Welding joint residual fatigue life assessment method based on multi-source heterogeneous information fusion

Abstract

The application discloses a kind of welding node residual fatigue life evaluation methods based on multi-source heterogeneous information fusion, including with crack length as judging quantity to the uncertainty factor in the process of fatigue crack initiation and propagation Sensitivity analysis;According to the analysis result, based on bayesian theory, railway steel bridge welding node degradation model is established;Then crack length image, acoustic emission signal and strain signal are collected;Input to deep learning neural network, and obtain crack length;And the crack length is input into railway steel bridge welding node degradation model, and the residual fatigue life of weld point is obtained.The present application can comprehensively consider the uncertainty of monitoring data, material mechanics performance and the model of theory, thereby significantly improve the prediction accuracy of steel bridge weld point residual life.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of railway steel bridge welded joint fatigue life prediction and evaluation, more particularly to a welded joint residual fatigue life evaluation method based on multi-source heterogeneous information fusion. BACKGROUND

[0002] Fatigue cracking has always been a prominent problem faced by railway steel bridge design and operation, and fatigue short cracks or fatigue long cracks that penetrate the thickness direction of the welded joint are often detected, which brings serious safety hazards to the operation of railway steel bridges. Therefore, how to accurately obtain the current crack state at the position of the key welded joint and accurately evaluate its expansion trend and residual life has great research significance and engineering value for ensuring the safe operation of railway steel bridges.

[0003] Current residual life evaluation methods are mainly divided into methods based on fatigue tests and methods based on fracture mechanics theory;

[0004] The former mainly relies on fatigue test methods, i.e., evaluating the residual life of the welded joint based on the S-N curve measured by the test. Fatigue test data have significant random dispersion, and probability analysis needs to be performed on the basis of a sufficient number of test data to find statistical regularity. Due to the limit of research and development cycle and investment, the number of fatigue test specimens of various studies is very limited, and whether the obtained S-N curve meets the specific requirements needs to be demonstrated, and the research results lack the mastery of the regularity of cause and effect. Therefore, it is difficult to obtain the accurate value of the residual fatigue life of the railway steel bridge based on the fatigue test method.

[0005] The latter is to analyze and predict the fatigue long crack propagation of the key welded joint based on the theory of fracture mechanics, which can avoid the time-consuming and cost problems caused by fatigue tests, but since fracture mechanics cannot evaluate the fatigue crack initiation and short crack initiation stage life, it cannot give accurate results for the low stress amplitude fatigue life prediction of the railway steel bridge during service. At the same time, the service environment of the railway steel bridge is complex, which makes the fatigue long crack often affected by various uncertain factors during the expansion process, thereby causing low prediction accuracy of the fatigue long crack propagation.

[0006] Therefore, the fracture mechanics theory cannot be directly applied to the accurate value of the residual fatigue life of the railway steel bridge. SUMMARY

[0007] Therefore, the present application fuses the information of various information sources of the service steel bridge service health monitoring system, establishes a set of real-time accurate value evaluation method for the residual life of the welded joint based on the fatigue life theory, and comprehensively considers the uncertainty of the monitoring data, material mechanical properties and theoretical model.

[0008] In order to achieve the above object, the present application adopts the following technical solutions:

[0009] A welding joint residual fatigue life evaluation method based on multi-source heterogeneous information fusion, comprising,

[0010] Sensitivity analysis is performed on the uncertainty factors in the fatigue crack initiation and propagation process by taking the crack length as the judgment quantity; according to the analysis result, a railway steel bridge welding joint degradation model is established based on the Bayesian theory;

[0011] Collecting crack length images, acoustic emission signals and strain signals; inputting into a deep learning neural network to obtain the crack length; inputting the crack length into the railway steel bridge welding joint degradation model to obtain the residual fatigue life of the welding point.

