Fire window glass breakage early warning method based on strain mutation

Abstract

The present application relates to a kind of fireproof window glass breakage early warning method based on strain mutation.Specifically by collecting dynamic response sequence, the intrinsic mode function component set is obtained after processing, and instantaneous phase trajectory sequence is transformed;Theoretical phase reference curve is constructed, and the spatial-temporal diagram of debonding strength is calculated from residual phase sequence;Residual phase sequence recurrence plot reconstruction, extract three nonlinear dynamics indexes to construct three-dimensional phase chaos index vector, input lightweight graph attention network to output global breakage trend probability distribution, mapping generates final early warning instruction, completes fireproof window glass breakage early warning.This method can realize the multi-scale, real-time code of glass structure Damage trend prediction, improve the level of safety protection and response efficiency.

Description

Technical Field

[0001] This invention relates to the field of structural health monitoring and strain mutation feature identification technology, and in particular to a method for early warning of fireproof window glass breakage based on strain mutation. Background Technology

[0002] Current technological advancements in fireproof window glass breakage monitoring and early warning have led to several mainstream solutions centered on real-time perception and failure prediction for structural health monitoring. The industry primarily relies on methods such as strain signal amplitude abrupt change identification, single sensor point detection, and multimodal data stitching to determine glass material damage and breakage trends. Typical technologies include fiber optic grating sensor arrays for real-time acquisition of glass structural conditions, combined with signal amplitude threshold determination, high-frequency noise removal, and time-series modeling methods to process dynamic strain responses. Some solutions attempt to introduce Bayesian optimization or fractal dimension calculation to improve the accuracy of breakage trend prediction. However, these methods generally suffer from the following technical bottlenecks:

[0003] Existing mutation identification models, when dealing with the multi-scale strain characteristics of fireproof window glass, lack effective organization for the fusion of information between different physical mechanisms, leading to information loss or redundancy in the fusion process of strain amplitude characteristics and mutation signals. For example, by using amplitude threshold determination methods, the sensing system cannot separate multiple physical mechanisms such as intrinsic material creep, interface debonding, and microcrack generation, and can only output single-point anomalies or global alarms, lacking interpretability of the fracture evolution mechanism and making it difficult to capture the true physical process of the strain response of fireproof window glass under thermo-mechanical coupling conditions. Summary of the Invention

[0004] This application provides a method for early warning of fireproof window glass breakage based on strain mutation, aiming to solve the problem mentioned in the background art of lacking interpretability of the breakage evolution mechanism and difficulty in capturing the real physical process of strain response of fireproof window glass under thermo-mechanical coupling conditions.

[0005] This application provides a method for early warning of fireproof window glass breakage based on strain mutation, specifically including:

[0006] S1: Dynamic response sequences are collected using sensors deployed on fireproof window glass;

[0007] S2: Process the dynamic response sequence to obtain the intrinsic mode function component set, calculate and transform the intrinsic mode function component set to obtain the instantaneous phase trajectory sequence;

[0008] S3: Based on the instantaneous phase trajectory sequence, a theoretical phase reference curve is constructed. The residual phase sequence is calculated by comparing the instantaneous phase trajectory sequence with the theoretical phase reference curve. The residual phase sequence is then processed to obtain a spatiotemporal diagram of debonding intensity.

[0009] S4: Perform recursive graph reconstruction on the residual phase sequence and extract three nonlinear dynamic indices to construct a three-dimensional phase chaos exponential vector;

[0010] S5: Input the debonding strength spatiotemporal map and the three-dimensional phase chaotic exponential vector into the lightweight graph attention network, and output the global fracture trend probability distribution;

[0011] S6: Utilize the global breakage trend probability distribution to perform mapping calculations, generate a final warning command, and issue a warning for the breakage of the fireproof window glass based on the final warning command.

