CN111476311A - Anchor chain flash welding quality online detection method based on incremental learning - Google Patents
Anchor chain flash welding quality online detection method based on incremental learning Download PDFInfo
- Publication number
- CN111476311A CN111476311A CN202010309872.1A CN202010309872A CN111476311A CN 111476311 A CN111476311 A CN 111476311A CN 202010309872 A CN202010309872 A CN 202010309872A CN 111476311 A CN111476311 A CN 111476311A
- Authority
- CN
- China
- Prior art keywords
- sample
- samples
- unqualified
- electrode position
- qualified
- 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
- 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
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24133—Distances to prototypes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/30—Computing systems specially adapted for manufacturing
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Analysing Materials By The Use Of Radiation (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses an increment learning-based online detection method for the flash welding quality of an anchor chain, which comprises the following steps of: collecting M unqualified samples and N qualified samples; making the signal lengths of the samples consistent through piecewise linear interpolation; carrying out normalization processing on the signals; calculating the distance between unqualified samples and establishing a distance matrix; synthesizing a new sample after randomly extracting unqualified samples, and calculating the distance between the new sample and the existing unqualified samples; judging whether the number of unqualified samples is the same as that of qualified samples; constructing a convolutional neural network, and training a model; outputting the prediction result of the model on the test set; training the model by using the newly added sample; and outputting the prediction result of the model after the incremental learning on the test set. The invention effectively increases the number of unqualified samples, solves the problem of unbalance of the samples and improves the generalization capability of the model; the problem that the number of samples for the flash welding of the anchor chain is increased day by day is solved.
Description
Technical Field
The invention belongs to a welding quality detection method, and particularly relates to an anchor chain flash welding quality online detection method based on incremental learning.
Background
With the development of marine engineering and shipping industry, anchor chains are developed from original plant ropes, forged anchor chains and cast anchor chains to the current welded anchor chains. The anchor chain is formed by connecting a plurality of chain rings, is a special chain for buffering external force borne by ships and ocean engineering equipment, needs to keep good mechanical property for a long time in a severe seawater environment, and the quality of the anchor chain can directly influence the safety of lives and properties of related personnel. At present, the quality of the anchor chain is mainly detected through a tension test before leaving a factory, and for the whole anchor chain, only one chain link is broken due to the quality problem, the whole anchor chain can lose efficacy, and the loss caused by the chain link is huge.
In the production process of the anchor chain, flash welding is a core link of anchor chain production, and key mechanical property indexes of tensile load, breaking load, impact load and the like of the anchor chain are determined to a great extent. Because a series of complex physical and chemical reactions occur during the flash welding of the anchor chain, a simple and effective mathematical model is difficult to establish, and the surface characteristics after welding are not obvious, so that the timely and effective detection of the welding quality is a technical problem which puzzles the industry for a long time, the Chinese patent CN109242023A discloses 'an on-line assessment method for the flash welding quality of the anchor chain based on DTW and MDS', the invention realizes the visualization of data based on DTW and MDS, and realizes the on-line assessment of the flash welding quality of the anchor chain by a Dirichlet process mixed model. The disadvantage is that the invention does not take into account that in the actual production of the anchor chain, the number of qualified samples is much greater than the number of rejected samples, which is a typical data imbalance. The Chinese patent CN109886298A discloses a 'weld quality detection method based on a convolutional neural network', and the method carries out weld quality analysis on a weld image based on the convolutional neural network, avoids complicated detection steps, can automatically position and capture a weld area, and realizes effective detection on the weld quality. The method has the disadvantages that the anchor chain flash welding joint is sealed in the base metal, an effective welding seam image cannot be obtained, the number of welding samples is not considered to be increased day by day, the increased samples bring new welding characteristics, and a fixed and unchangeable model is difficult to adapt to the changes, so that the long-term detection of the welding quality is not facilitated.
Disclosure of Invention
The purpose of the invention is as follows: the invention aims to provide an on-line detection method for the flash welding quality of an anchor chain based on incremental learning, which solves the problems that a sample is unbalanced, cannot adapt to a newly added sample and has insufficient detection accuracy.