[0012] Specifically, the welding joint residual fatigue life evaluation method based on multi-source heterogeneous information fusion disclosed in the present application has the beneficial effects compared with the prior art, including:

[0013] 1) A high-precision fatigue crack real-time data recognition method based on a deep learning neural network is established, which greatly improves the fatigue crack recognition efficiency and precision of the railway steel bridge;

[0014] 2) In order to realize the fatigue damage evolution of the railway steel bridge welding joint under the action of random load in time series, a dynamic Bayesian network model facing the fatigue crack initiation and propagation is established, the uncertainty factors existing in the fatigue crack initiation and propagation process are taken as the state nodes of the model, and the precise evaluation of the fatigue crack size and residual fatigue life under random load is realized;

[0015] 3) In order to reduce the influence of the uncertainty factors in the fatigue crack initiation and propagation process on the evaluation precision, an uncertainty parameter reasoning algorithm based on the instrumented indentation method and the particle filter reasoning method is established, the key material data of the fatigue dangerous part of the welding joint collected by the instrumented indentation method are used to realize the step-by-step reduction of the influence of multiple uncertainty factors on the fatigue crack initiation and propagation process;

[0016] 4) Based on the real fatigue crack data collected by multiple sensors, the uncertainty parameters in the fatigue crack initiation and propagation process are updated step by step, and the dynamic following of the fatigue crack propagation and the real-time accurate evaluation of the residual fatigue life are realized. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the accompanying drawings in the following description only aim at the embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort on the basis of the provided drawings.

[0018] Figure 1 A flow chart of a welding joint residual fatigue life evaluation method based on multi-source heterogeneous information fusion;

[0019] Figure 2 A flow chart for solving a cyclic constitutive curve;

[0020] Figure 3 A flow chart for solving a cyclic resistance curve;

[0021] Figure 4 A flow chart for solving a crack propagation rate curve;

[0022] Figure 5 A flow chart of a multi-source information crack size identification algorithm based on deep learning. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without any creative effort do not deviate from the scope of the present application.

[0024] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, however, the present application can be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the concept of the present application, therefore, the present application is not limited to the specific embodiments disclosed below.

[0025] In view of the deficiencies of the existing crack life evaluation method, the present application respectively adopts deep learning, dynamic Bayesian network and fatigue short crack nonlinear prediction theory.

[0026] Specifically, in order to reduce the influence of uncertain factors on the evaluation of the residual fatigue life of the railway steel bridge welding joint, the present embodiment provides a welding joint fatigue residual life evaluation method based on multi-source information fusion, as shown in Figure 1 , which comprises:

[0027] The sensitivity of the uncertainty factors in the fatigue crack initiation and propagation process is analyzed by taking the crack length as the judgment variable; and a railway steel bridge welded joint degradation model is established based on the Bayesian theory according to the analysis results.

[0028] The crack length image, acoustic emission signal and strain signal are collected; the crack length is obtained by inputting into the deep learning neural network; and the residual fatigue life of the weld point is obtained by inputting the crack length into the railway steel bridge welded joint degradation model.

[0029] Embodiment one: establishing a railway steel bridge welded joint degradation model

[0030] In this embodiment, the fatigue crack under variable amplitude load is taken as a specific object, the local stress-strain method, IBESS method and high-precision fatigue long crack propagation prediction method are used to construct the physical state model of fatigue crack initiation, short crack and long crack propagation, so as to characterize the multi-stage degradation process of fatigue crack initiation, short crack propagation and long crack propagation of the welded joint under variable amplitude load; then the relationship between the typical uncertainty factors in the fatigue crack initiation and propagation process is analyzed, and the dynamic performance degradation model of the fatigue crack initiation and propagation is established based on the dynamic Bayesian network. The steps include:

[0031] Step one, the sensitivity of the uncertainty factors in the fatigue crack initiation and propagation process is analyzed by taking the crack length as the judgment variable; including:

[0032] When the crack length is less than 0.1mm, the crack is in the crack initiation stage, and the local stress-strain method is used to solve the fatigue life;

[0033] When the crack length is between 0.1mm and 0.5mm, the IBESS method is used to solve the fatigue life;

[0034] When the crack length is greater than 0.5mm, the long crack propagation prediction algorithm is used to solve the fatigue life.

