Steel plate crack positioning method based on convolutional neural network and particle swarm optimization algorithm

Abstract

The invention relates to the technical field of railway bridge crack detection and positioning, in particular to a steel plate crack positioning method based on a convolutional neural network and a particle swarm optimization algorithm, which comprises the following steps: drawing grid lines in a corresponding region of a steel plate, detecting acoustic emission signals of specific coordinate positions in the corresponding region of the steel plate based on a convolutional neural network model, obtaining a steel plate crack first positioning result; wherein the convolutional neural network model adopts a global mean pooling layer to execute average pooling operation on each channel in the two-dimensional feature map after convolution processing, and outputs a one-dimensional feature vector; or introducing an adjustable learning factor and an inertia weight into a particle swarm optimization algorithm, and carrying out traversal search on the acoustic emission signals of the corresponding areas of the steel plate in combination with the crossing thought of a genetic algorithm to obtain a second positioning result of the steel plate crack; and taking the first positioning result or the second positioning result as a final positioning result of the steel plate crack. According to the invention, automatic detection and accurate positioning of bridge fatigue cracks can be realized.

Description

Technical Field

[0001] The present invention relates to the technical field of railway bridge crack detection and positioning, and more specifically to a steel plate crack positioning method based on convolutional neural network and particle swarm optimization algorithm. Background Art

[0002] In terms of bridge fatigue crack detection and positioning, acoustic emission signal processing and CNN convolutional neural network are generally combined to perform feature recognition, learning and positioning of cracks, but there are some shortcomings. First, in terms of acoustic emission crack positioning, signal acquisition and extraction have a high professional and technical threshold, relying on expensive professional equipment and experienced operators, which makes the detection cost high and difficult to promote and apply. Secondly, in the structure of the convolutional neural network, the fully connected layer can capture the global information of the input signal, but in the case of a small amount of data in the database, a large number of parameters can easily lead to model overfitting and high computational costs. In addition, during the actual operation of railway bridges, full-time acoustic emission monitoring will be subject to complex environmental noise and interference signals. Since the bridge is affected by changes in factors such as load and temperature, it is difficult to build a training database. In addition, due to the limitations of environmental conditions, such as weather and light, the accuracy and stability of the detection are affected, and the complex crack morphology and location may be limited, making it difficult to achieve accurate positioning.

[0003] Therefore, how to realize the automated detection and precise positioning of bridge fatigue cracks is a problem that needs to be solved urgently by those skilled in the art. Summary of the invention

[0004] In view of this, the present invention provides a steel plate crack location method based on convolutional neural network and particle swarm optimization algorithm, which can realize automatic detection and precise location of bridge fatigue cracks.

[0005] In order to achieve the above object, the present invention adopts the following technical solution:

[0006] A steel plate crack location method based on convolutional neural network and particle swarm optimization algorithm comprises the following steps:

[0007] Collect the acoustic emission signals of the corresponding area of ​​the steel plate;

[0008] Draw grid lines in the corresponding area of ​​the steel plate, detect the acoustic emission signals at specific coordinate positions in the corresponding area of ​​the steel plate based on the convolutional neural network model, and obtain the first positioning result of the crack in the steel plate;

[0009] The convolutional neural network model uses multiple convolutional layers to perform convolution processing on the acoustic emission signal of the steel plate, and uses a global mean pooling layer to perform average pooling operation on each channel in the two-dimensional feature map after convolution processing, and outputs a one-dimensional feature vector; the one-dimensional feature vector is predicted to be a crack category, and the category prediction result is mapped to the corresponding coordinate position of the steel plate;

[0010] or:

[0011] In the particle swarm optimization algorithm, an adjustable learning factor and inertia weight are introduced, and the crossover idea of ​​the genetic algorithm is combined to traverse and search the acoustic emission signals in the corresponding area of ​​the steel plate to obtain the second location result of the steel plate crack.

[0012] The first positioning result or the second positioning result is used as the final positioning result of the steel plate crack.

[0013] Furthermore, the convolutional neural network model includes: an input layer, multiple convolutional layers, a pooling layer, a global mean pooling layer and an output layer;

[0014] The input data of the input layer is a one-dimensional acoustic emission signal of the corresponding area of ​​the steel plate;

[0015] The pooling layer is connected between the first convolution layer and the second convolution layer, and the other convolution layers are cascaded in sequence; the global mean pooling layer is connected to the output of the last convolution layer, and converts the two-dimensional feature map output by the last convolution layer into a one-dimensional feature vector;

[0016] The output layer uses a Softmax function to process the one-dimensional feature vector to obtain a first crack positioning result of a corresponding area of ​​the steel plate.