The technical scheme is as follows: the invention relates to an increment learning-based online detection method for the flash welding quality of an anchor chain, which comprises the following steps of:
(1) collecting M unqualified samples S ═ S1,S2,…,Si,…,SMT and N qualified samples T ═ T1,T2,…,Ti,…,TN};
(2) Enabling the signal lengths of the unqualified samples and the qualified samples to be consistent through piecewise linear interpolation;
(3) carrying out normalization processing on signals of unqualified samples and qualified samples;
(4) calculating the distance between unqualified samples and establishing a distance matrix;
(5) randomly extracting samples from the unqualified samples, finding the sample closest to the extracted sample according to the distance matrix, extracting the sample and the sample closest to the sample to synthesize a new unqualified sample, and calculating the distance between the new sample and the existing unqualified sample;
(6) judging whether the number of the new unqualified samples is the same as that of the qualified samples, if so, executing the step (7); if not, executing the step (5);
(7) constructing a convolutional neural network, selecting half of new unqualified samples and half of qualified samples as a training set to train the model, substituting the rest of new unqualified samples and qualified samples as a test set into the model, and outputting a prediction result of the model on the test set;
(8) judging whether a new sample is added, if so, executing the step (9); if not, completing model training;
(9) training the model by using the newly added sample, and updating the parameters of the model after the training is finished;
(10) and substituting unqualified samples and qualified samples in the test set into the model after incremental learning to obtain the prediction probability of the samples belonging to each category, and outputting the prediction result of the model after incremental learning.
Wherein the step (2) specifically comprises:
l will be mixedi(1)An integer of successive incrementsAs a non-conforming sample SiAt the interpolation node in the pre-heating stage,is an electrode position signal A of the interpolation node corresponding to the preheating stagei(1)The piecewise linear interpolation function r (x) is expressed as:
whereinFor preheating stage electrode position signal Ai(1)The interpolation basis function of (a) is specifically expressed as:
if x ∈ [ xj,xj+1]Then the electrode position signal A in the preheating stagei(1)The piecewise linear interpolation function r (x) of (a) is expressed as:
wherein, unqualified sample S ═ { S ═ S1,S2,…,Si,…,SMT and T ═ T1,T2,…,Ti,…,TNThe signal of the ith sample is denoted as Si=[Ai,Bi]The lengths in the preheating stage, the continuous flashing stage and the upsetting stage are L respectivelyi(1)、Li(2)And Li(3)Electrode position signal A of a rejected sampleiIs shown as Ai=[Ai(1),Ai(2),Ai(3)]The electrode position signal during the preheating phase is expressed asThe electrode position signal of the successive flash phases is represented asThe electrode position signal during the upset phase is shown asCurrent signal BiIs represented as Bi=[Bi(1),Bi(2),Bi(3)]The current signal during the preheating phase is represented asThe current signal for successive flash phases is represented asCurrent signal representation during upset forgingThe ith qualified sample is denoted as Ti=[Ci,Di]The lengths of the preheating stage, the continuous polishing stage and the upsetting stage are Qi(1)、Qi(2)And Qi(3)Electrode position signal C of the qualified sampleiIs represented as Ci=[Ci(1),Ci(2),Ci(3)]The electrode position signal during the preheating phase is expressed asThe electrode position signal of the continuous flash is expressed asThe upset electrode position signal is shown asCurrent signal DiIs denoted by Di=[Di(1),Di(2),Di(3)]The current signal during the preheating phase is represented asThe current signal for successive flash phases is represented asThe current signal during the upset phase is shown as
The length of the unqualified sample after the piecewise linear interpolation is G, and the lengths of the preheating stage, the continuous flashing stage and the upsetting stage are G respectively1、G2And G3The electrode position signal in the preheating stage after interpolation isContinuous flashing stage by the same principleThe electrode position signal ofThe electrode position signal in the upset stage isA′i=[A′i(1),A′i(2),A′i(3)]Is a defective sample SiElectrode position signal AiThe result of piecewise linear interpolation is like B'i=[B′i(1),B′i(2),B′i(3)]Is a defective sample SiCurrent signal BiResult of piecewise linear interpolation, S'i=[A′i,B′i]Is a defective sample SiAnd performing piecewise linear interpolation on the qualified sample according to the formula after the piecewise linear interpolation, wherein the length of the qualified sample is G, and the lengths of the preheating stage, the continuous flashing stage and the upsetting stage are G respectively1、G2And G3,C′i=[C′i(1),C′i(2),C′i(3)]As a qualified sample TiElectrode position signal CiThe result after piecewise linear interpolation is D'i=[D′i(1),D′i(2),D′i(3)]Is a qualified sample SiCurrent signal D ofiPiecewise linear interpolated result, T'i=[C′i,D′i]As a qualified sample TiAnd (5) performing piecewise linear interpolation.
