In-board damage TFM positioning and detection method based on multiple networks

The multi-network based TFM positioning and detection method addresses inefficiencies in existing metal plate damage assessment by using CNN-BiLSTM and ResNet networks to analyze Lamb wave signals, achieving precise and efficient damage identification and quantification.

CN119064451APending Publication Date: 2024-12-03NANJING TECH UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410613554.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-11-04
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

Existing methods for assessing metal plate damage, such as in bridges, buildings, ships, and wind turbines, are inefficient and inaccurate, posing safety risks due to their inability to effectively identify and quantify damage areas and sizes.

Method used

A multi-network based TFM positioning and detection method using CNN-BiLSTM and ResNet networks to analyze Lamb wave signals from sensor arrays, combining traditional Lamb wave technology with deep learning to identify damage areas and sizes in metal plates.

Benefits of technology

Improves the efficiency and accuracy of damage location and size assessment by using data-driven deep learning, enhancing the stability and precision under various conditions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119064451A_ABST
    Figure CN119064451A_ABST
Patent Text Reader

Abstract

The invention discloses an in-board damage TFM positioning and detection method based on multiple networks, and relates to the technical field of material detection. The method comprises the following steps: firstly, selecting a proper mode as a Lamb wave excitation signal and modulating; emitting an excitation signal in a target area in the metal plate, receiving an echo signal of the target area through a sensor array, and taking sensor data as an area identification original data set to train a CNN-BiLSTM network model to carry out circular hole damage identification; according to the method, the traditional Lamb wave technology and the deep learning technology are combined, so that the area identification of the damage and the identification of the damage radius can be realized at the same time; the defects and limitation of the existing method on corresponding aspects are overcome; compared with a traditional Lamb wave analysis method, the method has the advantages that the in-plate damage positioning efficiency and the size evaluation capability are improved, and stability is achieved for various operation and environmental conditions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of material detection, and more specifically to a multi-network based TFM positioning and detection method for plate damage. Background Art

[0002] CNN-BiLSTM stands for the combination of convolutional neural network and bidirectional long-term memory network. Convolutional neural network is an artificial neural network specially used to process data with grid structure. Bidirectional long-term memory network is an extended form of long short-term memory network, which is a neural network model used to process sequence data. It has hidden states in both forward and backward directions, and can capture the timing information and contextual relationships in sequence data. Metal plate materials are widely used in bridges, buildings, ships, aircraft, wind turbines and pipelines due to their high strength, light weight and corrosion resistance. However, when these materials are damaged by impact, wear and corrosion due to long-term use, it will seriously affect the safety of the structure. Therefore, efficient and accurate assessment of metal plate damage is crucial.

[0003] In view of the above problems, the present invention proposes a solution. Summary of the invention

[0004] In order to overcome the above-mentioned defects of the prior art, the present invention provides a multi-network-based TFM positioning and detection method for plate damage to solve the problems existing in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] The multi-network based TFM positioning and detection method for plate damage includes the following steps:

[0007] Step SO1: Obtain the Lamb wave dispersion curve of the metal plate, select a suitable mode as the Lamb wave excitation signal and modulate the selected Lamb wave excitation signal;

[0008] Step SO2: setting a target area in the metal plate, and placing an excitation sensor at the target area, and symmetrically transmitting the Lamb wave excitation signal described in step SO1 on the upper and lower surfaces of the metal plate at the excitation sensor;

[0009] Step SO3: a group of circular sensor arrays are arranged at designated positions on the surface of the metal plate to receive echo signals from the target area except for the excitation sensor;

[0010] Step SO4: changing the distribution of circular holes in the circular sensor array, repeating the excitation in step SO2 and step SO3, and using each sensor data obtained as the original data set for region recognition;

[0011] Step SO5: dividing the region recognition original data set obtained in step SO4 into a region recognition training data set, a region recognition test data set and a region recognition verification data set;

[0012] Step SO6: Use the training data set obtained in step SO5 to train the parameters of the CNN-BiLSTM network model, and use the region recognition verification data set for verification to select the model with the best performance;

[0013] Step SO7: using the model with the best performance obtained in step SO6 to identify the area to which the circular hole damage in the metal plate belongs to the area recognition test data set in step SO5;