[0012] The fireproof window glass breakage early warning method based on strain mutation provided in this application has the following beneficial effects:

[0013] This application effectively overcomes the shortcomings of traditional amplitude threshold or wavelet packet energy methods, which are susceptible to noise interference, have high false alarm rates, and are difficult to distinguish between different failure modes such as debonding and microcracks, by extracting nonlinear dynamic indicators to form a three-dimensional chaotic exponential vector. This significantly improves the physical interpretability and accuracy of damage assessment. At the same time, by introducing a lightweight graph attention network to discretize the glass surface into grid nodes and construct adaptive edge weights based on geometric adjacency and historical phase coherence, a global fracture trend probability distribution is output on low-cost hardware at the edge, realizing low-latency and high-efficiency real-time fracture early warning. This provides a highly reliable and easily integrated technical solution for the safe operation and maintenance of fireproof windows in key scenarios such as high-rise buildings and rail transit. Attached Figure Description

[0014] Figure 1 This is the main flowchart of a fireproof window glass breakage early warning method based on strain mutation.

[0015] Figure 2 This is a sub-flowchart of a method for early warning of fireproof window glass breakage based on strain mutation.

[0016] Figure 3 This is another sub-flowchart of a method for early warning of fireproof window glass breakage based on strain mutation. Detailed Implementation

[0017] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0018] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0019] like Figure 1 As shown, this application provides a method for early warning of fireproof window glass breakage based on strain mutation, specifically including:

[0020] S1: Dynamic response sequences are collected using sensors deployed on fireproof window glass.

[0021] S2: Process the dynamic response sequence to obtain the intrinsic mode function component set, calculate and transform the intrinsic mode function component set to obtain the instantaneous phase trajectory sequence.

[0022] S3: Based on the instantaneous phase trajectory sequence, a theoretical phase reference curve is constructed. The residual phase sequence is obtained by calculating the instantaneous phase trajectory sequence and the theoretical phase reference curve. The residual phase sequence is then processed to obtain a spatiotemporal diagram of debonding intensity.

[0023] S4: Perform recursive graph reconstruction processing on the residual phase sequence to extract three nonlinear dynamic indices to construct a three-dimensional phase chaos index vector.

[0024] S5: Input the debonding strength spatiotemporal map and the three-dimensional phase chaotic exponential vector into the lightweight graph attention network, and output the global rupture trend probability distribution.

[0025] S6: Utilize the global breakage trend probability distribution to perform mapping calculations, generate a final warning command, and issue a warning for the breakage of the fireproof window glass based on the final warning command.

[0026] like Figure 2 As shown, step S1 involves using sensors deployed on the fireproof window glass to collect a dynamic response sequence. Specifically, this includes:

[0027] S1.1: Based on the physical topology of the key monitoring area of ​​the fireproof window glass, the deployment coordinates of the micro fiber grating array sensor are determined according to the distribution law of thermal stress concentration. The sensor nodes are then solidified and installed at the interface of the glass surface and the frame adhesive layer to construct a distributed strain sensing network covering the entire glass surface.

[0028] Obtain the physical topological parameters of the fireproof window glass, including glass size, thickness, and frame constraints, as a geometric reference for sensor deployment.

[0029] Based on the finite element thermo-mechanical coupling simulation model, the temperature field distribution and stress concentration areas on the glass surface under fire conditions are simulated, and the corners and edges are identified as high thermal stress sensitive areas.

[0030] Based on the stress gradient distribution cloud map, the key monitoring coordinates of the miniature fiber optic array sensor are determined, focusing on covering the stress concentration peak points and areas where strain changes occur frequently.

[0031] After cleaning the glass surface, UV-cured optical adhesive is used to precisely attach the fiber grating sensor nodes along a preset coordinate path to ensure that the sensor axis is parallel to the principal stress direction in order to maximize strain sensitivity.

[0032] At the interface between the glass and the frame adhesive layer, a flexible encapsulated fiber optic grating sensing unit is embedded and fixed by a microstructure slot to resist shear deformation, thereby enabling direct sensing of the interface debonding state.

[0033] Optical path connectivity tests and initial wavelength calibrations are performed on the deployed sensor network. The center wavelength reference value of each node under no-load conditions is recorded to construct the physical underlying layer of the distributed strain sensing network.