[0017] Furthermore, the configuration of multiple convolutional layers in the convolutional neural network model is as follows:

[0018] The first convolution layer sets 64 convolution kernels, each with a size of 16*1, a step size of 8, and a padding value of 12. After the one-dimensional acoustic emission signal is processed by the first convolution layer, the dimension of the generated feature map is 130*64.

[0019] After being processed by the LeakyRelu activation function and the batch normalization layer, the data is input into the pooling layer. The pooling layer is provided with 64 convolution kernels, each of which has a size of 3*1. The dimension of the feature map generated after being processed by the pooling layer is 43*64.

[0020] The second convolutional layer sets 96 convolution kernels, each with a size of 3*1, a step size of 1, and no zero padding. After being processed by the second convolutional layer, the dimension of the generated feature map is 41*96. After being activated by LeakyRelu and processed by the batch normalization layer, the data is passed to the third convolutional layer.

[0021] The third convolutional layer sets 128 convolution kernels, each with a size of 3*1, a step size of 2, and no zero padding. After being processed by the third convolutional layer, the dimension of the generated feature map is 20*128. After being activated by LeakyRelu and processed by the batch normalization layer, the data is passed to the fourth convolutional layer.

[0022] The fourth convolutional layer sets 192 convolution kernels, each with a size of 3*1, a step size of 1, and no zero padding. After being processed by the fourth convolutional layer, the dimension of the generated feature map is 18*192. After being activated by LeakyRelu and processed by the batch normalization layer, the data is passed to the fifth convolutional layer.

[0023] The fifth convolutional layer sets 256 convolution kernels, each of which has a size of 3*1, a step size of 2, and no zero padding. After processing by the fifth convolutional layer, the generated feature map dimension is 8*256. After LeakyRelu activation and batch normalization layer processing, the data is flattened and input into the global mean pooling layer. The output data dimension is 1*256.

[0024] Furthermore, when the convolutional neural network model is trained, the initial learning rate is 0.001, the batch size is 64, and the number of iterations is 50 rounds.

[0025] Furthermore, in the particle swarm optimization algorithm, the adjustable inertia weight expression is:

[0026]

[0027] in, For the tth 迭代次数 The inertia weight of the particle during iteration, ω min is the minimum inertia weight, ω max is the maximum inertia weight, T max is the maximum number of iterations.

[0028] Furthermore, in the particle swarm optimization algorithm, the expression of the adjustable learning factor is:

[0029]

[0030] Among them, c1 is the individual learning factor, c2 is the social learning factor,

[0031] Furthermore, in the particle swarm optimization algorithm, the crossover idea of ​​the genetic algorithm is combined to cross-exchange two particles in the group. The crossover formula is:

[0032]

[0033] in, represents particle i at t 迭代次数 The coordinate vector of represents particle j at t 迭代次数 The coordinate vector under , α is a constant value or a random number in the range of 0 to 1.

[0034] Furthermore, the value of α is determined by taking the ratio of the fitness function values ​​of the two particles:

[0035]

[0036] in, is particle i at t 迭代次数 The fitness function value under iteration, is particle j at t 迭代次数 The fitness function value under ;

[0037] When the ratio of the fitness function values ​​of two different particles meets the above conditions, the two particles will be cross-exchanged.

[0038] Furthermore, after collecting the acoustic emission signals of the corresponding area of ​​the steel plate, it also includes:

[0039] The median filter, wavelet transform or EEMD method is used to reduce the noise of the steel plate acoustic emission signal. It can be seen from the above technical solution that compared with the prior art, the present invention has the following beneficial effects:

[0040] 1. The present invention optimizes the structure of the convolutional neural network model, adopts a global mean pooling layer to replace the traditional fully connected layer and pooling layer, and performs an average pooling operation on each channel of the convolutional layer feature map, thereby converting the two-dimensional feature map into a one-dimensional feature vector. This conversion not only significantly reduces the number of model parameters, but also improves the model's adaptability to image size, can effectively prevent overfitting, and thus enhances the generalization ability of the model.

[0041] 2. The present invention improves the particle swarm optimization algorithm. On the one hand, by introducing functions such as variable learning factors and inertia weights, the search ability of particles is balanced. On the other hand, the genetic algorithm is combined with the particle swarm optimization algorithm, and the GA crossover idea is introduced to weaken the problem of the PSO algorithm falling into local convergence early, and enrich the diversity of particles in the algorithm. Based on the improved PSO algorithm, even under the influence of complex internal properties of steel plate materials and environmental factors such as noise and temperature, the location of cracks can be effectively searched. By continuously optimizing the search process, the automatic detection and precise positioning of bridge fatigue cracks can be achieved, which reduces labor costs and detection time, and improves the accuracy and reliability of detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0043] Figure 1 A flow chart of a steel plate crack location method based on a convolutional neural network and a particle swarm optimization algorithm provided by the present invention;

