CN112200000A - Welding stability recognition model training method and welding stability recognition method - Google Patents
Welding stability recognition model training method and welding stability recognition method Download PDFInfo
- Publication number
- CN112200000A CN112200000A CN202010952930.2A CN202010952930A CN112200000A CN 112200000 A CN112200000 A CN 112200000A CN 202010952930 A CN202010952930 A CN 202010952930A CN 112200000 A CN112200000 A CN 112200000A
- Authority
- CN
- China
- Prior art keywords
- signal
- welding
- segmentation
- signals
- model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
- G06F2218/04—Denoising
- G06F2218/06—Denoising by applying a scale-space analysis, e.g. using wavelet analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Signal Processing (AREA)
- Image Analysis (AREA)
Abstract
The invention relates to a welding stability recognition model training method and a welding stability recognition method, wherein in the model training method, firstly, a segmentation scale is optimally designed according to a welding signal, then, a plurality of groups of segmentation signals are obtained according to a segmentation strategy, and multi-scale feature vectors of the segmentation signals are extracted and used for training a machine learning model to obtain a welding stability recognition model; in the identification method, firstly, signal segmentation is carried out according to the segmentation scale determined by the model to obtain a group of identification signals, and the stability prediction result can be obtained by extracting the characteristics of the identification signals and inputting the characteristics to the welding stability identification model. Compared with the prior art, the method increases the identification performance of the welding stability identification model to various welding unstable states, enhances the adaptability to unstable fluctuation of different time frequencies and forming defects in the welding process, enables the model to identify more levels of welding instability factors, and has stronger generalization performance in the face of complex and unknown identification problems.
Description
Technical Field
The invention relates to the field of signal feature extraction and machine learning, in particular to a welding stability recognition model training method and a welding stability recognition method.
Background
Modern automated and intelligent welding processes place new requirements on the stability of the welding process, weld formation, and identification of surface defects, etc. In recent years, with the rapid development of the machine learning field, various machine learning algorithms are applied to the field of intelligent monitoring and identification of welding processes. The machine learning model collects various signals collected in the welding process, calculates and analyzes the relationship between the various signals and the stability, the weld forming and the welding defects of the welding process, and fits to form a model for monitoring the welding process and predicting the welding result.
Due to the complexity of the physical and chemical processes involved in the welding process, the wide variety, instability, and noise-rich nature of the signals produced by the welding process, it is often difficult to achieve the desired results using simple machine learning models. Theoretically, the monitoring and recognizing system of the welding process based on the machine learning can be improved by starting from two aspects of signal feature extraction and a machine learning algorithm. However, the improvement of the machine learning algorithm has higher requirements on mathematical theory, programming capability and computer technology, and is difficult in the field of industrial application, so researchers often try various signal processing and feature extraction methods to select appropriate signal features to form a feature data set, so that the performance of the welding process monitoring machine learning model is improved.
The general processing process comprises the steps of segmenting signals collected in the welding process, extracting characteristic parameters of each segment, using the characteristic parameters for training a machine learning model, and identifying the unstable forming defect of a welding seam formed in the welding process by the trained machine learning model.
In the paper "EMD-PNN based influenced plasma electrical signals" (Journal of Manufacturing Processes,2019, 45: 642-51), Huang Yiming and other signal processing methods based on wavelet packet decomposition and empirical mode decomposition perform noise reduction and reconstruction processing on plasma voltage signals acquired in the laser welding process, and perform feature extraction on the reconstructed signals. In the process of feature extraction, the width ranges of signal fluctuation caused by weld forming defects in the time domain are compared, and the average number of 5 wave troughs in each 1000 sampling points is found, so that the original signals are segmented according to a mode that the length of each segment is 200 sampling points, statistical parameters are obtained for each segment of signals and used for training a machine learning model, and the weld forming unstable defects formed in the welding process are identified. The comprehensive recognition accuracy of the finally trained machine learning model is 90.16%. However, in the model, the length of the signal segment is selected based on manual analysis and only contains a fixed value, and if the fluctuation period of the signal is not unique, or the fluctuation period of the signal changes with the welding process, or the signal fluctuation is complex, and it is difficult to manually determine the period range of the signal fluctuation, it will be difficult to perform the signal segmentation step when using the method. In addition, the sampling frequency of the acquisition equipment used by the model is 100kHz, the actual time length of 200 sampling points is 0.002 s, and the use of the extremely short segment length can possibly make the fluctuation of a larger time period not be distinguished, so that the model finally ignores the fluctuation of a signal with a large period, and the accuracy of the model can possibly be reduced.
