Steel bridge deck welding pore defect identification method and system

Abstract

The invention discloses a steel bridge deck welding pore defect identification method and system. The system comprises a signal acquisition module, a signal processing module, a signal classification module, an initial classification module, an intermediate classification module, a target classification module, a final classification module and a result output module. The signal acquisition module is used for acquiring a plurality of acoustic emission signals with no damage and pore defects on the test steel bridge deck, and the acoustic emission signals acquired on the test steel bridge deck form a source domain data set; a plurality of acoustic emission signals are collected on the to-be-tested steel bridge deck, and the acoustic emission signals collected on the to-be-tested steel bridge deck form a target domain data set; and the signal processing module is used for processing the acoustic emission signal in combination with the adaptive analysis mode and Hilbert transform to obtain an adaptive marginal spectrum of the acoustic emission signal. According to the method, the process of welding the steel bridge deck can be monitored on line, the air hole defect in the welding process can be accurately recognized, and the recognition efficiency is high.

Description

Technical Field

[0001] The present invention relates to a defect identification method, and in particular to a steel bridge deck welding porosity defect identification method and system. Background Art

[0002] Orthotropic steel bridge decks are complex, thin-walled structures welded together from panels, diaphragms, and longitudinal U-ribs. Due to their complex structural characteristics, welding can easily produce porosity defects. Under bridge service loads, these defects can lead to stress concentration and fatigue crack propagation, seriously threatening the safe service performance of bridge structures.

[0003] Existing common techniques for detecting porosity weld defects include visual, ultrasonic, eddy current, infrared, and X-ray inspections. Visual inspection can only detect surface defects in welds, making it difficult to identify subtle defects, resulting in low overall accuracy. Ultrasonic, eddy current, infrared, and X-ray inspections offer penetrating and non-destructive properties, but they suffer from low efficiency and high labor costs. Summary of the Invention

[0004] Purpose of the invention: The first purpose of the present invention is to provide a method for identifying porosity defects in steel bridge deck welding with high accuracy and efficiency.

[0005] The second object of the present invention is to provide a steel bridge deck welding porosity defect identification system.

[0006] Technical solution: The present invention discloses a method for identifying porosity defects in steel bridge deck welding, comprising the following steps:

[0007] S1: Multiple acoustic emission signals, including those with and without damage, are collected on the test steel bridge deck to form a source domain dataset; multiple acoustic emission signals are collected on the steel bridge deck to be tested to form a target domain dataset;

[0008] S2: Processing the acoustic emission signal by combining the adaptive analytical mode and Hilbert transform to obtain the adaptive marginal spectrum of the acoustic emission signal;

[0009] S3: The target domain dataset is divided into a target domain labeled dataset and a target domain unlabeled dataset. The acoustic emission signals contained in the source domain labeled dataset and the target domain labeled dataset are labeled with "no damage" and "with pore defects". The target domain labeled dataset contains acoustic emission signals without damage and with pore defects.

[0010] S4: Adjust the architecture of Wav2Vec2.0 to obtain an improved AEWav2Vec2.0 model, and use the adaptive marginal spectrum of the source domain annotated dataset to train the AEWav2Vec2.0 model to obtain an initial classification model for identifying the label category of the acoustic emission signal;

[0011] S5: The target domain annotated dataset is divided into a training set and a validation set according to a preset ratio. It is ensured that both the training set and the validation set contain acoustic emission signals of both damage-free working conditions and those with pore defects. The training set and the source domain annotated dataset form a mixed annotated dataset. The adaptive marginal spectrum of the mixed annotated dataset is used to train the initial model to obtain an intermediate classification model.

[0012] S6: Input the adaptive marginal spectrum of the unlabeled dataset of the target domain into the intermediate classification model. The intermediate classification model outputs two classification results with confidence for the corresponding acoustic emission signals. The acoustic emission signals with confidence higher than the preset value Q1 are screened, and the screened acoustic emission signals and the mixed labeled dataset constitute a joint training dataset. The intermediate classification model is iteratively optimized using the joint training dataset until the preset condition is met and the iterative optimization is stopped, and the target classification model is obtained.

[0013] S7: Input the adaptive marginal spectrum of the validation set into the target classification model, and the target classification model outputs two classification results with confidence for the corresponding acoustic emission signals. The acoustic emission signals with confidence higher than the preset value Q4 are screened to form a predicted and labeled data set. The predicted and labeled data set is added to the joint training data set to form a final training data set. The target classification model is iteratively optimized using the final training data set until the same preset condition as in step S6 is reached, and the iterative optimization is stopped to obtain the final classification model.

[0014] S8: Inputting the adaptive marginal spectrum of the target domain unlabeled dataset into the final classification model, the final classification model outputs the classification result of each acoustic emission signal of the target domain unlabeled dataset.

[0015] Furthermore, in step S1, the method for collecting the acoustic emission signal of the weld with porosity defects on the test steel bridge deck is: spraying the surface of the weld before welding.

[0016] Furthermore, the step of obtaining the adaptive marginal spectrum of the acoustic emission signal in step S2 is:

[0017] Adaptive analytical mode is used to decompose the acoustic emission signal into spectrally mutually exclusive components;

[0018] Perform Hilbert transform on all components of the acoustic emission signal to generate analytical signals;

[0019] Calculate the instantaneous amplitude and instantaneous frequency of all analytical signals, and construct the time-frequency spectrum of the corresponding acoustic emission signal based on all the instantaneous amplitudes and instantaneous frequencies of each acoustic emission signal;

[0020] The time-frequency spectrum is integrated along the time axis to obtain the adaptive marginal spectrum of each acoustic emission signal.

[0021] In step S4, the architecture of Wav2Vec2.0 is adjusted as follows: based on the dimension of the adaptive marginal spectrum of the acoustic emission signal, the convolution kernel, step size and number of input channels of the input convolution layer of the Wav2Vec2.0 model are modified; Dropout and Pooling are added after the convolution layer of Wav2Vec2.0, the number of encoder layers of the Transformer is set to 8, and the output layer of the Transformer is replaced with a binary classification fully connected layer.

