Construction method of fracture width inversion model based on ratio characteristic parameters
Through the fracture width inversion model based on ratio characteristic parameters, the nonlinear mapping relationship between the fracture width and the detection height is constructed using a deep network, which solves the problem of unavailability of calibration data when the detection conditions change in the prior art, and realizes the difference The quantization of wire breakage under detection conditions reduces the workload and improves the flexibility and accuracy of detection.
Patent Information
- Application Number
- CN202210868029.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-07-21
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2042-07-21
AI Technical Summary
When the existing magnetic leakage detection method changes, the calibration data is unavailable, resulting in a large workload during the detection process and cannot be generalized under different cable specifications and probe specifications.
Using the fracture width inversion model based on ratio characteristic parameters, the model is constructed through a deep network, and the ratio characteristic parameter R is calculated using the characteristic magnetic field signal to establish a nonlinear mapping relationship between the fracture width and the detection height to achieve accurate prediction of the fracture width and avoid The problem of recalibration due to changes in the detection environment.
The wire break quantization without recalibration is achieved when the detection conditions change, which reduces the calibration workload during the detection process, ensures the availability of calibration data, and improves the flexibility and accuracy of detection.
Smart Images

Figure CN115236175B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of nondestructive testing, and more specifically, relates to a method for constructing a fracture width inversion model based on ratio characteristic parameters. Background Art
[0002] As the main load-bearing component of a cable-stayed bridge, the cable is directly related to the safe operation of the bridge. In the built bridges, due to the long-term combined effects of huge dead loads, live loads, alternating stresses and vibrations on the cables, fatigue wire breakage is very likely to occur inside the cables, which in turn has a great impact on the safe operation of the bridge. It is very necessary to timely detect the health of the steel wire inside the cables. As a mature non-destructive testing method, magnetic leakage testing has been used for non-destructive testing of broken wire damage in in-service cables due to its simple detection principle, high sensitivity to ferromagnetic material defects, strong online detection capabilities, and low detection costs.
[0003] The core part of the magnetic flux leakage detection system is the detection probe, which consists of two parts: a magnetizer and a magnetic sensitive element. The basic principle is to use the magnetizer to saturate the local wire bundle inside the cable. When there are no defects such as broken wires in the internal wire bundle, the magnetic field is evenly distributed inside the wire bundle and the magnetic flux lines are parallel to the surface of the test piece. There is almost no magnetic flux line overflowing from the surface of the wire bundle. When there are defects such as broken wires in the internal wire bundle, part of the magnetic field passing through the defective area will be refracted into the air to form a leakage magnetic field, and then a magnetic anomaly signal will appear. The magnetic sensitive element can detect the magnetic anomaly signal under a certain lift-off to determine and quantify the broken wire.
[0004] The interior of the cable body is composed of multiple parallel galvanized high-carbon steel wires tied by a tape, and the outer layer is a polyethylene (PE) sheath. The object of the leakage magnetic detection is the parallel steel wire bundle inside. In the existing research on leakage magnetic detection methods, the method of calibration first and then quantification is often used to quantify the broken wires of the cable. However, the calibration data must strictly correspond to the cable specifications, probe specifications and detection lift-off during calibration. The calibration signals under different cable specifications, probe specifications or detection lift-off cannot be universal when the broken wire information is inverted. At the same time, there are too many cable specifications. The workload of calibrating the broken wire signals of all specifications of cables before detection is very large. Moreover, when the detection conditions change, such as replacing the magnetizer, adjusting the magnetizer lift-off distance or the lift-off distance from the magnetic sensitive element in the probe to the surface of the PE sheath, the previously calibrated data are no longer universal, and recalibration will bring a huge workload. Therefore, it is necessary to study a new method for quantifying broken wires. Summary of the invention
[0005] In view of the defects and improvement needs of the prior art, the present invention proposes a method for constructing a fracture width inversion model based on ratio characteristic parameters, aiming to solve the problem that pre-calibrated data is unavailable when the detection conditions change.
