Method and device for detecting position and depth of blind hole of 304 steel plate under insulating layer

Abstract

The invention provides a PEC detection method and device for the position and depth of a blind hole of a 304 steel plate under an insulating layer. The method comprises the following steps: firstly, placing a PEC probe above an original point of a to-be-detected area, detecting a blind hole along an X axis, and recording a distance R1 and a depth H1; then, the probe moves R / 2 to a detection point 2 in the positive direction of the X axis, and R2 and H2 are recorded; then, moving R / 2 to a detection point 3 along the positive direction or the negative direction of the Y axis, and recording R3 and H3; and obtaining the coordinates and the depth H of the blind hole through calculation. In the process, Z-Score standardization and PCA dimension reduction are carried out on PEC signals, a CI-I GA-BP model 1 is used for evaluating the blind hole distance R1 to obtain a blind hole distance evaluation value, features obtained after PCA dimension reduction are extracted and fused with the blind hole distance evaluation value, a CV-I GA-BP model 2 is used for evaluating fused data, and finally a blind hole depth evaluation value is obtained. The detection device comprises a rack; provided is an encoder track detection device. The technical scheme has the advantages of high accuracy, high efficiency, environment friendliness and the like when the blind hole of the 304 stainless steel plate is detected.

Description

Technical Field

[0001] The present invention relates to the technical field of nondestructive testing, and in particular to a method and a device for detecting the position and depth of a blind hole of a 304 steel plate under an insulating layer. Background Art

[0002] 304 steel plate has been widely used in major equipment such as offshore platforms, petrochemical facilities, and LNG carriers due to its excellent corrosion resistance, strength, heat resistance, and good processing performance. For reasons of safety and energy saving, the surface of 304 steel plate is generally painted or covered with an insulating layer such as anti-rust paint, rock wool, or polyurethane. Some insulating layers have gaps; some insulating layers may have gaps due to vibration, collision, thermal expansion and contraction. Rainwater or water vapor in the air contacts the 304 steel plate through the gap, forming a primary battery that continuously corrodes and invades, and it is easy to form pitting corrosion on the surface of the 304 steel plate, and gradually develop into a corrosion blind hole. If the location and depth of the corrosion blind hole are not detected in time, it is difficult to ensure the safe operation of major equipment. Since the gaps in the insulation layer are generally narrow, it is difficult to directly observe the corrosion of the 304 steel plate under the insulation layer. It is often necessary to remove the insulation layer for observation and detection, and then restore the insulation layer. However, removing the insulation layer is mostly a high-risk operation, with harsh implementation conditions, labor-intensive and time-consuming, and gaps are also prone to occur at the junction of the new and old insulation layers. Therefore, there is an urgent need for a method for detecting the position and depth of blind holes in 304 steel plates without damaging the insulating layer.

[0003] The existing PEC detection methods for local defects of ferromagnetic components such as carbon steel are difficult to be directly used for the detection of the position and depth of the corrosion blind holes of 304 stainless steel (non-ferromagnetic material) plates under the insulating layer. For methods that require the signal of the defect-free area as a reference signal, it is difficult to select the defect-free area without removing the insulating layer. For methods that require the scanning path to be as close as possible to or even directly above the defect, or the sensor to cover the area where the defect is located, when the defect location is unknown, the only way to reduce missed detection is to encrypt the scanning points, which not only has low detection efficiency, but also leads to missed detection in areas that are inaccessible to the sensor. In addition, for machine learning methods such as random forests and artificial neural networks, the detection effect of certain specific defects is not as good as that of artificial defects, and the artificially selected features are greatly affected by experience. Among them, the single-value point features are also easily affected by noise, which is difficult to be applied to the detection of the position and depth of the corrosion blind holes of 304 stainless steel plates under the insulating layer. The optimization of BP neural network based on genetic algorithm (GA) can realize the detection of crack depth and width of ferromagnetic components, but the selection of signal features and model parameters used in this method depends on experience, and a single model is difficult to accurately evaluate multiple parameters to be tested. The prior art generally only intercepts the detection signal corresponding to the falling edge of the rectangular wave excitation and thereafter for processing, and processes less the full waveform received signal, while the latter contains more information and is more conducive to the simultaneous detection of multiple parameters.

[0004] Pulsed eddy current testing (PEC) technology has the ability to detect local defects of conductive components through insulating layers. Chinese invention patent ZL202110684761.3 discloses a corrosion defect multi-layer penetration imaging detection system and method, which realizes the imaging detection of pitting defects in carbon steel plates. Since the product of the electrical conductivity and magnetic permeability of 304 steel is 1-2 orders of magnitude smaller than the product of the electrical conductivity and magnetic permeability of carbon steel, the dissipation rate of eddy current in 304 steel is much greater than that of carbon steel. The PEC signal corresponding to 304 steel decays faster and is more susceptible to noise. It is difficult to screen out enough available data based on this method. Chinese invention patent ZL202111248520.0 uses the signal collected from the air as the first reference signal, and the signal collected from the defect-free test block as the second reference signal. The signal collected from the sample is differentially processed with the first reference signal and the second reference model, and the defect state of the coating layer pipeline is judged according to the signal after differential processing. Because in the process of testing the sample, it is difficult to ensure that it is consistent with the first reference signal when it is affected by environmental noise, and it is also difficult to ensure that the sample is different from the second reference signal only in defects, but also includes differences in environmental noise, temperature, and stress. Chinese invention patent ZL202110684761.3 discloses a defect depth identification method and system based on marginal spectrum centroid detection, which realizes the detection of defect depth on high-speed steel, but this method also requires differential processing of the reference signal. Without removing the insulating layer, it is generally difficult to ensure that there is only a difference between the test point and the reference point in the presence or absence of defects. Chinese invention patent ZL202110684761.3 discloses a U-shaped magnetic conductor focusing probe and its pulse eddy current detection method, which realizes the detection of rectangular flat-bottomed blind hole defects in the center of 304L steel plate. However, this method does not quantitatively give the defect depth, and when the defect location is unknown, the scanning points need to be encrypted to make the scanning path as close as possible or even directly above the defect. Reference 1 (Wang, Z., Ye, P., Qiu, F., Tian, ​​G., & Woo, WL (2020). Crack characterization in ferromagnetic steels by pulsed eddy current technique based on GA-BP neural network model. Journal of magnetism and magnetic materials, 500, 166412) established the relationship between crack depth, width and the difference value of magnetic induction intensity in the Z direction by constructing a GA-BP neural network model. This method not only relies on experience for feature selection, but also on model parameters such as population size, crossover probability, and mutation probability. Summary of the invention

[0005] The purpose of the present invention is to provide a PEC detection method and device for the position and depth of blind holes in 304 steel plates under an insulating layer, in response to the deficiencies in the prior art. The present invention provides a PEC detection method and device for the position and depth of blind holes in 304 steel plates under an insulating layer, with the support of the National Natural Science Foundation Youth Project "Research on the Mechanism and Method of Pulsed Eddy Current Detection of Local Defects in Non-Ferromagnetic Metal Components with Covering Layers" (62003056), so as to automatically select signal features and model parameters, generate associations between multiple target features, and improve the accuracy of the final evaluation effect.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is a detection method of a PEC detection device for the position and depth of blind holes in a 304 steel plate under an insulating layer, which is characterized by comprising the following steps:

[0007] S1: Place the PEC probe above the origin of the current inspection area of ​​the 304 steel plate with an insulating layer, perform PEC inspection along the positive direction of the X-axis, and determine whether a blind hole is detected;

[0008] S2: If a blind hole is detected, the distance R1 between the blind hole and the center axis of the probe and the blind hole depth H1 are recorded, and the current position of the PEC probe is recorded as detection point 1;

[0009] S3: Move the PEC probe along the positive direction of the X-axis by R1 / 2, record the current position of the PEC probe as detection point 2, perform PEC detection at detection point 2, and obtain the distance R2 between the blind hole and the probe and the blind hole depth H2;

[0010] S4: Return the PEC probe to the detection point 1, move R1 / 2 along the positive direction of the Y axis, and perform PEC detection;

[0011] S5: If a blind hole is detected, the distance R3 between the blind hole and the central axis of the probe and the depth H3 of the blind hole are recorded, and the current position of the PEC probe is recorded as detection point 3;

[0012] S6: If no blind hole is detected, the PEC probe is returned to the detection point 1, and moved R1 / 2 along the negative direction of the Y axis, and the PEC detection is performed to obtain the distance R3 between the blind hole and the central axis of the probe and the depth D3 of the blind hole, and the current position of the PEC probe is recorded as the detection point 3;

[0013] S7: From this, the coordinates of the blind hole relative to the center axis of the probe are obtained as follows:

[0014] The blind hole depth is

[0015] S8: If no blind hole is detected and the edge of the probe detection range has reached the edge of the positive direction of the X-axis of the area to be inspected, the detection in the current area to be inspected is completed.

[0016] It should be noted that the distance between the current area to be inspected and the edge of the 304 steel plate should be greater than or equal to the probe detection radius r to avoid the situation where the edge cannot be detected during the detection process, and the y-direction size of the current area to be inspected should be less than or equal to the probe detection diameter 2r, and the origin of the detection area is located at the intersection of the edge of the current area to be inspected in the X-axis direction and the center line of the y-direction to ensure that every position can be detected.

[0017] In order to obtain the actual value of the target feature blind hole distance (R1) and the blind hole depth (H1) in the above steps as a training data set, the present invention also provides the following detection method steps:

[0018] S21, using a device to obtain full-wave time series PEC signal data of the stainless steel plate, and obtaining a data set X containing PEC detection signals;

[0019] S22, performing Z-Score standardization and PCA (Principal Component Analysis) dimensionality reduction on the data set X, and obtaining data x containing PCA features of the PEC detection signal pca ;

[0020] S23, use x pca , the distance R1 between the blind hole and the PEC probe The CV-IGA-BP (Cross-Validation-Improved Genetic Algorithm-Back Propagation: BP neural network based on cross-validation and improved genetic algorithm optimization) model was used to Train to obtain CV-IGA-BP model 1 and obtain the evaluation value x of the blind hole distance pred_std , perform Z-Score standardization on the evaluation value of blind hole distance;

[0021] S24, the normalized blind hole distance evaluation value x pred_std PCA feature dataset x of PEC detection signal pca Perform data fusion to obtain data set D;

[0022] S25, train using CV-IGA-BP model D to obtain CV-IGA-BP model 2, and finally obtain the evaluation value x of the defect depth pred_std .

[0023] As a preferred embodiment, S1 includes:

[0024] By processing the simulation data, the pulsed eddy current detection signal of the blind hole position and depth of the 304 steel plate under the insulating layer with a signal-to-noise ratio of 60dB is obtained, and the actual value of the target feature blind hole distance (R1) and the blind hole depth (H1) are obtained as the training data set.

[0025] The pulsed eddy current signal of 304 steel plate in the experimental environment is obtained by using the equipment as the test data set X.

[0026] As a preferred embodiment, S21 includes:

[0027] The full-wave time series characteristics of the pulsed eddy current detection signal in the training data x = [x 11 …x ij ]i,j=1…m.

[0028] To prevent gradient disappearance and gradient explosion during data training: where μ j is the mean of the corresponding column feature, m is the total number of each feature, x ij is the jth data of the i-th feature, σ j is the standard deviation of the corresponding column feature, y ij is the jth data of the newly generated i-th feature.

[0029] Use PCA method to correspond to y ij Feature data is processed for dimensionality reduction: Taking a feature as an example, a i for y i The variance of a feature in is calculated as follows:

[0030] b i For the remaining features, their covariance is calculated as follows:

[0031] Full-wave time series characteristic covariance matrix of pulsed eddy current testing signal x pca =μX.

[0032] Finally, the eigenvalue of the full-wave time series characteristics of the entire PECT (pulsed eddy current detection) signal is calculated, and the characteristic sequence μ is deduced and multiplied by the covariance X matrix to obtain the new characteristic data x after dimensionality reduction. pca .

[0033] As a preferred embodiment, S23 includes:

[0034] S231, assuming that the pre-processed pulsed eddy current detection signal data set is x pca , the target feature is the blind hole distance. The population size of the genetic algorithm is M, and the number of evolutions is GEN.

[0035] S232, build BP neural network in is the net input to the hidden layer, is the weight on the connection starting from input unit 1 (input unit 1: the first neuron in the input layer, which receives the value of the first input feature, which is the first feature in the features after PCA dimensionality reduction.), is the bias term, and “h” refers to the quantity on the hidden layer. Among them I j is the output of the hidden layer and, is the activation function. pca is the input feature and defect distance is the target feature.

[0036] The entire dataset is randomly divided into 5 different subsets, each of which is called a fold. The model is then trained and evaluated 5 times, with one fold selected as the test set each time and the other 4 folds as the training set. This way, each subset has the opportunity to be used as a test set, resulting in 5 evaluation results, which fully evaluates the generalization ability of the model and avoids overfitting. 20% of the training set is randomly divided as a validation set for early stopping and to check for overfitting during training. The range of hyperparameters is as follows:

[0037]

[0038] S233, the weight and threshold range of the BP neural network is [-1, 1], so real numbers (floating point numbers) are used for encoding. In step S32, we have obtained the weight vector and threshold vector in the optimal hyperparameters, and initialized them as: num = len(W) + len(b), where W is the weight matrix of the BP neural network, b is the bias matrix of the BP neural network, and num is the total number of parameters of the BP neural network calculated.

[0039] S234, varbound = [(-1,1)] × num, varbound is the parameter vector of weights and thresholds for initializing the BP neural network, and also serves as the individual genome of the genetic algorithm.

[0040] S235, taking the inverse of the mean square error of the BP neural network as the fitness value of the improved genetic algorithm (IGA). F i =1 / E i , E i is the mean square error of the individual in the ith population, F i is the fitness value of the i-th individual in the population, which is used as the maximum optimization objective of the genetic algorithm.

[0041] S236, four-level selection function, assuming the population size is 40, sort by size, the first ten are directly retained as the first level, and the rest are divided into three levels. Two copies of the second level are retained, one copy of the third level is retained, and the fourth level is directly eliminated. Simulate natural elimination and optimize the population.

[0042] S237, Adaptive Crossover Function Among them, F is the maximum fitness of the two individuals to be crossed, F mean is the average fitness of the entire population, n is the number of evolutionary iterations, n max is the maximum number of iterations, P jmax Set to 0.9, P jmin Set to 0.4. Repeat step GEN*M times.