The step (3) is specifically as follows:
for each unqualified sample S'iNormalization is performed using the following formula:
to each notQualified sample S'iNormalization is performed using the following formula:
wherein, a'i,maxIs an electrode position signal A'iMaximum value of a'i,minIs an electrode position signal A'iMinimum value of (A ″)i=[a″i,1,a″i,″2,…,a″i,j,…,a″i,G]Is an electrode position signal A'iNormalizing the result; b'i,maxIs a current signal B'iMaximum value of, b'i,minIs a current signal B'iMinimum value of (1), B ″)i=[b″i,1,b″i,2,…,b″i,j,…,b″i,G]Is a current signal B'iNormalized result, S ″)i=[A″i,B″i]Is unqualified sample S'iNormalizing the signal of the qualified sample according to the formula, wherein the normalized electrode position signal is C ″)i=[c″i,1,c″i,2,…,c″i,j,…,c″i,G]The current signal is Di=[d″i,1,d″i,2,…,d″i,j,…,d″i,G],T″i=[C″i,D″i]Is qualified sample T'iAnd (5) normalizing the result.
The step (4) is specifically as follows:
unqualified sample SiAnd S ″)jThe distance between them is calculated as follows:
establishing a distance matrix H of M according to the calculated distance:
wherein h isi,jDenotes a failed sample S ″)iAnd S ″)jThe distance between them; a ″)i,kDenotes a failed sample S ″)iElectrode position signal A ″)iThe kth number of (1); a ″)j,kDenotes a failed sample S ″)jElectrode position signal A ″)jThe kth number of (1); b ″)i,kDenotes a failed sample S ″)iCurrent signal B ″)iThe kth number of (1); b ″)j,kDenotes a failed sample S ″)jCurrent signal B ″)jThe kth number in (1).
The specific steps of calculating the distance between the new sample and the existing unqualified sample in the step (5) are as follows:
for unqualified sample SiAnd S ″)jThe electrode position signals of (a):
r=rand(0.1,0.9)
Anew=r×A″i+(1-r)×A″j
wherein r represents a randomly generated number between 0.1 and 0.9, AnewRepresenting the electrode position signal after synthesis;
for unqualified sample SiAnd S ″)jThe current signals are synthesized:
Bnew=r×B″i+(1-r)×B″j
Bnewrepresenting the current signal after synthesis;
Snew=[Anew,Bnew]adding a row and a column to a distance matrix H for storing new samples after synthesis
SnewDistance from an existing rejected sample.
The step (7) is specifically as follows:
constructing a convolutional neural network model, wherein the convolutional neural network model comprises 1 input layer, 2 convolutional layers, 2 pooling layers, 1 full-connection layer and 1 output layer; wherein 2 convolution layers respectively adopt 8 channels and 16 channels, and the sizes of convolution kernels are all 1 x 3; 2 pooling layers adopt maximum pooling, and the size of each pooling core is 1 x 2; the number of the neurons of the full connection layer is 3200; 2 categories of the output layer are qualified and unqualified;
updating parameters of the model by adopting back propagation and gradient descent, enabling the drop rate dropout to be 0.2 during training, and storing the model after iterative training to obtain a model for detecting the flash welding quality of the anchor chain;
and substituting unqualified samples and qualified samples in the test set into the model to obtain the prediction probability of each class of unqualified and qualified samples, wherein the class with higher probability is the prediction result of the model.
The step (9) is specifically as follows:
extracting M 'unqualified samples and N' qualified samples from the newly added samples, and performing piecewise linear interpolation processing on the signals of the samples according to the step (2);
normalizing the signals of the samples according to the step (3);
calculating the distance between M 'newly added unqualified samples according to the step (4), calculating the distance between the M' newly added unqualified samples and the existing unqualified samples, and adding M 'rows and M' columns to a distance matrix H for storing the calculated distance;
extracting a sample from M' newly-added unqualified samples, finding a sample closest to the sample to synthesize a new sample, and adding a row and a column to a distance matrix H for storing the distance between the synthesized new sample and the existing unqualified sample;
judging whether the number of unqualified samples is the same as that of qualified samples, if so, executing the next step, and if not, executing the previous step;
and inputting the newly added sample into the model for training, and updating the parameters of the model after the training is finished.
Has the advantages that: the method utilizes a piecewise linear interpolation method to enable the data lengths of the samples to be consistent, normalizes the samples to eliminate dimensional influence between electrode position signals and current signals, calculates the distance between unqualified samples by adopting Euclidean distance, synthesizes new samples by randomly extracting the unqualified samples, effectively increases the number of the unqualified samples, solves the problem of unbalance of the samples, and improves the generalization capability of the model; a convolution neural network is established, so that the real-time detection of the flash welding quality of the anchor chain is realized; the model is trained by adopting an incremental learning method, so that the problem that the number of anchor chain flash welding samples is increased day by day is solved, and the problem that the model is forgotten catastrophically is avoided; the invention can effectively detect the quality of the anchor chain flash welding in time with higher accuracy, improves the production efficiency, reduces the production cost and ensures the safety of the ship navigation and ocean engineering platform.
