permutation importance and having an important impact on the welding process, and greatly reduces the dimensionality of the feature space.

[0038] (4) The feature analysis combining monotonic correlation degree and permutation importance is beneficial to reducing the redundancy of the feature space and the complexity of the analysis process, and is also beneficial to improving the accuracy of the analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 is a schematic diagram of the evaluation process of welding feature quantities based on monotonic correlation degree and permutation importance;

[0040] Figure 2 is the calculation result of the Spearman monotonic correlation degree in a 19-dimensional feature space;

[0041] Figure 3 It is based on the results of 5-fold cross-validation and hyperparameter optimization by grid search;

[0042] Figure 4 It is a flowchart for calculating permutation importance;

[0043] Figure 5 It is the permutation importance of 15-dimensional feature variables. Specific implementation manner

[0044] The present invention will be further described in detail below.

[0045] Based on the voltage signal and sound signal in the K-TIG welding process, combined with signal analysis and feature extraction methods, the signal characteristics of different penetration states are analyzed and evaluated, and the selected features are fused at the feature level.

[0046] Figure 1 It is a schematic diagram of the evaluation process of welding feature quantities based on monotonic correlation degree and permutation importance. The evaluation process includes:

[0047] Step S1: Extract the n-dimensional welding signal feature set Ω from the original welding signal, number the features, and calculate the Spearman monotonic correlation degree of the feature variables in the feature set Ω;

[0048] Step S2: Screen out the feature set Ω with a relatively high monotonic correlation degree a (a = 1, 2,... and a ≤ n / 2), calculate the monotonic correlation degree distance between each feature variable and the entire feature space, select the feature variable with the largest monotonic correlation degree distance, discard the other feature variables in the feature set Ω a and discard the corresponding feature variables in the feature set Ω as well, so as to obtain the feature set Ω1 after eliminating redundancy;

[0049] Step S3: Use the feature set Ω1 and combine it with K-fold cross-validation and network search method to train the SVM model, solve each set of hyperparameter combinations of the model, and then determine the SVM model;

[0050] Step S4: Evaluate the relative importance of each feature variable in the feature set Ω1 to the SVM model based on the permutation importance method, sort the feature variable importance in descending order, select the feature variables ranked in the front and with the sum of importance exceeding 95%, discard the other feature variables, and form a new feature set Ω' for describing the welding dynamic process.

[0051] To obtain sufficient voltage signal and sound signal data of different penetration states, by changing the welding current or welding speed, that is, changing the heat input, weldments with different penetration states can be obtained. Finally, the voltage signal and sound signal in the welding process under different penetration states are obtained. After signal preprocessing and feature extraction, the data set (xi , y i ), where \(i\) is the sample subscript, \(i = 1, 2, 3, \cdots, N\); \(N\) is the total number of samples, \(N = 3018\), including 1000 samples in the incomplete penetration state, 809 samples in the full penetration state, and 1209 samples in the over-penetration state; \(y_i\in\{1, 2, 3\}\) is the sample label, representing the incomplete penetration state, full penetration state, and over-penetration state respectively; \(x\ i \in R n is the characteristic variable, and \(n\) is the number of characteristic dimensions. The characteristic variables used in the present invention are a total of 19 dimensions, which are the time-domain statistical characteristics \(D1\), \(R1\), and \(K1\) of the voltage signal, the time-domain statistical characteristics \(R2\) and \(K2\) of the acoustic signal, the characteristic frequencies \(f2\), \(f4\), and \(f5\) of the 2nd, 4th, and 5th frequency bands in the frequency domain of the voltage signal, the frequency band energy \(E\) of the acoustic signal in the frequency domain, and the mel-frequency cepstral coefficients \(MFCC1\sim MFCC10\), and the 19-dimensional characteristic variables are numbered as shown in Table 1.

[0052] Table 1 Characteristic Variables and Their Numbers

[0053]

[0054] In step S1, the calculation formula for the Spearman monotonicity correlation degree of the feature space is

[0055]

[0056] In the formula, \(\rho\ ij is the Spearman monotonicity correlation degree between the characteristic variables \(x\ i and \(x\ j , \(x\ i , \(x\ j (i, j = 1, 2, 3, \cdots, n)\) are any two characteristic variables in the feature set \(\Omega\) respectively, each characteristic variable is an \(N\times1\) vector, \(n\) is the number of characteristic dimensions, \(N\) is the number of samples, \(X\ i and \(X\ j are the position averages after descending order arrangement of the characteristic variables \(x\ i and \(x\ j respectively, \(X\ i,p is the \(p\)th position number after descending order arrangement of the characteristic variable \(x\ i , \(X\ j,p is the \(p\)th position number after descending order arrangement of the characteristic variable \(x\ j ;

