A Rail Damage Detection Method and System Based on Multi-Domain Information Collaborative Transfer Learning

Abstract

This invention discloses a rail damage detection method and system based on multi-domain information collaborative transfer learning. The method includes: developing a data augmentation algorithm for random damage based on the scattering simulation signal of a pre-constructed benchmark rail finite element model to obtain source domain simulated ultrasonic guided wave samples; pre-training a multi-branch convolutional network (CBN) with multi-domain information collaborative transfer learning using the source domain simulated ultrasonic guided wave samples; adjusting the pre-trained CBN using transfer learning based on a preset number of measured samples; and detecting rail ultrasonic guided wave signals obtained in engineering applications using the adjusted CBN to obtain rail damage detection results. This invention can enhance the diversity of source domain simulated data to complete high-precision model pre-training and improve the model's generalization ability in measured signals by transferring the data to measured data, thereby improving the classification ability of the convolutional network for rail damage locations. It can be widely applied in the field of non-destructive testing technology.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing technology, and in particular to a rail damage detection method and system based on multi-domain information collaborative transfer learning. Background Technology

[0002] High-speed railways, as a crucial component of modern public transportation, inevitably suffer rail damage due to complex and variable coupled excitations and harsh operating environments. Numerous non-destructive testing (NDT) technologies have been applied to rail damage detection. Methods such as eddy current testing, magnetic flux leakage testing, and alternating current field measurement perform well in detecting surface and near-surface damage, but are limited by testing speed and have significant limitations in identifying internal damage. Visual inspection methods, utilizing high-speed cameras or optical sensors, offer high detection efficiency but cannot identify invisible damage and are highly dependent on ambient lighting conditions. Ultrasonic bulk wave-based detection techniques, while capable of detecting internal damage, are less efficient, have limited ability to detect transverse cracks, and surface defects may reflect ultrasonic waves, hindering their penetration and leading to missed internal defects.

[0003] In contrast, ultrasonic guided wave detection technology, which propagates through specific modes guided by the rail boundary, offers numerous advantages, including long propagation distance, low attenuation, strong penetration, good directionality, wide coverage within the waveguide cross-section, and sensitivity to small defects. These characteristics enable ultrasonic guided waves to acoustically propagate across the entire rail, allowing for efficient simultaneous detection of internal and surface damage over long distances. However, due to the dispersion and multimodal characteristics of guided waves, their propagation modes are often extremely complex. Combined with the uncertainties in rail geometry, defect orientation, and measurement noise, this significantly complicates damage feature extraction, impacting detection efficiency and accuracy. Therefore, despite the numerous advantages of ultrasonic guided waves, relatively few rail damage detection methods have been developed based on them. As an end-to-end approach, deep learning excels at discovering potential damage features in complex data while minimizing the impact of measurement noise and human intervention. This allows for the automatic and accurate extraction of damage features from guided wave signals, significantly improving the accuracy and efficiency of rail damage detection. However, the development of high-precision deep learning models with good generalization ability often requires a large number of labeled samples, and obtaining sufficient high-quality labeled data is a significant challenge in practical applications. Summary of the Invention

[0004] This invention aims to at least partially address the limitations of related technologies. To this end, this invention proposes a rail damage detection method and system based on multi-domain information collaborative transfer learning, which can accurately detect rail damage.

[0005] On one hand, embodiments of the present invention provide a rail damage detection method based on multi-domain information collaborative transfer learning, comprising the following steps:

[0006] A data augmentation algorithm for random damage was developed based on the scattering simulation signal of a pre-constructed benchmark rail finite element model to obtain simulated ultrasonic guided wave samples in the source domain; the benchmark rail finite element model was corrected based on the residual optimization results between the measured guided wave and the model-predicted guided wave dispersion characteristics.

[0007] Pre-training of a multi-branch convolutional network that coordinates multi-domain information is achieved by simulating ultrasonic guided wave samples in the source domain.

[0008] Based on a preset number of measured samples, the pre-trained multi-branch convolutional network is adjusted through transfer learning; the adjusted multi-branch convolutional network is used to detect the ultrasonic guided wave signals of rails obtained in engineering applications, and the rail damage detection results are obtained.

[0009] Optionally, a data augmentation algorithm for random damage is developed based on the scattering simulation signal of a pre-constructed benchmark rail finite element model to obtain source domain simulated ultrasonic guided wave samples, including the following steps:

[0010] By optimizing the residual between the measured guided wave and the predicted guided wave of the rail, the corrected physical parameters of the model are obtained, and then a reference rail finite element model is constructed based on the corrected results.

[0011] Based on the finite element model of the benchmark rail, damage is randomly added within the damage label area, and ultrasonic guided wave scattering simulation is completed to obtain the benchmark reflected wave signal of the corresponding damage label area.

[0012] A data augmentation algorithm based on scattering simulation was developed using a reference reflected wave signal to perform random damage data augmentation processing, resulting in source domain simulated ultrasonic guided wave samples.

[0013] Optionally, by optimizing the residual between the measured guided wave and the predicted guided wave of the rail, the corrected physical parameters of the model are obtained, and then a reference rail finite element model is constructed based on the corrected results, including the following steps:

[0014] Based on the rail cross-sectional parameters, an initial finite element model is constructed, and the physical model parameters of the rail to be corrected are determined. The physical model parameters of the rail to be corrected include the elastic modulus, Poisson's ratio, and damping loss coefficient.

[0015] Based on the parameters of the physical model of the rail to be corrected, the mass and stiffness matrices of the periodic waveguide structure of the rail under the corresponding parameters are determined.

[0016] Based on the mass and stiffness matrix of the rail periodic waveguide structure, the predicted ultrasonic wave number of the rail under the corresponding physical model parameters of the rail to be corrected is obtained by wave-finite element method analysis and calculation.

[0017] The ultrasonic guided wave signal of the actual rail is obtained by the sensor, and then the spatiotemporal matrix of the ultrasonic guided wave signal of the actual rail is constructed.

[0018] Two-dimensional fast Fourier transform was performed on the spatiotemporal matrix of the ultrasonic guided wave signal to obtain the measured ultrasonic guided wave number of the actual rail.

[0019] Based on the predicted ultrasonic guided wavenumber and the measured ultrasonic guided wavenumber, the wavenumber prediction error is determined and the likelihood function is constructed.

[0020] The likelihood function is embedded into the Bayesian formula, and random sampling is performed using asymptotic Markov Monte Carlo to obtain the posterior distribution and optimal values ​​of the physical model parameters of the actual rail.

[0021] The initial finite element model is corrected based on the optimal values ​​of the physical model parameters to obtain the reference rail finite element model.

[0022] Optionally, a data augmentation algorithm based on scattering simulation is developed using a reference reflected wave signal to perform random damage data augmentation processing, obtaining source domain simulated ultrasonic guided wave samples, including the following steps:

[0023] A weighted fusion process is performed on a preset number of reference reflected wave signals set sequentially in different damage label areas to obtain the reflected wave signal of damage occurring at any position in the damage label area corresponding to the rail cross section.

[0024] The amplitude and phase of the reflected wave signal of damage appearing at any position within the damage label area corresponding to the rail cross section are adjusted to obtain the reflected wave signal of damage of any size at any longitudinal position within the damage label area corresponding to the rail.

[0025] By adding undamaged guided wave signals to the reflected wave signals of any longitudinal position with damage of any size within the corresponding damage label area of ​​the rail, the ultrasonic guided wave signals of random damage within the corresponding damage label area of ​​the rail are obtained, thus completing the acquisition of source domain simulation ultrasonic guided wave samples.

[0026] Optionally, a multi-branch convolutional network for multi-domain information collaboration is pre-trained using source domain simulated ultrasonic guided wave samples, including the following steps:

[0027] The source domain ultrasonic guided wave signals corresponding to the simulated source domain ultrasonic guided wave samples are preprocessed, and then multi-domain signal processing results are obtained through multi-domain signal feature extraction. The preprocessing includes bandpass filtering and normalization.