[0006] To achieve the above object, the present invention provides a method for constructing a fracture width inversion model based on ratio characteristic parameters, which comprises the following steps:
[0007] S1. Determine the diameter D of the broken wire to be detected and the interval d between the gradient detection points, respectively change the fracture width w and the detection height h of the single broken wire specimen, obtain the magnetic leakage signals at different fracture widths w and the corresponding multiple detection heights set at equal intervals d based on the detection height h, and establish a rectangular coordinate system with the symmetry axis of the leakage magnetic signal as the y-axis and the axis of the broken wire to be detected as the x-axis;
[0008] S2, obtaining a denoised signal through the acquired leakage magnetic signal, and extracting the magnetic field value of the denoised signal on the y-axis of the rectangular coordinate system as a characteristic magnetic field signal B corresponding to the fracture width and the detection height;
[0009] S3, using the acquired characteristic magnetic field signal B to calculate the corresponding ratio characteristic parameter R, to form sample data, and randomly divide the sample data into training sample data and test sample data;
[0010] S4. Establish an initial network model based on the deep network, train the initial network model with training sample data to obtain a trained network model, thereby completing the establishment of the inversion model.
[0011] In one embodiment, the ratio characteristic parameter R=(R h ,R h+d ,R h+2d , .., R h+kd ), R contains at least 2 values,
[0012] and
[0013] Where k is an integer starting from 0.
[0014] In one embodiment, the characteristic magnetic field signal B=(B h ,B h+d ,B h+2d ,..,B h+nd ), B contains at least 3 values,
[0015] and
[0016] In the formula, B h+nd represents the magnetic field value of the denoised signal on the y-axis of the rectangular coordinate system, Q represents the point magnetic charge of equal value distributed on both sides of the fracture, μ0 represents the vacuum magnetic permeability, g represents the sum of the wire diameter D and the fracture width w, and n is an integer starting from 0.
[0017] In one of the embodiments, step S5 is further included: substituting the test sample data into the trained network model to verify its effectiveness.
[0018] In one of the embodiments, it is characterized in that the denoised signal is a signal with a background field subtracted.
[0019] In one embodiment, in step S1, when the fracture width w changes, the detection height h is kept unchanged; when the detection height h changes, the fracture width w is kept unchanged.
[0020] In one embodiment, the input of the initial network model is the R value, and the output is the detection height h and the fracture width w; the fracture width w in the data set is set to j cases, namely w1~w j , j ≥ 2, the detection height h is set in i cases in the training set, i.e. h1~h i , i ≥ 2, and the deep neural network is trained according to the R values obtained under different fracture width w and detection height h conditions to establish the mapping relationship between the ratio characteristic parameter R and the two parameters.
[0021] In one embodiment, the initial network model is a two-layer feedforward network with Sigmoid hidden neurons and linear output neurons.
[0022] In one embodiment, the initial network model is trained using the Bayesian regularization method, and the loss function in the training process is the mean square error (MSE), which is expressed as:
[0023]
[0024] Among them, y k Indicates the true value of the fracture width w (or detection height h) output in the kth group in each round of training; Represents y k In each corresponding round of training, the output value of the fracture width w (or detection height h) output by the network model in the kth group is n, and n is the total number of groups.
[0025] The above technical solution conceived by the present invention has the following advantages compared with the prior art:
[0026] 1. The present invention provides a fracture width inversion method based on the characteristic parameter of the ratio of the leakage magnetic field of the broken wire. Through a single calibration of the characteristic parameter R of a single type of cable, the characteristic parameter R presents a complex nonlinear change with the parameters w and h. A fracture width inversion model based on the parameter R is constructed using a deep network. After the training is completed, the leakage magnetic characteristic parameter R of the test piece is input to achieve an accurate prediction of the fracture width w. It is achieved that there is no need to calibrate the inversion network again when changing the lift-off height, replacing the magnetizer, changing the specification of the detection cable, etc., thereby ensuring the availability of the data before calibration and greatly reducing the calibration workload during the detection process.
[0027] 2. By constructing the ratio characteristic parameter R, its value is nonlinearly related to the broken wire diameter D, the detection height h and the broken wire width w, and the broken wire diameter D is a constant value. Therefore, the ratio characteristic parameter R is a value that changes with the change of the detection height h and the short wire width w. The change of the environment does not affect the determination of the R value to a certain extent, thereby avoiding the problem of recalibration due to changes in the detection environment.