[0057] Calculate the monotonic correlation degrees of the 19-dimensional characteristic variables in Table 1 in turn, and the calculation results are as Figure 2 shown. When the monotonic correlation degree \(|\rho\ ij |\geq0.8\), the characteristic variables \(x\ i and \(x\ jHas a high monotonic correlation degree and is marked as attribute A1; when 0.3 < |ρ ij | < 0.8, the feature variable x i and x j Have a general monotonic correlation degree and are marked as attribute A2; when |ρ ij | ≤ 0.3, the feature variable x i and x j Have a low monotonic correlation degree and are marked as attribute A3; when |ρ ij | = 0, the feature variable x i and x j Are completely uncorrelated and are marked as attribute A4. As can be seen from Figure 2 , there are two feature sets with relatively high monotonic correlation degrees in the 19-dimensional feature space, namely 4 - 9 - 10, that is, R2 - E - MFCC1, and 6 - 7 - 8, that is, f2 - f4 - f5. The two cluster classes are respectively extracted from the arc sound signal and the arc voltage signal. The former is related to the arc sound energy, and the latter is related to the main frequency of the characteristic frequency band of the arc voltage signal. It is not difficult to understand that their monotonic correlation degrees are relatively high.

[0058] Based on the analysis of the monotonic correlation degree of the feature space, a feature screening method based on the monotonic correlation distance is proposed to suppress the monotonic correlation and redundancy of the feature space. For an n-dimensional feature space, if there is a feature set with a relatively high monotonic correlation degree, then the number of feature variables k in the feature set Ω a with attribute A1 satisfies 2 ≤ k ≤ n - 1.

[0059] Expand ρ ij to an n × n feature space, then ρ ij measures the monotonic correlation distance between two feature variables. When the number of feature variables k = 2, calculate the monotonic correlation distance between the two feature variables in the feature set Ω a , and retain the feature variable with the largest monotonic correlation distance. The calculation formulas for the monotonic correlation distances between the two feature variables and the entire feature space are

[0060]

[0061] When 3 ≤ k ≤ n - 1, there is a strong monotonic correlation degree between each feature variable in the feature set Ω a . Each feature set Ω a selects the feature variable with the largest monotonic correlation distance as the representative of the feature set Ω a , and the calculation formula is

[0062]

[0063] Finally, the feature screening method based on monotonic correlation distance calculated the monotonic correlation distance of the feature variables in the two feature sets, and the results are shown in Table 2. Therefore, the feature variables f2 and R2 are retained in each feature set, and the monotonic correlation and redundancy of the 19-dimensional feature space are reduced, reducing it to a 15-dimensional feature space. The new feature variables and their encoding are shown in Table 3.

[0064] Table 2 Calculation results of monotonic correlation distance of feature variables in two feature sets

[0065]

[0066] Table 3 Feature variables and their encoding after redundancy reduction

[0067]

[0068] After redundancy reduction in the feature space and determination of the SVM model, the Permutation Importance (PI) method can be used to obtain an evaluation of the importance of the features after the model is trained. First, the 15-dimensional feature variable data in Table 3 is used to train the SVM model in combination with 5-fold cross-validation and grid search. For the hyperparameters C and γ, log2C = [–10, 10], log2γ = [–10, 10], with a step size of 0.2, and a grid search is performed. For each set of hyperparameter combinations (C, γ), a 5-fold cross-validation is used to obtain the model accuracy, and the final result is as follows: Figure 3 As shown in the 100×100×1 (C,γ,Accuracy) hyperplane, the best model accuracy is 99.79%, which occurs when the hyperparameter combination of (C,γ)=(13.337,1.632).

[0069] Permutation importance is a commonly used type of feature importance. Its core idea is that if features are replaced with randomly arranged values, the accuracy of the model will decrease. After obtaining the trained SVM model, the importance of the features to the model is evaluated based on the permutation importance method. The permutation importance method described in step S4 is as follows: Figure 4 As shown, it includes the following processes:

[0070] Step S41, input the feature set Ω1 into the SVM model for prediction, and calculate the accuracy s of the SVM model in the feature set Ω1;

[0071] Step S42, calculate the permutation importance of each feature variable in the feature set Ω1, the calculation process is,

[0072] Step S421, randomly permuting the order of the feature variables w to be evaluated in the feature set Ω1 so that they are consistent with the sample label y wThe corresponding relationship is disrupted to obtain the permuted and disordered feature set Ω t,w , where t is a marker for the number of disruptions in each loop. The disordered feature set Ω t,w is re-input into the SVM model, and then the accuracy rate s of the SVM model in the disordered feature set Ω t,w is obtained t,w ;

[0073] Step S422: Repeat step S421 for T times to obtain T accuracy rates s for the feature variable w t,w (t = 1, 2,..., T). The permutation importance PI of the feature variable w w is calculated as

[0074]