[0014] Step SO8: using the echo signal collected by the circular sensor array arranged in step SO3, the area of ​​the circular hole damage in the metal plate identified in step SO7 is imaged and located using the TFM algorithm;

[0015] Step SO9: by changing the circular hole distribution in the circular sensor array, repeating the step SO8, using the obtained TFM imaging image as the radius recognition original data set, and then dividing the radius recognition original data set into a radius recognition training data set, a radius recognition test data set, and a radius recognition verification data set;

[0016] Step S10: Use the radius recognition training data set obtained in the step SO9 to train the ResNet network model parameters, and use the radius recognition verification data set for verification, and select the model with the best performance. Use the obtained model with the best performance to identify the radius of circular hole damage in the plate using the radius recognition test data set in the step SO9 to verify the effectiveness of the algorithm model.

[0017] Preferably, the step of selecting a suitable mode as the Lamb wave excitation signal is to select a Lamb wave S0 mode signal with low dispersion and high speed as the excitation according to the Lamb wave dispersion curve of the metal plate.

[0018] Preferably, the modulation formula for the selected Lamb wave excitation signal is: Where H(t) is a step function, f c It represents the center frequency of the excitation signal, φ represents the phase of the signal, and N represents the number of peaks of the modulated sinusoidal signal.

[0019] Preferably, the step of training the CNN-BiLSTM network model parameters using the training data set obtained in step S06 is:

[0020] The region recognition dataset is divided into a region recognition training set, a region recognition validation set, and a region recognition test set. The region recognition training set is used for training the CNN-BiLSTM network, the region recognition validation set is used for optimizing and selecting model parameters, and the region recognition test set is used to evaluate the performance of the model.

[0021] The CNN-BiLSTM network model is trained using the stochastic gradient descent algorithm to initialize the network parameters, set the initial learning rate, and update the network parameters through the back propagation algorithm during the training process to minimize the loss function.

[0022] After each round of training, the model performance is evaluated using the region recognition validation set, and the learning rate and other parameters are adjusted based on the validation results to improve the generalization ability and accuracy of the model;

[0023] Repeat the above two steps until the early stopping condition is met or the preset maximum number of training rounds is reached, and finally select the model with the best performance on the validation set as the final model. The model training optimization formula is: Where θ represents the model parameter, α represents the learning rate, and L(θ) represents the loss function. It is expressed as the gradient of the loss function with respect to the model parameters.

[0024] Preferably, the step of imaging and locating the area to which the circular hole damage in the metal plate identified in step SO7 belongs using the TFM algorithm is:

[0025] The detection area is divided into grid points and each point is regarded as a focused target. The image is reconstructed using the full matrix echo signal captured by the phased array, the signal strength of each point is calculated, and a high-definition focused image is generated;

[0026] Set the array point (x i ,x i ) is the excitation point, and the corresponding other array point (x j ,x j ) The received induction signal is used by U ij (x i ,x j ,t) indicates that the intensity expression of the comprehensive signal obtained by superimposing the delayed signals corresponding to all excitation points and receiving points is: In this process, the signal is first emitted from the excitation source, reflected by the damage point, and finally received by the array sensor. The total duration of the whole process can be expressed as where r ing represents the displacement vector from the excitation point to the damage point, and r back Then it is represented as the displacement vector from the damage point to the receiving point, and C is represented as the speed of the sound wave propagating in the plate;

[0027] An imaging map is drawn according to the I(P) value of each focal point P, the data in the map is normalized, and the point with the maximum amplitude in the local area of ​​the target area in the imaging map is taken as the damage location result.

[0028] Preferably, the step S10 of training the ResNet network model parameters is as follows:

[0029] The radius recognition dataset is divided into a radius recognition training set, a radius recognition validation set, and a radius recognition test set. The radius recognition training set is used for model training, the radius recognition validation set is used for adjusting and optimizing model parameters, and the radius recognition test set is used to evaluate the accuracy and generalization ability of the model.

[0030] The stochastic gradient descent algorithm is used to train the model, the cross entropy loss function is used as the optimization target, and the early stopping mechanism is used to prevent overfitting.

[0031] Beneficial effects of the present invention:

[0032] The method of the present invention combines the traditional Lamb wave technology with the deep learning technology to simultaneously realize the identification of the damaged area and the damaged radius, thus making up for the shortcomings and limitations of the existing methods in the corresponding aspects.