[0028] 3. The characteristic magnetic field signal B is the magnetic field value on the y-axis of the rectangular coordinate system. The leakage magnetic signal is obtained by setting the detection height at equal intervals d based on the detection height h. The ratio between the leakage magnetic signals with a height difference of d is R. This ratio method eliminates some parameters, so that the ratio R is not affected by other factors. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] Figure 1 It is a flow chart of a method for constructing a fracture width inversion model based on ratio characteristic parameters according to an embodiment of the present invention;
[0030] Figure 2 It is a magnetic dipole model of a broken wire in a cable according to an embodiment of the present invention;
[0031] Figure 3 The leakage magnetic field distribution of the broken wire of the cable is calculated based on the magnetic dipole model in the embodiment of the present invention;
[0032] Figure 4 It is a schematic diagram of the arrangement of the cable magnetic leakage detection element in an embodiment of the present invention;
[0033] Figure 5 It is a structural schematic diagram of performing magnetic flux leakage detection on a cable in an embodiment of the present invention;
[0034] Figure 6 Graph showing the variation of R value in the embodiments of the present invention.
[0035] Throughout the drawings, the same reference numerals are used to denote the same elements or structures, wherein:
[0036] 1. Parallel steel wire bundle; 2. PE sheath; 3. Permanent magnet; 4. Magnetic yoke; 5. Magnetic sensitive element; 6. Broken wire to be tested. DETAILED DESCRIPTION
[0037] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0038] The principle of the present invention is to use a deep neural network to construct a method for constructing a fracture width inversion model based on ratio characteristic parameters. The characteristic parameter values under different detection heights and fracture widths are obtained as training sets by simulating actual working conditions through finite element simulation. The deep neural network model for width inversion is trained, and the characteristic parameters are used to achieve accurate prediction of the fracture width. The effectiveness and accuracy of the network model for fracture width inversion are verified using additional test data. Since the characteristic parameters are only related to the fracture width and the height of the detection point, there is no need to recalibrate when the specifications of the detection cable or other detection conditions change. In the prior art, due to the changes in various detection condition parameters during the detection process, the previously calibrated data is unavailable, which brings about the heavy work of recalibrating the data.
[0039] like Figure 1 As shown, the present invention provides a method for constructing a fracture width inversion model based on ratio characteristic parameters, which comprises the following steps:
[0040] S1. Determine the diameter D of the broken wire to be detected and the interval d between the gradient detection points, respectively change the fracture width w and the detection height h of the single broken wire specimen, obtain the magnetic leakage signals at different fracture widths w and the corresponding multiple detection heights set at equal intervals d based on the detection height h, and establish a rectangular coordinate system with the symmetry axis of the leakage magnetic signal as the y-axis and the axis of the broken wire to be detected as the x-axis;
[0041] S2, obtaining a denoised signal through the acquired leakage magnetic signal, and extracting the magnetic field value of the denoised signal on the y-axis of the rectangular coordinate system as a characteristic magnetic field signal B corresponding to the fracture width and the detection height;
[0042] S3, using the acquired characteristic magnetic field signal B to calculate the corresponding ratio characteristic parameter R, to form sample data, and randomly divide the sample data into training sample data and test sample data;
[0043] S4. Establish an initial network model based on the deep network, train the initial network model with training sample data to obtain a trained network model, thereby completing the establishment of the inversion model.
[0044] In fact, in engineering practice, a final step can be added: S5, substituting the test sample data into the trained network model to verify its effectiveness.
[0045] Some calculation methods in the above method will be explained in the following content.
[0046] First, the magnetic charge model of the broken wire magnetic field is used to analyze the leakage magnetic field distribution of the broken wire, such as Figure 2 and Figure 3 As shown in the figure, the model assumes that there are equal point magnetic charges -Q and +Q at both ends of the fracture, respectively located at the center of the inscribed circle of the fracture, and the coordinates of the two points are P1(-g / 2,0) and P2(g / 2,0). According to the magnetic charge theory and superposition theorem, the magnitude of the magnetic field at any position P(x,h) under the action of two magnetic charges is expressed as the vector sum of the two magnetic charges acting alone. That is, the horizontal magnetic induction intensity generated by the point magnetic charges -Q and +Q at point P(x,h) is B x :
[0047]
[0048] In formula (1), B 1x and B 2x are the components of the two magnetic charges in the horizontal direction, B x It represents the total magnetic flux density in the horizontal direction under the action of two magnetic charges. The parameters h and x represent the vertical height and horizontal relative position of the measuring point from the broken wire point magnetic charge, respectively. The parameter g represents the distance between the point magnetic charges -Q and +Q. For the specific cable to be measured, the wire diameter D is a fixed value, and only and Two specifications. The magnetic charge parameter Q is mainly related to the magnetization state of the steel wire. g is the sum of the steel wire diameter D and the fracture width w, that is: g = w + D.
