Deep learning-based ultrasonic guided wave signal and key feature extraction method, system and device, and medium

Abstract

The invention discloses an ultrasonic guided wave signal and key feature extraction method, system and device based on deep learning and a medium, and belongs to the technical field of ultrasonic guided wave signal and key feature extraction, and the method comprises the steps: collecting a guided wave signal of a steel strand, and carrying out the preprocessing of the guided wave signal; converting the pre-processed guided wave signal into a time-frequency spectrum, extracting a frequency peak value, envelope energy and group velocity features, and performing dimension reduction processing on a feature set to generate a feature vector; inputting the feature vectors into a deep learning network model, extracting convolution features and time sequence features of the signals, and completing model parameter updating based on a training sample; the trained deep learning network model is used for reasoning the steel strand guided wave signals, classification labels of corrosion types and regression prediction values of corrosion depths are output, and the specific positions and corrosion degrees of defects are judged. According to the method, the positioning precision, the recognition accuracy and the cross-scene robustness of the ultrasonic guided waves on the corrosion defect of the steel strand under the complex working condition are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of ultrasonic guided wave signal and key feature extraction technology, specifically to a method, system, device, and medium for ultrasonic guided wave signal and key feature extraction based on deep learning. Background Technology

[0002] Currently, common methods for corrosion detection of overhead ground wire steel strands include impedance measurement and ultrasonic measurement. However, impedance measurement is limited by environmental factors and the complex structure of overhead ground wires, resulting in low detection accuracy and the inability to achieve online, real-time monitoring. While ultrasonic measurement offers high sensitivity, the propagation of ultrasonic guided waves within the steel strands involves modal dispersion, energy attenuation, and complex scattering effects, leading to difficult signal extraction and unstable characteristics, thus affecting the accuracy of corrosion detection.

[0003] Existing feature extraction methods for ultrasonic guided wave detection technology largely rely on human experience and signal processing techniques (such as time-domain analysis, frequency-domain analysis, and time-frequency analysis). However, these methods suffer from incomplete feature extraction information and high feature redundancy. Furthermore, manually designed features often fail to fully utilize the complex temporal and modal information in ultrasonic guided wave signals, resulting in low automation of corrosion detection, susceptibility to noise and environmental interference, and ultimately affecting recognition accuracy.

[0004] While deep learning methods offer certain advantages in ultrasonic guided wave signal analysis, existing research largely focuses on simple pattern recognition or classification problems, failing to fully consider the complex propagation characteristics and corrosion modes of ultrasonic guided waves in steel strands. For example, although convolutional neural networks perform well in extracting spatial features, they are insufficient in handling the temporal dependencies and dynamic changes in ultrasonic guided wave signals. While recurrent neural networks, long short-term memory networks, or Transformers can capture temporal features, when faced with the complex multimodal characteristics and strong nonlinear effects of ultrasonic guided waves, the extracted features still suffer from information loss and classification confusion.

[0005] Existing ultrasonic guided wave-based corrosion detection methods also have shortcomings in database construction and feature matching. Due to the complexity of ultrasonic guided wave signals and the diversity of corrosion morphologies, the existing databases are insufficient in size and diversity, which limits the generalization ability of the algorithms, especially when facing different materials, environments, and corrosion types, resulting in insufficient robustness and adaptability.

[0006] This invention aims to address the aforementioned technical bottlenecks by combining deep learning with ultrasonic guided wave propagation theory, improving feature extraction methods, enhancing the algorithm's adaptability in complex environments, and improving the accuracy and stability of corrosion detection by constructing a large-scale, diverse corrosion feature database. Summary of the Invention

[0007] In view of the above-mentioned problems, the present invention is proposed.

[0008] Therefore, the technical problem solved by this invention is: how to address the fact that existing feature extraction methods for ultrasonic guided wave detection technology rely heavily on human experience and signal processing techniques (such as time-domain analysis, frequency-domain analysis, and time-frequency analysis), which suffer from incomplete feature extraction information and high feature redundancy. Furthermore, manually designed features often fail to fully utilize the complex temporal and modal information in ultrasonic guided wave signals, resulting in low automation levels in corrosion detection and susceptibility to noise and environmental interference.

[0009] To address the aforementioned technical problems, this invention provides the following technical solution: a method for extracting ultrasonic guided wave signals and key features based on deep learning, comprising: acquiring guided wave signals from steel strands; preprocessing the guided wave signals; converting the preprocessed guided wave signals into a time-frequency spectrum, extracting peak frequency, envelope energy, and group velocity features, and performing dimensionality reduction processing on the feature set to generate feature vectors; inputting the feature vectors into a deep learning network model containing a three-layer convolutional neural network and a two-layer bidirectional long short-term memory network, extracting convolutional features and temporal features of the signal, and updating the model parameters based on training samples; using the trained deep learning network model to infer the guided wave signals from the steel strands, outputting classification labels for corrosion types and regression prediction values ​​for corrosion depth, and determining the specific location and degree of corrosion of defects.