[0022] Furthermore, the initial classification model in step S4 is optimized, and the optimization goal is to minimize the cross entropy loss function L cls , L cls The expression is: where y i is the true label, p(y i |x i ) is the y predicted by the initial classification model i N refers to the number of acoustic emission signals contained in the source domain annotation dataset.

[0023] Furthermore, in step S5, during the training of the initial model, the feature extraction layer parameters of the initial classification model are retained, the last three layers of the Transformer encoder are unfrozen, and the parameters of the last three layers of the unfrozen Transformer encoder are allowed to be updated during the training process.

[0024] Furthermore, the steps for setting the preset conditions in step S6 are as follows:

[0025] Set the joint loss function L of the supervision loss and the intermediate classification model prediction label loss joint , and L joint The expression is L joint =λL sup +(1-λ)L predict Among them, L predict is the cross entropy loss of the acoustic emission signal filtered based on confidence, L sup is the cross entropy loss of the mixed labeled dataset, and λ is the dynamic weight;

[0026] When the dynamic weight λ decays to the preset value Q2, the iterative optimization stops.

[0027] Furthermore, the final classification model in step S8 also outputs whether each acoustic emission signal of the target domain dataset belongs to the predicted annotation dataset.

[0028] Furthermore, in step S8, the acoustic emission signals whose confidence level output by the final classification model is higher than the preset value Q4 are screened, and the screened acoustic emission signals are added to the final training data set. The target classification model is iteratively optimized again using the final training data set until the same preset conditions as in step S6 are reached and the iterative optimization is stopped.

[0029] Based on the same inventive concept, the present invention also discloses a steel bridge deck welding porosity defect identification system, comprising:

[0030] A signal acquisition module is used to collect multiple acoustic emission signals including those without damage and those with porosity defects on the test steel bridge deck, and the acoustic emission signals collected on the test steel bridge deck constitute a source domain data set; and to collect multiple acoustic emission signals on the steel bridge deck to be tested, and the acoustic emission signals collected on the steel bridge deck to be tested constitute a target domain data set;

[0031] A signal processing module is used to process the acoustic emission signal by combining the adaptive analytical mode and the Hilbert transform to obtain the adaptive marginal spectrum of the acoustic emission signal;

[0032] A signal classification module is used to divide the target domain dataset into a target domain labeled dataset and a target domain unlabeled dataset, and to label the acoustic emission signals contained in the source domain labeled dataset and the target domain labeled dataset using "no damage" and "with pore defects". The target domain labeled dataset contains acoustic emission signals without damage and with pore defects.

[0033] The initial classification module is used to adjust the architecture of Wav2Vec2.0 to obtain an improved AEWav2Vec2.0 model. The AEWav2Vec2.0 model is trained using the adaptive marginal spectrum of the source domain annotated dataset to obtain an initial classification model for identifying the label category of the acoustic emission signal.

[0034] The intermediate classification module is used to divide the target domain annotated dataset into a training set and a validation set according to a preset ratio, and ensure that both the training set and the validation set contain acoustic emission signals of both damage-free working conditions and those with pore defects. The training set and the source domain annotated dataset form a mixed annotated dataset, and the adaptive marginal spectrum of the mixed annotated dataset is used to train the initial model to obtain the intermediate classification model.

[0035] The target classification module is used to input the adaptive marginal spectrum of the unlabeled dataset of the target domain into the intermediate classification model. The intermediate classification model outputs two classification results with confidence for the corresponding acoustic emission signals, and selects acoustic emission signals with confidence higher than the preset value Q1. The selected acoustic emission signals and the mixed labeled dataset constitute a joint training dataset. The intermediate classification model is iteratively optimized using the joint training dataset until the preset conditions are met and the iterative optimization is stopped, and the target classification model is obtained.

[0036] The final classification module is used to input the adaptive marginal spectrum of the validation set into the target classification model. The target classification model outputs two classification results with confidence for the corresponding acoustic emission signals. The acoustic emission signals with confidence higher than the preset value Q4 are screened to form a predicted annotation dataset. The predicted annotation dataset is added to the joint training dataset to form the final training dataset. The target classification model is iteratively optimized using the final training dataset. When the preset conditions are met, the iterative optimization is stopped and the final classification model is obtained.

[0037] The result output module is used to input the adaptive marginal spectrum of the unlabeled data set in the target domain into the final classification model, and the final classification model outputs the classification result and confidence of each acoustic emission signal of the unlabeled data set in the target domain.

[0038] Beneficial Effects: Compared with existing technologies, this invention offers the following significant advantages: By collecting acoustic emission signals during welding, it enables online monitoring of the steel bridge deck welding process, accurately identifying porosity defects during welding with high efficiency. This can be used to guide the steel bridge deck welding process, thereby improving overall welding quality. Furthermore, acoustic emission signals utilize transient elastic waves released by materials during processes such as plastic deformation and fracture to diagnose structural damage, offering the advantages of high detection efficiency and sensitivity.

[0039] The present invention combines the adaptive analytical mode and Hilbert transform to obtain the adaptive marginal spectrum of the acoustic emission signal. On the one hand, it avoids the modal mixing phenomenon of traditional empirical mode analysis. On the other hand, it adaptively extracts the different frequency components contained in the acoustic emission signal, can obtain a time-frequency spectrum with clearer physical meaning, and effectively characterizes the frequency domain energy distribution of the acoustic emission signal.