[0034] Through the above-mentioned deployment and solidification process, the physical structure of the glass is transformed into a distributed strain sensing network with spatial resolution, realizing high-precision data acquisition covering the thermal stress concentration area and the interface damage area, providing a high signal-to-noise ratio raw data source for subsequent phase spectrum reconstruction.

[0035] For example, simulations of tempered fireproof glass with dimensions of 1200mm × 800mm × 6mm show that the corner stress concentration factor reaches 1.8. Eight fiber Bragg grating sensors are arranged 50mm from the four corners and at the midpoint of the long side, with a spacing of 100mm. The sensors are cured using refractive index-matched adhesive, and the initial wavelength calibration error is controlled within ±0.01nm. This arrangement effectively captures the strain abrupt change signal during the initiation stage of microcracks at the corners. Compared to a uniform arrangement, it significantly improves the sensitivity to early damage, ensuring the data integrity for subsequent phase feature extraction.

[0036] S1.2: Based on each miniature fiber optic grating array sensor in the distributed strain sensing network, a wavelength demodulator is used to perform high-speed synchronous trigger control with a sampling rate of not less than 2kHz to capture the amount of reflected wavelength drift under heat or external force in real time, so as to generate a multi-channel raw optical signal stream containing timestamp alignment.

[0037] For example, 16 fiber Bragg grating sensor nodes were deployed on a 1200mm × 800mm fireproof window. The wavelength demodulator was configured with a sampling rate of 5kHz, far exceeding the minimum requirement of 2kHz. Synchronization trigger accuracy was controlled within 1μs. In a simulated fire heating scenario, the demodulator recorded wavelength drift at a resolution of 0.1pm. Assuming a node's initial wavelength was 1550.00nm, and it drifted to 1550.05nm after heating, the demodulator completed a full scan and peak lock within 200μs. Verification showed that the timestamp deviation of the 16 channels was less than 0.5μs, and the effective data rate reached 99.9%, ensuring the spatiotemporal synchronization required for subsequent phase analysis.

[0038] S1.3: Perform wavelength-strain linear mapping transformation on the multi-channel original optical signal stream, and use a pre-calibrated sensitivity coefficient to convert the reflected wavelength drift into a physical strain value, so as to output an initial strain time series set characterizing the degree of local deformation of the glass.

[0039] Receive the time-base corrected multi-channel raw optical signal stream and extract the drift sequence of the center wavelength of each channel as a function of time.

[0040] Based on the fiber grating sensing principle, a linear mapping model between wavelength drift and physical strain is established, which follows the variation law of Bragg condition.

[0041] The specific strain sensitivity coefficient of each sensor node is read from the pre-stored calibration database. This coefficient is obtained by calibrating a tensile testing machine under constant temperature conditions.

[0042] The above mapping operation is performed in parallel on multi-channel data to eliminate sensitivity deviations caused by manufacturing differences between different sensors.

[0043] The calculated discrete strain values ​​are reassembled in time stamp order to construct a multidimensional strain matrix.

[0044] The strain matrix is ​​dimensionally standardized by normalizing the strain units of all channels to microstrain.

[0045] By using wavelength-strain linear mapping transformation, the optical signal drift from the previous step is converted into an initial strain time series set characterizing the degree of local deformation of the glass, thus realizing the accurate quantification of optical signals into mechanical physical quantities and providing a high-fidelity strain data foundation for subsequent phase spectrum reconstruction.

[0046] For example, for the five fiber Bragg grating nodes deployed at the four corners and center of the fireproof window, the initial reference wavelength is 1550.00 nm. The strain sensitivity coefficient is set to 1.2 pm / . When a node is subjected to thermal stress, and the real-time wavelength drifts to 1550.12 nm, the strain value is calculated to be 100 με using the formula. The system continuously acquires data at a frequency of 2 kHz, generating an initial strain time series set containing five channels, each with a channel length of N. The data accuracy reaches 0.1 με, effectively preserving the high-frequency dynamic characteristics of glass micro-deformation.

