Laser shock signal timing prediction method, computer readable storage medium and electronic device

Abstract

This invention discloses a method for time-series prediction of laser shock signals, comprising the following steps: Step S0, constructing and training a time-series prediction model; Step S1, performing time-series decomposition and baseline prediction; Step S2, completing multi-dimensional feature extraction and micro / nano-level numerical protection; Step S3, completing feature-aware fusion gating generation; Step S4, performing selective execution and dynamic computational routing; Step S5, completing prediction fusion and inverse reconstruction, outputting the final time-series prediction result of the laser shock signal. The corresponding computer-readable storage medium and electronic device are also disclosed. This invention accurately quantifies the predictability confidence of extreme physical signals through multi-dimensional orthogonal features, dynamically determines the routing execution strategy of the compensation prediction branch, and strongly embeds an anti-collapse numerical protection mechanism and a pseudo-inverse solution strategy in the underlying mathematical operations. While safeguarding the stability of the baseline prediction, it achieves high-precision adaptive prediction of laser shock signals.

Description

Technical Field

[0001] This invention relates to the field of laser shock application technology, and specifically to a method for predicting the timing of laser shock signals. Background Technology

[0002] Laser shock blasting technology can be applied to product processing and strengthening, as well as non-destructive testing of equipment. This technology typically utilizes a high-energy pulsed laser to act on the surface of a material, inducing transient plasma expansion under the combined action of an absorption layer and a confinement layer. This generates high-amplitude stress waves on and near the surface of the target material, which then propagate into the material's interior. Based on this mechanism, laser shock blasting technology has been widely used in aerospace, energy and power, rail transportation, precision mold making, and high-end equipment manufacturing to improve the surface integrity of components, introduce residual compressive stress, and enhance fatigue resistance and stress corrosion resistance. Simultaneously, by analyzing the propagation, reflection, and attenuation characteristics of stress waves within the structure, non-destructive testing and health diagnosis of material conditions, structural integrity, and potential damage can be achieved.

[0003] In laser shock peening (LSP) processes, the stress wave timing signal generated by the target material is closely related to the evolution of shock pressure, the redistribution of internal stress, and the formation of the final residual stress layer. Therefore, it serves as a crucial basis for evaluating the strengthening effect, optimizing process parameters, and implementing closed-loop control. In nondestructive testing (NDT) or condition diagnostics scenarios, the stress wave timing signal carries a variety of information, including material microstructure, geometric boundaries, interface bonding states, internal defects, and service damage, making it an important carrier for fault identification and performance evaluation. With the development of hardware such as high-speed oscilloscopes, photodetectors, piezoelectric sensors, and high-speed data acquisition cards, the sampling rate of laser shock signals in related systems has generally reached the GHz level, typically reaching 10 GHz, with a corresponding sampling time interval as low as 0.1 ns. Consequently, the acquisition and processing of stress wave signals has evolved from traditional low-speed feature extraction to a problem of fine-grained timing analysis with ultra-high temporal resolution.

[0004] In existing technologies, the analysis and processing of laser shock signals typically employ methods such as time-domain peak extraction, envelope analysis, Fourier transform, wavelet analysis, short-time spectral analysis, empirical mode decomposition, AR-like linear prediction models, state-space models, and Kalman filtering. In recent years, some research has also attempted to use deep learning models such as convolutional neural networks, recurrent neural networks, long short-term memory networks, and Transformers to identify, fit, reconstruct, or extrapolate stress wave signals in a short time. While these methods can achieve signal feature extraction and modeling under certain conditions, they still have significant limitations in environments with ultra-high sampling rates, strong non-stationarity, strong noise, strong coupling propagation boundaries, and industrial online applications.

[0005] It is particularly important to note that although the laser emitter can control excitation parameters such as pulse triggering time, repetition frequency, nominal pulse energy, and pulse width, the stress wave timing signal ultimately acquired by the sensor is not equivalent to the laser emission control signal itself, nor is it solely determined by the laser emission frequency. The laser "emission frequency" in engineering typically reflects the repetition cycle or trigger period of the laser pulse, determining when the excitation is applied. The stress wave actually formed inside the target and propagating to the measurement point, however, represents the dynamic response of the entire system: excitation—material / structure—propagation path—sensor. This response is influenced not only by the laser source parameters but also by a combination of factors including laser energy jitter, spot size and overlap rate, absorption layer state, constraint layer thickness, material inhomogeneity, geometry, boundary conditions, initial residual stress, reflection / refraction / scattering during propagation, measurement point location, sensor mounting and coupling method, acquisition link bandwidth, and environmental electromagnetic noise. Therefore, even if the parameters of the laser emitter remain unchanged, the measured stress wave waveforms obtained under different workpieces, different positions, different batches of materials, and different working conditions may still have significant differences in peak value, arrival delay, attenuation envelope, dominant frequency, and transient disturbance characteristics.

[0006] In other words, the object of prediction in this field is not whether or when a laser will emit, but rather, under the condition of only obtaining a limited observation window, predicting the actual stress wave response evolution of the target material in a very short subsequent time. This type of prediction has clear engineering necessity: First, industrial closed-loop control often requires predicting subsequent peak values, energy distribution, or abnormal change trends before the complete waveform has been fully acquired, in order to compensate for the inherent time delays caused by sensing, transmission, computation, and execution, thereby achieving more timely process adjustments; Second, in scenarios with high temperature, high vibration, strong electromagnetic interference, or limited space, sensor deployment and acquisition conditions are often restricted, and the measured signal may have missing segments, dead zones, saturated segments, or local distortions, requiring time-series prediction to achieve waveform completion and anomaly warning; Third, in the non-destructive diagnosis of complex structures, it is often difficult to directly place sensors in key areas, requiring estimation of the stress wave state in subsequent time periods or unmeasurable areas based on the historical response of measurable locations. Therefore, predicting the stress wave time-series signal is a forward-looking estimation and compensation of the actual response, belonging to the modeling requirement of complex system outputs, rather than a repetitive calculation of the transmitting end control quantity.

[0007] However, under the aforementioned ultra-high spatiotemporal resolution conditions, existing technologies still face the following challenges in performing high-precision stress wave time series prediction.

[0008] First, laser shock signals typically exhibit significant multimodal mixing characteristics, making it difficult for a single prediction architecture to balance accuracy, robustness, and computational cost. Real laser shock stress wave signals are complex in morphology, containing both low-frequency mechanical resonances or intrinsic standing wave envelopes formed by internal material propagation and boundary reflections (e.g., regular low-frequency components around 6.5 MHz) and high-frequency broadband random transient components caused by surface roughness, microstructure inhomogeneity, scattering effects, and measurement noise. For the former, traditional linear models may be effective within a short time window, but error accumulation easily occurs when extrapolating over long time steps at a distance. For the latter, if a complex deep neural network is forcibly used for uniform fitting, the model is prone to severe overfitting to the random broadband noise components, leading to decreased prediction performance and additional consumption of computational resources at the industrial edge.

[0009] Secondly, weak bias signals in industrial settings can easily cause instability or even collapse of underlying numerical calculations. The raw physical electrical signals output by laser shock sensors typically have extremely weak bias characteristics, with amplitudes often ranging from... Fluctuations around the order of magnitude, often with a near-zero mean, mean variance before standardization may drop to [missing value]. Order of magnitude. When the input signal contains sensor acquisition dead zones, near-zero energy segments across the entire frequency band, or abnormally flat segments, existing models are prone to triggering division-zero overflow, numerical divergence, or singular matrix errors during low-level calculations such as variance ratio calculation, normalization processing, frequency domain spectral entropy probability logarithmic operations, correlation analysis, and physical basis least squares solutions. This can occur due to denominators approaching zero, invalid logarithmic terms, or matrix multicollinearity, leading to online monitoring system interruptions and failing to meet the requirements of continuous operation and high reliability in industrial applications.

[0010] Third, existing deep hybrid models generally lack physical prior fusion and dynamic computational routing mechanisms. Most existing pure data-driven models learn the mapping relationship between input and output in a black-box manner, failing to explicitly introduce objective physical priors such as the inherent mechanical frequency of materials, the reflection law of structural boundaries, and typical attenuation characteristics into the model base. At the same time, they also lack a mechanism to make conditional judgments based on the "predictability" or "complexity" of the signal within the current observation window. It is difficult to dynamically allocate the model depth and computational load required for subsequent predictions based on whether the current signal is dominated by regular resonant components or random broadband transient components. Therefore, it is difficult to simultaneously take into account prediction accuracy, numerical stability, and industrial edge deployment efficiency.

[0011] Therefore, while existing technologies can acquire and analyze stress wave signals to a certain extent, they still lack a technical solution for ultra-high sampling rate stress wave time series prediction in laser shock strengthening processes and high-end equipment non-destructive diagnostic scenarios. This solution should effectively integrate physical priors under complex multimodal signal conditions, intelligently allocate prediction computing power according to different signal characteristics, and provide industrial-grade numerical stability guarantees in the underlying computational path. Based on this, there is an urgent need in this field to propose new stress wave time series prediction methods to meet the comprehensive requirements of high precision, high real-time performance, and high reliability for online evaluation, anomaly early warning, and closed-loop control. Summary of the Invention

[0012] To address the shortcomings of the existing technologies, the technical problem to be solved by this invention is: how to provide a laser shock signal timing prediction method, computer-readable storage medium, and electronic device that can adaptively sense signal predictability and selectively route prediction calculations for the variable physical conditions of laser shock ultra-high frequency signals, while maintaining the security of underlying numerical values ​​under extremely weak signal conditions, thereby improving prediction accuracy and adaptability.

[0013] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0014] A method for time-series prediction of laser shock signals, characterized by comprising the following steps:

[0015] Step S0: Construct and train a time series prediction model: Construct a time series prediction model that includes a basic linear decomposition network and multiple independent compensation prediction expert branches, and train and initialize the time series prediction model using historical training data.

[0016] Step S1: Perform time series decomposition and baseline prediction: Obtain the collected laser shock time series data to be predicted and input it into the trained time series prediction model. Perform reversible instance normalization on the model. Use the physical decomposition kernel in the basic linear decomposition network that is adapted to the target material's physical resonance fundamental frequency period to perform low-pass filtering time series decomposition on the normalized time series, extract the basic trend component and the seasonal fluctuation component, and then predict and project the basic trend component and the seasonal fluctuation component respectively through dual-line linear mapping and sum them to obtain the baseline prediction result. The kernel length of the physical decomposition kernel is determined according to the target material's mechanical resonance fundamental frequency: the sampling point period corresponding to the fundamental frequency oscillation is obtained by rounding down the ratio of the sampling rate to the resonance fundamental frequency. Take no less than An odd number plus one is used as the kernel length to ensure that the moving average low-pass passivity of the physical decomposition kernel for the resonant fundamental frequency signal is reduced to below 0.1%, so that the resonant fundamental frequency signal can fully enter the seasonal fluctuation component.

[0017] Step S2, complete multi-dimensional feature extraction and micro / nano-level numerical protection: Based on the seasonal fluctuation component and basic trend component obtained from the decomposition in step S1, multi-dimensional predictability features are extracted. These multi-dimensional predictability features include at least: time-domain autocorrelation features, frequency-domain spectral entropy features, and energy signal-to-noise ratio features. In the underlying calculation paths involving variance calculation, total power spectrum energy calculation, and normalized power spectrum logarithmic operation involved in extracting the above multi-dimensional predictability features, a micro / nano-level lower bound soft truncation protection mechanism is embedded. This mechanism performs lower bound truncation processing on input values ​​or intermediate state values ​​during the calculation process, with a threshold value not less than a preset minimum value, to accommodate signal amplitude standard deviations within a certain range. The numerical accuracy requirements for extremely weak bias laser shock signals at or below the order of magnitude are to prevent numerical underflow or division by zero anomalies caused by sensor dead values ​​or extremely weak biases; wherein, the preset minimum threshold value is no greater than Positive numbers;

[0018] Step S3, complete feature-aware fusion gating generation: cross-fuse the multidimensional predictability features extracted in step S2 to generate activation control weights for controlling the multiple independent compensation prediction expert branches respectively.

[0019] Step S4, selective execution and dynamic computational power routing: During forward prediction inference, the maximum value of the activation control weight corresponding to each independent compensation prediction expert branch is independently compared with a preset dynamic bypass threshold, triggering a soft routing mechanism. The soft routing mechanism includes: if the maximum value of the activation control weight corresponding to an independent compensation prediction expert branch is lower than the dynamic bypass threshold, it is determined that the current signal is in a high-frequency broadband transient noise condition that cannot be effectively modeled for that branch, and the forward feature extraction and prediction calculation of that branch are actively skipped to prevent overfitting and save end-side computational power; if the maximum value of the activation control weight corresponding to an independent compensation prediction expert branch is not lower than the dynamic bypass threshold, the branch is activated to perform forward calculation to generate compensation prediction results; wherein, the dynamic bypass threshold ranges from [value missing]. to ;

[0020] Step S5, complete prediction fusion and reverse restoration: The baseline prediction result obtained in step S1 and the branch compensation prediction results after element-wise dynamic weighting by the corresponding activation control weights in step S4 are added and fused together, and the inverse normalization operation corresponding to the reversible instance normalization is performed to restore to the original physical scale, and the final laser shock signal timing prediction result is output.

[0021] In this way, the present invention accurately quantifies the predictability confidence of extreme physical signals through multidimensional orthogonal features, dynamically determines the routing execution strategy of the compensation prediction branch, and powerfully embeds an anti-collapse numerical protection mechanism and a pseudo-inverse solution strategy in the underlying mathematical operations. Under the premise of safeguarding the stability of the benchmark prediction, it achieves high-precision adaptive prediction of laser shock signals, which greatly improves the accuracy and adaptability of the prediction.

[0022] Furthermore, the specific steps for extracting multidimensional predictability features in step S2 include:

[0023] The ratio of the autocorrelation covariance of the seasonal fluctuation component at each prediction step to the variance after processing by the micro / nano-level lower bound soft truncation protection mechanism is calculated, and the time-domain autocorrelation feature is obtained through nonlinear mapping. The variances of the seasonal fluctuation component and the basic trend component after processing by the micro / nano-level lower bound soft truncation protection mechanism are calculated respectively, and the logarithm of the ratio of their variances is calculated. The energy signal-to-noise ratio feature is obtained through nonlinear mapping. The power spectrum of the seasonal fluctuation component is obtained by frequency domain transformation. After lower bound truncation of the total energy of the power spectrum based on the micro / nano-level lower bound soft truncation protection mechanism, the normalized power probability distribution is calculated. The logarithmic probability is obtained by applying the micro / nano-level lower bound soft truncation protection mechanism to the normalized power probability distribution again, and the Shannon information entropy is calculated. The frequency domain spectral entropy feature is obtained through inverse nonlinear mapping.

[0024] In this way, the linear autoregressive predictability of the signal at each prediction step is quantified using time-domain autocorrelation characteristics, reflecting the degree of continuation of historical time-series patterns at future steps; the energy signal-to-noise ratio (SNR) is used to quantify the energy advantage of periodic oscillation components relative to the global trend drift, reflecting the ability of effective signal components to overcome background noise; and the frequency-domain spectral entropy is used to quantify the concentration and regularity of the power spectrum energy distribution. Lower spectral entropy indicates that the energy is more concentrated in a few discrete frequencies and the signal structure is more regular. These three aspects characterize the predictability of the signal from three orthogonal dimensions: time-domain correlation, SNR, and frequency-domain structural complexity. Simultaneously, the micro / nano-level lower bound soft truncation protection mechanism sets independent lower bound truncation barriers in variance calculation, total power spectrum energy calculation, and normalized power probability distribution logarithm calculation, forming a progressively layered three-level numerical security defense line. Even when the signal amplitude standard deviation is within a certain range... Even in extremely weak scenarios with magnitudes or even lower, the entire feature extraction computation chain can be guaranteed to be free from numerical underflow or division by zero, thus providing reliable and physically meaningful feature inputs for subsequent gating decisions.

[0025] Furthermore, in step S3, the plurality of independent compensation prediction expert branches include at least a nonlinear residual expert branch and a periodic extrapolation expert branch, and the specific steps for generating the activation control weights of each branch include:

[0026] For the nonlinear residual expert branch: the time-domain autocorrelation feature, the energy signal-to-noise ratio feature, and the frequency-domain spectral entropy feature are multiplied and fused element-wise to generate nonlinear activation control weights;

[0027] For the periodic extrapolation expert branch: extract autocorrelation peak features at multiple sampling point periods corresponding to the physical resonance fundamental frequency of the target material and its various higher-order harmonics (such as the sampling point period corresponding to the 6.5MHz fundamental frequency and the sampling point period corresponding to higher-order harmonics such as 13MHz and 19.5MHz), multiply the autocorrelation peak features element-wise with the energy signal-to-noise ratio features, and then multiply by the square of the frequency domain spectral entropy features to generate periodic activation control weights.

[0028] In this way, element-wise multiplication fusion achieves logical AND-gated semantics: the product result will significantly deviate from zero only when all features involved in the fusion reach a high level simultaneously, and the corresponding expert branch will be effectively activated. Any feature approaching zero can pull the entire gating weight close to zero, thus achieving strict multi-condition joint admission judgment. For nonlinear residual expert branches, the activation condition requires the signal to simultaneously possess a significant time-domain autocorrelation structure, sufficient signal-to-noise energy advantage, and low frequency-domain complexity; all three are indispensable. This avoids forcibly activating the branch and causing overfitting when there is a lack of predictability support in any dimension. For the periodic extrapolation expert branch, by extracting the autocorrelation peak at the periodic delay of a specific number of sampling points corresponding to the target material's resonant fundamental frequency and its various harmonics, the gating criterion is narrowed from generalized time-domain correlation to the precise detection of periodic structures pointing to the physical resonant frequency. At the same time, the squaring of the frequency domain spectral entropy features is used in the fusion, making the requirements for frequency domain regularity more stringent for this branch. Only when the power spectrum energy is highly concentrated in a few discrete spectral lines and the spectral entropy is significantly low can the squared-amplified spectral entropy features be sufficient to make the product gating weights reach an effective activation level, thereby avoiding misjudging broadband random noise as an extrapolable periodic signal.

[0029] Furthermore, the periodic extrapolation expert branch adopts a gradient-free deterministic extrapolation design, specifically as follows:

[0030] In the model building stage of step S0, the physical fundamental frequency and its harmonics (such as 6.5MHz, 13MHz, 19.5MHz, etc.) of the specific mechanical resonance standing wave of the laser shock target are explicitly injected as the sampling point period as a deterministic period constant. Based on this, the Fourier input basis matrix and output basis matrix composed of sine and cosine functions are constructed. When solving the projection coefficient matrix corresponding to the Fourier input basis matrix using the least squares method, the Moore-Penrose Pseudo-inverse algorithm with singular value tolerance truncation parameter is used to replace the standard inverse matrix solution operation.

[0031] In this approach, the injected physical fundamental frequency and its corresponding Fourier sine and cosine basis vectors may exhibit an approximate linear correlation due to the rounding effect of the sampling point number. This leads to high condition number or singularity in the coefficient matrix (i.e., the Gram matrix of the basis matrix) of the least squares equation system. Using the standard inverse matrix for solving this problem will cause numerical overflow or drastically unstable calculation results. The Moore-Ponros generalized pseudo-inverse algorithm with singular value tolerance truncation parameters is adopted. Internally, it decomposes the coefficient matrix into a product of orthogonal basis and singular value diagonal matrices through singular value decomposition. Then, it sets small singular values ​​below the tolerance threshold to zero instead of taking their reciprocals. Therefore, it can eliminate the risk of matrix solution collapse caused by multicollinearity from the underlying mathematical mechanism, ensuring a stable and energy-minimum least squares solution under any data conditions. This better guarantees the numerical robustness and reliability of the periodic extrapolation expert branch in engineering deployment.

[0032] Furthermore, in the nonlinear residual expert branch, a patch embedding segmentation mechanism is used to segment the seasonal fluctuation component into multiple local time windows. The sliding step size between adjacent patches is set to no more than half the patch length to ensure that the overlap rate between adjacent patches is not less than 50% (e.g., when the patch length is 64 sampling points, the sliding step size is set to 32 sampling points, and the overlap rate is 50%). The patch embedding segmentation mechanism is an existing technique for segmenting input time series data into fixed or adaptive length segments and mapping them to a high-dimensional embedding space. After embedding and mapping, the multiple local time windows are fed as token sequences into a multi-head self-attention network to capture local nonlinear features, including high-frequency transient spikes and gradient abrupt changes. The linear mapping layer at the output of the nonlinear residual expert branch adopts an all-zero initialization strategy during model initialization in step S0.

[0033] In this way, adjacent patches maintain an overlap rate of at least 50%, ensuring that each sampling time is covered by at least two adjacent patches. This avoids the information breaks and feature discontinuities that occur at patch boundaries in non-overlapping segmentation, which is beneficial for multi-head self-attention networks to establish smooth temporal contextual relationships between patches, thereby more accurately capturing local nonlinear features such as high-frequency transient spikes and gradient abrupt changes. Simultaneously, the linear mapping layer at the output of the nonlinear residual expert branch adopts a zero-initialization strategy, ensuring that all weight parameters of this layer are zero at the start of training. Therefore, the output of this expert branch is always a zero vector during forward propagation, without any additive perturbation to the baseline prediction result. This ensures that the expert output tends to zero in the early stages of training, eliminating interference with the convergence of the linear decomposition baseline. As training progresses, the weights of this layer gradually learn meaningful non-zero mappings from zero values, and the compensation contribution of the expert branch smoothly and gradually increases.

[0034] Furthermore, in step S0, a two-stage dynamic scrambling reduction joint training strategy is used to jointly train the time series prediction model, specifically including:

[0035] First training phase (benchmark warm-up): Disable the forward computation of all compensated prediction expert branches and their corresponding feature-aware fusion gating computation, and use only historical training data to independently fit and train the basic linear decomposition network that generates the benchmark prediction results until convergence and stable basic network weights are obtained.

[0036] The second training phase (joint optimization) involves reactivating all compensated prediction expert branches and their corresponding feature-aware fusion gating computations, and jointly training them with the basic linear decomposition network. During joint training, the learning rate of the basic linear decomposition network is reduced to less than one-tenth of the learning rate of the compensated prediction expert branches to protect the converged weights of the basic network from significant disturbance. Simultaneously, an auxiliary regularization penalty loss term is introduced into the global total loss function. This penalty loss term includes independent penalty terms for each compensated prediction expert branch. Each independent penalty term is constructed by multiplying the difference between the value 1 and the activation control weight corresponding to that branch by the square of the predicted energy of that branch's output, thus guiding each expert branch to suppress its output when the gating activation is low.

[0037] In this way, the first training phase disables all compensation expert branches and gating computations, allowing the basic linear decomposition network to independently optimize to convergence in a pure gradient environment without any compensation interference. This achieves stable and reliable baseline prediction capabilities and establishes a safe baseline for global prediction. In the second training phase, when all expert branches are reactivated for joint training, the learning rate of the basic linear decomposition network is reduced to less than one-tenth of the learning rate of the compensation expert branches. This limits the update magnitude of the backpropagation gradient on the converged basic weights during the joint optimization phase, preventing the gradient signals of the expert branches from causing catastrophic disturbances to the baseline prediction capability. Simultaneously, in the auxiliary regularization penalty loss term, for a given expert branch, when its corresponding activation control weight is close to zero, the difference between the value 1 and the weight is close to 1, resulting in the maximum penalty coefficient, forcing the predicted output energy of that branch to converge to zero. Conversely, when the activation control weight is close to 1, the difference is close to zero, the penalty coefficient tends to disappear, and the branch is allowed to freely output compensated predictions. Therefore, this penalty mechanism explicitly forces the predicted outputs of channels that are not highly activated by gating to converge to zero, decoupling gradient conflicts between branches and ensuring the baseline safety of the model in a weakly regularized environment.

[0038] The present invention also discloses a computer-readable storage medium storing a computer program that, when executed by a processor, implements the laser shock signal timing prediction method as described above.

[0039] The present invention also discloses an electronic device, including a processor and a memory; the memory stores a computer program, which, when run by the processor, causes the processor to perform the laser shock signal timing prediction method as described above.

[0040] Compared with existing technologies, the present invention has the following significant advantages in industrial application and technical effects:

[0041] 1) Multi-dimensional feature perception and adaptive gating enhance prediction stability under complex conditions. The multi-dimensional evidence fusion gating of this invention can effectively distinguish between low-frequency mechanical resonance and high-frequency transient shot noise. For broadband noise signals with poor regularity and prone to overfitting (such as in Experiment Example 2 described later), the features extracted by the system (high spectral entropy, low autocorrelation) cause the gating weights to actively approach zero, thereby bypassing the expert path. This avoids overfitting of complex models to invalid noise and, through the residual compensation mechanism, reduces the average prediction error by 37.95%.

[0042] 2) Integrating physical priors significantly improves the extrapolation prediction accuracy for long sequences. For signal segments containing significant mechanical resonances (such as in Example 1), the gating mechanism can effectively identify the periodicity of the acoustic standing waves of the target material and assign higher weights to the physical prior periodic extrapolation expert branch (using a deterministic harmonic basis such as 6.5MHz). This mechanism effectively suppresses the error divergence trend of the purely linear model at long prediction step sizes. Experiments show that compared with the classical linear benchmark, the prediction accuracy (MSE index) of this invention is improved by an average of 53.75%, with a local maximum improvement of 91.35%.

[0043] 3) Dynamic computing power allocation on demand reduces computational overhead for edge devices. This invention introduces selective execution and dynamic routing mechanisms. Through pre-gated gating computation, when the prediction confidence is lower than a set threshold, the system automatically skips computationally intensive attention expert networks and extrapolation matrix operations. This strategy of allocating computing power on demand effectively reduces inference latency and floating-point computation load during online deployment at the edge.

[0044] 4) Optimization of underlying numerical calculations enhances system robustness. This is geared towards practical data acquisition. For signals with bias magnitudes of minute size, this invention introduces numerical truncation protection (Clamp) in the underlying variance, logarithm, and division operators, and uses pseudo-inverse (pinv) instead of standard matrix inversion. These designs, from an algorithmic logic perspective, avoid computational interruptions caused by abnormal sensor values, zero energy input, or near-collinearity of matrices, ensuring stable operation of the system in complex industrial application scenarios.

[0045] In summary, this method performs reversible instance normalization and time series decomposition based on physical resonance priors on ultra-high frequency laser shock time series to obtain baseline predictions. It extracts time-domain correlation, frequency-domain spectral entropy, and energy comparison features, and introduces lower-bound soft truncation protection and pseudo-inverse solving during variance, power spectrum, logarithm, and matrix solving processes. Based on these features, it generates expert branch activation weights, dynamically bypasses or executes periodic extrapolation experts, and nonlinear residual experts. The compensated predictions are fused with the baseline predictions and inversely normalized to output the final result. This method is suitable for the prediction and online monitoring of laser shock signals. Attached Figure Description

[0046] Figure 1 This is a flowchart of the method of the present invention;

[0047] Figure 2 In the experimental verification of this invention, a thermal comparison chart of the improvement (%) of the MSE (mean square error) of the method of this invention (ZetaTime) compared with the linear baseline model (LinearBaseline) under multiple prediction lengths for low-frequency resonant dominant signal (Type 1) and high-frequency transient dominant signal (Type 2) is presented.

[0048] Figure 3 This is a bar chart comparing the MSE prediction accuracy of the Type 1 signal at prediction lengths of 128 points (13ns), 256 points (26ns), and 512 points (51ns) in Experiment Example 1 of this invention.

[0049] Figure 4 This is a bar chart comparing the MSE prediction accuracy of the Type 2 signal across different prediction length spans in Experiment Example 2 of this invention.

[0050] Figure 5 This is a comparison of the MSE error accumulation trend curves for the low-frequency mechanical resonance dominant signal (Type 1) in Experiment Example 1, along with the prediction physics step size (ns).

[0051] Figure 6 This is a comparison of the cumulative MSE error trend curves for the high-frequency transient dominant signal (Type 2) in Experiment Example 2, along with the prediction physics step size (ns). Detailed Implementation

[0052] The present invention will be further described in detail below with reference to specific embodiments and experimental examples.

[0053] A time-series prediction method for laser shock signals based on multi-dimensional feature perception and selective execution is characterized by including the following steps (see...). Figure 1 ):

[0054] Step S0: Construct and train a time series prediction model: Construct a time series prediction model that includes a basic linear decomposition network and multiple independent compensation prediction expert branches, and train and initialize the time series prediction model using historical training data.

[0055] Step S1: Perform time series decomposition and baseline prediction: Obtain the collected laser shock time series data to be predicted and input it into the trained time series prediction model. Perform reversible instance normalization on the model. Use the physical decomposition kernel in the basic linear decomposition network that is adapted to the target material's physical resonance fundamental frequency period to perform low-pass filtering time series decomposition on the normalized time series, extract the basic trend component and the seasonal fluctuation component, and then predict and project the basic trend component and the seasonal fluctuation component respectively through dual-line linear mapping and sum them to obtain the baseline prediction result. The kernel length of the physical decomposition kernel is determined according to the target material's mechanical resonance fundamental frequency: the sampling point period corresponding to the fundamental frequency oscillation is obtained by rounding down the ratio of the sampling rate to the resonance fundamental frequency. Take no less than An odd number plus one is used as the kernel length to ensure that the moving average low-pass passivity of the physical decomposition kernel for the resonant fundamental frequency signal is reduced to below 0.1%, so that the resonant fundamental frequency signal can fully enter the seasonal fluctuation component.

[0056] Step S2, complete multi-dimensional feature extraction and micro / nano-level numerical protection: Based on the seasonal fluctuation component and basic trend component obtained from the decomposition in step S1, multi-dimensional predictability features are extracted. These multi-dimensional predictability features include at least: time-domain autocorrelation features, frequency-domain spectral entropy features, and energy signal-to-noise ratio features. In the underlying calculation paths involving variance calculation, total power spectrum energy calculation, and normalized power spectrum logarithmic operation involved in extracting the above multi-dimensional predictability features, a micro / nano-level lower bound soft truncation protection mechanism is embedded. This mechanism performs lower bound truncation processing on input values ​​or intermediate state values ​​during the calculation process, with a threshold value not less than a preset minimum value, to accommodate signal amplitude standard deviations within a certain range. The numerical accuracy requirements for extremely weak bias laser shock signals at or below the order of magnitude are to prevent numerical underflow or division by zero anomalies caused by sensor dead values ​​or extremely weak biases; wherein, the preset minimum threshold value is no greater than Positive numbers;

[0057] Step S3, complete feature-aware fusion gating generation: cross-fuse the multidimensional predictability features extracted in step S2 to generate activation control weights for controlling multiple independent compensation prediction expert branches.

[0058] Step S4, selective execution and dynamic computational power routing: During forward prediction inference, the maximum value of the activation control weight corresponding to each independent compensation prediction expert branch is independently compared with a preset dynamic bypass threshold, triggering a soft routing mechanism. The soft routing mechanism includes: if the maximum value of the activation control weight corresponding to an independent compensation prediction expert branch is lower than the dynamic bypass threshold, it is determined that the current signal is in a high-frequency broadband transient noise condition that cannot be effectively modeled for that branch, and the forward feature extraction and prediction calculation of that branch are actively skipped to prevent overfitting and save end-side computational power; if the maximum value of the activation control weight corresponding to an independent compensation prediction expert branch is not lower than the dynamic bypass threshold, the branch is activated to perform forward calculation to generate compensation prediction results; wherein, the dynamic bypass threshold ranges from [value missing]. to ;

[0059] Step S5, complete prediction fusion and reverse restoration: The baseline prediction result obtained in step S1 and the branch compensation prediction results after element-wise dynamic weighting by the corresponding activation control weights in step S4 are added and fused together, and the inverse normalization operation corresponding to the reversible instance normalization is performed to restore to the original physical scale, and the final laser shock signal timing prediction result is output.

[0060] In this way, the present invention accurately quantifies the predictability confidence of extreme physical signals through multidimensional orthogonal features, dynamically determines the routing execution strategy of the compensation prediction branch, and powerfully embeds an anti-collapse numerical protection mechanism and a pseudo-inverse solution strategy in the underlying mathematical operations. Under the premise of safeguarding the stability of the benchmark prediction, it achieves high-precision adaptive prediction of laser shock signals, which greatly improves the accuracy and adaptability of the prediction.

[0061] The specific steps for extracting multidimensional predictability features in step S2 include:

[0062] The ratio of the autocorrelation covariance of the seasonal fluctuation component at each prediction step to the variance after processing by the micro / nano-level lower bound soft truncation protection mechanism is calculated, and the time-domain autocorrelation feature is obtained through nonlinear mapping. The variances of the seasonal fluctuation component and the basic trend component after processing by the micro / nano-level lower bound soft truncation protection mechanism are calculated respectively, and the logarithm of the ratio of their variances is calculated. The energy signal-to-noise ratio feature is obtained through nonlinear mapping. The power spectrum of the seasonal fluctuation component is obtained by frequency domain transformation. After lower bound truncation of the total energy of the power spectrum based on the micro / nano-level lower bound soft truncation protection mechanism, the normalized power probability distribution is calculated. The logarithmic probability is obtained by applying the micro / nano-level lower bound soft truncation protection mechanism to the normalized power probability distribution again, and the Shannon information entropy is calculated. The frequency domain spectral entropy feature is obtained through inverse nonlinear mapping.

[0063] In this way, the linear autoregressive predictability of the signal at each prediction step is quantified using time-domain autocorrelation characteristics, reflecting the degree of continuation of historical time-series patterns at future steps; the energy signal-to-noise ratio (SNR) is used to quantify the energy advantage of periodic oscillation components relative to the global trend drift, reflecting the ability of effective signal components to overcome background noise; and the frequency-domain spectral entropy is used to quantify the concentration and regularity of the power spectrum energy distribution. Lower spectral entropy indicates that the energy is more concentrated in a few discrete frequencies and the signal structure is more regular. These three aspects characterize the predictability of the signal from three orthogonal dimensions: time-domain correlation, SNR, and frequency-domain structural complexity. Simultaneously, the micro / nano-level lower bound soft truncation protection mechanism sets independent lower bound truncation barriers in variance calculation, total power spectrum energy calculation, and normalized power probability distribution logarithm calculation, forming a progressively layered three-level numerical security defense line. Even when the signal amplitude standard deviation is within a certain range... Even in extremely weak scenarios with magnitudes or even lower, the entire feature extraction computation chain can be guaranteed to be free from numerical underflow or division by zero, thus providing reliable and physically meaningful feature inputs for subsequent gating decisions.

[0064] In step S3, the plurality of independent compensation prediction expert branches include at least a nonlinear residual expert branch and a periodic extrapolation expert branch. The specific steps for generating the activation control weights for each branch include:

[0065] For the nonlinear residual expert branch: the time-domain autocorrelation feature, the energy signal-to-noise ratio feature, and the frequency-domain spectral entropy feature are multiplied and fused element-wise to generate nonlinear activation control weights;

[0066] For the periodic extrapolation expert branch: extract autocorrelation peak features at multiple sampling point periods corresponding to the physical resonance fundamental frequency of the target material and its various higher-order harmonics (such as the sampling point period corresponding to the 6.5MHz fundamental frequency and the sampling point period corresponding to higher-order harmonics such as 13MHz and 19.5MHz), multiply the autocorrelation peak features element-wise with the energy signal-to-noise ratio features, and then multiply by the square of the frequency domain spectral entropy features to generate periodic activation control weights.

[0067] In this way, element-wise multiplication fusion achieves logical AND-gated semantics: the product result will significantly deviate from zero only when all features involved in the fusion reach a high level simultaneously, and the corresponding expert branch will be effectively activated. Any feature approaching zero can pull the entire gating weight close to zero, thus achieving strict multi-condition joint admission judgment. For nonlinear residual expert branches, the activation condition requires the signal to simultaneously possess a significant time-domain autocorrelation structure, sufficient signal-to-noise energy advantage, and low frequency-domain complexity; all three are indispensable. This avoids forcibly activating the branch when there is a lack of predictability support in any dimension, which could lead to overfitting. For the periodic extrapolation expert branch, by extracting the autocorrelation peak at the periodic delay of a specific number of sampling points corresponding to the target material's resonant fundamental frequency and its various harmonics, the gating criterion is narrowed from generalized time-domain correlation to the precise detection of periodic structures pointing to the physical resonant frequency. At the same time, the squaring of the frequency domain spectral entropy features is used in the fusion, making the requirements for frequency domain regularity more stringent for this branch. Only when the power spectrum energy is highly concentrated in a few discrete spectral lines and the spectral entropy is significantly low can the squared-amplified spectral entropy features be sufficient to make the product gating weights reach an effective activation level, thereby avoiding misjudging broadband random noise as an extrapolable periodic signal.

[0068] The periodic extrapolation expert branch employs a gradient-free deterministic extrapolation design, specifically as follows:

[0069] In the model building phase, the physical fundamental frequency and its harmonics (such as 6.5MHz, 13MHz, 19.5MHz, etc.) of the specific mechanical resonance standing wave of the laser shock target are explicitly injected as deterministic periodic constants. Based on this, the Fourier input basis matrix and output basis matrix composed of sine and cosine functions are constructed. When solving the projection coefficient matrix corresponding to the Fourier input basis matrix using the least squares method, the Moore-Penrose Pseudo-inverse algorithm with singular value tolerance truncation parameter is used to replace the standard inverse matrix solving operation.

[0070] In this approach, the injected physical fundamental frequency and its corresponding Fourier sine and cosine basis vectors may exhibit an approximate linear correlation due to the rounding effect of the sampling point number. This leads to high condition number or singularity in the coefficient matrix (i.e., the Gram matrix of the basis matrix) of the least squares equation system. Using the standard inverse matrix for solving this problem will cause numerical overflow or drastically unstable calculation results. The Moore-Ponros generalized pseudo-inverse algorithm with singular value tolerance truncation parameters is adopted. Internally, it decomposes the coefficient matrix into a product of orthogonal basis and singular value diagonal matrices through singular value decomposition. Then, it sets small singular values ​​below the tolerance threshold to zero instead of taking their reciprocals. Therefore, it can eliminate the risk of matrix solution collapse caused by multicollinearity from the underlying mathematical mechanism, ensuring a stable and energy-minimum least squares solution under any data conditions. This better guarantees the numerical robustness and reliability of the periodic extrapolation expert branch in engineering deployment.

[0071] In the nonlinear residual expert branch, a patch embedding segmentation mechanism is used to divide the seasonal fluctuation component into multiple local time windows. The sliding step size between adjacent patches is set to no more than half the patch length to ensure that the overlap rate between adjacent patches is not less than 50% (e.g., when the patch length is 64 sampling points, the sliding step size is set to 32 sampling points, and the overlap rate is 50%). The patch embedding segmentation mechanism is an existing technique for segmenting input time series data into fixed or adaptive length segments and mapping them to a high-dimensional embedding space. After embedding and mapping, the multiple local time windows are fed as token sequences into a multi-head self-attention network to capture local nonlinear features such as high-frequency transient spikes and gradient abrupt changes. The linear mapping layer at the output of the nonlinear residual expert branch adopts an all-zero initialization strategy during model initialization.

[0072] In this way, adjacent patches maintain an overlap rate of at least 50%, ensuring that each sampling time is covered by at least two adjacent patches. This avoids the information breaks and feature discontinuities that occur at patch boundaries in non-overlapping segmentation, which is beneficial for multi-head self-attention networks to establish smooth temporal contextual relationships between patches, thereby more accurately capturing local nonlinear features such as high-frequency transient spikes and gradient abrupt changes. Simultaneously, the linear mapping layer at the output of the nonlinear residual expert branch adopts a zero-initialization strategy, ensuring that all weight parameters of this layer are zero at the start of training. Therefore, the output of this expert branch is always a zero vector during forward propagation, without any additive perturbation to the baseline prediction result. This ensures that the expert output tends to zero in the early stages of training, eliminating interference with the convergence of the linear decomposition baseline. As training progresses, the weights of this layer gradually learn meaningful non-zero mappings from zero values, and the compensation contribution of the expert branch smoothly and gradually increases.

[0073] This invention also discloses a joint training method for a time-series prediction model to implement the above-mentioned laser shock signal time-series prediction method, which adopts a two-stage dynamic descrambling joint training strategy:

[0074] First training phase (benchmark warm-up): Disable the forward computation of all compensated prediction expert branches and their corresponding feature-aware fusion gating computation, and use only historical training data to independently fit and train the basic linear decomposition network that generates the benchmark prediction results until convergence and stable basic network weights are obtained.

[0075] The second training phase (joint optimization) involves reactivating all compensated prediction expert branches and their corresponding feature-aware fusion gating computations, and jointly training them with the basic linear decomposition network. During joint training, the learning rate of the basic linear decomposition network is reduced to less than one-tenth of the learning rate of the compensated prediction expert branches to protect the converged weights of the basic network from significant disturbance. Simultaneously, an auxiliary regularization penalty loss term is introduced into the global total loss function. This penalty loss term includes independent penalty terms for each compensated prediction expert branch. Each independent penalty term is constructed by multiplying the difference between the value 1 and the activation control weight corresponding to that branch by the square of the predicted energy of that branch's output, thus guiding each expert branch to suppress its output when the gating activation is low.

[0076] In this way, the first training phase disables all compensation expert branches and gating computations, allowing the basic linear decomposition network to independently optimize to convergence in a pure gradient environment without any compensation interference. This achieves stable and reliable baseline prediction capabilities and establishes a safe baseline for global prediction. In the second training phase, when all expert branches are reactivated for joint training, the learning rate of the basic linear decomposition network is reduced to less than one-tenth of the learning rate of the compensation expert branches. This limits the update magnitude of the backpropagation gradient on the converged basic weights during the joint optimization phase, preventing the gradient signals of the expert branches from causing catastrophic disturbances to the baseline prediction capability. Simultaneously, in the auxiliary regularization penalty loss term, for a given expert branch, when its corresponding activation control weight is close to zero, the difference between the value 1 and the weight is close to 1, resulting in the maximum penalty coefficient, forcing the predicted output energy of that branch to converge to zero. Conversely, when the activation control weight is close to 1, the difference is close to zero, the penalty coefficient tends to disappear, and the branch is allowed to freely output compensated predictions. Therefore, this penalty mechanism explicitly forces the predicted outputs of channels that are not highly activated by gating to converge to zero, decoupling gradient conflicts between branches and ensuring the baseline safety of the model in a weakly regularized environment.

[0077] The present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the laser shock signal timing prediction method as described above, or implements the timing prediction model joint training method as described above.

[0078] The present invention also discloses an electronic device, including a processor and a memory; the memory stores a computer program, which, when run by the processor, causes the processor to execute the laser shock signal timing prediction method as described above, or to implement the timing prediction model joint training method as described above.

[0079] To better verify the effectiveness of the present invention, the applicant further conducted experimental verification of the present invention according to the following requirements, based on the above-described requirements for implementing the embodiments.

[0080] I. Experimental operating environment and underlying parameter configuration.

[0081] The system model construction and predictive inference verified in the experimental examples of this invention rely on cloud and edge computing systems equipped with NVIDIA L4 GPUs. The verification input uses real single-channel ultra-high precision .dat format physical waveform data, and the acquisition standard is 10GHz ultra-high frequency sampling (sampling time interval accurate to 0.1ns).

[0082] To accurately adapt to these extreme frequency band characteristics, the system's underlying core path is configured with highly targeted hyperparameters:

[0083] (1) Physical filtering prior kernel: The moving average decomposition kernel of the time-series time-domain decomposition module is specialized to decomp_kernel = 1539. The physical basis for this is that the dominant acoustic mechanical standing wave sampling period at 6.5MHz frequency on the target material in the 10GHz band is approximately 1538 points. Based on the signal filtering transfer function... When the kernel size K=1539 and the signal period T=1538, the transfer rate is... This means that up to 99.96% of the key mechanical resonance signals will be intercepted by the linear moving average, thus entering the "seasonal component" with extremely high precision, enabling the activated period extrapolation experts to model the standing wave pattern with the most perfect data source.

[0084] (2) Global observation horizon: The historical sequence input window is configured to seq_len = 4096 points (physical duration 409.6ns), covering approximately 2.66 complete 1538-point standing wave cycles, effectively ensuring sufficient repeating waveform basis when extracting time-domain autocorrelation (ACF) evidence and preventing spectral estimation distortion.

[0085] (3) Strict industrial-grade numerical boundary protection: When the signal mean value drops to In cases of extremely weak amplitude, the model forcibly embeds safe_clamp_min = 1e-10 and pseudo-inverse solution anti-singularity truncation tolerance pinv_rcond = 1e-5.

[0086] II. Experimental Laser Dataset and Validation Scenario Setup

[0087] This invention selects two of the most representative extreme physical forms of laser shock stress waves to construct a group of implementation examples. Autoregressive prediction tests are uniformly conducted on three time spans (long, medium, and short) (prediction lengths of 128 points (12.8 ns), 256 points (25.6 ns), and 512 points (51.2 ns)). The classical time-series decomposition lightweight linear model (DLinear) is used as the baseline for comparison.

[0088] Experiment Example 1: Low-frequency mechanical resonance-dominated signal scenario (Type 1). The signal waveform has a distinct 6.5MHz sustained standing wave envelope (prominent physical characteristics). The focus is on verifying whether the system's "sensory gating" can powerfully capture periodic laws, activate with high confidence, and utilize Fourier extrapolation experts to achieve a significant improvement in accuracy.

[0089] Experiment Example 2: High-frequency broadband transient-dominated signal scenario (Type 2). The signal mainly consists of high-energy transient spikes from disordered stress wave scattering and random broadband noise, lacking periodicity. Traditional deep large models often produce "catastrophic overfitting" in this region. The focus is on verifying whether the "selective execution" mechanism of this invention can sense high spectral entropy characteristics, automatically approach zero through gating confidence, trigger "physical bypass defense," and thus safeguard the prediction lower bound of the linear benchmark.

[0090] III. Analysis of Experimental Results and Demonstration of Beneficial Effects

[0091] The complete results of the 12 sets of experimental tests and internal control monitoring data are shown in Table 1. The evaluation index is the mean squared error (MSE), with a smaller value indicating a lower prediction error; the improvement is defined as the percentage reduction in the MSE of this invention (ZetaTime) compared to the DLinear benchmark MSE.

[0092]

[0093] (1) Experimental Example 1: According to Table 1, Figure 2 (Heat distribution map) and Figure 3(Type 1 MSE histogram) shows that, under the highly predictable low-frequency mechanical resonance condition (Type 1), the system of this invention achieves a significant performance improvement compared to the linear reference model. Compared to the linear reference, the average prediction error is significantly reduced by 53.75%, with the MSE decreasing by as much as 91.35% when the prediction length is 256 points. This performance advantage is mainly due to the system's effective capture of strong periodic signals. Under this condition, the time-domain autocorrelation function (ACF) and energy signal-to-noise ratio (SNR) features extracted by the system produce a significant synergistic effect, jointly driving the "periodic gating" (… The activation confidence level of ")" averages 0.6404. This high gating weight allows the Fourier extrapolation expert module, which embeds the "intrinsic constants of target mechanics," to dominate time series prediction. For example... Figure 5 As shown in the (Type 1 error accumulation line graph), unlike the linear model where the MSE diverges over time, this invention, through the effective function of the expert module, successfully constrains the long-term accumulation of error, significantly smoothing and reducing the slope of error growth.

[0094] (2) Experimental Example 2: Under the high-frequency transient-dominated operating condition (Type 2), the signal mainly consists of high-frequency random scattering noise and high-energy spikes. As shown in Table 1, Figure 4 (Type 2 MSE bar chart) and Figure 6 As shown in the (Type 2 error cumulative line graph), the system of this invention not only did not experience prediction collapse on this type of dataset (Type 2), which is prone to overfitting traditional deep networks, but also achieved a positive defensive performance improvement of +37.95% compared to the baseline on the extremely noisy test set by relying on residual fine-tuning. Its core principle lies in the effectiveness of the soft-routing defense mechanism of this invention: after the system's underlying layer senses extremely high random spectral entropy, it actively applies "periodic gating" (…). The confidence level of permissions was drastically reduced, with the mean value plummeting to an extremely weak activation range of only 0.1389 (non-linear gating). It also synchronously regresses to the order of 0.23. This intelligent scheduling of "tactical weights approaching zero" essentially performs a sophisticated "algorithm bypass defense," forcing the complex deep extrapolation network to actively retreat, allowing the inference path to safely fall back and become highly reliable on the linear base. This mechanism avoids the model's forced fitting to broadband random noise, thus ensuring the safety and stability of the prediction. In summary, this invention, through multi-dimensional feature perception fusion and gated soft and hard routing architecture, enables the model to switch freely in an adaptive balance of "using it extensively when it is useful and bypassing it when it is not." Combined with micro-nano-level minimum soft truncation protection embedded throughout the entire lifecycle and Moore-Penrose pseudo-inverse solving technology, the system has withstood the test. The online computation of extremely weak, 10GHz high-dimensional time-series data has met the high-precision, high-frequency, real-time operation requirements of industrial laser intelligent equipment.

[0095] (III) Summary of Overall Technical Effects

[0096] Based on the test results of the two experimental examples, four datasets, and a total of twelve sets of experiments mentioned above, the technical effects of the method of this invention can be summarized as follows:

[0097] (1) Adaptive improvement in prediction accuracy. In low-frequency resonant signal scenarios with deterministic periodic structures, multidimensional evidence fusion gating accurately identifies highly predictable evidence, driving dual-path expert branches to fully participate in prediction correction, achieving an average MSE reduction of 53.75%. In high-frequency transient signal scenarios with weak periodicity, gating automatically attenuates the intensity of expert contributions, achieving an average positive improvement of 37.95% while ensuring baseline safety.

[0098] (2) Effective protection of benchmark performance by the gating mechanism. In all 12 sets of experiments, the MSE of the method of this invention was better than that of the linear benchmark model, verifying the reliability of the gating defense strategy of "use when it is advantageous and withdraw when it is not" on real data. In the face of broadband transient noise, a high-risk region for overfitting in traditional deep models, the autonomous and significant decay of the gating weights effectively suppressed the inappropriate intervention of expert branches.

[0099] (3) The benefits of far-end extrapolation by physical prior injection. In the mid-to-long-range prediction configurations (such as 256-point and 512-point) of low-frequency resonance scenarios, the periodic extrapolation expert, which incorporates the intrinsic mechanical constants of the target material, plays a significant role in correcting deviations and effectively curbs the cumulative divergence of errors that occur in the linear reference model after more than one resonance period.

[0100] (4) Numerical stability and robustness to industrial deployment. The underlying soft truncation protection and pseudo-inverse solution mechanism did not trigger numerical anomalies throughout the experiment, proving the effectiveness of the scheme under extremely weak bias signal conditions. In addition, the dynamic computing power routing mechanism reduces redundant computation by weight decay under low confidence conditions, providing a feasible technical path for computing power optimization in actual industrial deployment scenarios.