[0033] The method of the present invention improves the efficiency of plate damage location and size assessment capability by using a data-driven deep learning method compared to the traditional Lamb wave analysis method;

[0034] The method of the present invention realizes the positioning and size evaluation of the damage in the plate by arranging a group of sensor arrays, improves the accuracy of positioning and evaluating different damages, and has stability under various operating and environmental conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is the overall flow chart of the present invention.

[0036] Figure 2 The present invention order allows modification of the index calculation flow chart.

[0037] Figure 3 Schematic diagram of the TFM algorithm of the method of the present invention.

[0038] Figure 4 This is a schematic diagram of the damaged area classification work of the method of the present invention.

[0039] Figure 5 Schematic diagram of the aluminum plate model of the method of the present invention.

[0040] Figure 6 It is the time domain diagram and normalized spectrum diagram of the excitation signal of the method of the present invention.

[0041] Figure 7This is the LSTM network structure diagram of the method of the present invention.

[0042] Figure 8 This is a structural diagram of the BiLSTM method of the present invention.

[0043] Fig. 9 This is the overall workflow diagram of CNN-BiLSTM of the method of the present invention.

[0044] Fig.10 This is the CNN-BiLSTM network TFM damage localization imaging image of the method of the present invention.

[0045] Fig.11 This is a network structure diagram of damage size classification detection using the method of the present invention.

[0046] Fig.12 This is a dispersion curve diagram of the method of the present invention.

[0047] Fig.13 This is the loss function and accuracy diagram of the ResNet network training of the method of the present invention. DETAILED DESCRIPTION

[0048] The technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the present invention. In addition, the forms of the various structures recorded in the following embodiments are merely illustrative. The multi-network-based TFM positioning and detection method for plate damage involved in the present invention is not limited to the various structures recorded in the following embodiments. All other implementations obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0049] The present invention provides a multi-network based TFM positioning and detection method for plate damage, such as Figure 1 As shown, the following steps are included:

[0050] Step SO1: Obtain the Lamb wave dispersion curve of the metal plate, select a suitable mode as the Lamb wave excitation signal and modulate the selected Lamb wave excitation signal;

[0051] Step SO2: setting a target area in the metal plate, and placing an excitation sensor at the target area, and symmetrically transmitting the Lamb wave excitation signal described in step SO1 on the upper and lower surfaces of the metal plate at the excitation sensor;

[0052] Step SO3: a group of circular sensor arrays are arranged at designated positions on the surface of the metal plate to receive echo signals from the target area except for the excitation sensor;

[0053] Step SO4: changing the distribution of circular holes in the circular sensor array, repeating the excitation in step SO2 and step SO3, and using each sensor data obtained as the original data set for region recognition;

[0054] Step SO5: dividing the region recognition original data set obtained in step SO4 into a region recognition training data set, a region recognition test data set and a region recognition verification data set;

[0055] Step SO6: Use the training data set obtained in step SO5 to train the parameters of the CNN-BiLSTM network model, and use the region recognition verification data set for verification to select the model with the best performance;

[0056] Step SO7: using the model with the best performance obtained in step SO6 to identify the area to which the circular hole damage in the metal plate belongs to the area recognition test data set in step SO5;

[0057] Step SO8: using the echo signal collected by the circular sensor array arranged in step SO3, the area of ​​the circular hole damage in the metal plate identified in step SO7 is imaged and located using the TFM algorithm;

[0058] Step SO9: by changing the circular hole distribution in the circular sensor array, repeating the step SO8, using the obtained TFM imaging image as the radius recognition original data set, and then dividing the radius recognition original data set into a radius recognition training data set, a radius recognition test data set, and a radius recognition verification data set;

[0059] Step S10: Use the radius recognition training data set obtained in the step SO9 to train the ResNet network model parameters, and use the radius recognition verification data set for verification, and select the model with the best performance. Use the obtained model with the best performance to identify the radius of circular hole damage in the plate using the radius recognition test data set in the step SO9 to verify the effectiveness of the algorithm model.

