A method for evaluating the interfacial bond strength of plant fiber reinforced plastics

Abstract

This invention provides a method for evaluating the interfacial bonding strength of plant fiber reinforced plastics, belonging to the field of bio-based composite materials technology. This invention employs Dijkstra's shortest path algorithm to calculate the path cost from the sample center to each candidate emission point and selects the candidate emission point with the minimum path cost as the optimized emission point. At the optimized emission point, a fine search is performed based on the transmission signal propagation time and frequency attenuation value to determine the optimal emission direction angle and optimal emission frequency. Finally, these optimal parameters are used to perform ultrasonic transmission tests on all test points to establish an ultrasonic characteristic response matrix, which is then input into an interface quality assessment model to output a predicted value of the interfacial bonding strength. This method solves the technical problem of insufficient detection accuracy caused by the difficulty in adaptively optimizing ultrasonic detection parameters when evaluating the interfacial bonding strength of plant fiber reinforced plastics.

Description

Technical Field

[0001] This invention belongs to the field of bio-based composite materials technology, and more specifically, relates to a method for evaluating the interfacial bonding strength of plant fiber reinforced plastics. Background Technology

[0002] In the field of quality evaluation of plant fiber reinforced plastics, interfacial bonding strength is a key indicator determining the mechanical properties of composite materials. Traditional techniques mainly employ ultrasonic transmission to perform non-destructive testing of interfacial quality, indirectly assessing the interfacial bonding state by analyzing the amplitude attenuation and propagation time of the transmitted signal. This method has been widely used in the testing of homogeneous materials and has formed a relatively mature testing standard system. However, traditional ultrasonic testing technology faces significant difficulties when applied to plant fiber reinforced plastics. Due to the heterogeneity of wood flour fiber distribution in the matrix, the ultrasonic wave propagation path is complex and variable. Different transmission parameters are required for different testing locations to obtain effective signals. However, existing methods generally use fixed transmission point positions and uniform transmission parameters, failing to adaptively adjust parameters according to the local characteristics of the sample, resulting in poor signal quality or even ineffective penetration in some areas. In other words, existing technologies suffer from the technical problem of insufficient accuracy due to the difficulty in adaptively optimizing ultrasonic testing parameters when evaluating the interfacial bonding strength of plant fiber reinforced plastics. Summary of the Invention

[0003] In view of this, the present invention provides a method for evaluating the interfacial bonding strength of plant fiber reinforced plastics, which can solve the technical problem in the prior art where the ultrasonic detection parameters are difficult to adaptively optimize, resulting in insufficient detection accuracy.

[0004] This invention is implemented as follows: A method for evaluating the interfacial bonding strength of plant fiber reinforced plastics is provided. Plant fiber reinforced plastic samples are prepared by mixing wood flour with a polyethylene matrix. Multiple test points are selected on the sample surface and their coordinate positions are marked. A spatial coordinate system for the sample is established. A grid is created on the sample surface, and candidate emission points are set at each grid node. An ultrasonic transmitting device and an ultrasonic receiving device are set up. Multiple sets of initial ultrasonic transmission parameters are set for each candidate emission point, and preliminary ultrasonic transmission tests are performed. The amplitude, propagation time, and frequency attenuation of the transmitted signal are collected. An optimized ultrasonic parameter graph is established. The weighted sum of spatial Euclidean distance and the similarity of the transmitted signal is used as the graph edge weights. Dijkstra's shortest path algorithm is employed. The method calculates the path cost from the center of the sample to each candidate emission point, selects the candidate emission point with the minimum path cost as the optimal emission point, and performs a fine search at the optimal emission point to determine the optimal emission direction angle and optimal emission frequency. Ultrasonic transmission tests are performed on all test points using the optimal emission point coordinates, optimal emission direction angle, and optimal emission frequency to establish an ultrasonic characteristic response matrix. The ultrasonic characteristic response matrix, sample temperature, wood flour particle size, and polyethylene molecular weight are input into the interface quality assessment model to output the interface bonding strength prediction value. The interface bonding strength prediction value is fitted with a Weibull distribution, and a Monte Carlo method is used for random sampling simulation to calculate the strength confidence interval and failure probability. The interface bonding strength evaluation results and reliability index are output.

[0005] The grid spacing is 5mm.

[0006] Specifically, the initial ultrasonic emission parameters are set by setting the candidate emission direction angle every 15° within the interval [0°, 90°] for each candidate emission point, and setting the candidate emission frequency every 100kHz within the interval [100kHz, 1000kHz].

[0007] Specifically, the calculation of the transmission signal similarity involves extracting the transmission signal amplitude for any two candidate emission points, calculating the absolute value of the difference between the transmission signal amplitudes, dividing the absolute value of the difference between the transmission signal amplitudes by the maximum value of the transmission signal amplitudes of all candidate emission points to obtain the normalized transmission signal difference, and subtracting the transmission signal difference from 1 to obtain the transmission signal similarity.

[0008] Specifically, the establishment of the ultrasonic parameter optimization graph involves taking all candidate emission points as graph nodes, calculating the spatial Euclidean distance and the transmission signal similarity for any two candidate emission points, normalizing the spatial Euclidean distance to the interval [0, 1], and calculating the sum of the normalized spatial Euclidean distance multiplied by 0.6 and the transmission signal similarity multiplied by 0.4 as the graph edge weight.

[0009] Specifically, the Dijkstra shortest path algorithm calculates the path cost by taking the candidate launch point closest to the center of the sample as the starting node, initializing the path cost from the starting node to itself to 0, and setting the path cost from the starting node to all other graph nodes to infinity. The graph node with the smallest path cost among the unvisited graph nodes is selected as the current node, and the path cost from the starting node to the adjacent graph nodes is calculated for all adjacent graph nodes of the current node.

[0010] Specifically, the determination of the optimal transmission direction angle involves finding the candidate transmission direction angle with the shortest transmission signal propagation time among all candidate transmission direction angles at the optimized transmission point, and then performing a fine search with a step size of 1° within the interval [candidate transmission direction angle minus 7°, candidate transmission direction angle plus 7°] centered on the candidate transmission direction angle.

[0011] Specifically, the determination of the optimal transmission frequency involves finding the candidate transmission frequency with the smallest transmission signal frequency attenuation value among all candidate transmission frequencies at the optimized transmission point, and then performing a fine search within the interval [the candidate transmission frequency minus 50kHz, the candidate transmission frequency plus 50kHz] with a step size of 10kHz, centered on the candidate transmission frequency.

[0012] Specifically, the establishment of the ultrasonic characteristic response matrix involves numbering all test points on the sample surface according to spatial coordinates, performing ultrasonic transmission tests on each test point using the optimal emission point coordinates, optimal emission direction angle, and optimal emission frequency, and arranging the optimal emission point coordinates, optimal emission direction angle, optimal emission frequency, transmission signal amplitude, transmission signal propagation time, and transmission signal frequency attenuation value of each test point according to the test point number order to form an ultrasonic characteristic response matrix.

[0013] The interface quality assessment model has an encoder-decoder architecture. The encoder contains three convolutional layers and two sparse attention layers. The input of the first convolutional layer is the ultrasonic feature response matrix, the kernel size is 3×3, and the number of output channels is 64. The kernel size of the second convolutional layer is 3×3, and the number of output channels is 128. The kernel size of the third convolutional layer is 3×3, and the number of output channels is 256. The decoder contains two fully connected layers.

[0014] The sparse attention layer contains multiple attention heads. Each attention head uses a local window mechanism to calculate the attention weight between each test point and test points in its spatial neighborhood. The radius of the local window is the square root of the total number of test points, rounded down. The sparse attention layer achieves sparsity of attention weights through a mask matrix. The attention weights of test point pairs whose spatial distance exceeds the radius of the local window in the mask matrix are set to zero.

[0015] The interface quality assessment model employs a knowledge retention mechanism based on continuous learning. It protects important parameters by solidifying them with elastic weights, maintains historical performance by utilizing experience replay, and uses progressive network expansion to allocate capacity for new tasks while constraining knowledge forgetting through regularization.

[0016] Specifically, the elastic weight solidification involves calculating the gradient of each network parameter on the historical task, accumulating the squares of the gradients to obtain the diagonal elements of the Fisher information matrix, and applying an elastic weight solidification penalty term to important network parameters during subsequent training. The penalty term is the square of the difference between the current value of the network parameter and the value of the network parameter when the historical task training was completed, multiplied by the diagonal elements of the Fisher information matrix.

[0017] Specifically, the progressive network extension involves calculating the Wasserstein distance between the data feature distributions of the new task and the historical task. When the Wasserstein distance is greater than 0.5, it is determined that the new task is significantly different from the historical task. A new parallel network branch is added to the interface quality assessment model. The parameters of the first two convolutional layers of the encoder in the new parallel network branch are shared with the original network branch and are frozen without being updated. The third convolutional layer and the sparse attention layer of the new parallel network branch are initialized independently.