[0049] Extract the magnetic flux density of each gradient detection point on the ordinate axis (x = 0) in formula (1) whose vertical distances from the electromagnetic charge are h, h + d, h + 2d, ..., h + nd, that is, the magnetic flux density at (x = 0, y = h + nd), and use it as the magnetic field characteristic quantity B = (B h ,B h+d ,B h+2d ,..,B h+nd ), and B includes at least 3 values, among which,
[0050]
[0051] The adjacent gradient magnetic field characteristic quantities in formula (2) are ratio-processed to obtain a new characteristic parameter R, which is shown in the following formula:
[0052]
[0053] In the formula, R h , R h+d , .., R h+(n-1)d They represent the R values at the vertical heights of h, h+d, .., h+(n-1)d from the fracture point. Substituting several B values in formula (2) into formula (3), we can get the new characteristic parameter R=(R h ,R h+d ,..,R h+(n-1)d ), and R includes at least 2 values. From the above (n-1) analytical expressions of R, it can be seen that the characteristic parameter R is only related to the parameters D, d, w and h. In actual detection, the wire diameter D and the gradient detection point spacing d are both fixed values, and the characteristic parameter R is only a nonlinear function of the parameters w and h.
[0054] In addition, the present invention uses a deep neural network method to determine the nonlinear relationship between the characteristic parameter R and the parameters w and h. Deep learning refers to machine learning based on a deep artificial neural network architecture. A typical deep neural network architecture includes an input layer, two or more hidden layers, and an output layer. After training with a large number of samples, the network can be used to predict the corresponding parameters to be determined. The above is the fracture width inversion method based on the characteristic parameter of the broken wire leakage magnetic field ratio provided by the present invention, and each step is described in detail below.
[0055] Specifically, in step S1, the diameter D of the broken wire to be detected and the gradient detection point spacing d are determined, and the fracture width w and the detection height h of the single broken wire specimen are changed respectively, and the magnetic leakage signals under different fracture widths w and corresponding multiple detection heights set at equal intervals d based on the detection height h are obtained, and the symmetry axis of the leakage magnetic signal is used as the y-axis, and the axis of the broken wire to be detected is used as the x-axis to establish a rectangular coordinate system. Among them, the steel wire diameter D to be determined in step S1 has two sizes of φ5mm and φ7mm, and the measurement point spacing d is determined in combination with the actual size of the magnetic sensitive element 5. Further, in S1, when the fracture width w changes, the detection height h is kept unchanged; when the detection height h changes, the fracture width w is kept unchanged, that is, the fracture width w is kept unchanged and the detection height h is kept unchanged.
[0056] In step S2, a denoised signal is obtained through the acquired magnetic leakage signal, and the magnetic field value of the denoised signal on the y-axis of the rectangular coordinate system is extracted as a characteristic magnetic field signal B corresponding to the fracture width and the detection height.
[0057] After all the leakage magnetic signals corresponding to the detection height under different fracture widths are obtained in step S1, the respective denoised signals are obtained through the obtained leakage magnetic signals. The extracted denoised signals are signals with the background field subtracted. The magnetic field value of each denoised signal on the symmetry axis (x=0) is extracted as the characteristic magnetic field signal B=(B h ,B h+d ,B h+2d ,..,B h+nd ),
[0058] and
[0059] In formula (4), B h+nd represents the magnetic field value of the signal minus the background field on the ordinate of the rectangular coordinate system, Q represents the point magnetic charge of equal value distributed on both sides of the fracture, μ0 represents the magnetic permeability of vacuum, g represents the sum of the wire diameter D and the fracture width w, and n is an integer starting from 0.