[0010] As a preferred embodiment of the deep learning-based ultrasonic guided wave signal and key feature extraction method described in this invention, the guided wave signal includes an ultrasonic sensor with good coupling to the steel strand, and is acquired by exciting a high-frequency narrowband pulse signal through an external excitation source.

[0011] As a preferred embodiment of the deep learning-based ultrasonic guided wave signal and key feature extraction method described in this invention, the external excitation source includes an excitation signal whose center frequency is set according to the diameter of the steel strand, the expected detection distance, and the sensitivity requirements.

[0012] As a preferred embodiment of the deep learning-based ultrasonic guided wave signal and key feature extraction method described in this invention, the preprocessing includes noise suppression and signal amplitude standardization in the guided wave signal.

[0013] As a preferred embodiment of the deep learning-based ultrasonic guided wave signal and key feature extraction method described in this invention, the step of converting the preprocessed guided wave signal into a time-frequency spectrum includes mapping the time characteristics of the guided wave signal into a time-frequency domain distribution using short-time Fourier transform or continuous wavelet transform, extracting the frequency peak, envelope energy, and group velocity as key parameters reflecting defect characteristics, and performing dimensionality reduction processing on the key parameter set to remove redundant information and retain the main discriminative features, thereby generating a feature vector that can be learned by a deep learning network model.

[0014] This preferred solution can effectively map a one-dimensional time series into a two-dimensional time-frequency graph, significantly improving the distinguishability between defect waveforms and background noise. Furthermore, it extracts key physical features such as frequency peaks, envelope energy, and group velocity, and combines dimensionality reduction processing to remove redundant information, which helps improve model training efficiency and discrimination accuracy. In particular, it can improve recognition stability in situations with high data dimensionality and a small number of samples.

[0015] As a preferred embodiment of the deep learning-based ultrasonic guided wave signal and key feature extraction method described in this invention, the deep learning network model includes a three-layer convolutional neural network for convolutional feature extraction and a two-layer bidirectional long short-term memory network for temporal feature extraction. The three-layer convolutional neural network is configured with 32, 64, and 128 convolutional kernels respectively, with a kernel size of 3×3. After each convolutional layer, a batch normalization layer, a ReLU activation function, and a 2×2 max pooling layer are sequentially connected to extract the convolutional features of the guided wave signal, including frequency distribution and signal peak values. The first layer of the bidirectional long short-term memory network is configured with 128 units, and the second layer with 64 units. Dropout layers are added between layers to reduce overfitting and are used to extract the temporal features of the guided wave signal in the propagation direction and between defect echoes. The three-layer convolutional neural network and the two-layer bidirectional long short-term memory network form a cascaded structure. The deep learning network model backpropagates by using the cross-entropy loss function or the mean square error loss function to measure the error between the input feature vector and the labeled samples, and updates the parameters using the Adam optimizer.

[0016] This preferred solution constructs a deep learning network model structure that combines a three-layer convolutional neural network and a two-layer bidirectional long short-term memory network. The former can automatically extract the spatial characteristics of the frequency and energy distribution of the ultrasonic guided wave signal, while the latter can effectively capture the temporal correlation of the signal during propagation, thus achieving joint modeling of time and frequency information.

[0017] As a preferred embodiment of the deep learning-based ultrasonic guided wave signal and key feature extraction method described in this invention, the following steps are included: The trained deep learning network model is used to infer the guided wave signal of the steel strand. This includes: the trained deep learning network model receives the feature vector of the guided wave signal to be detected, outputs a classification label including the corrosion type and a regression prediction value of the corrosion depth; the classification label is determined by the maximum value of the category probability calculated by the Softmax activation function, and the regression prediction value is output through a linear output layer; the physical location of the defect in the steel strand is determined by combining the time information of the excitation time and the echo peak time, as well as the propagation speed of the guided wave; the corrosion location and depth identification results are used to determine whether a preset threshold is exceeded, and an evaluation report containing the defect location, corrosion type, and corrosion degree is output.

[0018] This preferred solution introduces a dual-task output mechanism to classify corrosion types and predict corrosion depth. It also combines the excitation-echo time difference and propagation speed to calculate defect location. This method can not only identify the existence of defects but also quantify the degree of corrosion and its specific location. It is particularly suitable for long-distance inspection scenarios such as power steel strands. It can achieve remote, non-contact, and high-precision corrosion location and alarm, and has strong engineering practical value and deployment adaptability.

[0019] Another objective of this invention is to provide a deep learning-based ultrasonic guided wave signal and key feature extraction system.

