A guided wave signal feature enhancement method based on learnable anti-noise activation

Abstract

This invention relates to the field of pattern recognition technology and discloses a guided wave signal feature enhancement method based on learnable noise-resistant activation to solve the problems of poor noise resistance and low accuracy in guided wave signal pattern recognition. The method first acquires ultrasonic guided wave scattering signals, extracts envelope features through Hilbert transform, and completes signal preprocessing. It then constructs a feature enhancement network containing four levels of depthwise separable convolutional modules and fully connected regression modules. Based on a hierarchical structure of depthwise convolution and pointwise convolution, guided wave features are extracted layer by layer. A learnable threshold linear unit noise-resistant activation function is designed to adaptively suppress noise and enhance anomalous features. A layered Dropout regularization mechanism is employed to adapt to network layer differences and suppress overfitting. The output of anomalous location coordinates completes the localization performance evaluation, achieving pattern recognition. This invention can accurately mine effective anomalous features under strong noise conditions, improve the accuracy of anomalous location coordinate localization, and has extremely strong robustness.

Description

Technical Field

[0001] This invention belongs to the field of pattern recognition technology and relates to a method for enhancing the features of guided wave signals based on learnable noise-resistant activation. Background Technology

[0002] With the development of pattern recognition and deep learning technologies, neural network-based temporal signal feature enhancement and pattern recognition technologies have become core technologies for industrial sensor signal processing and anomaly localization. Pipelines, as critical infrastructure for transporting oil, gas, and chemical fluids, are prone to corrosion and cracking during long-term service. The ultrasonic guided wave temporal signals collected by piezoelectric sensors in pipelines carry complete characteristic information of abnormal signals. However, the complex industrial environment inevitably subjects the collected guided wave signals to various industrial interferences such as electromagnetic noise, environmental vibration, and sudden pulses, causing abnormal features to be completely submerged by strong noise, resulting in extremely low signal-to-noise ratios. Therefore, effectively enhancing and accurately locating anomalies from high-dimensional, long-sequence guided wave signals with noise interference through pattern recognition technology, and achieving robust anomaly localization, is a core challenge that urgently needs to be addressed.

[0003] Current pattern recognition technologies for guided wave signals are mainly divided into two categories: physical model-driven and data-driven. Physical model-driven methods manually extract the time-domain and frequency-domain features (such as time-of-flight and correlation coefficients) of the guided wave signal using digital signal processing methods such as time-frequency analysis and wavelet transform. These features are then combined with traditional pattern recognition algorithms or geometric localization methods (such as delay accumulation and probabilistic imaging algorithms) to locate the target position. However, in noisy environments, these manually designed features are extremely sensitive to noise interference and completely lack adaptive feature enhancement capabilities.

[0004] Data-driven approaches, centered on deep learning models such as Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and Transformers, automatically perform feature learning and pattern recognition of guided wave signals through data-driven methods, representing the mainstream technology in this field. However, existing models still exhibit significant shortcomings in noisy scenarios. The activation functions commonly employ fixed thresholds and slopes, and the parameters of the Dropout regularization strategy remain constant, lacking a noise-adaptive feature enhancement mechanism. Furthermore, in an effort to improve noise resistance, the network depth and width are often blindly increased. This not only fails to enhance features at abnormal locations but also results in severely insufficient generalization ability of the model in unknown, noisy scenarios. Summary of the Invention

[0005] This invention provides a guided wave signal feature enhancement method based on learnable noise-resistant activation to solve the problems of poor noise resistance and low accuracy in guided wave signal pattern recognition in the prior art.

[0006] This invention provides the following technical solution: a method for enhancing the features of guided waves based on learnable noise-resistant activation, comprising the following steps: S1. Set up an experimental platform, delineate the target area, obtain the ultrasonic guided wave response signals under normal and abnormal conditions, and record the corresponding spatial coordinates under normal and abnormal conditions. S2. Perform differential operation on the ultrasonic guided wave response signal to extract the scattered signal, obtain the Hilbert envelope, and then perform downsampling and fusion with multipath signals to complete the initial feature enhancement preprocessing. S3. Construct a deep separable convolutional neural network with an integrated learnable noise-resistant activation function. The deep separable convolutional neural network includes a learnable threshold linear unit noise-resistant activation function, a hierarchical Dropout regularization mechanism, a deep separable convolutional module, and a fully connected regression module. The network is trained end-to-end using the data obtained after preprocessing in S2. S4. Based on the trained, learnable noise-resistant activation function, a deep separable convolutional neural network performs feature enhancement processing on the guided wave signal, outputs the coordinates of the abnormal position, and completes the localization performance evaluation.

[0007] S1 specifically includes: The experimental platform consists of the structural component under test, a piezoelectric sensor array, an ultrasonic waveguide, and a host computer system. The piezoelectric sensor uses a lead zirconate titanate piezoelectric ceramic wafer. All sensors are evenly spaced along the circumferential direction, and the axial interval between two rows of sensors is defined as the target area. The ultrasonic guided wave instrument outputs an excitation signal, and a five-cycle sinusoidal pulse modulated by a Hanning window is selected as the excitation signal. The piezoelectric sensor is sequentially excited using a polling excitation method, while the other sensors receive the guided wave signal, thus acquiring the signal under normal conditions. ; The target area was divided into several equally sized grids. Cylindrical damping putty was used to artificially simulate an abnormal state, and the spatial coordinates under the abnormal state were recorded. Using the same excitation parameters and acquisition procedures as the normal state, the signals under the abnormal state were acquired. ; The abnormal signal is subjected to noise simulation processing. Let the original abnormal signal be... The signal after adding noise is The method of adding noise is uniformly represented as follows: ; In the formula, This represents the maximum amplitude of the original abnormal signal. These are noise level control parameters. It is a random variable that follows a specific distribution; Gaussian noise, Laplace noise, and pink noise are superimposed onto the original anomalous signal, respectively. The random variable corresponding to the Gaussian noise satisfies... That is, it follows a standard normal distribution. The random variable corresponding to Laplace noise follows a Laplace distribution. Pink noise is obtained by spectral shaping of Gaussian white noise.