[0044] Figure 2 A schematic diagram of the change in accuracy during the training of the convolutional neural network model provided by the present invention;

[0045] Figure 3 A schematic diagram of the change in the loss function value during the training of the convolutional neural network model provided by the present invention;

[0046] Figure 4 A diagram of the division of the pre-segment point area provided by the present invention;

[0047] Figure 5 This is a schematic diagram of the regional diagnosis results of lead break points based on the convolutional neural network model;

[0048] Figure 6 The present invention provides a point position error diagram for positioning the original signal;

[0049] Figure 7 This is a positioning point error map after median filtering and noise reduction provided by the present invention;

[0050] Figure 8 This is the positioning point error map after wavelet denoising provided by the present invention;

[0051] Fig. 9 This is the positioning point error diagram after EEMD noise reduction provided by the present invention. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] The embodiment of the present invention discloses a steel plate crack location method based on a convolutional neural network and a particle swarm optimization algorithm, comprising the following steps:

[0054] Collect the acoustic emission signals of the corresponding area of ​​the steel plate;

[0055] Draw grid lines in the corresponding area of ​​the steel plate, detect the acoustic emission signal of the specific coordinate position in the corresponding area of ​​the steel plate based on the convolutional neural network model, and obtain the first positioning result of the crack of the steel plate; wherein the convolutional neural network model uses multiple convolutional layers to perform convolution processing on the acoustic emission signal of the steel plate, and uses a global mean pooling layer to perform an average pooling operation on each channel in the two-dimensional feature map after the convolution processing, and outputs a one-dimensional feature vector; predict the crack category of the one-dimensional feature vector, and map the category prediction result to the corresponding coordinate position of the steel plate;

[0056] or:

[0057] In the particle swarm optimization algorithm, an adjustable learning factor and inertia weight are introduced, and the crossover idea of ​​the genetic algorithm is combined to traverse and search the acoustic emission signals in the corresponding area of ​​the steel plate to obtain the second location result of the steel plate crack.

[0058] The first positioning result or the second positioning result is used as the final positioning result of the steel plate crack.

[0059] Among them, when detecting the acoustic emission signals of the corresponding areas of the steel plate based on the convolutional neural network model, the Nielsen lead breaking experiment is first designed to construct a low-cost and high-quality training data set; then, the convolutional neural network model is built to optimize the model structure and hyperparameters; finally, fatigue crack detection and positioning of the railway bridge are carried out based on the trained model.

[0060] The present invention adopts a global average pooling layer, which can greatly reduce the number of parameters compared to the fully connected layer, and helps to reduce overfitting. The global average pooling layer converts the two-dimensional feature map into a one-dimensional feature vector by performing an average pooling operation on each channel of the convolutional layer feature map. This conversion not only significantly reduces the number of model parameters, but also improves the model's adaptability to image size, which is an effective strategy to prevent overfitting.

[0061] Compared with traditional pooling layers, a significant advantage of global mean pooling layers is their globality. Traditional pooling layers reduce feature dimensions by maximizing or averaging operations in local neighborhoods, while global mean pooling layers retain more global information by processing the entire feature map. This globality enables the global mean pooling layer to capture more comprehensive contextual information in the image, which is particularly important for tasks such as image classification and object detection.

[0062] In comparison with fully connected layers, the global mean pooling layer shows its structural superiority. The fully connected layer is usually located at the end of the CNN architecture and is used to convert feature maps into the final output vector. Although the fully connected layer plays a key role in the model, it also introduces a large number of parameters, increasing the complexity of the model and the risk of overfitting. In contrast, the global mean pooling layer significantly reduces the number of parameters by simplifying the feature conversion process, while also improving the model's adaptability to different input sizes, thereby enhancing the model's generalization ability. The use of the global mean pooling layer also promotes the interpretability of the model. By mapping the prediction of each category back to a specific area of ​​the input, researchers can more easily understand and explain the decision-making process of the model. This feature is crucial to improving the transparency and credibility of the model, especially in application scenarios that require a high degree of interpretability.

[0063] Specifically, the convolutional neural network model includes: an input layer, five convolutional layers, a pooling layer, a global mean pooling layer, and an output layer; the pooling layer is connected between the first convolutional layer and the second convolutional layer, and the other convolutional layers are cascaded in sequence; the global mean pooling layer is connected to the output of the last convolutional layer, and the two-dimensional feature map output by the last convolutional layer is converted into a one-dimensional feature vector; the output layer is connected to the last convolutional layer, and outputs the first crack positioning result of the corresponding area of ​​the steel plate. The parameters of each layer are shown in Table 1.