[0043] S238, adaptive variogram Among them, F is the maximum fitness among the individuals to be crossed, F mean is the average fitness of the entire population, n is the number of evolutionary iterations, n max is the maximum number of iterations, P jmax Set to 0.1, P jmin Set to 0.006. Repeat step GEN*M times.

[0044] S239, the optimal solution of the weight vector and the optimal solution of the threshold vector obtained by the improved genetic algorithm (IGA) are used as the initialization weight vector of the BP neural network and initialize the threshold vector

[0045] S2310, PCA features of pulsed eddy current detection signals and blind hole distance datasets using BP neural network Processing is performed to obtain the final optimal solution of the weight vector and the final threshold vector Optimal solution and CV-IGA-BP model1.

[0046] S2311, use CV-IGA-BP model 1 to process to obtain the evaluation value of the blind hole distance, and perform Z-Score standardization on the evaluation value of the blind hole distance for subsequent processing.

[0047] As a preferred embodiment, S24 includes:

[0048] S241, let the normalized blind hole distance evaluation value be x pred_std , the PCA feature data set of the pulsed eddy current detection signal is x pca .

[0049] S242, x pred_stdIn the form of column repeat completion, ensure that x pred_std and x pca The matrix shape remains consistent.

[0050] S243,x pred_std and x pca The linear weight sum between them is 1, let r be x pca The weight of x is 1-r. pred_std The weight of D = r × x pca +(1-r)×x pred_std , D is the training data for CV-IGA-BP model 2.

[0051] As a preferred embodiment, S5 includes:

[0052] S51, taking the data set D as the input feature and the blind hole depth as the target feature.

[0053] S52, as in the training step of the model 1 of the CV-IGA-BP in S23, the model 2 of the CV-IGA-BP is trained, and finally the evaluation value of the blind hole depth is obtained.

[0054] Among them, the PEC detection device for the position and depth of blind holes in 304 steel plates under the insulation layer includes:

[0055] A rack for carrying components of the testing device;

[0056] An encoder trajectory detection device, the encoder trajectory detection device comprises a bracket, a coupling, a meter wheel and an encoder, the bracket is mounted on the frame, the encoder is rotatably mounted on the bracket, the meter wheel is connected to the shaft end of the encoder through a coupling, two sets of directional wheel walking devices are mounted on both sides of the bottom of the frame, and the three-axis position display device is mounted on the top of the frame;

[0057] A regionalized pulsed eddy current testing device comprises an X-axis mechanism, a Y-axis mechanism and a Z-axis mechanism, and the detection unit is installed on the Y-axis mechanism.

[0058] Preferably, the X-axis mechanism, the Y-axis mechanism and the Z-axis mechanism all include a mounting plate, a first motor, a screw rod, a threaded slide, a slide rail and a connecting frame; the first motor is fixedly mounted on one side of the mounting plate, the screw rod is mounted on the output shaft of the first motor, the threaded slide is threadedly connected to the surface of the screw rod, and the threaded slide is slidably connected to the mounting plate via a slide rail; the connecting frame in the Y-axis mechanism is mounted on the connecting frame in the X-axis mechanism, and the connecting frame in the Z-axis mechanism is mounted on the mounting plate in the Y-axis mechanism.

[0059] Preferably, it also includes a second X-axis mechanism and a second Y-axis mechanism, the second X-axis mechanism includes a first fixed plate, a first reciprocating screw and a second threaded slide, the first reciprocating screw is installed inside the first fixed plate, and the second threaded slide is threadedly connected to the second reciprocating screw; the second Y-axis mechanism includes a second fixed plate, a second reciprocating screw and a third threaded slide, the second fixed plate is installed on the third threaded slide, the second reciprocating screw is fixed inside the second fixed plate, the third threaded slide is threadedly connected to the second reciprocating screw, and the third threaded slide is slidably connected to the second fixed plate.

[0060] Preferably, a toothed roller is installed on the lower side of the first fixed plate, one end of the toothed roller is connected to the first reciprocating screw through a one-way transmission member, the one-way transmission member includes two transmission wheels, two one-way bearings and two driving wheels, the two one-way bearings are installed on the above, the two driving wheels are respectively installed on the two one-way bearings, the two transmission wheels are respectively installed on one end of the first reciprocating screw and one end of the toothed roller, the two driving wheels are respectively connected to the two transmission wheels through transmission belts; a sliding member is installed on the lower side of the first fixed plate and on one side of the toothed roller, the sliding member includes a sliding arm and a sliding frame, the lower end of the sliding frame is sleeved on the surface of the sliding arm, the sliding frame is fixedly connected to the second threaded slide through a connecting arm, and a driving member is installed inside the sliding frame.

[0061] Preferably, the driving member includes a rotating shaft, a gear, a main bevel gear and a transmission mechanism, the rotating shaft is rotatably connected to the sliding frame, the main bevel gear and the gear are fixedly connected to the rotating shaft, and the gear is meshed with the gear roller; one end of the second reciprocating screw is connected to a slave bevel gear, and the slave bevel gear is connected to the main bevel gear through a transmission wheel of the transmission mechanism.

[0062] Compared with the related art, the method and device for detecting the position and depth of blind holes in 304 steel plates provided by the present invention have the following beneficial effects:

[0063] Compared with the prior art, the present invention uses a cross-validation BP neural network to obtain the optimal hyperparameters of the model, thereby avoiding the subjectivity of artificial hyperparameters, reducing manual operations, and improving efficiency and accuracy; at the same time, the initialization weight vector and initialization threshold vector of the BP neural network are optimized based on an improved genetic algorithm. Compared with the traditional genetic algorithm to optimize the BP neural network, its global search ability and local search ability are improved, and the average absolute error is smaller; the proposed feature fusion model, compared with the traditional method - directly merging the trained target features behind the training features, is more suitable for situations with more training features, and can achieve better results than the original data set. In addition, the PEC detection device for the position and depth of the corrosion blind hole of the 304 steel plate, the encoder trajectory detection device transmits the real-time displacement and trajectory to the host computer, and displays it through an image, which can be accurately positioned to determine the location of the corrosion blind hole;

[0064] Among them, the regionalized pulsed eddy current detection device can perform regional detection through the movement of the X, Y, and Z axes, and can measure 304 steel plates of different heights, reducing the time and labor intensity of manual operation, and does not need to disassemble the plates to be tested, making the operation simpler and more convenient, and significantly improving the detection efficiency. The design of the X, Y, and Z axis mechanisms enables the device to adapt to 304 steel plates of different sizes and shapes, improving the versatility and flexibility of the detection device. The displacement and trajectory data transmitted in real time can be recorded and stored, which is convenient for subsequent data analysis and quality traceability. The optimal hyperparameters of the model are automatically obtained by cross-validating the BP neural network, avoiding the subjectivity of artificially adjusting the hyperparameters and improving the efficiency of model training. In this process, the cross-validation BP neural network and the BP neural network optimized by the improved genetic algorithm are used to improve the generalization ability and accuracy of the model, making the detection results more accurate. It also reduces manual operation and disassembly of the plates to be tested, reduces the impact on the environment, and is more environmentally friendly. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 A flowchart of the PEC detection method and device for the position and depth of blind holes in 304 steel plates under the insulating layer provided by the present invention;