[0008] S2 includes S2.1, which involves subtracting the acquired abnormal signal from the normal baseline signal point by point to obtain the scattered signal; Regarding the first One propagation path, scattered signal for: ; In the formula, For the first Scattered signals corresponding to each propagation path For the first The abnormal state guided wave response signal corresponding to each propagation path For the first The normal baseline guided wave response signal corresponding to each propagation path.

[0009] S2 includes S2.2, which is the scattering signal for each path. First, a Hilbert transform is performed to construct an analytic signal, and then the magnitude of the analytic signal is taken as the envelope signal. : ; In the formula, For the first The envelope signal of the path-scattered signal. For the first Scattered signals corresponding to each propagation path For the first Scattered signal corresponding to each propagation path The result of the Hilbert transform.

[0010] S2 includes S2.3, which uses an equal-interval sampling method to perform dimensionality reduction processing on the envelope signal, retaining one in every 100 consecutive sampling points. After the downsampling is completed, the compressed envelope signals of all paths are concatenated end to end in ascending order of their numbers to form a one-dimensional feature vector.

[0011] S3 includes S3.1, which dynamically divides the input signal into a low-amplitude region and a high-amplitude region through a learnable amplitude threshold, and assigns independent learnable linear gains to each region. The expression for the learnable threshold linear unit noise-resistant activation function is: ; In the formula, The output feature signal of the learnable threshold linear unit. The input feature signal of the activation function, This refers to the learnable slope parameter corresponding to the low amplitude range. This refers to the learnable slope parameter corresponding to the high amplitude range. For learnable positive threshold and ,in These are the original trainable parameters. A symbolic function, used to guarantee that the function... Continuous; During network forward propagation, the Softplus function is used. Will Mapped to positive values This ensures that the learnable positive threshold is always greater than 0, completing the forward mapping calculation of the threshold parameters; during backpropagation, the gradient is backpropagated to the original trainable parameters via the Softplus function. The Adam optimizer iteratively updates all parameters; each feature channel is configured with an independent set of trainable parameters. The number of trainable parameters is three times the number of feature channels; The learnable threshold linear unit noise-resistant activation function is located after each convolutional layer of the depthwise separable convolutional module and deployed before the pooling operation and the hierarchical Dropout regularization mechanism. The number of learnable threshold linear unit parameters of each depthwise separable convolutional module matches the number of output channels of the depthwise separable convolutional module.

[0012] S3 includes S3.2, whereby the hierarchical Dropout regularization mechanism is used to adapt to the features of different network layers; the hierarchical Dropout regularization mechanism configures the Dropout rate differently according to the network layer depth, with the Dropout rate of the convolutional layer increasing linearly with the network depth and the Dropout rate of the fully connected layer decreasing linearly with the network depth. For convolutional layers, the total number of convolutional layers is set to... , No. The formula for calculating the basic Dropout rate of a convolutional layer is: ; In the formula, For the first The base Dropout rate of each convolutional layer This represents the initial Dropout rate of the convolutional layer. The dropout rate at which the convolutional layer terminates. This represents the total number of depthwise separable convolutional modules in the network. For fully connected layers, the total number of fully connected layers is set to [number]. , No. The base Dropout rate of each fully connected layer The calculation formula is: ; In the formula, For the first The base Dropout rate of a fully connected layer This is the initial Dropout rate for the fully connected layer. The Dropout rate is the termination factor for fully connected layers. The specific layered Dropout regularization mechanism involves setting Dropout layers after the four-level depthwise separable convolutional modules to achieve progressively increasing regularization strength at each convolutional level, ensuring that shallow convolutions retain the original signal features and deep convolutions enhance noise overfit suppression; and setting Dropout layers after the three fully connected layers to achieve progressively decreasing regularization strength at each fully connected level.

[0013] S3 includes S3.3, wherein the depthwise separable convolutional neural network is equipped with a depthwise separable convolutional noise-resistant feature extractor, and the depthwise separable convolutional noise-resistant feature extractor completes feature extraction through a series of operations of depthwise convolution and pointwise convolution. Set the input guided wave characteristic sequence as , Where L is the number of input channels and L is the sequence length, the output of a standard convolution is calculated as follows: ; In the formula, For standard one-dimensional convolution at time 1 , No. Output characteristic values ​​of each output channel This is the kernel position index. For convolution kernel weights, The kernel size is [size]. For bias, For the input channel index of the convolution, This refers to the output channel index of the convolution. For standard convolution The bias term for each output channel; The total number of parameters in a standard convolution is ; The depthwise convolution operation process involves matching the number of convolution kernel groups to the number of input channels, and performing a convolution operation individually for each input channel. The formula for calculating depthwise convolution is: ; In the formula, For the first Single-channel features extracted from each input channel via depthwise convolution. Here are the kernel weights for the corresponding channels of the depthwise convolution. The total number of parameters for the depthwise convolution is... ; The pointwise convolution uses a 1×1 convolution kernel and performs a linear combination operation on the output features of the depthwise convolution along the channel dimension. The calculation formula for the pointwise convolution output is as follows:

[0014] ; In the formula, For pointwise convolution at time... , No. The fused feature values ​​output by each output channel. For the trainable weights of pointwise convolution, For pointwise convolution, the first The bias term for each output channel; The total number of parameters for pointwise convolution is The total number of parameters for depthwise separable convolution is ; The depthwise separable convolutional noise-resistant feature extractor includes four levels of depthwise separable convolutional modules. Each level of convolutional module performs operations in a fixed order: depthwise separable convolution operation, batch normalization, learnable threshold linear unit activation, max pooling downsampling, and Dropout regularization. The feature map output by the four-level deep separable convolutional module is flattened and converted into a one-dimensional feature vector, which is then input into the fully connected regression module. The fully connected regression module contains three fully connected layers, and each fully connected layer performs operations in a fixed order: linear transformation, batch normalization, learnable threshold linear unit activation, and Dropout regularization. In a fully connected layer, the number of bias parameters is equal to the length of the one-dimensional feature vector output by the fully connected layer. The output layer of the fully connected regression module is set to a linear layer, without a learnable threshold linear unit noise-resistant activation function. The predicted values ​​of the outlier coordinates are directly output through the output layer. .

[0015] S3 includes S3.4, which performs preprocessing on the acquired raw guided wave abnormality signal to obtain a preprocessed set of valid samples. First, an independent test set is separated from the set of samples. The samples in the independent test set are sealed throughout the process and do not participate in any training. All the remaining valid samples after removing the test set are randomly divided in a 7:3 ratio to construct a model training set and a model validation set respectively. The mean squared error loss function is used as the basis for parameter optimization supervision. for: ; In the formula, The model predicts the first The outlier coordinates of each sample and , For the first The true coordinates corresponding to each sample and , The total number of samples; The Adam optimizer is used to iteratively update the trainable parameters of the network. A learning rate decay strategy is configured during training, and an early stopping mechanism is introduced. By analyzing the training loss curve, validation loss curve, and coefficient of determination... Quantitatively evaluate the regression fit of the model; By setting up repeated training experiments to eliminate random interference caused by random initialization of network parameters and random partitioning of the dataset, the average of multiple independent training results is taken to obtain the trained ensemble anomaly localization model.

[0016] S4 specifically includes: The independent test set samples from S3.4 are input into the trained deep separable convolutional neural network, and the anomaly localization model outputs the predicted coordinates of the corresponding anomaly location. The Euclidean distance between the predicted anomaly and the actual damage anomaly, converted from cylindrical coordinates to rectangular coordinates, is used as the positioning error. The formula for calculating positioning error is: ; In the formula, For the radius of the structural component, Predict the axial components of cylindrical coordinates for outliers. For the axial component of the true cylindrical coordinates of the outlier point, Predict the angular components of the cylindrical coordinates for outliers. The angle components of the true cylindrical coordinates of the outlier point; For the test set Based on the single-sample positioning error, the mean absolute error, standard deviation, maximum positioning error, and minimum positioning error of the abnormal test samples are statistically calculated to complete the quantitative statistics of the model's positioning accuracy and stability. Gaussian noise, pink noise, and Laplace noise are superimposed onto the test set samples respectively, and the abnormal coordinate prediction and positioning error calculation process is repeatedly executed to achieve pattern recognition and positioning of abnormal locations.

[0017] Compared with the prior art, the present invention has the following beneficial effects: This invention employs an original learnable thresholded linear unit (LTLU) feature enhancement core mechanism. By using a learnable threshold, the input signal is dynamically divided into low-amplitude noise regions and high-amplitude feature regions, each assigned an independent, learnable slope parameter. This allows the network to adaptively suppress noise while directionally enhancing subtle anomalous features. Simultaneously, the channel decoupling design of depthwise separable convolution enables independent feature enhancement for each channel, avoiding interference from cross-channel noise contamination. Combining these multi-level feature enhancement innovations, this invention achieves significantly superior localization accuracy compared to existing methods in noisy environments. On an independent test set, the maximum localization error of this invention's model is only 20.82 mm, the minimum error is as low as 0.15 mm, the mean absolute error is 6.84 mm, the standard deviation is 3.61 mm, and 99.3% of the samples have errors below 20 mm. In comparison, the existing DCAS-Transformer model has a mean absolute error of 10.17 mm, a maximum error of 34.18 mm, and only 88.5% of the samples have errors below 20 mm, demonstrating significant anomalous feature enhancement and a substantial improvement in localization accuracy under strong noise conditions.

[0018] This invention abandons the inefficient path of blindly increasing network size to improve noise resistance in existing methods. Instead, it uses a lightweight, depthwise separable convolutional feature enhancement module to replace the standard one-dimensional convolution. The convolution operation is decomposed into two serial feature enhancement steps: depthwise convolution and pointwise convolution. While ensuring that the feature extraction effect is not reduced, the number of model parameters and the amount of computation are greatly reduced. It can be easily deployed in embedded edge computing devices in industrial sites to meet the positioning requirements based on real-time feature enhancement.