Drawings
FIG. 1 is a schematic flow diagram of the present invention;
FIG. 2 is a graph of electrode position signals for rejected samples versus qualified samples;
FIG. 3 is a graph of current signals for a fail sample versus a pass sample;
FIG. 4 is a graph of two electrode position signals after piecewise linear interpolation and normalization and a synthesized electrode position signal;
FIG. 5 is a graph of two current signals after piecewise linear interpolation and normalization and a synthesized current signal;
FIG. 6 is a block diagram of a convolutional neural network;
FIG. 7 is a graph of the predicted results of incremental learning on a test set.
Detailed Description
The invention is further described below with reference to the accompanying drawings.
The invention discloses an online detection method for the flash welding quality of an anchor chain based on incremental learning, which comprises the following steps of:
(1) 40 unqualified samples were collected, S ═ S1,S2,…,Si,…,S40And 400 qualified samples T ═ T1,T2,…,Ti,…,T400}。
(2) The signal lengths of the samples are made uniform by piecewise linear interpolation.
The signals for the 40 failed samples and the 400 pass samples are both comprised of electrode position signals and current signals. The ith failing sample is denoted Si=[Ai,Bi](as shown in fig. 2 and 3), L for the preheating stage, the continuous flashing stage, and the upsetting stage, respectivelyi(1)、Li(2)And Li(3). Electrode position signal AiIs shown as Ai=[Ai(1),Ai(2),Ai(3)]The electrode position signal during the preheating phase is expressed asThe electrode position signal of the successive flash phases is represented asThe electrode position signal during the upset phase is shown asCurrent signal BiIs represented as Bi=[Bi(1),Bi(2),Bi(3)]The current signal during the preheating phase is represented asThe current signal for successive flash phases is represented asThe current signal during the upset phase is shown asThe ith qualified sample is denoted as Ti=[Ci,Di]The lengths of the preheating stage, the continuous polishing stage and the upsetting stage are Qi(1)、Qi(2)And Qi(3)Electrode position signal C of the qualified sampleiIs represented as Ci=[Ci(1),Ci(2),Ci(3)]Electrode position signal during preheating phaseIs shown asThe electrode position signal of the successive flash phases is represented asThe electrode position signal during the upset phase is shown asCurrent signal DiIs denoted by Di=[Di(1),Di(2),Di(3)]The current signal during the preheating phase is represented asThe current signal of the continuous flash is represented asThe current signal during the upset phase is shown as
L will be mixedi(1)An integer of successive incrementsAs a non-conforming sample SiAt the interpolation node in the pre-heating stage,is an electrode position signal A of the interpolation node corresponding to the preheating stagei(1)The piecewise linear interpolation function r (x) is expressed as:
whereinFor preheating stage electrode position signal Ai(1)The interpolation basis function of (a) is specifically expressed as:
if x ∈ [ xj,xj+1]Then the electrode position signal A in the preheating stagei(1)The piecewise linear interpolation function r (x) of (a) is expressed as:
the length of the unqualified sample after the piecewise linear interpolation is G, and the lengths of the preheating stage, the continuous flashing stage and the upsetting stage are G respectively1、G2And G3The electrode position signal in the preheating stage isThe electrode position signals during the successive flash phases areThe electrode position signal in the upset stage isA′i=[A′i(1),A′i(2),A′i(3)]Is a defective sample SiElectrode position signal AiThe result of piecewise linear interpolation is like B'i=[B′i(1),B′i(2),B′i(3)]Is a defective sample SiCurrent signal BiResult of piecewise linear interpolation, S'i=[A′i,B′i]Is a defective sample SiAnd (5) performing piecewise linear interpolation. Performing piecewise linear interpolation on the qualified sample, wherein the length of the qualified sample is G, and the lengths of the preheating stage, the continuous flashing stage and the upsetting stage are G respectively1、G2And G3,C′i=[C′i(1),C′i(2),C′i(3)]As a qualified sample TiElectrode position signal CiThe result after piecewise linear interpolation is D'i=[D′i(1),D′i(2),D′i(3)]Is a qualified sample SiCurrent signal D ofiPiecewise linear interpolated result, T'i=[C′i,D′i]As a qualified sample TiAnd (5) performing piecewise linear interpolation.