[0028] Construct a multi-branch convolutional network based on the processing results of different types of signal domains in the multi-domain signal processing results;

[0029] In this system, the number of signal domains in the multi-domain signal processing results is the same as and corresponds one-to-one with the number of branch structures in the multi-branch convolutional network. The multi-branch convolutional network with multi-domain information collaboration includes sequentially connected branch structures, feature fusion layers, fully connected layers, and classification layers, with multiple branch structures connected in parallel to the feature fusion layer. Each branch structure includes a convolutional module, a spatial attention mechanism, and a Dropout layer. The kernel scale of the convolutional module in each branch structure is determined by the data dimension of the processing results of the corresponding signal domain.

[0030] The results of multi-domain signal processing are divided into a first training set and a first validation set.

[0031] Based on the first training set, the multi-branch convolutional network is pre-trained using a self-adjusting moment estimation optimization algorithm;

[0032] The first validation set is used to validate the pre-trained multi-branch convolutional network.

[0033] Optionally, the multi-domain signal processing results include a one-dimensional envelope, a one-dimensional spectral curve, and a two-dimensional time-frequency plot; the multi-domain signal processing results are obtained through multi-domain signal feature extraction, including the following steps:

[0034] The preprocessed source domain ultrasonic guided wave signal was subjected to time-domain feature extraction using the Hilbert transform method to obtain the one-dimensional envelope of the source domain ultrasonic guided wave signal.

[0035] The source domain ultrasonic guided wave signal was subjected to frequency domain feature extraction processing by fast Fourier transform to obtain the one-dimensional spectrum curve of the source domain ultrasonic guided wave signal.

[0036] The time-frequency energy features of the preprocessed source domain ultrasonic guided wave signal are extracted by synchronous squeezing wavelet transform, resulting in a two-dimensional time-frequency diagram of the source domain ultrasonic guided wave signal.

[0037] Optionally, the pre-trained multi-branch convolutional network is adjusted through transfer learning based on a preset number of test samples, including the following steps:

[0038] Ultrasonic guided wave sensors were used to acquire measured guided wave signals of added damage at different locations on laboratory steel rails as experimental samples.

[0039] The measured samples are divided into a second training set and a second validation set based on a preset ratio.

[0040] Among them, the multi-branch convolutional network with multi-domain information collaboration includes multiple branch structures, a feature fusion layer, a fully connected layer and a classification layer connected in sequence, with multiple branch structures connected in parallel to the feature fusion layer;

[0041] The hyperparameters of all network layers in the pre-trained multi-branch convolutional network are frozen. Then, starting from the classification layer, the hyperparameters of each network layer are gradually released forward. The released hyperparameters of the network layers are adjusted multiple times with a preset learning rate using the measured data in the second training set to complete the transfer of the deep learning model for rail damage detection.

[0042] The second validation set is used to validate and evaluate the damage detection performance of the deep learning model for rail damage detection after the transfer is completed.

[0043] On the other hand, embodiments of the present invention provide a rail damage detection system based on multi-domain information collaborative transfer learning, comprising:

[0044] The first module is used to develop a data augmentation algorithm for random damage based on the scattering simulation signal of the pre-built benchmark rail finite element model to obtain source domain simulation ultrasonic guided wave samples; the benchmark rail finite element model is obtained by correcting the residual optimization results between the measured guided wave and the model predicted guided wave dispersion characteristics.

[0045] The second module is used to pre-train a multi-branch convolutional network that coordinates multi-domain information by simulating ultrasonic guided wave samples in the source domain.

[0046] The third module is used to adjust the pre-trained multi-branch convolutional network based on a preset number of measured samples through transfer learning; the adjusted multi-branch convolutional network is used to detect the ultrasonic guided wave signals of the rail obtained in engineering applications to obtain the rail damage detection results.

[0047] On the other hand, embodiments of the present invention provide an electronic device, including: a processor and a memory; the memory is used to store a program; the processor executes the program to implement the above-mentioned rail damage detection method based on multi-domain information collaborative transfer learning.

[0048] On the other hand, embodiments of the present invention provide a computer storage medium storing a processor-executable program, which, when executed by the processor, is used to implement the above-described rail damage detection method based on multi-domain information collaborative transfer learning.

[0049] This invention develops a data augmentation algorithm for random damage based on the scattering simulation signal of a pre-constructed benchmark rail finite element model to obtain source domain simulated ultrasonic guided wave samples. The benchmark rail finite element model is corrected based on the residual optimization results between the measured guided wave and the model-predicted guided wave dispersion characteristics. A multi-branch convolutional network for multi-domain information collaboration is pre-trained using the source domain simulated ultrasonic guided wave samples. The pre-trained multi-branch convolutional network is adjusted through transfer learning based on a preset number of measured samples. The adjusted multi-branch convolutional network is used to detect the rail ultrasonic guided wave signal obtained in engineering applications to obtain the rail damage detection result. This invention has the following beneficial effects:

[0050] 1. Enhancing Data Diversity and Accuracy: By developing a data augmentation algorithm for random damage based on the scattering simulation signal of a pre-built benchmark rail finite element model, a rich sample of simulated ultrasonic guided waves in the source domain can be obtained. These samples not only increase the diversity of the data, but also ensure the accuracy of the data because the benchmark model is obtained by correcting the residual optimization results between the measured guided waves and the predicted guided wave dispersion characteristics of the rail, thus laying a solid foundation for subsequent network training.

[0051] 2. Enhancing Model Generalization Ability: Pre-training a multi-branch convolutional network that coordinates multi-domain information using source-domain simulated ultrasonic guided wave samples allows the model to learn important features of rail ultrasonic guided wave signals in the early stages of training. This pre-training method helps improve the model's generalization ability in subsequent transfer learning, enabling it to better adapt to rail damage detection in different scenarios.

[0052] 3. Reduced reliance on real-world samples and lower detection costs: By using a pre-set number of real-world samples for transfer learning, the pre-trained multi-branch convolutional network can be further adjusted to better meet actual detection needs. This approach reduces reliance on a large number of real-world samples, lowers detection costs, and improves detection efficiency.

[0053] 4. Achieving efficient and accurate rail damage detection: The multi-branch convolutional network, after pre-training and transfer learning adjustments, can efficiently and accurately detect ultrasonic guided wave signals from rails acquired in engineering applications and derive rail damage detection results. This helps improve the automation level of rail safety monitoring and provides strong support for the safe operation of railway transportation.

[0054] In summary, the embodiments of the present invention effectively improve the accuracy and efficiency of rail damage detection and reduce detection costs through data augmentation, pre-training, and transfer learning, and have significant practical application value. Attached Figure Description

[0055] The accompanying drawings are provided to further understand the technical solutions of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the technical solutions of the present invention, and do not constitute a limitation on the technical solutions of the present invention.

[0056] Figure 1 This is a schematic diagram of an implementation environment for rail damage detection based on multi-domain information collaborative transfer learning, provided in an embodiment of the present invention.

[0057] Figure 2 This is a flowchart illustrating a rail damage detection method based on multi-domain information collaborative transfer learning provided in an embodiment of the present invention.

[0058] Figure 3 This is a schematic diagram illustrating the principle and logic of the rail damage detection method based on multi-domain information collaborative transfer learning provided in an embodiment of the present invention.

[0059] Figure 4 A schematic diagram of the structure of the finite element model of the rail provided in an embodiment of the present invention;

[0060] Figure 5 This is a schematic diagram illustrating an example of damage detection results of a deep learning model before source domain simulation sample migration, as provided in an embodiment of the present invention.

[0061] Figure 6 This is a schematic diagram illustrating an example of damage detection results of a deep learning model after source domain simulation sample migration, as provided in an embodiment of the present invention.

[0062] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention;

[0063] Figure 8 This is a schematic diagram of the structure of a rail damage detection system based on multi-domain information collaborative transfer learning, provided in an embodiment of the present invention.

[0064] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0066] It should be noted that although functional modules are divided in the system diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the system or the order in the flowchart. The terms "first / S100," "second / S200," etc., in the specification, claims, and the aforementioned figures are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0067] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0068] It is understood that the rail damage detection method based on multi-domain information collaborative transfer learning provided in this embodiment of the invention can be applied to any computer device with data processing and computing capabilities, and this computer device can be various terminals or servers. When the computer device in the embodiment is a server, the server is an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. Optionally, the terminal can be a smartphone, tablet, laptop, or desktop computer, but it is not limited to these.

[0069] like Figure 1 The diagram shown is a schematic representation of an implementation environment provided by an embodiment of the present invention. (Refer to...) Figure 1 The implementation environment includes at least one terminal 102 and a server 101. The terminal 102 and the server 101 can be connected via a network, either wirelessly or via a wired connection, to complete data transmission and exchange.

[0070] Server 101 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0071] Additionally, server 101 can also be a node server in a blockchain network. Blockchain is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms.

[0072] Terminal 102 can be a smartphone, tablet computer, laptop computer, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. Terminal 102 and server 101 can be directly or indirectly connected via wired or wireless communication, and this embodiment of the invention does not impose any limitations.

[0073] Exemplary based on Figure 1 The implementation environment shown in this embodiment of the invention provides a rail damage detection method based on multi-domain information collaborative transfer learning. The following description uses the application of this rail damage detection method based on multi-domain information collaborative transfer learning in server 101 as an example. It can be understood that this rail damage detection method based on multi-domain information collaborative transfer learning can also be applied in terminal 102.

[0074] Reference Figure 2 , Figure 2 This is a flowchart illustrating a rail damage detection method based on multi-domain information collaborative transfer learning applied to a server, as provided in an embodiment of the present invention. The executing entity of this rail damage detection method based on multi-domain information collaborative transfer learning can be any of the aforementioned computer devices (including servers or terminals). (Refer to...) Figure 2 The method includes the following steps:

[0075] S100. A data augmentation algorithm for random damage is developed based on the scattering simulation signal of a pre-constructed benchmark rail finite element model to obtain source domain simulation ultrasonic guided wave samples.

[0076] Among them, the reference rail finite element model is obtained by correcting the residual optimization results between the measured guided wave and the model predicted guided wave dispersion characteristics;

[0077] It should be noted that in some embodiments, step S100 may include the following steps: optimizing the residual between the dispersion characteristics of the measured guided wave and the predicted guided wave of the rail to obtain the correction result of the physical parameters of the model, and then constructing a reference rail finite element model based on the correction result; randomly adding damage in the damage label area based on the reference rail finite element model, and completing the ultrasonic guided wave scattering simulation to obtain the reference reflected wave signal of the corresponding damage label area; developing a data augmentation algorithm based on scattering simulation driven by the reference reflected wave signal to perform random damage data augmentation processing to obtain source domain simulated ultrasonic guided wave samples.

[0078] For example, in some specific implementations, developing a data augmentation algorithm driven by guided wave scattering simulation of a benchmark model to obtain diverse source domain simulation samples includes the following steps: First, constructing and optimizing an objective function based on the dispersion characteristics of the measured guided wave signal and the model prediction characteristics to update the physical parameters of the rail finite element model, thereby obtaining a high-precision and reliable benchmark finite element model; Second, randomly deleting elements in each labeled region of the benchmark finite element model to simulate damage in that region of the rail, and setting signal excitation and receiving sensors at multiple locations at the rail head, rail web, and rail bottom, while simultaneously exciting ultrasonic guided wave modes sensitive to the entire rail cross-section, completing guided wave scattering simulation under random damage, extracting the reflected wave signals from each sensor due to damage, and constructing the benchmark reflected wave signal for that labeled region; Then, developing a random damage data augmentation algorithm based on the benchmark reflected wave signal to generate a large number of randomly damaged guided wave signals to expand the source domain samples, thereby completing the source domain data diversity enhancement.

[0079] In some embodiments, the residual between the measured guided wave and the predicted guided wave of the rail is optimized to obtain the corrected physical parameters of the model. Based on the corrected results, a reference rail finite element model is constructed. This process may include the following steps: constructing an initial finite element model based on the rail cross-sectional parameters; determining the physical model parameters of the rail to be corrected, including the elastic modulus, Poisson's ratio, and damping loss coefficient; determining the mass and stiffness matrices of the rail periodic waveguide structure under the corresponding parameters based on the physical model parameters; and calculating the rail's mass and stiffness under the corresponding parameters using wave-finite element method analysis based on the mass and stiffness matrices of the rail periodic waveguide structure. The process involves predicting the ultrasonic guided wavenumber under the physical model parameters of the actual rail; acquiring the ultrasonic guided wave signal of the actual rail through sensors, and then constructing the spatiotemporal matrix of the ultrasonic guided wave signal of the actual rail; performing a two-dimensional fast Fourier transform on the spatiotemporal matrix of the ultrasonic guided wave signal to obtain the measured ultrasonic guided wavenumber of the actual rail; determining the wavenumber prediction error and constructing a likelihood function based on the predicted and measured ultrasonic guided wavenumbers; embedding the likelihood function into Bayes' theorem and performing random sampling through asymptotic Markov Monte Carlo to obtain the posterior distribution and optimal values ​​of the physical model parameters of the actual rail; and correcting the initial finite element model based on the optimal values ​​of the physical model parameters to obtain the reference rail finite element model.

[0080] For example, in some specific implementations, taking Bayesian model correction as an example, the wavenumber of ultrasonic guided waves of the rail under different physical model parameters to be corrected is quickly predicted using WFE (wave-finite element) theory. This prediction is then combined with the wavenumber extracted from the measured guided wave signal via 2DFFT (two-dimensional fast Fourier transform) to construct a likelihood function. Finally, a Bayesian inference framework is embedded to obtain the posterior distribution of the parameters to be updated. The specific process includes:

[0081] Based on the cross-sectional parameters of CHN60 rails, a finite element model of the rail is constructed in ABAQUS as follows: Figure 3 As shown, the mass and stiffness matrices of the periodic waveguide structure (a substructure with a longitudinal length of 2 mm) of the rail constructed by the finite element model under different physical model parameters (elastic modulus, Poisson's ratio, and damping loss coefficient) are extracted. Using WFE theory, the wavenumber of the rail at frequency ω under the corresponding physical model parameters is quickly calculated.

[0082] By collecting ultrasonic guided wave signals of corresponding modes at equally spaced measurement points along the longitudinal direction of the rail at the rail head, rail web, and rail foot in the laboratory, a signal spatiotemporal matrix of ultrasonic guided waves propagating in the measured rail is constructed. Based on 2DFFT theory, and by performing Fast Fourier Transform on the signal spatiotemporal matrix in the time and spatial domains, the ultrasonic guided wavenumbers of the corresponding rail modes are obtained. The relationship between the measured wavenumber and the predicted wavenumber can be expressed as:

[0083]

[0084] Where ε ω The prediction error is represented by a zero mean and a variance of σ. 2 Gaussian white noise. Therefore, the likelihood function can be expressed as:

[0085]

[0086] In the above formula, p(D|(θ,σ)) represents the likelihood function, D represents all measured wavenumbers, σ represents the standard deviation of the prediction error, ω1 and ω2 represent the cutoff frequencies of the measured wavenumbers, θ represents the rail physical model parameters to be updated, including the rail's elastic modulus, Poisson's ratio, and damping loss coefficient, and ω represents the frequency. This represents the measured ultrasonic guided wave number. This indicates the predicted ultrasonic guided wave number.

[0087] Embedding the objective function into Bayes' formula, its expression is:

[0088]

[0089] Where λ = {θ, σ} represents the parameters to be updated, p(λ|D) is the posterior distribution of the parameters to be updated, p(λ) is the prior distribution of each parameter, which is generally assumed to be uniform, and p(D) represents the evidence value, which is usually a normalization constant.

[0090] By using the TMCMC (Progressive Markov Monte Carlo) random sampling algorithm to generate a series of samples that progressively approximate the posterior probability distribution of the parameters, the corresponding likelihood functions are evaluated, and the posterior distribution of each parameter is finally obtained. The optimal values ​​of each parameter are then used to construct a rail finite element model, resulting in a corrected rail reference finite element model. This model minimizes the differences between source and target domain signal characteristics, improving the transfer performance of subsequent models.

[0091] In some optional implementations, the steps, based on a reference rail finite element model, involve randomly adding damage within the damage label area and performing ultrasonic guided wave scattering simulation to obtain the reference reflected wave signal for the corresponding damage label area. Specifically, this can be achieved as follows:

[0092] Damage was randomly added to each labeled area of ​​the updated rail finite element model, and the corresponding reference reflected wave signal was collected.

[0093] Specifically, such as Figure 4 As shown, the corresponding rail damage labels are first set as follows: no damage condition (H), one damage condition at any position of the rail head (T), rail web (Y), and rail foot (J), two damage conditions simultaneously at any position of the rail head and rail web (YT), rail head and rail foot (JT), and rail web and rail foot (YT), and three damage conditions simultaneously at any position of the rail head, rail web, and rail foot (JYT), for a total of eight damage type labels. Next, four pairs of PZT piezoelectric ceramic sensors are added to the top of the rail head, the middle of the rail web, and both sides of the rail foot. A sinusoidal signal with a ten-peak Hanning window center frequency of 100kHz is simultaneously applied to the four sensors on the right to excite ultrasonic guided wave signals throughout the entire cross-section of the rail, which propagate longitudinally along the rail. Upon encountering damage, reflected wave signals are generated. The four sensors on the left are used to receive the ultrasonic guided wave signals throughout the entire process. Therefore, the ultrasonic guided wave signal S under the damaged state... i (t) is the superposition of the guided wave signal s0(t) received by the sensor in the undamaged state and the reflected wave signal s due to damage. i (t). Therefore, the reflected wave signal can be obtained by subtracting the undamaged guided wave signal from the damaged guided wave signal.

[0094] s i (t)=S i (t)-s0(t)

[0095] Damage was simulated by randomly deleting units from each label area on the rail. Guided wave signals from each sensor were then collected, and the corresponding reflected wave signals were extracted to form the reference reflected wave signals for each label area. Furthermore, damage at different locations resulted in a larger amplitude of the reflected wave signal received by the corresponding sensor compared to other sensors. The relative relationship between the reflected wave amplitudes extracted from the signals received by the four sensors determined the location of the damage, providing a theoretical basis for data augmentation based on guided wave scattering.

[0096] In some embodiments, a data augmentation algorithm based on scattering simulation is developed using reference reflected wave signals to perform random damage data augmentation processing to obtain source domain simulated ultrasonic guided wave samples. This process may include the following steps: weighting and fusing a preset number of reference reflected wave signals sequentially set in different damage label regions to obtain reflected wave signals showing damage at any location within the damage label region corresponding to the rail cross-section; adjusting the amplitude and phase of the reflected wave signals showing damage at any location within the damage label region corresponding to the rail cross-section to obtain reflected wave signals showing damage of any size at any longitudinal location within the damage label region corresponding to the rail; adding undamaged guided wave signals to the reflected wave signals showing damage of any size at any longitudinal location within the damage label region corresponding to the rail to obtain random damage ultrasonic guided wave signals within the damage label region corresponding to the rail, thus completing the acquisition of source domain simulated ultrasonic guided wave samples.

[0097] For example, in some specific implementations, a data augmentation algorithm based on scattering simulation is developed by using a reference guided wave reflected wave signal to generate a large number of simulated ultrasonic guided wave signals, thereby enhancing the diversity of source domain data.

[0098] In some specific application scenarios, since the ultrasonic guided wave signal under damaged conditions is generated by superimposing the reflected wave signal caused by the damage onto the guided wave signal under undamaged conditions, the final enhanced guided wave signal can be obtained by generating the reflected wave signal corresponding to random damage. However, since the relative relationship between the amplitudes of the reflected waves received by different sensors is related to the damage area, directly adjusting the reflected wave signals of each sensor may cause changes in the newly generated sample labels. Therefore, the signals from the four sensors are combined into a group for unified adjustment. In addition, if the reference reflected wave signal acquired from a single damage is adjusted, the generated signal can only reflect the characteristics of that damage location and cannot complete the data enhancement of damages at other locations. Therefore, assuming that the reflected wave space of each label area is continuous, by sequentially setting n damages in each label area and extracting the n reference reflected wave signals corresponding to each damage as the basis signal of this continuous space, the reflected wave signal s at any location in this space can be generated through weighted fusion. f(t), and can ensure that the damage label corresponding to the newly generated reflected wave signal does not change, wherein the weighted fusion process is represented by the following formula:

[0099] s f (t)=a1×s1(t)+a2×s2(t)+…+a n ×s n (t)

[0100] a1 + a2 + ... + a n =1

[0101] Where a n This represents the weighting coefficients for each reference reflected wave signal. When n damages are uniformly distributed across the entire tag area, damage reflected wave signals at any location within that area can be generated by adjusting the weighting coefficients. Furthermore, the time of the first appearance of the reflected wave signal is determined by the distance between the damage and the sensor, while the amplitude of the reflected wave reflects the severity of the damage. Therefore, by adjusting the amplitude and phase of the generated reflected wave signal and then adding an undamaged guided wave signal, sample data of random spatial damage to the tag can be obtained, expressed as:

[0102] S(t)=s0(t)+b×s f (tc)

[0103] Where S(t) represents the newly synthesized guided wave time-domain signal, b represents the reflected wave amplitude variation coefficient, and c represents the reflected wave phase adjustment coefficient.

[0104] The data augmentation algorithm driven by the guided wave scattering simulation of this benchmark model can be used to augment simulation data of diverse source domain random damage.

[0105] S200: Pre-train a multi-branch convolutional network that coordinates multi-domain information by simulating ultrasonic guided wave samples in the source domain;

[0106] It should be noted that, in some embodiments, step S200 may include the following steps: preprocessing the source domain ultrasonic guided wave signal corresponding to the source domain simulated ultrasonic guided wave sample, and then obtaining the multi-domain signal processing result through multi-domain signal feature extraction processing; the preprocessing includes bandpass filtering and normalization; constructing a multi-branch convolutional network based on the processing results of different types of signal domains in the multi-domain signal processing result; dividing the multi-domain signal processing result into data to obtain a first training set and a first validation set; and pre-training the multi-branch convolutional network based on the first training set using a self-adjusting moment estimation optimization algorithm;

[0107] The number of signal domains in the multi-domain signal processing results corresponds one-to-one with the number of branch structures in the multi-branch convolutional network. The multi-branch convolutional network with multi-domain information collaboration includes sequentially connected branch structures, feature fusion layers, fully connected layers, and classification layers, with multiple branch structures connected in parallel to the feature fusion layer. Each branch structure includes a convolutional module, a spatial attention mechanism, and a Dropout layer. The kernel scale of the convolutional module in each branch structure is determined by the data dimension of the processing results of the corresponding signal domain. The first validation set is used to validate the pre-trained multi-branch convolutional network.

[0108] For example, in some specific implementations, the main steps of pre-training a multi-branch convolutional network with multi-domain information collaboration are as follows: First, the guided wave signal is bandpass filtered and normalized, and then multi-domain signal feature extraction processing is performed on the signal. Second, a multi-branch convolutional network is constructed based on the different signal processing results, and the convolution kernel scale is determined according to the signal characteristics. A spatial attention mechanism is added to capture the interrelationships and global features of different parts of the signal, while the number of convolutional modules in each branch network is determined through Bayesian optimization. After the damage features extracted by the branch networks are fused, the damage location can be classified through a SoftMax classification layer. Finally, the enhanced source domain simulation samples are divided into a training set and a validation set. The multi-branch convolutional network with multi-domain information collaboration is trained in a supervised manner using the training set, and the optimal network hyperparameters are determined using the validation set to obtain the final pre-trained model.

[0109] In some specific application scenarios, the time-domain envelope and frequency-domain spectrum curves are one-dimensional data, while the time-frequency energy distribution map is two-dimensional data. Therefore, to solve the matching problem of data with different dimensions, a multi-branch convolutional network is constructed. The kernel scale of different branches is determined based on the periodicity and fluctuation characteristics of different data types, and an attention mechanism is added to capture the interrelationships and global features of different parts of the signal. The basic architecture of each branch network is as follows: Figure 3 As shown, the system consists of varying numbers of convolutional modules, an attention mechanism, Dropout layers, fully connected layers, and a SoftMax classification layer. The number of convolutional blocks in each branch network is determined through Bayesian optimization. After the damage features extracted by each branch network are fused, the classification of rail damage conditions can be completed through the SoftMax layer.

[0110] For example, the feature extraction results corresponding to the simulated guided wave signals in the source domain can be divided into a training set and a test set in a 7:3 ratio. The multi-branch convolutional network for coordinating multi-domain information on rail damage is then trained based on the training set data. The Adaptive Moment Estimation (ADAM) optimization algorithm is used to optimize backpropagation of errors, while the initial learning rate, regularization coefficient, and the number of convolutional modules in each branch are determined through Bayesian optimization. By continuously optimizing the hyperparameters of the deep learning model and evaluating its performance in the damage location classification task using validation set data, the optimal pre-trained multi-branch convolutional network model for coordinating multi-domain information is finally obtained.

[0111] In some embodiments, the multi-domain signal processing results include a one-dimensional envelope, a one-dimensional spectral curve, and a two-dimensional time-frequency graph. Obtaining the multi-domain signal processing results through multi-domain signal feature extraction can include the following steps: performing time-domain feature extraction processing on the preprocessed source-domain ultrasonic guided wave signal using the Hilbert transform method to obtain the one-dimensional envelope of the source-domain ultrasonic guided wave signal; performing frequency-domain feature extraction processing on the preprocessed source-domain ultrasonic guided wave signal using the fast Fourier transform method to obtain the one-dimensional spectral curve of the source-domain ultrasonic guided wave signal; and performing time-frequency energy feature extraction processing on the preprocessed source-domain ultrasonic guided wave signal using the synchronous squeezing wavelet transform method to obtain the two-dimensional time-frequency graph of the source-domain ultrasonic guided wave signal.

[0112] For example, in some specific implementations, guided wave signals can be preprocessed by time-domain Hilbert transform, frequency-domain fast Fourier transform, and time-frequency domain synchronous squeezing wavelet transform to obtain the corresponding one-dimensional envelope, one-dimensional spectrum curve, and two-dimensional time-frequency diagram.

[0113] Specifically, the reflected wave signal caused by damage has lower energy compared to the signal received by the sensor in the undamaged state, making it difficult to directly extract damage feature information. The time-domain envelope contains key damage features such as signal stability and amplitude variation, the frequency-domain spectrum curve reflects nonlinear information such as higher-order harmonics caused by damage, and the time-frequency plot reveals the damage effect of energy attenuation. By employing multiple data processing methods in the time, frequency, and time-frequency domains, the hidden damage feature information in the guided wave signal can be comprehensively extracted from different perspectives, completing the guided wave signal preprocessing. Therefore, after bandpass filtering and normalization of the acquired ultrasonic guided wave signal in the 80kHz-120kHz range, its time-domain envelope is first extracted using Hilbert transform. Then, based on the length of the guided wave signal received by each sensor, the signal is re-divided into four parts, and a fast Fourier transform is performed on each part to obtain the corresponding spectrum curve, which is then synthesized into a complete set of spectrum curves. Finally, a time-frequency energy distribution map of the ultrasonic guided wave signal is generated through synchronous squeezing wavelet transform.

[0114] S300: Based on a preset number of measured samples, the pre-trained multi-branch convolutional network is adjusted through transfer learning; the adjusted multi-branch convolutional network is used to detect the ultrasonic guided wave signal of the rail obtained in engineering applications to obtain the rail damage detection result.

[0115] It should be noted that in some embodiments, adjusting the pre-trained multi-branch convolutional network through transfer learning based on a preset number of measured samples may include the following steps: using an ultrasonic guided wave sensor to acquire measured guided wave signals of added damage at different locations on a laboratory rail as measured samples; dividing the measured samples into a second training set and a second validation set based on a preset ratio; freezing the hyperparameters of all network layers of the pre-trained multi-branch convolutional network, and then gradually releasing the hyperparameters of each network layer from the classification layer forward, and using the measured data in the second training set to adjust the released hyperparameters of the network layers multiple times with a preset learning rate to complete the transfer of the rail damage detection deep learning model;

[0116] The multi-branch convolutional network with multi-domain information collaboration includes multiple branch structures, a feature fusion layer, a fully connected layer, and a classification layer connected in sequence, with multiple branch structures connected in parallel to the feature fusion layer; the second validation set is used to validate and evaluate the damage detection performance of the deep learning model for rail damage detection after the transfer.

[0117] For example, in some specific implementations, the transfer of a pre-trained model based on measured signals mainly includes the following steps: First, using an ultrasonic guided wave sensor to acquire measured guided wave signals with added damage at different locations on the rail, completing the acquisition of measured samples in the intermediate experimental domain, and performing bandpass filtering and normalization on the acquired measured signals to complete signal preprocessing in the time domain, frequency domain, and time-frequency domain; Second, by gradually freezing and releasing the hyperparameters of different layers of the pre-trained network, the hyperparameters of the released layers are retrained using a small number of measured samples. Simultaneously, the remaining measured data is used to verify the transfer performance of the model, optimize the transfer strategy, finally obtain the optimal transfer model, and evaluate its classification accuracy and generalization ability on the measured samples. Finally, the final transferred model can be directly applied to the damage location classification and detection of rails in service in actual engineering.

[0118] In some specific application scenarios, the excitation and acquisition settings of guided wave signals in the rail reference finite element model can be used as a reference to design an experimental scheme for acquiring guided wave signals of rail damage. A voltage signal is excited by a signal function generator, amplified by a signal amplifier, and transmitted to four PZT piezoelectric ceramic sensors on the right end. The voltage signal is converted into an ultrasonic guided wave signal that propagates in the rail. Different masses of copper blocks are attached to arbitrary positions on the left end of the rail to simulate different types of rail damage, generating corresponding reflected wave signals. These ultrasonic guided wave signals are received by the four PZT piezoelectric ceramic sensors on the left side, converted back into voltage signals, displayed on an oscilloscope, and saved to a computer. The excitation signal of the signal function generator also uses a 100kHz sinusoidal signal with a ten-peak Hanning window center frequency and a sampling frequency of 2MHz. By continuously changing the attachment position of the copper blocks, ultrasonic guided wave signals with different types of rail damage are acquired, completing the acquisition of measured samples in the intermediate experimental domain. Next, the acquired ultrasonic guided wave signals are processed using the signal preprocessing method described in the aforementioned specific implementation to obtain the corresponding preprocessing results for each guided wave signal.

[0119] The transfer of a multi-branch convolutional network for rail damage location classification using experimental intermediate domain measured signals was achieved. The accuracy of the transferred network's damage location classification was verified based on the measured signals. For example, a small number of experimental intermediate domain measured samples were divided into training and validation sets in a 6:4 ratio. All hyperparameters of the pre-trained multi-branch convolutional network were frozen. Then, starting from the classification layer, the hyperparameters of the corresponding layers were gradually released forward. The released layer hyperparameters were adjusted with a small learning rate using experimental intermediate domain training set measured data. Through multiple adjustments, the transfer of the rail damage detection deep learning model was completed. The damage detection performance of the transferred model was evaluated using experimental intermediate domain validation set measured samples, thus determining the optimal transfer model. Furthermore, to verify the performance improvement of the deep learning model after pre-training on source domain simulation samples and transferring it to experimental samples, all hyperparameters of a deep learning model with the same architecture were trained using only experimental intermediate domain training set measured data, and the damage location classification results of this model on the validation set were obtained. A comparative analysis was conducted on the results of the multi-domain information collaborative multi-branch convolutional network before and after transfer learning for classifying damage locations on experimental intermediate domain validation set measured samples. Figure 5 as well as Figure 6 As shown, where Figure 5 The overall accuracy of classifying rail damage locations without transfer learning was 78.6%. Figure 6After transfer learning, the overall accuracy of rail damage location classification was improved to 90.6%. This shows that the multi-branch convolutional network with multi-domain information collaboration, which is pre-trained with source domain data, can improve the classification performance of the damage location classification network on the measured signal after transfer to the experimental intermediate domain measured waveguide signal. It can effectively solve the problem of poor model generalization ability caused by insufficient measured training data, and can provide important support for the detection of measured rail damage defects.

[0120] In some specific application scenarios, consistent with the guided wave signal acquisition scheme for rail damage in the laboratory, four pairs of sensors are used on in-service rails to excite a 100kHz sinusoidal signal with a ten-peak Hanning window center frequency and to receive the guided wave signal. These four pairs of sensors can be mounted on an automatic clamping robotic arm to quickly excite and acquire the guided wave signal. A multi-channel acquisition card can replace an oscilloscope for signal acquisition and storage. Then, using the same signal preprocessing method as the laboratory signal, the corresponding multi-domain information collaborative multi-branch convolutional network input signal is obtained and directly imported into the transferred model to classify the damage location of the in-service rail.

[0121] To explain in detail the principle of the technical solution of the present invention, the overall process of the present invention will be described below with reference to some specific embodiments. It is easy to understand that the following is an explanation of the technical principle of the present invention and should not be regarded as a limitation of the present invention.

[0122] First, it should be noted that the reference... Figure 7 The technical solution of this invention mainly consists of data augmentation driven by guided wave scattering simulation of the benchmark model, pre-training of a multi-branch convolutional network with multi-domain information collaboration, and transfer of the pre-trained model based on measured signals. This invention transfers the deep learning model knowledge learned from source domain samples to the experimental intermediate domain, and uses measured data from the experimental intermediate domain to adjust the model to adapt to the classification of damage locations of measured guided wave signals, thereby achieving the goal of classifying the damage locations of measured rails in the engineering target domain. This effectively overcomes the problem of insufficient measured samples and improves the model's detection accuracy, generalization ability, and practicality in measured data. Detailed steps of this invention are as follows: Figure 3 ( Figure 3 The data tables mentioned are only for illustrative purposes, illustrating data samples and logical relationships. Specific values ​​will vary depending on the actual application, and the values ​​do not affect the technical logic of this invention.

[0123] S1: Developing a data augmentation algorithm driven by guided wave scattering simulation of a benchmark model to obtain diverse source domain simulation samples includes the following steps: First, construct and optimize an objective function based on the dispersion characteristics of measured guided wave signals and the model prediction characteristics to update the physical parameters of the rail finite element model, thereby obtaining a high-precision and reliable benchmark finite element model; Second, randomly delete elements in a specified area of ​​the benchmark finite element model to simulate damage in that area of ​​the rail, and set up signal excitation and receiving sensors at multiple locations at the rail head, rail web, and rail bottom, simultaneously exciting ultrasonic guided wave modes sensitive to the entire rail cross-section to complete guided wave scattering simulation under random damage, extracting the reflected wave signals caused by damage from each sensor, and constructing the benchmark reflected wave signal for the labeled area; Then, develop a random damage data augmentation algorithm based on the benchmark reflected wave signal to generate a large number of randomly damaged guided wave signals to expand the source domain samples, thereby enhancing the diversity of source domain data.

[0124] S2: The main steps of pre-training a multi-branch convolutional network with multi-domain information collaboration are as follows: First, the guided wave signal is bandpass filtered and normalized. Then, the signal undergoes time-domain Hilbert transform, frequency-domain fast Fourier transform, and time-frequency domain synchronous squeeze wavelet transform to obtain the corresponding one-dimensional envelope, spectral curve, and two-dimensional time-frequency plot. Second, a multi-branch convolutional network is constructed based on the results of different signal processing. The convolution kernel scale is determined according to the signal characteristics, and a spatial attention mechanism is added to capture the interrelationships and global features of different parts of the signal. The number of convolutional modules in each branch network is determined through Bayesian optimization. After the damage features extracted by the branch networks are fused, the damage location can be classified through a SoftMax classification layer. Finally, the enhanced source domain simulation samples are divided into training and validation sets. The multi-branch convolutional network with multi-domain information collaboration is trained in a supervised manner using the training set, and the optimal network hyperparameters are determined using the validation set to obtain the final pre-trained model.

[0125] S3: The transfer of the pre-trained model based on measured signals mainly includes the following steps: First, ultrasonic guided wave sensors are used to acquire measured guided wave signals with added damage at different locations on the rail, completing the acquisition of measured samples in the intermediate experimental domain. The acquired measured signals are then bandpass filtered and normalized, completing time-domain, frequency-domain, and time-frequency-domain preprocessing. Second, by gradually freezing and releasing the hyperparameters of different layers of the pre-trained network, the hyperparameters of the released layers are retrained using a small number of measured samples. Simultaneously, the remaining measured data is used to verify the model's transfer performance, optimize the transfer strategy, and finally obtain the optimal transfer model. Its classification accuracy and generalization ability on the measured samples are then evaluated. Finally, the transferred model can be directly applied to the damage location classification and detection of rails in service in actual engineering projects.

[0126] It should be noted that the detailed explanation of each step from S1 to S3 is provided in the example description of the aforementioned specific implementation method, and will not be repeated here.

[0127] In summary, to address the issues in existing technologies, this invention proposes a deep transfer learning-based technical solution for ultrasonic guided wave rail damage location classification. This solution primarily comprises two modules: data augmentation driven by a benchmark model guided wave scattering simulation and pre-training of a multi-branch convolutional network with multi-domain information collaboration. The benchmark model guided wave scattering simulation-driven data augmentation mainly involves optimizing the residual between the dispersion characteristics of the measured guided wave and the model-predicted guided wave to obtain corrected physical parameters of the model. This corrected result is then used to construct a high-precision and reliable benchmark finite element model for the rail, laying the foundation for improving transfer learning performance. Subsequently, cells are randomly deleted from a specified region of the benchmark finite element model to complete ultrasonic guided wave damage scattering simulation, obtaining the benchmark reflected wave signal for that region. By analyzing the intrinsic relationship between the amplitude of the reflected wave obtained from different sensors and the damaged region, a random damage data augmentation algorithm based on weighted fusion of benchmark reflected wave signals, amplitude and phase adjustment, and superposition of lossless signals is developed. This avoids the label mutation problem that may occur during random damage signal generation and increases the diversity and sample size of the source domain simulation signals. Multi-branch convolutional network pre-training, which involves multi-domain information collaboration, primarily utilizes key damage features such as signal stability and amplitude variation in the one-dimensional time-domain envelope of the source domain signal, nonlinear information such as higher-order harmonics caused by damage in the one-dimensional frequency-domain spectral curve, and the damage effect of energy attenuation in the two-dimensional time-frequency diagram. This comprehensively extracts damage feature information from different angles, avoiding the problem of reduced model detection accuracy caused by the lack of damage feature information when using only a single type of signal. Furthermore, by constructing a multi-branch convolutional network, the matching problem of data from different dimensions is solved. Different scale branches are designed according to the characteristics of different types of signal data, and an attention mechanism is introduced to capture the interrelationships and global features of different parts of the signal, further enhancing the extraction capability of key damage feature information. By fusing the damage features extracted from each branch network, accurate classification of rail damage locations is achieved. A pre-trained model is obtained using rich source domain augmented simulation signals. Subsequently, based on the pre-trained model, guided wave measured signals under a small number of laboratory rail damage conditions are used. Different layers of the pre-trained network are gradually frozen and released, and the released layers are retrained and evaluated with a low learning rate to determine the optimal transfer learning strategy, completing the transfer of the rail damage detection network. In practical applications, the acquired operational guided wave signals are bandpass filtered and normalized, and after time-domain, frequency-domain, and time-frequency-domain preprocessing, they are directly imported into the migrated rail damage detection network to obtain the actual rail damage status, demonstrating high engineering practicality.

[0128] It should be noted that the rail damage detection technical solution provided by this invention, based on benchmark model guided wave scattering simulation-driven data augmentation and multi-domain information collaborative multi-branch convolutional network transfer learning, fully utilizes the unique advantage of ultrasonic guided waves in detecting large-area, small-scale rail damage. Only four pairs of piezoelectric ceramic sensors are needed to excite and receive guided wave signals sensitive to damage at any location, achieving efficient classification of random damage areas while significantly reducing equipment costs. The benchmark model guided wave scattering simulation-driven data augmentation algorithm developed in this invention avoids the label change problem that may be caused by signal augmentation at any location, and can quickly generate high-quality, diverse source domain data, greatly improving the generalization ability of the damage detection network. The multi-domain information collaborative multi-branch convolutional network constructed in this invention comprehensively extracts weak damage features from complex guided wave signals from multiple angles, avoiding feature information loss and effectively solving the matching problem of data from different dimensions. Simultaneously, the introduction of an attention mechanism further enhances the extraction capability of key features, significantly improving the damage classification accuracy of the model. This method pre-trains a damage location classification network by enhancing the source domain signal and then transfers the model using a small number of experimental intermediate domain measured samples. This overcomes the challenge of obtaining large amounts of measured data and improves the model's detection accuracy and generalization ability in measured signals. The transferred model can be directly applied to engineering target domains, greatly enhancing the method's engineering applicability. This method is simple to operate, provides efficient and accurate detection results, and exhibits good robustness. It is suitable for damage detection of in-service rails and has broad application prospects.

[0129] Compared with existing methods, the deep transfer learning-based ultrasonic guided wave damage location classification method for rails proposed in this invention is mainly used to determine whether damage exists in the rail and the regions in the rail cross-section where the damage occurs, such as the rail head, rail web, and rail foot. Compared with existing technologies, this invention has at least the following beneficial effects:

[0130] 1. This invention utilizes source-domain enhanced simulated ultrasonic guided wave signals to train a high-precision, highly generalizable, multi-branch convolutional network for rail damage location classification, integrating multi-domain information. The pre-trained network is fine-tuned using a small number of experimental intermediate domain measured ultrasonic guided wave signals, enabling transfer learning. This effectively improves the model's accuracy, generalization ability, and applicability in classifying actual rail damage locations in both the experimental intermediate domain and the engineering target domain. It overcomes the limitation of obtaining a large number of measured damage samples, significantly enhancing the model's performance and application value.

[0131] 2. This invention updates the main physical model parameters of the rail finite element model based on the dispersion characteristics extracted from the measured ultrasonic guided wave signal, obtains a reliable rail reference finite element model, improves the simulation accuracy of ultrasonic guided wave scattering, reduces the difference between the simulated signal and the measured signal characteristics, and improves the transfer performance of the subsequent deep learning model.

[0132] 3. Based on the principle of reflected wave signal generation when there is damage at different locations on the rail, this invention develops a data augmentation algorithm driven by guided wave scattering simulation of the benchmark model by using the benchmark reflected wave signals of each damage label region extracted from the updated finite element model. This enables the generation of guided wave signals when there is damage at any location on the rail, greatly enhancing the diversity of source domain simulation data and improving the generalization ability of the model.

[0133] 4. This invention develops a multi-domain information collaborative multi-branch convolutional network, which fully utilizes the characteristics of different types of signals in the time domain, frequency domain, and time-frequency domain after preprocessing. Combined with the attention mechanism, it comprehensively extracts and fuses the hidden damage features in the ultrasonic guided wave signal, thereby further improving the convolutional network's ability to classify rail damage locations.

[0134] 5. By installing four pairs of ultrasonic guided wave sensors at the rail head, rail foot, and rail web, this invention can excite and receive ultrasonic guided wave signals from all directions across the entire rail cross section, enabling the detection and classification of damage at any location on the rail.

[0135] On the other hand, such as Figure 8 As shown, this embodiment of the invention provides a rail damage detection system 900 based on multi-domain information collaborative transfer learning, which may include:

[0136] The first module 901 is used to develop a data augmentation algorithm for random damage based on the scattering simulation signal of the pre-built reference rail finite element model to obtain source domain simulation ultrasonic guided wave samples; the reference rail finite element model is obtained by correcting the residual optimization results between the measured guided wave and the model predicted guided wave dispersion characteristics.

[0137] The second module 902 is used to pre-train a multi-branch convolutional network that coordinates multi-domain information by simulating ultrasonic guided wave samples in the source domain.

[0138] The third module 903 is used to adjust the pre-trained multi-branch convolutional network based on a preset number of measured samples through transfer learning; the adjusted multi-branch convolutional network is used to detect the ultrasonic guided wave signal of the rail obtained in the engineering application to obtain the rail damage detection result.

[0139] The content of the method embodiments of the present invention is applicable to the system embodiments. The specific functions implemented in the system embodiments are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above methods.

[0140] On the other hand, embodiments of the present invention also provide an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the aforementioned rail damage detection method based on multi-domain information collaborative transfer learning. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0141] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0142] like Figure 9 As shown, Figure 9 The hardware structure of an electronic device 1000 according to another embodiment is illustrated. The electronic device 1000 includes:

[0143] The processor 1001 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (aSIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present invention.

[0144] The memory 1002 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RaM). The memory 1002 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1002 and is called and executed by the processor 1001 to execute the network node population optimization method of the embodiments of this invention.

[0145] Input / output interface 1003 is used to implement information input and output;

[0146] The communication interface 1004 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0147] Bus 1005 transmits information between various components of the device (e.g., processor 1001, memory 1002, input / output interface 1003, and communication interface 1004);

[0148] The processor 1001, memory 1002, input / output interface 1003 and communication interface 1004 are connected to each other within the device via bus 1005.

[0149] The electronic device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0150] The content of the method embodiments of the present invention is applicable to the embodiments of the present electronic device. The specific functions implemented by the embodiments of the present electronic device are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above methods.

[0151] Another aspect of this invention provides a computer-readable storage medium storing a program that is executed by a processor to implement the aforementioned method.

[0152] It should be noted that the computer-readable medium shown in the embodiments of the present invention can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD to ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In the present invention, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, wherein computer-readable program code is carried. Such transmitted data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.

[0153] The content of the method embodiments of the present invention is applicable to the computer-readable storage medium embodiments. The specific functions implemented by the computer-readable storage medium embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above methods.

[0154] This invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the aforementioned method.

[0155] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0156] It should be noted that although several modules for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to embodiments of the present invention, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0157] Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of the present invention can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, portable hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, touch terminal, or network device, etc.) to execute the method according to the embodiments of the present invention.

[0158] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this invention are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0159] Furthermore, although the invention has been described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding the invention. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of conventional skill of an engineer. Therefore, those skilled in the art can implement the invention as set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and not intended to limit the scope of the invention, which is determined by the full scope of the appended claims and their equivalents.

[0160] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0161] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution means, apparatus, or device (such as a computer-based device, a processor-including device, or other means that can fetch and execute instructions from, or in conjunction with, an instruction execution means, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution means, apparatus, or device.

[0162] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0163] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution device. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0164] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0165] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0166] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of the present invention.