[0019] This invention constructs a training set covering typical noise scenarios in actual industrial environments by adding various types and intensities of noise to the signal during the training phase. During training, a hierarchical Dropout strategy, coordinated with the feature enhancement process, adjusts the regularization intensity based on network depth, effectively avoiding noise overfitting and underfitting of anomalous features, thus ensuring the generalization ability of the feature enhancement mechanism. With a noise level control parameter set to 0.2, under three different types of noise interference—Gaussian noise, pink noise, and Laplace noise—the errors of the model in this invention were less than 20 mm in 98.7%, 99.3%, and 98.3% of samples, respectively, consistently exceeding 98%, significantly better than the 93%–97% level of existing models. This demonstrates the strong robustness of the feature enhancement mechanism and its good adaptation to unknown noise types and intensities, indicating that the adaptive feature enhancement mechanism of this invention has extremely strong robustness in high-noise environments. Attached Figure Description

[0020] Figure 1 This is a flowchart of the technology of the present invention; Figure 2 This is a schematic diagram of the noise-resistant activation function curve for the Learnable Threshold Linear Unit (LTLU). Figure 3 This is a schematic diagram of the model's anomaly location prediction results based on test set samples. Detailed Implementation

[0021] The present invention will be further illustrated below with reference to embodiments. These embodiments are for illustrative purposes only and are not intended to limit the invention in any way. It should be understood that the described embodiments are merely some, not all, of the embodiments described in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0022] like Figure 1 As shown, Figure 1 This is a flowchart of the technical process of the present invention. A method for enhancing guided wave signal features based on learnable noise-resistant activation is described in this embodiment. An experimental platform for ultrasonic guided wave pipeline pattern recognition is constructed. The platform mainly consists of a stainless steel pipeline under test, a piezoelectric sensor array, an ultrasonic guided wave instrument, and a host computer system. The stainless steel pipeline under test has an outer diameter of 204 mm and a wall thickness of 2 mm, meeting the operating requirements of conventional industrial pipelines. The sensor array uses 16 lead zirconate titanate (PZT) piezoelectric ceramic wafers. The sensor array consists of multiple sensors, which are evenly arranged at 45° intervals along the circumference of the pipeline's outer wall, divided into two columns. The axial section between the two columns of sensors is defined as the abnormal region. Each piezoelectric ceramic wafer is firmly attached to the outer wall of the pipeline using epoxy resin adhesive, ensuring stable excitation and reception of the guided wave signal. The excitation signal is output from an ultrasonic guided wave instrument, using a five-cycle sinusoidal pulse modulated by a Hanning window as the excitation signal. The signal center frequency is 150kHz, the excitation voltage amplitude is ±60V, and the data acquisition rate is set to 10MHz. A polling excitation method is used to sequentially excite 16 PZT sensors. During a single sensor excitation, the other 15 sensors synchronously receive the helical guided wave signal propagating from the pipe. A single complete poll can acquire waveform data from 120 independent propagation paths, and a single time-domain signal contains 5500 sampling points. First, the normal guided wave signal of the pipeline is acquired, and the acquired baseline normal signal is denoted as... , Subsequently, the target area of ​​the pipeline was divided into 640 equal-sized grids, with each grid measuring 20mm axially and 25mm circumferentially. Cylindrical damping putty with a diameter of 20mm was used to simulate abnormal states. Simulated abnormal locations were sequentially placed at each grid position, and the spatial cylindrical coordinates corresponding to each abnormal state were recorded synchronously. Under completely consistent excitation parameters and acquisition procedures, the guided wave response signal under the abnormal state was acquired and recorded as the abnormal signal. , ; To simulate complex interference environments in actual engineering projects and verify the model's noise resistance and generalization capabilities, multi-type noise simulation experiments were conducted on the original anomalous signal. Let the original anomalous signal be... The signal after adding noise is The method of adding noise is uniformly represented as follows: ; In the formula, This represents the maximum amplitude of the original anomalous signal, used to adaptively correlate noise intensity with the signal's own intensity. These are noise level control parameters. It is a random variable that follows a specific distribution; This embodiment adds three typical types of noise: Gaussian noise, corresponding to... Simulate sensor electronic noise and circuit thermal noise; Laplace noise follows probability density. The Laplace distribution is used to simulate sudden pulse-type interference in the field; pink noise is passed through Gaussian white noise. The spectrum is shaped to simulate natural low-frequency disturbances such as wind vibration and fluid disturbances in the environment.

[0023] This embodiment performs standardized preprocessing on the collected normal and abnormal signals to achieve preliminary enhancement of abnormal features and removal of redundant information. The specific steps are as follows: The abnormal scattering signal is extracted by subtracting the collected abnormal signal from the normal baseline signal point by point to obtain the scattering signal. The scattering signal can clearly reflect the impact of the structural anomaly on the propagation of guided waves, that is, the waveform difference between the normal signal and the abnormal signal. At the same time, it can also effectively remove unrelated interferences such as normal direct waves and boundary reflections. Regarding the first One propagation path, scattered signal for: ; In the formula, For the first Scattered signals corresponding to each propagation path For the first Guided wave response signal of the pipeline under abnormal state corresponding to each propagation path For the first The guided wave response signal of the pipeline at the normal baseline corresponding to each propagation path; Extract the envelope signal and the scattered signal for each path. First, a Hilbert transform is performed to construct an analytic signal, and then the magnitude of the analytic signal is taken as the envelope signal. : ; In the formula, For the first The envelope signal of the path-scattered signal. For the first Scattered signals corresponding to each propagation path For the first Scattered signal corresponding to each propagation path The Hilbert transform result; Downsampling and multipath signal fusion are performed by using an equal-interval downsampling method. One effective sampling point is extracted from every 100 consecutive sampling points of the smooth envelope signal to achieve signal dimensionality reduction while preserving the core abnormal features. After downsampling, the compressed envelope signals of 120 paths are concatenated end to end according to the path number from smallest to largest to obtain a one-dimensional feature vector, which is used as the standard input sample data for the subsequent neural network.