(3) For each unqualified sample S'iNormalization is performed by using the following formula:
wherein a'i,maIs an electrode position signal A'iMaximum value of a'i,minIs an electrode position signal A'iMinimum value of (A ″)i=[a″i,1,a″i,2,…,a″i,j,…,a″i,G]Is an electrode position signal A'iNormalizing the result; b'i,maxIs a current signal B'iMaximum value of, b'i,minIs a current signal B'iMinimum value of (1), B ″)i=[b″i,1,b″i,2,…,b″i,j,…,b″i,G]Is a current signal B'iNormalizing the result; s ″)i=[A″i,B″i]Is unqualified sample S'iThe normalized results are shown in fig. 4 and 5. Normalizing the signal of the qualified sample according to the formula, and normalizing the electrode position signalIs Ci=[c″i,1,c″i,2,…,c″i,j,…,c″i,G]The current signal is Di=[d″i,1,d″i,2,…,d″i,j,…,d″i,G],T″i=[C″i,D″i]Is qualified sample T'iAnd (5) normalizing the result.
(4) Calculate the unqualified sample S ″)iAnd S ″)jThe distance between:
the distances between 40 rejected samples are calculated according to the formula above, and a 40 x 40 distance matrix H is established:
wherein h isi,jDenotes a failed sample S ″)iAnd S ″)jThe distance between them; a ″)i,kDenotes a failed sample S ″)iElectrode position signal A ″)iThe kth number of (1); a ″)j,kDenotes a failed sample S ″)jElectrode position signal A ″)jThe kth number of (1); b ″)i,kDenotes a failed sample S ″)iCurrent signal B ″)iThe kth number of (1); b ″)j,kDenotes a failed sample S ″)jCurrent signal B ″)jThe kth number in (1).
(5) Randomly extracting an unqualified sample S ″ from unqualified samplesiFinding out unqualified sample S ″, based on distance matrix HiOne unqualified sample S' nearest to the samplej。
(6) Synthesis of a New sample S from an unqualified samplenew=[Anew,Bnew]Adding one row and one column to the distance matrix H for storing a new sample S, as shown in fig. 4 and 5newThe distance between the sample and an existing unqualified sample is specifically as follows:
(6.1) for unqualified sample S ″)iAnd S ″)jThe electrode position signals of (a):
r=rand(0.1,0.9)
Anew=r×A″i+(1-r)×A″j
wherein r represents a randomly generated number between 0.1 and 0.9, AnewRepresenting the electrode position signal after synthesis.
(6.2) for unqualified sample S ″)iAnd S ″)jThe current signals are synthesized:
Bnew=r×B″i+(1-r)×B″j
Bnewrepresenting the current signal after synthesis.
(6.3)Snew=[Anew,Bnew]Adding a row and a column to a distance matrix H for storing a new sample S for the new sample after synthesisnewDistance from an existing rejected sample.
(7) Judging whether the number of unqualified samples is the same as that of qualified samples, if so, executing the step 8; and if not, executing the step 5 and the step 6.
(8) A convolutional neural network was constructed as shown in fig. 6, and the model was trained.
Constructing a convolutional neural network model, wherein the convolutional neural network model comprises 1 input layer, 2 convolutional layers, 2 pooling layers, 1 full-connection layer and 1 output layer; wherein 2 convolution layers respectively adopt 8 channels and 16 channels, and the sizes of convolution kernels are all 1 x 3; 2 pooling layers adopt maximum pooling, and the size of each pooling core is 1 x 2; the number of the neurons of the full connection layer is 3200; the 2 categories of the output layer are pass and fail.
And randomly selecting half of unqualified samples and half of qualified samples from the sample set as a training set, and using the rest unqualified samples and qualified samples as a testing set.
And updating parameters of the model by adopting back propagation and gradient descent, keeping the drop rate dropout to be 0.2 during training in order to prevent over-fitting of the network, and storing the model after iterative training to obtain the model for detecting the flash welding quality of the anchor chain.
(9) And substituting unqualified samples and qualified samples in the test set into the model to obtain the prediction probability of each class of unqualified and qualified samples, wherein the class with higher probability is the prediction result of the model.
(10) Judging whether a newly added sample is added, if so, executing the step 11; and if not, finishing the model training.
(11) And training the model by using the newly added sample.
And (11.1) extracting 20 unqualified samples and 200 qualified samples from the newly added samples, and performing piecewise linear interpolation processing on signals of the samples.
And (11.2) carrying out normalization processing on the signals of the samples.
(11.3) calculating the distance between 20 newly-added unqualified samples, calculating the distance between the 20 newly-added unqualified samples and the existing unqualified samples, and adding 20 rows and 20 columns to a distance matrix H for storing the calculated distance.
(11.4) extracting a sample from the 20 newly-added unqualified samples, finding a sample closest to the sample to synthesize a new sample, and adding a row and a column to a distance matrix H for storing the distance between the synthesized new sample and the existing unqualified sample.
(11.5) judging whether the number of unqualified samples is the same as that of qualified samples, if so, executing the step 11.6; if not, go to step 11.4.
(11.6) inputting the newly added samples into the current model for training, and updating the parameters of the model after the training is finished.