[0064] Table 1

[0065]

[0066]

[0067] The input data of the input layer is a one-dimensional acoustic emission signal of the corresponding area of ​​the steel plate, and the dimension is set to 1024x1.

[0068] The first convolution layer sets 64 convolution kernels, each with a size of 16*1, a step size of 8, and a padding value of 12. After the one-dimensional acoustic emission signal is processed by the first convolution layer, the dimension of the generated feature map is 130*64.

[0069] After being processed by the LeakyRelu activation function and the batch normalization layer, the data is input into the pooling layer. The pooling layer is provided with 64 convolution kernels, each of which has a size of 3*1. The dimension of the feature map generated after being processed by the pooling layer is 43*64.

[0070] The second convolutional layer sets 96 convolution kernels, each with a size of 3*1, a step size of 1, and no zero padding. After being processed by the second convolutional layer, the generated feature map has a dimension of 41*96. After being activated by LeakyRelu and processed by the batch normalization layer, the data is passed to the third convolutional layer.

[0071] The third convolutional layer is set with 128 convolution kernels, each with a size of 3*1, a step size of 2, and no zero padding. After being processed by the third convolutional layer, the dimension of the generated feature map is 20*128. After being activated by LeakyRelu and processed by the batch normalization layer, the data is passed to the fourth convolutional layer.

[0072] The fourth convolutional layer sets 192 convolution kernels, each with a size of 3*1, a step size of 1, and no zero padding. After being processed by the fourth convolutional layer, the generated feature map has a dimension of 18*192. After being activated by LeakyRelu and processed by the batch normalization layer, the data is passed to the fifth convolutional layer.

[0073] The fifth convolutional layer sets 256 convolution kernels, each of which has a size of 3*1, a step size of 2, and no zero padding. After processing by the fifth convolutional layer, the generated feature map dimension is 8*256. After LeakyRelu activation and batch normalization layer processing, the data is flattened and input into the global mean pooling layer. The output data dimension is 1*256.

[0074] Afterwards, during training, the output layer uses the Softmax function to process the data to obtain 51 probability values, calculates the loss function with the output 51 probability values ​​and the true labels, and then updates the model weights through back propagation, and finally saves the best model.

[0075] In the positioning based on the convolutional neural network model, CNN automatically extracts the features of the input data (acoustic emission signals) through its multi-layer structure. These features are gradually extracted through consecutive convolutional layers, activation layers, and pooling layers to obtain high-level representations of the signals. The positioning logic is based on classification. During training, the network learns how to map the extracted features to predefined categories. Each category corresponds to a specific location or range of locations. The training of CNN is usually supervised and requires a large amount of labeled data (i.e., acoustic emission signals and their corresponding precise source locations).

[0076] The loss function is an operational function used to measure the degree of difference between the model's predicted value f(x) and the true value Y. It is a non-negative real-valued function, usually expressed as L(Y,f(x)). A smaller loss function value means that the model has better robustness. The loss function is mainly used in the training phase of the model. In each training batch, the training data is input into the model for forward propagation to generate a predicted value. Then, the loss function calculates the difference between the predicted value and the true value, that is, the loss value. Through the back-propagation algorithm, the model uses the loss value to update each parameter to reduce the difference between the true value and the predicted value, so that the model's predicted value gradually approaches the true value, thereby achieving the goal of learning. Choosing an appropriate loss function is crucial for model training and optimization. Different tasks and model types may require different loss functions. Commonly used loss functions are the mean square error function and the cross-entropy loss function. The specific formula is shown below.

[0077]

[0078] Among them, y m represents the true value, represents the model prediction value, M represents the number of categories, and N represents the number of samples; y ic The value range is 0 or 1. Specifically, y is equal to 0 if and only if the model prediction category of sample i is exactly the same as its true category c. ic The value of is set to 1, otherwise it is 0. ic Represents the predicted probability value of the model for sample i belonging to category c. Considering the complexity and diversity of classification situations in the field of fault diagnosis, the present invention adopts the cross entropy loss function. This choice is based on the superior performance of the cross entropy loss function in dealing with multi-category problems, especially its ability to provide detailed and targeted error feedback for model learning, thereby effectively improving the accuracy and reliability of fault diagnosis.

[0079] In order to speed up the improvement of the accuracy of the diagnostic model, the present invention selects key hyperparameters such as learning rate, batch size and number of iterations. If the learning rate is small, it may lead to slow convergence, and if the learning rate is large, the model may fall into a local optimal solution. Through numerical calculation methods, the optimizer can obtain the network parameters that minimize the loss function. A larger batch size will increase the training time, but a smaller batch size may cause overfitting. The initial learning rate of the present invention is set to 0.001, the Adam optimizer is used, and the batch size is set to 64 and the number of iterations is 50 rounds.