[0018] This invention selects optimal emission points by constructing an ultrasonic parameter optimization graph and using Dijkstra's shortest path algorithm to calculate the path cost from the sample center to each candidate emission point. Then, at this optimal emission point, a fine search is performed based on the transmission signal propagation time and frequency attenuation value to determine the optimal emission direction angle and optimal emission frequency. This solves the problem of fixed ultrasonic testing parameters in traditional techniques, which cannot adapt to material heterogeneity. This invention uses spatial Euclidean distance and transmission signal similarity as graph edge weights to establish the parameter optimization graph. This ensures that the selected optimal emission point considers not only the representativeness of its spatial location but also the typicality of its signal propagation characteristics, thereby ensuring that the starting point for subsequent fine parameter search has global optimum. This avoids the local optimum trap caused by improper initial parameter selection in traditional methods. Based on the optimized emission point, a fine search with small step sizes further improves the accuracy of the emission parameters, enabling ultrasonic testing to adaptively configure parameters for the heterogeneous characteristics of plant fiber reinforced plastics. In summary, this invention solves the technical problem mentioned in the background art of insufficient detection accuracy due to the difficulty in adaptively optimizing ultrasonic testing parameters when evaluating the interfacial bonding strength of plant fiber reinforced plastics. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention.

[0020] Figure 2 This is a diagram showing the amplitude distribution of the ultrasonic transmission test signal in the embodiment.

[0021] Figure 3 This is a Weibull distribution fitting diagram of the interface bonding strength in the embodiment.

[0022] Figure 4 This is a visualization of the Weibull distribution feature matrix in the embodiment. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0024] like Figure 1 The diagram shows a flowchart of a method for evaluating the interfacial bonding strength of plant fiber reinforced plastics provided by this invention. This method includes the following steps:

[0025] S01. Prepare plant fiber reinforced plastic samples by mixing wood flour with polyethylene matrix, select multiple test points on the sample surface and mark their coordinate positions to establish a sample spatial coordinate system.

[0026] S02. Establish a grid on the sample surface with a grid spacing of 5mm. Set up candidate emission points at each grid node, and set up an ultrasonic transmitter and an ultrasonic receiver. The ultrasonic receiver is arranged on the opposite side of the sample.

[0027] S03. For each candidate emission point, set the candidate emission direction angle every 15° in the interval [0°, 90°], and set the candidate emission frequency every 100kHz in the interval [100kHz, 1000kHz], and combine them to form multiple sets of initial ultrasonic emission parameters.

[0028] S04. Perform preliminary ultrasonic transmission tests on each set of initial ultrasonic emission parameters, collect the amplitude of the transmitted signal, the propagation time of the transmitted signal, and the frequency attenuation value of the transmitted signal, and record the transmission signal data of all candidate emission points.

[0029] S05. Establish an ultrasonic parameter optimization graph, take all candidate emission points as graph nodes, calculate the spatial Euclidean distance and transmission signal similarity between any two candidate emission points, take the weighted sum of spatial Euclidean distance and transmission signal similarity as graph edge weight, use Dijkstra's shortest path algorithm to calculate the path cost from the sample center to each candidate emission point, and select the candidate emission point with the minimum path cost as the optimized emission point.

[0030] S06. In optimizing the position of the transmission point, based on the transmission signal propagation time and transmission signal frequency attenuation value recorded in step S04, a fine search is performed in the interval near the candidate transmission direction angle corresponding to the shortest transmission signal propagation time with a step size of 1° to determine the optimal transmission direction angle. A fine search is performed in the interval near the candidate transmission frequency corresponding to the minimum transmission signal frequency attenuation value with a step size of 10kHz to determine the optimal transmission frequency.

[0031] S07. Using the optimal emission point coordinates, optimal emission direction angle and optimal emission frequency, perform ultrasonic transmission tests on all test points, record the transmission signal amplitude, transmission signal propagation time and transmission signal frequency attenuation value of each test point, and establish an ultrasonic characteristic response matrix.

[0032] S08. Input the ultrasonic characteristic response matrix, sample temperature, wood flour particle size and polyethylene molecular weight into the interface quality assessment model, process it, and output the predicted value of the interface bonding strength at each test point.

[0033] S09. Fit the predicted interface strength values ​​of all test points to a Weibull distribution, calculate the Weibull characteristic strength and Weibull shape parameters, and establish the interface strength probability distribution function.

[0034] S10. The Monte Carlo method is used to randomly sample and simulate the interface strength probability distribution function, generating a random number sequence of length 10000. The strength confidence interval and failure probability at a 95% confidence level are calculated. When the failure probability is greater than 0.05, the number of test points is increased and steps S07 to S09 are repeated. When the failure probability is less than or equal to 0.05, the interface combined strength evaluation results and reliability indicators are output.

[0035] The steps for establishing the ultrasonic characteristic response matrix specifically include: numbering all test points on the sample surface according to spatial coordinate order; performing ultrasonic transmission tests on each test point using the optimal emission point coordinates, optimal emission direction angle, and optimal emission frequency, and recording the transmission signal amplitude, transmission signal propagation time, and transmission signal frequency attenuation value; arranging the optimal emission point coordinates, optimal emission direction angle, optimal emission frequency, transmission signal amplitude, transmission signal propagation time, and transmission signal frequency attenuation value of each test point according to the test point number order to form an ultrasonic characteristic response matrix; the number of rows in the ultrasonic characteristic response matrix corresponds to the number of test points, and the number of columns in the ultrasonic characteristic response matrix is ​​9, corresponding to the optimal emission point horizontal coordinate, optimal emission point vertical coordinate, optimal emission direction angle, optimal emission frequency, transmission signal amplitude, transmission signal propagation time, transmission signal frequency attenuation value, and predicted value of interface bonding strength, respectively.

[0036] The specific steps for calculating the similarity of the transmitted signals include: for any two candidate emission points, extracting the amplitude of the transmitted signals recorded in step S04; calculating the absolute value of the difference between the amplitudes of the transmitted signals of the two candidate emission points; dividing the absolute value of the difference between the amplitudes of the transmitted signals by the maximum value of the amplitudes of the transmitted signals of all candidate emission points to obtain the normalized difference in the transmitted signals; subtracting the difference in the transmitted signals from 1 to obtain the similarity of the transmitted signals, where the value range of the similarity of the transmitted signals is the interval [0, 1].

[0037] The steps for establishing the ultrasonic parameter optimization graph specifically include: using all candidate emission points as graph nodes of the ultrasonic parameter optimization graph; calculating the spatial Euclidean distance between any two candidate emission points; calculating the transmission signal similarity between two candidate emission points according to the transmission signal similarity calculation steps based on the transmission signal data recorded in step S04; normalizing the spatial Euclidean distance to the interval [0, 1] by dividing the spatial Euclidean distance by the maximum value of the spatial Euclidean distance between all pairs of candidate emission points; calculating the sum of the normalized spatial Euclidean distance multiplied by 0.6 and the transmission signal similarity multiplied by 0.4 as the graph edge weight; when the graph edge weight is less than 0.7, connecting the corresponding two candidate emission points in the ultrasonic parameter optimization graph to form the edge set of the graph.

[0038] The specific steps of Dijkstra's shortest path algorithm for calculating path cost include: selecting the candidate emission point closest to the sample center as the starting node; initializing the path cost from the starting node to itself to 0, and the path cost from the starting node to all other graph nodes to infinity; selecting the graph node with the smallest path cost from the unvisited graph nodes as the current node; calculating the path cost from the starting node to the adjacent graph nodes for all adjacent graph nodes of the current node, where the path cost is the sum of the path cost from the starting node to the current node and the edge weights from the current node to the adjacent graph nodes; updating the path cost of the adjacent graph nodes if the calculated path cost is less than the original path cost of the adjacent graph nodes; marking the current node as visited, and repeating the above steps until all graph nodes have been visited; and selecting the candidate emission point corresponding to the graph node with the smallest path cost as the optimized emission point.

[0039] The steps for determining the optimal transmission direction angle specifically include: finding the candidate transmission direction angle with the shortest transmission signal propagation time among all candidate transmission direction angles recorded in step S04; setting a fine search angle with a step size of 1° within the interval [the candidate transmission direction angle minus 7°, the candidate transmission direction angle plus 7°] centered on the candidate transmission direction angle; performing ultrasonic transmission tests on each fine search angle and recording the transmission signal propagation time; and selecting the fine search angle corresponding to the minimum transmission signal propagation time as the optimal transmission direction angle.

[0040] The specific steps for determining the optimal transmission frequency include: finding the candidate transmission frequency with the smallest transmission signal frequency attenuation value among all candidate transmission frequencies recorded in step S04; setting a fine search frequency with a step size of 10kHz within the interval [the candidate transmission frequency minus 50kHz, the candidate transmission frequency plus 50kHz] centered on the candidate transmission frequency; performing ultrasonic transmission tests on each fine search frequency and recording the transmission signal frequency attenuation value; and selecting the fine search frequency corresponding to the minimum transmission signal frequency attenuation value as the optimal transmission frequency.

[0041] The specific steps of the Weibull distribution fitting include: arranging the predicted interfacial bonding strength values ​​of all test points in ascending order; calculating the cumulative failure probability of the i-th predicted interfacial bonding strength value as i divided by the total number of test points plus 1; taking the natural logarithm of each predicted interfacial bonding strength value and its corresponding cumulative failure probability, and then taking the natural logarithm of the cumulative failure probability to complete the double logarithmic transformation; fitting a straight line using the natural logarithm of the predicted interfacial bonding strength value as the abscissa and the double logarithmic value of the cumulative failure probability as the ordinate, obtaining the slope and intercept of the fitted line; the Weibull shape parameter is equal to the slope of the fitted line, and the Weibull characteristic strength is equal to the power of the natural constant, with the power being the intercept of the fitted line divided by the negative of the Weibull shape parameter; establishing the interfacial strength probability distribution function, the cumulative distribution function expression of which includes the Weibull characteristic strength and the Weibull shape parameter.