[0040] The acoustic emission signal of the steel bridge deck to be tested of the present invention is divided into three parts to participate in the training and optimization of the classification model. On the one hand, it can improve the overall feature recognition of the acoustic emission signal of the steel bridge deck to be tested, which is conducive to improving the accuracy of the final recognition result; on the other hand, the model can be trained during the test process to achieve continuous optimization of the classification model parameters, which is conducive to improving the long-term stability and accuracy of the model during the application process. That is, as the number of uses and the duration of use increase, the accuracy of the classification results of the present invention also increases accordingly. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 is a flow chart of the method of the present invention;

[0042] Figure 2 A framework diagram of an initial classification model according to an embodiment of the present invention;

[0043] Figure 3 A top view of a steel bridge deck according to an embodiment of the present invention;

[0044] Figure 4 This is a front view of a steel bridge deck according to an embodiment of the present invention;

[0045] Figure 5 is a time domain waveform of an undamaged acoustic emission signal of a steel bridge deck according to an embodiment of the present invention;

[0046] Figure 6 is a time domain waveform of an acoustic emission signal of a steel bridge deck with a porosity defect according to an embodiment of the present invention;

[0047] Figure 7 For the present invention Figure 5 The original waveform and component waveforms of the damage-free acoustic emission signal;

[0048] Figure 8 For the present invention Figure 6 The original waveform and component waveforms of the acoustic emission signal of the pore defect;

[0049] Figure 9 The adaptive marginal spectrum of the acoustic emission signal under non-destructive working conditions according to an embodiment of the present invention;

[0050] Figure 10 This is an adaptive marginal spectrum of the acoustic emission signal under non-destructive working conditions according to an embodiment of the present invention. DETAILED DESCRIPTION

[0051] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0052] Example 1

[0053] like Figure 1 As shown, the present invention discloses a method for identifying porosity defects in steel bridge deck welding, comprising the following steps:

[0054] S1: Multiple acoustic emission signals, including those with and without damage, are collected on the test steel bridge deck to form a source domain dataset. Multiple acoustic emission signals are collected on the steel bridge deck to be tested to form a target domain dataset.

[0055] Preferably, when acquiring the initial sample acoustic emission signals, both non-destructive welding and welding with porosity defects are performed on the test steel bridge deck. Acquiring the acoustic emission signals with porosity defects involves spraying the weld surface with moisture before welding. The generation of water vapor during welding generates acoustic emission signals with porosity defects. Acquiring acoustic emission signals with porosity defects is not limited to this method; other methods may also be used to create porosity defects in actual use.

[0056] Preferably, acoustic emission sensors are installed at the four corners of the test steel bridge deck and the steel bridge deck to be tested, and the frequency range of the acoustic emission sensors is 100-1000KHZ. Preferably, the signal-to-noise ratio of the acoustic emission signal is improved by a preamplifier.

[0057] S2: All acoustic emission signals are processed by combining the adaptive analytical mode and Hilbert transform to obtain the adaptive marginal spectrum of the acoustic emission signal.

[0058] S21: The adaptive analytical mode is used to decompose the acoustic emission signal into spectrally mutually exclusive components. The decomposition expression is as follows:

[0059]

[0060] Where x(t) refers to the acoustic emission signal, The cutoff frequency of the modal decomposition can be adaptively adjusted based on the actual situation or needs.

[0061] S22: Perform Hilbert transform on each component of the acoustic emission signal to generate an analytical signal:

[0062]

[0063] Among them, z i (t) refers to the analytical signal, H[x] refers to the acoustic emission signal x(t) Perform Hilbert transform, where j is the imaginary unit.

[0064] S23: Calculate the instantaneous amplitude A of all analytical signals i (t) and instantaneous frequency f i (t):

[0065]

[0066] The instantaneous amplitude and frequency domain information of the analytical signals of all acoustic emission signals are superimposed on the time-frequency plane to construct a time-frequency spectrum, that is, to construct an analytical mode decomposition - Hilbert spectrum:

[0067]

[0068] Where f is frequency; t is time; H(f,t) is the time-frequency spectrum; and δ[·] is the Dirac function, indicating that energy is concentrated at the instantaneous frequency. This method combines analytical mode decomposition and the Hilbert transform to obtain mutually exclusive frequency components and adaptively extracts the different frequency components contained in the acoustic emission signal, resulting in a time-frequency spectrum with clearer physical meaning.

[0069] S24: Integrate the time-frequency spectrum along the time axis to obtain the adaptive marginal spectrum of each acoustic emission signal. The expression of the adaptive marginal spectrum M(f) of the acoustic emission signal is:

[0070]

[0071] Where T is the total duration of the acoustic emission signal; the adaptive marginal spectrum can reflect the cumulative energy-frequency distribution of the acoustic emission signal within the total duration and can be used to identify the characteristic frequency of welding porosity defects.

[0072] S3: The target domain dataset is divided into a target domain labeled dataset and a target domain unlabeled dataset. The acoustic emission signals contained in the source domain labeled dataset and the target domain labeled dataset are labeled with "undamaged" and "with pore defects." The target domain labeled dataset contains acoustic emission signals of both undamaged and pore defects. The target domain labeled dataset is labeled based on manual experience to identify acoustic emission signals in the target domain dataset that are clearly undamaged or clearly defective.

[0073] Preferably, the ratio of the acoustic emission signal data contained in the source domain standard dataset, the target domain annotated dataset, and the target domain unannotated dataset is 6:1:3. In actual use, this preset ratio can be adjusted according to actual conditions and needs. For example, in actual use, when annotating the acoustic emission signals in the target domain dataset, the obviously intact and defective acoustic emission signals are usually annotated, that is, the target domain annotated dataset accounts for a relatively small proportion.

[0074] S4: The architecture of Wav2Vec2.0 is adjusted to obtain an improved AEWav2Vec2.0 model, and the AEWav2Vec2.0 model is trained using the adaptive marginal spectrum of the source domain annotated dataset to obtain an initial classification model for identifying the label category of the acoustic emission signal.

[0075] Preferably, the adaptive marginal spectra of all acoustic emission signals are normalized in advance to obtain the model input tensor of the Wav2Vec2.0 model.

[0076] The way to adjust the architecture of Wav2Vec2.0 is as follows:

[0077] Based on the dimensionality of the adaptive marginal spectrum of the acoustic emission signal, the convolution kernel of the input convolution layer of the Wav2Vec2.0 model is modified to adapt to the long sequence characteristics of the adaptive marginal spectrum, the step size of the input convolution layer of the Wav2Vec2.0 model is modified to reduce the amount of computation, and the number of input channels of the input convolution layer of the Wav2Vec2.0 model is modified. The convolution kernel size is typically smaller, such as 30, to extract detailed features in longer acoustic emission signals, such as when extracting detailed features in an adaptive marginal spectrum of dimension 32768×1. Preferably, a step size of 10 is selected to achieve a balance between computational efficiency and information retention. The number of input channels is typically set to 1, as each signal of the acoustic emission signal is typically a single channel.