[0080] The present invention works by simulating the dynamics of a group of particles in the search space of potential solutions in the location detection of steel plate cracks based on the particle swarm optimization algorithm. Each particle represents a potential position solution and is optimized according to its performance (fitness function). In terms of iterative updates, particles update their positions based on their best historical positions and the best positions in the group. This update is based on the fitness evaluation during the optimization process. Unlike CNN, the particle swarm algorithm does not require a training data set. It works directly on the fitness function, which measures the quality of each position solution.

[0081] Due to the complex internal properties of steel plate materials and the influence of environmental factors such as noise and temperature, signal acquisition positioning errors are unavoidable, and the positioning error can only be weakened as much as possible to improve positioning accuracy. To this end, the present invention proposes an improved particle swarm optimization algorithm, which uses the improved algorithm to traverse and search the area to find the optimal solution as the estimated value of the real coordinate point of the broken lead.

[0082] The specific steps include:

[0083] (1) Particle swarm initialization

[0084] m particles are randomly initialized in the pre-set target area, where each particle represents a potential solution, that is, the location of an acoustic emission signal. The initial population is set to P0, and the initial iteration number is t 迭代次数 =0, maximum number of iterations T max The optimal solution pbest of particle i in the position it passes through i , particle swarm optimal solution gbest.

[0085] (2) Determine the fitness function

[0086] Mean Absolute Error (MAE) and Mean Squared Error (MSE) are two commonly used indicators to measure the performance of prediction models. Compared with MSE, MAE pays more attention to the absolute difference between the predicted value and the true value. It can calculate the average value of the absolute error between the predicted value and the true value, which can reflect the average degree of deviation of the predicted value from the true value. The present invention adopts MAE, and its calculation formula is:

[0087]

[0088] is the real coordinate value of the lead break position, y iis the predicted coordinate value; in the process of minimizing the error value, the estimated coordinate value will approach its true coordinate value. The fitness function value is inversely proportional to the position estimation result. The smaller the function value, the smaller the distance error between the particle and the actual position, and thus the better the position estimation result. In the initialization phase, each particle is evaluated to determine its fitness level, where the initial position of each particle is regarded as its individual best position. Then, all particles are sorted in ascending order according to the fitness level, and the particle with the lowest fitness is identified as the best performance, and its position becomes the optimal solution of the entire group, which is marked as gbest. In the subsequent iterative update process, the current fitness value of each particle is compared with its previously recorded best value, and its speed and position are updated. If the current value is found to be better, that is, the fitness is lower, then the record of the particle will be updated. At the same time, based on the latest fitness value of the particle, the entire group is sorted again to ensure that the optimal solution gbest of the group can be updated in real time.

[0089] The MAE fitness function is applicable when the artificial acoustic emission source is constructed by collecting the broken lead signal and the real coordinates of the broken lead signal are known. In the actual bridge fatigue crack detection, MAE is not applicable because the real position of the crack is unknown. Maximum likelihood estimation (MLE) is a method for estimating model parameters based on observed data. In the actual bridge fatigue crack detection, it can be applied to the process of locating cracks based on acoustic emission signals or other sensor data. The key to this method is to find a set of parameters (crack locations) so that the probability (likelihood) of the observed data is maximized given this set of parameters.

[0090] (3) Particle velocity and position update

[0091] When the algorithm does not reach the termination condition, the speed and position of the particles in the particle swarm need to be updated. The relevant formula is as follows:

[0092]

[0093] In the formula, Indicates t 迭代次数 +1 The particle velocity vector at iteration 1. Indicates the tth 迭代次数 The particle velocity vector at the iteration. represents the inertia weight, c1, c2 are learning factors, and r1, r2 are random numbers in the range of (0, 1). represents particle i at t 迭代次数 The coordinate vector below.

[0094] According to the above formula, the change of particle velocity is affected by the inertia weight and the learning factor. This will result in faster particle exploration, increasing its ability to conduct extensive searches. It will slow down the exploration speed and make the local search more detailed. Choosing a suitable inertia weight is crucial to maintain a balance between extensive search and precise search. The learning factor also affects the particle search speed.

[0095] The present invention improves the particle swarm optimization algorithm from two aspects. On the one hand, by introducing variable learning factors, inertia weight and other functions, the search ability of particles is balanced. The adjustable inertia weight expression is:

[0096]

[0097] in, For the tth 迭代次数 The inertia weight of the particle during iteration; ω min is the minimum inertia weight, ω max is the maximum value of inertia weight, usually ω max =0.9,ω min =0.4; T max is the maximum number of iterations.

[0098] It can be clearly seen from the above formula that the larger the result in the formula, that is, when the number of iterations is small, the particles have a stronger search speed and search ability, which is obviously beneficial to the global search; in the later stage of iteration, when the particles are in their respective optimal positions in the region, the result in the formula becomes smaller, that is, the particles have a more detailed local search ability, which is obviously beneficial to the optimization convergence of the algorithm.