[0024] This embodiment constructs a deep separable convolutional neural network that integrates a learnable noise-resistant activation function. The core structure of the network includes a learnable threshold linear unit noise-resistant activation function, a hierarchical Dropout regularization mechanism, a deep separable convolutional module, and a fully connected regression layer.

[0025] The LTLU noise-resistant activation function used in this embodiment can achieve channel-level adaptive feature enhancement and noise suppression. The function expression is as follows: ; In the formula, The output feature signal of the learnable threshold linear unit. The input feature signal of the activation function, This refers to the learnable slope parameter corresponding to the low amplitude range. This refers to the learnable slope parameter corresponding to the high amplitude range. For learnable positive threshold and ,in These are the original trainable parameters. A symbolic function, used to guarantee that the function... Continuous; During forward propagation in the network, the softplus function is used. Will Mapped to positive values This ensures that the learnable positive threshold is always greater than 0; during backpropagation, the gradient is propagated back to the softplus function. Parameters are updated using optimizers such as Adam. and They are also defined as trainable parameters, which do not require additional positive constraints and directly participate in gradient descent optimization. A symbolic function, used to guarantee that the function... Continuous; Each feature channel has its own set The number of new trainable parameters is only 3 × the number of channels, resulting in extremely low computational overhead. During training, the network automatically adjusts the threshold and slope of each channel based on the loss function of coordinate regression under abnormal conditions, suppressing low-amplitude noise components and enhancing high-amplitude abnormal components, thereby achieving adaptive feature-enhanced noise-resistant activation. The learnable thresholded linear unit (LTLU) noise-resistant activation function is located after each convolutional layer in the depthwise separable convolutional module and before pooling operations and hierarchical Dropout regularization. Specifically, in the first to fourth levels of the depthwise separable convolutional module, after each level completes the convolution operation, batch normalization and LTLU activation are performed sequentially. The number of learnable thresholded linear unit parameters in each depthwise separable convolutional module matches the number of output channels. After the fully connected layer completes the linear transformation, batch normalization and LTLU activation are performed sequentially. In the fully connected layer, the number of bias parameters is equal to the length of the one-dimensional feature vector output by the fully connected layer. The curve shape of the LTLU activation function is as follows: Figure 2 As shown, Figure 2 Learnable threshold Low amplitude slope and slope of high amplitude area Through this design, the network automatically suppresses low-amplitude noise and enhances high-amplitude anomaly features during training, significantly improving localization performance under abnormal conditions in noisy environments.

[0026] The hierarchical Dropout regularization mechanism is used to adapt to the features of different network layers, suppress model overfitting in noisy environments, and improve the model's generalization ability and anomaly localization stability. The hierarchical Dropout regularization mechanism configures the Dropout rate differently according to the network layer depth. The Dropout rate of the convolutional layer increases linearly with the network depth, while the Dropout rate of the fully connected layer decreases linearly with the network depth. For convolutional layers, the total number of convolutional layers is set to... , No. The formula for calculating the basic Dropout rate of a convolutional layer is: ; In the formula, For the first The base Dropout rate of each convolutional layer This represents the initial Dropout rate of the convolutional layer. The dropout rate at which the convolutional layer terminates. This represents the total number of depthwise separable convolutional modules in the network. For fully connected layers, the total number of fully connected layers is set to [number]. , No. The base Dropout rate of each fully connected layer The calculation formula is: ; In the formula, For the first The base Dropout rate of a fully connected layer, This is the initial Dropout rate for the fully connected layer. The Dropout rate is the termination factor for fully connected layers. The specific layered Dropout regularization mechanism involves setting Dropout layers after each of the four levels of deep separable convolutional modules, thereby increasing the regularization strength at each convolutional level. This allows shallow convolutions to retain the original signal features while deep convolutions enhance noise overfit suppression. Additionally, Dropout layers are set after each of the three fully connected layers, resulting in a gradual decrease in the regularization strength at each fully connected layer. This ensures a smooth transition of high-dimensional mixed features to the coordinate regression output of anomalies, guaranteeing model prediction stability and improving anomaly localization accuracy and generalization ability in noisy environments.