[0078] Considering the limited number of acoustic emissions, which can easily lead to overfitting in the early stages, Dropout is added after the convolutional layer of Wav2Vec2.0 to prevent overfitting and enhance the generalization ability of the model. Preferably, the Dropout rate is 0.1. Pooling is added after the convolutional layer to reduce the dimensionality through downsampling, so that the subsequent layers of Wav2Vec2.0 only need to process refined features, thereby reducing the 7 convolutional layers in the original Wav2Vec2.0 model to 3 layers without affecting the performance of the Wav2Vec2.0 model. The number of Transformer encoder layers is set to 8 to reduce the complexity of the Wav2Vec2.0 model to adapt to the situation where there are fewer acoustic emission signal samples. The Transformer output layer is replaced with a two-class fully connected layer to distinguish between the two labels of "no damage" and "with pore defects".

[0079] Preferably, the AEWav2Vec2.0 model is trained using the adaptive marginal spectrum of the source domain annotated dataset for supervised training, and the training optimization goal is to minimize the cross entropy loss function L cls , where the cross entropy loss function L cls The expression is

[0080]

[0081] Among them, y i is the real label of the acoustic emission signal in advance, p(y i |x i ) is the y predicted by the initial classification model i The label of the source domain is N, which refers to the number of acoustic emission signals contained in the source domain annotation dataset. Minimize the cross entropy loss function L cls , you can optimize the weights and biases of the convolutional layers in the feature extractor, the weight matrix of the attention layer in the Transformer encoder, and the classification weights of the output layer.

[0082] In summary, if Figure 2 As shown in the figure, the initial classification model obtained after training includes an input one-dimensional convolution module, a Transformer encoder, a Pooling layer, and a binary classification output layer, which can perform preliminary recognition and classification of the adaptive marginal spectrum of the acoustic emission signal.

[0083] S5: The target domain annotated dataset is divided into a training set and a validation set according to a preset ratio, and both the training set and the validation set are guaranteed to contain acoustic emission signals of both damage-free working conditions and those with pore defects. The training set and the source domain annotated dataset constitute a mixed annotated dataset, and the adaptive marginal spectrum of the mixed annotated dataset is used to train the initial model to obtain an intermediate classification model.

[0084] Preferably, the preset ratio of the target domain annotated dataset is training set: validation set = 1:1; the ratio between the two can be adjusted according to actual needs in actual use. In actual use, the preset ratio can be adjusted according to actual conditions and needs.

[0085] Because the acoustic emission signals of the source and target domain annotated datasets differ, training with only the source domain annotated dataset can easily lead to overfitting of the acoustic emission signal characteristics of the source domain annotated dataset, resulting in poor recognition of the acoustic emission signal characteristics of the target domain annotated dataset. Combining the training set of the target domain annotated dataset with the source domain annotated dataset to construct a hybrid annotated dataset helps the model gradually adapt to the characteristics of the target domain data and improves the accuracy and stability of the intermediate classification model's cross-domain recognition.

[0086] Preferably, during the training of the initial model, the feature extraction layer parameters of the initial classification model are retained, the last three layers of the Transformer encoder are unfrozen, and the parameters of the last three layers of the unfrozen Transformer encoder are allowed to be updated during the training process to improve the adaptability of the intermediate classification model to the characteristics of the acoustic emission signals in the target domain annotated dataset.

[0087] S6: The adaptive marginal spectrum of the unlabeled dataset in the target domain is input into the intermediate classification model. The intermediate classification model outputs two classification results with confidence for the corresponding acoustic emission signals. The acoustic emission signals with confidence higher than the preset value Q1 are screened. The screened acoustic emission signals and the mixed labeled dataset constitute a joint training dataset. The intermediate classification model is iteratively optimized using the joint training dataset until the preset conditions are met and the iterative optimization is stopped, and the target classification model is obtained.

[0088] Preferably, the preset value Q1 of the screening confidence is 0.9; in actual use, the preset value Q1 can be adjusted according to needs.

[0089] Set the joint loss function of the supervision loss and the intermediate classification model prediction label loss, and set the dynamic weight control mechanism, where the joint loss function Ljoint The expression is

[0090] L joint =λL sup +(1-λ)L predict

[0091] Among them, L predict is the cross entropy loss of the acoustic emission signal filtered based on confidence, L sup is the cross-entropy loss for the mixed annotated dataset, and λ is the dynamic weight. The iterative optimization of the intermediate classification model stops when the dynamic weight λ decays to a preset value Q2, preferably set to 0.3. Using the joint loss function to iteratively train the fine-tuning model helps preserve the model parameters that achieve the best validation performance.

[0092] S7: The adaptive marginal spectrum of the validation set is input into the target classification model, and the target classification model outputs two classification results with confidence for the corresponding acoustic emission signals. The acoustic emission signals with confidence higher than the preset value Q4 are screened to form a predicted annotation dataset, and the predicted annotation dataset is added to the joint training dataset to form the final training dataset. The target classification model is iteratively optimized using the final training dataset until the same preset conditions as in step S6 are reached, and the iterative optimization is stopped, and the final classification model is obtained.

[0093] Preferably, the preset value Q4 is set to 0.9, and Q4 can be adjusted according to needs in actual use. Preferably, Q4 is equal to Q1, and actual values of the two can be set separately in actual application.

[0094] S8: Inputting the adaptive marginal spectrum of the target domain unlabeled dataset into the final classification model, the final classification model outputs the classification result and confidence of each acoustic emission signal of the target domain unlabeled dataset.

[0095] When the confidence level of a certain acoustic emission signal with a porosity defect exceeds a preset value Q3, the acoustic emission signal is determined to have a porosity defect. The final classification model also outputs whether each acoustic emission signal in the target domain dataset belongs to the predicted annotation dataset.

[0096] Preferably, the acoustic emission signals output by the final classification model with a confidence level higher than a preset value Q4 are screened, and the screened acoustic emission signals are added to the final training dataset. The target classification model is iteratively optimized using the final training dataset again until the same preset conditions as in step S6 are met, and the iterative optimization is stopped. After the final classification model outputs the results, the acoustic emission signals with high confidence levels higher than the preset value Q4 in the unlabeled dataset of the target domain are screened and used again for iterative optimization of the target classification model, which can further improve the accuracy of the final recognition result.