[0099] The learning factor also affects the particle search speed. In the particle swarm optimization algorithm, each particle represents a potential solution in the problem space, and it adjusts its search direction and speed based on individual experience and group experience. The learning factor is an important parameter for adjusting this process, which usually includes two main factors: individual learning factor c1 and social learning factor c2.

[0100] Among them, the adjustable learning factor formula is as follows:

[0101]

[0102] in, c1 is the individual learning factor, which affects the ability of particles to adjust their speed based on their own historical best positions. Larger values ​​make particles more dependent on their own experience, which helps particles to search more carefully in the good areas they find themselves. This may increase local search capabilities, but if it is too high, it may cause particles to be trapped in local optima. c2 is the social learning factor, which determines the extent to which particles rely on group experience when adjusting their speed. Larger values ​​encourage particles to move toward the currently known global best position, increasing the ability of the particle group as a whole to explore the search space. This helps the algorithm converge to a good solution quickly, but if it is too large, it may lead to premature convergence and miss a better solution.

[0103] On the other hand, the genetic algorithm (GA) is combined with the particle swarm optimization algorithm, and the GA crossover idea is introduced to reduce the problem of the PSO algorithm falling into local convergence prematurely and enrich the diversity of particles in the algorithm. The genetic algorithm is an optimization algorithm inspired by the evolutionary process in nature. It simulates the process of biological evolution and searches for the optimal solution through genetic operations on individuals (such as selection, crossover and mutation). Its basic idea is to evolve from generation to generation through competition and cooperation between individuals in the population, and finally find the solution with the highest fitness. In the genetic algorithm (GA), chromosomes are visualized through the encoding process, and new encoding entities are generated through a partial exchange mechanism. Different from the GA approach, the particle swarm optimization algorithm adopts a more abstract particle concept and does not involve specific encoding steps. Therefore, in the particle swarm optimization algorithm, the crossover operation principle of the genetic algorithm is introduced. By selecting two specific particles i and j in the population, the crossover formula is as follows:

[0104]

[0105] in, represents particle i at t 迭代次数 The coordinate vector of represents particle j at t 迭代次数 The coordinate vector under , α is a constant value or a random number in the range of 0 to 1.

[0106] The value of α is determined by taking the ratio of the fitness function values ​​of the two particles:

[0107]

[0108] in, is particle i at t 迭代次数 The fitness function value under is particle j at t 迭代次数 The fitness function value under ;

[0109] When the ratio of the fitness function values ​​of two different particles meets the above conditions, the two particles will crossover. This method helps to speed up the evolution of low-fitness particles and slow down the evolution of high-fitness particles, thereby achieving a balance in the rate of particle evolution and enhancing the diversity among particles.

[0110] The convolutional neural network model of the present invention is based on the Pytorch 1.8.0 deep learning framework and is implemented on the Windows 10 operating system. The hardware configuration used includes an NVIDIA GTX1050ti graphics card, an Intel i7

[0111] 12400F processor, 24G memory.

[0112] Next, the performance of the above method of the present invention is verified.

[0113] 1. Verification of the positioning of broken lead acoustic emission signals based on convolutional neural networks.

[0114] The accuracy of experimental results and learning efficiency are important indicators for evaluating model performance. Figure 2 It shows the change of accuracy during model training. Figure 3 It represents the changing trend of the loss function value in the training step. Figure 2 It can be seen that as the number of training cycles increases, the accuracy of the model rises rapidly to a high level close to 1.0, and stabilizes after the initial few cycles. This trend shows that the model can quickly learn the characteristics of the acoustic emission signal in the early stage of training, and the later training consolidates the accuracy of the model with almost no signs of overfitting. The high level of model accuracy proves the effectiveness of convolutional neural networks in dealing with the problem of locating broken lead acoustic emission signals; Figure 3 The loss function curve in shows that the loss value during the training process gradually decreases and stabilizes in a lower value range, which is a normal performance of the model during the optimization process. In the early stage of training, the loss value drops sharply, indicating that the model quickly adapts to the distribution of the data set. As the number of training steps increases, the curve tends to be smooth, indicating that the model has undergone gradual adjustment and optimization during this period, the rate of loss value reduction slows down, and the model begins to converge. The lower stable loss value indicates that the model has good generalization ability and small prediction error.

[0115] In general, the accuracy curve and the loss function curve jointly verify that the proposed convolutional neural network model has demonstrated efficient learning ability and high stability on the lead-breaking acoustic emission signal training set. The curves of these two performance indicators provide a quantitative basis for evaluating model performance, and also lay the foundation for subsequent model verification, testing and deployment in practical applications.