[0020] To address the aforementioned technical problems, this invention provides the following technical solution: a deep learning-based ultrasonic guided wave signal and key feature extraction system, comprising: a signal acquisition module, a feature extraction module, a deep learning network model module, and a defect identification module; the signal acquisition module is used to acquire guided wave signals from steel strands and preprocess the guided wave signals; the feature extraction module is used to convert the preprocessed guided wave signals into a time-frequency spectrum, extract frequency peaks, envelope energy, and group velocity features, and perform dimensionality reduction processing on the feature set to generate feature vectors; the deep learning network model module is used to input the feature vectors into a deep learning network model containing a three-layer convolutional neural network and a two-layer bidirectional long short-term memory network, extract the convolutional features and temporal features of the signal, and update the model parameters based on training samples; the defect identification module is used to infer the guided wave signals of steel strands using the trained deep learning network model, output a classification label for corrosion type and a regression prediction value for corrosion depth, and determine the specific location and degree of corrosion of the defects.

[0021] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the computer program to implement the steps of the deep learning-based ultrasonic guided wave signal and key feature extraction method.

[0022] The present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the deep learning-based ultrasonic guided wave signal and key feature extraction method.

[0023] The beneficial effects of this invention are as follows: This invention combines ultrasonic guided wave propagation theory with deep learning modeling technology, enabling the simultaneous extraction and utilization of the spatial and temporal features of ultrasonic guided wave signals, overcoming the limitations of traditional methods based on manual feature engineering or single time-frequency analysis. By employing convolutional neural networks to extract the time-frequency features of guided wave signals and combining this with long short-term memory networks to learn the temporal characteristics of guided wave signals, the model can effectively address the dispersion, attenuation, and mode conversion problems of ultrasonic guided waves propagating in steel strands. Compared to traditional time-domain and frequency-domain methods, deep learning methods can automatically extract and learn the complex mode information implicit in ultrasonic guided wave signals, improving the accuracy of defect location and corrosion degree assessment.

[0024] This invention introduces an attention mechanism to dynamically adjust the model's focus on different temporal features and spectral regions during the model learning process, significantly improving the model's feature extraction capabilities from weak echo signals and complex background noise. This adaptive feature extraction method effectively overcomes the effects of energy attenuation, signal aliasing, and noise interference in ultrasonic guided wave signals, enabling the model to maintain high recognition accuracy even under low signal-to-noise ratio conditions. Furthermore, the invention employs transfer learning and data augmentation strategies (such as signal perturbation, translation, and amplitude scaling) to expand the model's adaptability and robustness under different operating conditions, solving problems such as uneven data distribution and diverse defect types in complex environments.

[0025] This invention also constructs a multimodal ultrasonic guided wave feature database, covering different materials, environments, and corrosion modes, overcoming the shortcomings of existing ultrasonic guided wave detection methods in terms of database size and diversity. Through the combination of deep learning and a large-scale database, the model can maintain stable detection accuracy under different working conditions and possesses strong cross-scenario generalization ability. In actual testing, the invented model exhibits high accuracy and stability in corrosion localization, depth prediction, and classification tasks. Especially under conditions of modal aliasing and complex interference, the model still maintains superior recognition accuracy and fast response capability compared to traditional methods.

[0026] Therefore, this invention enhances the learning and modeling capabilities of ultrasonic guided wave signals under complex working conditions by introducing deep learning modeling, adaptive feature extraction, and an attention mechanism, thus overcoming the bottlenecks of traditional ultrasonic guided wave detection methods in terms of accuracy, stability, and generalization. It exhibits significant advantages in corrosion defect localization accuracy, identification accuracy, real-time response capability, and cross-scene adaptability, demonstrating broad prospects for engineering applications. Attached Figure Description

[0027] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 The above is a flowchart of an ultrasonic guided wave signal and key feature extraction method based on deep learning, provided as an embodiment of the present invention.

[0029] Figure 2 The flowchart illustrates an algorithm for a deep learning-based method for extracting ultrasonic guided wave signals and key features, as provided in one embodiment of the present invention.

[0030] Figure 3 This is a deep learning network model structure diagram of a deep learning-based ultrasonic guided wave signal and key feature extraction method provided in one embodiment of the present invention.

[0031] Figure 4 This is a schematic diagram of a deep learning-based ultrasonic guided wave signal and key feature extraction system provided in one embodiment of the present invention. Detailed Implementation

[0032] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0033] Example 1, referring to Figures 1-3 This is one embodiment of the present invention, which provides a method for extracting ultrasonic guided wave signals and key features based on deep learning, including:

[0034] S1. Acquire the guided wave signal of the steel strand and preprocess the guided wave signal.

[0035] S2. Convert the preprocessed guided wave signal into a time spectrum, extract the frequency peak, envelope energy, and group velocity features, and perform dimensionality reduction on the feature set to generate feature vectors.

[0036] S3. Input the feature vector into a deep learning network model containing a three-layer convolutional neural network and a two-layer bidirectional long short-term memory network, extract the convolutional features and temporal features of the signal, and update the model parameters based on the training samples.

[0037] S4. Use the trained deep learning network model to infer the guided wave signal of the steel strand, output the classification label of corrosion type and the regression prediction value of corrosion depth, and determine the specific location and degree of corrosion of the defect.