[0097] In this embodiment, in step S5, a mixed labeled dataset including a training set is used to optimize the initial model for training. In step S6, a mixed labeled dataset including unlabeled data in the target domain is used to optimize the intermediate classification model. In step S7, a joint mixed labeled data including a validation set is used to iteratively optimize the target classification model. This joint training can effectively reduce the distribution differences between domains, which is conducive to improving the accuracy of the overall identification of air hole defects. In this embodiment, the acoustic emission signals of the steel bridge deck to be tested are divided into three parts: a labeled training set, a labeled validation set, and an unlabeled unlabeled dataset in the target domain. The acoustic emission signals of these three parts are used in the training and optimization of the classification model in steps. On the one hand, this can improve the overall feature recognition of the acoustic emission signals of the steel bridge deck to be tested, which is conducive to improving the accuracy of the final recognition result. On the other hand, the model can be trained during the testing process to achieve continuous optimization of the classification model parameters, which is conducive to improving the long-term stability and accuracy of the model during application. That is, as the number of uses and the length of use increase, the accuracy of the classification results output by the final classification model also increases.

[0098] Example 2

[0099] The present invention discloses a method for identifying porosity defects in steel bridge deck welding during experimental verification.

[0100] S1: Multiple acoustic emission signals, including those with and without damage, are collected on the test steel bridge deck to form a source domain dataset. Multiple acoustic emission signals are collected on the steel bridge deck to be tested to form a target domain dataset.

[0101] In this embodiment, the welding type is robot gas shielded welding. The experiment set up seven steel bridge panels. On five of them, non-destructive welding was performed, and 700 non-destructive acoustic emission signals were collected. The weld surface of the other two steel bridge panels was sprayed with wet treatment, and 300 acoustic emission signals with porosity defects were collected. The structural diagram of the steel bridge panel is shown in the figure. Figure 3 and Figure 4 As shown, the dimensions of the steel bridge deck are as follows: length and width are 500mm×600mm, and the plate thickness is 20mm; the lower edge of the U-shaped longitudinal rib is 300mm wide, 280mm high, and the plate thickness is 8mm.

[0102] In this embodiment, four broadband acoustic emission sensors with a frequency range of 100-1000kHz are selected to collect acoustic emission signals, and the four acoustic emission sensors are as follows: Figure 3 As shown in S1-S4 in the figure, they are installed at the top corners of the steel bridge deck, and the installation position of the acoustic emission sensor is 30 mm away from the edge of the steel bridge deck. It is fixed by coupling agent and magnetic clamps. The signal-to-noise ratio of the acoustic emission signal is improved by a preamplifier, the amplification factor is set to 40 dB, the trigger sampling threshold is 40 dB, and the sampling frequency is 5 MHz.

[0103] In this embodiment, a total of 700 intact acoustic emission signals and 300 pore defect acoustic emission signals are obtained, and 600 of them are randomly selected to form a source domain data set, and the source domain data set contains both intact acoustic emission signals and pore defect acoustic emission signals. The other 400 acoustic emission signals constitute a target domain data set, and the target domain data set contains both intact acoustic emission signals and pore defect acoustic emission signals. The time domain waveform of one of the intact acoustic emission signals received by the acoustic emission sensor in this embodiment is as follows: Figure 5 As shown, the time domain waveform of the acoustic emission signal received by the acoustic emission sensor of this embodiment is as follows: Figure 6 shown.

[0104] S2: All acoustic emission signals are processed by combining the adaptive analytical mode and Hilbert transform to obtain the adaptive marginal spectrum of the acoustic emission signal.

[0105] Adaptive analytical mode is used to decompose the acoustic emission signal collected in step S1 into multiple frequency-exclusive components:

[0106]

[0107] Where x(t) is the acoustic emission signal, are n mutually exclusive components of the spectrum; Figure 5 Taking the non-destructive acoustic emission signal of as an example, the adaptive analytical mode decomposition method is used to decompose it into 5 components. The original waveform of the non-destructive acoustic emission signal and the waveforms of the 5 components are as follows: Figure 7 shown; Figure 6 Taking the acoustic emission signal with pore defects as an example, the adaptive analytical mode decomposition method is used to decompose it into 5 components. The original waveform of the acoustic emission signal with pore defects and the waveforms of the 5 components are as follows: Figure 8 shown.

[0108] Perform Hilbert transform on the decomposed components of each acoustic emission signal to generate an analytical signal:

[0109]

[0110] Among them, z i (t) is the analytical signal, and H[x] represents the Hilbert transform of the signal x.

[0111] Calculate the instantaneous amplitude A of the analytical signal i (t) and instantaneous frequency f i (t):

[0112]

[0113] The instantaneous amplitudes of all components of the acoustic emission signal are superimposed with the frequency domain information in the time-frequency plane to construct a time-frequency spectrum (analytical mode decomposition - Hilbert spectrum):

[0114]

[0115] where H(f,t) is the analytical mode decomposition—the Hilbert spectrum, and δ[·] is the Dirac function, indicating that the energy is concentrated at the instantaneous frequency.

[0116] By integrating the time-frequency spectrum in the time dimension, the adaptive marginal spectrum M(f) of the acoustic emission signal is obtained, which is expressed as follows:

[0117]

[0118] Where T is the total duration of the signal; the adaptive marginal spectrum reflects the cumulative energy-frequency distribution of the acoustic emission signal in the entire period, which is used to identify the characteristic frequency of welding porosity defects. The adaptive marginal spectrum of the acoustic emission signal of the steel bridge deck welding non-damage working condition in this embodiment is as follows: Figure 9 As shown in the figure, the adaptive marginal spectrum of the acoustic emission signal of the steel bridge deck welding with porosity defects is as follows: Figure 10 shown.