[0075] Step S43: Repeat step S42 until the permutation importance of each feature variable in the feature set Ω1 is traversed and calculated

[0076] Figure 4 The number of repetitions T is equal to the number of feature variables in the feature set Ω1, T = number(Ω1). When the feature variable w is more important, the value of the accuracy rate s t,w is smaller, and its permutation importance PI w is higher and closer to the original accuracy rate s. Conversely, when the value of the accuracy rate s t,w is larger, its permutation importance PI w is lower

[0077] According to the calculation process of the permutation importance, it is not difficult to find that when the correlation degree among multiple features is relatively high, when calculating the permutation importance of one feature, the remaining features with strong correlation can still play an important role in the model, and the permutation importance of multiple feature variables with high correlation degrees will be relatively low. Therefore, it is extremely necessary to reduce the redundancy of the feature space based on the monotonic correlation degree. At this time, repeat step S1 and step S2 to eliminate multiple feature variables with high monotonic correlation degrees and reduce the redundancy of the feature space

[0078] Considering the suppression of accidental factors, in the present invention, the number of repetitions T is taken as 10 times, and the permutation importance of the 15-dimensional feature space is calculated as shown Figure 5 below Figure 5The first four features with the highest permutation importance are 6-2-10-11, namely f2-R1-MFCC5-MFCC6. There are two features each extracted from the arc voltage signal and the arc sound signal. The f2 feature is the most important and is the key feature for analysis, containing the information of the oscillation of the weld pool; the second most important is the R1 feature, representing the energy of the arc voltage signal. Both the time-domain and frequency-domain features of the arc voltage signal are of great significance for identifying the penetration state of the weldment. The third and fourth important features are the Mel-frequency cepstral coefficients MFCC5 and MFCC6 of the arc sound signal. It should be noted that the time-domain features R2 and K2 of the arc sound signal are ranked 6th and 14th respectively, and their permutation importance PI < 0.01. The frequency features of the arc sound signal are of great significance for identifying the penetration state. Further analyzing the physical meaning of the Mel-frequency cepstral coefficients, that is, the Mel-frequency cepstral coefficients reflect the envelope of the spectrum of the arc sound signal in the Mel frequency domain, highlighting the position and size of the resonance peaks. The feature ranked fifth in terms of permutation importance is MFCC10. It is found that MFCC5, MFCC6, and MFCC10 respectively correspond to two obvious resonance peaks, and the frequency bands where the resonance peaks corresponding to MFCC5 and MFCC6 are located overlap with the frequency band of f2 in the frequency domain of the arc voltage signal.

[0079] In summary, the present invention proposes a method for evaluating the monotonic correlation degree of the feature space based on the Spearman monotonic correlation coefficient, reducing the 19-dimensional feature space to 15 dimensions, which greatly suppresses the redundancy of the feature space and the monotonic correlation between features. Inputting the 15-dimensional features with reduced redundancy into the SVM penetration recognition model, it is found that when the hyperparameter combination is (C,γ)=(13.337,1.632), the model accuracy is the highest, reaching 99.79%. The present invention also proposes a feature evaluation method based on permutation importance, and it is found that the top four important features for the penetration recognition task are f2-R1-MFCC5-MFCC6, which come from the arc voltage signal and the arc sound signal respectively. Both the time-domain and frequency-domain features of the arc voltage signal are of high importance for the penetration recognition task, and the frequency-domain features of the arc sound signal are more important than its time-domain features, which is also consistent with the phenomenon that sound signals are extremely important for welders to perform high-quality welding operations in actual welding. The feature with the highest importance for the penetration recognition task is the feature frequency f2 of the arc voltage signal, representing the feature of the oscillation frequency of the weld pool surface.

[0080] The above embodiments are the preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent replacement methods and are all included in the protection scope of the present invention.