[0047] S1.4: Based on the initial strain time series set, a sliding window denoising method is used to remove high-frequency electronic noise and environmental micro-vibration interference, and time base correction is performed to eliminate transmission delay differences, so as to generate a cleaned and aligned dynamic response sequence.

[0048] Receive the initial strain time series set after wavelength-strain mapping transformation, which contains raw physical strain data acquired synchronously through multiple channels.

[0049] A sliding window mean filtering process is performed on the initial strain time series set. The window length is set to 50 sampling points. The arithmetic mean of the data within the window is calculated to smooth high-frequency electronic noise interference and generate a preliminary denoised strain sequence.

[0050] An outlier removal method based on standard deviation is adopted to calculate the standard deviation of strain data within a sliding window. Outliers exceeding three times the standard deviation are replaced with linear interpolation results of adjacent valid data to eliminate transient spike pulses caused by environmental micro-vibrations.

[0051] Cross-correlation analysis was performed on the denoised multi-channel strain sequence. A reference channel signal was selected, and the peak positions of the cross-correlation functions between the other channel signals and the reference signal were calculated to determine the relative time delay between each channel.

[0052] A time-base correction matrix is ​​constructed, and phase compensation shifting is performed on the strain sequences of each channel based on the determined relative time delay. The data timestamps of all channels are aligned using cubic spline interpolation resampling technology.

[0053] Through the above sliding window denoising and time-base correction processing, the initial strain data containing noise and timing bias is transformed into a dynamic response sequence with high signal-to-noise ratio and strict time alignment, thus achieving the high-precision data input conditions required for subsequent phase spectrum reconstruction analysis.

[0054] For example, 16 fiber Bragg grating sensor nodes are deployed on the surface of a fireproof window glass, with a sampling frequency of 2kHz. The sliding window length is set to 50 points, corresponding to a time span of 25ms. The initial strain data of channels 1 to 16 are mean-filtered to remove electronic noise with frequencies higher than 40Hz. The standard deviation of each window is calculated, and abnormal peaks with amplitudes exceeding ±5με are removed. Using channel 1 as a reference, it is found that channel 8 has a time delay of 3ms and channel 15 has a time delay of 1.5ms. Cubic spline interpolation is used to shift channel 8 forward by 6 sampling points and channel 15 forward by 3 sampling points to complete the time base alignment of all channels. After processing, the signal-to-noise ratio is improved from 20dB to 35dB, and the time synchronization error between channels is controlled within 0.1ms, meeting the strict requirements of phase calculation for data consistency.

[0055] like Figure 3 As shown, step S2 involves processing the dynamic response sequence to obtain a set of intrinsic mode function components, calculating and transforming the set of intrinsic mode function components to obtain an instantaneous phase trajectory sequence. Specifically, this includes:

[0056] S2.1: Perform extreme point detection and envelope fitting processing on the dynamic response sequence. Construct upper and lower envelopes based on cubic spline interpolation. Calculate the mean sequence of the upper and lower envelopes and subtract the mean sequence from the dynamic response sequence to obtain the first-order residual strain component containing high-frequency oscillation components.

[0057] It receives a cleaned and aligned dynamic response sequence and uses it as the input data source for extremum detection and envelope fitting.

[0058] A local extremum search is performed on the strain response sequence, traversing the sampling points on the time axis to identify all sets of maxima and minima that satisfy the condition of the first derivative sign change.

[0059] For the identified set of maxima, a smooth upper envelope curve is constructed using cubic spline interpolation to ensure that the curve is continuous and second-differentiable at the extrema.

[0060] For the identified set of minimum points, a smooth lower envelope curve is constructed using cubic spline interpolation, forming an accurate approximation of the original signal fluctuation boundary.

[0061] Calculate the arithmetic mean of the upper and lower envelope curves at each time stamp to generate a mean sequence representing the local DC component or trend term of the signal.

[0062] The mean sequence is subtracted point by point from the original dynamic response sequence to remove the low-frequency trend component and obtain the first-level residual strain component containing high-frequency oscillation components.

[0063] By performing the envelope mean removal process described above, the non-stationary strain signal is transformed into a zero-mean oscillating component, achieving preliminary separation of the signal frequency band and providing standardized input for subsequent iterative screening.

[0064] For example, for a strain sequence with a sampling rate of 2 kHz, the extremum determination window length is set to 5 sampling points. A total of 450 maxima and 448 minima were detected between seconds 100 and 1000. Cubic spline interpolation was used to construct upper and lower envelopes, and the mean sequence was calculated. The original strain value at t=500s was 120 με, corresponding to upper and lower envelope mean values ​​of 125 με and 115 με, respectively, with a total mean of 120 με. Subtraction yielded a first-order residual component of 0 με. This processing effectively removed baseline drift caused by slow temperature changes, reducing the standard deviation of the residual component from 15 με in the original signal to 3 με, significantly improving the convergence speed of subsequent mode decomposition.

[0065] S2.2: Perform iterative screening convergence judgment processing on the first-level residual strain components. Based on the preset screening termination criterion, repeatedly perform envelope mean elimination operation until the residual strain components satisfy the balance condition of zero mean characteristic and number of extreme points of the intrinsic mode function, thereby generating a set of intrinsic mode function components with clear physical frequency band meaning.

[0066] Iterative screening convergence judgment is performed on the first-level residual strain components, and the envelope mean removal operation is repeatedly performed based on the preset screening termination criterion. The coordinate set of all local maxima and minima of the current residual strain component is obtained. The upper and lower envelope curves are fitted using cubic spline interpolation. The arithmetic mean of the upper and lower envelopes at each sampling time is calculated to generate a local mean sequence. This local mean sequence is subtracted point by point from the current residual strain component to obtain the updated residual component. The number of extreme points and the number of zero crossings of the updated residual component are counted. The absolute value of the difference between the number of extreme points and the number of zero crossings is calculated, and it is determined whether it is less than or equal to 1. If the difference is greater than 1 or there is a negative amplitude, it is determined that the intrinsic mode function condition is not met, and the updated residual component is used as a new input to repeat the above envelope construction and mean removal process. If the difference is less than or equal to 1 and there is no negative amplitude, it is determined that the screening termination criterion is met, and the component is marked as a standard intrinsic mode function component and stored. Through iterative filtering, non-stationary strain signals are decomposed into a set of eigenmode function components that satisfy orthogonality and completeness, thereby achieving accurate separation of multi-scale physical frequency bands.

[0067] For example, the screening termination criterion is set as the difference between the number of extreme points and the number of zero-crossing points being less than or equal to 1. Initial screening of the residual strain component with a length of 4096 points detects 120 extreme points and 118 zero-crossing points, with a difference of 2, which does not meet the condition. After performing envelope mean rejection and re-detection, the number of extreme points becomes 115 and the number of zero-crossing points becomes 114, with a difference of 1, which meets the condition. This component is output as the first-order intrinsic mode function, preserving the high-frequency microcrack initiation characteristics and significantly improving the signal-to-noise ratio.

[0068] S2.3: Perform Hilbert transform processing on each intrinsic mode function component in the intrinsic mode function component set, construct an orthogonal conjugate signal sequence using Cauchy principal value integral, and combine the intrinsic mode function component with its orthogonal conjugate signal sequence to form an analytic signal vector in the complex domain, so as to obtain a complex analytic signal representation containing instantaneous amplitude and instantaneous phase information.

[0069] For example, a Hilbert transform is performed on the third-order intrinsic mode function components with a sampling rate of 2 kHz, and the integration window length is set to 1024 points. After calculating the complex analytic signal, the instantaneous phase sequence is extracted. In the initial stage of glass heating, the phase change rate is 0.05 rad / s; when microcrack initiation occurs, the phase abrupt change slope increases significantly to 2.5 rad / s. This processing method effectively captures early weak damage characteristics that are difficult to identify with traditional amplitude analysis, and significantly improves the sensitivity of fracture trend prediction.

[0070] S2.4: Perform arctangent phase solution processing on the complex analytic signal representation, calculate the principal value of the phase angle based on the ratio of the imaginary part data to the real part data of the complex analytic signal representation, and eliminate phase jump ambiguity through the phase expansion method, thereby extracting a continuous and smooth instantaneous phase trajectory sequence that characterizes the instantaneous state change of the material.

[0071] The complex analytic signal generated by S2.3 is received, and its real and imaginary data are extracted as the basic input variables for phase calculation.

[0072] A principal value calculation model for the phase angle is constructed based on the arctangent function, and the initial angle distribution of the instantaneous phase is determined by the ratio of the imaginary part to the real part.

[0073] The principal value of the phase angle is calculated using the following formula:

[0074]

[0075] in, for The principal value of the phase angle at time t. Let be the imaginary part of the complex analytic signal. Let be the real part of the complex analytic signal.

[0076] Identify the ±π phase jump points in the calculation results caused by the range limitation of the arctangent function, and mark the timestamp positions of the phase discontinuities.

[0077] The phase unwinding method is executed to detect the absolute value of the phase difference between adjacent sampling points. If it exceeds a preset threshold, it is determined to be phase winding.

[0078] Add or subtract integer multiples of 2π to the phase sequence that is determined to be wound to eliminate artificial phase discontinuities caused by periodic truncation.

[0079] By iteratively correcting all transition points, an absolute phase trajectory with temporal continuity is reconstructed, thus restoring the true phase history of glass strain evolution.

[0080] Through the above arctangent solution and phase expansion processing, the analytic signal in the complex domain is transformed into a continuous and smooth instantaneous phase trajectory sequence, realizing an unambiguous characterization of the microscopic deformation state of the glass material and providing a high-precision data foundation for subsequent residual phase separation.

[0081] S2.5: Perform multi-scale temporal alignment and outlier removal processing on the instantaneous phase trajectory sequence, identify and correct phase spikes caused by sensor noise based on the statistical characteristics of the sliding window, and use the corrected instantaneous phase trajectory sequence as the final output data to support the subsequent generation of residual phase sequence and quantitative analysis of interface debonding degree.

[0082] The instantaneous phase trajectory sequence after phase unrolling is received. Although the 2π jump ambiguity has been eliminated, the sequence still contains high-frequency random disturbances caused by sensor shot noise and environmental micro-vibrations.

[0083] A local statistical feature extraction mechanism based on a sliding window is constructed. The window length is set to M, and the phase data of M / 2 sampling points before and after the current time t are extracted to form a local analysis subset.

[0084] Calculate the mean and standard deviation of phase data within the local analysis subset, establish a dynamic confidence interval using the three-standard-deviation criterion, and identify outlier phase points that exceed the upper or lower limit of the interval as potential noise spikes.

[0085] For the identified outlier phase points, the Lagrange interpolation method is used to reconstruct the numerical values ​​using adjacent normal phase points, and the outliers are replaced to restore the continuity of the phase trajectory.

[0086] A multi-scale temporal alignment strategy is introduced to decompose the corrected phase sequence into high-frequency detail components and low-frequency trend components. Smoothing is performed using sliding windows of different widths for each component. A narrow window is used to preserve abrupt changes in the high-frequency components, while a wide window is used to suppress baseline drift in the low-frequency components.

[0087] Through the above-mentioned multi-scale filtering and outlier correction processing methods, the noisy original instantaneous phase trajectory is transformed into a corrected instantaneous phase trajectory sequence with high signal-to-noise ratio and clear physical meaning. This achieves accurate preservation of weak strain abrupt change signals and effective removal of noise interference, providing a high-quality data foundation for subsequent residual phase generation and interface debonding quantization.

[0088] For example, the sliding window length is set to 50, corresponding to a 25ms time span at a 2kHz sampling rate. For a phase sequence at a certain monitoring point, the mean μ = 1.2 rad and the standard deviation σ = 0.05 rad within the window are calculated, and the threshold is set to μ ± 3σ, i.e., [1.05, 1.35] rad. A phase value of 1.42 rad is detected at t = 100 ms, which is determined to be a noise spike. Using the normal phase point from t = 98 ms to 102 ms, a cubic Lagrange interpolation is performed to reconstruct the phase at this point as 1.21 rad. A window smoothing of M = 10 is used for high-frequency components, and a window smoothing of M = 100 is used for low-frequency components. After processing, the standard deviation of the phase sequence is reduced from 0.05 rad to 0.01 rad, significantly improving the signal-to-noise ratio and ensuring the accuracy of subsequent debonding identification.

[0089] Step S3: Based on the instantaneous phase trajectory sequence, a theoretical phase reference curve is constructed. The residual phase sequence is calculated by comparing the instantaneous phase trajectory sequence with the theoretical phase reference curve. The residual phase sequence is then processed to obtain a spatiotemporal diagram of the debonding intensity. Specifically, this includes:

[0090] S3.1: Based on the corrected instantaneous phase trajectory sequence and synchronously acquired temperature field data, the theoretical phase reference curve is constructed using the parameter inversion method of the generalized Maxwell rheological model (a classical mechanical model describing the viscoelastic behavior of materials) to obtain the theoretical phase reference curve characterizing the intrinsic creep behavior of the material in a damage-free state.

[0091] The system receives the corrected instantaneous phase trajectory sequence after cleaning and alignment in step S2, and simultaneously acquires the temperature field data at the corresponding timestamp, which serves as the dual-mode input source for constructing the theoretical phase reference curve.

[0092] To address the thermoviscoelastic properties of fireproof window glass materials, a generalized Maxwell rheological model is established. This model consists of a Hooke spring and multiple Maxwell elements connected in parallel, and is used to describe the intrinsic creep behavior of the material under thermo-mechanical coupling.

[0093] The relaxation time in the model was corrected using the WLF equation (an equation describing the relationship between relaxation time and temperature) based on temperature field data. The shift factor at different temperatures was calculated to reflect the effect of temperature change on the viscoelastic response rate of the material. The formula is as follows:

[0094]

[0095] in, For shift factor, For real-time temperature, For reference temperature, and is a material constant.

[0096] Based on the corrected relaxation time, a theoretical phase expression for the strain response is constructed. The model parameters are then inverted using the least squares method to minimize the fitting error between the theoretical phase curve and the measured phase trajectory in the undamaged stage.

[0097] Substituting the optimal model parameters obtained from the inversion into the generalized Maxwell model, a theoretical phase reference curve representing the intrinsic creep of the material under no-damage conditions driven only by thermal stress is generated.

[0098] Through the above parameter inversion and model building processes, the complex strain response coupled with multiple physics fields is transformed into a quantifiable theoretical phase reference, realizing the accurate stripping of intrinsic creep phase drift and providing a high-purity background reference for the independent extraction of interface debonding features.

[0099] S3.2: Perform point-by-point difference operation on the corrected instantaneous phase trajectory sequence and the theoretical phase reference curve to remove the influence of intrinsic creep phase drift of the material and generate a residual phase sequence containing only interface damage and nonlinear disturbance information.

[0100] Read the corrected instantaneous phase trajectory sequence output by S2.5 and the theoretical phase reference curve generated by S3.1, and ensure that the two are strictly aligned on the time axis and have the same number of sampling points.

[0101] A point-by-point differential calculation model is established. For the same time step t, the measured instantaneous phase value is subtracted from the theoretical reference phase value to eliminate the inherent viscoelastic creep effect of glass material caused by temperature change.

[0102] The residual phase is calculated using the following formula:

[0103] in, For residual phase, This is the corrected measured instantaneous phase. This is the theoretical reference phase.

[0104] By iterating through all time steps within the entire monitoring period, the above subtraction operation is performed sequentially to construct a complete set of residual phase time series.