[0060] In step S3, the corresponding ratio characteristic parameter R is calculated using the acquired characteristic magnetic field signal B to form sample data, and the sample data is randomly divided into training sample data and test sample data. Since the R value is the ratio of the magnetic field signal B at two adjacent detection heights, the corresponding ratio characteristic parameter R can be calculated by the acquired characteristic magnetic field signal B, and a set of R values, i.e., sample data, is formed. The acquired sample data is randomly divided into training sample data and test sample data for subsequent training and testing. Specifically, the ratio characteristic parameter R=(R h ,R h+d ,R h+2d , .., R h+kd ), R includes at least 2 values, where:
[0061]
[0062] In formula (5), k is an integer starting from 0. Since the R value is the ratio of two magnetic field intensities, it has a nonlinear relationship with w and D, and the h value and d value are both known, so the R value is not affected by other parameters. Therefore, after the data is calibrated, even if the calibration environment changes, such as cable specifications, probe specifications and detection lift-off, these data calibration environment changes will not affect the R value, and there is no need to recalibrate. The introduction of the ratio characteristic parameter R makes the calibration data universal and has good stability.
[0063] In order to facilitate data input or output, the collected data are formed into a data set. In the data set, the width of the cutout w is set in j cases, namely w1~wj, j≥2, and the detection height h is set in i cases, namely h1~hi, i≥2. The deep neural network is trained according to the R values obtained under different combinations of w and h to establish the mapping relationship between the feature parameter R and the two parameters. For example, R h+kd When k=2 in , when the input and output formats of the training samples are as follows, and:
[0064]
[0065] The input and output formats of the training samples are obtained from formula (6). For different fracture widths w and detection heights h, there are i*j observation values of R:
[0066]
[0067] The parameter mapping relationship obtained by i*j observations is obtained by formula (7).
[0068] In step S4, an initial network model based on a deep neural network is established, and the initial network model is trained by training sample data to obtain a trained network model, thereby completing the establishment of an inversion model. The initial network model is a two-layer forward feedback network with Sigmoid hidden neurons and linear output neurons. Then, the training sample data is substituted into the initial network model for training to obtain a mature trained network model. The Bayesian regularization method (BR algorithm) is used for training, and the loss function in this regression problem selects the mean square error (MSE), which is expressed as:
[0069]
[0070] In formula (8), y k Indicates the true value of the fracture width w (or detection height h) output in the kth group in each round of training; Represents y k In each corresponding round of training, the output value of the fracture width w (or detection height h) output by the network model in the kth group. The trained network can determine the nonlinear mapping relationship between the characteristic parameter R and the parameters w and h, and realize the inversion of the fracture width w by detecting the R value, and n represents the total number of groups.
[0071] Furthermore, the method further includes S5, testing the test sample data through the trained network model to complete the inversion of the fracture width. The prediction variable of the training sample data is the R value, and the R value of the prediction variable is substituted into the trained network model after the training, and the output response is the detection height h and the fracture width w, that is, the inversion process with the R value as input and the detection height h and the fracture width w as output is completed.
[0072] In order to explain the present invention more clearly, the following examples are used to illustrate that this embodiment uses COMSOL Multiphysics simulation software to simulate the magnetic flux leakage signal and create an inversion database. Among them, it is important to understand the meaning of "detection height h". The detection height h is the height from the broken wire surface to the detection point to be detected. It is based on Figure 4 The cross-sectional diagram of the cable to be tested is defined as Figure 4 Schematic diagram of the arrangement of cable magnetic leakage detection elements in an embodiment of the present invention.
[0073] like Figure 4 As shown, the cable includes a parallel steel wire bundle 1 and a PE sheath 2 arranged outside the parallel steel wire bundle 1. The parallel steel wire bundle 1 contains a broken wire 6 to be detected. The broken wire 6 to be detected is located in the surface steel wire just below the center of the magnetic yoke. The magnetic sensitive element 5 is located above the broken wire 6 to be detected. Figure 5 The permanent magnet 3 and the yoke 4 connected together are arranged on the surface of the cable to magnetize the parallel steel wire bundle 1 in the cable.
[0074] The process of fracture width inversion mainly includes the following steps:
[0075] Step 1: Determine the D value: Combined with the description of the cable wire specifications in the national standard "GB / T18365-2018", the steel wire used has only two sizes of φ5mm and φ7mm. In this embodiment, φ5mm steel wire is used for simulation; Determine the d value: The d value is the spacing between the gradient detection points. The d value can be reasonably designed according to the actual detection conditions. In this embodiment, the detection spacing d is set to 1mm.