[0066] Figure 2 for Figure 1 Flowchart of the framework for evaluating blind hole distance and blind hole depth using the CV-IGA-BP model of the method and device;

[0067] Figure 3 for Figure 1 The training flow chart of CV-IGA-BP model 1 in the method and apparatus;

[0068] Figure 4 for Figure 1The training flow chart of CV-IGA-BP model 2 in the method and apparatus;

[0069] Figure 5 for Figure 1 A schematic diagram of the PEC detection process of the method and device (front view);

[0070] Figure 6 for Figure 1 Schematic diagram of the PEC detection process of the method and device (top view);

[0071] Figure 7 for Figure 1 PEC detection signal diagram of the method and device;

[0072] Figure 8 for Figure 1 PCA dimension reduction principal component index diagram of the method and device;

[0073] Fig. 9 for Figure 1 Schematic diagram of PCA feature data set of pulsed eddy current detection signal of the method and device;

[0074] Fig.10 for Figure 1 Data feature graph after fusion of new data set D of the method and device;

[0075] Fig.11 for Figure 1 IGA algorithm flow chart of the method and apparatus;

[0076] Fig.12 for Figure 1 A four-stage selection flow chart in the IGA of the method and apparatus;

[0077] Fig.13 for Figure 1 A flow chart of the adaptive crossover function in the IGA of the method and apparatus;

[0078] Fig.14 for Figure 1 A flowchart of the adaptive variogram in the IGA of the method and apparatus;

[0079] Fig.15 for Figure 1 Blind hole defect distance assessment diagram of 304 steel plate using the method and apparatus;

[0080] Fig.16 for Figure 1 Defect depth comparison evaluation diagram of 304 steel plate by the method and the device; (a) is the data set x pca (a) is the evaluation result obtained by training with dataset D, and (b) is the evaluation result obtained by training with dataset D;

[0081] Fig.17 for Figure 1 Comparative evaluation diagram of defect depth of stainless steel plate at different defect distances using the method and device; (a) is the data set x pca (a) is the evaluation result obtained by training with dataset D, and (b) is the evaluation result obtained by training with dataset D;

[0082] Fig.18 for Figure 1 The structural diagram of the device in FIG.

[0083] Fig.19 for Fig.18 A schematic diagram of the structure of a regionalized pulsed eddy current testing device;

[0084] Fig. 20 for Fig.18 A schematic diagram of the structure of the directional wheel walking device;

[0085] Fig.21 for Fig.18 A schematic structural diagram of a third embodiment of a PEC detection device;

[0086] Fig. 22 for Fig.21 A side view of the PEC detection device in FIG.

[0087] Fig.23 for Fig. 22 Enlarged view of part A in FIG.

[0088] Numbers in the figure: 1, frame, 2, encoder trajectory detection device, 21, bracket, 22, coupling, 23, meter wheel, 24, encoder, 3, directional wheel walking device, 4, three-axis position display device, 5, regional pulse eddy current detection device, 51, X-axis mechanism, 52, Y-axis mechanism, 53, Z-axis mechanism, 511, mounting plate, 512, first motor, 513, screw rod, 514, threaded slide, 515, slide rail, 516, connecting frame, 6, second X-axis mechanism, 61, first fixed plate, 62, first reciprocating wire Rod, 63, second threaded slide; 7, second Y-axis mechanism, 71, second fixed plate, 72, second reciprocating screw, 73, third threaded slide, 8, gear roller, 9, driving member, 91, rotating shaft, 92, gear, 93, main bevel gear, 94 transmission mechanism, 10, supporting slide, 11, sliding member, 111, sliding arm, 112, sliding frame, 12, one-way transmission member, 121, transmission wheel, 122, transmission belt, 123, one-way bearing, 124 driving wheel, 13, from bevel gear, 14, connecting arm, 15, second motor. DETAILED DESCRIPTION

[0089] Figure 1This is a step diagram of the PEC detection device for the position and depth of blind holes in 304 steel plates provided by the present invention. The present invention uses an artificial neural network algorithm in machine learning to establish a model. The model results are as follows: Figure 2 As shown in the figure, it is called the CV-IGA-BP (BP neural network based on cross-validation and improved genetic algorithm optimization) model. The model is divided into two parts. The first part is the search for the best hyperparameters based on the cross-validation BP neural network, and the second part is the BP neural network based on the improved genetic algorithm to obtain the optimal solution of the initial weight vector and the optimal solution of the initial threshold vector. Finally, the CV-IGA-BP model is integrated to evaluate the blind hole distance of 304 steel plate. For the evaluation of the blind hole depth, the concept of feature fusion framework is added to the input features, which further improves the stability and accuracy of the blind hole depth evaluation. Figure 1-4 The following is the overall framework flowchart and the training flowchart of the two models.

[0090] First embodiment:

[0091] The method for detecting the position and depth of blind holes of 304 steel plates under the insulating layer of the present invention comprises the following steps:

[0092] S1: Place the PEC probe above the origin of the current inspection area of ​​the 304 steel plate with an insulating layer, perform PEC inspection along the positive direction of the X-axis, and determine whether a blind hole is detected;

[0093] S2: If a blind hole is detected, the distance R1 between the blind hole and the center axis of the probe and the blind hole depth H1 are recorded, and the current position of the PEC probe is recorded as detection point 1;

[0094] S3: Move the PEC probe along the positive direction of the X-axis by R1 / 2, record the current position of the PEC probe as detection point 2, perform PEC detection at detection point 2, and obtain the distance R2 between the blind hole and the probe and the blind hole depth H2;

[0095] S4: Return the PEC probe to the detection point 1, move R1 / 2 along the positive direction of the Y axis, and perform PEC detection;

[0096] S5: If a blind hole is detected, the distance R3 between the blind hole and the central axis of the probe and the depth H3 of the blind hole are recorded, and the current position of the PEC probe is recorded as detection point 3;

[0097] S6: If no blind hole is detected, the PEC probe is returned to the detection point 1, and moved R1 / 2 along the negative direction of the Y axis, and the PEC detection is performed to obtain the distance R3 between the blind hole and the central axis of the probe and the depth D3 of the blind hole, and the current position of the PEC probe is recorded as the detection point 3;

[0098] S7: From this, the coordinates of the blind hole relative to the center axis of the probe are obtained as follows:

[0099] The blind hole depth is

[0100] S8: If no blind hole is detected and the edge of the probe detection range has reached the edge of the positive direction of the X-axis of the area to be inspected, the detection in the current area to be inspected is completed.

[0101] It should be noted that the distance between the current inspection area and the edge of the 304 steel plate should be greater than or equal to the probe detection radius r to avoid the situation where the edge cannot be detected during the detection process, and the y-direction size of the current inspection area should be less than or equal to the probe detection diameter 2r, and the origin of the detection area is located at the intersection of the X-axis edge of the current inspection area and the y-direction center line to ensure that every position can be detected;

[0102] In order to obtain the actual value of the target characteristic corrosion blind hole distance (R1) and the blind hole depth (H1) in the above steps as a training data set, the present invention also provides the following detection method steps: S11, obtaining the pulse eddy current detection signal data set includes: first, by processing the simulation data to obtain the pulse eddy current detection signal of the stainless steel plate defect with a signal-to-noise ratio of 60dB, the target characteristic blind hole distance (R1) actual value and the blind hole depth (H1) actual value are obtained as the training data set, and then the equipment is used to obtain the pulse eddy current signal of the stainless steel plate in the experimental environment as the test data set. Among them, the training set samples are 230, the validation set samples are 66, and the test set samples are 141, such as Figure 7 shown.