[0035] 1) When the local stress-strain method is used to solve the fatigue life, the load spectrum and weld size need to be measured, according to the knowledge of structural mechanics, the nominal stress spectrum at the fatigue crack of the railway steel bridge welded joint is obtained, combined with the cyclic constitutive curve of the steel and the Neuber criterion, the local elastic-plastic stress spectrum data at the crack is obtained, and the residual fatigue life is obtained combined with the cumulative damage criterion. The related formula is as follows:

[0036] The fatigue crack initiation life of the welded joint in the two-stage model is calculated by the strain-fatigue life formula of the smooth specimen. The Basquin-Coffin-Manson formula describes the strain-fatigue life of the smooth specimen, which can be expressed as:

[0037]

[0038] where Δε is the plastic strain amplitude, E is the elastic modulus, N i is the fatigue crack initiation life, ε' f is the fatigue ductility coefficient, σ' f is the fatigue strength coefficient, c is the fatigue ductility exponent, and b is the fatigue strength exponent. When the effect of mean stress on the strain-fatigue life of smooth specimens is considered, the equation (1) needs to be modified, and the modified equation is:

[0039]

[0040] where σ m is the mean stress, and the other symbols are the same as in equation (1).

[0041] When the steel material is under cyclic loading, its stress-strain relationship can be given by the Ramberg-Osgood equation, which is:

[0042]

[0043] where Δε is the strain amplitude, Δσ is the stress amplitude, E is the elastic modulus, K' is the cyclic strength coefficient, and n' is the cyclic strain hardening coefficient.

[0044] There is stress concentration at the root of the notch, and the stress is not directly equal to the far-field stress. Topper et al. [by modifying the Neuber rule, the relationship between the cyclic stress amplitude, strain amplitude at the root of the notch, and the far-field stress amplitude is as follows:

[0045]

[0046] where Δε is the notch strain amplitude, Δσ is the notch stress amplitude, E is the elastic modulus, K f is the fatigue notch coefficient, and S is the far-field stress amplitude. The fatigue notch coefficient K f is the ratio of the fatigue strength of the smooth specimen to the fatigue strength of the notched specimen.

[0047] 2) The IBESS method considers both short crack and long crack propagation processes when solving fatigue life. Given the measured load spectrum and weld size, the internal force spectrum is obtained based on structural mechanics knowledge, and then the stress intensity factor spectrum is obtained based on fracture mechanics knowledge. Combining the crack closure effect and the cyclic resistance curve of the short crack threshold, the plastic modified stress intensity factor amplitude at the fatigue crack is obtained. According to the "Cycle-by-Cycle" algorithm, the remaining life is obtained. The relevant formulas include:

[0048] Short crack crack propagation rate expression:

[0049]

[0050] where da / dN is the crack growth rate; ΔK is the total driving force applied at the crack; ΔK th is the crack growth threshold; C and m are constants determined by the material.

[0051] The effective threshold value ΔK th,eff for long cracks is used to estimate the threshold value ΔK dR for short crack growth in the Chapetti model expression: th

[0052] ΔK th = ΔK dR + (ΔK thR - ΔK dR )[1 - e -k(a-d) ] (6)

[0053] where ΔK thR is the long crack stress intensity factor amplitude threshold; a is the crack length, and k is a material constant given by:

[0054]

[0055] where d is the distance from the surface of the strongest microstructural barrier; ΔK thR is the long crack stress intensity factor amplitude threshold; ΔK dR is the microstructural threshold for crack growth.

[0056] 3) The long crack growth prediction algorithm is used to solve the fatigue life, which is based on the measured load spectrum and the weld size, the internal force spectrum is obtained according to the knowledge of structural mechanics, and then the stress intensity factor spectrum is obtained according to the knowledge of fracture mechanics, and the effective stress intensity factor spectrum is obtained based on the crack closure parameter; According to the effective stress intensity factor spectrum combined with the crack propagation curve, the remaining life is obtained according to the Cycle-by-Cycle algorithm. The related formula is as follows:

[0057] The SIF formula of the semi-elliptical surface crack plate can be expressed by the following hybrid formula:

[0058]

[0059] where K is the stress intensity factor (SIF); F E is the crack shape correction factor; F S is the crack front correction factor; F T is the finite thickness correction factor; F W is the finite width correction factor; F D is the stress gradient correction factor; σ is the stress at the crack tip; a is the crack depth. ​

[0060] Long crack crack propagation rate expression:

[0061]

[0062] In the formula: da / dN is the crack propagation rate; ΔK is the total driving force applied at the crack; ΔK th is the crack propagation threshold value; C and m are constants determined by the material.

[0063] Step two, according to the analysis results, based on Bayesian theory, the degradation model of railway steel bridge welded joint is established; Specifically including: setting high sensitivity parameters as state nodes in dynamic Bayesian network, setting low sensitivity parameters as fixed nodes, establishing railway steel bridge welded joint degradation model for crack initiation and propagation.