Claims

1. A rail damage detection method based on multi-domain information collaborative transfer learning, characterized in that, Includes the following steps: A data augmentation algorithm for random damage is developed based on the scattering simulation signal of a pre-constructed benchmark rail finite element model to obtain simulated ultrasonic guided wave samples in the source domain; the benchmark rail finite element model is obtained by correcting the residual optimization results between the measured guided wave and the model-predicted guided wave dispersion characteristics. The source domain simulated ultrasonic guided wave samples were used to pre-train a multi-branch convolutional network that coordinates multi-domain information. Based on a preset number of measured samples, the pre-trained multi-branch convolutional network is adjusted through transfer learning; the adjusted multi-branch convolutional network is used to detect the ultrasonic guided wave signal of the rail obtained in engineering applications to obtain the rail damage detection result. 2.The method of claim 1, wherein, The method for developing a data augmentation algorithm for random damage based on the scattering simulation signal of a pre-constructed benchmark rail finite element model to obtain source domain simulated ultrasonic guided wave samples includes the following steps: By optimizing the residual between the measured guided wave of the rail and the predicted guided wave of the model, the correction results of the physical parameters of the model are obtained, and then the reference rail finite element model is constructed based on the correction results. Based on the reference rail finite element model, damage is randomly added within the damage label area, and ultrasonic guided wave scattering simulation is completed to obtain the reference reflected wave signal corresponding to the damage label area. A data augmentation algorithm based on scattering simulation was developed using the reference reflected wave signal to perform random damage data augmentation processing, thereby obtaining the source domain simulated ultrasonic guided wave sample. 3.The method of claim 2, wherein, The process of optimizing the residual between the measured guided wave of the rail and the predicted guided wave of the model to obtain the corrected physical parameters of the model, and then constructing the reference rail finite element model based on the corrected results, includes the following steps: Based on the rail cross-sectional parameters, an initial finite element model is constructed, and the physical model parameters of the rail to be corrected are determined. The physical model parameters of the rail to be corrected include the elastic modulus, Poisson's ratio, and damping loss coefficient. Based on the physical model parameters of the rail to be corrected, the mass and stiffness matrices of the rail periodic waveguide structure under the corresponding parameters are determined. Based on the mass and stiffness matrix of the rail periodic waveguide structure, the predicted ultrasonic wave number of the rail under the corresponding physical model parameters of the rail to be corrected is obtained by wave-finite element method analysis and calculation. The ultrasonic guided wave signal of the actual rail is acquired by a sensor, and then the spatiotemporal matrix of the ultrasonic guided wave signal of the actual rail is constructed. The measured ultrasonic guided wave number of the actual rail is obtained by performing a two-dimensional fast Fourier transform on the spatiotemporal matrix of the ultrasonic guided wave signal. Based on the predicted ultrasonic guided wavenumber and the measured ultrasonic guided wavenumber, the wavenumber prediction error is determined and a likelihood function is constructed. The likelihood function is embedded into the Bayesian formula, and random sampling is performed using asymptotic Markov Monte Carlo to obtain the posterior distribution and optimal value of the physical model parameters of the actual rail. The initial finite element model is corrected based on the optimal values ​​of the physical model parameters to obtain the reference rail finite element model.

4. The rail damage detection method based on multi-domain information collaborative transfer learning according to claim 2, characterized in that, The step of developing a data augmentation algorithm based on scattering simulation driven by the reference reflected wave signal to perform random damage data augmentation processing and obtain the source domain simulated ultrasonic guided wave sample includes the following steps: A weighted fusion process is performed on a preset number of reference reflected wave signals sequentially set in different damage label areas to obtain reflected wave signals of damage occurring at any position within the damage label area corresponding to the rail cross section. The amplitude and phase of the reflected wave signal of damage appearing at any position within the damage label area corresponding to the cross-section of the rail are adjusted to obtain the reflected wave signal of damage of any size at any longitudinal position within the damage label area corresponding to the rail. By adding a non-damaging guided wave signal to the reflected wave signal of any longitudinal position with damage of any size within the damage label area corresponding to the rail, a random damage ultrasonic guided wave signal within the damage label area corresponding to the rail is obtained, thus completing the acquisition of the source domain simulation ultrasonic guided wave sample.

5. The rail damage detection method based on multi-domain information collaborative transfer learning according to claim 1, characterized in that, The pre-training of a multi-branch convolutional network that coordinates multi-domain information using the source domain simulated ultrasonic guided wave samples includes the following steps: The source domain ultrasonic guided wave signal corresponding to the source domain simulated ultrasonic guided wave sample is preprocessed, and then multi-domain signal processing results are obtained through multi-domain signal feature extraction; the preprocessing includes bandpass filtering and normalization. The multi-branch convolutional network is constructed based on the processing results of different types of signal domains in the multi-domain signal processing results; The number of signal domains in the multi-domain signal processing results corresponds one-to-one with the number of branch structures in the multi-branch convolutional network. The multi-branch convolutional network for multi-domain information collaboration includes sequentially connected branch structures, feature fusion layers, fully connected layers, and classification layers, with multiple branch structures connected in parallel to the feature fusion layer. Each branch structure includes a convolutional module, a spatial attention mechanism, and a Dropout layer. The kernel scale of the convolutional module in each branch structure is determined by the data dimension of the processing results of the corresponding signal domain. The multi-domain signal processing results are divided into a first training set and a first validation set. Based on the first training set, the multi-branch convolutional network is pre-trained using a self-adjusting moment estimation optimization algorithm; The first validation set is used to validate the pre-trained multi-branch convolutional network.

6. The rail damage detection method based on multi-domain information collaborative transfer learning according to claim 5, characterized in that, The multi-domain signal processing results include a one-dimensional envelope, a one-dimensional spectral curve, and a two-dimensional time-frequency plot; the process of obtaining the multi-domain signal processing results through multi-domain signal feature extraction includes the following steps: The source domain ultrasonic guided wave signal is subjected to time domain feature extraction processing by the Hilbert transform method to obtain the one-dimensional envelope of the source domain ultrasonic guided wave signal; The source domain ultrasonic guided wave signal is subjected to frequency domain feature extraction processing by fast Fourier transform to obtain a one-dimensional spectrum curve of the source domain ultrasonic guided wave signal. The source domain ultrasonic guided wave signal is subjected to time-frequency energy feature extraction processing by synchronous squeezing wavelet transform method to obtain a two-dimensional time-frequency diagram of the source domain ultrasonic guided wave signal.

7. The rail damage detection method based on multi-domain information collaborative transfer learning according to claim 1, characterized in that, The process of adjusting the pre-trained multi-branch convolutional network using transfer learning based on a preset number of test samples includes the following steps: The measured guided wave signals of added damage at different locations on the laboratory rail were obtained using an ultrasonic guided wave sensor as the measured samples. The measured samples are divided into a second training set and a second validation set based on a preset ratio. The multi-branch convolutional network with multi-domain information collaboration includes multiple branch structures, a feature fusion layer, a fully connected layer and a classification layer connected in sequence, with the multiple branch structures connected in parallel to the feature fusion layer; The hyperparameters of all network layers of the pre-trained multi-branch convolutional network are frozen, and then the hyperparameters of each network layer are gradually released from the classification layer forward. The released hyperparameters of the network layers are adjusted multiple times with a preset learning rate using the measured data in the second training set to complete the transfer of the deep learning model for rail damage detection. The second validation set is used to validate and evaluate the damage detection performance of the deep learning model for rail damage detection after the transfer is completed.

8. A rail damage detection system based on multi-domain information collaborative transfer learning, characterized in that, include: The first module is used to develop a data augmentation algorithm for random damage based on the scattering simulation signal of a pre-built benchmark rail finite element model, and to obtain source domain simulation ultrasonic guided wave samples. The reference rail finite element model is obtained by correcting the residual optimization results between the measured guided wave and the model predicted guided wave dispersion characteristics. The second module is used to pre-train a multi-branch convolutional network that coordinates multi-domain information using the source domain simulated ultrasonic guided wave samples. The third module is used to adjust the pre-trained multi-branch convolutional network based on a preset number of test samples through transfer learning; The adjusted multi-branch convolutional network is used to detect the ultrasonic guided wave signal of the rail obtained in engineering applications, and the rail damage detection results are obtained.

9. An electronic device, characterized in that, Including the processor and memory; The memory is used to store programs; The processor executes the program to implement the method as described in any one of claims 1 to 7.

10. A computer storage medium storing a processor-executable program, characterized in that, The processor-executable program, when executed by the processor, is used to implement the method as described in any one of claims 1 to 7.

Images