Acoustic emission source positioning method based on artificial intelligence deep learning

Abstract

The invention provides a sound emission source positioning method based on artificial intelligence deep learning, and aims to solve the problems that a traditional sound emission source positioning method is low in positioning precision in a complex structure and is easily influenced by sound wave propagation characteristics. The method comprises the steps of grid division, acoustic emission sensor arrangement, data acquisition, artificial intelligence deep learning neural network model construction, data preprocessing, label data generation and model training, model verification, acoustic emission source positioning and the like. By constructing an artificial intelligence deep learning neural network model comprising an input layer, at least two hidden layers capable of dynamically adjusting the number of nodes and an output layer, multi-dimensional features of acoustic emission signals are extracted and analyzed, and high-precision positioning of an acoustic emission source in a complex structure is achieved. The method does not need to depend on a complex mathematical analysis algorithm, has the advantages of being high in positioning precision, robustness and adaptability, high in automation degree and the like, is suitable for sound emission source positioning of various complex structures such as a welding structure, a metal material and a composite material, and provides powerful support for structural health monitoring and safety evaluation.

Description

Technical Field

[0001] The present invention relates to the field of nondestructive testing technology, and in particular to an acoustic emission source positioning method based on artificial intelligence deep learning. The method uses acoustic emission signals and deep learning models to perform high-precision positioning of acoustic emission sources in complex structures, and is suitable for structural health monitoring of welded structures, metal materials, composite materials, etc. Background Art

[0002] As an important non-destructive testing method, acoustic emission technology is widely used to evaluate the internal defects and structural integrity of materials. Traditional acoustic emission source positioning methods are usually based on the time difference method (TDOA) or the arrival time method (TOA), relying on analytical mathematical algorithms to locate the acoustic emission source. However, in complex structures, sound wave propagation is affected by many factors, such as material inhomogeneity, geometry, boundary conditions, etc. Traditional methods are difficult to obtain accurate positioning results, especially when there are multipath effects and signal attenuation, the error increases significantly.

[0003] In recent years, with the development of artificial intelligence technology, signal processing technology based on deep learning has gradually been applied to the field of acoustic emission source localization. Through deep learning models, features in complex signals can be effectively extracted to improve the accuracy and robustness of acoustic emission source localization. Summary of the invention

[0004] In order to solve the problem that the traditional acoustic emission source positioning method has low positioning accuracy in complex structures and is easily affected by the propagation characteristics of sound waves, this paper proposes an acoustic emission source positioning method based on artificial intelligence deep learning. This method does not rely on complex mathematical analytical algorithms, but achieves high-precision positioning of acoustic emission sources by building and training deep learning models.

[0005] The following is the detailed content of the present invention: (1) Grid division: The first step of the present invention is to divide the structure to be tested into multiple grid areas according to the geometric shape, material properties and sound wave propagation characteristics of the structure to be tested. Each grid represents a region to be tested. The fineness of the grid division is adjusted according to the positioning accuracy requirements. The finer the grid division, the higher the accuracy of acoustic emission source positioning. For scenes with high positioning accuracy requirements, the grid division can be more refined, so that the area of ​​each grid is smaller, so as to improve the positioning accuracy of the acoustic emission source. By training a deep learning model and using the acoustic emission signals of fine grids, high-precision acoustic emission source positioning can be achieved in complex structures.

[0006] (2) AE sensor placement: After the grid division is completed, the present invention places multiple AE sensors at key locations of the structure to be tested. The placement strategy of these sensors is crucial, and it is necessary to ensure that the training signal of each grid can be collected by at least three different AE sensors. The specific location of the sensor is usually selected near key locations prone to cracks, stress concentration or other defects, so as to accurately capture the AE signal.

[0007] (3) Data collection: In order to build an effective deep learning model, the present invention uses a lead break or signal generator to generate pulse signals in the grid and collects these signals as training data. At least 100 pulse signals are collected in each grid to ensure the diversity and representativeness of the data. The collected data is then divided into a test set and a validation set, which are used for model training and validation, respectively.