[0038] Current research on using ultrasonic guided waves to detect corrosion of overhead ground wire steel strands struggles to adequately address the complex propagation characteristics of multi-mode guided waves, such as dispersion and attenuation. Traditional methods often rely solely on single feature analysis in the time or frequency domains, leading to information gaps or misclassification. Furthermore, capturing critical features sensitive to defects often requires extensive manual experience or repeated parameter tuning, lacking both systematic adaptability and the ability to achieve accurate and rapid corrosion diagnosis across a wider range. In addition, the service environment of steel strands is highly variable, with numerous defect types, and existing models lack adaptability across different scenarios and devices. Once the signal characteristics differ significantly from those in the training phase, the detection accuracy drops markedly.

[0039] Against this backdrop, this invention addresses key challenges commonly encountered in practical engineering, such as noise interference and uneven distribution of monitoring data, by proposing an adaptive feature extraction and intelligent evaluation method. This method combines various signal processing techniques and deep learning models. First, in the preprocessing stage, it performs noise reduction and redundancy removal operations on the raw ultrasonic guided wave signals from different operating conditions, mitigating the impact of complex environmental noise on feature extraction. Subsequently, during feature selection and model training, an adaptive algorithm mechanism is introduced to fully learn and represent dispersion, multimodal features, and waveform instability, effectively addressing training biases caused by limited or unbalanced test sample numbers. Finally, the algorithm, driven by data, improves the accuracy of defect location and corrosion degree identification in overhead ground wire steel strands, while ensuring good robustness and generalization ability in real-time online monitoring. Through this adaptive evaluation system, users can promptly grasp the development trend of corrosion defects and obtain continuous and stable detection results in various scenarios, providing reliable technical support for the formulation and implementation of subsequent maintenance strategies.

[0040] like Figure 2 As shown, the entire algorithm consists of six core stages:

[0041] Data collection phase:

[0042] In this stage, an ultrasonic guided wave excitation source (typically a narrowband sinusoidal pulse with a center frequency of 80kHz to 200kHz) is configured using a magnetostrictive or piezoelectric sensor, and the echo signal is recorded at a high sampling rate (above 1MHz). To ensure a high signal-to-noise ratio, a suitable sensor coupling method (such as surround or bonded coupling) is configured.

[0043] Signal preprocessing stage:

[0044] The acquired raw guided wave signals are often affected by environmental noise and equipment errors, requiring noise reduction methods such as wavelet thresholding and Empirical Mode Decomposition (EMD). Low-frequency and high-frequency interference signals are then eliminated using bandpass filtering or adaptive filtering. To enhance signal stability, signal normalization and amplitude scaling can also be performed.

[0045] Feature extraction stage:

[0046] In feature extraction, the time-domain signal is converted into a time-spectrum using Short-Time Fourier Transform (STFT) or Continuous Wavelet Transform (CWT) to extract features such as frequency peaks, envelope energy, and group velocity. Since the number of features is usually large, Principal Component Analysis (PCA) or Linear Discriminant Analysis (LDA) is used to reduce redundant information and perform dimensionality reduction on the feature set, generating relatively stable feature vectors.

[0047] Deep learning modeling stage:

[0048] A convolutional neural network (CNN) is used to extract spatial features of the signal, such as frequency distribution and signal peak values. A long short-term memory network (LSTM) is used to capture the temporal dynamics of the ultrasound signal, such as phase changes and group velocity. The model output includes corrosion location and corrosion level (regression or classification).

[0049] Training and optimization phase:

[0050] The loss function can be chosen from cross-entropy or mean squared error (MSE) depending on the task type. The Adam or Stochastic Gradient Descent (SGD) optimizer is used, combined with a learning rate decay strategy to adjust the network parameters. To address the data imbalance problem, data augmentation (such as translation, noise addition, amplitude perturbation, etc.) or transfer learning strategies can be introduced to accelerate training using existing models.

[0051] Reasoning and defect assessment phase:

[0052] Through model inference, the physical location of defects (based on echo delay and propagation velocity) and the degree of corrosion (based on echo peak and envelope energy) are located. Finally, an evaluation report is generated, outputting detailed parameters including corrosion location, corrosion depth, and damage level.

[0053] like Figure 3 As shown, the input to this deep learning network model is the time-domain signal (such as a sampling sequence of length 2048×2048) acquired from ultrasonic guided waves. It is converted into a time spectrum through short-time Fourier transform or continuous wavelet transform, forming a two-dimensional matrix, which is used as the input to the CNN.

[0054] Part 1: Convolutional Feature Extraction

[0055] The convolutional layer section uses a three-layer convolutional network with 32, 64, and 128 convolutional kernels respectively, all with a size of 3×3. Batch normalization is applied after each convolutional layer to accelerate convergence, and the ReLU activation function is used to introduce nonlinearity. The max-pooling layer has a window size of 2×2 to reduce the dimensionality and redundancy of the feature map. Through these three convolutional and pooling operations, the network can progressively extract spatial features related to ultrasonic guided wave characteristics, including peak frequencies and envelope energy.