[0103] Prevent pulse eddy current detection signal x=[x 11 …x ij ]i,j=1…m, gradient explosion / vanishing occurs during training, and Z-Score normalization is performed. where μ j is the mean of the corresponding column feature, m is the total number of each feature, x ij is the jth data of the i-th feature, σ j is the standard deviation of the corresponding column feature, y ij is the jth data of the newly generated i-th feature.

[0104] Use PCA method to correspond to y ij The feature data is processed for dimensionality reduction, y ij The feature dimension is 488: Taking one feature as an example, a i for y i The variance of a feature in is calculated as follows: b i For the remaining features, their covariance is calculated as follows: Full-wave time series characteristic covariance matrix of pulsed eddy current testing signal x pca =μX.

[0105] We choose the cumulative variance contribution rate to be 98%, and the final x pca The dimension of is 23, such as Figure 8 shown.

[0106] S231, assuming that the pre-processed pulsed eddy current detection signal data set is x pca ,like Fig. 9 As shown, the target feature is the defect distance. The genetic algorithm population size is M = 40, and the number of evolutions is GEN = 50.

[0107] S232, build BP neural network in is the net input to the hidden layer, is the weight on the connection starting from input unit 1, is the bias term, and h refers to the quantity on the hidden layer. Among them I j is the output of the hidden layer and, is the activation function. pca is the input feature and defect distance is the target feature.

[0108] The entire dataset is randomly divided into 5 different subsets, each of which is called a fold. The model is then trained and evaluated 5 times. Each time, one fold is selected as the test set, and the other 4 folds are used as the training set. In this way, each subset has the opportunity to be used as a test set, resulting in 5 evaluation results, which fully evaluates the generalization ability of the model and avoids overfitting. 20% of the training set is randomly divided as a validation set for early stopping and to check overfitting during training. The range of hyperparameters is as follows:

[0109] Table 1

[0110]

[0111] Model1 best hyperparameters

[0112] Table 2

[0113] Hidden layer size hidden_layer_sizes (16,16) Activation Function activation 'tanh Learning rate initialization learning_rate_init 0.1 Maximum number of iterations max_iter 1000

[0114]

[0115] S233, the weight and threshold range of BP neural network is [-1,1], so real number (floating point number) encoding is adopted. In S32, we have obtained the weight vector and threshold vector in the optimal hyperparameters, and initialized: num=len(W)+len(b), where W is the weight matrix of BP neural network, b is the bias matrix of BP neural network, and num is the total number of parameters of BP neural network calculated.

[0116] S234, varbound = [(-1,1)] × num, varbound is the parameter vector of weights and thresholds for initializing the BP neural network, and also serves as the individual genome of the genetic algorithm.

[0117] S325, IGA algorithm Fig.11 As shown, the inverse of the mean square error of the BP neural network is used as the fitness value of the improved genetic algorithm (IGA). F i =1 / E i , E i is the mean square error of the individual in the ith population, F i is the fitness value of the i-th individual in the population, which is used as the maximum optimization objective of the genetic algorithm.

[0118] S236, four-level selection function, assuming the population size is 40, sort by size, the first ten are directly retained as the first level, and the rest are divided into three levels. Two copies of the second level are retained, one copy of the third level is retained, and the fourth level is directly eliminated. Simulate natural elimination and optimize the population. Fig.12 Shown

[0119] S237, Adaptive Crossover Function Among them, F is the maximum fitness of the two crossover individuals, F mean is the average fitness of the entire population, n is the number of evolutionary iterations, n max is the maximum number of iterations, P jmax Set to 0.9, P jmin Set to 0.4. The maximum number of repetitions is GEN×M times. Fig.13 Shown

[0120] S238, adaptive variogram Among them, F is the maximum fitness of a crossover individual, F mean is the average fitness of the entire population, n is the number of evolutionary iterations, n max is the maximum number of iterations, P jmax Set to 0.1, P jmin Set to 0.006. The maximum number of repetitions is GEN×M times. Fig.14 Shown

[0121] S239, the optimal solution of the weight vector and the optimal solution of the threshold vector obtained by the improved genetic algorithm (IGA) are used as the initialization weight vector of the BP neural network and initialize the threshold vector

[0122] S2310, using BP neural network to process the data set composed of PCA features and defect distance of pulsed eddy current detection signal, and obtain the final optimal solution of weight vector and the final threshold vector Optimal solution and CV-IGA-BP model1.

[0123] S2311, use CV-IGA-BP model 1 to get the estimated value x of the blind hole distance pred_std ,like Fig.15 As shown, the evaluation value of the blind hole distance R1 is Z-Score standardized to facilitate subsequent processing.

[0124] S24, performing data fusion on the standardized blind hole distance evaluation value and the PCA feature data of the pulsed eddy current detection signal.

[0125] x pred_std In the form of column repeat completion, ensure that x pred_std and x pca The matrix shape remains consistent.

[0126] Because x pca The size is 469×23, and x pred_std The shape size is 469×1, and after processing x pred_std The shape size is 469×23.

[0127] x pred_std and x pca The linear weight sum between them is 1, let r be x pca The weight of x is 1-r. pred_std The weight of D = r × x pca +(1-r)×x pred_std Finally, it was found that the model had the highest accuracy when r = 0.4. Fig.10 As shown, the optimal hyperparameters of CV-IGA-BP model 2 are obtained:

[0128] Table 3

[0129] Hidden layer size hidden_layer_sizes (64,32) Activation Function activation 'tanh Learning Rate learning_rate_init 0.1 Maximum number of iterations max_iter 1000

[0130]

[0131] The fused data set D is shown in Figure x and will be used for training the CV-IGA-BP model 2.

[0132] S5 includes:

[0133] S51, taking the data set D as the input feature and the blind hole depth (H) as the target feature.

[0134] S52, as in the training steps of model 1 of CV-IGA-BP, train model 2 of CV-IGA-BP, and finally obtain the evaluation value x of the defect depth pred_stl At the same time I will use x pca The dataset trains a model2_pca for comparison with model 2 of CV-IGA-BP.

[0135] The evaluation indicators of the model of the present invention are MSE, R2, MAE, and RMSE, and the calculation formula is as follows:

[0136] y out_i is the actual output, Output for evaluation.

[0137] y out_i is the average evaluation output.

[0138]

[0139] Table 4

[0140] Model 1 of CV-IGA-BP MSE R2 MAE RMSE Corrosion blind hole distance 0.0002268699 0.999996789 0.0079663690 0.0150622017 Model 2 of CV-IGA-BP MSE R2 MAE RMSE Depth of Corrosion Blind Hole 0.0005698479 0.999803667 0.0130881762 0.0238714887 Model2_pca MSE R2 MAE RMSE Depth of Corrosion Blind Hole 0.0023015642 0.9992070325 0.0218214889 0.0479746206

[0141] According to the data in Table 4, Fig.16 , 17 (a) and Fig.16 , 17 By comparing (b), we can see that the effect of blind hole depth of 304 steel plate becomes very good under feature fusion, and the model also has better robustness, which shows that our method is of practical significance.