[0116] After that, the test set and network model are loaded to test the actual effect of the model. All the lead-breaking test points are divided into four areas: ABCD. Figure 4 shown.

[0117] Taking the lower left corner of the region division map as the origin (0,0) and the upper right corner as (10,4), each coordinate in the confusion matrix can be mapped to a specific spatial location. This mapping is crucial for analyzing how the model processes signals in a specific area. Each entry in the confusion matrix represents the comparison between the model's prediction and the true value, such as Figure 5 The values ​​on the diagonal represent the number of correctly located signals, while the values ​​on the off-diagonal represent the cases where the model incorrectly locates the signals to other coordinates.

[0118] The analysis reveals the model's ability to locate in a specific area. In particular, the classification accuracy at the edge of the area is of particular concern, as signals in the edge area may be more likely to be misclassified due to feature similarity. Evaluate the accuracy of the acoustic emission signal localization model. The confusion matrix data from area A to area D shows that most signals can be accurately located to their corresponding coordinates. For example, in the confusion matrix of area A, the signal at coordinate (0,0) is almost completely correctly located. However, there are a few errors, such as one signal at coordinate (0,1) being mislocated to coordinate (0,3), which may be due to the slight difference in the acoustic characteristics of these two locations. When looking closer to the boundary, the error increases slightly, which is particularly evident in the confusion matrices of areas C and D. For example, in the confusion matrix of area C, one signal at coordinate (6,0) was misclassified as coordinate (6,1). This means that the localization accuracy at the edge of the area may be affected by the similarity of acoustic emission signal characteristics between areas. For area D, the confusion matrix shows more misclassifications at coordinates (9,1) and (9,2). This may be because the signal characteristics in this area are more similar to those in other locations. Nevertheless, most signals can still be accurately located, and the model's positioning accuracy is about 98.8235%, confirming the model's excellent performance in positioning.

[0119] 2. Verification of positioning performance based on improved PSO algorithm

[0120] By collecting the acoustic emission signal of the broken lead and implementing three different noise reduction methods, the noise-reduced signal is obtained. Subsequently, these four sets of signals are used as the input of the PSO algorithm to predict the broken lead point. The simulation environment is a 500mm×200mm rectangular area constructed by python. The area is divided into 10×4 squares, with a total of 55 intersections. 9 real points are randomly selected in this area. The material parameters are taken as steel density 7,850kg / m 3, the elastic modulus is 210GPa. In the PSO algorithm, the initial population size is set to M = 100, the maximum number of iterations is set to Q = 3000, and the experimental results are obtained by positioning the point map (such as Figures 6 to 9 ) and the standard deviation data of each group of signal predicted position (predictedposition) and true position (true position) (as shown in Table 2) for characterization and analysis.

[0121] Table 2 Prediction differences and standard deviations of four positioning points

[0122]

[0123] It can be observed from the positioning error graph that the original signal without noise reduction processing has the largest positioning error. This may be because the original signal contains a lot of random noise and interference components, which affect the judgment of the positioning algorithm. The signal processed by median filtering shows a certain degree of improvement. As a nonlinear filtering method, median filtering can remove noise and make the signal smoother, thereby helping to improve positioning accuracy.

[0124] The signal denoised by wavelet transform shows more concentrated positioning error points, indicating that wavelet transform can better separate noise and signal itself when decomposing the signal, enhancing the extraction of useful information in the signal. Finally, the positioning error graph of the EEMD method shows the smallest error fluctuation, reflecting that EEMD can more accurately decompose noise when processing complex signals, retain the core characteristics of the signal, and has a significant auxiliary effect on the positioning algorithm.

[0125] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0126] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A steel plate crack location method based on convolutional neural network and particle swarm optimization algorithm, characterized in that: The following steps are involved: Collect the acoustic emission signals of the corresponding area of ​​the steel plate; Draw grid lines in the corresponding area of ​​the steel plate, detect the acoustic emission signals at specific coordinate positions in the corresponding area of ​​the steel plate based on the convolutional neural network model, and obtain the first positioning result of the crack in the steel plate; The convolutional neural network model uses multiple convolutional layers to perform convolution processing on the acoustic emission signal of the steel plate, and uses a global mean pooling layer to perform average pooling operation on each channel in the two-dimensional feature map after convolution processing, and outputs a one-dimensional feature vector; the one-dimensional feature vector is predicted to be a crack category, and the category prediction result is mapped to the corresponding coordinate position of the steel plate; or: In the particle swarm optimization algorithm, an adjustable learning factor and inertia weight are introduced, and the crossover idea of ​​the genetic algorithm is combined to traverse and search the acoustic emission signals in the corresponding area of ​​the steel plate to obtain the second location result of the steel plate crack. The first positioning result or the second positioning result is used as the final positioning result of the steel plate crack.