(12) Substituting unqualified samples and qualified samples of the test set into the model after incremental learning to obtain the prediction probability of the samples belonging to each category, and outputting the prediction result of the model after incremental learning, wherein the result is shown in fig. 7.
Claims (7)
1. The method for online detecting the flash welding quality of the anchor chain based on incremental learning is characterized by comprising the following steps of:
(1) collecting M unqualified samples S ═ S1,S2,…,Si,…,SMT and N qualified samples T ═ T1,T2,…,Ti,…,TN};
(2) Enabling the signal lengths of the unqualified samples and the qualified samples to be consistent through piecewise linear interpolation;
(3) carrying out normalization processing on signals of unqualified samples and qualified samples;
(4) calculating the distance between unqualified samples and establishing a distance matrix;
(5) randomly extracting samples from the unqualified samples, finding the sample closest to the extracted sample according to the distance matrix, extracting the sample and the sample closest to the sample to synthesize a new unqualified sample, and calculating the distance between the new sample and the existing unqualified sample;
(6) judging whether the number of the new unqualified samples is the same as that of the qualified samples, if so, executing the step (7); if not, executing the step (5);
(7) constructing a convolutional neural network, selecting half of new unqualified samples and half of qualified samples as a training set to train the model, substituting the rest of new unqualified samples and qualified samples as a test set into the model, and outputting a prediction result of the model on the test set;
(8) judging whether a new sample is added, if so, executing the step (9); if not, completing model training;
(9) training the model by using the newly added sample, and updating the parameters of the model after the training is finished;
(10) and substituting unqualified samples and qualified samples in the test set into the model after incremental learning to obtain the prediction probability of the samples belonging to each category, and outputting the prediction result of the model after incremental learning.
2. The online detection method for the flash welding quality of the anchor chain based on the incremental learning as claimed in claim 1, wherein the step (2) is specifically as follows:
l will be mixedi(1)An integer of successive incrementsAs a non-conforming sample SiAt the interpolation node in the pre-heating stage,is an electrode position signal A of the interpolation node corresponding to the preheating stagei(1)The piecewise linear interpolation function r (x) is expressed as:
whereinFor preheating stage electrode position signal Ai(1)The interpolation basis function of (a) is specifically expressed as:
if x ∈ [ xj,xj+1]Then the electrode position signal A in the preheating stagei(1)The piecewise linear interpolation function r (x) of (a) is expressed as:
wherein, unqualified sample S ═ { S ═ S1,S2,…,Si,…,SMT and T ═ T1,T2,…,Ti,…,TNThe signal of the ith sample is composed of an electrode position signal and a current signalThis is represented by Si=[Ai,Bi]The lengths in the preheating stage, the continuous flashing stage and the upsetting stage are L respectivelyi(1)、Li(2)And Li(3)Electrode position signal A of a rejected sampleiIs shown as Ai=[Ai(1),Ai(2),Ai(3)]The electrode position signal during the preheating phase is expressed asThe electrode position signal of the successive flash phases is represented asThe electrode position signal during the upset phase is shown asCurrent signal BiIs represented as Bi=[Bi(1),Bi(2),Bi(3)]The current signal during the preheating phase is represented asThe current signal for successive flash phases is represented asThe current signal during the upset phase is shown asThe ith qualified sample is denoted as Ti=[Ci,Di]The lengths of the preheating stage, the continuous polishing stage and the upsetting stage are Qi(1)、Qi(2)And Qi(3)Electrode position signal C of the qualified sampleiIs represented as Ci=[Ci(1),Ci(2),Ci(3)]The electrode position signal during the preheating phase is expressed asThe electrode position signal of the continuous flash is expressed asThe upset electrode position signal is shown asCurrent signal DiIs denoted by Di=[Di(1),Di(2),Di(3)]The flow signal during the preheating phase is represented asThe current signal for successive flash phases is represented asCurrent signal meter in upsetting stage
The length of the unqualified sample after the piecewise linear interpolation is G, and the lengths of the preheating stage, the continuous flashing stage and the upsetting stage are G respectively1、G2And G3The electrode position signal in the preheating stage after interpolation isIn the same way, the electrode position signals in the successive flashing phases areThe electrode position signal in the upset stage isA′i=[A′i(1),A′i(2),A′i(3)]Is a defective sample SiElectrode position signal AiThe result of piecewise linear interpolation is like B'i=[B′i(1),B′i(2),B′i(3)]Is a defective sample SiCurrent signal BiResult of piecewise linear interpolation, S'i=[A′i,B′i]Is a defective sample SiAnd performing piecewise linear interpolation on the qualified sample according to the formula after the piecewise linear interpolation, wherein the length of the qualified sample is G, and the lengths of the preheating stage, the continuous flashing stage and the upsetting stage are G respectively1、G2And G3,C′i=[C′i(1),C′i(2),C′i(3)]As a qualified sample TiElectrode position signal CiThe result after piecewise linear interpolation is D'i=[D′i(1),D′i(2),D′i(3)]Is a qualified sample SiCurrent signal D ofiPiecewise linear interpolated result, T'i=[C′i,D′i]As a qualified sample TiAnd (5) performing piecewise linear interpolation.