[0060] The present invention uses finite element analysis software ABAQUS to establish a three-dimensional aluminum plate simulation model, and uses display dynamic analysis to solve and analyze the propagation process of Lamb waves when there is damage in the three-dimensional aluminum plate. The parameters of the aluminum plate model created in the simulation are shown in Table 1:

[0061]

[0062] Table 1

[0063] The metal plate is set with a 60×60 detection area. In this study, the damage detection area is subdivided into nine sub-areas, and the damage setting of each sub-area follows a certain rule. Specifically, the damage is simulated by setting circular holes with diameters of 1mm, 1.5mm and 2mm in a specific area in sequence. The center position of each circular hole is precisely defined. The layout of area one takes the point (-27,27)mm as the starting position, and the circular holes are set in sequence at a distance of 1mm along the x-axis until the position of the 15th circular hole (-13,27)mm. Subsequently, the layout of the 16th circular hole is moved down 1mm in the y-axis direction to (-27,26)mm, and the circular holes are set at the same x-axis interval. This sequence continues until the layout of the area is completed, and a total of 100 circular holes are set. The layout of areas two to nine follows the same setting pattern as area one, but the starting positions are different. The coordinates of the center of the specific starting circular holes are: Region 2 (-7, 27) mm, Region 3 (13, 27) mm, Region 4 (-27, 7) mm, Region 5 (-7, 7) mm, Region 6 (13, 7) mm, Region 7 (-27, -13) mm, Region 8 (-7, -13) mm, and Region 9 (13, -13) mm. The arrangement of the circular holes in each region along the x-axis and y-axis adopts the same strategy as Region 1, and during the setting process, the size (radius) of the circular holes will be changed according to the experimental requirements to accurately simulate the effect of damage of different sizes on material properties. The use of this method can systematically study the effect of damage in different regions on material properties, and provide a standardized description method for damage distribution, which is convenient for comparison and analysis in subsequent studies. The boundary condition of the metal plate is a free boundary, that is, there is no constraint or displacement. A 1×7 phased array sensor array is set below the area to be detected, with a sensor spacing of 5 mm and a distance of 20 mm from the damage detection area, such as Figure 5 As shown in Figure 2, the sensor array can be used as an excitation source and a receiver at the same time to generate and receive Lamb wave signals.

[0064] In this embodiment, it should be specifically explained that Figure 6As shown, the step of selecting a suitable mode as the Lamb wave excitation signal is to select a low-dispersion, fast-speed Lamb wave S0 mode signal as excitation according to the Lamb wave dispersion curve of the metal plate. The dispersion curve of the Lamb wave describes the change of the propagation speed of the Lamb wave with the frequency under given material and geometric conditions. Usually, a lower frequency in the dispersion curve corresponds to a faster propagation speed, and a higher frequency corresponds to a slower propagation speed. The S0 mode signal is a basic mode in the Lamb wave, and usually has a lower dispersion rate and a faster propagation speed. Since the frequency of the S0 mode is low, it has a faster propagation speed in the metal plate and has better penetration performance, and can better penetrate the target object and propagate to a distant place.

[0065] In this embodiment, it should be specifically explained that the modulation of the selected Lqmb wave excitation signal uses a five-cycle 200KHz Hanning windowed sine wave signal as the excitation source, and its formula is: Where H(t) is a step function, f c =200kHz is the center frequency of the excitation signal, φ represents the phase of the signal, and N=5 represents the number of peaks of the modulated sinusoidal signal. Figure 6 As shown, the time domain waveform and normalized spectrum of the excitation signal are displayed.

[0066] At the same point of the aluminum plate, the upper and lower surfaces are respectively subjected to reverse symmetric excitation to generate a single S0 mode Lamb wave. The center of the aluminum plate is taken as the origin of the coordinate system, as shown in Figure 5 As shown, a linear sensor array consisting of seven points is set on the upper surface of the board, distributed on the left and right sides of the (0, -50) mm position. These sensors are used to excite and receive signals in sequence. The interval between them is 5 mm, and they are marked as S1, S2, ..., S7 from left to right. The position coordinates of sensor S1 are (-15, -50) mm. When any of the seven sensors is used as an excitation source, we place another sensor on the lower surface of its corresponding position to jointly apply an inversely symmetric Lamb wave excitation signal. At this time, the remaining six sensors on the upper surface are used as receivers to receive the reflected signals from the target area and use them as the original data set.