Claims

1. A method for evaluating welding feature quantities based on monotonic correlation degree and permutation importance, characterized in that Including the following steps: Step S1: Extract an n-dimensional welding signal feature set Ω from the original welding signal, number the features, and perform a Spearman monotonic correlation analysis on the feature set Ω to obtain the monotonic correlation degree of each feature variable in the feature set Ω. The calculation formula is: where ρ ij is the Spearman monotonic correlation between the feature variable x i and x j , x i , x j are any two feature variables in the feature set Ω, where i, j = 1, 2, 3, ···, n, each feature variable is an N×1 vector, n is the feature dimension, N is the number of samples, X i and X j are the average positions after descending order of the feature variables x i and x j respectively, X i,p is the p-th position number after descending order of the feature variable x i , X j,p is the p-th position number after descending order of the feature variable x j ; Step S2, after calculating the Spearman monotonic correlation in step S1, filter out the feature set Ω with a relatively high monotonic correlation a , where a = 1, 2,... and a ≤ n / 2, calculate the monotonic correlation distance between the feature variables in each feature set Ω a and the entire feature space, retain one feature variable with the largest monotonic correlation distance in each feature set Ω a , discard the other feature variables in the feature set Ω a , and at the same time, the feature set Ω also discards the corresponding feature variables, thereby obtaining the feature set Ω1 after eliminating redundancy; Step S3: Use the feature set Ω1, combine it with K-fold cross-validation and the network search method to train the SVM model, solve each set of hyperparameter combinations of the model, and then determine the SVM model. Step S4: Evaluate the relative importance of each feature variable in the feature set Ω1 to the SVM model based on the permutation importance method, sort each feature variable in descending order according to the importance, determine and retain the feature variables ranked in the front and whose sum of importance exceeds 95%, discard other feature variables, and form a new feature set Ω'. Each feature variable in the new feature set Ω' is the main feature variable describing the welding dynamic process. The permutation importance method described in Step S4 includes the following process: Step S41: Input the feature set Ω1 into the SVM model for prediction, and calculate the accuracy s of the SVM model in the feature set Ω1. Step S42: Calculate the permutation importance of each feature variable in the feature set Ω1. The calculation process is: Step S421: Randomly permute the order of the feature variable w to be evaluated in the feature set Ω1 so that its correspondence with the sample label y w is disrupted, and the permuted and shuffled feature set Ω t,w is obtained. t is a marker for the number of disruptions in each loop. Input the shuffled feature set Ω t,w back into the SVM model, and then obtain the accuracy rate s of the SVM model in the shuffled feature set Ω t,w ; t,w ; Step S422: Repeat step S421 for T times to obtain T accuracies s of the feature variable w, where t = 1, 2,..., T, and the permutation importance PI of the feature variable w t,w is calculated as follows w by the following formula Step S43: Repeat Step S42 until the permutation importance of each feature variable in the feature set Ω1 is traversed and calculated.

2. The evaluation method of welding feature quantity based on monotonic correlation degree and permutation importance according to claim 1, wherein When the monotonic correlation degree |ρ ij | ≥ 0.8 in step S1, the feature variable x i and x j have a high monotonic correlation degree and are marked as attribute A1; when 0.3 < |ρ ij | < 0.8, the feature variable x i and x j have a general monotonic correlation degree and are marked as attribute A2; when |ρ ij | ≤ 0.3, the feature variable x i and x j have a low monotonic correlation degree and are marked as attribute A3; when |ρ ij | = 0, the feature variable x i and x j are completely uncorrelated and are marked as attribute A4.

3. The welding feature quantity evaluation method based on monotonic correlation degree and permutation importance according to claim 2, characterized in that For the n-dimensional feature space, the number k of feature variables in the feature set Ω with the attribute A1 in step S2 a satisfies 2 ≤ k ≤ n - 1.

4. A method for evaluating welding feature quantities based on monotonic correlation degree and permutation importance according to claim 1, characterized in that Expand ρ ij to an n×n feature space, then ρ ij measures the monotonic correlation distance between two feature variables; When the number of feature variables k = 2, calculate the feature set Ω a The monotonic correlation distance between two feature variables, retain the feature variable with the largest monotonic correlation distance. The calculation formulas for the monotonic correlation distances between two feature variables and the entire feature space are When \(3\leq k\leq n - 1\), for the feature set \(\Omega\) a there is a strong monotonic correlation between each pair of feature variables. For each feature set \(\Omega\) a select the feature variable with the maximum monotonic correlation distance as the representative of the feature set \(\Omega\) a and the calculation formula is 5. The evaluation method of welding feature quantity based on monotonic correlation degree and permutation importance according to claim 1, characterized in that The number of repetitions T of Step S42 is equal to the number of feature variables in the feature set Ω1, T = number(Ω1).

6. The welding feature quantity evaluation method based on monotonic correlation degree and permutation importance according to claim 1, wherein When the feature variable w is more important, the accuracy rate s t,w has a smaller value, and its permutation importance PI w is higher and closer to the original accuracy rate s. On the contrary, when the accuracy rate s t,w has a larger value, its permutation importance PI w is lower.

7. The welding feature quantity evaluation method based on monotonic correlation degree and permutation importance according to claim 1, characterized in that Sample label y w ∈ {1, 2, 3}, where 1 represents the incomplete penetration state, 2 represents the full penetration state, and 3 represents the excessive penetration state.

8. A method for evaluating welding feature quantities based on monotonic correlation degree and permutation importance according to claim 1, characterized in that When the correlation degree between multiple features is relatively high, when calculating the permutation importance of one feature, the remaining features with relatively strong correlation can still play an important role in the model. The permutation importance of multiple feature variables with relatively high correlation degrees will be relatively low. At this time, repeat Step S1 and Step S2 to eliminate multiple feature variables with relatively high monotonic correlation degrees and reduce the redundancy of the feature space.

Images