[0042] The Monte Carlo simulation steps specifically include: generating a random number sequence of length 10000 using the inverse transformation sampling method based on the interface strength probability distribution function; counting the number of values ​​less than the design strength requirement in the random number sequence; calculating the failure probability as the number of values ​​divided by 10000; arranging the random number sequence in ascending order, extracting the 475th value as the lower limit of the strength confidence interval, and extracting the 9525th value as the upper limit of the strength confidence interval, with the lower and upper limits of the strength confidence interval corresponding to a 95% confidence level; and outputting the Weibull characteristic strength, Weibull shape parameters, lower and upper limits of the strength confidence interval, and failure probability as the interface combined strength evaluation results and reliability indicators.

[0043] The interface quality assessment model employs an encoder-decoder architecture. The encoder consists of three convolutional layers and two sparse attention layers. The first convolutional layer takes the ultrasonic feature response matrix as input, has a 3×3 kernel, and outputs 64 channels. The second and third convolutional layers each have a 3×3 kernel and output 128 channels. The sparse attention layers capture the global dependencies between different test points. The decoder consists of two fully connected layers. The first fully connected layer has an input dimension of 256 and an output dimension of 128. The second fully connected layer has an input dimension of 128 and an output dimension of 1, outputting the predicted interface binding strength for each test point. The interface quality assessment model uses a knowledge retention mechanism based on continuous learning. It protects important parameters by using elastic weights, maintains historical performance through experience replay, allocates capacity for new tasks through progressive network expansion, and constrains knowledge forgetting through regularization.

[0044] The sparse attention layer specifically includes the following structure: The sparse attention layer contains multiple attention heads, the number of which is determined during the training of the interface quality assessment model based on the performance of the validation subset; each attention head employs a local window mechanism, calculating only the attention weights between each test point and test points within its spatial neighborhood; the radius of the local window is the square root of the total number of test points, rounded down; the sparse attention layer achieves sparsity of attention weights through a mask matrix, setting the attention weights of test point pairs whose spatial distance exceeds the radius of the local window to zero; each attention head contains a query matrix, a key matrix, and a value matrix, the initial weights of which are obtained by mapping the input feature vector constructed from the sample temperature, wood flour particle size, and polyethylene molecular weight through a fully connected layer; the output of the sparse attention layer is the concatenation of the outputs of all attention heads, mapped to the encoder's hidden dimension of 256 through a linear transformation.

[0045] The steps for establishing the training dataset for the interface quality assessment model specifically include: preparing plant fiber reinforced plastic samples with different wood flour contents, different wood flour particle sizes, and different polyethylene molecular weights, with no fewer than 500 samples; recording the sample temperature, wood flour particle size, and polyethylene molecular weight for each sample group; performing ultrasonic transmission tests on each sample group according to steps S01 to S07 and acquiring the ultrasonic characteristic response matrix; performing tensile tests on each sample group and measuring the actual interfacial bonding strength; using the ultrasonic characteristic response matrix, sample temperature, wood flour particle size, and polyethylene molecular weight as input data, and the actual interfacial bonding strength as label data to construct the training dataset; and dividing the training dataset into a training subset and a validation subset in an 8:2 ratio.

[0046] The specific steps for training the interface quality assessment model include: initializing the network parameters of the interface quality assessment model, and initializing the number of attention heads in the sparse attention layer to 8; for each sample in the training subset, extracting the sample temperature, wood flour particle size, and polyethylene molecular weight to construct an input feature vector; mapping the input feature vector through a fully connected layer to obtain the initial weights of the query matrix, key matrix, and value matrix of the sparse attention layer; inputting the ultrasonic feature response matrix of the training subset into the interface quality assessment model, extracting local features through a convolutional layer, capturing global dependencies through a sparse attention layer, and outputting the interface combined with the intensity prediction value through a decoder. The mean squared error between the predicted and actual interface bonding strength is calculated. Backpropagation and the Adam optimizer are used to update network parameters to minimize the mean squared error. After each training epoch, the interface quality assessment model performance is evaluated using a validation subset, and the mean squared error on the validation subset is recorded as the validation loss. Attention heads of 4, 6, 8, 10, 12, 14, and 16 are tested. The interface quality assessment model is trained for each number of attention heads, and the validation loss is recorded. The number of attention heads corresponding to the minimum validation loss is selected as the final number of attention heads for the sparse attention layer. When the validation loss is continuously below 10 training epochs... Training stops when the training cycle no longer decreases, and the network parameters of the interface quality assessment model are saved. During the training of the interface quality assessment model, for the types of samples that have been trained, a Fisher information matrix for each network parameter in relation to the historical task is calculated through an elastic weight solidification mechanism. The diagonal elements of the Fisher information matrix represent the importance of the network parameter, and the update magnitude of highly important network parameters in subsequent training is constrained by a penalty term. A historical sample buffer is established through an experience replay mechanism, with a capacity of 10% of the total training dataset. In each training batch, samples are randomly drawn from the historical sample buffer and compared with the current task. The samples are mixed for training at a ratio of 1:9. When the data feature distribution of the new task is greater than 0.5 from the Wasserstein distance of the historical task, a new parallel network branch is added to the interface quality assessment model through a progressive network expansion mechanism. The new parallel network branch shares the first two convolutional layers of the encoder with the original network branch, while the third convolutional layer and the sparse attention layer are independent. The decoder output is fused by learningable weights. A regularization term is added to the loss function through regularization constraints. The regularization term is the L2 norm between the network parameters of the new parallel network branch and the network parameters of the original network branch, and the regularization coefficient is 0.01.

[0047] The specific steps of the knowledge retention mechanism based on continuous learning include: In the initial stage of training the interface quality assessment model, the gradient of each network parameter on historical tasks is calculated, and the squares of the gradients are accumulated to obtain the diagonal elements of the Fisher information matrix. The larger the value of the diagonal element of the Fisher information matrix, the more important the network parameter is to the historical task. In subsequent training, an elastic weight solidification penalty term is applied to important network parameters. The penalty term is the square of the difference between the current value of the network parameter and the value of the network parameter when the historical task training is completed, multiplied by the diagonal element of the Fisher information matrix. Minimizing the penalty term limits the update amplitude of important network parameters; this process is elastic weight solidification. A historical sample buffer is established, which uses a reservoir sampling algorithm to store representative samples from historical tasks, ensuring that the proportion of samples from each historical task in the historical sample buffer is equal. In each training batch, one-ninth of the number of samples from the historical sample buffer is randomly selected from the current task sample buffer, and the historical samples are mixed with the current task samples and input into the interface quality assessment model for training; this process is experience replay. The Wasserstein distance is calculated between the data feature distributions of the new task and the historical task. When the Wasserstein distance is greater than 0.5, the new task is considered significantly different from the historical task. A new parallel network branch is added to the interface quality assessment model. The parameters of the first two convolutional layers of the encoder of the new parallel network branch are shared with the original network branch and frozen without being updated. The third convolutional layer and sparse attention layer of the new parallel network branch are initialized independently. The decoder of the new parallel network branch is initialized independently. The output of the decoder of the original network branch and the output of the decoder of the new parallel network branch are weighted and summed through a learnable weight vector to obtain the final output. This process is called progressive network expansion. A regularization term is added to the loss function. The regularization term calculates the L2 norm between the network parameters of the independent layer of the new parallel network branch and the network parameters of the corresponding layer of the original network branch. The regularization term is multiplied by a regularization coefficient of 0.01 and added to the total loss function. When the total loss function is minimized by gradient descent, the regularization term constrains the network parameters of the new parallel network branch to prevent them from deviating too much from the network parameters of the original network branch. This process is called regularization constraint knowledge forgetting.

[0048] The technical effects of the knowledge retention mechanism based on continuous learning are as follows: In the actual production process of plant fiber reinforced plastics, the raw material batches, processing technology, and environmental conditions are constantly changing, leading to a dynamic evolution trend in the interface bonding characteristics. Traditional static models are difficult to adapt to these changes, requiring frequent retraining and consuming a lot of time and computing resources. Through the elastic weight solidification mechanism, the interface quality assessment model can identify and protect network parameters that are crucial to historical tasks, avoiding damage to these network parameters when learning new tasks. This ensures that the interface quality assessment model maintains accurate identification of interface characteristics under historical production conditions. The elastic weight solidification mechanism quantifies the importance of network parameters through the Fisher information matrix, with higher-importance network parameters receiving stronger penalties, thereby automatically balancing the performance requirements of new and old tasks when updating parameters. Through the experience replay mechanism, the interface quality assessment model periodically reviews historical samples when learning new tasks, strengthening the memory of historical data patterns and preventing the complete forgetting of knowledge from old tasks due to focusing on new tasks. The experience replay mechanism uses a reservoir sampling algorithm to ensure the representativeness and balance of historical samples, avoiding excessive forgetting or over-reinforcement of certain historical tasks. Through the progressive network expansion mechanism, the interface quality assessment model... The model allocates dedicated network capacity for significantly different new tasks, avoiding interference between new and old tasks in the same network space. Simultaneously, it retains existing network branches to handle historical tasks, achieving knowledge isolation and coexistence. The progressive network expansion mechanism achieves knowledge transfer through shared low-level feature extraction layers, task isolation through independent high-level feature extraction layers, and task collaboration through learnable weight fusion. Through regularization constraints, the interface quality assessment model maintains the correlation between network parameters and existing branches when expanding new parallel network branches, creating a flexible connection between new and old knowledge. This allows for both learning new features and inheriting historical experience. Regularization constraints penalize network parameter deviations using the L2 norm, ensuring that new parallel network branches do not completely deviate from the knowledge space of existing branches. This knowledge retention mechanism enables the interface quality assessment model to continuously learn and optimize its performance with the accumulation of production data, without needing to be completely rebuilt. It adapts to the diverse raw material sources, dynamic adjustment of process parameters, and continuous expansion of product specifications in the production of plant fiber reinforced plastics, significantly reducing the maintenance cost of the interface quality assessment model, improving the long-term stability and reliability of interface strength evaluation, and providing continuous and effective technical support for quality control in the production process.