In the article "A quality diagnosis method of GMAW based on improved Empirical Mode Decomposition and Empirical learning machine" (Journal of Manufacturing Processes 2020, 54: 120-8), Huang Yong et al designs a machine learning electric arc welding surface shaping monitoring system based on the signal processing method of Adaptive Noise set Empirical Mode Decomposition With Adaptive Noise, CEEMDAN. The CEEMDAN method is used by the CEEMDAN method to separate a signal into sub-signals of different frequency bands, the energy ratio of the sub-signals relative to a total signal and the energy entropy value of the total signal are respectively obtained and used as a feature vector training machine learning model, and the prediction accuracy rate of 98.75% is obtained. In this model, the length of the signal segment is directly set to 4096 sample points, corresponding to an acquisition time of 0.4096 s. Under the characteristic extraction strategy of the signal segmentation length and the sub-signal energy ratio, short-time signal sudden change or fluctuation is ignored by a characteristic extraction algorithm, so that a machine learning model cannot predict welding forming defects caused by the signal instantaneous sudden change or fluctuation.
In the paper EMD-based pulsed TIG welding process position detection and position diagnosis using GA-SVM (Journal of Materials Processing Technology,2017,239: 92-102), Huang Yiming et al extract the intensity signal of the arc spectral characteristic line in the aluminum alloy short-circuit arc welding process for training the identification model of the welding air hole defect. When building a data set, they simply collect 15 sampling points in each pulse peak area as a segment, extract their statistical information, and finally obtain a model with a prediction accuracy of 92.5%. In the feature extraction method, each segment of the signal is obtained by sampling from a peak area, the nonuniformity of the width of the peak area and the signal fluctuation of which the period exceeds the peak width are not considered, and the information of the signal fluctuation with large period cannot be obtained.
In the prior art, the length of a signal segment is often set manually, and for a specific problem, the length of the signal segment can be selected according to sampling frequency, signal characteristics and the like. However, in more cases, when a researcher constructs a data set for machine learning based on the collected signals, a specific strategy for determining the length of signal segments is not provided, a randomly set method is usually used, experimental rigor is lacked, and certain uncertain factors are easily brought to the adaptability of a prediction model. The single signal segment or the randomly divided signal segments may cause the problems of insufficient signal feature extraction, insufficient model applicability, and the like. The existing machine learning prediction model has limited recognition capability and low recognition accuracy when facing signal instability of different scales.
Disclosure of Invention
The invention aims to overcome the defects in the prior art and provide a welding stability recognition model training method and a welding stability recognition method, which use multiple scales to segment signals and extract parameters, use the signal segment parameters of the multiple scales as the parameters of model training, increase the recognition performance of the model on multiple welding unstable states, enhance the adaptability of the model on unstable fluctuation and forming defects of different time frequencies in the welding process, enable the model to recognize more layers of welding unstable factors, and have stronger generalization performance in the face of complex and unknown recognition problems.
The purpose of the invention can be realized by the following technical scheme:
a welding stability recognition model training method comprises the following steps:
a 1: the method comprises the steps of signal marking, namely obtaining an original welding signal and a welding seam corresponding to the original welding signal, marking a time area corresponding to a stably formed welding seam in the original welding signal by using a stable label, and marking a time area corresponding to an unstably formed welding seam in the original welding signal by using an unstable label to obtain an original data set;
a 2: signal preprocessing, namely reserving an actual welding signal corresponding to an actual welding process in the original welding signal, and filtering the actual welding signal to obtain a stable welding signal;
a 3: determining segmentation scales, determining n segmentation scales and a signal segment length value of each segmentation scale: l1,l2,…,ln；
a 4: signal segmentation, namely segmenting the stable welding signals by using n segmentation scales respectively to obtain m groups of segmentation signals, wherein each group of segmentation signals comprises n signal segments with different lengths;
a 5: extracting features, namely obtaining the feature parameters of n signal segments in each group of segmented signals through feature extraction, wherein all the feature parameters in one group of segmented signals are integrated to be used as multi-scale feature vectors of the group of segmented signals, and the m multi-scale feature vectors form a feature data set;
a 6: and (3) model training, namely dividing the characteristic data set into a training set and a testing set, training the machine learning identification model and testing the identification performance, if the identification performance is unqualified, executing the step a3, otherwise, ending the training to obtain the welding stability identification model.
Further, the raw weld signal is one or more one-dimensional signals.
Further, the step a3 includes the following steps:
a 31: determining to use n segmentation scales according to the type of the welding signal, setting an initial length value of each segmentation scale, and acquiring a section of test welding signal from the stable welding signal;
a 32: updating the current length values of n segmentation scales, segmenting the test welding signals by using the n segmentation scales respectively to obtain a plurality of groups of segmentation signals, obtaining multi-scale feature vectors of each group of segmentation signals through feature extraction, wherein all the multi-scale feature vectors form a feature data set, dividing the feature data set into a training set and a testing set, training and testing the recognition performance of the machine learning recognition model, and recording the performance of the model;
a 33: controlling the length values of other segment scales to be unchanged, adjusting the length value of the segment scale to be optimized, and executing the step a32 until the n segment scales are optimized;
a 34: selecting a length value combination of the segmentation scale with the best model performance, and determining the length values of different signal segments of n segmentation scales: l1,l2,…,ln。
Still further, the model performance includes a time consumption T including a data partitioning time consumption and a model training time consumption, and a model accuracy a ═ nc/ng) Wherein n isgNumber of samples, n, representing test samples for model testingcRepresenting the number of correctly classified samples in the test sample.