[0056] Part Two: Temporal Feature Extraction

[0057] The flattened feature map is fed into a bidirectional LSTM (BiLSTM) layer to learn the temporal dependence and nonlinear relationships of the signal. The first BiLSTM layer contains 128 units, followed by a Dropout layer (0.5) to reduce the risk of overfitting. The second BiLSTM layer contains 64 units for further extraction of sequence features. Through the bidirectional structure, the LSTM can simultaneously capture forward and backward temporal features, enhancing the model's sensitivity to signal propagation direction and defect echoes.

[0058] Part Three: Output and Classification / Regression

[0059] The output of the BiLSTM layer is fed into two fully connected layers, containing 128 and 64 neurons respectively, using the ReLU activation function to introduce non-linearity. After the second fully connected layer, the network generates the final result through an output layer. Depending on the task, the output layer can be a softmax layer (for erosion type classification) or a linear layer (for erosion depth regression).

[0060] A CNN was used to extract spatial pattern information of ultrasonic guided waves, while an LSTM was used to capture the temporal dependence of sequence data. Classification and regression outputs were then performed using a fully connected network. Batch normalization and Dropout were employed to enhance generalization ability, and a bidirectional LSTM network was used to strengthen the temporal modeling capability of the signal. Due to the use of STFT or CWT conversion, the network has a stronger ability to learn and recognize complex signal features such as dispersion effects and waveform distortion of ultrasonic guided waves. Ultimately, high-precision corrosion detection and classification were achieved in practical engineering.

[0061] It should be noted that steel strands are widely used in overhead power lines, bridge structures, and prestressed engineering. During operation, they are susceptible to environmental corrosion and mechanical fatigue, leading to the accumulation of hidden damage. This invention, through joint modeling using ultrasonic guided wave signals and deep learning technology, can identify the corrosion type and quantitatively predict the corrosion depth of steel strands without damaging the structure. Furthermore, by combining propagation delay information, it can infer the location of defects. This invention is suitable for scenarios such as remote inspection of power lines and monitoring of bridge cables, helping to improve structural operational safety and maintenance response efficiency.

[0062] Example 2, an embodiment of the present invention, provides a deep learning-based method for extracting ultrasonic guided wave signals and key features, based on the previous embodiment, including:

[0063] In this embodiment of the application, the guided wave signal of the steel strand is acquired in step S1, and the guided wave signal is preprocessed.

[0064] Preprocessing includes noise suppression in the guided wave signal and normalization of the signal amplitude.

[0065] During ultrasonic guided wave signal acquisition, an ultrasonic sensor (such as a magnetostrictive or piezoelectric type) with good coupling to the steel strand is required, and a high-frequency narrowband pulse signal is excited by an external excitation source. The center frequency of the excitation pulse can be selected in the range of tens to hundreds of kilohertz, and the specific value is determined by the diameter of the steel strand, the expected detection distance, and the sensitivity requirements. To ensure that the signal has a certain energy intensity when propagating in the steel strand, the excitation amplitude can be controlled between tens of volts and hundreds of volts, and the number of pulses is usually selected as 5 to 7 peaks. To simplify the propagation process of the ultrasonic guided wave in the steel strand, the following wave equation can be used to approximately describe the motion law of the guided wave:

[0066]

[0067] Where u represents displacement, and c is the propagation speed of the ultrasonic wave in the steel strand, which is closely related to the material's elastic modulus and density. The acquisition end can use a high-speed data acquisition card (sampling rate above 1MHz) to acquire the echo signal, and use a time-domain window to capture a duration of 0.5ms to 2ms to ensure that defect reflection information of the entire steel strand can be captured.

[0068] In signal preprocessing, the focus is on denoising and feature extraction. For suppressing complex noise, methods such as wavelet thresholding and empirical mode decomposition can be used to enhance the effective echo component by filtering out high-frequency pulses and low-frequency interference. Subsequently, in time-frequency analysis, short-time Fourier transform or continuous wavelet transform can be used to map the one-dimensional time-domain signal onto a two-dimensional time-frequency spectrum, allowing for a more intuitive observation of energy accumulation and phase changes caused by defects. To quantify the relationship between the reflected wave amplitude and the amplitude of the main wave peak at the defect, a reflection coefficient can be defined:

[0069]

[0070] Where R is the reflection coefficient, A defect A represents the amplitude of the main peak of the echo signal generated at the defect. excitation The reference amplitude is used for the excitation pulse. Besides the amplitude, features such as arrival time and envelope energy should also be extracted, and redundant information should be removed using dimensionality reduction techniques such as PCA or LDA. For imbalanced data under actual operating conditions, data augmentation operations such as random noise addition and shifting can be considered at this stage to provide more diverse feature samples for subsequent deep learning modeling.

[0071] In one optional implementation, the preprocessing step employs the EMD method to decompose the original signal into multiple intrinsic mode functions (IMFs). Based on energy distribution and spectral characteristics, reconstructed components are selected to filter out high-frequency interference and systematic low-frequency drift. Subsequently, the EMD-reconstructed signal is normalized, adjusting its amplitude to the range of [-1, 1] to improve signal amplitude consistency and model input stability in subsequent feature extraction stages.