[0076] The cable model was established using COMSOL Multiphysics simulation software. The material used for the permanent magnet 3 was NdFeB N35, and the coercive force Hc was set to 876 kA / m. The magnetizer was 10 mm away from the surface of the parallel steel wire bundle 1, simulating the magnetization of the internal steel wire bundle under the thicker PE sheath 2 structure. In the simulation, the steel wire diameter was 5 mm, and a total of 55 steel wires formed the internal parallel steel wire bundle 1. The total length of the steel wire bundle was 500 mm, and the broken wire was set at the midpoint of the axial position of the steel wire bundle. The magnetizer size parameters are shown in Table 1
[0077] Table 1 Magnetizer parameters
[0078]
[0079] The fracture width w = 2, 4, 6, 8, 10, 12, 14, 16, 18, 20 mm was changed, and the complete leakage magnetic signal along the scanning path was obtained based on d = 1 mm when the detection height h was 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14 mm. The total length of the scanning path was about 40 mm before and after the fracture.
[0080] Step 2: Extract the signal value at the position of the symmetric axis of the broken wire leakage magnetic field under different fracture widths w and detection heights h, and subtract the respective background fields to obtain the unbiased value, as the magnetic field characteristic value B under the fracture width w and detection height h = (B h_w ,B h+d_w ,B h+2d_w ,B h+3d_w ).
[0081] Step 3: Using the magnetic field characteristic value signal B = (B h_w ,B h+d_w ,B h+2d_w ,B h+3d_w ), calculate the characteristic value R under different fracture width w and detection height h = (R h_w , R h+d_w , R h+2d_w ) is used as the data for network training, and the final R value change diagram is as follows Figure 6 As shown, these data are combined as a training set for the deep neural network.
[0082] The left side of the following data set is the input, which has 3 features and i observations. The three features are the R values when the fracture width is w and the detection height is h, h+d and h+2d, respectively denoted as R h_w , R h+d_w and R h+2d_w The right side of the following data set is the output, which has 2 features and i observations, which are the detection height h and fracture width w corresponding to the input variables. Combined with the existing finite element results, the width w1~w j They are set to 2, 4, 6, 8, 10, 12, 14, 16, 18, and 20 respectively. There are 10 detection height values under the same width w, namely h1~h i They are set to 3, 4, 5, 6, 7, 8, 9, 10, 11, and 12 respectively, so the final training data set has a total of 100 observations. The training set format is as follows. Replace the specific values obtained above to complete the production of the training set.
[0083]
[0084] 15 samples are randomly selected from the 100 samples in the training set as test samples.
[0085] Step 4: This embodiment uses a two-layer feedforward network with Sigmoid hidden neurons and linear output neurons. In view of the slow convergence speed and easy to fall into the local optimum during the training process, the Bayesian regularization method (BR algorithm) is used for training. This method can prevent the occurrence of overfitting during the training process and improve the generalization ability of the model. The initial model of the deep neural network is designed with two hidden layers, and the number of neurons in each layer is 20. The training algorithm uses the Bayesian regularization algorithm, and the loss function in this regression problem uses the mean square error (MSE).
[0086] The fit of the trained model reached 0.9999 and the mean square error reached 0.0024. The test results of the test sample set showed that the fit reached 0.9898 and the mean square error was 0.2302. In order to further test the generalization ability of the trained network, two sets of simulation tests were conducted.
[0087] Step 5: To simulate the working condition of adjusting the lifting distance of the magnetic sensitive element 5 during the detection, the R value data at four detection heights of h = 9, 10, 11, and 12 mm were simulated respectively when the fracture width w = 2.5, 4.5, and 6.5 mm and the equal spacing d = 1 mm, as an additional test set to predict the fracture width w information. The mean square error (MSE) of the trained network predicting the fracture width reached 0.0029, and the maximum absolute error and relative error of the width quantization were 0.109 mm and 2.85% respectively. The prediction results are relatively accurate. The results are shown in Table 2 below:
[0088] Table 2 Calculation results
[0089]
[0090] In order to simulate the working condition that the magnetization state of the steel wire is affected by the change of the magnetizer or cable specifications during the test, the magnetization conditions in the model are changed, and the coercive force of the magnet is set to Hc = 976kA / m, 1076kA / m and 1176kA / m respectively. The R value data under h = 9, 10, 11, and 12mm when the fracture width w = 10mm is calculated as an additional test set to predict the fracture width w. The results are shown in Table 3. From the error results, it can be seen that the mean square error of the trained network predicting the fracture width w reaches 0.0452, and the maximum absolute error and relative error of the quantization are 0.380mm and 3.80% respectively. The prediction results are relatively accurate.