Further, the step a4 includes the following steps:
a 41: obtaining signal segment length values of n segment scales: l1,l2,…,lnAnd wherein the maximum signal segment length value lmaxAnd a minimum signal segment length value lmin；
a42：The length of the stable welding signal is obtained to be L, namely the stable welding signal comprises L data points (0,1,2, …, L), the stable welding signal is respectively segmented by using n segmentation scales, and the length value of the signal segment is LiThe segment size of (i ═ 1,2,3, …, n) divides the stable weld signal into m segments, where the starting data point of the k-th signal segment (k ═ 1,2,3, …, m) is [ l ═ lmax+(k-1)lmin-li]End data point is [ lmax+ (k-1)lmin]If the situation that the complete segmentation cannot be obtained occurs at the tail of the stable welding signal, the incomplete part is abandoned;
a 43: so as to obtain m groups of cutting signals,
each group of the segmented signals comprises n signal segments with different lengths, the starting data points of the n signal segments in each group of the segmented signals are different, and the ending data points are the same.
Further, the kth group of split signals corresponds to a data point [ l ] in the stable welding signalmax+(k-1)lmin]The time of day.
Further, the machine in the step a6 learns the identification model as a support vector machine classification model.
A welding stability identification method comprises the following steps:
b 1: taking the current moment as the ending moment of the signal, and forwardly intercepting a section of actual welding signal;
b 2: filtering the actual welding signal to obtain a stable welding signal;
b 3: respectively segmenting the stable welding signals by using n preset segmentation scales to obtain a group of identification signals, wherein the group of identification signals comprises n signal segments with different lengths;
b 4: obtaining characteristic parameters of n signal segments in the group of identification signals through characteristic extraction, and taking the characteristic parameters after being integrated as multi-scale characteristic vectors of the group of identification signals;
b 5: and inputting the multi-scale feature vector into a welding stability recognition model to obtain a stability recognition result.
Further, in the step b3, the preset n segment scales and the signal segment length value of each segment scale are the n segment scales and the signal segment length value of each segment scale determined in the welding stability recognition model training process.
Further, the step b3 includes the following steps:
b 31: acquiring n segmentation scales and a signal segment length value of each segmentation scale: l1,l2,…,ln；
b 32: intercepting n signal segments on the stable welding signal, wherein the length values of the n signal segments are respectively l1,l2,…,lnThe end data points of the n signal segments are tail data points of the stable welding signal;
b 33: the n signal segments with different lengths and the same end data point form a group of identification signals.
Compared with the prior art, the invention has the following beneficial effects:
(1) the method has the advantages that the signal segmentation is carried out by using multiple scales, the parameters are extracted, the signal segment parameters of the multiple scales are used as the parameters of model training, the recognition performance of the model to multiple welding unstable states is improved, the adaptability of the model to unstable fluctuation and forming defects of different time frequencies in the welding process is enhanced, the model can recognize more levels of welding unstable factors, and the method has stronger generalization performance when facing complex and unknown recognition problems.
(2) The segmentation scale is designed according to specific problems and different signals in the welding process, the length value of the segmentation scale is optimized, the segmentation scale combination with the best model performance is selected, the theoretical explanatory performance is increased for identifying the model, and the pertinence of the model to the specific problems is improved.
(3) A multi-scale signal segmentation strategy is designed, multiple scale signals exist at the same time, signal mutation or fluctuation within a short time and signal change and period within a long time can be fully considered, and the accuracy of the model is greatly improved.
Drawings
FIG. 1 is a schematic flow chart of the present invention;
FIG. 2 is a schematic diagram of a multi-scale signal segmentation strategy according to the present invention;
FIG. 3 is a schematic diagram illustrating the formation result and unstable area division of the welding surface in the embodiment;
FIG. 4 is a schematic diagram of an original welding signal collected in an embodiment;
FIG. 5 is a diagram illustrating a stable welding signal after preprocessing in an embodiment;
FIG. 6 is a diagram illustrating multi-scale signal segmentation in an embodiment;
FIG. 7 is the model accuracy for different short segment lengths after the long segment is fixed in the example;
FIG. 8 is a graph showing the model calculation times for different short segment lengths after the long segment is fixed in the embodiment;
FIG. 9 is a graph showing the model accuracy for different long segment lengths after the short segment is fixed in the example;
FIG. 10 is a graph showing the model calculation times for different long segment lengths after the short segment is fixed in the embodiment.