[0064] In order to further optimize the above technical scheme, the particle filtering inference algorithm is used to input the measured fatigue crack observation data to the dynamic performance degradation model, correct the prediction results, reduce the influence of uncertain factors, and obtain the real-time accurate value of the remaining fatigue life.

[0065] Specifically, the monotonic constitutive curve of the fatigue crack dangerous position of the welded joint is obtained based on the hardness method; Based on the monotonic constitutive curve, the initial cyclic constitutive curve is obtained through the Uniform materials method, the initial cyclic resistance curve is obtained through the method of Leonetti, and the initial crack propagation curve is obtained through the iLAPS method;

[0066] The obtained initial curve is used as the initial parameter value of the particle filtering algorithm, the posterior probability distribution of the uncertain node is represented by a set of random particles with weights according to the principle of particle filtering algorithm, the particle set weight is updated by comparing the input observation value with the estimated value, and the real posterior distribution is gradually approached, so as to reduce the influence of the uncertainty of the parameters.

[0067] In an exemplary embodiment, the detailed steps are as follows:

[0068] (1) Using instrumented indentation method, the hardness value HV of the fatigue cracking prone position of the railway steel bridge welded joint is measured, and according to the empirical formula of hardness HV, cyclic strength coefficient K', cyclic yield strength σ yc , cyclic strain hardening index', these data are obtained, and according to the steel cyclic constitutive curve formula, the initial cyclic constitutive curve is obtained by using the above obtained data, the steps refer to Figure 2 , and the initial cyclic constitutive curve will be used for local stress strain method, and the above parameters can be used as the initial value during analysis.

[0069] For the initial cyclic constitutive curve, in order to give the median curve of high cycle fatigue crack growth rate of most metals more accurately and reliably, the crack closure parameter U is introduced, and a new high cycle fatigue crack growth model is proposed. That is:

[0070]

[0071] In the formula: da / dN is the crack growth rate, U is the long crack closure parameter, c is the fatigue ductility index, E is the material elastic modulus, ε' f is the fatigue ductility coefficient, ΔK is the stress intensity factor amplitude, σ yc is the cyclic yield strength, n' is the cyclic strain hardening index, ΔK th is the stress intensity factor amplitude threshold.

[0072] The monotonic tensile curve of metal materials is different, so the corresponding cyclic stress-strain relationship should be determined by the following relationship, that is, the Ramberg-Osgood equation:

[0073]

[0074] In the formula: Δε is the strain amplitude, Δσ is the stress amplitude, E is the elastic modulus, K' is the cyclic strength coefficient, n' is the cyclic strain hardening index. Here, the definition method of monotonic yield strength can be used to calibrate the cyclic yield strength σ yc , and the expression is as follows:

[0075] σ yc = K'(0.002) n′ (13)

[0076] And the cyclic strain-life relationship can be described by the Coffin-Manson model, that is:

[0077]

[0078] In the formula: Δε is the plastic strain amplitude, Δσ is the stress amplitude, Ni is the fatigue crack initiation life, ε' f is the fatigue ductility coefficient, σ' f is the fatigue strength coefficient, c is the fatigue ductility index, b is the fatigue strength index.

[0079] Muralidharan et al. proposed a correlation model that can be compared with the general slope method in accuracy, that is:

[0080]

[0081] In the formula: ε f is the fracture strain E is the material elastic modulus, ε' f is the fatigue ductility coefficient, σ' fσ' = σ (1 + ε'c), ε' = εb f where %RA is defined as the reduction of area and is related to ε

[0082]

[0083] The homogeneous material method can be called a slope method. The difference is that the estimated values of the parameters used in the two types of materials are significantly different. The parameter ε' f for titanium alloys and aluminum alloys is considered to be a constant value of 0.35; the parameter ε' f for steel materials is related to i.e.

[0084]

[0085] where ε f is the fracture strain, the monotonic tensile strength σ f , E is the modulus of elasticity of the material, ε' f is the fatigue ductility coefficient, σ' f is the fatigue strength coefficient, c is the fatigue ductility exponent, and b is the fatigue strength exponent.

[0086] Further, in order to better use the four-point method model, the parameters are modified, i.e.