[0008] (4) Construction of artificial intelligence deep learning neural network model: In the present invention, the constructed artificial intelligence deep learning neural network model includes an input layer, at least two hidden layers and an output layer. The input layer receives acoustic emission parameters collected by multiple acoustic emission sensors, such as arrival time, amplitude, ring count, energy, etc. The number of nodes in the hidden layer is dynamically adjusted according to the number of nodes in the input layer and the output layer. In order to adapt to different data scales and positioning accuracy requirements, the number of nodes in the hidden layer can be optimized and adjusted by methods such as geometric mean or cross-validation. The activation function can choose ReLU or Sigmoid to improve the nonlinear fitting ability and generalization ability of the model. The output layer includes nodes corresponding to the number of grids and an additional node for indicating positioning failure or error state. Finally, the probability value of each node is output by the Softmax or Sigmoid activation function, and the specific location of the acoustic emission source is determined according to the maximum probability value. In addition, regularization techniques such as Dropout and Batch Normalization are introduced in the present invention to prevent overfitting of the model and improve the generalization ability of the model. Hyperparameter tuning is performed by grid search or random search to further improve the performance and stability of the model.

[0009] (5) Data preprocessing: Before model training, the present invention preprocesses the collected acoustic emission signal data. The preprocessing step includes normalizing the parameters of the input nodes, that is, converting parameters of different dimensions and ranges to the same scale to ensure the consistency of the input data and the convergence of the model. The normalization method adopts the minimum-maximum scaling method, and the formula is as follows: , in, is the minimum arrival time, is the maximum arrival time, The maximum amplitude value is 100dB.

[0010] (6) Label data generation and model training: In order to train the deep learning model, the present invention generates label data of the corresponding grid area from the normalized test set acoustic emission signal data. These label data are used as supervision information for model training to guide the model to learn the mapping relationship between acoustic emission signals and grid positions. During the training process, optimization algorithms such as the Adam optimization algorithm, the stochastic gradient descent (SGD) optimization algorithm, or the RMSprop optimization algorithm are used to update the model parameters, and a learning rate decay strategy is used to improve the convergence speed and stability of the model.

[0011] (7) Model verification: After the model training is completed, the present invention uses the normalized validation set acoustic emission signal data to verify the model. The verification process determines whether the model meets the positioning accuracy requirements by judging the consistency between the output grid number and the actual grid number. At the same time, the model performance is comprehensively evaluated through indicators such as accuracy, precision, recall rate and F1 score. These indicators can reflect the performance of the model in different aspects and provide strong support for further optimization of the model.

[0012] (8) Acoustic emission source location: Finally, the present invention uses the trained deep learning model to analyze the unknown acoustic emission signal data. The model calculates and outputs the grid number based on the input acoustic emission parameters, and the grid number is the predicted acoustic emission source location. In this way, the present invention achieves high-precision positioning of acoustic emission sources in complex structures, providing strong support for structural health monitoring and safety assessment. Beneficial Effects

[0013] Compared with the prior art, the present invention has the following significant advantages: Improved positioning accuracy: Through the deep learning model, the multi-dimensional features of the acoustic emission signal are extracted and analyzed, which achieves high-precision acoustic emission source positioning in complex structures, overcoming the limitations of traditional methods that rely on analytical algorithms.

[0014] Enhanced robustness: The present invention can cope with material inhomogeneity, geometric complexity and multipath propagation problems, ensuring that high positioning accuracy can be maintained in complex structures.

[0015] Strong adaptability: This method is applicable to various types of complex structures and materials, including metal materials, composite materials and engineering structures, and has broad industrial application prospects.

[0016] High degree of automation: Through the automated processing of deep learning models, the necessity of human participation is reduced, greatly improving the efficiency of acoustic emission source positioning.

[0017] In summary, the acoustic emission source localization method based on artificial intelligence deep learning of the present invention provides a new solution for the acoustic emission source localization of complex structures, which has broad application prospects and important engineering value. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 The present invention provides a flow chart of an acoustic emission source localization method based on artificial intelligence deep learning.

[0019] Figure 2 A 400mm x 400mm metal plate grid division and sensor layout diagram provided for a specific embodiment.

[0020] Figure 3 A schematic diagram of an artificial neural network structure provided for a specific embodiment.

[0021] Figure 4 The Loss function curve provided for a specific embodiment.

[0022] Figure 5 The Score function curve provided for a specific embodiment.