Detailed Description
The invention is described in detail below with reference to the figures and specific embodiments. The present embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation manner and a specific operation process are given, but the scope of the present invention is not limited to the following embodiments.
Example 1:
the specific flow of a training method of a welding stability recognition model and a welding stability recognition method is shown in fig. 1, and the segmentation strategy of multi-scale signal segmentation is shown in fig. 2.
In the embodiment, the voltage signal and the current signal are used as welding signals, a machine learning model is trained by collecting the welding current and arc voltage signals in the plasma lap welding process, and a monitoring and identifying model of the surface forming stability of the plasma lap welding is established.
In the embodiment, the adopted welding material is 304L stainless steel, the plate thickness is 1.2mm, the diameter of a welding gun nozzle is 1.6mm, the diameter of a tungsten electrode is 1.5mm, the internal shrinkage of the tungsten electrode is 3mm, the welding current is set to be 40A, the welding speed is 20cm/min, the shielding gas and the ion gas are 99.99% argon, the gas flow is 12L/min, and an Intel Core i5-8300H CPU is used by a computer for model building, training and testing. Meanwhile, in order to train the detection capability of the machine learning model on unstable welding, factors causing welding instability are increased by methods of artificially changing the arc length, changing the gap between two plates and the like, and the forming of a plurality of unstable welding seam surfaces is obtained, wherein the forming of one welding seam surface is shown in fig. 3.
The electric signal acquisition result of the welding process corresponding to the weld joint in fig. 3 is shown in fig. 4, and the sampling frequency of the signal is 20000 Hz.
The method comprises the steps of firstly marking an original welding signal, and analyzing the forming state of a welding seam according to actual engineering requirements. And marking the unstable region, finding out a welding time period corresponding to the unstable region and a corresponding original signal segment, and marking the segment of signal as the unstable signal region. In this embodiment, the label of the unstable signal region is denoted by "1", and the labels of the remaining stable regions are denoted by "0".
The analysis of the welding seam forming state can be manually distinguished according to experience, and can also identify an unstable area in a welding seam image through an image processing method.
And then, preprocessing the original welding signal, wherein the preprocessing comprises two steps of effective signal interception and signal filtering. And intercepting an effective area in the original welding signal by an observation method, namely removing useless signals measured before the welding starts and after the welding ends, and reserving an actual welding signal corresponding to an actual welding process.
In addition to manual observation, the interception of the valid signal can also set a threshold value through the change of the signal, and the actual welding signal is directly reserved by using a computer.
The method for filtering the actual welding signal obtained by interception by using a wavelet filtering process comprises the following steps:
1) decomposing the actual welding signal into a series of wavelets by using a discrete wavelet decomposition mode;
2) and for each wavelet, filtering out components with higher frequency by adopting a threshold method, and finally recombining and reconstructing each filtered wavelet into a filtered waveform.
The filtering parameters need to be optimized through manual testing, and stable welding signals meeting the standard or the precision requirement are obtained. After a series of attempts, the fundamental wave of the wavelet filtering is finally set as a "db 4" wavelet, the decomposition layer number is 11, the Threshold method is a global Threshold (Universal Threshold) method, and the Threshold type is a Soft Threshold (Soft Threshold) type. In addition, in order to remove the residual part of the impact noise after the wavelet filtering, the signal is further filtered by adopting a median filtering with a window size of 25. The waveform of the stable welding signal obtained after the preprocessing is shown in fig. 5.
And then, determining the appropriate segmentation scale for multi-scale division of the signal, wherein the main idea is to test the model performance of a series of segmentation scales by adopting a control variable method and select the segment length collocation with the best model performance. The model performance indexes selected by the embodiment are model classification time consumption T and model accuracy A, the time consumption T comprises data division time consumption and model training time consumption, and the model accuracy A is (n)c/ng) Wherein n isgNumber of samples, n, representing test samples for model testingcRepresenting the number of correctly classified samples in the test sample.
Because the classification task of the stability recognition model trained by using the voltage signal and the current signal is simpler, only two segment scales are adopted, namely a long segment scale (the length value is recorded as l)1) And short segment length (length value l)2. In order to reduce the calculation time, only one segment of the stable welding signal is selected as the test welding signal.
In order to optimize the length values of the two segment scales, the length value of the long segment scale is fixed to be a fixed value, which is set to be 12800, the length value of the short segment scale is changed, and the test welding signal is segmented by using the two segment scales.