3. The on-line detection method for the flash welding quality of the anchor chain based on the incremental learning as claimed in claim 1, wherein the step (3) is specifically as follows:
for each unqualified sample S'iNormalization is performed using the following formula:
wherein, a'i,maxIs an electrode position signal A'iMaximum value of a'i,minIs an electrode position signal A'iMinimum value of (A ″)i=[a″i,1,a″i,2,…,a″i,j,…,a″i,G]Is an electrode position signal A'iNormalizing the result; b'i,maxIs an electric currentSignal B'iMaximum value of, b'i,minIs a current signal B'iMinimum value of (1), B ″)i=[b″i,1,b″i,2,…,b″i,j,…,b″i,G]Is a current signal B'iNormalized result, S ″)i=[A″i,B″i]Is unqualified sample S'iNormalizing the signal of the qualified sample according to the formula, wherein the normalized electrode position signal is C ″)i=[c″i,1,c″i,2,…,c″i,j,…,c″i,G]The current signal is Di=[d″i,1,d″i,2,…,d″i,j,…,d″i,G],T″i=[C″i,D″i]As a qualified sample Ti' normalized results.
4. The on-line detection method for the flash welding quality of the anchor chain based on the incremental learning as claimed in claim 1, wherein the step (4) is specifically as follows:
unqualified sample SiAnd S ″)jThe distance between them is calculated as follows:
establishing a distance matrix H of M according to the calculated distance:
wherein h isi,jDenotes a failed sample S ″)iAnd S ″)jThe distance between them; a ″)i,kDenotes a failed sample S ″)iElectrode position signal A ″)iThe kth number of (1); a ″)j,kDenotes a failed sample S ″)jElectrode position signal A ″)jThe kth number of (1); b ″)i,kDenotes a failed sample S ″)iCurrent signal ofB″iThe kth number of (1); b ″)j,kDenotes a failed sample S ″)jCurrent signal B ″)jThe kth number in (1).
5. The on-line detection method for the welding quality of the anchor chain flash welding based on the incremental learning of the claim 1, wherein the specific steps of calculating the distance between the new sample and the existing unqualified sample in the step (5) are as follows:
for unqualified sample SiAnd S ″)jThe electrode position signals of (a):
r=rand(0.1,0.9)
Anew=r×A″i+(1-r)×A″j
wherein r represents a randomly generated number between 0.1 and 0.9, AnewRepresenting the electrode position signal after synthesis;
for unqualified sample SiAnd S ″)jThe current signals are synthesized:
Bnew=r×B″i+(1-r)×B″j
Bnewrepresenting the current signal after synthesis; snew=[Anew,Bnew]Adding a row and a column to a distance matrix H for storing a new sample S for the new sample after synthesisnewDistance from an existing rejected sample.
6. The on-line detection method for the flash welding quality of the anchor chain based on the incremental learning as claimed in claim 1, wherein the step (7) is specifically as follows:
constructing a convolutional neural network model, wherein the convolutional neural network model comprises 1 input layer, 2 convolutional layers, 2 pooling layers, 1 full-connection layer and 1 output layer; wherein 2 convolution layers respectively adopt 8 channels and 16 channels, and the sizes of convolution kernels are all 1 x 3; 2 pooling layers adopt maximum pooling, and the size of each pooling core is 1 x 2; the number of the neurons of the full connection layer is 3200; 2 categories of the output layer are qualified and unqualified;
updating parameters of the model by adopting back propagation and gradient descent, enabling the drop rate dropout to be 0.2 during training, and storing the model after iterative training to obtain a model for detecting the flash welding quality of the anchor chain;
and substituting unqualified samples and qualified samples in the test set into the model to obtain the prediction probability of each class of unqualified and qualified samples, wherein the class with higher probability is the prediction result of the model.