Claims

1. A method for time-series prediction of laser shock signals, characterized in that, Includes the following steps: Step S0: Construct and train a time series prediction model: Construct a time series prediction model that includes a basic linear decomposition network and multiple independent compensation prediction expert branches, and train and initialize the time series prediction model using historical training data. Step S1: Perform time series decomposition and baseline prediction: Obtain the collected laser shock time series data to be predicted and input it into the trained time series prediction model. Perform reversible instance normalization on the model. Use the physical decomposition kernel in the basic linear decomposition network that is adapted to the target material's physical resonance fundamental frequency period to perform low-pass filtering time series decomposition on the normalized time series, extract the basic trend component and the seasonal fluctuation component, and then predict and project the basic trend component and the seasonal fluctuation component respectively through dual-line linear mapping and sum them to obtain the baseline prediction result. The kernel length of the physical decomposition kernel is determined according to the target material's mechanical resonance fundamental frequency: the sampling point period corresponding to the fundamental frequency oscillation is obtained by rounding down the ratio of the sampling rate to the resonance fundamental frequency. Take no less than An odd number plus one is used as the kernel length to ensure that the moving average low-pass passivity of the physical decomposition kernel for the resonant fundamental frequency signal is reduced to below 0.1%, so that the resonant fundamental frequency signal can fully enter the seasonal fluctuation component. Step S2, complete multi-dimensional feature extraction and micro / nano-level numerical protection: Based on the seasonal fluctuation component and basic trend component obtained from the decomposition in step S1, multi-dimensional predictability features are extracted. These multi-dimensional predictability features include at least: time-domain autocorrelation features, frequency-domain spectral entropy features, and energy signal-to-noise ratio features. In the underlying calculation paths involving variance calculation, total power spectrum energy calculation, and normalized power spectrum logarithmic operation involved in extracting the above multi-dimensional predictability features, a micro / nano-level lower bound soft truncation protection mechanism is embedded. This mechanism performs lower bound truncation processing on input values ​​or intermediate state values ​​during the calculation process, with a threshold value not less than a preset minimum value, to accommodate signal amplitude standard deviations within a certain range. The numerical accuracy requirements for extremely weak bias laser shock signals at or below the order of magnitude are to prevent numerical underflow or division by zero anomalies caused by sensor dead values ​​or extremely weak biases; wherein, the preset minimum threshold value is no greater than Positive numbers; Step S3, complete feature-aware fusion gating generation: cross-fuse the multidimensional predictability features extracted in step S2 to generate activation control weights for controlling the multiple independent compensation prediction expert branches respectively. Step S4, selective execution and dynamic computational power routing: During forward prediction inference, the maximum value of the activation control weight corresponding to each independent compensation prediction expert branch is independently compared with a preset dynamic bypass threshold, triggering a soft routing mechanism. The soft routing mechanism includes: if the maximum value of the activation control weight corresponding to an independent compensation prediction expert branch is lower than the dynamic bypass threshold, it is determined that the current signal is in a high-frequency broadband transient noise condition that cannot be effectively modeled for that branch, and the forward feature extraction and prediction calculation of that branch are actively skipped to prevent overfitting and save end-side computational power; if the maximum value of the activation control weight corresponding to an independent compensation prediction expert branch is not lower than the dynamic bypass threshold, the branch is activated to perform forward calculation to generate compensation prediction results; wherein, the dynamic bypass threshold ranges from [value missing]. to ; Step S5, complete prediction fusion and reverse restoration: The baseline prediction result obtained in step S1 and the branch compensation prediction results after element-wise dynamic weighting by the corresponding activation control weights in step S4 are added and fused together, and the inverse normalization operation corresponding to the reversible instance normalization is performed to restore to the original physical scale, and the final laser shock signal timing prediction result is output.

2. The laser shock signal timing prediction method as described in claim 1, characterized in that, The specific steps for extracting multidimensional predictable features in step S2 include: The ratio of the autocorrelation covariance of the seasonal fluctuation component at each prediction step to the variance after processing by the micro / nano-level lower bound soft truncation protection mechanism is calculated, and the time-domain autocorrelation feature is obtained through nonlinear mapping. The variances of the seasonal fluctuation component and the basic trend component after processing by the micro / nano-level lower bound soft truncation protection mechanism are calculated respectively, and the logarithm of the ratio of their variances is calculated. The energy signal-to-noise ratio feature is obtained through nonlinear mapping. The power spectrum of the seasonal fluctuation component is obtained by frequency domain transformation. After lower bound truncation of the total energy of the power spectrum based on the micro / nano-level lower bound soft truncation protection mechanism, the normalized power probability distribution is calculated. The logarithmic probability is obtained by applying the micro / nano-level lower bound soft truncation protection mechanism to the normalized power probability distribution again, and the Shannon information entropy is calculated. The frequency domain spectral entropy feature is obtained through inverse nonlinear mapping.

3. The laser shock signal timing prediction method as described in claim 1, characterized in that, In step S3, the multiple independent compensation prediction expert branches include at least a nonlinear residual expert branch and a periodic extrapolation expert branch. The specific steps for generating the activation control weights for each branch include: For the nonlinear residual expert branch: the time-domain autocorrelation feature, the energy signal-to-noise ratio feature, and the frequency-domain spectral entropy feature are multiplied and fused element-wise to generate nonlinear activation control weights; For the periodic extrapolation expert branch: extract autocorrelation peak features at multiple sampling points corresponding to the fundamental frequency of the target material's physical resonance and its higher harmonics, multiply the autocorrelation peak features element-wise with the energy signal-to-noise ratio features, and then multiply by the square of the frequency domain spectral entropy features to generate periodic activation control weights.

4. The laser shock signal timing prediction method as described in claim 3, characterized in that, The periodic extrapolation expert branch adopts a gradient-free deterministic extrapolation design, specifically: In the model building stage of step S0, the physical fundamental frequency of the specific mechanical resonance standing wave of the laser shock target and the sampling point period corresponding to each harmonic are explicitly injected as deterministic periodic constants. Based on this, the Fourier input basis matrix and output basis matrix composed of sine and cosine functions are constructed. When solving the projection coefficient matrix corresponding to the Fourier input basis matrix using the least squares method, the Moore-Penrose generalized pseudo-inverse algorithm with singular value tolerance truncation parameter is used to replace the standard inverse matrix solving operation.

5. The laser shock signal timing prediction method as described in claim 3, characterized in that, In the nonlinear residual expert branch, the seasonal fluctuation component is divided into multiple local time windows using a patch embedding segmentation mechanism. The sliding step size between adjacent patches is set to be no greater than half the patch length to ensure that the overlap rate between adjacent patches is no less than 50%. After embedding and mapping, the multiple local time windows are fed into the multi-head self-attention network as a token sequence to capture local nonlinear features, including high-frequency transient spikes and gradient abrupt changes. The linear mapping layer at the output of the nonlinear residual expert branch adopts an all-zero initialization strategy during model initialization in step S0.

6. The laser shock signal timing prediction method according to any one of claims 1 to 5, characterized in that, In step S0, a two-stage dynamic perturbation reduction joint training strategy is used to jointly train the time series prediction model, specifically including: First training phase: Disable the forward computation of all compensated prediction expert branches and their corresponding feature-aware fusion gating computation, and use only historical training data to independently fit and train the basic linear decomposition network that generates the benchmark prediction results until convergence and stable basic network weights are obtained. Second training phase: Reactivate all compensated prediction expert branches and their corresponding feature-aware fusion gating computations, and perform joint training with the basic linear decomposition network; during joint training, reduce the learning rate of the basic linear decomposition network to less than one-tenth of the learning rate of the compensated prediction expert branches to protect the converged weights of the basic network from significant disturbance; simultaneously, introduce an auxiliary regularization penalty loss term into the global total loss function; the penalty loss term includes independent penalty terms for each compensated prediction expert branch, each independent penalty term is constructed based on the difference between the value 1 and the activation control weight corresponding to that branch multiplied by the square of the predicted result energy term of that branch, to guide each expert branch to suppress its output when the gating activation is low.

7. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is executed by the processor, it implements the laser shock signal timing prediction method as described in any one of claims 1 to 6.

8. An electronic device comprising a processor and a memory; wherein the memory stores a computer program, characterized in that, When the computer program is run by the processor, the processor performs the laser shock signal timing prediction method as described in any one of claims 1 to 6.

Images