As shown in fig. 6, three groups of segmented signals are obtained, each group of segmented signals includes two signal segments with different lengths and the same end time, and the first group of segmented signals: long fragment [0, l1]Short segment, [ l ]1-l2,l1]And, a second set of split signals: long section [ l2,l1+l2]Short segment, [ l ]1,l1+l2]And, third group division signal: long segment [2l2,l1+2l2]Short segment, [ l ]1+l2,l1+2l2]。
And respectively extracting the principal component of each signal segment in each group of the segmentation signals as a characteristic parameter by using a principal component analysis method, obtaining 12 signal segments because the welding signals are voltage signals and current signals, and extracting 12 characteristic parameters in total, wherein the principal component analysis processes of the short segments and the long segments are independent.
And summarizing the 12 characteristic parameters according to groups, summarizing the characteristic parameters of all signal segments in each group of segmented signals to obtain a multi-scale characteristic vector of the group of segmented signals, wherein the ending time of all the signal segments in each group of segmented signals is the same, and the multi-scale characteristic vector corresponds to the ending time. All multi-scale feature vectors together constitute a feature data set.
For ease of training and testing, the feature data set is divided into a test set and a training set on a 70% and 30% scale. In fact, the training set and data set may be divided in other proportions, or cross-tested.
And inhibiting overfitting by adopting a K-fold cross validation method in the training process of the model, wherein K is 5, and recording the time consumption T of training the model and the model accuracy A.
In this embodiment, the machine learning model adopts a support vector machine classification model of a radial basis kernel, the hyper-parameters C and γ of the model are optimized by using a grid search method, and the parameter search range is set as: {0.1,1,10,100,1000}. According to the actual problem to be solved, a neural network model and the like can be selected for training, and parameters and an optimization method can be changed.
The length value of the long segment scale is fixed to be 12800, the length value of the short segment scale is changed, and the test results of various performances of the model are shown in fig. 7 and 8. It can be found from the figure that when the length value of the long segment scale is fixed, the model accuracy a increases with the decrease of the length value of the short segment scale, when the length value of the short segment scale and the length value of the long segment scale are close, the model accuracy a is low and unstable, and as the length value of the short segment scale decreases, the time consumption T of the model also increases greatly. The length value of the short segmentation scale of the model is properly set to be about 256 by comprehensively considering the factors of model accuracy and model time consumption.
Similarly, the length value of the short segment scale of the model is set to 256, the length value of the long segment scale of the model is changed, the model is built and trained by the same method, and the test results of various performances of the model are shown in fig. 9 and 10. It can be seen from the figure that the accuracy a of the model increases with the increase of the length value of the long segment scale, and when the length value of the long segment scale approaches the length value of the short segment scale, the accuracy a of the model is low and unstable. As the length value of the long segmentation scale increases, the model elapsed time T increases by a small amount. The performance index of the model is comprehensively considered, and the length of the long section is determined to be about 12800, which is more appropriate.
Determining a length value l for a long segment scale 112800, length value of short segment size l2Is 256.
Next, the actual welding signal is divided.
In practice, the strategy of dividing the welding signal using n segment scales may be referred to as "long segment timing instants, short segment timing offsets".
For a length value of l1The actual welding signal of length L is divided into m segments, the starting data point of the k-th signal segment (k ═ 1,2,3, …, m) being [ (k-1) L2]End data point is [ l1+(k-1)l2]。
For a length value of l2The actual welding signal of length L is divided into m segments, the k-th signal segment (k 1,2,3, …,m) starting data point is [ l1+(k-2)l2]End data point is [ l1+(k-1)l2]。
And if the situation that the complete segment cannot be obtained occurs at the end of the signal, discarding the incomplete part.
And respectively segmenting the current data and the voltage data by using 256 and 12800 obtained through optimized screening as length values of segmentation scales to obtain m groups of segmentation signals, wherein each group of segmentation information comprises signal segments with different lengths. And extracting the characteristic parameters of each segment, and integrating all the characteristic parameters in each group of segmented signals to form a multi-scale characteristic vector, wherein the multi-scale characteristic vector represents the characteristics of a certain specific moment, and the specific moment is the moment when all signal segments in the group of segmented signals end.
And putting all multi-scale feature vectors into the feature data set, training and testing the model by using the multi-scale feature vectors, wherein the training method is the same as that when the length value of the segmentation scale is optimized, and finally obtaining the welding stability recognition model.
And testing the performance of the welding stability recognition model, and if the welding stability recognition model is not qualified, redesigning the type of the segmentation scale, the length value of the segmentation scale and retraining the model until the performance of the welding stability recognition model reaches the standard.
The welding stability recognition model is divided from the data set to the training and testing completion, the total consumed time is 112.984s, most of the time consumed by large-scale data division and model training is used here, and the consumed time of the actual detection process is very little.