[0027] The deep separable convolutional neural network is equipped with a deep separable convolutional noise-resistant feature extractor, which extracts features sequentially through depthwise convolution and pointwise convolution. Set the input guided wave characteristic sequence as , Input the number of channels. Given the sequence length, the standard convolution output is calculated as follows: ; In the formula, For standard one-dimensional convolution at time 1 , No. Output characteristic values ​​of each output channel This is the kernel position index. For convolution kernel weights, The kernel size is [size]. For bias, For the input channel index of the convolution, This refers to the output channel index of the convolution. For standard convolution The bias term for each output channel; The total number of parameters in a standard convolution is ; The depthwise convolution operation process involves matching the number of convolution kernel groups to the number of input channels, and performing a convolution operation individually for each input channel. The formula for calculating depthwise convolution is: ; In the formula, For the first Single-channel features extracted from each input channel via depthwise convolution. Here are the kernel weights for the corresponding channels of the depthwise convolution; the total number of parameters for the depthwise convolution is... ; The pointwise convolution uses a 1×1 convolution kernel to perform linear combination operations on the output features of the depthwise convolution along the channel dimension, realizing cross-channel fusion and secondary enhancement of the enhanced abnormal features. The calculation formula for the pointwise convolution output is as follows: ; In the formula, For pointwise convolution at time... , No. The fused feature values ​​output by each output channel. For the trainable weights of pointwise convolution, For pointwise convolution, the first The bias term for each output channel; The total number of parameters for pointwise convolution is The total number of parameters for depthwise separable convolution is ; when and When all parameters are greater than 1, the number of parameters in a depthwise separable convolution is approximately that of a standard convolution. Overall, the actual computational cost of this network is reduced to approximately [a fraction of] that of a standard convolution. ; The depthwise separable convolutional noise-resistant feature extractor includes four levels of depthwise separable convolutional modules. Each level of the depthwise separable convolutional module performs operations in a fixed order: depthwise separable convolution, batch normalization, learnable thresholded linear unit activation, max pooling downsampling, and Dropout regularization. The feature map output by the four-level deep separable convolutional module is flattened and converted into a one-dimensional feature vector, which is then input into the fully connected regression module. The fully connected regression module contains three fully connected layers, and each fully connected layer performs operations in a fixed order: linear transformation, batch normalization, learnable threshold linear unit activation, and Dropout regularization. The output layer of the fully connected regression module is set to a linear layer without an activation function, and the coordinate prediction values ​​under pipeline anomalies are directly output through the output layer. .

[0028] The original guided wave anomaly signals were preprocessed to obtain a set of effective samples. First, an independent test set was separated from the set of samples. The samples in the independent test set were sealed and not used for any training. All the remaining effective samples after removing the test set were randomly divided in a 7:3 ratio to construct the model training set and the model validation set respectively. The mean squared error loss function is used as the basis for parameter optimization supervision. for: ; In the formula, The model predicts the first The outlier coordinates of each sample and , For the first The true coordinates corresponding to each sample and , The total number of samples; The Adam optimizer is used to iteratively update the trainable parameters of the network. A learning rate decay strategy is configured during training, and an early stopping mechanism is introduced. By analyzing the training loss curve, validation loss curve, and coefficient of determination... Quantitatively evaluate the regression fit of the model; This paper conducts five independent experiments to eliminate random interference caused by random initialization of network parameters and random partitioning of the dataset. The average of the five independent training results is taken to obtain the trained integrated anomaly localization model.

[0029] The deep separable convolutional neural network is trained by inputting samples from a separate independent test set, and the model outputs the predicted coordinates of the corresponding anomaly locations. The Euclidean distance between the predicted anomaly and the actual anomaly points (after converting their cylindrical coordinates to rectangular coordinates) is used as the localization error. Mean Absolute Error (MAE), Standard Deviation (STD), Maximum Error, and Minimum Error are used to evaluate positioning accuracy and stability. The positioning error calculation formula is as follows: ; In the formula, For the pipe radius, Predict the axial components of cylindrical coordinates for outliers. For the axial component of the true cylindrical coordinates of the outlier point, Predict the angular components of the cylindrical coordinates for outliers. The angle components of the true cylindrical coordinates of the outlier point; For the test set Based on the single-sample positioning error, the mean absolute error, standard deviation, maximum positioning error, and minimum positioning error of the abnormal test samples are statistically calculated to complete the quantitative statistics of the model's positioning accuracy and stability. Gaussian noise, pink noise, and Laplace noise are superimposed on the test set samples to construct multiple types of noise test samples. Each type of noise test sample is then input into the trained model, and the abnormal coordinate prediction and positioning error calculation process is repeatedly executed to achieve pattern recognition and positioning of abnormal locations.

[0030] Figure 2 The curve characteristics of the learnable threshold linear unit (LTLU) noise-resistant activation function designed in this invention are intuitively demonstrated. This activation function abandons the fixed threshold design of traditional activation functions and achieves adaptive suppression of low-amplitude noise signals and directional enhancement of high-amplitude abnormal features by using a learnable threshold and a dual-interval trainable slope parameter. It solves the technical problem of confusion between abnormal features and noise features and loss of effective features in complex noise environments from the feature activation level, and provides core support for the model to accurately extract abnormal features. Figure 3 The results of the trained deep separable convolutional neural network predicting the location of pipeline anomalies are presented intuitively. It can be clearly seen that the coordinates of the anomalies predicted by the model are highly consistent with the actual locations of the anomalies. By combining a lightweight deep separable convolutional structure and a hierarchical Dropout regularization mechanism, this invention effectively improves the stability of pattern recognition of ultrasonic guided wave signal anomalies while significantly reducing the number of model parameters and computational overhead and avoiding model overfitting. This proves that the method can adapt to the noise interference scenario of pipelines in complex service, and can efficiently and accurately locate pipeline anomalies, which is significantly better than traditional methods.