[0119] S3: The target domain dataset is divided into a target domain labeled dataset and a target domain unlabeled dataset. The acoustic emission signals contained in the source domain labeled dataset and the target domain labeled dataset are labeled with "undamaged" and "with pore defects." The target domain labeled dataset contains acoustic emission signals with both intact and pore defects. The target domain labeled dataset includes 100 acoustic emission signals, and the target domain unlabeled dataset includes 300 acoustic emission signals. In this embodiment, the user actually knows whether the 1000 acoustic emission signals are intact or have pore defects, but the user only labels 600 acoustic emission signals in the source domain dataset and 100 acoustic emission signals in the target domain labeled dataset. The user does not label the 300 acoustic emission signals in the target domain unlabeled dataset. The above data label settings simulate scenarios in actual applications.

[0120] S4: The architecture of Wav2Vec2.0 is adjusted to obtain an improved AEWav2Vec2.0 model, and the AEWav2Vec2.0 model is trained using the adaptive marginal spectrum of the source domain annotated dataset to obtain an initial classification model for identifying the label category of the acoustic emission signal.

[0121] The adaptive marginal spectra of all acoustic emission signals are normalized in advance to obtain the model input tensor of the Wav2Vec2.0 model.

[0122] The way to adjust the architecture of Wav2Vec2.0 is as follows:

[0123] Based on the dimensionality of the adaptive marginal spectrum of the acoustic emission signal, the convolution kernel of the input convolution layer of the Wav2Vec2.0 model is modified to adapt to the long sequence characteristics of the adaptive marginal spectrum, the step size of the input convolution layer of the Wav2Vec2.0 model is modified to reduce the amount of computation, and the number of input channels of the input convolution layer of the Wav2Vec2.0 model is modified. Considering the limited number of acoustic emissions, which can easily lead to overfitting in the early stages, Dropout is added after the convolution layer of Wav2Vec2.0 to prevent overfitting and enhance the generalization ability of the model. Preferably, the Dropout rate is 0.1. Pooling is added after the convolution layer to reduce the dimensionality through downsampling, so that the subsequent layers of Wav2Vec2.0 only need to process the refined features. This reduces the number of convolution layers in the original Wav2Vec2.0 model from 7 to 3, without affecting the performance of the Wav2Vec2.0 model. The number of Transformer encoder layers is set to 8 to reduce the complexity of the Wav2Vec2.0 model to adapt to the situation with fewer acoustic emission signal samples. The Transformer output layer is replaced with a binary fully connected layer to distinguish between the two labels of "no damage" and "with pore defects".

[0124] Preferably, the adaptive marginal spectrum of the source domain annotated dataset is used to train the Wav2Vec2.0 model for supervised training, and the training optimization goal is to minimize the cross entropy loss function L cls , where the cross entropy loss function L cls The expression is

[0125]

[0126] Among them, y i is the real label of the acoustic emission signal in advance, p(y i |x i ) is the y predicted by the initial classification model i The label of the source domain is N, which refers to the number of acoustic emission signals contained in the source domain annotation dataset. Minimize the cross entropy loss function L cls , you can optimize the weights and biases of the convolutional layers in the feature extractor, the weight matrix of the attention layer in the Transformer encoder, and the classification weights of the output layer.

[0127] S5: The target domain annotated dataset is divided into a training set and a validation set according to a preset ratio. Both the training set and the validation set contain acoustic emission signals from both undamaged conditions and those with porosity defects. The training set and the source domain annotated dataset form a hybrid annotated dataset. The adaptive marginal spectrum of the hybrid annotated dataset is used to train the initial model to obtain an intermediate classification model. The preset ratio of the target domain annotated dataset is training set: validation set = 1:1, that is, the training set and validation set each contain 150 acoustic emission signals.

[0128] During the training of the initial model, the feature extraction layer parameters of the initial classification model are retained, the last three layers of the Transformer encoder are unfrozen, and the parameters of the last three layers of the unfrozen Transformer encoder are allowed to update during the training process to improve the adaptability of the intermediate classification model to the characteristics of the acoustic emission signals in the target domain annotated dataset.

[0129] S6: The adaptive marginal spectrum of the target domain unlabeled dataset is input into the intermediate classification model. The intermediate classification model outputs two classification results with confidence levels for the corresponding acoustic emission signals. Acoustic emission signals with confidence levels higher than a preset value Q1 are screened. These screened acoustic emission signals and the mixed labeled dataset form a joint training dataset. The intermediate classification model is iteratively optimized using the joint training dataset until the preset condition is met, and the iterative optimization is terminated to obtain the target classification model. The preset value Q1 is 0.9.

[0130] Set the joint loss function of the supervision loss and the intermediate classification model prediction label loss, and set the dynamic weight control mechanism, where the joint loss function L joint The expression is

[0131] L joint =λL sup +(1-λ)L predict

[0132] Among them, L predict is the cross entropy loss of the acoustic emission signal filtered based on confidence, L sup is the cross entropy loss of the mixed annotated dataset, and λ is the dynamic weight. The iterative optimization of the intermediate classification model is terminated when the dynamic weight λ decays to a preset value Q2, which is set to 0.3. In this embodiment, the initial value of the dynamic weight λ is 0.7, and the iterative optimization is terminated when the dynamic weight λ decays to 0.3.

[0133] S7: Input the adaptive marginal spectrum of the validation set into the target classification model, and the target classification model outputs two classification results with confidence for the corresponding acoustic emission signals. The acoustic emission signals with confidence higher than the preset value Q4 are screened to form a predicted and labeled data set. The predicted and labeled data set is added to the joint training data set to form a final training data set. The target classification model is iteratively optimized using the final training data set until the same preset condition as in step S6 is reached, and the iterative optimization is stopped to obtain the final classification model.

[0134] S8: The adaptive marginal spectrum of the target domain unlabeled dataset is input into the final classification model. The final classification model outputs the classification result and confidence level of each acoustic emission signal in the target domain unlabeled dataset. The final classification model also outputs whether each acoustic emission signal in the target domain dataset belongs to the predicted labeled dataset.

[0135] When the confidence level for a particular acoustic emission signal to have a porosity defect exceeds 0.5, the signal is determined to have a porosity defect. The classification results ("undamaged" or "with porosity defect") for 300 acoustic emission signals from the target domain unlabeled dataset, output by the final classification model, are compared with the true labels of these 300 acoustic emission signals in step S1. This example ultimately achieved a 96% recognition rate for undamaged samples and a 95% recognition rate for samples with porosity defects in the target domain unlabeled dataset, validating the accuracy, stability, and adaptability of the method for identifying porosity defects in steel bridge deck welds.