[0072] In another alternative implementation, for guided wave signals with severe edge waveforms, a local weighted regression smoothing method is used to smooth the signal curve to enhance its shape preservation capability, preserving the main peak structure and eliminating local oscillation components. Combined with amplitude scaling, the signal amplitude is standardized to a set engineering range, improving the comparability of test data from different batches under the same standard.

[0073] This invention provides a clearer and more stable signal foundation for subsequent feature extraction, improving the training efficiency and generalization ability of the recognition model.

[0074] In this embodiment of the application, in step S2, the preprocessed guided wave signal is converted into a time spectrum, the frequency peak, envelope energy, and group velocity features are extracted, and the feature set is dimensionality reduced to generate a feature vector.

[0075] The time characteristics of the guided wave signal are mapped to a time-frequency domain distribution using short-time Fourier transform or continuous wavelet transform. The peak frequency, envelope energy and group velocity are extracted as key parameters reflecting the defect characteristics. The set of key parameters is then dimensionality-reduced to remove redundant information and retain the main discriminative features, generating feature vectors that can be learned by deep learning network models.

[0076] In one alternative implementation, the guided wave signal is decomposed using Empirical Wavelet Transform (EWT) to obtain amplitude envelopes and instantaneous energy information of multiple frequency bands, which are used to construct a multidimensional feature set. During the dimensionality reduction process, the Minimum Redundancy Maximum Relevance (mRMR) algorithm is used to select several feature dimensions that are most relevant to the corrosion category and have the lowest redundancy, forming a compact feature vector for model input.

[0077] In another alternative implementation, the original guided wave signal is decomposed into several intrinsic mode functions (IMFs) based on empirical mode decomposition. For each IMF component, its envelope energy, dominant frequency, and instantaneous phase are calculated to form an original feature set. Then, linear discriminant analysis is used for dimensionality reduction to enhance the separability between different categories of features, thereby improving the classification performance of subsequent deep learning network models.

[0078] This invention transforms non-stationary guided wave signals into two-dimensional time-spectrum diagrams, which can significantly enhance the visibility of defect-induced frequency drift and energy focusing. Combined with PCA for dimensionality reduction, it can retain dominant features while compressing the impact of high-dimensional features on model training efficiency. Thus, it improves the computational efficiency and generalization ability of the model while maintaining recognition accuracy, and is especially suitable for typical engineering scenarios with uneven sample distribution in industrial environments.

[0079] In this embodiment of the application, in step S3, the feature vector is input into a deep learning network model containing a three-layer convolutional neural network and a two-layer bidirectional long short-term memory network to extract the convolutional features and temporal features of the signal, and the model parameters are updated based on the training samples.

[0080] The deep learning network model includes a three-layer convolutional neural network for convolutional feature extraction and a two-layer bidirectional long short-term memory network for temporal feature extraction. The three-layer convolutional neural network is configured with 32, 64 and 128 convolutional kernels respectively, with a kernel size of 3×3. After each convolutional layer, a batch normalization layer, a ReLU activation function and a 2×2 max pooling layer are connected in sequence to extract the convolutional features of the guided wave signal, including frequency distribution and signal peak values.

[0081] The first layer of the bidirectional long short-term memory network is configured with 128 units, and the second layer is configured with 64 units. Dropout layers are added between layers to reduce overfitting and are used to extract the temporal characteristics of the guided wave signal in the propagation direction and between the defect echo.

[0082] A three-layer convolutional neural network and a two-layer bidirectional long short-term memory network form a cascaded structure. The deep learning network model uses the error between the input feature vector and the labeled sample, and performs backpropagation using the cross-entropy loss function or the mean squared error loss function, and uses the Adam optimizer to update the parameters.

[0083] Specifically, in model training and optimization, this invention uses a deep learning network to associate feature parameters with the degree of erosion or defect type. If it is necessary to simultaneously identify defect location and erosion depth, the time-frequency features can first be converted into a matrix (similar to a grayscale image), and then a multi-layer convolutional neural network can be used for automatic feature extraction and classification. If it is necessary to preserve strong temporal relationships, a long short-term memory network or a bidirectional LSTM can be used to process the transformed time-domain sequence. Taking defect binary classification ("erosion" or "no erosion") as an example, a typical cross-entropy loss function can be written as:

[0084]

[0085] Where N is the number of samples, y i It's a real label. These are model predictions. If corrosion depth is to be predicted using regression, the mean squared error loss function can be used instead.

[0086]

[0087] These represent the actual corrosion depth and the predicted corrosion depth, respectively. The network parameters are iteratively updated using optimizers such as Adam and SGD, and learning rate decay strategies or early stopping strategies are incorporated to prevent overfitting.