[0031] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for enhancing the features of guided waves based on learnable noise-resistant activation, characterized in that, Includes the following steps: S1. Set up an experimental platform, delineate the target area, obtain the ultrasonic guided wave response signals under normal and abnormal conditions, and record the corresponding spatial coordinates under normal and abnormal conditions. The experimental platform consists of the structural component under test, a piezoelectric sensor array, an ultrasonic waveguide, and a host computer system. S2. Perform differential operation on the ultrasonic guided wave response signal to extract the scattered signal, obtain the Hilbert envelope, and then perform downsampling and fusion with multipath signals to complete the initial feature enhancement preprocessing. S3. Construct a deep separable convolutional neural network with an integrated learnable noise-resistant activation function. The deep separable convolutional neural network includes a learnable threshold linear unit noise-resistant activation function, a hierarchical Dropout regularization mechanism, a deep separable convolutional module, and a fully connected regression module. The network is trained end-to-end using the data obtained after preprocessing in S2. S3 includes S3.1, which dynamically divides the input signal into a low-amplitude region and a high-amplitude region through a learnable amplitude threshold, and assigns independent learnable linear gains to each region. The expression for the learnable threshold linear unit noise-resistant activation function is: ； In the formula, The output feature signal of the learnable threshold linear unit. The input feature signal of the activation function, This refers to the learnable slope parameter corresponding to the low amplitude range. This refers to the learnable slope parameter corresponding to the high amplitude range. For learnable positive threshold and ,in These are the original trainable parameters. A symbolic function, used to guarantee that the function... Continuous; During network forward propagation, the Softplus function is used. Will Mapped to positive values This ensures that the learnable positive threshold is always greater than 0, completing the forward mapping calculation of the threshold parameters; during backpropagation, the gradient is backpropagated to the original trainable parameters via the Softplus function. And the Adam optimizer completes the iterative update of all parameters; Each feature channel is configured with an independent set of trainable parameters. The number of trainable parameters is three times the number of feature channels; The learnable threshold linear unit noise-resistant activation function is located after each convolutional layer of the depthwise separable convolutional module and deployed before the pooling operation and the hierarchical Dropout regularization mechanism. The number of learnable threshold linear unit parameters of each depthwise separable convolutional module matches the number of output channels of the depthwise separable convolutional module. S4. Based on the trained, learnable noise-resistant activation function, a deep separable convolutional neural network performs feature enhancement processing on the guided wave signal, outputs the coordinates of the abnormal position, and completes the localization performance evaluation.

2. The guided wave signal feature enhancement method based on learnable noise-resistant activation according to claim 1, characterized in that, The piezoelectric sensor uses a lead zirconate titanate piezoelectric ceramic wafer. All sensors are evenly spaced along the circumferential direction, and the axial interval between two rows of sensors is defined as the target area. The ultrasonic guided wave instrument outputs an excitation signal, and a five-cycle sinusoidal pulse modulated by a Hanning window is selected as the excitation signal. The piezoelectric sensor is sequentially excited using a polling excitation method, while the other sensors receive the guided wave signal, thus acquiring the signal under normal conditions. ; The target area was divided into several equally sized grids. Cylindrical damping putty was used to artificially simulate an abnormal state, and the spatial coordinates under the abnormal state were recorded. Using the same excitation parameters and acquisition procedures as the normal state, the signals under the abnormal state were acquired. ; The abnormal signal is subjected to noise simulation processing. Let the original abnormal signal be... The signal after adding noise is The method of adding noise is uniformly represented as follows: ； In the formula, This represents the maximum amplitude of the original abnormal signal. These are noise level control parameters. It is a random variable that follows a specific distribution; Gaussian noise, Laplace noise, and pink noise are superimposed onto the original anomalous signal, respectively. The random variable corresponding to the Gaussian noise satisfies... That is, it follows a standard normal distribution. The random variable corresponding to Laplace noise follows a Laplace distribution. Pink noise is obtained by spectral shaping of Gaussian white noise.

3. The guided wave signal feature enhancement method based on learnable noise-resistant activation according to claim 1, characterized in that, S2 includes S2.1, which involves subtracting the acquired abnormal signal from the normal baseline signal point by point to obtain the scattered signal; Regarding the first One propagation path, scattered signal for: ； In the formula, For the first Scattered signals corresponding to each propagation path For the first The abnormal state guided wave response signal corresponding to each propagation path For the first The normal baseline guided wave response signal corresponding to each propagation path.

4. The guided wave signal feature enhancement method based on learnable noise-resistant activation according to claim 1, characterized in that, S2 includes S2.2, which is the scattering signal for each path. First, a Hilbert transform is performed to construct an analytic signal, and then the magnitude of the analytic signal is taken as the envelope signal. : ； In the formula, For the first The envelope signal of the path-scattered signal. For the first Scattered signals corresponding to each propagation path For the first Scattered signal corresponding to each propagation path The result of the Hilbert transform.

5. The guided wave signal feature enhancement method based on learnable noise-resistant activation according to claim 1, characterized in that, S2 includes S2.3, which uses an equal-interval sampling method to perform dimensionality reduction processing on the envelope signal, retaining one in every 100 consecutive sampling points. After the downsampling is completed, the compressed envelope signals of all paths are concatenated end to end in ascending order of their numbers to form a one-dimensional feature vector.

6. The guided wave signal feature enhancement method based on learnable noise-resistant activation according to claim 1, characterized in that, S3 includes S3.2, whereby the hierarchical Dropout regularization mechanism is used to adapt to the features of different network layers; the hierarchical Dropout regularization mechanism configures the Dropout rate differently according to the network layer depth, with the Dropout rate of the convolutional layer increasing linearly with the network depth and the Dropout rate of the fully connected layer decreasing linearly with the network depth. For convolutional layers, the total number of convolutional layers is set to... , No. The formula for calculating the basic Dropout rate of a convolutional layer is: ； In the formula, For the first The base Dropout rate of each convolutional layer This represents the initial Dropout rate of the convolutional layer. The dropout rate at which the convolutional layer terminates. This represents the total number of depthwise separable convolutional modules in the network. For fully connected layers, the total number of fully connected layers is set to [number]. , No. The base Dropout rate of each fully connected layer The calculation formula is: ； In the formula, For the first The base Dropout rate of a fully connected layer This is the initial Dropout rate for the fully connected layer. The Dropout rate is the termination factor for fully connected layers. The specific layered Dropout regularization mechanism involves setting Dropout layers after the four-level depthwise separable convolutional modules to achieve progressively increasing regularization strength at each convolutional level, ensuring that shallow convolutions retain the original signal features and deep convolutions enhance noise overfit suppression; and setting Dropout layers after the three fully connected layers to achieve progressively decreasing regularization strength at each fully connected level.