[0087] σ' b = σ f (1 + ε f ), ε' f

[0088]

[0089]

[0090] where ε f is the fracture strain, the monotonic tensile strength σ b , E is the modulus of elasticity of the material, ε' f is the fatigue ductility coefficient, σ' f is the fatigue strength coefficient, c is the fatigue ductility exponent, and b is the fatigue strength exponent.

[0091] At the same time, the functional relationship between HB and the tensile strength σ b is:

[0092]

[0093] where ε' f is the fatigue ductility coefficient, σ' f is the fatigue strength coefficient, and HB is the hardness.

[0094] In the median method, Meggiolaro et al. proposed that LCF parameter estimation can be performed for aluminum alloys and steels, i.e.:

[0095] σ' f = 1.5σ b , ε' f = 0.45

[0096] b = -0.09, c = -0.59 (20)

[0097] Park et al. proposed that only monotonic tensile strength σ b is used to estimate the relevant LCF model parameters, i.e.:

[0098]

[0099] The cyclic strain hardening index n' was found by Hu et al. using mathematical transformation to be related to 1-σ yc / K', i.e.:

[0100] n' = 0.34 (1-σ yc / K') - 0.05 (22)

[0101] For the cyclic strength coefficient K', some scholars proposed the following formula:

[0102] K' = 57K 0.545 - 1220 (23)

[0103] In the formula, K is the monotonic strength coefficient, and K' is the cyclic strength coefficient.

[0104] (2) The hardness value HV of the fatigue cracking-prone position of the welded joint of a railway steel bridge is measured by using the instrumented indentation method, the relationship between the hardness HV and the defect size The intrinsic stress intensity factor amplitude threshold ΔK th，eff and the maximum defect size are obtained, the cyclic resistance curve calibration parameters are obtained according to the empirical relationship between the cyclic resistance curve calibration parameters and the above-mentioned parameters, the stress ratio at the welded joint is obtained by combining the load spectrum and the knowledge of structural mechanics, the long crack stress intensity factor amplitude threshold in the existing technical specification, and the cyclic resistance curve calibration parameter k, the variation law of the stress intensity factor amplitude threshold ΔK th，eff with the crack size is obtained, that is, the initial cyclic resistance curve, the steps are referred to Figure 3 , and the initial cyclic constitutive curve will be used in the IBESS method, and the above-mentioned parameters can be used as initial values in analysis.

[0105] For the initial cyclic resistance curve, the stress intensity factor amplitude threshold and the crack velocity function were modified to simulate the lower stress intensity factor amplitude threshold and higher crack growth rates exhibited by short cracks relative to long cracks. The dependence on crack size was made explicit. The cumulative model for the external effect associated with long cracks is given by:

[0106] K th = ΔK th,eff + (ΔK th,lc - ΔK th,eff )[1 - exp(-k(a - d))] (24)

[0107] where ΔK th is the stress intensity factor amplitude threshold, ΔK th，eff is the intrinsic stress intensity factor amplitude threshold, ΔK th，lc is the long crack stress intensity factor amplitude threshold, k is a cyclic resistance curve calibration parameter, a is the crack depth, and d is the distance from the defect free surface.

[0108] where the parameter k controls the formation of the external effect and the length scale at which this occurs, 1 / k represents the crack length at which some initial value of the external effect is reduced to 1 / e of its initial value, where e is Euler's number. Chapetti estimates d, the distance from the defect free surface, using k as follows:

[0109]

[0110] This relationship was derived by introducing micro-defects of different sizes in polished specimens. For smaller defect sizes, the threshold stress range was found to be equal to the fatigue limit. This leads to the definition of a critical value of the parameter, i.e. represents the maximum defect size for which the defect does not propagate at stress ranges below the fatigue limit. When the threshold stress coincides with the fatigue limit, the existing formula cannot be used and a hardness is used to estimate it, as follows:

[0111]

[0112] where Hv is the material hardness and the maximum defect size

[0113] To estimate the size of the non-propagating crack, considering that the defect is semicircular, a can be converted to crack depth, assuming a crit is a reasonable estimate of d, a crit can be used instead of d in the formula as follows:

[0114]

[0115] (3) The hardness value HV of the fatigue-prone area of ​​the welded joint of the railway steel bridge was measured by the instrumented indentation method. Based on the relationship between hardness HV and cyclic strength coefficient K′ and cyclic yield strength σ, yc Based on the empirical relationship between the fatigue ductility index n′, the fatigue ductility coefficient c, and the cyclic strain hardening index n′, the above parameter values ​​were obtained. The initial crack propagation rate curve was then obtained using the iLAPS model, following the steps outlined below. Figure 4 The initial cyclic constitutive curves will be used in the IBESS method and the long crack propagation prediction algorithm, and the above parameters can be used as initial values ​​for the analysis.

