Method for measuring the thickness of a multilayer metal

Abstract

The application discloses a kind of layered metal layered thickness measurement method, this method first carries out multi-electrode electrical measurement method experiment on the surface of layered metal material, obtains the apparent resistivity of measuring point under different electrode spacing;Then the relationship between the measurement depth of accurate measuring point in multi-electrode electrical measurement method and electrode spacing is calculated, further obtain the measurement depth of measuring point under different electrode spacing in experiment;A one-dimensional convolutional neural network model for layered metal material thickness and resistivity inversion is constructed, training samples are created to train the model, and finally the apparent resistivity of measuring point under different electrode spacing and the measurement depth data obtained are input into the model after training, and the layered thickness information and resistivity information of layered metal material are obtained by inversion.The application does not need to be pre-calibrated before use, and the thickness of each layer of layered metal can be obtained using the method, and the material type can be roughly judged by the resistivity obtained by inversion under the premise of unknown material type.

Description

Technical Field

[0001] This invention relates to the field of metal thickness measurement technology, and in particular to a method for measuring the thickness of multilayer metal layers. Background Technology

[0002] With the increasing demands for structural safety and reliability in high-end manufacturing, aerospace, energy and chemical industries, rail transportation, and pressure vessels, the importance of online service status monitoring and non-destructive testing technologies for materials and components is becoming increasingly prominent. In practical engineering, the use of multilayer metal materials is becoming more and more common. Their thickness parameters are not only related to load-bearing capacity and safety margin, but are also key indicators for assessing failure modes such as corrosion, wear, and peeling. Existing thickness measurement methods have common shortcomings. Most schemes require pre-calibration using samples or parameters before use. Although X-ray thickness measurement is non-contact and has good stability, the equipment is expensive and there are radiation safety issues. Its key parameters, such as the linear attenuation coefficient, often need to be obtained through calibration, and it is difficult to measure in-service conditions such as pressure vessels with sealed cavities. Eddy current thickness measurement has the advantages of being fast and highly sensitive, but the measurement range is extremely narrow and it is highly sensitive to material properties (such as magnetic permeability and electrical conductivity). It must be calibrated with a standard sample that is completely consistent with the material being measured. In ultrasonic methods, the sound velocity also needs to be pre-calibrated. For unknown materials or multi-layered composite systems, it is often necessary to back-calculate using test blocks that are completely identical to the test object in terms of "material / number of layers / layout sequence / thickness ratio", which further limits engineering applications. Summary of the Invention

[0003] In view of the problems and defects mentioned in the background art, the present invention provides a calibration-free method for measuring the thickness of multilayer metal layers.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A method for measuring the thickness of multilayer metal layers mainly includes the following steps:

[0006] S1. Conduct a multi-electrode electrical measurement experiment on the surface of a multilayer metal material, including setting four electrodes on the material surface, the four electrodes are symmetrically arranged, the two outer electrodes inject positive and negative currents into the material respectively, the two inner electrodes are used as voltage measurement points, the positions of the two inner electrodes are fixed, the distance between the two outer electrodes is continuously increased, and the apparent resistivity of the measurement points under different electrode distances is calculated.

[0007] S2. Obtain the relationship between the measurement depth of the precise measurement point and the electrode spacing in the multi-electrode electrical measurement method;

[0008] S3. Based on the relationship between the measurement depth of the measuring point and the electrode spacing obtained in step S2, obtain the measurement depth of the measuring point under different electrode spacings in the experiment;

[0009] S4. Construct a one-dimensional convolutional neural network model for the inversion of the thickness and resistivity of multilayer metals, create training samples to train the one-dimensional convolutional neural network model, input the apparent resistivity and measurement depth data of the measurement points under different electrode spacings obtained in steps S1 and S3 into the trained one-dimensional convolutional neural network model, and invert the layer thickness information and resistivity information of the multilayer metal material.

[0010] Preferably, in step S1, the formula for calculating the apparent resistivity of the measuring points under different electrode spacings is:

[0011] ;

[0012] in, The apparent resistivity of the measuring point. For device coefficients, , The potential difference between electrodes M and N is To inject current; in the experiment, the four electrodes A, M, N, and B are arranged from left to right. The distance between electrodes A and M is... The distance between electrodes B and M is... The distance between electrodes A and N is... The distance between electrodes B and N is denoted as .

[0013] Preferably, in step S2, the relationship between the measurement depth of the measuring point and the electrode spacing is as follows:

[0014] ;

[0015] in, The measured depth of the measuring point.