[0067] Since the present invention uses a single-frequency excitation source, a suitable frequency can be selected according to the dispersion curve to excite a specific Lamb wave mode. The present invention selects 200kHz as the center frequency of the excitation to reduce the influence of multi-mode and dispersion phenomena in Lamb wave propagation, such as Fig.12As shown in the Lamb wave dispersion curve, the S0 mode with smaller dispersion and faster propagation speed is selected as the excitation signal. This choice helps to alleviate the impact of multi-mode and dispersion on signal analysis.

[0068] By changing the location and size of the defect, data will be collected from sensors other than the excitation location, and the corresponding lossless data will be subtracted from these data to obtain the scattered signal, and then the corresponding defect area will be marked, accumulating a total of 37,800 sets of data. These data are assigned to the neural network for in-depth experimental analysis, of which the training set contains 70% of the data, a total of 26,460 sets of data; 20% is assigned to the cross-validation set, a total of 7,560 sets of data; the remaining 10% is used as a test set, a total of 3,780 sets of data.

[0069] In this embodiment, it should be specifically explained that Figure 2 As shown, the steps of training the CNN-BiLSTM network model parameters using the training data set obtained in step SO5 are:

[0070] In the model training stage, the region recognition dataset is divided into a region recognition training set, a region recognition validation set, and a region recognition test set in a ratio of 7:2:1. The region recognition training set is used for training the CNN-BiLSTM network, the region recognition validation set is used for optimizing and selecting model parameters, and the region recognition test set is used to evaluate the performance of the model.

[0071] The CNN-BiLSTM network model is trained by using a stochastic gradient descent algorithm, the network parameters are initialized, the initial learning rate is set, and during the training process, the network parameters are updated by a back propagation algorithm to minimize the loss function, wherein the stochastic gradient descent is a commonly used optimization algorithm for training machine learning models, including neural network models, and the goal of the gradient descent algorithm is to minimize the loss function so as to minimize the error between the predicted value and the actual value by adjusting the model parameters, and the stochastic gradient descent is a commonly used optimization algorithm for training machine learning models, including neural network models, and it is an iterative optimization algorithm that realizes the training of the model by continuously updating the model parameters to minimize the loss function;

[0072] After each round of training, the validation set is used to evaluate the model performance, and the learning rate and other parameters are adjusted according to the validation results to improve the generalization ability and accuracy of the model;

[0073] Repeat the above two steps until the early stopping condition is met or the preset maximum number of training rounds is reached, and finally select the model with the best performance on the validation set as the final model. The model training optimization method is: Where θ represents the model parameter, α represents the learning rate, and L(θ) represents the loss function. It is expressed as the gradient of the loss function with respect to the model parameters.

[0074] In this embodiment, it should be specifically explained that the loss function formula is: Where N represents the number of samples in the region recognition dataset, y i Represents the deed label of sample i, Represents the probability predicted by the model.

[0075] In the model prediction stage, the test data is extracted from the regional identification test data set and sent to the selected optimal model. At this time, the internal parameters of the network are already the optimal parameters. The model will preprocess the test data and extract the damage distribution characteristic information for different regions to realize the identification of the region to which the circular hole damage belongs.

[0076] In this embodiment, it should be specifically explained that the step of imaging and locating the area to which the circular hole damage in the metal plate identified in step SO7 belongs using the TFM algorithm is as follows:

[0077] By subdividing the detection area into grid points and treating each point as a focus target, the image is reconstructed using the full matrix echo signal captured by the phased array, thereby calculating the signal strength of each point and generating a high-definition focus image. The schematic diagram of the algorithm for reconstructing the image using the full matrix echo signal captured by the phased array is shown in FIG. Figure 3 As shown;

[0078] Set the array point (x i ,x i ) is the excitation point, and the corresponding other array point (x j ,x j ) The received induction signal can be used ij (x i ,x j ,t) indicates that the delayed signals corresponding to all excitation points and receiving points are superimposed, and the intensity expression of the comprehensive signal can be obtained as follows: In this process, the signal is first emitted from the excitation source, reflected by the damage point, and finally received by the array sensor. The total duration of the whole process can be expressed as where r ing represents the displacement vector from the excitation point to the damage point, and r back Then it is represented as the displacement vector from the damage point to the receiving point, and C is represented as the speed of the sound wave propagating in the plate;

[0079] The TFM algorithm is used to locate and image the damage area identified by the trained CNN-BiLSTM network model.