[0116] For the initial crack propagation curve, based on the parameter σ yc Monotonic yield strength σ y With monotonic tensile strength σ b Scatter plots of ratios; Fatemi et al. found that the parameter σ yc Existence with The data points are clearly segmented. Based on the fitting of the data points, the segmented estimation formula for the cyclic plasticity parameters used in the new model iLAPS is obtained:

[0117]

[0118] The cyclic strain hardening exponent n′ is related to the piecewise distributed cyclic yield strength σ. yc It is related to the parameter K′. Through some data analysis, the following expression is obtained:

[0119]

[0120] Based on the experimental data processing results, it can be seen that the estimation error is too large for some metallic materials. After refitting, the following results are obtained:

[0121] K′=8K 0.719 (30)

[0122] Referring to the calculation formula of Hu et al., the strain hardening coefficient n is expressed through the monotonic yield strength σ. y and monotonic tensile strength σ b The estimate is:

[0123]

[0124] Cyclic yield strength σ yc It can be obtained from equation (28), and referring to equation (31), it can be seen that n′ can be obtained by using the cyclic yield strength σ. yc Based on the estimation of the cyclic strength coefficient K′, the following formula is obtained through numerical analysis:

[0125]

[0126] Fatigue ductility coefficient ε' f Generally, this can be determined by the nominal fracture strain and ε. f To characterize it, Genel believed that ε f With monotonic tensile strength σ b and monotonic yield strength σ y The ratios of ε' and ε' are related, therefore ε' f We should also consider The influence of the value on itself was determined by fitting experimental data, leading to the following conclusions:

[0127]

[0128] Some scholars have shown that the parameter c and ε f ,σ b The fatigue ductility index c is related to the elastic modulus E. From equation (18), the fatigue ductility index c can be derived from the fracture strain and ε. f Monotonic tensile strength σ b The ratio of the elastic modulus E to the two variables is used to represent this, and it is also necessary to consider... After correcting the final result and fitting the parameters, the following result was obtained:

[0129]

[0130] The crack closure parameter U for long cracks can be defined as a function of the stress ratio R. This is because, in the long crack stage, fatigue crack propagation is no longer affected by microstructural barriers and is mainly related to the applied load. The crack closure function (UNC) is obtained by transforming the Codrington closure function, as shown in the following formula:

[0131] U = 0.52 + 0.35R + 0.14R 2 (R≥-2) (35)

[0132] U = 0.45 + 0.37R + 0.20R 2 (R≥0) (36)

[0133] From the two equations above, it can be seen that different calculation formulas are used under different stress ratios R. Therefore, by fitting the two functions numerically, a new function (New-God) is obtained, yielding the following results:

[0134] U = 0.49 + 0.35R + 0.17R 2 (37)

[0135] Substitute the values ​​into the formula to calculate the crack propagation rate, and obtain the initial crack propagation curve through the calculation.

[0136] (4) Based on the initial value, the probability distribution of the state node in the dynamic Bayesian network is set, and the distribution particle group of the plurality of uncertain parameters is generated according to the random combination method.

[0137] (5) Based on the crack length information of the current welding node obtained by the deep learning system analysis, the likelihood of each particle is estimated.

[0138] (6) The particle is updated in weight, and the state node parameter value in the updated dynamic Bayesian network is output.

[0139] Example two: welding point residual life prediction

[0140] This embodiment is based on a global dynamic measurement method of fatigue cracks of multiple sensors and deep learning. A laboratory typical welding node fatigue crack propagation test is used to collect acoustic emission signals, strains and image information at different fatigue crack lengths. Finite element method is used to simulate acoustic emission signals, strains and image information at different fatigue crack lengths.