[0023] Figure 6 The Margin function curve provided for a specific embodiment.

[0024] Figure 7 Entropy function curve provided for a specific embodiment. Specific embodiments

[0025] The following is a detailed description of the acoustic emission source localization method based on artificial intelligence deep learning of the present invention through specific embodiments. In this embodiment, a carbon steel plate with a size of 400mm×400mm×2mm is selected as the structure to be tested, and the material is Q235 steel.

[0026] 5.1 Experimental Preparation Experimental Materials: Measured object: Carbon steel plate with dimensions of 400mm×400mm×2mm, made of Q235 steel.

[0027] Experimental equipment: Acoustic Emission Sensors: 4 GI150 acoustic emission sensors with a frequency response range of 100kHz to 1MHz and a sensitivity greater than 70dB.

[0028] Acoustic emission collector: SAEU3H multi-channel acoustic emission collector, sampling rate is 10MHz.

[0029] Signal generator: used to generate a standard triangular waveform acoustic emission signal with a frequency of 50kHz and an amplitude of 100dB.

[0030] Data processing equipment: Computers equipped with high-performance CPUs and GPUs for processing and analyzing data, and training and validating deep learning models.

[0031] 5.2 Experimental procedures Step 1: Meshing The steel plate is divided into 16 equidistant grids of 4×4, each with a size of 100mm×100mm. Figure 2 As shown, each grid represents a potential AE source location.

[0032] Step 2: Sensor Placement Four acoustic emission sensors are arranged at the four corners of the steel plate and numbered as sensor 1 to sensor 4. Figure 2 The sensors are arranged in such a way that the signals of each grid can be collected by at least three sensors.

[0033] Step 3: Signal Acquisition A signal generator was used to generate a triangular waveform acoustic emission signal at the center of each grid, and the signal was repeated 500 times in each grid, and a total of 32,000 sets of data were collected. The collected signal characteristics include arrival time (TOA) and amplitude.

[0034] Step 4: Data Preprocessing The collected acoustic emission signal data is normalized to ensure the consistency of input data and the convergence of the model.

[0035] Arrival time normalization: Sort the arrival times from large to small, subtract the minimum arrival time from the current arrival time, and divide it by the maximum time difference to scale the arrival time difference to the range of 0 to 1. The formula is as follows: in, is the minimum arrival time, is the maximum arrival time.

[0036] Amplitude normalization: Divide the amplitude value by 100 to normalize it to the range of 0 to 1. The formula is as follows: in, The maximum amplitude value is 100dB.

[0037] Step 5: Artificial Intelligence Deep Learning Neural Network Model Construction In this embodiment, an artificial intelligence deep learning neural network model including an input layer, a hidden layer and an output layer is constructed. The input layer is used to receive signal parameters collected by the acoustic emission sensor, and the parameters collected by each sensor include arrival time and amplitude information.

[0038] The number of nodes in the hidden layer is not fixed, and can be calculated according to the geometric mean of the number of input nodes and output nodes, or dynamically adjusted by cross-validation. In this embodiment, two hidden layers are initially selected, with 50 to 100 nodes in each layer, but the specific number can be optimized and adjusted according to experimental requirements and data scale. We experimentally determined the optimal number of hidden layer nodes to be 50 nodes, and used the ReLU activation function to enhance the nonlinear fitting ability of the model, further improving the positioning accuracy of the model.

[0039] The output layer is set to 17 nodes (16 grids + 1 node representing the error state) according to the grid division of the steel plate, and the Softmax activation function is used to normalize the probability of each grid. The model outputs the grid number where the acoustic emission source is located according to the maximum probability value.

[0040] During the model training process, we also introduced the Dropout technology to prevent the model from overfitting. At the same time, we used the grid search method to optimize the model's hyperparameters to further improve the model's performance.

[0041] like Figure 3 shown.

[0042] Step 6: Model Training The normalized test set acoustic emission signal data is used to generate label data for the corresponding grid area, and the model is trained using the label data.

[0043] Training algorithm: The Adam optimization algorithm is used. The initial value of the learning rate is 0.001, and the learning rate decays after every 100 iterations.

[0044] Loss function: cross entropy loss function is used.