[0080] For the CNN-LSTM hybrid model, the CNN layer extracts signal features from the received signal. Next, these extracted features are integrated into a two-dimensional array so that they can be passed as input to the LSTM layer to parse the feature information in the received signal time series. LSTM effectively regulates the flow of information by introducing three unique gating mechanisms - forget gate, input gate and output gate, thereby improving the network's ability to handle long-term dependencies. The specific architecture of the LSTM model is as follows: Figure 7 As shown. In the LSTM architecture, the input at each time step is represented as. Core component C t It plays a crucial role in LSTM and is responsible for maintaining state information across time steps. The dynamics inside an LSTM cell can be described by the following equation: Among them, σ represents the Sigmoid activation function, and its calculation formula is Tanh represents the hyperbolic tangent activation function, which represents the Hadamard product (element-wise product), and its calculation formula is The input gate controls the introduction of new information, the forget gate determines which old information in the memory cell should be forgotten, and the output gate controls the output from the LSTM cell to the next state.

[0081] Compared with the CNN-LSTM model combining convolutional neural network (CNN) and long short-term memory network (LSTM), the CNN-BiLSTM model of bidirectional long short-term memory network (BiLSTM) used in the present invention provides some significant advantages and benefits, especially in processing one-dimensional time series data. The network structure of BiLSTM is as follows: Figure 8 Shown: L i represents the forward LSTM, L i ' represents reverse LSTM, s and s' are the time series information transmitted in the LSTM unit. CNN-BiLSTM can capture the contextual information in the time series more comprehensively by considering both the forward and backward information of the time series. This bidirectional processing mechanism gives CNN-BiLSTM a significant advantage in understanding the overall structure and long-term dependencies of the time series. The advantage of CNN-BiLSTM over CNN-LSTM is mainly reflected in its bidirectional processing mechanism, which not only improves the depth and breadth of the model's understanding of time series data, but also enhances the model's predictive performance and robustness. The overall workflow of CNN-BiLSTM is as follows: Fig. 9 shown.

[0082] The experiment compares the imaging speed of TFM based on CNN-BiLSTM network with that of traditional TFM imaging method. A thin aluminum plate with damage is used as the test sample, and a linear array sensor with 7 units is used to complete the excitation and reception of the signal. By changing the damage position and measuring the imaging time of CNN-LiLSTM network imaging and traditional TFM imaging method, the imaging performance is compared. Fig.10 (a) with Fig.10 (b) TFM positioning imaging of the lesions located at (-27, 27) mm and (1, 6) mm, respectively.

[0083] Table 2 shows the imaging speed of defects at different positions. At different defect positions, the average speed of TFM imaging based on the CNN-BiLSTM network is 46.4% higher than that of traditional TFM imaging.

[0084]

[0085] Table 2

[0086] In this embodiment, it should be specifically explained that the steps of training the ResNet network model parameters are:

[0087] The ResNet network of the deep residual learning framework is used to process the images after TFM imaging to identify and classify the size of the damage radius. The ResNet network overcomes the gradient vanishing problem in deep network training by optimizing the residual mapping between input and output. The ResNet network adopts the design of residual learning units to overcome the gradient vanishing problem encountered in deep network training, thereby deepening the network structure without affecting the training effect. Unlike the traditional strategy of directly learning the underlying mapping, the ResNet optimization strategy is achieved by learning the residual mapping between input and output. This structural innovation not only promotes the network's feature learning at a deeper level and improves the performance of image classification tasks, but also optimizes the network's training process, ensuring effective training even in extremely deep network structures. In addition, ResNet's wide applicability and friendliness to transfer learning make it a powerful tool for handling various image classification problems, and it performs well in both data-rich and data-scarce scenarios;

[0088] The radius recognition dataset is divided into a radius recognition training set, a radius recognition validation set, and a radius recognition test set in a ratio of 7:2:1. The radius recognition training set is used for model training, the radius recognition validation set is used for adjusting and optimizing model parameters, and the radius recognition test set is used to evaluate the accuracy and generalization ability of the model.