[0136] Example 3

[0137] The present invention discloses a steel bridge deck welding porosity defect recognition system, which includes a signal acquisition module, a signal processing module, a signal classification module, an initial classification module, an intermediate classification module, a target classification module, a final classification module and a result output module.

[0138] The signal acquisition module is configured to collect multiple acoustic emission signals, including those from a test steel bridge deck with both intact and porous defects, from the test steel bridge deck. The acoustic emission signals collected from the test steel bridge deck constitute a source domain dataset. Furthermore, the signal acquisition module is configured to collect multiple acoustic emission signals from the steel bridge deck to be tested. The acoustic emission signals collected from the steel bridge deck to be tested constitute a target domain dataset. The signal acquisition module executes step S1 in Example 1.

[0139] The signal processing module is used to process the acoustic emission signal by combining the adaptive analytical mode and the Hilbert transform to obtain the adaptive marginal spectrum of the acoustic emission signal. The signal processing module executes step S2 in embodiment 1.

[0140] A signal classification module is configured to divide the target domain dataset into a target domain labeled dataset and a target domain unlabeled dataset, and to label the acoustic emission signals contained in the source domain labeled dataset and the target domain labeled dataset using the terms "no damage" and "with pore defects," wherein the target domain labeled dataset includes acoustic emission signals of both no damage and with pore defects. The signal classification module executes step S3 in Example 1.

[0141] The initial classification module is used to adjust the Wav2Vec2.0 architecture to obtain an improved AEWav2Vec2.0 model. The AEWav2Vec2.0 model is trained using the adaptive marginal spectrum of the source domain annotated dataset to obtain an initial classification model for identifying the label category of the acoustic emission signal. The initial classification module performs step S4 in Example 1.

[0142] An intermediate classification module is configured to divide the target domain annotated dataset into a training set and a validation set according to a preset ratio, ensuring that both the training set and the validation set contain acoustic emission signals from both damage-free conditions and those with porosity defects. The training set and the source domain annotated dataset form a hybrid annotated dataset, and the adaptive marginal spectrum of the hybrid annotated dataset is used to train the initial model to obtain an intermediate classification model. The intermediate classification module executes step S5 in Example 1.

[0143] The target classification module is configured to input the adaptive marginal spectrum of the target domain unlabeled dataset into an intermediate classification model. The intermediate classification model outputs two classification results corresponding to the acoustic emission signals with confidence levels. Acoustic emission signals with confidence levels greater than a preset value Q1 are screened, and the screened acoustic emission signals and the mixed labeled dataset form a joint training dataset. The intermediate classification model is iteratively optimized using the joint training dataset until a preset condition is met, and the iterative optimization is terminated to obtain a target classification model. The target classification module executes step S6 in Example 1.

[0144] The final classification module is configured to input the adaptive marginal spectrum of the validation set into the target classification model. The target classification model outputs two classification results corresponding to the acoustic emission signals with confidence levels. Acoustic emission signals with confidence levels higher than a preset value Q4 are screened to form a predicted and labeled dataset. The predicted and labeled dataset is added to the joint training dataset to form a final training dataset. The target classification model is iteratively optimized using the final training dataset. The iterative optimization is terminated when a preset condition is met, and the final classification model is obtained. The final classification module executes step S7 in Example 1.

[0145] The result output module is configured to input the adaptive marginal spectrum of the target domain dataset into the final classification model. The final classification model outputs two classification results with confidence levels. When the confidence level that a particular acoustic emission signal contains a pore defect exceeds a preset value Q3, the acoustic emission signal is determined to contain a pore defect. The result output module executes step S8 in Example 1.

Claims

1. A method for identifying porosity defects in steel bridge deck welding, characterized by: The following steps are included: S1: Multiple acoustic emission signals, including those with and without damage, are collected on the test steel bridge deck to form a source domain dataset; multiple acoustic emission signals are collected on the steel bridge deck to be tested to form a target domain dataset; S2: Processing the acoustic emission signal by combining the adaptive analytical mode and Hilbert transform to obtain the adaptive marginal spectrum of the acoustic emission signal; S3: The target domain dataset is divided into a target domain labeled dataset and a target domain unlabeled dataset. The acoustic emission signals contained in the source domain labeled dataset and the target domain labeled dataset are labeled with "no damage" and "with pore defects". The target domain labeled dataset contains acoustic emission signals without damage and with pore defects. S4: Adjust the architecture of Wav2Vec2.0 to obtain an improved AEWav2Vec2.0 model, and use the adaptive marginal spectrum of the source domain annotated dataset to train the AEWav2Vec2.0 model to obtain an initial classification model for identifying the label category of the acoustic emission signal; S5: The target domain annotated dataset is divided into a training set and a validation set according to a preset ratio. It is ensured that both the training set and the validation set contain acoustic emission signals of both damage-free working conditions and those with pore defects. The training set and the source domain annotated dataset form a mixed annotated dataset. The adaptive marginal spectrum of the mixed annotated dataset is used to train the initial model to obtain an intermediate classification model. S6: Input the adaptive marginal spectrum of the unlabeled dataset of the target domain into the intermediate classification model. The intermediate classification model outputs two classification results with confidence for the corresponding acoustic emission signals. The acoustic emission signals with confidence higher than the preset value Q1 are screened, and the screened acoustic emission signals and the mixed labeled dataset constitute a joint training dataset. The intermediate classification model is iteratively optimized using the joint training dataset until the preset condition is met and the iterative optimization is stopped, and the target classification model is obtained. S7: Input the adaptive marginal spectrum of the validation set into the target classification model, and the target classification model outputs two classification results with confidence for the corresponding acoustic emission signals. The acoustic emission signals with confidence higher than the preset value Q4 are screened to form a predicted and labeled data set. The predicted and labeled data set is added to the joint training data set to form a final training data set. The target classification model is iteratively optimized using the final training data set until the same preset condition as in step S6 is reached, and the iterative optimization is stopped to obtain the final classification model. S8: Inputting the adaptive marginal spectrum of the target domain unlabeled dataset into the final classification model, the final classification model outputs the classification result and confidence of each acoustic emission signal of the target domain unlabeled dataset.