[0045] Monitor the loss value and accuracy changes on the validation set, and prevent model overfitting through cross-validation and Dropout techniques.

[0046] Each training is repeated for 1000 iterations until the model converges.

[0047] Step 7: Model Validation The model was validated using the normalized validation set acoustic emission signal data (100 signal data for each grid).

[0048] The model positioning accuracy is evaluated by judging the consistency between the output grid number and the actual grid number.

[0049] Calculate and evaluate the accuracy, precision, recall, and F1 score of the model.

[0050] Credibility metrics such as Score, Margin and Entropy are introduced to further verify the stability and reliability of the model in positioning tasks.

[0051] Step 8: Acoustic Emission Source Location The trained deep learning model is used to analyze the unknown acoustic emission signal data and output the grid number to achieve high-precision positioning of the acoustic emission source.

[0052] 5.3 Experimental Results and Analysis 5.3.1 Training and Validation Results During the training process, the cross entropy loss function is used for optimization, and the changes in the loss function during the training and validation phases are monitored. The Loss function calculation formula is as follows: loss=-(\sum ^{i=16}_{i=1} {(Ture[i]\times log(Predict[i])+(1-Ture[i])\times log(1-Predict[i])))\div 17} Figure 4 The figure shows the change of Loss function with the number of iterations in the training and validation stages. It can be seen that the training loss decreases rapidly in the early stage, while the validation loss tends to stabilize after about 1000 iterations. The loss value stabilizes at a low level in the later stage of training, indicating that the model converges well and has strong generalization ability.

[0053] The score function is defined as: Score = the maximum value among the current 17 output nodes, which indicates the confidence of the model in predicting the correct category. Figure 5 The Score function curve is shown in Figure 2. It can be seen that the Score value gradually increases during the training process and finally stabilizes at around 0.99. This value is close to 1, indicating that the model has a very high confidence in the prediction of most categories.

[0054] The Margin function is defined as: Margin = the maximum value of the current 17 output nodes – the second largest value of the current 17 output nodes, which represents the difference between the optimal category and the second-best category. The larger the Margin, the stronger the confidence of the model in the classification result. Figure 6 It is the curve of Margin function change. As can be seen from the figure, as the training progresses, the Margin value gradually increases and finally stabilizes above 0.9. This shows that the model can effectively distinguish the optimal category from the suboptimal category when making classification decisions, reducing the occurrence of misclassification.

[0055] The Entropy function is used to measure the distribution of the model's prediction results. A smaller Entropy value indicates that the prediction results are more concentrated. It is defined as: Entropy=-\sum ^{i=16}_{i=0} {output[i]\times log2(output[i])} Figure 7 The Entropy function curve is shown in the figure. As can be seen from the figure, the Entropy value gradually approaches about 0.1 with the change of training rounds. This means that the prediction distribution of the model is concentrated and has strong classification stability.

[0056] 5.3.2 Model Performance Rating The performance of the model on the test set is evaluated using indicators such as accuracy, precision, recall, and F1 score.

[0057] TP (True Positive): True positive, the model predicts it is a positive example, and it is actually a positive example.

[0058] FP (False Positive): False positive example, the model predicts it to be a positive example, but it is actually a negative example.

[0059] TN (True Negative): True negative example, the model predicts it to be a negative example, and it is actually a negative example.

[0060] FN (False Negative): False negative example, the model predicts a negative example, but it is actually a positive example.

[0061] The grid numbering starts from 0. The recognition errors in the experiment are as follows: 4 data of grid No. 1 are mistakenly recognized as grid No. 2; 2 data of grid No. 7 are mistakenly recognized as grid No. 3.

[0062] In order to better analyze the results, it is assumed that grids 2 and 3 are negative examples and the other grids are positive examples. Therefore, the number of positive examples is 100. 14, the number of negative examples is 100 2. According to the experimental results, the calculation of relevant indicators is as follows: True Positives (TP): The number of positive examples correctly classified, a total of 100 There are 14 positive examples, 6 of which were misidentified, so: True Negatives (TN): The number of correctly classified negative examples: False Negatives (FN): The number of positive examples that are misclassified as negative examples: False Positives (FP): The number of negative examples that are misclassified as positive examples: Based on these data, the following performance indicators can be calculated: Recall: Accuracy: Precision: F1 Score: The results are summarized in the following table: Table 1: Positioning accuracy, precision, recall and F1 score Accuracy Accuracy Recall F1 score 99.625% 100% 99.57% 99.78% From the above results, we can see that the model performs very well on the test set, with extremely high accuracy, precision, recall and F1 score. The model can accurately locate the positions of most acoustic emission sources with almost no false positives.