7. The on-line detection method for the flash welding quality of the anchor chain based on the incremental learning as claimed in claim 1, wherein the step (9) is specifically as follows:
extracting M 'unqualified samples and N' qualified samples from the newly added samples, and performing piecewise linear interpolation processing on the signals of the samples according to the step (2);
normalizing the signals of the samples according to the step (3);
calculating the distance between M 'newly added unqualified samples according to the step (4), calculating the distance between the M' newly added unqualified samples and the existing unqualified samples, and adding M 'rows and M' columns to a distance matrix H for storing the calculated distance;
extracting a sample from M' newly-added unqualified samples, finding a sample closest to the sample to synthesize a new sample, and adding a row and a column to a distance matrix H for storing the distance between the synthesized new sample and the existing unqualified sample;
judging whether the number of unqualified samples is the same as that of qualified samples, if so, executing the next step, and if not, executing the previous step;
and inputting the newly added sample into the model for training, and updating the parameters of the model after the training is finished.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010309872.1A CN111476311B (en) | 2020-04-20 | 2020-04-20 | Anchor chain flash welding quality online detection method based on increment learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010309872.1A CN111476311B (en) | 2020-04-20 | 2020-04-20 | Anchor chain flash welding quality online detection method based on increment learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN111476311A true CN111476311A (en) | 2020-07-31 |
CN111476311B CN111476311B (en) | 2023-05-26 |
Family
ID=71755948
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010309872.1A Active CN111476311B (en) | 2020-04-20 | 2020-04-20 | Anchor chain flash welding quality online detection method based on increment learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111476311B (en) |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108038471A (en) * | 2017-12-27 | 2018-05-15 | 哈尔滨工程大学 | A kind of underwater sound communication signal type Identification method based on depth learning technology |
CN109239301A (en) * | 2018-09-12 | 2019-01-18 | 江苏科技大学 | A kind of anchor chain flash welding quality online evaluation method |
CN109242023A (en) * | 2018-09-12 | 2019-01-18 | 江苏科技大学 | A kind of anchor chain flash welding quality online evaluation method based on DTW and MDS |
-
2020
- 2020-04-20 CN CN202010309872.1A patent/CN111476311B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108038471A (en) * | 2017-12-27 | 2018-05-15 | 哈尔滨工程大学 | A kind of underwater sound communication signal type Identification method based on depth learning technology |
CN109239301A (en) * | 2018-09-12 | 2019-01-18 | 江苏科技大学 | A kind of anchor chain flash welding quality online evaluation method |
CN109242023A (en) * | 2018-09-12 | 2019-01-18 | 江苏科技大学 | A kind of anchor chain flash welding quality online evaluation method based on DTW and MDS |
Also Published As
Publication number | Publication date |
---|---|
CN111476311B (en) | 2023-05-26 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Dai et al. | Deep learning assisted vision inspection of resistance spot welds | |
CN110739031B (en) | Supervised prediction method and device for metallurgical sintering process and storage medium | |
CN109141881A (en) | A kind of rotating machinery health evaluating method of depth autoencoder network | |
Liu et al. | Data anomaly detection for structural health monitoring using a combination network of GANomaly and CNN | |
CN108665339B (en) | E-commerce product reliability index based on subjective emotion measure and implementation method thereof | |
CN110992349A (en) | Underground pipeline abnormity automatic positioning and identification method based on deep learning | |
CN113695713B (en) | On-line monitoring method and device for welding quality of water heater liner | |
CN110470729B (en) | Eddy current-based nondestructive testing method for oil field casing pipe defects | |
CN101850410A (en) | Continuous casting breakout prediction method based on neural network | |
CN110287583B (en) | Industrial equipment residual life prediction method based on cyclic neural network | |
Nasiri et al. | Online damage monitoring of SiC f-SiC m composite materials using acoustic emission and deep learning | |
CN112149750A (en) | Water supply network pipe burst identification data driving method | |
CN110608884A (en) | Rolling bearing state diagnosis method based on self-attention neural network | |
CN117312816B (en) | Special steel smelting effect evaluation method and system | |
CN113592939A (en) | Deep learning method for judging size of narrow blood vessel based on coronary angiography image | |
CN115796002A (en) | Strip steel coil defect prediction and diagnosis method based on deep learning | |
CN115673596A (en) | Welding abnormity real-time diagnosis method based on Actor-Critic reinforcement learning model | |
CN114898821A (en) | Alloy performance prediction method based on machine learning model fusion | |
CN111476311A (en) | Anchor chain flash welding quality online detection method based on incremental learning | |
Bauer et al. | Multi-Modal Artificial Intelligence in Additive Manufacturing: Combining Thermal and Camera Images for 3D-Print Quality Monitoring. | |
CN104504265A (en) | Method for safety evaluation of monitoring information of in-service bridge | |
CN112802011A (en) | Fan blade defect detection method based on VGG-BLS | |
CN109242023B (en) | DTW and MDS-based anchor chain flash welding quality online evaluation method | |
CN115457307A (en) | Chemical process fault diagnosis method based on improved residual error network | |
CN109239301B (en) | Anchor chain flash welding quality online evaluation method |
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 |