[0016] Preferably, in step S4, the method for creating the training samples includes: constructing a large-scale dataset of random resistivity and layer thickness combinations based on the five-point difference method. Each set of data in the dataset includes an apparent file and a True file. The apparent file contains two columns of data: the first column is the measured depth of the measuring point, and the second column is the apparent resistivity of the measuring point. The True file contains two columns of data: the first column is the layer thickness information of the multilayer metal material, and the second column is the layer resistivity information of the multilayer metal material.

[0017] Preferably, in step S4, the inversion process of the one-dimensional convolutional neural network model for layer thickness information and resistivity information includes:

[0018] S41. Input data is standardized, features are extracted through multi-layer convolution, and output through a fully connected layer. The relative thickness ratio of each layer is output in the SoftMax layer:

[0019] ;

[0020] in, For the first The relative thickness ratio of the layers This represents the total number of floors. The first output of the fully connected layer logit of layer thickness The first output of the fully connected layer The logit of layer thickness is a dimensionless metric output by the model after the preceding convolution, pooling, and flattening operations.

[0021] S42. Thickness calculation and boundary accumulation:

[0022] ;

[0023] in, For the first The thickness of the layer, For the total thickness of the material, For the first Depth of each layer For the first The depth of each layer;

[0024] S43. Construct depth-dependent mask functions:

[0025] ;

[0026] in, This represents the activation function. , For depth The mask function value at that location, For depth, For interface smoothing scale;

[0027] S44, Constructed mask function for depth Predicting the resistivity at a given location:

[0028] ;

[0029] in, For depth Predicted resistivity at [location] The depth after sharpening The mask function value at that location, This is the sharpening factor. The first output of the model The resistivity of the layer;

[0030] S45. Perform TV regularization:

[0031] ;

[0032] ;

[0033] in, For total variation loss, This represents the number of depth sampling points. and The first The and the first The weight of each depth point and The first The and the first Predicted resistivity values ​​at the depth of each depth point. This represents the depth value at the nth depth point. Indicates depth as The mask function value at that location.

[0034] The principle of this invention: Multilayered metals have a multi-layered structure. The resistivity of the same layer remains constant at different depths, except for a step change at the boundary. Due to this resistivity step between layers, the multi-electrode electrical measurement method changes the electrode arrangement to make the current field respond to different depth ranges within the material, thus obtaining an apparent resistivity curve that varies with depth. The trend changes and inflection point characteristics in the apparent resistivity curve reflect the layered characteristics of the material's internal electrical structure. Further analysis using a one-dimensional convolutional neural network model can determine the resistivity parameters and thickness information of each layer.

[0035] Compared with the prior art, the present invention has the following advantages:

[0036] This invention creatively proposes a calibration-free method for measuring the layer thickness of multilayer metals. A novel and precise relationship between electrode spacing and measurement depth is established through four-electrode sensitivity. A convolutional neural network framework suitable for multilayer metal layer thickness measurement is constructed. Potential is measured using a multi-electrode electrical measurement method, and the layer thickness and resistivity of the upper and lower layers of the multilayer metal are obtained through inversion. The inversion results show good agreement with actual values. This method can not only determine the thickness of each layer of a multilayer metal, but also roughly determine the material type using the inverted resistivity when the material type is unknown. Attached Figure Description

[0037] Figure 1 (a) is the equivalent model of the multilayered metal material of the present invention. Figure 1 (b) is a schematic diagram of the multi-electrode electrical sounding method for depth measurement;

[0038] Figure 2 This is a schematic diagram showing the deflection of current at the interface layer.

[0039] Figure 3 (a) Schematic diagram of three-dimensional electrode arrangement in multi-electrode electrical measurement method. Figure 3 (b) is a schematic diagram of the electrode planar arrangement;

[0040] Figure 4 This is a schematic diagram of the potential change caused by local resistivity perturbation under four electrodes.

[0041] Figure 5 This invention provides a one-dimensional convolutional neural network model framework for inverting the thickness and resistivity of multilayer metals.