The performance test results of the welding stability identification model are shown in the following table:
| type (B) | Accuracy of measurement | Recall rate | F1-value | Number of test samples |
| Label 0 | 99.963％ | 100.000％ | 99.982％ | 2731 |
| Label 1 | 100.000％ | 99.814％ | 99.907％ | 539 |
| Tag averaging | 99.982％ | 99.907％ | 99.944％ | 3270 |
| Weighted average | 99.969％ | 99.969％ | 99.969％ | 3270 |
In the table, the accuracy P indicates the proportion of correct results in all unstable results, the recall ratio R indicates the proportion of successful detection in all unstable regions, the F1-value is a comprehensive index, and the specific value is determined by the following formula:
it can be seen that, in the embodiment, the machine learning model is trained on the basis of multi-scale signal segmentation, and the welding stability model obtained through training has high prediction accuracy. According to the training results of different segmentation scales, when the model is subjected to multi-scale segmentation with different lengths and each segmentation length has a certain difference, the accuracy of the model is remarkably improved, the sequence of the segmentation lengths is designed and optimized for specific problems, the adaptability of the model to the specific problems can be improved, and the comprehensive performance of the model for welding stability detection is improved.
After the model training is finished, the practical application can be carried out. And predicting the stability of the welding seam at a certain moment in the welding process by using the same engineering background, and acquiring the voltage value and the current value in the welding process in real time.
And selecting a certain moment, taking the moment as the ending moment of the signal, forwardly intercepting a section of actual welding signal, and filtering the actual welding signal to obtain a stable welding signal.
And respectively intercepting the signal segments of the current data and the voltage data by using 256 and 12800 obtained by optimized screening as length values of segment scales.
And taking the selected time as the end time of the signal segment, and forwards intercepting the signal segment with the length of 256 as a short message segment. And taking the selected time as the end time of the signal segment, and cutting the signal segment with the length of 12800 forwards to be used as a long signal segment. Since the welding signals are voltage signals and current signals, 4 signal segments are obtained to form a group of identification signals.
And respectively extracting principal components of each signal segment by using a principal component analysis method to serve as characteristic parameters, and synthesizing the characteristic parameters to serve as a multi-scale characteristic vector.
And inputting the multi-scale feature vector into a welding stability model to obtain a prediction result of the welding seam stability at the selected moment.
Claims (10)
1. A welding stability recognition model training method is characterized by comprising the following steps:
a 1: the method comprises the steps of signal marking, namely obtaining an original welding signal and a welding seam corresponding to the original welding signal, marking a time area corresponding to a stably formed welding seam in the original welding signal by using a stable label, and marking a time area corresponding to an unstably formed welding seam in the original welding signal by using an unstable label to obtain an original data set;
a 2: signal preprocessing, namely reserving an actual welding signal corresponding to an actual welding process in the original welding signal, and filtering the actual welding signal to obtain a stable welding signal;
a 3: determining segmentation scales, determining n segmentation scales and a signal segment length value of each segmentation scale: l1,l2,…,ln；
a 4: signal segmentation, namely segmenting the stable welding signals by using n segmentation scales respectively to obtain m groups of segmentation signals, wherein each group of segmentation signals comprises n signal segments with different lengths;
a 5: extracting features, namely obtaining the feature parameters of n signal segments in each group of segmented signals through feature extraction, wherein all the feature parameters in one group of segmented signals are integrated to be used as multi-scale feature vectors of the group of segmented signals, and the m multi-scale feature vectors form a feature data set;
a 6: and (3) model training, namely dividing the characteristic data set into a training set and a testing set, training the machine learning identification model and testing the identification performance, if the identification performance is unqualified, executing the step a3, otherwise, ending the training to obtain the welding stability identification model.
2. The welding stability recognition model training method of claim 1, wherein the raw welding signals are one or more one-dimensional signals.
3. The welding stability recognition model training method of claim 1, wherein the step a3 comprises the following steps:
a 31: determining to use n segmentation scales according to the type of the welding signal, setting an initial length value of each segmentation scale, and acquiring a section of test welding signal from the stable welding signal;
a 32: updating the current length values of n segmentation scales, segmenting the test welding signals by using the n segmentation scales respectively to obtain a plurality of groups of segmentation signals, obtaining multi-scale feature vectors of each group of segmentation signals through feature extraction, wherein all the multi-scale feature vectors form a feature data set, dividing the feature data set into a training set and a testing set, training and testing the recognition performance of the machine learning recognition model, and recording the performance of the model;
a 33: controlling the length values of other segment scales to be unchanged, adjusting the length value of the segment scale to be optimized, and executing the step a32 until the n segment scales are optimized;
a 34: selecting a length value combination of the segmentation scale with the best model performance, and determining the length values of different signal segments of n segmentation scales: l1,l2,…,ln。
4. The welding stability recognition model training method of claim 3, wherein the model performance comprises a time T and a model accuracy A, the time T comprises a data partitioning time and a model training time, and the model accuracy A ═ nc/ng) Wherein n isgNumber of samples, n, representing test samples for model testingcRepresenting the number of correctly classified samples in the test sample.