7. The guided wave signal feature enhancement method based on learnable noise-resistant activation according to claim 1, characterized in that, S3 includes S3.3, wherein the depthwise separable convolutional neural network is equipped with a depthwise separable convolutional noise-resistant feature extractor, and the depthwise separable convolutional noise-resistant feature extractor completes feature extraction through a series of operations of depthwise convolution and pointwise convolution. Set the input guided wave characteristic sequence as , Where L is the number of input channels and L is the sequence length, the output of a standard convolution is calculated as follows: ； In the formula, For standard one-dimensional convolution at time 1 , No. Output characteristic values ​​of each output channel This is the kernel position index. For convolution kernel weights, The kernel size is [size]. For bias, For the input channel index of the convolution, This refers to the output channel index of the convolution. For standard convolution The bias term for each output channel; The total number of parameters in a standard convolution is ; The depthwise convolution operation process involves matching the number of convolution kernel groups to the number of input channels, and performing a convolution operation individually for each input channel. The formula for calculating depthwise convolution is: ； In the formula, For the first Single-channel features extracted from each input channel via depthwise convolution. Here are the kernel weights for the corresponding channels of the depthwise convolution. The total number of parameters for the depthwise convolution is... ; The pointwise convolution uses a 1×1 convolution kernel and performs a linear combination operation on the output features of the depthwise convolution along the channel dimension. The calculation formula for the pointwise convolution output is as follows: ； In the formula, For pointwise convolution at time... , No. The fused feature values ​​output by each output channel. For the trainable weights of pointwise convolution, For pointwise convolution, the first The bias term for each output channel; The total number of parameters for pointwise convolution is The total number of parameters for depthwise separable convolution is ; The depthwise separable convolutional noise-resistant feature extractor includes four levels of depthwise separable convolutional modules. Each level of convolutional module performs operations in a fixed order: depthwise separable convolution operation, batch normalization, learnable threshold linear unit activation, max pooling downsampling, and Dropout regularization. The feature map output by the four-level deep separable convolutional module is flattened and converted into a one-dimensional feature vector, which is then input into the fully connected regression module. The fully connected regression module contains three fully connected layers, and each fully connected layer performs operations in a fixed order: linear transformation, batch normalization, learnable threshold linear unit activation, and Dropout regularization. In a fully connected layer, the number of bias parameters is equal to the length of the one-dimensional feature vector output by the fully connected layer. The output layer of the fully connected regression module is set to a linear layer, without a learnable threshold linear unit noise-resistant activation function. The predicted values ​​of the outlier coordinates are directly output through the output layer. .

8. The guided wave signal feature enhancement method based on learnable noise-resistant activation according to claim 1, characterized in that, S3 includes S3.4, which performs preprocessing on the acquired raw guided wave abnormality signal to obtain a preprocessed set of valid samples. First, an independent test set is separated from the set of samples. The samples in the independent test set are sealed throughout the process and do not participate in any training. All the remaining valid samples after removing the test set are randomly divided in a 7:3 ratio to construct a model training set and a model validation set respectively. The mean squared error loss function is used as the basis for parameter optimization supervision. for: ； In the formula, The model predicts the first The outlier coordinates of each sample and , For the first The true coordinates corresponding to each sample and , The total number of samples; The Adam optimizer is used to iteratively update the trainable parameters of the network. A learning rate decay strategy is configured during training, and an early stopping mechanism is introduced. By analyzing the training loss curve, validation loss curve, and coefficient of determination... Quantitatively evaluate the regression fit of the model; By setting up repeated training experiments to eliminate random interference caused by random initialization of network parameters and random partitioning of the dataset, the average of multiple independent training results is taken to obtain the trained ensemble anomaly localization model.

9. The guided wave signal feature enhancement method based on learnable noise-resistant activation according to claim 1, characterized in that, S4 specifically includes: The independent test set samples from S3.4 are input into the trained deep separable convolutional neural network, and the anomaly localization model outputs the predicted coordinates of the corresponding anomaly location. The Euclidean distance between the predicted anomaly and the actual anomaly, converted from cylindrical coordinates to rectangular coordinates, is used as the localization error. The formula for calculating positioning error is: ； In the formula, For the radius of the structural component, Predict the axial components of cylindrical coordinates for outliers. For the axial component of the true cylindrical coordinates of the outlier point, Predict the angular components of the cylindrical coordinates for outliers. The angle components of the true cylindrical coordinates of the outlier point; For the test set Based on the single-sample positioning error, the mean absolute error, standard deviation, maximum positioning error, and minimum positioning error of the abnormal test samples are statistically calculated to complete the quantitative statistics of the model's positioning accuracy and stability. Gaussian noise, pink noise, and Laplace noise are superimposed onto the test set samples respectively, and the abnormal coordinate prediction and positioning error calculation process is repeatedly executed to achieve pattern recognition and positioning of abnormal locations.

Images