[0091] Figure 3 Calculation results
[0092]
[0093] It can be seen from the above related results that the proposed deep neural network model based on the characteristic parameter R is relatively accurate for the inversion of the fracture width information, which proves the effectiveness and accuracy of the constructed fracture width inversion method based on the characteristic parameter of the broken wire leakage magnetic field ratio, and at the same time provides a new solution to the complex calibration problem when the detection conditions change.
[0094] It will be easily understood by those skilled in the art that 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 in the protection scope of the present invention.
Claims
1. A method for constructing a fracture width inversion model based on ratio characteristic parameters, characterized in that: The following steps are involved: S1. Determine the diameter D of the broken wire to be detected and the interval d between the gradient detection points, respectively change the fracture width w and the detection height h of the single broken wire specimen, obtain the magnetic leakage signals at different fracture widths w and the corresponding multiple detection heights set at equal intervals d based on the detection height h, and establish a rectangular coordinate system with the symmetry axis of the leakage magnetic signal as the y-axis and the axis of the broken wire to be detected as the x-axis; S2, obtaining a denoised signal through the obtained leakage magnetic signal, and extracting the magnetic field value of the denoised signal on the y-axis of the rectangular coordinate system as a characteristic magnetic field signal B corresponding to the fracture width and the detection height; S3, using the acquired characteristic magnetic field signal B to calculate the corresponding ratio characteristic parameter R, to form sample data, and randomly divide the sample data into training sample data and test sample data; S4. Establish an initial network model based on a deep neural network, and use sample data to train the initial network model to obtain a trained network model, thereby completing the establishment of an inversion model.
2. The method for constructing a fracture width inversion model based on ratio characteristic parameters according to claim 1, characterized in that: Ratio characteristic parameter R = (R h ,R h+d ,R h+2d , .., R h+kd ), R contains at least 2 values, and Here, k is an integer starting from 0.
3. The method for constructing a fracture width inversion model based on ratio characteristic parameters according to claim 1, characterized in that: Characteristic magnetic field signal B=(B h ,B h+d ,B h+2d ,..,B h+nd ), B contains at least 3 values, and In the formula, B h+nd represents the magnetic field value of the denoised signal on the y-axis of the rectangular coordinate system, Q represents the point magnetic charge of equal value distributed on both sides of the fracture, μ0 represents the vacuum magnetic permeability, g represents the sum of the wire diameter D and the fracture width w, and n is an integer starting from 0.
4. The method for constructing a fracture width inversion model based on ratio characteristic parameters according to claim 1, characterized in that: The method also includes step S5: substituting the test sample data into the trained network model to verify its effectiveness.
5. The method for constructing a fracture width inversion model based on ratio characteristic parameters according to any one of claims 1 to 4, characterized in that: The denoised signal in step S2 is a signal with the background field subtracted.
6. The method for constructing a fracture width inversion model based on ratio characteristic parameters according to claim 5, characterized in that: In step S1, when the fracture width w changes, the detection height h is kept unchanged; when the detection height h changes, the fracture width w is kept unchanged.
7. The method for constructing a fracture width inversion model based on ratio characteristic parameters according to claim 6, characterized in that: In step S4, the input of the initial network model is the R value, and the output is the detection height h and the fracture width w; the fracture width w in the data set is set in j cases, namely w1~w j , j ≥ 2, the detection height h in the training set is set to i cases, i.e. h1~h i , i ≥ 2, and the deep neural network is trained according to the R values obtained under different fracture width w and detection height h conditions to establish the mapping relationship between the ratio characteristic parameter R and the two parameters.
8. The method for constructing a fracture width inversion model based on ratio characteristic parameters according to claim 7, characterized in that: In step S4, the initial network model is a two-layer feedforward network with Sigmoid hidden neurons and linear output neurons.
9. The method for constructing a fracture width inversion model based on ratio characteristic parameters according to claim 8, characterized in that: The Bayesian regularization method is used for initial network model training, and the loss function in the training process is the mean square error MSE, which is expressed as: Among them, y k Indicates the true value of the fracture width w or detection height h output in the kth group in each round of training; Represents y k Correspondingly, in each round of training, the output value of the fracture width w or the detection height h output by the network model in the kth group is output, and n represents the total number of groups.