[0063] 5.3.3 Comparison with the traditional calculation results based on the time difference method (TDOA) To further verify the superiority of the deep learning method, we compared it with the traditional acoustic emission time difference positioning method (TDOA). The accuracy of the traditional time difference positioning algorithm is defined as: the number of acoustic emission positioning sources falling within the corresponding grid / the number of pulse signals emitted 500. The accuracy of AI deep learning acoustic emission source positioning is defined as: the number of times the current grid is correctly identified / the total number of data identifying the current grid.

[0064] In this test, the results of AI deep learning acoustic emission source positioning are as follows: Grid No. 2 identified a total of 100 data, of which only 4 were recognized incorrectly; Grid No. 8 identified a total of 100 data, of which only 2 were recognized incorrectly; and the remaining grids identified 100 grid data, all of which were recognized correctly.

[0065] The results are shown in Table 2.

[0066] Table 2 Comparison of the accuracy of traditional time difference positioning method and deep learning method Grid Number Traditional acoustic emission accuracy AI deep learning acoustic emission source positioning accuracy 01 100% 100% 02 99% 96% 03 50% 100% 04 99% 100% 05 79% 100% 06 89% 100% 07 100% 100% 08 90% 98% 09 61% 100% 10 100% 100% 11 99% 100% 12 79% 100% 13 100% 100% 14 50% 100% 15 81% 100% 16 100% 100% Since the experiment used a small metal plate, the acoustic signal is easily affected by the superposition of reflected waves, so the positioning accuracy of the traditional time difference algorithm is low. However, it can be seen from the table that the positioning accuracy of the deep learning method in each grid is much higher than that of the traditional TDOA method, especially in the difficult-to-position grids 03, 05, 09, 12, 14 and 15, where the traditional time difference positioning accuracy is not high, the deep learning method still maintains a high accuracy of 100%. This shows that the deep learning method has stronger robustness and the ability to handle complex scenes.

[0067] In summary, this embodiment verifies the feasibility and superiority of the deep learning-based acoustic emission source positioning method through specific experiments. This method achieves high-precision positioning of acoustic emission sources in complex structures by constructing and training a deep learning model, significantly improving positioning accuracy and robustness, and providing a new solution for the field of non-destructive testing.

Claims

1. A method for locating acoustic emission sources based on artificial intelligence deep learning, characterized in that: The method comprises the following steps: (1) Grid division: The structure to be measured is divided into multiple grid areas according to its geometric shape, material properties and acoustic wave propagation characteristics. Each grid represents a region to be measured. The fineness of the grid division is adjusted according to the positioning accuracy requirements. The finer the grid division, the higher the acoustic emission source positioning accuracy. (2) AE sensor placement: Multiple AE sensors are placed at key locations of the structure to be tested to ensure that the training signal of each grid can be collected by at least three different AE sensors; (3) Data collection: Generate pulse signals in the grid using a lead cutter or signal generator. Collect at least 100 pulse signals in each grid, and divide the collected data into a test set and a validation set. (4) Construction of artificial intelligence deep learning neural network model: Construct an artificial intelligence deep learning neural network model including an input layer, at least two hidden layers, and an output layer. The input nodes of the input layer are based on multiple acoustic emission parameters, the number of hidden layer nodes is in a geometric mean relationship with the number of input nodes, and the activation function can be selected as ReLU or Sigmoid to enhance the nonlinear fitting ability of the model; (5) Data preprocessing: preprocessing the collected acoustic emission signals, wherein the preprocessing includes normalizing the parameters of the input nodes; (6) Label data generation and model training: Generate label data of the corresponding grid area from the normalized test set acoustic emission signal data, and use the label data to train the model until convergence; (7) Model verification: The model is verified using the normalized validation set acoustic emission signal data. The consistency between the output grid number and the actual grid number is used to determine whether the model meets the positioning accuracy requirements. The model is evaluated using the accuracy, precision, recall and F1 score indicators. (8) Acoustic emission source positioning: Use the trained deep learning model to analyze the unknown acoustic emission signal data and output the grid number, which is the location of the acoustic emission source.