5. The welding stability recognition model training method of claim 1, wherein the step a4 comprises the following steps:
a 41: obtaining signal segment length values of n segment scales: l1,l2,…,lnAnd wherein the maximum signal segment length value lmaxAnd a minimum signal segment length value lmin；
a 42: the length of the obtained stable welding signal is L, namely the stable welding signal contains L numbersAccording to point (0,1,2, …, L), the stable welding signal is signal segmented using n segmentation scales, respectively, with a signal segment length value of LiThe segment size of (i ═ 1,2,3, …, n) divides the stable weld signal into m segments, where the starting data point of the k-th signal segment (k ═ 1,2,3, …, m) is [ l ═ lmax+(k-1)lmin-li]End data point is [ lmax+(k-1)lmin]If the situation that the complete segmentation cannot be obtained occurs at the tail of the stable welding signal, the incomplete part is abandoned;
a 43: so as to obtain m groups of cutting signals,
each group of the segmented signals comprises n signal segments with different lengths, starting data points of the n signal segments in each group of the segmented signals are different, and ending data points are the same.
6. The welding stability recognition model training method of claim 5, wherein the kth group of segmentation signals correspond to data points [ l ] in the stable welding signalmax+(k-1)lmin]The time of day.
7. The method for training the welding stability recognition model according to claim 1, wherein the machine learning recognition model in the step a6 is a support vector machine classification model.
8. A welding stability identification method is characterized by comprising the following steps:
b 1: taking the current moment as the ending moment of the signal, and forwardly intercepting a section of actual welding signal;
b 2: filtering the actual welding signal to obtain a stable welding signal;
b 3: respectively segmenting the stable welding signals by using n preset segmentation scales to obtain a group of identification signals, wherein the group of identification signals comprises n signal segments with different lengths;
b 4: obtaining the characteristic parameters of n signal segments in the group of identification signals through characteristic extraction, and taking the synthesized characteristic parameters as the multi-scale characteristic vector of the group of identification signals;
b 5: inputting the multi-scale feature vector into a trained welding stability recognition model to obtain a stability recognition result, wherein the welding stability recognition model is obtained by training according to the training method of any one of claims 1 to 7.
9. The welding stability recognition method of claim 8, wherein in step b3, the preset signal segment length values of the n segment scales and each segment scale are the signal segment length values of the n segment scales and each segment scale determined in the welding stability recognition model training process.
10. The welding stability identification method of claim 8, wherein the step b3 comprises the steps of:
b 31: acquiring n segmentation scales and a signal segment length value of each segmentation scale: l1,l2,…,ln；
b 32: intercepting n signal segments on the stable welding signal, wherein the length values of the n signal segments are respectively l1,l2,…,lnThe end data points of the n signal segments are tail data points of the stable welding signal;
b 33: the n signal segments with different lengths and the same end data point form a group of identification signals.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010952930.2A CN112200000B (en) | 2020-09-11 | 2020-09-11 | Welding stability recognition model training method and welding stability recognition method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010952930.2A CN112200000B (en) | 2020-09-11 | 2020-09-11 | Welding stability recognition model training method and welding stability recognition method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112200000A true CN112200000A (en) | 2021-01-08 |
| CN112200000B CN112200000B (en) | 2022-10-11 |
Family
ID=74014729
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010952930.2A Active CN112200000B (en) | 2020-09-11 | 2020-09-11 | Welding stability recognition model training method and welding stability recognition method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112200000B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113408483A (en) * | 2021-07-14 | 2021-09-17 | 宝钢日铁汽车板有限公司 | Method for detecting abnormal state of welding process of welding machine based on convolution self-coding network |
| CN114167189A (en) * | 2021-12-08 | 2022-03-11 | 华中科技大学鄂州工业技术研究院 | Welding defect detection method, electronic device and storage medium |
| CN115255566A (en) * | 2022-09-26 | 2022-11-01 | 苏芯物联技术(南京)有限公司 | Welding deviation real-time intelligent detection method based on high-quality time domain characteristics |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1546995A (en) * | 2003-12-12 | 2004-11-17 | 上海交通大学 | Active infrared CCD weld seam detecting method |
| CN108334663A (en) * | 2018-01-05 | 2018-07-27 | 扬州东升汽车零部件股份有限公司 | A kind of hollow stabilizer rod of automobile multi-pass welding technique towards accurate Deformation control |