2. The method for identifying porosity defects in steel bridge deck welding according to claim 1, characterized in that: In step S1, the method for collecting the acoustic emission signal of the weld with porosity defects on the test steel bridge deck is: spraying the surface of the weld before welding.

3. The method for identifying steel bridge deck welding porosity defects according to claim 1, characterized in that: The steps of obtaining the adaptive marginal spectrum of the acoustic emission signal in step S2 are: Adaptive analytical mode is used to decompose the acoustic emission signal into spectrally mutually exclusive components; Perform Hilbert transform on all components of the acoustic emission signal to generate analytical signals; Calculate the instantaneous amplitude and instantaneous frequency of all analytical signals, and construct the time-frequency spectrum of the corresponding acoustic emission signal based on all the instantaneous amplitudes and instantaneous frequencies of each acoustic emission signal; The time-frequency spectrum is integrated along the time axis to obtain the adaptive marginal spectrum of each acoustic emission signal.

4. The method for identifying porosity defects in steel bridge deck welding according to claim 1, characterized in that: In step S4, the architecture of Wav2Vec2.0 is adjusted as follows: based on the dimension of the adaptive marginal spectrum of the acoustic emission signal, the convolution kernel, step size and number of input channels of the input convolution layer of the Wav2Vec2.0 model are modified; Dropout and Pooling are added after the convolution layer of Wav2Vec2.0, the number of encoder layers of the Transformer is set to 8, and the output layer of the Transformer is replaced with a binary classification fully connected layer.

5. The method for identifying porosity defects in steel bridge deck welding according to claim 1, characterized in that: Optimize the initial classification model in step S4, and the optimization goal is to minimize the cross entropy loss function L cls , L cls The expression is: where y i is the true label, p(y i |x i ) is the y predicted by the initial classification model i N refers to the number of acoustic emission signals contained in the source domain annotation dataset.

6. The method for identifying porosity defects in steel bridge deck welding according to claim 1, characterized in that: In step S5, during the training of the initial model, the feature extraction layer parameters of the initial classification model are retained, the last three layers of the Transformer encoder are unfrozen, and the parameters of the last three layers of the unfrozen Transformer encoder are allowed to be updated during the training process.

7. The method for identifying porosity defects in steel bridge deck welding according to claim 1, characterized in that: The steps for setting the preset conditions in step S6 are as follows: Set the joint loss function L of the supervision loss and the intermediate classification model prediction label loss joint , and L joint The expression is L joint =λL sup +(1-λ)L predict Among them, L predict is the cross entropy loss of the acoustic emission signal filtered based on confidence, L sup is the cross entropy loss of the mixed labeled dataset, and λ is the dynamic weight; When the dynamic weight λ decays to the preset value Q2, the iterative optimization stops.

8. The method for identifying steel bridge deck welding porosity defects according to claim 1, characterized in that: The final classification model in step S8 also outputs whether each acoustic emission signal of the target domain dataset belongs to the predicted annotation dataset.

9. The method for identifying porosity defects in steel bridge deck welding according to claim 1, characterized in that: In step S8, the acoustic emission signals whose confidence level output by the final classification model is higher than the preset value Q4 are screened, and the screened acoustic emission signals are added to the final training data set. The target classification model is iteratively optimized again using the final training data set until the same preset conditions as in step S6 are reached and the iterative optimization is stopped.

10. A steel bridge deck welding porosity defect identification system, characterized by: include, A signal acquisition module is used to collect multiple acoustic emission signals including those without damage and those with porosity defects on the test steel bridge deck, and the acoustic emission signals collected on the test steel bridge deck constitute a source domain data set; and to collect multiple acoustic emission signals on the steel bridge deck to be tested, and the acoustic emission signals collected on the steel bridge deck to be tested constitute a target domain data set; A signal processing module is used to process the acoustic emission signal by combining the adaptive analytical mode and the Hilbert transform to obtain the adaptive marginal spectrum of the acoustic emission signal; The signal classification module is used to divide the target domain dataset into a target domain labeled dataset and a target domain unlabeled dataset, and annotate the acoustic emission signals contained in the source domain labeled dataset and the target domain labeled dataset using "no damage" and "with pore defects". The target domain labeled dataset contains acoustic emission signals without damage and with pore defects. The initial classification module is used to adjust the architecture of Wav2Vec2.0 to obtain an improved AEWav2Vec2.0 model. The AEWav2Vec2.0 model is trained using the adaptive marginal spectrum of the source domain annotated dataset to obtain an initial classification model for identifying the label category of the acoustic emission signal. The intermediate classification module is used to divide the target domain annotated dataset into a training set and a validation set according to a preset ratio, and ensure that both the training set and the validation set contain acoustic emission signals of both damage-free working conditions and those with pore defects. The training set and the source domain annotated dataset form a mixed annotated dataset, and the adaptive marginal spectrum of the mixed annotated dataset is used to train the initial model to obtain the intermediate classification model. The target classification module is used to input the adaptive marginal spectrum of the unlabeled dataset of the target domain into the intermediate classification model. The intermediate classification model outputs two classification results with confidence for the corresponding acoustic emission signals, and selects acoustic emission signals with confidence higher than the preset value Q1. The selected acoustic emission signals and the mixed labeled dataset constitute a joint training dataset. The intermediate classification model is iteratively optimized using the joint training dataset until the preset conditions are met and the iterative optimization is stopped, and the target classification model is obtained. The final classification module is used to input the adaptive marginal spectrum of the validation set into the target classification model. The target classification model outputs two classification results with confidence for the corresponding acoustic emission signals. The acoustic emission signals with confidence higher than the preset value Q4 are screened to form a predicted annotation dataset. The predicted annotation dataset is added to the joint training dataset to form the final training dataset. The target classification model is iteratively optimized using the final training dataset. When the preset conditions are met, the iterative optimization is stopped and the final classification model is obtained. The result output module is used to input the adaptive marginal spectrum of the unlabeled data set in the target domain into the final classification model, and the final classification model outputs the classification result and confidence of each acoustic emission signal of the unlabeled data set in the target domain.

Images