[0049] The steps for establishing the ultrasonic parameter sensitivity matrix specifically include: randomly selecting 10% of all test points on the sample surface as representative test points; for each representative test point, fixing the optimal emission direction angle and optimal emission frequency, increasing the horizontal coordinate of the optimal emission point by 1 mm, and then performing an ultrasonic transmission test, recording the change in the amplitude of the transmitted signal, and calculating the change in the amplitude of the transmitted signal divided by 1 mm to obtain the sensitivity of the horizontal coordinate of the optimal emission point; repeating the above steps to calculate the sensitivity of the vertical coordinate of the optimal emission point; fixing the optimal emission point coordinates and optimal emission frequency, increasing the optimal emission direction angle by 1°, and then performing an ultrasonic transmission test, recording the change in the amplitude of the transmitted signal, and calculating the amplitude of the transmitted signal... The change in value is divided by 1° to obtain the optimal transmission direction angle sensitivity; the optimal transmission point coordinates and optimal transmission direction angle are fixed, and the optimal transmission frequency is increased by 10kHz before ultrasonic transmission test is performed. The change in transmission signal amplitude is recorded, and the change in transmission signal amplitude is calculated and divided by 10kHz to obtain the optimal transmission frequency sensitivity; the optimal transmission point abscissa sensitivity, optimal transmission point ordinate sensitivity, optimal transmission point vertical coordinate sensitivity, optimal transmission direction angle sensitivity, and optimal transmission frequency sensitivity of all representative test points are arranged in the order of representative test point numbers to form an ultrasonic parameter sensitivity matrix; the number of rows in the ultrasonic parameter sensitivity matrix corresponds to the number of representative test points, and the number of columns in the ultrasonic parameter sensitivity matrix is ​​5.

[0050] The steps for establishing the Weibull distribution characteristic matrix specifically include: dividing the sample surface into three segments along both the length and width directions, forming nine regions; extracting the predicted interfacial bonding strength values ​​of all test points within each region; fitting the predicted interfacial bonding strength values ​​within the region according to the Weibull distribution fitting steps, and calculating the Weibull characteristic strength and Weibull shape parameters of the region; simulating the region according to the Monte Carlo method simulation steps, and calculating the lower limit of the strength confidence interval and the failure probability of the region at a 95% confidence level; arranging the Weibull characteristic strength, Weibull shape parameters, lower limit of the strength confidence interval, and failure probability of the nine regions in order of region number to form the Weibull distribution characteristic matrix; the Weibull distribution characteristic matrix has nine rows and four columns, corresponding to the Weibull characteristic strength, Weibull shape parameters, lower limit of the strength confidence interval, and failure probability, respectively.

[0051] The optimized emission point is the candidate emission point with the minimum path cost from the nearest candidate emission point to all candidate emission points, calculated using Dijkstra's shortest path algorithm in the ultrasonic parameter optimization graph; the optimal emission point coordinates are the spatial coordinates corresponding to the optimized emission point; the input feature vector is a three-dimensional vector constructed based on sample temperature, wood flour particle size, and polyethylene molecular weight, used for initializing the weight mapping of the query matrix, key matrix, and value matrix of the sparse attention layer in the interface quality assessment model; the Fisher information matrix is ​​a diagonal matrix calculated during the training of the interface quality assessment model to quantify the importance of network parameters to historical tasks; Wasser The Stein distance is a metric used to measure the difference in characteristic distribution between new and historical task data; the Weibull feature strength represents the strength value corresponding to a cumulative failure probability of 63.2% for the interface bonding strength, reflecting the overall level of interface bonding strength; the Weibull shape parameter represents the dispersion of the interface bonding strength distribution, and a larger Weibull shape parameter indicates a more concentrated strength distribution and better consistency of interface bonding quality; the lower limit of the strength confidence interval represents the minimum guaranteed value of the interface bonding strength at a given confidence level, used for reliability design of interface bonding strength; the failure probability represents the probability that the interface bonding strength is lower than the design strength requirement, used to evaluate the reliability level of the interface bonding.

[0052] The specific implementation methods of the above steps are described in detail below.

[0053] The specific implementation of step S01 involves mixing wood flour and polyethylene matrix in a set ratio, and preparing plant fiber reinforced plastic samples through a hot pressing process. The sample dimensions are 100 mm in length, 100 mm in width, and 10 mm in thickness. Multiple test points are selected on the sample surface using a grid-based method. The number of test points is determined based on the sample area, generally 50 to 100. For each test point, a three-dimensional coordinate measuring instrument is used to mark its horizontal, vertical, and vertical coordinates. A spatial rectangular coordinate system is established with the geometric center of the sample as the origin. The purpose of this step is to provide a spatial position reference for subsequent ultrasonic transmission testing, ensuring the accurate positioning of the test points and the spatial correspondence of the data.

[0054] The specific implementation of step S02 involves using computer-aided design software to establish a grid on the sample surface. The grid adopts a square grid structure with a uniform grid spacing of 5 mm. Candidate emission points are set at each grid node position. The number of candidate emission points is calculated based on the sample size and grid spacing, generally 400 to 500. An ultrasonic transmitting device is set on the upper surface of the sample, and an ultrasonic receiving device is arranged at the corresponding position on the lower surface of the sample. The ultrasonic transmitting device uses a broadband ultrasonic transducer, and the ultrasonic receiving device uses a high-sensitivity ultrasonic sensor. The purpose of this step is to establish a candidate emission point network covering the entire surface of the sample, providing sufficient space for subsequent ultrasonic parameter optimization. The grid spacing of 5 mm is a balance value that comprehensively considers test accuracy and test efficiency.

[0055] The specific implementation of step S03 involves setting different emission direction angles and emission frequencies for each candidate emission point. The emission direction angles are set every 15° from 0° up to 90°, forming a total of 7 candidate emission direction angles. The emission frequencies are set every 100kHz up to 1000kHz from 100kHz up to 1000kHz, forming a total of 10 candidate emission frequencies. The 7 candidate emission direction angles and 10 candidate emission frequencies are then fully combined to form 70 sets of initial ultrasonic emission parameters. The purpose of this step is to establish an initial parameter set covering the emission parameter space, providing basic data for subsequent parameter optimization. The angle interval of 15° and the frequency interval of 100kHz are empirical values ​​based on the propagation characteristics of ultrasonic waves in plant fiber reinforced plastics.

[0056] The specific implementation of step S04 involves performing ultrasonic transmission tests on each set of initial ultrasonic emission parameters for each candidate emission point. The ultrasonic transmitting device emits ultrasonic signals according to the set emission direction angle and emission frequency. The ultrasonic receiving device receives the transmission signal and collects the transmission signal amplitude, transmission signal propagation time, and transmission signal frequency attenuation value. The transmission signal amplitude is obtained through a peak detection circuit. The transmission signal propagation time is calculated by the time difference between the emission time and the reception time. The transmission signal frequency attenuation value is calculated by comparing the power spectral density difference between the emission frequency and the reception frequency after performing a fast Fourier transform on the received signal. The transmission signal data of all candidate emission points under all initial ultrasonic emission parameters are recorded and stored. The purpose of this step is to obtain the transmission response characteristics of the sample under different ultrasonic parameters, providing a data basis for subsequent parameter optimization and interface quality assessment.

[0057] The specific implementation of step S05 involves establishing an ultrasonic parameter optimization graph with candidate emission points as nodes. First, all candidate emission points are treated as graph nodes. For any two candidate emission points, the Euclidean distance in three-dimensional space is calculated. The Euclidean distance is the square root of the sum of the squares of the differences in the horizontal, vertical, and vertical coordinates of the two points. Then, the transmitted signal amplitude recorded in step S04 is extracted between the two candidate emission points. The absolute value of the difference in the transmitted signal amplitude between the two points is calculated. This absolute value is divided by the maximum value of the transmitted signal amplitude of all candidate emission points to obtain the normalized transmitted signal difference. The difference is subtracted from 1 to obtain the transmitted signal similarity. The spatial Euclidean distance is normalized by dividing by the maximum value of the spatial Euclidean distance between all candidate emission point pairs. The weighted sum of the normalized spatial Euclidean distance multiplied by a weighting coefficient of 0.6 and the transmitted signal similarity multiplied by a weighting coefficient of 0.4 is calculated as the graph edge weight. When the graph edge weight is less than a threshold of 0.7, the edge weight is considered to be less than the threshold of 0.7. A graph edge connection is established between the candidate emission points. Dijkstra's shortest path algorithm is used, with the candidate emission point closest to the sample center as the starting node. The path cost from the starting node to itself is initialized to 0, and the path cost to other nodes is infinite. The node with the smallest path cost among the unvisited nodes is selected as the current node. The path cost from the starting node to the current node is calculated for all adjacent nodes of the current node. If the path cost is less than the original path cost of the adjacent node, it is updated. The current node is marked as visited and this process is repeated until all nodes are visited. The candidate emission point corresponding to the node with the smallest path cost is selected as the optimized emission point. The purpose of this step is to select the optimal ultrasonic emission position by comprehensively considering the spatial distribution and signal propagation characteristics. The weight coefficients of 0.6 and 0.4 reflect the relative importance of spatial distance and signal similarity, and the threshold of 0.7 is used to control the sparsity of the graph connection to improve computational efficiency.