[0142] Second embodiment:

[0143] See also Figure 18-Figure 20 , PEC detection device for the position and depth of blind holes in 304 steel plates under the insulation layer, including:

[0144] A frame 1 for carrying components of the detection device;

[0145] The encoder track detection device 2 comprises a bracket 21, a coupling 22, a meter wheel 23 and an encoder 24. The bracket 21 is mounted on the frame 1. The encoder 24 is rotatably mounted on the bracket 21. The meter wheel 23 is connected to the shaft end of the encoder 24 through the coupling 22. Two sets of directional wheel walking devices 3 are mounted on both sides of the bottom of the frame 1.

[0146] A three-axis position display device 4, wherein the three-axis position display device 4 is installed on the top of the frame 1;

[0147] The regionalized pulsed eddy current testing device 5 comprises an X-axis mechanism 51 , a Y-axis mechanism 52 and a Z-axis mechanism 53 , and a detection part for detection is installed on the Y-axis mechanism 52 .

[0148] The X-axis mechanism 51, the Y-axis mechanism 52 and the Z-axis mechanism 53 all include a mounting plate 511, a first motor 512, a screw rod 513, a threaded slide 514, a slide rail 515 and a connecting frame 516. The first motor 512 is fixedly mounted on one side of the mounting plate 511, the screw rod 513 is mounted on the output shaft of the first motor 512, the threaded slide 514 is threadedly connected to the surface of the screw rod 513, and the threaded slide 514 is slidably connected to the mounting plate 511 through the slide rail 515.

[0149] The connecting frame 516 in the Y-axis mechanism 52 is installed on the connecting frame 516 in the X-axis mechanism 51 . The connecting frame 516 in the Z-axis mechanism 53 is installed on the mounting plate 511 in the Y-axis mechanism 52 .

[0150] The bracket 21 is fixed to the frame 1 by bolts, providing structural stability;

[0151] The encoder 24 is fixed in the round hole of the bracket 21, and the tightness can be adjusted by adjusting the bolt, thereby ensuring the accurate installation and adjustment of the encoder 24;

[0152] The coupling 22 connects the shaft of the encoder 24 and the shaft of the meter wheel 23 to ensure synchronous rotation between the two.

[0153] The threaded slide 514 is provided with a threaded hole adapted to the screw rod 513. This design enables the slide to move accurately along the screw rod 513;

[0154] The encoder trajectory detection device 2 calculates the trajectory of movement based on the function of the encoder 24 and the size of the meter wheel 23; the regionalized pulse eddy current detection device 5 performs accurate pulse eddy current metal defect detection through the movement of the X, Y, and Z axes, and the three-axis position display device 4 can display the movement of the X, Y, and Z axes in real time to observe the location of the defect; the directional wheel walking device 3 can enable the entire machine to move on the object to be tested.

[0155] The detection part is a pulsed eddy current detector, which stimulates a pulsed magnetic field by passing a pulse current into the excitation coil, generating pulsed eddy currents in the conductor specimen in the magnetic field. The magnetic field generated by the pulsed eddy currents induces a voltage signal on the detection coil that changes with time, thereby achieving the detection purpose.

[0156] The working principle of this PEC detection device is as follows:

[0157] Above the 304 steel plate, the PEC detection device is dragged by manpower, using the directional wheel walking device 3 as a walking mechanism, and can move in the X-axis or Y-axis direction on the steel plate.

[0158] When pulse eddy current detection is performed, the first motor 512 on the Z-axis mechanism 53 drives the screw rod 513 to rotate. The rotation of the screw rod 513 causes the threaded slide 514 to drive the detection part to move up and down along the Z-axis to find a suitable distance.

[0159] During the detection, the X-axis mechanism 51 and the Y-axis mechanism 52 will also move. The movement principle is the same as that of the Z-axis. The positions of the X, Y, and Z axes will be displayed on the three-axis position display device 4, so that the position of the defect can be determined by the positions of the X and Y axes.

[0160] The first motor 512 is preferably a stepper motor, which can improve the energy efficiency of the entire device by utilizing its control accuracy and high efficiency.

[0161] Compared with the related art, the PEC detection device for the position and depth of blind holes in 304 steel plates provided by the present invention has the following advantages:

[0162] Beneficial effects:

[0163] The book search encoder track detection device 2 transmits the real-time displacement and track to the host computer and displays it through images, which can accurately locate and thus determine the location of the blind hole;

[0164] The regionalized pulsed eddy current testing device 5 can perform regionalized testing through the movement of the X and Y axes, and does not need to disassemble the plate to be tested, so the operation is simpler and more convenient.

[0165] Third embodiment:

[0166] See also Figure 16-23 Based on the PEC detection device for the position and depth of blind holes in 304 steel plates provided in the second embodiment of the present application, the third embodiment proposes another PEC detection device. The third embodiment is only a preferred method of the second embodiment and will not affect the independent implementation of the second embodiment.

[0167] Specifically, the difference of the PEC detection device for the position and depth of blind holes in 304 steel plates provided in the third embodiment of the present application is that it also includes a second X-axis mechanism 6 and a second Y-axis mechanism 7, and the second X-axis mechanism 6 includes a first fixed plate 61, a first reciprocating screw 62 and a second threaded slide 63, the first reciprocating screw 62 is installed inside the first fixed plate 61, and the second threaded slide 64 is threadedly connected to the second reciprocating screw 62.

[0168] The second Y-axis mechanism includes a second fixed plate 71, a second reciprocating screw 72 and a third threaded slide 73. The second fixed plate 71 is installed on the third threaded slide 73. The second reciprocating screw 72 is fixed inside the second fixed plate 71. The third threaded slide 73 is threadedly connected to the second reciprocating screw 72, and the third threaded slide 73 is slidably connected to the second fixed plate 71.

[0169] The threaded slide can reciprocate along the surface of the reciprocating screw.

[0170] A toothed roller 8 is installed on the lower side of the first fixed plate 61, and one end of the toothed roller 8 is connected to the first reciprocating screw rod 62 through a one-way transmission member 12. The one-way transmission member 12 includes two transmission wheels 121, a transmission belt 122, two one-way bearings 123 and two driving wheels 124. The two one-way bearings 123 are installed on the transmission belt 122, and the two driving wheels 124 are respectively installed on the two one-way bearings 123. The two transmission wheels 121 are respectively installed on one end of the first reciprocating screw rod 62 and one end of the toothed roller 8, and the two driving wheels 124 are respectively connected to the two transmission wheels 121 through the transmission belt 122.

[0171] The driving wheel 124, the transmission wheel 121 and the transmission belt 122 may be matched in a manner such as a gear and a toothed belt or a pulley and a belt.

[0172] A sliding member 11 is installed on the lower side of the first fixed plate 61 and on one side of the gear roller 8. The sliding member 11 includes a sliding arm 111 and a sliding frame 112. The lower end of the sliding frame 112 is sleeved on the surface of the sliding arm 111. The sliding frame 112 is fixedly connected to the second threaded slide 63 through a connecting arm 14. A driving member 9 is installed inside the sliding frame 112.

[0173] The cross section of the sliding arm 111 is rectangular.

[0174] The driving member 9 includes a rotating shaft 91 , a gear 92 , a main bevel gear 93 and a transmission mechanism 94 . The rotating shaft 91 is rotatably connected to the sliding frame 112 . The main bevel gear 93 and the gear 92 are connected to the rotating shaft 91 . The gear 92 is meshed with the gear roller 8 .