[0141] Based on the test and finite element simulation data, a deep learning neural network training sample set is established. Based on the program package Tensorflow, a deep learning neural network model is constructed and trained and verified, and a multi-source information fusion fatigue crack recognition deep learning model is established.

[0142] Using the deep learning model, the crack length is obtained according to the acoustic emission signal, strain and image information of the crack to be measured, and the crack length is input into the railway steel bridge welding node degradation model to predict the residual life of the welding point.

[0143] In a preferred embodiment, the deep learning model includes a first branch and a second branch; the first branch is used to extract acoustic emission signal and strain signal features, and the second branch is used to extract crack length image features.

[0144] Further, the first branch includes a first feature extraction module, and the second branch includes a second feature extraction module and a third feature extraction module connected in series; the output of the second feature module is interacted with the input data of the first branch, and then input into the first feature extraction module and the third feature extraction module respectively for feature extraction.

[0145] Then the output features of the first feature extraction module and the third feature extraction module are fused, and the crack size is obtained according to the fused features.

[0146] In an exemplary embodiment, the detailed steps are referred to Figure 5 , including:

[0147] (1)Carry out crack propagation tests of typical welded joints of railway steel bridges with various structural forms, plate thicknesses, and stress modes, accurately collect crack length images during crack propagation, and measure acoustic emission signals and strain signals in the plastic zone at the crack tip during crack propagation.

[0148] (2)Establish refined finite element models of typical welded joints of railway steel bridges with various structural forms, plate thicknesses, and stress modes, accurately simulate the initiation and propagation of fatigue cracks, and collect crack length images, acoustic emission signals, and strain signals in the plastic zone at the crack tip during crack propagation.

[0149] (3)Strain signals and acoustic emission signals are combined into one category as high-fidelity signals; crack length images are another category as low-fidelity signals.

[0150] (4)Establish a deep learning neural network model, including high-fidelity and low-fidelity neural networks. Both high-fidelity and low-fidelity neural networks are composed of two layers of fully connected neural networks. The low-fidelity neural network contains 128 neurons in each layer, and the high-fidelity neural network contains 8 neurons in each layer.

[0151] (5)Both low-fidelity data and high-fidelity data are used as input parameters for training the network. The low-fidelity data set uses the crack length results mentioned above, and the high-fidelity data set uses the finite element simulation results and experimental acoustic emission signals and strain signals results; the output y L of the low-fidelity neural network is used as the feedforward input of the high-fidelity neural network.

[0152] (6)After the low-fidelity neural network is trained, only the subsequent high-fidelity neural network needs to be trained to achieve fast output of the results. The output y H of the high-fidelity neural network has a residual connection with the output y L of the low-fidelity neural network, and its calculation formula is as follows

[0153] y H = y L + β (tanh α1· y H_L + tanh α2· y H_NL )

[0154] where β is the residual weight value, which is 0.1. y H_L and y H_NL are the output values of the linear high-fidelity neural network NNH_L and the nonlinear high-fidelity neural network NNH_NL, respectively, and α1 and α2 are two parameters that need to be trained.

[0155] (7) The acoustic emission signal, strain signal and image signal of the fatigue cracking position of the railway steel bridge welded joint are monitored by using images, acoustic emission and strain sensors, input into the trained multi-fidelity neural network, and the size of the measured crack, including the crack length, is obtained;

[0156] (8) The crack length is input into the railway steel bridge welded joint degradation model, and the cyclic constitutive parameters are corrected based on the crack length, the remaining fatigue life of the welded joint is predicted, and the dynamic evaluation of the health condition of the welded joint is realized.

[0157] In order to realize the fatigue damage evolution of the railway steel bridge welded joint under the action of random load in time series, a dynamic Bayesian network model facing fatigue crack initiation and propagation is established, the uncertainty factors existing in the fatigue crack initiation and propagation process are taken as the state nodes of the model, and the accurate evaluation of the fatigue crack size and the remaining fatigue life under random load is realized; in order to reduce the influence of the uncertainty factors on the evaluation accuracy in the fatigue crack initiation and propagation process, an uncertainty parameter inference algorithm based on instrumented indentation method and particle filter inference method is established, the key material data of the fatigue dangerous position of the welded joint collected by the instrumented indentation method are used to realize the step-by-step reduction of the influence of multiple uncertainty factors on the fatigue crack initiation and propagation process.

[0158] Further, based on the real fatigue crack data collected by multiple sensors, the uncertainty parameters in the fatigue crack initiation and propagation process are updated step by step, the dynamic following of the fatigue crack propagation and the real-time accurate evaluation of the remaining fatigue life are realized. At the same time, a high-precision identification method of fatigue crack measured data based on deep learning neural network is established, which greatly improves the identification efficiency and precision of the railway steel bridge fatigue crack.

[0159] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same and similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the related parts can be referred to the method part.

[0160] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to the embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A welding node residual fatigue life evaluation method based on multi-source heterogeneous information fusion, characterized in that, sensitivity analysis is performed on the uncertainty factors in the fatigue crack initiation and propagation process by taking the crack length as the judgment quantity; including: when the crack length is less than 0.1mm, the local stress-strain method is used to solve the fatigue life; when the crack length is between 0.1mm and 0.5mm, the IBESS method is used to solve the fatigue life; when the crack length is greater than 0.5mm, the long crack propagation prediction algorithm is used to solve the fatigue life; the local stress-strain method includes obtaining the nominal stress spectrum at the fatigue crack of the welding point according to the measured load and the weld size, combining the cyclic constitutive curve of the steel and the Neuber criterion to obtain the local stress spectrum, and based on the local stress spectrum and the cumulative damage criterion, the residual fatigue life of the welding point is obtained; the IBESS method includes determining the stress intensity factor spectrum according to the measured load and the weld size, combining the crack closure parameter and the cyclic resistance curve to obtain the plastic correction stress intensity factor amplitude at the fatigue crack; according to the plastic correction stress intensity factor amplitude and the crack propagation curve, the residual fatigue life of the welding point and the predicted crack length are obtained; the long crack propagation prediction algorithm includes determining the stress intensity factor spectrum according to the measured load and the weld size, and based on the crack closure parameter, the effective stress intensity factor spectrum is obtained; according to the effective stress intensity factor spectrum and the crack propagation curve, the residual fatigue life of the welding point and the predicted crack length are obtained; based on the analysis results, a railway steel bridge welding node degradation model is established based on the Bayesian theory; collecting crack length images, acoustic emission signals and strain signals; input to the deep learning neural network to obtain the crack length; input the crack length into the railway steel bridge welding node degradation model to obtain the residual fatigue life of the welding point.

2. The residual fatigue life evaluation method according to claim 1, characterized in that, obtaining the monotonic constitutive curve of the fatigue crack dangerous part of the welding node based on the hardness method; based on the monotonic constitutive curve, the initial cyclic constitutive curve is obtained by the Uniform materials method, the initial cyclic resistance curve is obtained by the Leonetti method, and the initial crack propagation curve is obtained by the iLAPS method; the obtained initial curve is used as the initial parameter value of the particle filtering algorithm, and the particle weight is iteratively updated by comparing the crack length observation value with the predicted crack length.

3. The method for evaluating the residual fatigue life according to claim 2, characterized by, high sensitivity parameters are set as state nodes in the dynamic Bayesian network, and low sensitivity parameters are set as fixed nodes to establish a railway steel bridge welding node degradation model for crack initiation and propagation.

4. The residual fatigue life assessment method according to claim 1, characterized by, The deep learning neural network includes a first branch and a second branch; the first branch is used to extract acoustic emission signal and strain signal features, and the second branch is used to extract crack length image features.

5. The method for evaluating the residual fatigue life according to claim 4, characterized by, The first branch includes a first feature extraction module, and the second branch includes a second feature extraction module and a third feature extraction module connected in series; the output of the second feature extraction module is interacted with the input data of the first branch and then input to the first feature extraction module and the third feature extraction module respectively for feature extraction. The output features of the first feature extraction module and the third feature extraction module are fused, and a crack size is obtained according to the fused features.

6. The method of assessing the residual fatigue life according to claim 5, characterized in that, The first feature extraction module, the second feature extraction module and the third feature extraction module are all composed of two layers of full connection neural networks.

7. The method of claim 5, wherein The feature fusion expression is: y H = y L + β(tanh α1 · y H_L + tanh α2 · y H_NL ) In the formula, y L is the output of the second feature extraction module, y H_L is the output of the first feature extraction module, y H_NL is the output of the third feature extraction module, and α1 and α2 are parameters that need to be trained, and β is a residual weight.

Images