[0088]

[0089] Among them, z i and These represent the actual corrosion depth and the predicted corrosion depth, respectively. The network parameters are iteratively updated using optimizers such as Adam and SGD, and learning rate decay strategies or early stopping strategies are incorporated to prevent overfitting.

[0090] In the embodiments of this application, in step S4, the trained deep learning network model is used to infer the guided wave signal of the steel strand, output the classification label of corrosion type and the regression prediction value of corrosion depth, and determine the specific location and degree of corrosion of the defect.

[0091] The trained deep learning network model receives the feature vector of the guided wave signal to be detected and outputs a classification label for the corrosion type and a regression prediction value for the corrosion depth. The classification label is determined by the maximum value of the category probability calculated by the Softmax activation function, and the regression prediction value is output through a linear output layer.

[0092] By combining the time information of the excitation time and the echo peak time, as well as the propagation speed of the guided wave, the physical location of the defect in the steel strand is determined.

[0093] The corrosion location and depth identification results are used to determine whether the preset threshold is exceeded, and an evaluation report containing the defect location, corrosion type and corrosion degree is output.

[0094] In defect location and corrosion identification applications, to accurately locate defects, the time delay and wave velocity can be combined to obtain the defect location:

[0095]

[0096] Where t0 is the excitation time, t1 is the time when the defect echo peak appears, c is the effective propagation speed of the corresponding guided wave mode, and the denominator coefficient 2 represents the round-trip path of the ultrasonic wave in this segment. If multiple sensor arrays are deployed, a multi-sensor fusion algorithm can be used to further improve the positioning accuracy. At the model inference end, this invention can determine whether the corrosion depth exceeds a certain threshold based on the regression results, and issue an alarm or priority maintenance suggestion based on the exceedance.

[0097] Example 3: An embodiment of the present invention provides a method for extracting ultrasonic guided wave signals and key features based on deep learning. To verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0098] To illustrate the feasibility and adaptability of this invention in practical applications, two different implementation environments are presented below.

[0099] The first implementation was conducted in a laboratory environment. Artificial grooving (groove depth approximately 1 mm) was performed on a 2 m long steel strand. Magnetostrictive sensors were deployed at both ends, and narrowband pulses with a center frequency of 150 kHz were applied to excite them. After sampling, wavelet denoising was applied to the echo signals, and the time-frequency energy distribution was extracted as input to a CNN. Finally, through cross-validation of the training and validation sets, the corrosion depth recognition accuracy reached over 93%.

[0100] The second implementation environment is based on an actual operational railway line. A pair of magnetostrictive sensors are deployed at each end and the middle of the line. The collected guided wave signals are transmitted back to the backend server via a wireless communication module. The backend server performs real-time inference on the signals based on a pre-trained bidirectional LSTM network, identifying a corrosion point approximately 3.2m from one end of the sensor, with a corrosion depth of about 13% of the total cross-section. Subsequent manual inspections verified a high degree of agreement between the defect location and the predicted results, demonstrating the feasibility and accuracy of the system.

[0101] This invention, based on ultrasonic guided wave detection, integrates signal preprocessing, time-frequency analysis, and deep learning techniques to effectively solve problems such as low detection accuracy caused by the complex structure of steel strands and noise interference, and insufficient model generalization ability due to data imbalance. By adaptively adjusting the excitation frequency, waveform, and network model structure, reliable and efficient defect location and corrosion assessment can be achieved in diverse application environments, providing technical support for online monitoring of overhead ground wires and guy wires in power systems.

[0102] Example 4, referring to 4, is an embodiment of the present invention. This embodiment provides a deep learning-based ultrasonic guided wave signal and key feature extraction system, including a signal acquisition module, a feature extraction module, a deep learning network model module, and a defect identification module.

[0103] The signal acquisition module is used to acquire the guided wave signal of the steel strand and to preprocess the guided wave signal.

[0104] The feature extraction module is used to convert the preprocessed guided wave signal into a time spectrum, extract frequency peaks, envelope energy, and group velocity features, and perform dimensionality reduction on the feature set to generate feature vectors.

[0105] The deep learning network model module is used to input feature vectors into a deep learning network model containing a three-layer convolutional neural network and a two-layer bidirectional long short-term memory network, extract convolutional features and temporal features of the signal, and update the model parameters based on training samples.

[0106] The defect identification module uses a trained deep learning network model to infer the guided wave signal of the steel strand, outputting a classification label of corrosion type and a regression prediction value of corrosion depth to determine the specific location and degree of corrosion of the defect.

[0107] This embodiment also provides an electronic device applicable to a deep learning-based ultrasonic guided wave signal and key feature extraction method, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the deep learning-based ultrasonic guided wave signal and key feature extraction method proposed in the above embodiment.

[0108] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements a deep learning-based method for extracting ultrasonic guided wave signals and key features as proposed in the above embodiments.