[0058] The specific implementation of step S06 involves performing a fine search for transmission parameters at the optimized transmission point location. First, from the transmission signal propagation times recorded in step S04 for all candidate transmission direction angles at the optimized transmission point, the candidate transmission direction angle corresponding to the shortest propagation time is identified. Using this angle as the center, a fine search angle is set in a step size of 1° within the range of this angle minus 7° to this angle plus 7°. Ultrasonic transmission tests are performed on each fine search angle, and the transmission signal propagation time is recorded. The fine search angle corresponding to the minimum transmission signal propagation time is selected as the optimal transmission direction angle. Then, from the optimized transmission point recorded in step S04 for all candidate transmission frequencies... The candidate transmission frequency corresponding to the minimum attenuation value is found among the transmitted signal frequency attenuation values. A fine search frequency is set with a step size of 10kHz in the range of 50kHz minus 50kHz to 50kHz plus 50kHz. Ultrasonic transmission test is performed for each fine search frequency and the transmitted signal frequency attenuation value is recorded. The fine search frequency corresponding to the minimum transmitted signal frequency attenuation value is selected as the optimal transmission frequency. The purpose of this step is to perform local fine optimization on the basis of coarse search to obtain the best ultrasonic transmission parameters. The search range of 7° and 50kHz and the search step size of 1° and 10kHz are a compromise choice that takes into account the optimization accuracy and computational cost.

[0059] The specific implementation of step S07 involves performing ultrasonic transmission tests on all test points on the sample surface using optimized emission point coordinates, optimal emission direction angle, and optimal emission frequency. For each test point, the amplitude of the transmitted signal, the propagation time of the transmitted signal, and the frequency attenuation value of the transmitted signal are recorded. All test points are numbered in spatial coordinate order. The optimized emission point horizontal coordinate, vertical coordinate, optimal emission direction angle, optimal emission frequency, amplitude of the transmitted signal, propagation time of the transmitted signal, frequency attenuation value of the transmitted signal, and the reserved interface bonding strength prediction value are arranged in the order of the test point numbers to form an ultrasonic characteristic response matrix. The number of rows in the matrix is ​​equal to the number of test points, and the number of columns in the matrix is ​​9. The purpose of this step is to obtain the full-surface transmission response data of the sample under optimal ultrasonic parameters and construct a structured data matrix for subsequent model processing.

[0060] The specific implementation of step S08 involves using the ultrasonic feature response matrix as the main input and the sample temperature, wood flour particle size, and polyethylene molecular weight as auxiliary inputs. After constructing the input feature vector, it is input into the interface quality assessment model for processing. The interface quality assessment model adopts an encoder-decoder architecture. The encoder contains three convolutional layers for extracting the local spatial patterns of the ultrasonic features. The first convolutional layer takes the ultrasonic feature response matrix as input, has a 3×3 kernel size, and 64 output channels. The second convolutional layer has a 3×3 kernel size and 128 output channels. The third convolutional layer has a 3×3 kernel size and 256 output channels. The encoder also includes two sparse attention layers for... To capture the global dependencies between different test points, the sparse attention layer contains multiple attention heads. Each attention head employs a local window mechanism to calculate the attention weights only between a test point and its spatial neighbors. The local window radius is the floor function of the square root of the total number of test points. Attention sparsity is achieved by resetting the attention weights of test point pairs exceeding the local window radius to zero using a mask matrix. Each attention head contains a query matrix, a key matrix, and a value matrix. The initial weights of these matrices are obtained by mapping the input feature vector constructed from the sample temperature, wood flour particle size, and polyethylene molecular weight through a fully connected layer. The output of the sparse attention layer is the concatenation of the outputs of all attention heads and linearly transformed. The mapping is shifted to a hidden dimension of 256. The decoder contains two fully connected layers: the first layer has an input dimension of 256 and an output dimension of 128, and the second layer has an input dimension of 128 and an output dimension of 1. The output is the interface quality prediction value for each test point. The interface quality assessment model employs a knowledge preservation mechanism based on continuous learning. A flexible weight solidification mechanism is used to calculate the sum of the squared gradients of each network parameter on historical tasks to obtain the diagonal elements of the Fisher information matrix, quantifying parameter importance. Penalty terms are applied to important parameters to constrain their update magnitude. An empirical replay mechanism is used to establish a historical sample buffer with a capacity of 10% of the total training dataset. Samples are randomly drawn from this buffer in each training batch. The samples are mixed with the current task samples at a ratio of 1:9 for training. When the Wasserstein distance between the data feature distributions of the new task and the historical task is greater than 0.5, a new parallel network branch is added to the model through a progressive network expansion mechanism. The new branch shares the first two convolutional layers of the encoder with the original branch, while the third convolutional layer and the sparse attention layer are independent. The decoder output is weighted and fused through learnable weights. A regularization term is added to the loss function, which is the L2 norm between the network parameters of the new branch and the original branch multiplied by a regularization coefficient of 0.01. The purpose of this step is to use the deep learning model to establish the mapping relationship between ultrasonic features and interface bonding strength and to realize the continuous learning capability of the model.

[0061] The specific implementation of step S09 involves fitting the predicted interface bond strength values ​​of all test points using a Weibull distribution. First, the predicted interface bond strength values ​​are arranged in ascending order. For the i-th predicted value, the cumulative failure probability is calculated as i divided by the total number of test points plus 1. The natural logarithm of each predicted value and its cumulative failure probability is taken, and then the natural logarithm of the cumulative failure probability is taken again to complete the double logarithmic transformation. A straight line is fitted using the natural logarithm of the predicted value as the abscissa and the double logarithm of the cumulative failure probability as the ordinate, obtaining the slope and intercept of the fitted line. The Weibull shape parameter is equal to the slope of the fitted line, and the Weibull characteristic strength is equal to the power of the natural constant e, with the power being the intercept of the fitted line divided by the negative of the Weibull shape parameter. An interface strength probability distribution function is established. The purpose of this step is to transform the discrete predicted interface bond strength values ​​into a continuous probability distribution model to describe the statistical characteristics of the interface strength. The Weibull distribution is suitable for describing the statistical distribution of material strength and can reflect the dispersion and reliability level of the strength.

[0062] The specific implementation of step S10 involves using the Monte Carlo method to perform random sampling simulation of the interface strength probability distribution function. Based on the interface strength probability distribution function, an inverse transformation sampling method is used to generate a random number sequence of length 10000. The number of values ​​less than the design strength requirement is counted in the random number sequence, and the failure probability is calculated as this number divided by 10000. The random number sequence is arranged in ascending order, and the 475th value is extracted as the lower limit of the strength confidence interval, and the 9525th value is extracted as the upper limit of the strength confidence interval. Corresponding to a 95% confidence level, the Weibull characteristic strength is... The Weibull shape parameters, lower limit of the strength confidence interval, upper limit of the strength confidence interval, and failure probability are used as the evaluation results and reliability indicators of the interface bonding strength. When the failure probability is greater than 0.05, the number of test points is increased and steps S07 to S09 are repeated. When the failure probability is less than or equal to 0.05, the final evaluation result is output. The purpose of this step is to evaluate the reliability level of the interface bonding strength and determine the confidence interval through Monte Carlo simulation. The failure probability threshold of 0.05 indicates that the maximum allowable failure risk is 5%. The random number sequence length of 10000 is the minimum sample size to ensure statistical accuracy.

[0063] It should be noted that the continuous changes in raw material batches, processing technology, and environmental conditions during the production of plant fiber reinforced plastics lead to dynamic evolution of interface properties. Traditional static models need to be completely retrained when faced with new production conditions and cannot retain the knowledge accumulated from historical production data. This invention protects network parameters that are important to historical tasks through an elastic weight solidification mechanism, strengthens memory by periodically reviewing historical samples through an experience replay mechanism, allocates dedicated network capacity for significantly different new tasks through a progressive network expansion mechanism, and maintains a flexible connection between new and old knowledge through regularization constraints. This mechanism enables the interface quality assessment model to have continuous learning capabilities, continuously optimize performance as production data accumulates without having to start from scratch, significantly reduce model maintenance costs, and improve the long-term stability of the evaluation system.

[0064] Specifically, the principle of this invention is as follows: This invention solves the technical problem of the difficulty in adaptively optimizing ultrasonic testing parameters. Its fundamental principle lies in transforming parameter optimization into a shortest path search problem in graph theory. By quantifying the spatial correlation and signal propagation similarity between candidate emission points, a weighted graph structure is constructed. Dijkstra's algorithm is used to find the optimal path from the sample center to each location. This path cost comprehensively reflects the representativeness of the spatial location and the stability of the signal quality. Therefore, the emission point corresponding to the node with the minimum path cost is necessarily the most suitable location as the detection benchmark globally. After determining the optimal emission point, this invention adopts a two-stage strategy combining coarse and fine search. First, a preliminary test with a large step size quickly identifies the parameter intervals corresponding to the shortest transmission signal propagation time and the minimum frequency attenuation. Then, a fine scan with a small step size is performed within these intervals. This strategy ensures global coverage of the parameter space to avoid missing optimal solutions, while improving the accuracy of parameter determination through local fine search. This allows the finally obtained emission direction angle and emission frequency to accurately match the acoustic characteristics of the sample at that location, thereby achieving adaptive optimization of ultrasonic testing parameters and ensuring accurate detection of heterogeneous plant fiber reinforced plastics.