[0175] The toothed roller 8 is provided with a strip-shaped tooth groove, and the gear 92 is meshed with the tooth groove on the toothed roller 8 .

[0176] One end of the second reciprocating screw rod 72 is connected to a slave bevel gear 13 , and the slave bevel gear 13 is connected to the main bevel gear 93 via a transmission wheel of the transmission mechanism 94 .

[0177] The transmission mechanism 94 includes a transmission shaft and transmission bevel gears located at both ends of the transmission shaft, and the two transmission bevel gears are respectively meshed with the main bevel gear 93 and the slave bevel gear 13 .

[0178] In the third embodiment, both the X-axis mechanism 51 and the Y-axis mechanism 52 use one first motor 512 .

[0179] The fourth embodiment is a preferred embodiment of the third embodiment. In this embodiment, a motor is provided to control the X-axis mechanism 51 and the Y-axis mechanism 52 to be implemented separately through the forward and reverse rotation of the motor output shaft, which is more energy-saving and environmentally friendly.

[0180] The working principle of the PEC detection device in the third embodiment and the fourth embodiment is as follows:

[0181] When the second X-axis mechanism 6 needs to be operated, the second motor 15 is controlled to rotate clockwise. At this time, the one-way bearing 123 located on the inner side of the transmission belt 122 drives the driving wheel 124 located on the inner side to rotate, and drives the transmission wheel 121 located on the first reciprocating screw 62 to rotate, thereby driving the first reciprocating screw 62 to rotate, thereby driving the second threaded slide 63 to follow the movement, and the second threaded slide 63 can drive the second Y-axis mechanism 6 to follow the movement, thereby driving the detection part located on the Z-axis mechanism 53 to move. At this time, the one-way bearing 123 located on the outer side of the transmission belt 122 will not be displaced, and the transmission wheel 121 on the gear roller 8 rotates;

[0182] When it is necessary to drive the Y-axis mechanism 52 to move, the second motor 15 rotates in the opposite direction, that is, counterclockwise. At this time, the one-way bearing 123 located on the outer side of the transmission belt 122 drives the driving wheel 124 located on the outer side to rotate, and drives the transmission wheel 121 located on the first reciprocating screw 62 to rotate, thereby driving the transmission wheel 121 located on the gear roller 8 to rotate, and at this time drives the gear roller 8 to rotate, and the gear roller 8 rotates, driving the gear 92 to rotate, and at this time, the slave bevel gear 13 located on the second reciprocating screw 72 is driven to rotate through the transmission mechanism 94, and the second reciprocating screw 72 can be driven to rotate;

[0183] Similarly, at this time, the first reciprocating screw rod 62 will not rotate;

[0184] When the first reciprocating screw 62 rotates, driving the second threaded slide 63 to rotate, the connecting arm 14 can drive the sliding frame 112 to follow the movement along the sliding arm 111, so that the driving member 9 and the Y-axis mechanism 52 can be kept relatively still, and the teeth of the gear 92 can move along the tooth grooves on the gear roller 8.

[0185] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the enlightenment of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, which all belong to the protection scope of the present invention.

Claims

1. A method for detecting the position and depth of blind holes in 304 steel plates under an insulating layer, characterized in that: The following steps are involved: S1: Place the PEC probe above the origin of the current inspection area of ​​the 304 steel plate with an insulating layer, perform PEC inspection along the positive direction of the X-axis, and determine whether a blind hole is detected; S2: If a blind hole is detected, the distance R1 between the blind hole and the center axis of the probe and the blind hole depth H1 are recorded, and the current position of the PEC probe is recorded as detection point 1; S3: Move the PEC probe along the positive direction of the X-axis by R1 / 2, record the current position of the PEC probe as detection point 2, perform PEC detection at detection point 2, and obtain the distance R2 between the blind hole and the probe and the blind hole depth H2; S4: Return the PEC probe to the detection point 1, move R1 / 2 along the positive direction of the Y axis, and perform PEC detection; S5: If a blind hole is detected, the distance R3 between the blind hole and the central axis of the probe and the depth H3 of the blind hole are recorded, and the current position of the PEC probe is recorded as detection point 3; S6: If no blind hole is detected, the PEC probe is returned to the detection point 1, and moved R1 / 2 along the negative direction of the Y axis, and the PEC detection is performed to obtain the distance R3 between the blind hole and the central axis of the probe and the depth D3 of the blind hole, and the current position of the PEC probe is recorded as the detection point 3; S7: From this, the coordinates of the blind hole relative to the center axis of the probe are obtained as follows: The blind hole depth is S8: If no blind hole is detected and the edge of the probe detection range has reached the edge of the positive direction of the X-axis of the area to be inspected, the detection in the current area to be inspected is completed.

2. The method for detecting the position and depth of blind holes of 304 steel plate under the insulating layer according to claim 1 is characterized in that: S2 includes: S21, using a device to obtain full-wave time series PEC signal data of a 304 steel plate, and obtaining a data set X containing a PEC detection signal; S22, perform Z-Score standardization and PCA dimension reduction on the data set X to obtain data x containing PCA features of PEC detection signals pca ; S23, use x pca , the distance R1 between the blind hole and the PEC probe Using CV-IGA-BP model Train to obtain CV-IGA-BP model 1 and obtain the evaluation value x of the blind hole distance pred_std , perform Z-Score standardization on the evaluation value of blind hole distance; S24, for the standardized x pred_std PCA feature dataset x of PEC detection signal pca Fusion is performed to obtain data set D; S25, training the data set D with the CV-IGA-BP model to obtain the CV-IGA-BP model 2, and obtaining the evaluation value x of the blind hole depth pred_stl .

3. The method for detecting the position and depth of blind holes of 304 steel plate under the insulating layer according to claim 2, characterized in that: S1 includes: S11, by processing the simulation data, a PEC detection signal of the blind hole position of the 304 steel plate with a signal-to-noise ratio of 60 dB is obtained, and the actual value of the target feature blind hole distance (R1) and the blind hole depth (H1) are obtained as a training data set; S12, use the device to obtain the PEC signal of 304 steel plate in the experimental environment as the test data set.

4. The method for detecting the position and depth of blind holes of 304 steel plate under the insulating layer according to claim 2, characterized in that: S21 includes: S211, full-wave time series characteristics of pulsed eddy current detection signal in training data x = [x 11 ..x ij ]i, j = 1...m, in order to prevent the gradient from disappearing and exploding during data training: where μ j is the mean of the corresponding column feature, m is the total number of each feature, x ij is the jth data of the i-th feature, σ j is the standard deviation of the corresponding column feature, y ij is the jth data of the newly generated i-th feature; S212, performing PCA dimensionality reduction processing on the standardized PEC detection signal to obtain a PCA feature data set x of the pulsed eddy current detection signal pca .