| CN108898158A (en) * | 2018-05-31 | 2018-11-27 | 华南理工大学 | A kind of weld joint recognition method based on sectional linear fitting |
| CN111242021A (en) * | 2020-01-10 | 2020-06-05 | 电子科技大学 | Distributed optical fiber vibration signal feature extraction and identification method |
| CN111414912A (en) * | 2020-03-31 | 2020-07-14 | 北京博清科技有限公司 | Method, device and equipment for identifying characteristic points of butt-joint type welding seam and storage medium |
-
2020
- 2020-09-11 CN CN202010952930.2A patent/CN112200000B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1546995A (en) * | 2003-12-12 | 2004-11-17 | 上海交通大学 | Active infrared CCD weld seam detecting method |
| CN108334663A (en) * | 2018-01-05 | 2018-07-27 | 扬州东升汽车零部件股份有限公司 | A kind of hollow stabilizer rod of automobile multi-pass welding technique towards accurate Deformation control |
| CN108898158A (en) * | 2018-05-31 | 2018-11-27 | 华南理工大学 | A kind of weld joint recognition method based on sectional linear fitting |
| CN111242021A (en) * | 2020-01-10 | 2020-06-05 | 电子科技大学 | Distributed optical fiber vibration signal feature extraction and identification method |
| CN111414912A (en) * | 2020-03-31 | 2020-07-14 | 北京博清科技有限公司 | Method, device and equipment for identifying characteristic points of butt-joint type welding seam and storage medium |
Non-Patent Citations (2)
| Title |
|---|
| XIANGDONG GAO ETAL.: ""Process stability analysis and weld formation evaluation during disk laser–mag hybrid welding"", 《OPTICS AND LASERS IN ENGINEERING》 * |
| 胥志强 等: ""一种基于分形插值原理的焊缝轨迹识别方法"", 《宁夏电力》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113408483A (en) * | 2021-07-14 | 2021-09-17 | 宝钢日铁汽车板有限公司 | Method for detecting abnormal state of welding process of welding machine based on convolution self-coding network |
| CN114167189A (en) * | 2021-12-08 | 2022-03-11 | 华中科技大学鄂州工业技术研究院 | Welding defect detection method, electronic device and storage medium |
| CN114167189B (en) * | 2021-12-08 | 2023-11-03 | 华中科技大学鄂州工业技术研究院 | Welding defect detection method, electronic equipment and storage medium |
| CN115255566A (en) * | 2022-09-26 | 2022-11-01 | 苏芯物联技术(南京)有限公司 | Welding deviation real-time intelligent detection method based on high-quality time domain characteristics |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112200000B (en) | 2022-10-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112200000B (en) | Welding stability recognition model training method and welding stability recognition method | |
| Chan et al. | Time-frequency sparsity map on automatic partial discharge sources separation for power transformer condition assessment | |
| CN110108992B (en) | Cable partial discharge fault identification method and system based on improved random forest algorithm | |
| CN108804731B (en) | Time series trend feature extraction method based on important point dual evaluation factors | |
| CN106021361A (en) | Sequence alignment-based self-adaptive application layer network protocol message clustering method | |
| CN110726898B (en) | Power distribution network fault type identification method | |
| CN108197647B (en) | Rapid clustering method for automobile starter endurance test data | |
| CN111089856B (en) | Post-processing method for extracting Raman spectrum weak signal | |
| CN111444784B (en) | Power equipment discharge signal separation and classification method based on kernel principal component analysis | |
| CN111331211B (en) | On-line penetration detection method and system for electric spark small hole machining | |
| CN111025100A (en) | Transformer ultrahigh frequency partial discharge signal mode identification method and device | |
| CN112528774B (en) | Intelligent unknown radar signal sorting system and method in complex electromagnetic environment | |
| CN113884844B (en) | Transformer partial discharge type identification method and system | |
| CN116861303A (en) | Digital twin multisource information fusion diagnosis method for transformer substation | |
| CN114397569A (en) | Circuit breaker fault arc detection method based on VMD parameter optimization and sample entropy | |
| CN117076955A (en) | Fault detection method and system for high-voltage frequency converter | |
| CN117708760B (en) | Multi-mode fusion-based switch cabinet multi-source partial discharge mode identification method and system | |
| CN116298881B (en) | Electrical signal motor health early warning method based on channel attention multi-module LMMD | |
| CN113523904A (en) | Cutter wear detection method | |
| CN116503710A (en) | GIS partial discharge type identification method based on self-adaptive convolutional neural network | |
| CN110117798B (en) | Aluminum oxide concentration estimation method and device for aluminum electrolysis | |
| CN109917245B (en) | Ultrasonic detection partial discharge signal mode identification method considering phase difference | |
| CN116008735A (en) | Partial discharge signal extraction method and system based on density | |
| CN112059725A (en) | Cutter wear monitoring method based on EMD-SVM | |
| CN112419304A (en) | One-dimensional data-oriented multi-stage target detection method and device |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |