A method for determining ultrasonic guided wave damage based on a neural network model, electronic equipment, dielectric, and software products.

Abstract

This application provides a method, electronic device, medium, and program product for determining ultrasonic guided wave damage based on a neural network model. The method includes: acquiring the time-domain signal of the ultrasonic guide; determining the update cycle of neurons according to a preset periodic function; determining the dynamic phase according to the update cycle; inputting the dynamic phase into a gating function and outputting a gating value; obtaining candidate states; determining the hidden state of the time step based on the gating value, the candidate states, and the hidden state of the time step; obtaining the hidden state sequence corresponding to the time-domain signal; and determining the damage result according to the hidden state sequence. The technical solution provided by this application improves classification accuracy and has excellent generalization robustness.

Description

Technical Field

[0001] This invention relates to the field of ultrasonic guided wave damage technology, and more specifically, to a method, electronic device, medium, and program product for determining ultrasonic guided wave damage based on a neural network model. Background Technology

[0002] Carbon fiber reinforced polymer (CFRP) composites, with their superior specific strength, excellent corrosion resistance, and designable anisotropy, have become key structural materials in modern aerospace, wind power generation, and high-end transportation equipment. However, due to the unique characteristics of laminated structures, CFRPs are highly susceptible to hidden damage such as matrix cracking, delamination, and fiber breakage during manufacturing and use. These damages often originate and propagate within the material, making them difficult to detect visually. Once they evolve into catastrophic failures, they pose a serious threat to structural safety. Therefore, developing efficient and accurate structural health monitoring (SHM) technology is of significant engineering importance for ensuring the safety of composite material structures and reducing maintenance costs. Among numerous non-destructive testing techniques, ultrasonic guided wave (UGW)-based monitoring methods are widely recognized as the benchmark for health monitoring of large-area plate and shell structures due to their advantages such as wide coverage of single-point excitation, long propagation distance, and sensitivity to minute changes in acoustic impedance.

[0003] Dispersion effects are a core bottleneck restricting the accuracy of guided wave detection. When a guided wave propagates in a waveguide medium, its group velocity is a function of frequency. This means that during propagation, different frequency components of a broadband guided wave packet will travel at different speeds. As the propagation distance increases, the wave packet, which was originally compact and sharp in the near field, undergoes severe nonlinear stretching and waveform distortion when it reaches the far field. High-frequency and low-frequency components gradually separate on the time axis, resulting in a significant increase in the time width of the wave packet, a substantial attenuation of the amplitude, and complex tailing phenomena. Therefore, how to maintain accurate capture of damage characteristics under the influence of different dispersion effects is a key problem that urgently needs to be solved in this field. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide an ultrasonic guided wave damage determination method, electronic device, medium and program product based on a neural network model, which can improve the problem that the existing technology still maintains inaccurate capture of damage characteristics under the action of different dispersion effects.

[0005] To achieve the above technical objectives, the technical solution adopted in this application is as follows:

[0006] In a first aspect, embodiments of this application provide a method for determining ultrasonic guided wave damage based on a neural network model, the method comprising:

[0007] S1: Obtain the time-domain signal X of the ultrasonic waveguide, where, n represents the number of sampling points;

[0008] S2: Based on Time step The neuron update cycle is determined based on a preset periodic function. The periodic function is obtained through and time step The product of these terms yields the update period. ,in This represents the adjustment factor, which is a hyperparameter of the neural network model. ;

[0009] S3: According to the update cycle Determine the time step Relative to update cycle dynamic phase ;

[0010] S4: The dynamic phase The input is fed into a gating function, which is based on the dynamic phase. Compared with the preset ratio range Output gate value To enable neurons to be in dynamic phase Located in the ratio range It is in an active state at the time;

[0011] S5: According to and time step Hidden state Based on the reset and update gates of the neural network model, candidate states are obtained. ;

[0012] S6: Based on the gate value Candidate status and time step Hidden state Determine the time step Hidden state ,in, All are 0;

[0013] S7: Repeat S2-S6 until the hidden state sequence H1 corresponding to the time-domain signal X is obtained. ;

[0014] S8: Determine the damage result based on the hidden state sequence H1, wherein the damage result includes the damage location coordinates and / or the damage category.

[0015] According to the first aspect, the periodic function is:

[0016]

[0017] Indicates the baseline update period for network initialization; It is a constant that controls the scaling factor.

[0018] According to the first aspect, the gate function is:

[0019]

[0020] This represents the leakage scaling constant.

[0021] According to the first aspect, determining the damage result based on the hidden state sequence H1 includes:

[0022] The hidden state sequence H1 is input into the fully connected layer of the neural network model, and the initial probability of each damage category is output through the fully connected layer;

[0023] Calculate the L2 norm of each hidden state in the hidden state sequence H1 to obtain the energy corresponding to each hidden state, thus obtaining the energy sequence E. ,in, They represent Energy;

[0024] Extract the time steps corresponding to the first q energies from the energy sequence E to obtain the set of active time peaks. , where q is an integer greater than or equal to 2 and less than or equal to n;

[0025] Based on the set of active time peaks The initial probability is corrected by the wave packet arrival time range for each damage category, wherein the wave packet arrival time range is obtained through preset or pre-training.

[0026] Based on the corrected initial probability, the damage category of the time-domain signal X is determined.

[0027] According to the first aspect, the set of active time peaks The initial probability is corrected by the wave packet arrival time range for each damage category, including:

[0028] Through the set of active time peaks A consistency score is determined based on the wave packet arrival time range for each damage category, and this consistency score is correlated with the set of active time peaks. The number of time steps corresponding to the first q energies of the wave packet arriving in the time range is positively correlated;

[0029] The consistency score and the initial probability are substituted into the preset model, and the output of the preset model is the corrected initial probability;

[0030] The preset model is:

[0031]

[0032] To avoid zeroing out items;

[0033] Indicates temperature parameter;

[0034] This represents the corrected initial probability;

[0035] Indicates the initial probability;

[0036] An index representing the damage category;

[0037] This indicates the total number of damage categories;

[0038] Represents the set of active time peaks The temporal consistency score with the k-th damage category.

[0039] According to the first aspect, determining the damage result based on the hidden state sequence H1 includes:

[0040] The hidden state sequence H1 is input into the linear activation layer of the neural network model, and the initial coordinates are output through the linear activation layer;

[0041] Calculate the L2 norm of each hidden state in the hidden state sequence H to obtain the energy corresponding to each hidden state, and then obtain the energy-based sequence E. ,in, They represent Energy;

[0042] The time step in which the maximum energy value in the energy sequence is located. Construct a peak feature vector centered at the core, wherein the time step index of the peak feature vector is... The peak feature vector is represented as ;

[0043] Each hidden state in the peak feature vector is input into a pre-trained decoder, which outputs the correction amount of the peak feature vector for each time step relative to the initial coordinates, as well as the confidence score.

[0044] Based on the correction amount and confidence score corresponding to all time steps in the feature vector, the initial coordinates are corrected to obtain the corrected initial coordinates, thereby obtaining the damage result.

[0045] Secondly, embodiments of this application provide an electronic device, which includes a processor and a memory coupled to each other. The memory stores a computer program, and when the computer program is executed by the processor, the electronic device performs the method described in the first aspect.

[0046] Thirdly, embodiments of this application provide a computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program, which, when run on a computer, causes the computer to perform the method described in the first aspect.

[0047] Fourthly, embodiments of this application propose a program product, characterized in that it includes a computer program, which, when executed by a processor, implements the method described in the first aspect.

[0048] The invention employing the above technical solution has the following advantages:

[0049] In the technical solution provided in this application, a learnable modulation factor is used. With time step of The product is a dynamic periodic function, and is based on the update period. Calculate dynamic phase Then, the time-varying gated value is output using a gating function. This approach breaks the limitation of traditional recurrent networks that update hidden states with a fixed step size, endowing each neuron with the ability to adaptively stretch or compress information update rhythm based on the local dispersion characteristics of the signal. In the temporal feature extraction process, the reset gate and candidate states inherit the excellent modeling capabilities of GRU, while the time-varying modulation gate precisely replaces the update gate, allowing the hidden state to... Substantial updates are performed only in critical phase intervals where wave packet energy is concentrated and closely related to damage, while historical memory is forcibly preserved in noisy and tailed intervals. This method, which learns the stretching pattern of wave packets intrinsically without any explicit dispersion compensation preprocessing, achieves end-to-end adaptive alignment of dispersion and features in the feature space, enhancing the physical interpretability of the model. Finally, the damage result is obtained by decoding the hidden state sequence H1, effectively suppressing the interference of dispersion effects in non-stationary, time-varying distorted guided wave signals on localization and classification, significantly reducing localization errors, improving classification accuracy, and exhibiting excellent generalization robustness. Attached Figure Description

[0050] This application can be further illustrated by the non-limiting embodiments given in the accompanying drawings. It should be understood that the following drawings only illustrate some embodiments of this application and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained from these drawings without any inventive effort.

[0051] Figure 1 A flowchart of an ultrasonic guided wave damage determination method based on a neural network model provided in an embodiment of this application.

[0052] Figure 2 This is a structural diagram of the TVMU unit provided in an embodiment of this application. Detailed Implementation

[0053] The present application will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that similar or identical parts are referred to by the same reference numerals in the drawings or description. Implementations not shown or described in the drawings are forms known to those skilled in the art. In the description of this application, terms such as "first" and "second" are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0054] Please refer to Figure 1 This application provides a method for determining ultrasonic guided wave damage based on a neural network model, which can be applied to electronic devices and whose steps can be executed or implemented by the electronic device. The electronic device can be, but is not limited to, personal computers, smartphones, and other electronic devices. The method for determining ultrasonic guided wave damage based on a neural network model may include the following steps:

[0055] S1: Obtain the time-domain signal X of the ultrasonic waveguide, where, n represents the number of sampling points;

[0056] S2: Based on Time step The neuron update cycle is determined based on a preset periodic function. The periodic function is obtained through and time step The product of these terms yields the update period. ,in This represents the adjustment factor, which is a hyperparameter of the neural network model. ;

[0057] S3: According to the update cycle Determine the time step Relative to update cycle dynamic phase ;

[0058] S4: The dynamic phase The input is fed into a gating function, which is based on the dynamic phase. Compared with the preset ratio range Output gate value So that the neuron is in dynamic phase Located in the ratio range It is in an active state at the time;

[0059] S5: According to and time step Hidden state Based on the reset and update gates of the neural network model, candidate states are obtained. ;

[0060] S6: Based on the gate value Candidate status and time step Hidden state Determine the time step Hidden state ,in, All are 0;

[0061] S7: Repeat S2-S6 until the hidden state sequence H1 corresponding to the time-domain signal X is obtained. ;

[0062] S8: Determine the damage result based on the hidden state sequence H1, wherein the damage result includes the damage location coordinates and / or the damage category.

[0063] In this embodiment, addressing the issue of nonlinear stretching and distortion of wave packets along the time axis due to dispersion effects in long-distance, large-area ultrasonic guided wave monitoring, which causes traditional deep learning models to fail in feature extraction, this embodiment breaks away from the limitations of traditional RNNs and their variants that rely on static, fixed-time-step updates by introducing a learnable dynamic modulation factor within the network. With time-varying gating modulation mechanism, the TVMU network can adaptively adjust the update cycle of neurons based on the local frequency characteristics and wave packet stretching of the input guided wave signal. This mechanism acts as a dispersion adaptive filter for physical perception during the feature extraction stage, achieving dynamic alignment between network information flow updates and physical wave packet distortion.

[0064] This embodiment constructs a complete TVMU ultrasonic guided wave damage localization network architecture. This architecture aims to directly extract highly dispersion-robust damage features from a one-dimensional raw signal sequence and ultimately decode the precise damage coordinates or region category. Its overall process includes the following three main modules: signal input and preprocessing module: the initial input of the model is the one-dimensional ultrasonic guided wave time-domain signal received by the multi-channel sensor. (n is the number of sampling points). Considering low-frequency drift interference, the signal is first processed through a high-pass filter; subsequently, to reduce sequence length, simplify features, and reduce computational redundancy in subsequent networks, the signal is compressed through a max-pooling layer. TVMU-based feature extraction module: The preprocessed sequence features are fed into the core TVMU layer. In this module, the network processes long sequence data in parallel, utilizing the time-varying gated modulation mechanism described above, and continuously optimizing the dynamic modulation factor based on the backpropagation algorithm. Each neuron adaptively stretches or compresses its temporal receptive field based on local frequency characteristics, accurately capturing the lesion waveform affected by dispersion during dynamic evolution. Feature fusion and linear decoding module: After processing by the TVMU layer, the extracted high-dimensional temporal feature vector is flattened and input into a series of fully connected layers. These layers are responsible for nonlinear mapping and deep information coupling of the extracted features. At the end of the architecture, decoding is performed according to different task requirements: for lesion region classification tasks, the network's end connects to a softmax layer to output the predicted class probability; for lesion localization tasks accurate to coordinates (x, y), a linear activation layer directly regresses and outputs the two-dimensional coordinates of the lesion.

[0065] The steps of the ultrasonic guided wave damage determination method based on a neural network model will be described in detail below:

[0066] In S1, the ultrasonic guided wave time-domain signal can be obtained in the following way. Acquisition: First, an array of piezoelectric ceramic sensors is arranged on the surface of the plate-like or tubular structure to be detected. Typically, one sensor is used as the excitation end, and the rest are used as the receiving ends. A narrowband excitation pulse with a center frequency modulated by a Hanning window is generated by an arbitrary waveform generator, amplified by a power amplifier, and applied to the excitation sensor. The inverse piezoelectric effect is used to excite guided wave modes such as Lamb waves in the structure. If the guided wave encounters damage such as cracks or corrosion while propagating in the structure, it will produce reflection, scattering, or energy attenuation, resulting in changes in the amplitude, phase, and mode of the received signal. The receiving sensor converts the mechanical vibration into a voltage signal based on the piezoelectric effect, which is synchronously acquired by a multi-channel data acquisition card at a fixed sampling frequency to obtain a discrete time-domain sequence. ,in The number of sampling points is given. This sequence will then undergo the aforementioned high-pass filtering and amplitude normalization preprocessing before being used as the input to the TVMU network.

[0067] In S2, the periodic function is expressed as:

[0068]

[0069] Indicates the baseline update period for network initialization; It is a constant that controls the scaling factor. The regulation factor represents the hyperparameters of the neural network model proposed in this application, which can be obtained through training.

[0070] In this embodiment, As a non-linear activation term, it controls the rate and direction of stretching or compressing during neuron update cycles. When the network learns... hour, Positive, update cycle It will happen over time It gradually increases in size over time. This aligns with the physical phenomena of far-field dispersive wave packets, such as decreased frequency, elongated waveform, and tailing, allowing neurons to slow down their update pace and effectively capture the low-frequency stretching characteristics of dispersive waves. When the network learns... As the period decreases, the network focuses more on capturing the high-frequency transient changes in the wave packet front. For example... Figure 2 The diagram shown is a structural diagram of the TVMU.

[0071] This embodiment uses a static neuron update cycle. Replace with an adaptive update cycle that changes dynamically over time. And it intrinsically matches the wave packet stretching mode through learnable parameters. Specifically, it uses learnable dynamic modulation factors. As a non-linear activation term, it controls the stretching and compression of the neuron's update cycle. The sign and magnitude of the wave packet are learned entirely by data, without the need for manual pre-setting of the wave packet shape. This allows the network to automatically discover statistical patterns of dispersion from a large number of labeled guided wave signals. When learning is positive, the update cycle is... As propagation time gradually increases, corresponding to the physical changes of decreased far-field wave packet frequency, elongated waveform, and intensified tailing, neurons slow down their update rhythm, allowing for a wider receptive field to fully cover the stretched wave packet, avoiding fragmentation and loss of low-frequency features; when When the learning period is negative, it gradually decreases, allowing neurons to effectively capture steep transient changes and dense detail fluctuations at the wave packet leading edge. This time-varying periodic adaptive mechanism enables the network to actively align temporal distortions caused by dispersion during the feature extraction stage without increasing the number of layers or parameters. It refocuses damage information that was originally scattered across different time scales onto a few key dimensions of the hidden state, reducing the mapping complexity of the fully connected decoding layer. Ultimately, this improves localization accuracy while enhancing the model's generalization ability to changes in propagation distance, modal mixing, and environmental interference.

[0072] In S3, after determining the dynamic update period, it is necessary to calculate the dynamic phase of the current time step relative to that period. This is used to determine the activity level of neurons at the current moment. The formula for calculating the phase is as follows:

[0073]

[0074] In the formula, This is the phase offset parameter.

[0075] In S4, the gate function is expressed as:

[0076]

[0077] This represents the leakage scaling constant.

[0078] In a TVMU unit, "activated state" refers to the state of a neuron in a dynamic phase. When the input signal falls within a preset ratio range (e.g., 0.2 to 0.7), the system operates in a mode that is highly sensitive to the input signal and actively updates information. At the gating level, the time-varying modulation gate... When the hidden state is opened to near 1, allowing free flow of information, at the state update level, the hidden state is almost entirely determined by the current candidate state. Neurons discard old memories and write the new features carried by the wave packet at this moment into their memory. At the physical function level, the activated state means that the neuron determines that it is currently in the phase where the guided wave packet energy is concentrated, and therefore acquires key damage information such as abrupt changes, peaks, or mode transitions in the waveform with the highest openness. In contrast, when the phase is outside the proportional range, the neuron is in a dormant state, the gating is closed to a tiny leakage value, the hidden state is hardly updated, thus effectively shielding against noise and redundant signal interference.

[0079] In this embodiment, the "activity level" of the neuron refers to the strength of the neuron's response to the input signal at the current moment, that is, whether it is ready and to what extent it is open to accepting and remembering the external information at this moment. This activity level is determined by the dynamic phase. Decision: When the phase falls within the preset active range, the time-varying modulation gate... When the modulation gate is turned up to near 1, it indicates that the neuron is in a "highly active" state. In this state, it actively receives input signals, writes the damage characteristics carried by the wave packet into the hidden state, and completes a crucial information update. Conversely, when the phase is outside the active region, the modulation gate... As the leakage constant approaches a very small value, the neuron enters an "inactive" or "dormant" state, where it responds almost no to new inputs, and the hidden state strictly maintains historical memory. Therefore, the activity level is essentially determined by the TVMU's matching with the physical rhythm of the wave packet, causing the neuron to window and sample only during the critical phases where the energy of the damaging wave packet is concentrated, while automatically shielding during periods of irrelevant noise and dispersion tails. This achieves highly selective acquisition of effective signals and efficient suppression of interference.

[0080] Through this design This creates a dynamic window similar to a bandpass filter, which opens only in the critical physical phases where the wave packet energy is concentrated, while remaining closed in the redundant background noise range.

[0081] In the time domain, this window is not of a fixed width or position, but rather varies according to the dynamic phase. Within the [0,1) interval, the phase is activated only in a preset active phase range, while maintaining a minimal leakage constant value in other ranges. This active phase range is not arbitrarily chosen but is determined using a learnable phase offset parameter. and adaptive cycle The combined adjustment allows the window to be panned and scaled on the timeline, providing full self-adaptation capabilities.

[0082] When the damaged wave packet reaches the sensor, its energy is not uniformly distributed, but concentrated within a continuous period of the main lobe, with weaker energy and noise at the leading and trailing edges. The TVMU's dynamic window opens precisely in the phase interval corresponding to the main lobe of the wave packet, allowing the hidden state to fully absorb the input features at the current moment and capture key information related to the damage location, mode transition, and dispersion. In the pure noise segment before the leading edge of the wave packet and the trailing oscillation segment after the trailing edge, the dynamic window closes automatically, effectively shielding the interference of environmental noise, boundary reflection clutter, and dispersion tail waves, preventing these useless signal components from being written into the hidden state.

[0083] Compared to traditional fixed-bandwidth digital filters, the opening timing of this window automatically shifts with the arrival time of the wave packet, eliminating the need to predict the location of damage. The time span of the window automatically changes with the dynamic period and propagation distance. At long distances, the window widens to enclose the elongated wave packet, while at short distances, the window narrows to maintain sensitivity to transient changes. The shape of the window achieves bandpass characteristics through a double-slope function, with the central region fully open, the transition region rapidly attenuating, and both ends remaining constant low, thus achieving a correspondence with the shape of the energy envelope wave packet.

[0084] In S5, the method proposed in this embodiment retains the excellent gated computation kernel of GRU in generating candidate states, but introduces a time-varying modulation gate in the final hidden state update stage. Implement weighted control.

[0085] Understandably, candidate states Generated based on the reset and update gates of existing neural network models, specifically:

[0086] The reset gate determines how much past information needs to be forgotten when generating candidate states. It will... and After concatenation, the result is obtained through a fully connected layer and a sigmoid activation function:

[0087]

[0088] This is the weight matrix;

[0089] It is paranoia;

[0090] For the Sogmoid function, the output value is in between.

[0091] Then use the reset door pair Selective forgetting is performed, and then the forgotten historical information is compared with the current input. Combining these states, candidate states are generated using the Tanh activation function:

[0092]

[0093] Represents the weight matrix. It indicates paranoia.

[0094] This represents candidate memory content extracted based on the current input and historical context.

[0095] Define the candidate update information calculated by the standard GRU unit based on the current input and the previous hidden state. Then the final hidden layer state update formula of TVMU is:

[0096]

[0097] From the above formula, it can be seen that when the phase is in the active phase (i.e.) When it approaches 1), the model receives The latest extracted features correspond to waveform abrupt changes or energy peaks that occur when the signal passes through the damage location; when the phase is in a dormant period, i.e. When the value is close to 0, the model rejects new input; ,at this time Neurons are forced to retain historical memories. Through Synchronous change with the degree of wave packet stretching, time-varying modulation gate It acts as a physically-aware dispersion adaptive filter, enabling the network to realign the dispersed frequency components in the feature space, greatly reducing the interference of dispersion effects on deep learning feature extraction.

[0098] The training process of this neural network model adopts an end-to-end supervised learning approach: First, using a training dataset consisting of multi-channel guided wave time-domain signals and corresponding damage coordinate labels (or category labels), the original signals are processed through anomaly removal, high-pass filtering, normalization, and max pooling dimensionality reduction to obtain a sequence of length L; in each iteration, the sequence is fed into the TVMU layer step by step, and each TVMU unit is trained according to the current input. Hidden state from the previous moment First, utilize the learnable dynamic modulation factor , benchmark period The adaptive update cycle and dynamic phase are calculated using parameters such as phase shift. A time-varying modulation gate is generated through a dual-slope function, while the standard GRU reset gate is retained to generate candidate states. The time-varying modulation gate is then used to replace the update gate to perform hidden state updates, resulting in a hidden state sequence after the loop. The hidden states are then flattened and fused with features through a fully connected layer. Finally, based on the task output classification probability or regression coordinates, the classification cross-entropy loss or mean squared error loss between the predicted value and the true label is calculated. Through automatic differentiation and gradient backpropagation, the loss is differentiated with respect to all weight matrices, biases, and hyperparameters of the TVMU, and the parameters are updated using optimizers such as Adam. This process is repeated in multiple training rounds until the localization error on the validation set converges. The resulting model parameters are then solidified and can be directly used for damage localization inference of new signals.

[0099] In S8, the specific scheme for obtaining the damage category based on the hidden state sequence H is as follows:

[0100] The hidden state sequence H1 is input into the fully connected layer of the neural network model, and the initial probability of each damage category is output through the fully connected layer; the L2 norm of each hidden state in the hidden state sequence H1 is calculated to obtain the energy corresponding to each hidden state, thus obtaining the energy sequence E. ,in, They represent The energy; extract the time steps corresponding to the first q energies in the energy sequence E to obtain the set of active time peaks. Where q is an integer greater than or equal to 2 and less than or equal to n; based on the set of active time peaks The initial probabilities are corrected based on the wave packet arrival time range for each damage category, wherein the wave packet arrival time range is obtained through preset or pre-training. Based on the corrected initial probabilities, the damage category of the time-domain signal X is determined.

[0101] Among them, the set of active time peaks Given the wave packet arrival time range for each damage category, the initial probability, after correction, can be expressed as:

[0102] Through the set of active time peaks A consistency score is determined based on the wave packet arrival time range for each damage category, and this consistency score is correlated with the set of active time peaks. The number of time steps corresponding to the first q energies of the wave packet arriving in the time range is positively correlated;

[0103] The consistency score and the initial probability are substituted into the preset model, and the output of the preset model is the corrected initial probability;

[0104] The preset model is:

[0105]

[0106] To avoid zeroing out items;

[0107] Indicates temperature parameter;

[0108] This represents the corrected initial probability;

[0109] Indicates the initial probability;

[0110] An index representing the damage category;

[0111] This indicates the total number of damage categories;

[0112] Represents the set of active time peaks The temporal consistency score with the k-th damage category.

[0113]

[0114] First, the time-domain signal of the guided wave to be tested is input into the pre-trained TVMU network. After forward propagation, the hidden state sequence H1 is obtained. Then, H1 is passed through a fully connected layer and a softmax layer to obtain the initial probability of each damage category. Next, the L2 norm of the hidden state at each time step is calculated to obtain the energy sequence E. The energy values ​​are sorted from largest to smallest, and the time step indices corresponding to the top q energy peaks are extracted to form a set of active time peaks. Subsequently, for each damage category k, the temporal consistency score is calculated using the wave packet arrival time range pre-determined through training set statistics or a group velocity physics model. Finally, the initial probability... Consistency score Substitute into the preset model The corrected probability distribution is obtained, and the category corresponding to the highest probability is used as the final damage category determination result.

[0115] In this embodiment, the TVMU's unique time-varying gating mechanism ensures that the energy of the hidden state increases significantly only during the wave packet arrival period, while approaching zero in the noise region. Therefore, the active time peak extracted from H1 can very cleanly correspond to the arrival time of the damage wave packet. Damage at different spatial locations has different guided wave propagation path lengths, resulting in a definite distribution range for wave packet arrival times. If the initially predicted category of the fully connected layer is physically unreasonable (e.g., predicted as a long-distance damage, but the wave packet arrival time is very early), the time consistency score for that category will be very low; conversely, physically reasonable categories will have high scores. The correction formula is achieved by using... Recalibrating the initial probabilities for weights is equivalent to introducing a posterior probability adjustment based on physical priors: it enhances the class probability of time-matching, suppresses the class probability of time-conflicting, and adjusts the temperature parameter... The strength of physical constraints was controlled. This preserves the powerful ability of neural networks to extract deep features from complex waveforms, while also eliminating erroneous predictions that violate the laws of guided wave propagation through explicit time window checks, thus achieving collaborative decision-making between data-driven approaches and physical laws.

[0116] This correction method significantly improves the accuracy and robustness of damage classification. By directly utilizing the fundamental physical quantity of wave packet arrival time, it effectively corrects "anti-physics" misclassifications caused by waveform distortion, mode aliasing, or noise interference in fully connected layers—for example, misclassifying near-end damage as far-end damage. In practical applications, only a very lightweight energy peak extraction and consistency calculation are added to the original network output, increasing inference time almost entirely, and eliminating the need to retrain the backbone network, allowing for plug-and-play reduction of classification error rates. Simultaneously, this approach enhances model interpretability: by observing changes in consistency scores, it's possible to clearly determine whether each classification decision is supported by the physical time window, thus providing a basis for model diagnosis and result confidence assessment. Even when sensor degradation or environmental changes alter the signal morphology, the physical time constraint remains stable as an invariant, giving the overall system stronger generalization ability and engineering applicability.

[0117] In this embodiment, the method for confirming the damage result can be:

[0118] The hidden state sequence H1 is input into the linear activation layer of the neural network model, and the initial coordinates are output through the linear activation layer. The L2 norm of each hidden state in the hidden state sequence H is calculated to obtain the energy corresponding to each hidden state, so as to obtain the energy sequence E. ,in, They represent The energy; the time step in which the maximum energy value in the energy sequence is located. Construct a peak feature vector centered at the core, wherein the time step index of the peak feature vector is... The peak feature vector is represented as Each hidden state in the peak feature vector is input into a pre-trained decoder. The decoder outputs the correction amount of the peak feature vector for each time step relative to the initial coordinates, as well as the confidence score. Based on the correction amount and confidence score corresponding to all time steps in the feature vector, the initial coordinates are corrected to obtain the corrected initial coordinates, thereby obtaining the damage result.

[0119] In this embodiment, the initial damage coordinates are directly output through regression using a linear activation layer. Then, the L2 norm of the hidden state at each time step is calculated. The energy sequence was obtained. And determine the time step corresponding to the maximum value in the sequence. .by Centered on, take from both the front and back Construct peak feature vectors at each time step. Its time step index set is .Will The hidden state at each time step is fed into a pre-trained lightweight decoder, which outputs three values ​​in parallel: the correction amount of the initial coordinates at the current time step. and the corrected confidence score The weights are obtained by performing Softmax normalization on the confidence scores at all time steps. Then, the corrected coordinates are obtained by weighted averaging.

[0120] This serves as the final result of damage localization.

[0121] The core of this scheme lies in treating the hidden states at multiple time steps near the energy peak of the wave packet as multiple independent snapshots of the same damage event, and achieving multi-view fusion and refinement through a self-learning confidence mechanism. The gating mechanism of the TVMU ensures that the energy of the hidden states only increases significantly during the arrival time of the damage wave packet, while tending to be calm in the noise region. Therefore, the time of the energy peak and its neighborhood contain the richest and most complete damage information of the wave packet. However, the contributions of different phases of the wave packet (leading edge, main lobe, trailing edge) to localization are not entirely the same: the leading edge is more sensitive to the arrival time, the main lobe is more sensitive to energy and modal features, and the trailing edge may carry dispersion stretching information. Through a pre-trained decoder, the network can automatically learn how to extract localization correction cues from the hidden states of each specific phase, and simultaneously output a confidence score to evaluate the reliability of the cues. The confidence-weighted fusion mechanism ensures that high-reliability time steps dominate the correction direction, while low-reliability or damage-irrelevant time steps are automatically downweighted, thus achieving precise coordinate calibration purely based on the implicit knowledge learned by the network without relying on external physical constants.

[0122] In this embodiment, since the correction only applies to a few time steps near the energy peak (typically only about ten steps), the decoder structure is extremely lightweight, with negligible computational overhead during inference. It can also be directly added to the trained backbone network as a plug-and-play module without modifying the original model parameters. The introduction of confidence scores not only improves the robustness of the fusion but also provides a natural uncertainty estimate for each localization result: by visualizing the weights of each time step, it is possible to intuitively determine which part of the wave packet the network relies on most, enhancing the model's interpretability and fault diagnosis capabilities. Compared to post-processing schemes that rely on physical constants such as wave velocity, this method is entirely data-driven, automatically adapting to different structures or operating conditions without additional calibration, and possesses stronger generalization performance and ease of engineering deployment.

[0123] The lightweight decoder employs a "backbone freezing, independent training" strategy to avoid compromising the learned TVMU feature extraction capabilities. First, the pre-trained TVMU localization network is used to propagate forward through each guided wave signal in the training set, extracting the initial coordinates of its linear activation layer output and the hidden state sequence. A peak feature vector centered on the energy peak is then constructed as the input sample for the decoder, following the previously described method. The decoder itself is a small fully connected network containing one or two hidden layers, with its terminals outputting three values ​​in parallel: coordinate correction and confidence score. During training, each hidden state is fed into the decoder individually to obtain the corresponding absolute coordinate prediction and confidence score. Softmax normalization is then used to obtain the weights, and a weighted sum is calculated to obtain the final refined coordinates of the sample. The loss function consists of three parts: the main loss is the mean squared error between the predicted coordinates and the ground truth labels, driving the refined coordinates to converge towards the true location; the spatial consistency regularization loss calculates the variance of the absolute coordinates at each time step, forcing the localization results of different phase snapshots to approach each other, improving fusion stability; and the anchoring loss constrains the deviation between the final coordinates and the initial coordinates, preventing over-correction that could lead to drift. The decoder requires only a few training epochs to converge, and its parameters are fixed after training, allowing it to be deployed as a plug-and-play post-processing module alongside the backbone network.

[0124] This embodiment designs different TVMU network structures for ultrasonic guided wave damage identification, classification and localization, as shown in Tables 1 and 2, which respectively represent TVMU network structures applied to different tasks.

[0125] Table 1. Specific structure of the TVMU network for the damage classification task.

[0126]

[0127] Table 2. Specific structure of TVMU network for damage localization task.

[0128]

[0129] Analysis of experimental results:

[0130] In the experiment, a total of five different excitation frequencies were used, namely: This study verifies the TVMU network's ability to capture and learn damage features from signals with different center frequencies under the same dispersion effect. Detailed accuracy figures are shown in Table 3. The results in Table 3 demonstrate that the TVMU network exhibits strong feature perception and extraction capabilities under the same dispersion effect, even when faced with signals with different feature distributions, achieving an accuracy of approximately 97%. This is because the TVMU uses a time-varying gating modulation mechanism to dynamically modulate the neuron update cycle to address the impact of dispersion on the signal waveform.

[0131] Table 3. Accuracy results of damage classification under different feature distributions.

[0132]

[0133] Table 4 shows the damage localization accuracy of different network models in the original baseline group, the mild dispersion group, and the severe dispersion group.

[0134] Table 4 Experimental results of various networks in the damage localization task.

[0135]

[0136] From the longitudinal comparison in Table 4, except for CNN and FCN, the other models all achieved high localization accuracy in the original benchmark group, but with the increase of the dispersion intensity coefficient... With increasing dispersion, the performance of each model showed varying degrees of decline. While CNN and FCN had relatively short training times, their lesion localization accuracy was low and could not meet the requirements of the lesion localization task. Although GRU had the ability to process time series data, its accuracy dropped to 80.07% under severe dispersion. While Phased LSTM showed good lesion localization accuracy in the original baseline and mild dispersion groups, its accuracy drop under severe dispersion was significantly greater than that of TVMU. DCSCNet, using the fusion characteristics of the attention mechanism, achieved good prediction results, but its training time was too long, and it could not effectively adapt to packet changes under severe dispersion conditions. Compared to all other models, TVMU showed the best stability and robustness across all test groups, with a relatively moderate training time. Faced with packet stretching due to dispersion, TVMU did not passively accept features but actively adjusted the neuron update cycle through a time-varying modulation gating mechanism. This allowed the network to slow down the update frequency to synchronize low-frequency trailing signals or speed up the update frequency to capture high-frequency signals. This mechanism achieves physical alignment between signals and features during the feature extraction stage, eliminating the impact of waveform distortion caused by dispersion on the feature extraction process of the neural network.

[0137] This application provides an electronic device that may include a processing module and a memory. The memory stores a computer program, which, when executed by the processor, enables the electronic device to perform the corresponding steps in the aforementioned method for determining ultrasonic guided wave damage based on a neural network model.

[0138] In this embodiment, the processor can be an integrated circuit chip with signal processing capabilities. For example, the processor can be a central processing unit (CPU), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application.

[0139] The memory can be, but is not limited to, random access memory, read-only memory, programmable read-only memory, erasable programmable read-only memory, electrically erasable programmable read-only memory, etc.

[0140] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the electronic device described above can be referred to the corresponding steps in the aforementioned method, and will not be elaborated further here.

[0141] This application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program that, when run on a computer, causes the computer to perform the ultrasonic guided wave damage determination method based on a neural network model as described in the above embodiments.

[0142] Computer-readable storage media may be magnetic disks, optical disks, read-only memory, random access memory, flash memory, USB flash drives, hard disks, or solid-state drives, etc., and may also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implement the methods shown in the above embodiments.

[0143] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the aforementioned method for determining ultrasonic guided wave damage based on a neural network model. The computer program product may exist in a computer-readable storage medium in forms including, but not limited to, source files, executable files, and installation package files.

[0144] Based on the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by hardware or by using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application can be embodied in the form of a software product. This software product can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause a computer device (such as a personal computer, electronic device, or network device, etc.) to execute the methods described in the various implementation scenarios of this application.

[0145] In the embodiments provided in this application, it should be understood that the disclosed methods can also be implemented in other ways. The method embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, program segment, or part of code, which includes one or more executable instructions for implementing a specified logical function. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions. Furthermore, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0146] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for determining ultrasonic guided wave damage based on a neural network model, characterized in that, The method includes: S1: Obtain the time-domain signal X of the ultrasonic waveguide, where, n represents the number of sampling points; S2: Based on Time step The neuron update cycle is determined based on a preset periodic function. The periodic function is obtained through and time step The product of these terms yields the update period. ,in This represents the adjustment factor, which is a hyperparameter of the neural network model. ; S3: According to the update cycle Determine the time step Relative to update cycle dynamic phase ; S4: The dynamic phase The input is fed into a gating function, which is based on the dynamic phase. Compared with the preset ratio range Output gate value To enable neurons to be in dynamic phase Located in the ratio range It is in an active state at the time; S5: According to and time step Hidden state Based on the reset and update gates of the neural network model, candidate states are obtained. ; S6: Based on the gate value Candidate status and time step Hidden state Determine the time step Hidden state ,in, All are 0; S7: Repeat S2-S6 until the hidden state sequence H1 corresponding to the time-domain signal X is obtained. ; S8: Determine the damage result based on the hidden state sequence H1, wherein the damage result includes damage location coordinates and / or damage category; The periodic function is: Indicates the baseline update period for network initialization; It is a constant that controls the scaling factor; The gate function is: Indicates the leakage scaling constant; Determining the damage result based on the hidden state sequence H1 includes: The hidden state sequence H1 is input into the fully connected layer of the neural network model, and the initial probability of each damage category is output through the fully connected layer; Calculate the L2 norm of each hidden state in the hidden state sequence H1 to obtain the energy corresponding to each hidden state, thus obtaining the energy sequence E. ,in, They represent Energy; Extract the time steps corresponding to the first q energies from the energy sequence E to obtain the set of active time peaks. , where q is an integer greater than or equal to 2 and less than or equal to n; Based on the set of active time peaks The initial probability is corrected by the wave packet arrival time range for each damage category, and the wave packet arrival time range is obtained through preset or pre-training. Based on the corrected initial probability, the damage category of the time-domain signal X is determined.

2. The method according to claim 1, characterized in that, The set of active time peaks The initial probability is corrected by the wave packet arrival time range for each damage category, including: Through the set of active time peaks A consistency score is determined based on the wave packet arrival time range for each damage category, and this consistency score is correlated with the set of active time peaks. The number of time steps corresponding to the first q energies of the wave packet arriving in the time range is positively correlated; The consistency score and the initial probability are substituted into the preset model, and the output of the preset model is the corrected initial probability; The preset model is: To avoid zeroing out items; Indicates temperature parameter; This represents the corrected initial probability; Indicates the initial probability; An index representing the damage category; This indicates the total number of damage categories; Represents the set of active time peaks Consistency score with the k-th damage category.

3. The method according to claim 1, characterized in that, Determining the damage result based on the hidden state sequence H1 includes: The hidden state sequence H1 is input into the linear activation layer of the neural network model, and the initial coordinates are output through the linear activation layer; Calculate the L2 norm of each hidden state in the hidden state sequence H to obtain the energy corresponding to each hidden state, and then obtain the energy-based sequence E. ,in, They represent Energy; The time step in which the maximum energy value in the energy sequence is located. Construct a peak feature vector centered at the core, wherein the time step index of the peak feature vector is... The peak feature vector is represented as ; Each hidden state in the peak feature vector is input into a pre-trained decoder, which outputs the correction amount of the peak feature vector for each time step relative to the initial coordinates, as well as the confidence score. Based on the correction amount and confidence score corresponding to all time steps in the feature vector, the initial coordinates are corrected to obtain the corrected initial coordinates, thereby obtaining the damage result.

4. An electronic device, characterized in that, The electronic device includes a processor and a memory coupled together, the memory storing a computer program that, when executed by the processor, causes the electronic device to perform the method as described in any one of claims 1 to 3.

5. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when run on a computer, causes the computer to perform the method as described in any one of claims 1 to 3.

6. A program product, characterized in that, It includes a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 3.

Images