[0042] Figure 6 The inversion results are obtained by using the one-dimensional convolutional neural network model constructed in this invention to invert models with different resistivity and different layer thicknesses. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0044] 1. Experiment on multi-electrode electrical measurement method

[0045] Multilayer metals have a multilayer structure (its equivalent model is as follows) Figure 1 (a) shows the figure. For clad metals, four electrodes are placed on the material surface, arranged symmetrically. The outer pair of electrodes A and B are injected with positive and negative currents respectively, and the inner pair of electrodes M and N serve as voltage measurement points (e.g., Figure 1 (b) As shown. When an applied current is injected into the interior of the cladding metal from the surface, it flows according to the conductivity distribution of each layer. However, at the interface layer, due to the difference in conductivity between the upper and lower layers, the current will be deflected when it flows through the interface layer (e.g. Figure 2 As shown in the figure, this leads to a difference in the surface potential of the composite material. According to the principle of multi-electrode electrical measurement, when the voltage measuring electrodes M and N are fixed, and current is injected into electrodes A and B to expand the measurement, the detection position will gradually penetrate into the interior of the composite metal along the midpoint of the line connecting M and N. Thus, the apparent resistivity at different depths of the composite metal material can be measured. The formula for calculating the apparent resistivity is shown in Equation 1:

[0046]

[0047] in, The apparent resistivity of the measuring point. For device coefficients, , The potential difference between electrodes M and N is To inject current, The distance between electrodes A and M is... The distance between electrodes B and M is... The distance between electrodes A and N is... The distance between electrodes B and N is denoted as .

[0048] However, most existing multi-electrode electrical measurement methods use empirical formulas for depth calculation, as shown in Equation 2:

[0049]

[0050] in, To characterize the feature lengths of the device (such as the distance between power supply electrodes) ), This is an empirical coefficient (typically ranging from 0.2 to 0.8). Such formulas are insufficient to provide accurate measurement depths; therefore, a precise method for calculating measurement depths, related to electrode arrangement, needs to be established.

[0051] 2. Precise depth calculation method

[0052] 1) Sensitivity calculation of multi-electrode electrical measurement method

[0053] A schematic diagram of the three-dimensional and planar arrangement of electrodes in a multi-electrode electrical measurement method is shown below. Figure 3 As shown, according to Poisson's equation for the electric field, the electric potential... satisfy:

[0054]

[0055] in, This represents the position vector of the potential observation point. Represents the position vector of the current source. Represents the Laplace operator. It represents the distribution of electrical conductivity inside an object. It represents the distribution of electric potential inside an object. The current intensity of the current source. The Dirac function (representing current only when the position vector is...) (Injection at the location)

[0056] When the resistivity is slightly disturbed At this point, the conductivity changes:

[0057]

[0058] in, This represents the perturbation of the resistivity inside the object. This represents the internal conductivity distribution of an object after resistivity perturbation. This represents the resistivity distribution inside the object before the disturbance. The resistivity distribution inside the object after the resistivity is disturbed;

[0059] Expanding equation 4 using Taylor series yields:

[0060]

[0061] At this moment, the electric potential changes:

[0062]

[0063]

[0064] in, This represents the internal electric potential distribution of an object after its resistivity is disturbed. This refers to the change in internal electric potential of an object caused by a disturbance in resistivity.

[0065] Expanding the left side of equation 7, we get:

[0066] (8)

[0067] in, This represents the change in conductivity caused by resistivity perturbation;

[0068] Based on this, subtracting equation 3 from equation 7 yields:

[0069]

[0070] Ignoring second-order minor quantities, we can obtain:

[0071]

[0072] Type 10, side-by-side escort field We can obtain:

[0073]

[0074] in, The accompanying field represents the potential distribution inside the object when a unit current is injected into the M and N electrodes. It is a virtual field (not the actual injected current).

[0075] Both sides simultaneously affect the entire area By performing volume integration, we can obtain:

[0076]

[0077] in, It is a volume infinitesimal element;

[0078] By divergence theorem ,make We can obtain:

[0079] (13)

[0080] (14)

[0081] in, For the region The boundary surface, For the area of ​​a micro-element, It is the surface normal vector;

[0082] From the boundary insulation condition, we can obtain:

[0083]

[0084]

[0085]

[0086] Accompanying field Satisfies the Poisson equation:

[0087]

[0088] in, Let M be the position vector. The position vector of point N;

[0089] Type 18, riding on both sides And for the region By performing a volume integral, we can obtain:

[0090] (19)

[0091] in, The potential change at point M after the resistivity perturbation occurs. The potential change at point N after the resistivity perturbation occurs;

[0092] From the divergence theorem and the insulation boundary condition equations, it becomes:

[0093]

[0094]

[0095] Where U represents the potential difference between points M and N, This represents the change in the potential difference between points M and N caused by resistivity perturbation.

[0096] From equation 17, we can obtain:

[0097]

[0098] Depend on We can obtain:

[0099]

[0100] Thus, the change in potential difference between points M and N caused by resistivity perturbation is obtained.

[0101] 2) Calculation of the relationship between electrode spacing and depth sounding

[0102] The potential change caused by local resistivity disturbance under four electrodes is as follows: Figure 4 As shown, the current electrode is at C The electric potential generated at the point is:

[0103]

[0104] in, Let A be the x-coordinate of point A. Let B be the x-coordinate. Let C be the distance between points A and C. The distance between point B and point C;

[0105] Voltage electrode at C The accompanying electric field potential generated at the point is:

[0106]

[0107] in, Let M be the x-coordinate of point M. Let N be the x-coordinate of point N. Let M be the distance between points M and C. The distance between point N and point C;

[0108] Then we have:

[0109]

[0110] in:

[0111] ;

[0112] We can obtain:

[0113]

[0114]

[0115]

[0116]

[0117] Calculate sensitivity in the z-axis direction by injecting a unit current. for:

[0118]

[0119] Integrating equation 32, we get:

[0120] (33)

[0121] in, The potential response generated at point M by the power supply electrode A is given. This represents the potential response generated at point N by the power supply electrode A. This represents the potential response generated at point M by the power supply electrode B. The potential response generated at point N by the power supply electrode B;

[0122] Weighted average along the depth direction using the sensitivity function as the weight:

[0123] (34)

[0124] The newly established precise relationship between electrode spacing and depth sounding was finally obtained:

[0125]

[0126] Based on the above formula, the measurement depth of the measuring point under different electrode spacings in the experiment can be obtained.

[0127] 3. Establish a convolutional neural network inversion framework

[0128] In electrical geophysical exploration, traditional one-dimensional true resistivity inversion methods typically employ iterative optimization algorithms such as the least squares method. These methods are computationally expensive, sensitive to the initial model, and have difficulty guaranteeing iterative convergence. Currently, with the development of deep learning algorithms, convolutional neural networks (CNNs) are widely used. Based on this, this invention constructs a one-dimensional convolutional neural network framework (e.g., ...) for the inversion of thickness and resistivity in multilayer metals. Figure 5 As shown, the model contains four convolutional layers of different dimensions. To improve the generalization ability and robustness of the constructed CNN method, a large-scale dataset of random resistivity and layer thickness combinations was constructed based on the five-point difference method to train the model. Each set of data in the dataset is divided into two files, where the apparent file contains two columns: the first column is the depth measurement of the measurement point, calculated using the formula... The second column shows the apparent resistivity of the measuring point, calculated using the following formula: The True file contains two columns: the first column shows the layer thickness information of the multilayer metal material; the second column shows the layer resistivity information of the multilayer metal material. Meanwhile, to improve the model's generalization ability to different depth sampling data, N depth grid numbers are uniformly set based on the original depth sampling data to obtain data with a unified dimension, thus facilitating batch training. To ensure that the output results conform to real physical laws, TV regularization and mask weights are added, strongly penalizing intra-layer variations while minimizing penalties for inter-layer steps.

[0129] Specifically, the inversion process of the constructed one-dimensional convolutional neural network model for layer thickness and resistivity information is as follows:

[0130] 1) Input data is standardized, features are extracted through multiple convolutional layers, and output through fully connected layers. The relative thickness ratio of each layer is output at the SoftMax layer:

[0131] ;

[0132] in, For the first The relative thickness ratio of the layers This represents the total number of floors. The first output of the fully connected layer logit of layer thickness The first output of the fully connected layer The logit of layer thickness is a dimensionless metric that is output by the model after a series of operations such as convolution, pooling, and flattening.

[0133] 2) Thickness calculation and boundary accumulation:

[0134] ;

[0135] in, For the first The thickness of the layer, For the total thickness of the material, For the first Depth of each layer For the first The depth of each layer.

[0136] 3) Construct depth-dependent mask functions:

[0137] ;

[0138] in, This represents the activation function. , For depth The mask function value at that location, For depth, For interface smoothing scale.

[0139] 4) Depth based on constructed mask function Predicting the resistivity at a given location:

[0140] ;

[0141] in, For depth Predicted resistivity at [location] The depth after sharpening The mask function value at that location, This is the sharpening factor. The first output of the model The resistivity of the layer.

[0142] 5) Perform TV regularization:

[0143] ;

[0144] ;

[0145] in, For total variation loss, This represents the number of depth sampling points. and The first The and the first The weight of each depth point and The first The and the first Predicted resistivity values ​​at the depth of each depth point. This represents the depth value at the nth depth point. Indicates depth as The mask function value at that location;

[0146] At this point, the total loss of the model during training is:

[0147] ;

[0148] in, The error between the model's predicted value and the actual value. The weight of the TV regularization term (taken as 0.01 here).

[0149] 4. Input data into the model for measurement.

[0150] The apparent resistivity and measurement depth data of the measurement points under different electrode spacings are input into the trained one-dimensional convolutional neural network model to invert the layer thickness and resistivity information of the multilayer metal material.

[0151] 5. Comparative Analysis of Measurement Results

[0152] Figure 6 This paper demonstrates the inversion results of four multilayer metal models with different resistivities and layer thicknesses, constructed using the one-dimensional convolutional neural network model of this invention. Each of the four multilayer metal models consists of two layers. Since the resistivity of the same material layer is constant throughout, a sudden change in resistivity occurs at the boundary between the two layers. Therefore, by observing at what depth (thickness) the resistivity abruptly changes in the inversion result curve and comparing it with the actual distribution curve, the accuracy of the layer thickness inversion can be determined. Similarly, by observing whether the resistivity of the two layers in the inversion result curve matches the resistivity in the actual distribution curve, the accuracy of the resistivity inversion can be determined. The results show that the error between the inverted resistivity and the actual resistivity does not exceed 10%, and the error between the inverted layer thickness and the actual layer thickness does not exceed 5%, indicating good agreement between the inversion results and reality.

[0153] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for measuring the thickness of a multilayer metal layer, characterized in that, Includes the following steps: S1. Conduct a multi-electrode electrical measurement experiment on the surface of a multilayer metal material, including setting four electrodes on the material surface, injecting positive and negative currents into the material into the two outer electrodes respectively, and using the two inner electrodes as voltage measurement points. The positions of the two inner electrodes are fixed, and the distance between the two outer electrodes is continuously increased to calculate the apparent resistivity of the measurement points under different electrode distances. S2. Obtain the relationship between the measurement depth of the precise measurement point and the electrode spacing in the multi-electrode electrical measurement method; S3. Based on the relationship between the measurement depth of the measuring point and the electrode spacing obtained in step S2, obtain the measurement depth of the measuring point under different electrode spacings in the experiment; S4. Construct a one-dimensional convolutional neural network model for the inversion of the thickness and resistivity of multilayer metals, create training samples to train the one-dimensional convolutional neural network model, input the apparent resistivity and measurement depth data of the measurement points under different electrode spacings obtained in steps S1 and S3 into the trained one-dimensional convolutional neural network model, and invert the layer thickness information and resistivity information of the multilayer metal material. The inversion process of the one-dimensional convolutional neural network model for layer thickness and resistivity information includes: S41. Input data is standardized, features are extracted through multi-layer convolution, and output through a fully connected layer. The relative thickness ratio of each layer is output in the SoftMax layer: ； in, For the first The relative thickness ratio of the layers This represents the total number of floors. The first output of the fully connected layer logit of layer thickness The first output of the fully connected layer The logit of layer thickness; S42. Thickness calculation and cumulative boundary: ； in, For the first The thickness of the layer, For the total thickness of the material, For the first Depth of each layer For the first The depth of each layer; S43. Construct depth-dependent mask functions: ； in, For mask function, For depth, For interface smoothing scale; S44, Constructed mask function for depth Predicting the resistivity at a given location: ； in, For depth Predicted resistivity at [location] This is the sharpened mask function. This is the sharpening factor. The first output of the model The resistivity of the layer; S45. Perform TV regularization: ； ； in, For total variation loss, This represents the number of depth sampling points. and The first The and the first The weight of each depth point and The first The and the first The predicted resistivity at the depth of each depth point.

2. The method for measuring the thickness of multilayer metal layers according to claim 1, characterized in that, In step S1, the formula for calculating the apparent resistivity of the measuring points under different electrode spacings is as follows: ； in, The apparent resistivity of the measuring point. For device coefficients, , The potential difference between electrodes M and N is To inject current; in the experiment, the four electrodes A, M, N, and B are arranged from left to right. The distance between electrodes A and M is... The distance between electrodes B and M is... The distance between electrodes A and N is... The distance between electrodes B and N is denoted as .

3. The method for measuring the thickness of multilayer metal layers according to claim 2, characterized in that, In step S2, the relationship between the measurement depth of the measuring point and the electrode spacing is as follows: ； in, The measured depth of the measuring point.

4. The method for measuring the thickness of multilayer metal layers according to claim 1, characterized in that, In step S4, the method for creating the training samples includes: constructing a large-scale dataset of random resistivity and layer thickness combinations based on the five-point difference method. Each set of data in the dataset includes an apparent file and a True file. The apparent file contains two columns of data: the first column is the measured depth of the measuring point, and the second column is the apparent resistivity of the measuring point. The True file contains two columns of data: the first column is the layer thickness information of the multilayer metal material, and the second column is the layer resistivity information of the multilayer metal material.

Images