2. The steel plate crack location method based on convolutional neural network and particle swarm optimization algorithm according to claim 1 is characterized in that: The convolutional neural network model includes: an input layer, multiple convolutional layers, a pooling layer, a global mean pooling layer and an output layer; The input data of the input layer is a one-dimensional acoustic emission signal of the corresponding area of ​​the steel plate; The pooling layer is connected between the first convolution layer and the second convolution layer, and the other convolution layers are cascaded in sequence; the global mean pooling layer is connected to the output of the last convolution layer, and converts the two-dimensional feature map output by the last convolution layer into a one-dimensional feature vector; The output layer uses a Softmax function to process the one-dimensional feature vector to obtain a first crack positioning result of a corresponding area of ​​the steel plate.

3. The steel plate crack location method based on convolutional neural network and particle swarm optimization algorithm according to claim 2 is characterized in that: The settings of multiple convolutional layers in the convolutional neural network model are as follows: The first convolution layer sets 64 convolution kernels, each with a size of 16*1, a step size of 8, and a padding value of 12. After the one-dimensional acoustic emission signal is processed by the first convolution layer, the dimension of the generated feature map is 130*64. After being processed by the LeakyRelu activation function and the batch normalization layer, the data is input into the pooling layer. The pooling layer is provided with 64 convolution kernels, each of which has a size of 3*1. The dimension of the feature map generated after being processed by the pooling layer is 43*64. The second convolutional layer sets 96 convolution kernels, each with a size of 3*1, a step size of 1, and no zero padding. After being processed by the second convolutional layer, the dimension of the generated feature map is 41*96. After being activated by LeakyRelu and processed by the batch normalization layer, the data is passed to the third convolutional layer. The third convolutional layer sets 128 convolution kernels, each with a size of 3*1, a step size of 2, and no zero padding. After being processed by the third convolutional layer, the dimension of the generated feature map is 20*128. After being activated by LeakyRelu and processed by the batch normalization layer, the data is passed to the fourth convolutional layer. The fourth convolutional layer is set with 192 convolution kernels, each with a size of 3*1, a step size of 1, and no zero padding; After being processed by the fourth convolutional layer, the generated feature map dimension is 18*192. After being processed by the LeakyRelu activation and batch normalization layer, the data is passed to the fifth convolutional layer; The fifth convolutional layer sets 256 convolution kernels, each of which has a size of 3*1, a step size of 2, and no zero padding. After processing by the fifth convolutional layer, the generated feature map dimension is 8*256. After LeakyRelu activation and batch normalization layer processing, the data is flattened and input into the global mean pooling layer. The output data dimension is 1*256.

4. The steel plate crack location method based on convolutional neural network and particle swarm optimization algorithm according to claim 1, characterized in that: When training the convolutional neural network model, the initial learning rate is 0.001, the batch size is 64, and the number of iterations is 50 rounds.

5. The steel plate crack location method based on convolutional neural network and particle swarm optimization algorithm according to claim 1, characterized in that: In the particle swarm optimization algorithm, the adjustable inertia weight expression is: in, For the tth 迭代次数 The inertia weight of the particle during iteration, ω min is the minimum inertia weight, ω max is the maximum inertia weight, T max is the maximum number of iterations.

6. The steel plate crack location method based on convolutional neural network and particle swarm optimization algorithm according to claim 5, characterized in that: In the particle swarm optimization algorithm, the expression of the adjustable learning factor is: Among them, c1 is the individual learning factor, c2 is the social learning factor, 7. The steel plate crack location method based on convolutional neural network and particle swarm optimization algorithm according to claim 1, characterized in that: In the particle swarm optimization algorithm, the crossover idea of ​​the genetic algorithm is combined to cross-exchange two particles in the group. The crossover formula is: Among them, X represents particle i at t 迭代次数 The coordinate vector under iteration, represents particle j at t 迭代次数 The coordinate vector under , α is a constant value or a random number in the range of 0 to 1.

8. The steel plate crack location method based on convolutional neural network and particle swarm optimization algorithm according to claim 7, characterized in that: The value of α is determined by taking the ratio of the fitness function values ​​of the two particles: in, is particle i at t 迭代次数 The fitness function value under is particle j at t 迭代次数 The fitness function value under ; When the ratio of the fitness function values ​​of two different particles meets the above conditions, the two particles will be cross-exchanged.

9. The steel plate crack location method based on convolutional neural network and particle swarm optimization algorithm according to claim 1, characterized in that: After collecting the acoustic emission signals of the corresponding area of ​​the steel plate, it also includes: The median filter, wavelet transform or EEMD method is used to reduce the noise of the steel plate acoustic emission signal.

Images