2. The method for locating an acoustic emission source based on artificial intelligence deep learning according to claim 1, characterized in that: The grid division method is adjusted according to the geometric shape, material properties and sound wave propagation characteristics of the structure to be measured. Refined grid division is used in scenarios that require high-precision positioning. For regular structures, the grid area is an equidistant square or rectangular grid; for irregular structures, the grid area is irregularly divided according to the curvature change of the geometric shape, the sound wave propagation path and material properties. The smaller the grid, the higher the positioning accuracy.

3. The method for locating an acoustic emission source based on artificial intelligence deep learning according to claim 1, characterized in that: The sensors are arranged near the locations where cracks are prone to occur, stress concentration locations or other key locations where defects are prone to occur in the structure to be tested.

4. The method for locating an acoustic emission source based on artificial intelligence deep learning according to claim 1, characterized in that: The acoustic emission parameters of the input node include but are not limited to arrival time and amplitude. The arrival time is used to calculate the time difference between the acoustic emission source and each sensor, and the amplitude information represents the energy attenuation experienced by the sound wave during the propagation process.

5. The input node acoustic emission parameters according to claim 4 further include ring count, energy, rise time, duration, RMS, ASL, centroid frequency, peak frequency, average frequency, initial frequency, and reverberation frequency.

6. The method for locating acoustic emission sources based on artificial intelligence deep learning according to claim 1, characterized in that: The normalization process in the preprocessing step adopts the minimum-maximum scaling method to scale the input node acoustic emission parameters to the range of 0 to 1 respectively, so as to ensure the consistency of input data and the convergence of the model.

7. The method for locating acoustic emission sources based on artificial intelligence deep learning according to claim 1, characterized in that: The structure and parameter configuration of the artificial neural network model are as follows: a) Input layer: The input layer includes N input nodes, where N is the product of the number of sensors and the number of selected acoustic emission parameters, and the input node is used to receive multiple acoustic emission parameters collected by each sensor; b) Hidden layer: contains at least two hidden layers. The number of nodes in each layer is calculated based on the geometric mean of the number of input nodes and output nodes or dynamically adjusted through cross-validation. The number of nodes can be optimized according to actual needs and computing resources. The hidden layer uses ReLU or Sigmoid activation function to enhance the nonlinear fitting ability and generalization ability of the model. The specific number of nodes and layers of the hidden layer can be flexibly adjusted according to the requirements of data scale, computing resources and model performance; c) Output layer: The output layer contains M+1 nodes, where M is the number of grids, and +1 node is used to indicate positioning failure or error state. The output layer uses Softmax or Sigmoid activation function to output the probability distribution of each grid area and error state, and finally determines the grid area where the acoustic emission source is located according to the maximum probability value; d) Model optimization: During the training of the neural network, regularization techniques are used to prevent overfitting, and network hyperparameters are optimized through hyperparameter optimization methods to improve the performance and stability of the model.

8. The method for locating acoustic emission sources based on artificial intelligence deep learning according to claim 1, characterized in that: During the training process of the deep learning model, the Adam optimization algorithm, the stochastic gradient descent (SGD) optimization algorithm or the RMSprop optimization algorithm are used to improve the convergence speed and stability of the model. During the training process, a learning rate decay strategy is adopted, with an initial learning rate of 0.001, and the learning rate is reduced according to a predetermined ratio after every 100 iterations.

9. The method for locating acoustic emission sources based on artificial intelligence deep learning according to claim 1, characterized in that: During the optimization process of the deep learning model, a cross entropy loss function is used for classification tasks or a mean square error (MSE) loss function is used for regression tasks to perform error calculation to improve the positioning accuracy of the model.

10. The method for locating acoustic emission sources based on artificial intelligence deep learning according to claim 1, characterized in that: The output layer calculates the output result through the selected activation function, the value of the output node represents the probability that the acoustic emission source is located in the corresponding grid area, and the node with the maximum probability is selected as the final acoustic emission source positioning result.

Images