[0089] The stochastic gradient descent algorithm is used to train the model, and the cross entropy loss function is used as the optimization target. The early stopping mechanism is used to prevent overfitting to ensure that the model is effectively trained in a deep network structure. The cross entropy loss function is a commonly used loss function for classification problems, which is usually used to measure the difference between two probability distributions. The early stopping mechanism is a simple and effective technique for preventing machine learning models from overfitting. The basic idea is to monitor the performance of the model on the validation set during the model training process and stop training when the model performance begins to decline on the validation set to avoid overfitting the training data.

[0090] The ResNet network adopts the design of residual learning units, aiming to overcome the gradient vanishing problem encountered in deep network training, so as to achieve the deepening of the network structure without affecting the training effect. Unlike the traditional strategy of directly learning the underlying mapping, the ResNet optimization strategy is achieved by learning the residual mapping between input and output. This structural innovation not only promotes the network's feature learning at a deeper level and improves the performance of image classification tasks, but also optimizes the network's training process, ensuring effective training even in extremely deep network structures. In addition, ResNet's wide applicability and friendliness to transfer learning make it a powerful tool for handling various image classification problems, and it performs well in both data-rich and data-scarce scenarios.

[0091] The ResNet network of the deep residual learning framework was used to process the images after TFM imaging, a total of 900 TFM images, to identify and classify the size of the damage radius. Fig.11 The network structure for classifying circular hole radius is presented. The network overcomes the gradient vanishing problem in deep network training by optimizing the residual mapping between input and output.

[0092] During the in-depth training of the ResNet network model, 200 training batches were used and an early stopping mechanism was introduced to prevent overfitting. The performance of the model was continuously optimized during the training process, such as Fig.13 As shown in the figure, the loss function decreases significantly, while the classification accuracy gradually improves.

[0093] During the evaluation of the test set, the damage radius detection model based on the ResNet network showed excellent performance, achieving a high accuracy of 100%. This remarkable result not only highlights the efficiency of the ResNet model in processing complex data sets, but also demonstrates its strong generalization ability. Through the deep learning method, the model can accurately estimate the size of the damage, which is crucial for non-destructive testing technology. Especially in the field of Lamb wave non-destructive testing, accurate damage size information not only helps to assess the health status of the structure, but also provides important decision-making basis for maintenance and reinforcement work. In addition, the successful application of the ResNet model further confirms the great potential of deep learning technology in improving the accuracy and efficiency of non-destructive testing.

[0094] According to the experimental results, compared with the traditional full-focus imaging method, the CNN-BiLSTM network proposed in the present invention shows significant advantages in the accuracy and efficiency of damage detection; at the same time, the damage size detection results based on the ResNet network are accurate and reliable, which fully confirms the effectiveness and accuracy of the present invention in damage size assessment.

[0095] Finally: 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, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

[0096] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A multi-network based TFM method for locating and detecting plate damage, characterized in that: The following steps are involved: Step SO1: Obtain the Lamb wave dispersion curve of the metal plate, select a suitable mode as the Lamb wave excitation signal and modulate the selected Lamb wave excitation signal; Step SO2: setting a target area in the metal plate, and placing an excitation sensor at the target area, and symmetrically transmitting the Lamb wave excitation signal described in step SO1 on the upper and lower surfaces of the metal plate at the excitation sensor; Step SO3: a group of circular sensor arrays are arranged at designated positions on the surface of the metal plate to receive echo signals from the target area except for the excitation sensor; Step SO4: changing the distribution of circular holes in the circular sensor array, repeating the excitation in step SO2 and step SO3, and using each sensor data obtained as the original data set for region recognition; Step SO5: dividing the region recognition original data set obtained in step SO4 into a region recognition training data set, a region recognition test data set and a region recognition verification data set; Step SO6: Use the training data set obtained in step SO5 to train the parameters of the CNN-BiLSTM network model, and use the region recognition verification data set for verification to select the model with the best performance; Step SO7: using the model with the best performance obtained in step SO6 to identify the area to which the circular hole damage in the metal plate belongs to the area recognition test data set in step SO5; Step SO8: using the echo signal collected by the circular sensor array arranged in step SO3, the area of ​​the circular hole damage in the metal plate identified in step SO7 is imaged and located using the TFM algorithm; Step SO9: by changing the circular hole distribution in the circular sensor array, repeating the step SO8, using the obtained TFM imaging image as the radius recognition original data set, and then dividing the radius recognition original data set into a radius recognition training data set, a radius recognition test data set, and a radius recognition verification data set; Step S10: Use the radius recognition training data set obtained in the step SO9 to train the ResNet network model parameters, and use the radius recognition verification data set for verification, and select the model with the best performance. Use the obtained model with the best performance to identify the radius of circular hole damage in the plate using the radius recognition test data set in the step SO9.