[0109] The storage medium proposed in this embodiment and the method for extracting ultrasonic guided wave signals and key features based on deep learning proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0110] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0111] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for extracting ultrasonic guided wave signals and key features based on deep learning, characterized in that: include, Acquire guided wave signals from steel strands and preprocess the guided wave signals; The preprocessed guided wave signal is converted into a time spectrum, and the frequency peak, envelope energy, and group velocity features are extracted. The feature set is then dimensionality reduced to generate a feature vector. Input the feature vector into a deep learning network model containing a three-layer convolutional neural network and a two-layer bidirectional long short-term memory network, extract the convolutional features and temporal features of the signal, and update the model parameters based on the training samples; The trained deep learning network model is used to infer the waveguide signal of the steel strand, and outputs the classification label of corrosion type and the regression prediction value of corrosion depth to determine the specific location and degree of corrosion of the defect.

2. The method for extracting ultrasonic guided wave signals and key features based on deep learning as described in claim 1, characterized in that: The guided wave signal includes an ultrasonic sensor with good coupling to the steel strand, and is acquired by exciting a high-frequency narrowband pulse signal through an external excitation source.

3. The method for extracting ultrasonic guided wave signals and key features based on deep learning as described in claim 2, characterized in that: The external excitation source includes the center frequency of the excitation signal of the external excitation source, which is set according to the diameter of the steel strand, the expected detection distance, and the sensitivity requirements.

4. The method for extracting ultrasonic guided wave signals and key features based on deep learning as described in claim 3, characterized in that: The preprocessing includes noise suppression in the guided wave signal and signal amplitude normalization.

5. The method for extracting ultrasonic guided wave signals and key features based on deep learning as described in claim 4, characterized in that: The process of converting the preprocessed guided wave signal into a time-frequency spectrum involves using short-time Fourier transform or continuous wavelet transform to map the time characteristics of the guided wave signal into a time-frequency domain distribution, extracting the frequency peak, envelope energy, and group velocity as key parameters reflecting defect characteristics, and performing dimensionality reduction processing on the set of key parameters to remove redundant information and retain the main discriminative features, thereby generating feature vectors that can be learned by deep learning network models.

6. The method for extracting ultrasonic guided wave signals and key features based on deep learning as described in claim 4, characterized in that: The deep learning network model includes a three-layer convolutional neural network for convolutional feature extraction and a two-layer bidirectional long short-term memory network for temporal feature extraction. The three-layer convolutional neural network is configured with 32, 64 and 128 convolutional kernels respectively, with a kernel size of 3×3. After each convolutional layer, a batch normalization layer, a ReLU activation function and a 2×2 max pooling layer are connected in sequence to extract the convolutional features of the guided wave signal, including frequency distribution and signal peak value. The first layer of the bidirectional long short-term memory network is configured with 128 units, and the second layer is configured with 64 units. Dropout layers are added between layers to reduce overfitting and are used to extract the temporal characteristics of the guided wave signal in the propagation direction and between the defect echo. The three-layer convolutional neural network and the two-layer bidirectional long short-term memory network form a series structure. The deep learning network model uses the error between the input feature vector and the labeled sample, and performs backpropagation using the cross-entropy loss function or the mean squared error loss function, and uses the Adam optimizer to update the parameters.

7. The method for extracting ultrasonic guided wave signals and key features based on deep learning as described in claim 4, characterized in that: The process of inferring the guided wave signal of the steel strand using a trained deep learning network model includes: The trained deep learning network model receives the feature vector of the waveguide signal to be detected and outputs a classification label for corrosion type and a regression prediction value for corrosion depth. The classification label is determined by the maximum value of the category probability calculated by the Softmax activation function, and the regression prediction value is output through a linear output layer. By combining the time information of the excitation time and the echo peak time, as well as the propagation speed of the guided wave, the physical location of the defect in the steel strand is determined. The corrosion location and depth identification results are used to determine whether the preset threshold is exceeded, and an evaluation report containing the defect location, corrosion type and corrosion degree is output.

8. A deep learning-based ultrasonic guided wave signal and key feature extraction system, employing the deep learning-based ultrasonic guided wave signal and key feature extraction method as described in any one of claims 1 to 7, characterized in that, include: Signal acquisition module, feature extraction module, deep learning network model module, and defect recognition module; The signal acquisition module is used to acquire guided wave signals from steel strands and to preprocess the guided wave signals. The feature extraction module is used to convert the preprocessed guided wave signal into a time spectrum, extract frequency peaks, envelope energy, and group velocity features, and perform dimensionality reduction processing on the feature set to generate feature vectors. The deep learning network model module is used to input feature vectors into a deep learning network model containing a three-layer convolutional neural network and a two-layer bidirectional long short-term memory network, extract convolutional features and temporal features of the signal, and update model parameters based on training samples. The defect identification module is used to infer the guided wave signal of the steel strand using a trained deep learning network model, outputting a classification label of corrosion type and a regression prediction value of corrosion depth to determine the specific location and degree of corrosion of the defect.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the deep learning-based ultrasonic guided wave signal and key feature extraction method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the deep learning-based ultrasonic guided wave signal and key feature extraction method according to any one of claims 1 to 7.

Images