5. According to the method for detecting the position and depth of blind holes of 304 steel plate under the insulating layer of claim 4, S23 comprises: S231, assuming that the pre-processed pulsed eddy current detection signal data set is x pca , the target feature is the blind hole distance, the genetic algorithm population size is M, and the number of evolutions is GEN; S232 includes: S2321, build BP neural network in is the net input to the hidden layer, is the weight on the connection starting from input unit 1, is the bias term, h′ refers to the quantity on the hidden layer; Among them I j is the output of the hidden layer and, is the activation function, x pca is the input feature, and the blind hole distance is the target feature; S2322, randomly split the entire dataset into 5 different subsets, each subset is called a fold, and then train and evaluate the model 5 times, each time selecting one fold as the test set and the other 4 folds as the training set; S2323, randomly divide 20% of the training set as a validation set for early stopping of training; S2324, check whether there is overfitting during the training process; S233, the weight and threshold range of the BP neural network is [1, 1], so real number (floating point number) encoding is used; the weight vector and threshold vector in the optimal hyperparameters have been obtained in S32, and initialized: num = len(W) + len(b), where W is the weight matrix of the BP neural network, b is the bias matrix of the BP neural network, and num is the total number of parameters of the BP neural network calculated: S234, varbound = [(-1, 1)] × num, varbound is the parameter vector of weights and thresholds for BP neural network initialization, and also serves as the genetic algorithm individual genome; S235, taking the inverse of the mean square error of the BP neural network as the fitness value of the improved genetic algorithm (IGA); F i =1 / E i , E i is the mean square error of the individual in the ith population, F i is the fitness value of the i-th population individual, which is the maximum optimization target of the genetic algorithm; S236, four-level selection function, assuming the population size is 40, sort by size, the first ten are directly retained as the first level, the rest are divided into three levels, the second level is copied twice and retained, the third level is copied once and retained, and the fourth level is directly eliminated, simulating natural elimination and optimizing the population; S237, Adaptive Crossover Function Among them, F is the maximum fitness of the two individuals to be crossed, F mean is the average fitness of the entire population, n is the number of evolutionary iterations, n max is the maximum number of iterations, P jmax Set to 0.9, P jmin Set to 0.

4. Repeat step GEN*M times; S238, adaptive variogram Among them, F is the maximum fitness among the individuals to be crossed, F mean is the average fitness of the entire population, n is the number of evolutionary iterations, n max is the maximum number of iterations, P jmax Set to 0.1, P jmin Set it to 0.006 and repeat step GEN*M times; S239, the optimal solution of the weight vector and the optimal solution of the threshold vector obtained by the improved genetic algorithm (IGA) are used as the initialization weight vector of the BP neural network and initialize the threshold vector S2310, using BP neural network to process the data set consisting of PCA features of pulsed eddy current detection signals and blind hole distances, and obtain the final optimal solution of weight vector and the final threshold vector Optimal solution and CV-IGA-BP model 1; S231 1, use CV-IGA-BP model 1 to obtain the evaluation value of the blind hole distance, and perform Z-Scor on the evaluation value of the blind hole distance e Standardization facilitates subsequent processing.

6. According to the method for detecting the position and depth of blind holes of 304 steel plate under the insulating layer of claim 2 or 4, S24 comprises: S241, let the normalized blind hole distance evaluation value be x pred_std , the PCA feature data set of the pulsed eddy current detection signal is x pca : S242, x pred_std In the form of column repeat completion, ensure that x pred_std and x pca The matrix shape remains consistent; S243,x pred_std and x pca The linear weight sum between them is 1, let r be x pca The weight of x is 1-r. pred_std The weight of D = r × x pca +(1-r)×x pred_std , D is the training data for CV-IGA-BP model 2.

7. According to the method for detecting the position and depth of blind holes of 304 steel plate under the insulating layer of claim 2, 5 or 6, S5 comprises: S51, taking the data set D as the input feature and the blind hole depth as the target feature; S52, as in the training step of the model 1 of CV-IGA-BP in S23, the data D is trained with CV-IGA-BP to obtain the CV-IGA-BP model 2, and finally the evaluation value of the blind hole depth is obtained.

8. A device for detecting the position and depth of blind holes in 304 steel plates under the insulating layer, characterized in that: include: A rack for carrying components of the testing device; An encoder trajectory detection device, the encoder trajectory detection device comprises a bracket, a coupling, a meter wheel and an encoder, the bracket is mounted on the frame, the encoder is rotatably mounted on the bracket, the meter wheel is connected to the shaft end of the encoder through a coupling, two sets of directional wheel walking devices are mounted on both sides of the bottom of the frame, and the three-axis position display device is mounted on the top of the frame; A regionalized pulsed eddy current testing device comprises an X-axis mechanism, a Y-axis mechanism and a Z-axis mechanism, and the detection unit is installed on the Y-axis mechanism.

9. The device for detecting the position and depth of blind holes of 304 steel plate under the insulating layer according to claim 8, characterized in that: The X-axis mechanism, the Y-axis mechanism and the Z-axis mechanism all include a mounting plate, a first motor, a screw, a threaded slide, a slide rail and a connecting frame. The first motor is fixedly mounted on one side of the mounting plate, the screw is mounted on the output shaft of the first motor, the threaded slide is threadedly connected to the surface of the screw, and the threaded slide is slidably connected to the mounting plate via a slide rail; the connecting frame in the Y-axis mechanism is mounted on the connecting frame in the X-axis mechanism, and the connecting frame in the Z-axis mechanism is mounted on the mounting plate in the Y-axis mechanism.

10. The device for detecting the position and depth of blind holes of 304 steel plate under the insulating layer according to claim 9, characterized in that: It also includes a second X-axis mechanism and a second Y-axis mechanism, the second X-axis mechanism includes a first fixed plate, a first reciprocating screw and a second threaded slide, the first reciprocating screw is installed inside the first fixed plate, and the second threaded slide is threadedly connected to the second reciprocating screw; the second Y-axis mechanism includes a second fixed plate, a second reciprocating screw and a third threaded slide, the second fixed plate is installed on the third threaded slide, the second reciprocating screw is fixed inside the second fixed plate, the third threaded slide is threadedly connected to the second reciprocating screw, and the third threaded slide is slidably connected to the second fixed plate.

11. The device for detecting the position and depth of blind holes of 304 steel plate under the insulating layer according to claim 10, characterized in that: A toothed roller is installed on the lower side of the first fixed plate, one end of the toothed roller is connected to the first reciprocating screw rod through a one-way transmission member, the one-way transmission member includes two transmission wheels, two one-way bearings and two driving wheels, the two one-way bearings are installed on the above, the two driving wheels are respectively installed on the two one-way bearings, the two transmission wheels are respectively installed on one end of the first reciprocating screw rod and one end of the toothed roller, and the two driving wheels are respectively connected to the two transmission wheels through transmission belts; a sliding member is installed on the lower side of the first fixed plate and on one side of the toothed roller, the sliding member includes a sliding arm and a sliding frame, the lower end of the sliding frame is sleeved on the surface of the sliding arm, the sliding frame is fixedly connected to the second threaded slide through a connecting arm, and a driving member is installed inside the sliding frame.

12. The device for detecting the position and depth of blind holes of 304 steel plate under the insulating layer according to claim 11, characterized in that: The driving member includes a rotating shaft, a gear, a main bevel gear and a transmission mechanism. The rotating shaft is rotatably connected to the sliding frame. The main bevel gear and the gear are fixedly connected to the rotating shaft, and the gear is meshed with the gear roller; one end of the second reciprocating screw is connected to a slave bevel gear, and the slave bevel gear is connected to the main bevel gear through a transmission wheel of the transmission mechanism.

Images