[0065] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0066] The specific implementation methods of steps S01-S03 and S06 are the same as those described above, and will not be repeated in detail here.

[0067] The specific implementation of step S04 involves performing ultrasonic transmission tests on each set of initial ultrasonic emission parameters for each candidate emission point. The formula for calculating the frequency attenuation value of the transmitted signal is as follows:

[0068] ;

[0069] In the formula, For the first The candidate launch points are at the The frequency attenuation value of the transmitted signal at each candidate transmission frequency is a dimensionless quantity. , The total number of candidate launch sites, typically 400 to 500; This corresponds to 10 candidate transmission frequencies; For the first The transmit power spectral density of each candidate transmit frequency, in units of The first digit is extracted by performing a Fast Fourier Transform on the transmitted signal. The power spectral density corresponding to each candidate transmission frequency is obtained; For the first The candidate launch points are at the The received power spectral density at each candidate transmit frequency, in units of The first digit is extracted by performing a Fast Fourier Transform on the received signal. The power spectral density corresponding to each candidate transmission frequency is obtained; The maximum value of the power spectral density difference across all candidate emission points and frequencies, in units of Calculate by traversing all candidate transmission points and all candidate transmission frequencies The maximum value is then taken. This formula normalizes the transmission signal frequency attenuation values ​​of different candidate emission points to a range of 0 to 1, facilitating subsequent comparison and optimization.

[0070] The specific implementation of step S05 involves establishing an optimized ultrasonic parameter map. The formula for calculating the spatial Euclidean distance is as follows:

[0071] ;

[0072] In the formula, For the first The candidate launch point and the first The spatial Euclidean distance between the candidate launch points, in units of ; and ; The first The x-coordinate, y-coordinate, and y-coordinate of each candidate launch point, in units of... The result is obtained by measuring with a three-dimensional coordinate measuring instrument in step S02. The first The x-coordinate, y-coordinate, and y-coordinate of each candidate launch point, in units of... Acquisition method and same.

[0073] The formula for calculating the similarity of transmitted signals is as follows:

[0074] ;

[0075] In the formula, For the first The candidate launch point and the first The similarity of the transmitted signals between candidate emission points is a dimensionless quantity. For the first The transmitted signal amplitude of each candidate emission point, in units of The peak value is acquired by the peak detection circuit in step S04. For the first The transmitted signal amplitude of each candidate emission point, in units of Acquisition method and same; The maximum amplitude of the transmitted signal at all candidate emission points, in units of This is obtained by traversing all candidate launch points.

[0076] The formula for calculating the edge weights of a graph is as follows:

[0077] ;

[0078] In the formula, For the first The candidate launch point and the first The graph edge weights between candidate launch points are dimensionless. The maximum spatial Euclidean distance between all candidate launch point pairs, in units of Calculate by traversing all candidate launch points The maximum value is then taken. This formula comprehensively considers spatial location relationships and signal propagation characteristics. The weighting coefficients 0.6 and 0.4 represent the relative importance of spatial distance and signal similarity. The graph edge weights are formed by weighted summation. Connect the two corresponding candidate emission points in the ultrasonic parameter optimization diagram.

[0079] The formula for calculating path cost in the shortest path algorithm is as follows:

[0080] ;

[0081] In the formula, From the starting node to the The path cost of each candidate launch point is a dimensionless quantity. From the starting node to the The path cost of each candidate launch point is a dimensionless quantity; when Less than Updated regularly The value of is initialized. The starting node is the candidate emission point closest to the center of the sample. The initial path cost of other nodes is infinite. This formula achieves shortest path search through iterative updates, selecting the candidate launch point with the minimum path cost as the optimized launch point.

[0082] The specific implementation of step S07 involves performing ultrasonic transmission tests on all test points using the optimal emission point coordinates, optimal emission direction angle, and optimal emission frequency. The ultrasonic characteristic response matrix is ​​represented as follows:

[0083] ;

[0084] In the formula, This is the characteristic response matrix of the ultrasonic wave; This refers to the total number of test points, typically between 50 and 100. The first The x-coordinate, y-coordinate, and vertices of the optimal launch point corresponding to each test point, in units of... The optimized launch point coordinates are obtained through step S05. ; For the first The optimal launch direction angle corresponding to each test point, in degrees, is determined through step S06. For the first The optimal transmission frequency corresponding to each test point, in units of This is determined through step S06; For the first The amplitude of the transmitted signal at each test point, in units of The peak value is acquired through a peak detection circuit. For the first The transmission signal propagation time at each test point, in units of This is obtained by calculating the time difference between the transmission and reception times; For the first The transmission signal frequency attenuation value at each test point is a dimensionless quantity, and its calculation method is the same as in step S04. same; For the first The interface combined with the intensity prediction value of each test point, in units of... The ultrasonic characteristic response matrix is ​​obtained through the output of the interface quality assessment model in step S08. The number of rows in the matrix equals the number of test points, and the number of columns is 9. The matrix is ​​arranged in the order of the test point numbers to form a structured data matrix.

[0085] The specific implementation of step S08 involves processing the ultrasonic characteristic response matrix, sample temperature, wood flour particle size, and polyethylene molecular weight into the interface quality assessment model. The input feature vector is represented as follows:

[0086] ;

[0087] In the formula, The input feature vector; The temperature of the sample is in °C and is obtained in real time by a thermocouple temperature sensor during the ultrasonic transmission test. The particle size of the wood flour is expressed in units of 1000 mm. The particle size was obtained by measuring with a laser particle size analyzer. This represents the molecular weight of polyethylene, in units of... The values ​​were obtained by gel permeation chromatography. The input feature vector was mapped through a fully connected layer to obtain the initial weights of the query matrix, key matrix, and value matrix of the sparse attention layer.

[0088] The formula for calculating the diagonal elements of the Fisher information matrix is ​​as follows:

[0089] ;

[0090] In the formula, For the Fisher information matrix Diagonal elements, unit is loss function The unit is usually a dimensionless quantity; For the first Network parameters, , This represents the total number of network parameters. For the first The loss function for each historical task is a dimensionless quantity, calculated using mean square error. This represents the total number of historical tasks. loss function For network parameters The gradient is calculated using the backpropagation algorithm. This formula quantifies the importance of network parameters by summing the squares of the gradients from previous tasks. A larger value indicates higher network parameters. The more important it is to the historical mission.

[0091] The formula for calculating the elastic weight fixation penalty term is as follows:

[0092] ;

[0093] In the formula, For elastic weights, a fixed penalty term is applied; the unit and loss function are defined. They are the same, and are dimensionless quantities. When training for historical missions is completed The values ​​of several network parameters. This penalty term prevents catastrophic forgetting by limiting the update magnitude of important network parameters.

[0094] The formula for calculating the Wasserstein distance is as follows:

[0095] ;

[0096] In the formula, Let be the Wasserstein distance, and be a dimensionless quantity. This is the normalized data feature vector for the new task, and it is a dimensionless quantity. The normalized data feature vector of historical tasks is a dimensionless quantity. for and The set of all possible joint distributions between them; It is a joint distribution; For mathematical expectation operators; For normalized data feature vectors and The Euclidean distance between them is a dimensionless quantity. When It is determined that the new task is significantly different from the historical task.

[0097] The weighted fusion formula for the decoder output in progressive network expansion is expressed as follows:

[0098] ;

[0099] In the formula, The final output is a vector of interface-integrated intensity prediction values, in units of ; This is the output of the original network branch decoder, in units of ; The output of the new parallel network branch decoder, in units of ; These are learnable weights, dimensionless quantities, obtained through gradient descent optimization, with an empirical value of 0.5. They enable adaptive combination of the outputs of the original network branches and the new parallel network branches. The contribution ratio of the two branches is determined through training optimization. This formula enables the model to make full use of historical knowledge and new task knowledge, thereby improving prediction accuracy and generalization ability.

[0100] The formula for calculating the regularization term is as follows:

[0101] ;

[0102] In the formula, This is a regularization term, with the same unit as the loss function, and is a dimensionless quantity. is the regularization coefficient, a dimensionless quantity with a value of 0.01; The number of independent layers for the new parallel network branches; For the new parallel network branch The network parameter vector of the layer, ; For the original network branch corresponding to the first The network parameter vector of the layer; is the square of the L2 norm, and is a dimensionless quantity. This regularization term constrains the network parameters of the new parallel network branch to prevent them from deviating too much from those of the original network branch.

[0103] The interface quality assessment model adopts an encoder-decoder architecture. The encoder contains 3 convolutional layers and 2 sparse attention layers, and the decoder contains 2 fully connected layers. It outputs the predicted value of the interface binding strength for each test point.

[0104] The specific implementation of step S09 involves fitting a Weibull distribution to the interface combined strength prediction values ​​of all test points. The formula for calculating the cumulative failure probability is as follows:

[0105] ;

[0106] In the formula, For the arranged number The cumulative failure probability of each interface combined with the strength prediction value is a dimensionless quantity. This represents the total number of test points. The index is the sequence number of the interface combined intensity prediction values ​​arranged from smallest to largest. .

[0107] The expression for the fitted line of the Weibull distribution is:

[0108] ;

[0109] In the formula, The slope of the fitted line is denoted as , which is a dimensionless quantity and is obtained by fitting using the least squares method. The intercept of the fitted line is a dimensionless quantity, obtained by fitting using the least squares method; For the first Each interface is combined with the intensity prediction value, in units of .

[0110] The formulas for calculating the Weibull shape parameters and the Weibull characteristic intensity are expressed as follows:

[0111] ;

[0112] ;

[0113] In the formula, Let be the Weibull shape parameter, which is a dimensionless quantity; The intensity of the Weibull characteristic is expressed in units of 1. The Weibull characteristic strength represents the strength value corresponding to a cumulative failure probability of 63.2% for the interfacial bond strength, while the Weibull shape parameter represents the dispersion of the interfacial bond strength distribution. The Weibull distribution fitting correlation formula transforms the discrete interfacial bond strength predictions into a continuous probability distribution function, and the cumulative failure probability formula... Assigning failure probability to each strength value, Weibull shape parameter and Weibull feature strength The cumulative distribution function describes the dispersion and central tendency of the intensity distribution, respectively. A quantitative relationship between strength and failure probability is established. This formula allows for an accurate description of the statistical characteristics of interfacial bonding strength, providing a theoretical basis for reliability assessment.

[0114] The cumulative distribution function expression of the interface intensity probability distribution function is:

[0115] ;

[0116] In the formula, For interface bonding strength The corresponding cumulative failure probability is a dimensionless quantity; Interfacial bonding strength, in units of This formula describes the probability distribution characteristics of the interface bonding intensity, which is used in subsequent Monte Carlo simulations.

[0117] The specific implementation of step S10 involves using the Monte Carlo method to randomly sample and simulate the interface intensity probability distribution function. The formula for generating a random number sequence using the inverse transformation sampling method is as follows:

[0118] ;

[0119] In the formula, For the first A sequence of random number values, in units of ; is a uniformly distributed random number within the interval 0 to 1, and is a dimensionless quantity; This formula generates a random number sequence of length 10000.

[0120] The formula for calculating the failure probability is expressed as follows:

[0121] ;

[0122] In the formula, Let be the failure probability, and be a dimensionless quantity. For a random number sequence, the strength requirement is less than the design strength requirement. The number of values, The design value for interface bonding strength is set according to engineering application requirements, and the unit is... The value is typically between 20 and 30. When Increase the number of test points and repeat steps S07 to S09. The output interface combines the strength evaluation results and reliability indicators. The lower limit of the strength confidence interval is the 475th value after the random number sequence is sorted, and the upper limit of the strength confidence interval is the 9525th value, corresponding to a 95% confidence level.

[0123] In the steps of establishing the ultrasonic parameter sensitivity matrix, the formula for calculating the sensitivity of the optimal emission point's abscissa is as follows:

[0124] ;

[0125] In the formula, Sensitivity of the x-coordinate of the optimal launch point, in units of ; Add to the x-coordinate of the optimal launch point The change in amplitude of the transmitted signal after transmission, in units of It is obtained by comparing the amplitude of the transmitted signal before and after the increase; This represents the increment on the horizontal axis. Similarly, the sensitivity of the vertical axis at the optimal launch point... and vertical axis sensitivity The calculation method is the same, and the units are all the same. .

[0126] The formula for calculating the optimal launch direction angle sensitivity is as follows:

[0127] ;

[0128] In the formula, The optimal launch direction angle sensitivity, in units of Spend; Increase the angle for optimal launch direction The change in amplitude of the transmitted signal after transmission, in units of ; Degrees are increments of angles.

[0129] The formula for calculating optimal transmit frequency sensitivity is as follows:

[0130] ;

[0131] In the formula, The optimal transmission frequency sensitivity is expressed in units of... ; To increase the optimal transmission frequency The change in amplitude of the transmitted signal after transmission, in units of ; This represents the frequency increment.

[0132] The ultrasonic parameter sensitivity matrix is ​​represented as follows:

[0133] ;

[0134] In the formula, This is the sensitivity matrix for ultrasonic parameters; The representative number of test points is usually 10% of the total number of test points. , This is the floor function. The number of rows in this matrix corresponds to the number of representative test cases, and the number of columns is 5.

[0135] In the steps of establishing the Weibull distribution characteristic matrix, the sample surface is divided into 9 regions. The Weibull characteristic intensity, Weibull shape parameter, lower limit of the strength confidence interval, and failure probability of each region are arranged in order of region number. The Weibull distribution characteristic matrix is ​​represented as follows:

[0136] ;

[0137] In the formula, The characteristic matrix of the Weibull distribution; For the first The Weibull feature intensity of each region, in units of ; For the first The Weibull shape parameters of each region are dimensionless. For the first The lower limit of the intensity confidence interval for each region, in units of ; For the first The failure probability of each region is a dimensionless quantity. This matrix, with 9 rows and 4 columns, is used to evaluate the distribution characteristics of interfacial bonding strength in different regions of the sample. The formula reveals the spatial variability of interfacial bonding strength, providing data support for identifying weak areas and optimizing production processes, thus improving the uniformity and consistency of plant fiber reinforced plastic products. When the Weibull shape parameters differ significantly between different regions, it indicates that the interfacial bonding quality within the sample is uneven, requiring adjustment of processing parameters.

[0138] To better understand and implement this invention, the following is an example 2 of a specific application scenario: A technical team evaluated the interfacial bonding strength of a batch of wood-plastic composite materials. This batch of materials was prepared by mixing poplar wood powder and high-density polyethylene, with a wood powder content of 40% and a wood powder particle size of 120 mm. The molecular weight of polyethylene is The technical team conducted an interface bonding strength evaluation according to the technical solution of this invention. The specific implementation process is as follows.

[0139] First, the technical team prepared plant fiber reinforced plastic samples with dimensions of 200mm × 150mm × 10mm. 120 test points were selected on the sample surface using a uniform grid pattern, and their coordinate positions were marked. A spatial coordinate system was established with the lower left corner of the sample as the origin, and the horizontal axis as... The axis, the longitudinal direction is The axis is perpendicular to the sample surface and pointing upwards. A mesh was created on the sample surface with a mesh spacing of 5 mm. Candidate emission points were set at each mesh node, resulting in a total of 882 candidate emission points. An ultrasonic transmitter and receiver were then installed, with the receiver positioned on the bottom surface of the sample.

[0140] For each candidate emission point, the technical team set candidate emission direction angles every 15° within the interval [0°, 90°], obtaining a total of 7 candidate angles: 0°, 15°, 30°, 45°, 60°, 75°, and 90°. Candidate emission frequencies were set every 100kHz within the interval [100kHz, 1000kHz], obtaining a total of 10 candidate frequencies. These were combined to form 6174 sets of initial ultrasonic emission parameters. Preliminary ultrasonic transmission tests were performed on each set of initial ultrasonic emission parameters, collecting the transmitted signal amplitude, transmission signal propagation time, and transmission signal frequency attenuation value. The sample thermometer showed the sample temperature to be 23℃.

[0141] The technical team established an ultrasonic parameter optimization graph, using all 882 candidate emission points as graph nodes. The spatial Euclidean distance between any two candidate emission points was calculated, with a maximum spatial distance of 242.5 mm. When calculating the transmission signal similarity, it was found that the transmission signal amplitudes of candidate emission points 156 and 157 were 1.28V and 1.31V respectively, with an absolute difference of 0.03V. The maximum transmission signal amplitude of all candidate emission points was 2.45V, the normalized transmission signal difference was 0.0122, and the transmission signal similarity was 0.9878. The spatial Euclidean distance was normalized to the interval [0, 1], and the weighted sum of the normalized spatial Euclidean distance multiplied by 0.6 and the transmission signal similarity multiplied by 0.4 was used as the graph edge weight. When the graph edge weight was less than 0.7, the corresponding two candidate emission points were connected in the ultrasonic parameter optimization graph, resulting in a graph containing 3256 edges.

[0142] Dijkstra's shortest path algorithm was used to calculate the path cost from the sample center to each candidate emission point. The sample center coordinates were (100mm, 75mm, 0mm), and the nearest candidate emission point coordinates were (100mm, 75mm, 0mm), which was used as the starting node. The path cost from the starting node to itself was initialized to 0, and the path cost from the starting node to all other graph nodes was set to infinity. After iterative calculation, the candidate emission point with the minimum path cost was selected as the optimized emission point, with coordinates (95mm, 70mm, 0mm) and a path cost of 0.186.

[0143] To optimize the launch point location, the technical team conducted a fine-grained search based on the recorded propagation time and frequency attenuation of the transmitted signal. The candidate launch direction angle corresponding to the shortest propagation time was 45°. A fine-grained search was performed within the interval [38°, 52°] with a step size of 1°, testing a total of 15 angles. The optimal launch direction angle was determined to be 43°, corresponding to a propagation time of 32.6 seconds. The candidate transmission frequency corresponding to the minimum transmission signal frequency attenuation value is 500kHz. A fine search was performed in the interval [450kHz, 550kHz] with a step size of 10kHz. A total of 11 frequencies were tested, and the optimal transmission frequency was determined to be 490kHz, with a corresponding transmission signal frequency attenuation value of 8.2dB.

[0144] like Figure 2 As shown, the technical team conducted ultrasonic transmission tests on all 120 test points using the optimal transmission point coordinates (95mm, 70mm, 0mm), the optimal transmission direction angle of 43°, and the optimal transmission frequency of 490kHz. The amplitude of the transmitted signal, the propagation time of the transmitted signal, and the frequency attenuation value of the transmitted signal were recorded for each test point, and an ultrasonic characteristic response matrix was established. The ultrasonic characteristic response matrix has 120 rows and 9 columns. The ultrasonic characteristic response data for some test points are shown in Table 1.

[0145] Table 1. Ultrasonic characteristic response data at some test points

[0146]

[0147] The technical team used ultrasonic characteristic response matrix, sample temperature 23℃, and wood flour particle size 120 mm. and the molecular weight of polyethylene The input interface quality assessment model is processed. The model employs an encoder-decoder architecture. The encoder consists of three convolutional layers and two sparse attention layers. The first convolutional layer has a 3×3 kernel size and 64 output channels; the second and third convolutional layers each have a 3×3 kernel size and 128 output channels. The sparse attention layers contain 10 attention heads with a local window radius of 10. The decoder consists of two fully connected layers. The first fully connected layer has an input dimension of 256 and an output dimension of 128; the second fully connected layer has an input dimension of 128 and an output dimension of 1. The model outputs a predicted interface binding strength value for each test point, ranging from 18.3 MPa to 26.7 MPa.

[0148] like Figure 3 As shown, the technical team fitted the predicted interface bonding strength values ​​for all 120 test points using a Weibull distribution. The predicted interface bonding strength values ​​were then arranged in ascending order. The data for the [number]th [test point] was then processed. The cumulative failure probability of each interface is calculated by combining the strength prediction value. Divide by 121. Take the natural logarithm of each predicted interface bond strength value and its corresponding cumulative failure probability, and then take the natural logarithm of the cumulative failure probability to complete the double logarithmic transformation. Plot the natural logarithm of the predicted interface bond strength value on the x-axis and the double logarithm of the cumulative failure probability on the y-axis, and fit a straight line using the least squares method. The slope of the fitted line is 8.65, and the intercept is -25.48. The Weibull shape parameter is equal to 8.65, and the Weibull characteristic strength is 23.1 MPa. Establish the interface strength probability distribution function.

[0149] The technical team used the Monte Carlo method to simulate the interface strength probability distribution function through random sampling. Based on the interface strength probability distribution function, an inverse transformation sampling method was used to generate a random number sequence of length 10000. The design strength requirement was 19.5 MPa. In the random number sequence, 328 values ​​less than 19.5 MPa were counted. The failure probability was calculated to be 0.0328. The random number sequence was arranged in ascending order, and the 475th value was extracted as the lower limit of the strength confidence interval, which was 20.2 MPa. The 9525th value was extracted as the upper limit of the strength confidence interval, which was 25.8 MPa. Since the failure probability of 0.0328 is less than 0.05, the technical team output the interface strength based on the strength evaluation results and reliability indicators: a Weibull characteristic strength of 23.1 MPa, a Weibull shape parameter of 8.65, a strength confidence interval of 20.2 MPa to 25.8 MPa, and a failure probability of 0.0328.

[0150] The technical team further established an ultrasonic parameter sensitivity matrix. Twelve representative test points were randomly selected from 120 test points on the sample surface. For each representative test point, with the optimal emission direction angle fixed at 43° and the optimal emission frequency at 490kHz, the horizontal coordinate of the optimal emission point was increased by 1mm before ultrasonic transmission testing. The change in the amplitude of the transmitted signal was recorded, and the sensitivity of the horizontal coordinate of the optimal emission point was calculated. The above steps were repeated to calculate the sensitivity of the vertical coordinate of the optimal emission point. With the optimal emission point coordinates and optimal emission frequency fixed, the optimal emission direction angle was increased by 1° before ultrasonic transmission testing, and the sensitivity of the optimal emission direction angle was calculated. With the optimal emission point coordinates and optimal emission direction angle fixed, the optimal emission frequency was increased by 10kHz before ultrasonic transmission testing, and the sensitivity of the optimal emission frequency was calculated. The sensitivity data of all representative test points were arranged in order of their representative test point numbers to form an ultrasonic parameter sensitivity matrix with 12 rows and 5 columns.

[0151] The technical team established a Weibull distribution characteristic matrix. The sample surface was divided into three segments along both the length and width directions, forming nine regions. For each region, the predicted interfacial bonding strength values ​​of all test points within that region were extracted, and Weibull distribution fitting and Monte Carlo simulations were performed. Figure 4 As shown in Table 2, the Weibull distribution characteristics of the nine regions are presented.

[0152] Table 2 Characteristic Matrix of Weibull Distribution

[0153]

[0154] The interface quality assessment model used in this embodiment employs a knowledge retention mechanism based on continuous learning during training. The Wasserstein distance between the data feature distributions of the current batch of materials and historical batches is 0.38, which is less than 0.5, therefore the progressive network expansion mechanism is not triggered. The model calculates the Fisher information matrix of each network parameter for historical tasks through an elastic weight solidification mechanism. The update magnitude of highly important network parameters in this training is constrained by a penalty term. The capacity of the historical sample buffer is 10% of the total training dataset. In each training batch, samples are randomly drawn from the historical sample buffer and mixed with the current task samples at a ratio of 1:9 for training. After 200 training epochs, the model training is complete when the verification loss no longer decreases for 10 consecutive training epochs.

[0155] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for evaluating the interfacial bonding strength of plant fiber reinforced plastics, characterized in that, Plant fiber reinforced plastic samples were prepared by mixing wood flour with a polyethylene matrix. Multiple test points were selected on the sample surface and their coordinate positions were marked. A spatial coordinate system was established for the sample. A grid was created on the sample surface, and candidate emission points were set at each grid node. An ultrasonic transmitting and receiving device was set up. Multiple sets of initial ultrasonic emission parameters were set for each candidate emission point, and preliminary ultrasonic transmission tests were performed. The amplitude, propagation time, and frequency attenuation of the transmitted signal were collected. An ultrasonic parameter optimization graph was established. The weighted sum of spatial Euclidean distance and transmitted signal similarity was used as the graph edge weights. The path cost from the sample center to each candidate emission point was calculated, and a path was selected. The candidate emission point with the minimum cost is selected as the optimal emission point. A fine search is performed at the optimal emission point to determine the optimal emission direction angle and the optimal emission frequency. Ultrasonic transmission tests are performed on all test points using the optimal emission point coordinates, optimal emission direction angle, and optimal emission frequency. An ultrasonic characteristic response matrix is ​​established. The ultrasonic characteristic response matrix, sample temperature, wood flour particle size, and polyethylene molecular weight are input into the interface quality assessment model, and the interface bonding strength prediction value is output. The interface bonding strength prediction value is fitted with a Weibull distribution, and random sampling simulation is performed using the Monte Carlo method to calculate the strength confidence interval and failure probability. The interface bonding strength evaluation results and reliability index are output. The interface quality assessment model employs a knowledge retention mechanism based on continuous learning. It protects important parameters by solidifying them with elastic weights, maintains historical performance by utilizing experience replay, and uses progressive network expansion to allocate capacity for new tasks while constraining knowledge forgetting through regularization.

2. The method according to claim 1, characterized in that, The grid spacing is 5mm.

3. The method according to claim 2, characterized in that, The initial ultrasonic emission parameters are set specifically by setting the candidate emission direction angle every 15° within the interval [0°, 90°] for each candidate emission point, and setting the candidate emission frequency every 100kHz within the interval [100kHz, 1000kHz].

4. The method according to claim 3, characterized in that, The calculation of the transmission signal similarity is specifically as follows: for any two candidate emission points, the amplitude of the transmission signal is extracted, the absolute value of the difference between the amplitudes of the transmission signals is calculated, the absolute value of the difference between the amplitudes of the transmission signals is divided by the maximum value of the amplitudes of the transmission signals of all candidate emission points to obtain the normalized transmission signal difference, and the transmission signal similarity is obtained by subtracting the transmission signal difference from 1.

5. The method according to claim 4, characterized in that, The establishment of the ultrasonic parameter optimization graph specifically involves taking all candidate emission points as graph nodes, calculating the spatial Euclidean distance and transmission signal similarity between any two candidate emission points, and normalizing and fusing them as graph edge weights.

6. The method according to claim 5, characterized in that, The path cost is calculated using Dijkstra's shortest path algorithm.

7. The method according to claim 6, characterized in that, The determination of the optimal transmission direction angle is specifically to find the candidate transmission direction angle with the shortest transmission signal propagation time among all candidate transmission direction angles from the optimized transmission point, and to perform a fine search in the interval [the candidate transmission direction angle minus 7°, the candidate transmission direction angle plus 7°] with a step size of 1°, centered on the candidate transmission direction angle.

8. The method according to claim 7, characterized in that, The determination of the optimal transmission frequency specifically involves finding the candidate transmission frequency with the smallest transmission signal frequency attenuation value among all candidate transmission frequencies at the optimized transmission point, and then performing a fine search with a step size of 10kHz within the interval [the candidate transmission frequency minus 50kHz, the candidate transmission frequency plus 50kHz] centered on the candidate transmission frequency.

9. The method according to claim 8, characterized in that, The establishment of the ultrasonic characteristic response matrix is ​​specifically achieved by numbering all test points on the sample surface according to spatial coordinate order, performing ultrasonic transmission tests on each test point using the optimal emission point coordinates, optimal emission direction angle, and optimal emission frequency, and arranging the optimal emission point coordinates, optimal emission direction angle, optimal emission frequency, transmission signal amplitude, transmission signal propagation time, and transmission signal frequency attenuation value of each test point according to the test point number order to form an ultrasonic characteristic response matrix.

10. The method according to claim 9, characterized in that, The interface quality assessment model has an encoder-decoder architecture. The encoder contains three convolutional layers and two sparse attention layers. The input of the first convolutional layer is the ultrasonic feature response matrix, the kernel size is 3×3, and the number of output channels is 64. The kernel size of the second convolutional layer is 3×3, and the number of output channels is 128. The kernel size of the third convolutional layer is 3×3, and the number of output channels is 256. The decoder contains two fully connected layers.

Images