2. The multi-network based TFM positioning and detection method for plate damage according to claim 1 is characterized in that: The step of selecting a suitable mode as the Lamb wave excitation signal in step S01 is to select the Lamb wave S0 mode signal with low dispersion and high speed as the excitation according to the Lamb wave dispersion curve of the metal plate.

3. The multi-network based TFM positioning and detection method for plate damage according to claim 1 is characterized in that: The method for modulating the selected Lamb wave excitation signal in step S01 is: Where H(t) is a step function, f c It represents the center frequency of the excitation signal, φ represents the phase of the signal, and N represents the number of peaks of the modulated sinusoidal signal.

4. The multi-network based TFM positioning and detection method for plate damage according to claim 1 is characterized in that: The steps of training the CNN-BiLSTM network model parameters using the training data set obtained in step S06 are: Step S061, dividing the region recognition data set into a region recognition training set, a region recognition verification set and a region recognition test set, wherein the region recognition training set is used for training the CNN-BiLSTM network, the region recognition verification set is used for optimizing and selecting model parameters, and the region recognition test set is used for evaluating the performance of the model; Step S062, using a stochastic gradient descent algorithm to train the CNN-BiLSTM network model, initialize network parameters, set an initial learning rate, and during the training process, update the network parameters through a back propagation algorithm to minimize the loss function; Step S063: After each round of training, the model performance is evaluated using the region recognition validation set, and the learning rate and other parameters are adjusted according to the validation results to improve the generalization ability and accuracy of the model; Repeat steps S062 and S063 until the early stopping condition is met or the preset maximum number of training rounds is reached, and finally select the model with the best performance on the validation set as the final model. The model training optimization formula is θ 新 =θ 旧 -α×▽ θ L(θ), where θ represents the model parameter, α represents the learning rate, and L(θ) represents the loss function. θ L(θ) is expressed as the gradient of the loss function with respect to the model parameters.

5. The multi-network based TFM positioning and detection method for plate damage according to claim 1 is characterized in that: The step of using the TFM algorithm to image and locate the area to which the circular hole damage in the metal plate identified in step SO7 belongs is: Step S071, dividing the detection area into grid points and treating each point as a focus target, reconstructing the image using the full matrix echo signal captured by the phased array, calculating the signal strength of each point, and generating a high-definition focus image; Step S072, set the array point (x i ,x i ) is the excitation point, and the corresponding other array point (x j ,x j ) The received induction signal is used by U ij (x i ,x j ,t) indicates that the intensity expression of the comprehensive signal obtained by superimposing the delayed signals corresponding to all excitation points and receiving points is: In this process, the signal is first emitted from the excitation source, reflected by the damage point, and finally received by the array sensor. The total duration of the whole process is expressed as where r ing represents the displacement vector from the excitation point to the damage point, and r back Then it is represented as the displacement vector from the damage point to the receiving point, and C is represented as the speed of the sound wave propagating in the plate; Step S073, draw an imaging map according to the I(P) value of each focus point P, normalize the data in the map, and take the maximum amplitude point in the local area of ​​the target area in the imaging map as the damage location result.

6. The multi-network based TFM positioning and detection method for plate damage according to claim 1 is characterized in that: The steps for training the ResNet network model parameters in step S10 are: Step S101, dividing the radius recognition data set into a radius recognition training set, a radius recognition verification set and a radius recognition test set, wherein the radius recognition training set is used for model training, the radius recognition verification set is used for adjusting and optimizing model parameters, and the radius recognition test set is used for evaluating the accuracy and generalization ability of the model; Step S102, adopting the stochastic gradient descent algorithm to train the model, using the cross entropy loss function as the optimization target, and using the early stopping mechanism to prevent overfitting.