[0105] The generated residual sequence was subjected to zero-mean verification to confirm that the long-term trend term had been completely removed, and only the local phase abrupt change component caused by interface debonding or microcracks was retained.

[0106] By using point-by-point differential processing, the measured phase and theoretical benchmark from the previous step are transformed into pure residual phase sequence data, effectively removing intrinsic creep interference and significantly improving the signal-to-noise ratio and specificity of subsequent interface damage identification.

[0107] S3.3: Based on the sliding window mechanism, perform local statistical feature extraction processing on the residual phase sequence, and obtain the standard deviation and skewness values ​​of the residual phase sequence within the window to generate a joint statistical feature vector characterizing the intensity of local phase fluctuations and the asymmetry of distribution.

[0108] For the residual phase sequence generated in S3.2, a local time-domain analysis window is constructed by setting the sliding window length and step size. A subset of residual phase data is extracted within the window, and its arithmetic mean is calculated as the reference center.

[0109] Based on the residual phase data after mean centering, the sum of squares of the deviations of each sampling point from the mean is calculated. The standard deviation is obtained by taking the square root after dividing by the degrees of freedom, which quantifies the dispersion and volatility of the local phase.

[0110] Simultaneously, the third central moment of the residual phase data is calculated, and the skewness value is obtained after normalization, which characterizes the asymmetry of the phase distribution relative to the normal distribution and the tendency of tail anomalies.

[0111] The standard deviation and skewness values ​​are arranged by time index and combined to generate a joint statistical feature vector, which fully describes the statistical morphology of the phase signal within the local window.

[0112] By traversing the entire sequence through a sliding window and updating the joint statistical feature vector in real time, we can achieve sensitive capture and quantitative description of non-Gaussian phase abrupt changes caused by interface debonding.

[0113] S3.4: Perform dual-threshold logic decision processing on the joint statistical feature vector, and locate the timestamps and spatial coordinates of local phase step events that meet the mutation conditions based on the preset standard deviation threshold and skewness threshold, so as to generate a set of local phase step events that mark the location of interface debonding.

[0114] For example, the reference noise standard deviation is set to 0.05 rad, the amplitude multiple to 3.0, and the skewness threshold to 1.5. At t=12.5s, the standard deviation of a certain monitoring point within the window is 0.18 rad, and the skewness is 2.1. The standard deviation ratio is calculated to be 3.6 > 3.0, and |2.1| > 1.5, thus it is determined to be a valid step event. The coordinates (150mm, 200mm) and the timestamp 12.5s are recorded to the event set.

[0115] S3.5: Based on the set of local phase step events, perform spatial interpolation and temporal cumulative mapping to transform discrete step events into a continuous spatial distribution matrix, so as to generate a spatiotemporal diagram of debonding intensity that intuitively represents the degree of interface debonding and its spatiotemporal evolution.

[0116] The system receives a set of local phase step events generated in the preceding steps. This set contains discrete timestamps, spatial coordinates, and corresponding phase abrupt change amplitude values. Based on the sensor grid topology of the fireproof window glass surface, a two-dimensional spatial interpolation matrix is ​​constructed to map discrete step event points to regular grid nodes. The Kriging interpolation method is used to calculate the estimated phase abrupt changes in areas without sensor deployment. A variogram model is used to quantify spatial correlation, generating an instantaneous debonding intensity distribution field covering the entire glass surface. The generated instantaneous debonding intensity distribution field undergoes temporal accumulation processing, introducing a forgetting factor mechanism to distinguish between historical damage and current newly formed damage. The cumulative debonding intensity value of each grid cell is updated using an exponentially weighted moving average method to suppress spurious abrupt changes caused by random noise and strengthen the weight of persistent debonding events. The updated cumulative debonding intensity values ​​are normalized and mapped to a range of 0 to 1, forming a standardized debonding intensity index. The standardized debonding intensity indices are stacked in time series to construct a three-dimensional spatiotemporal data cube, where two dimensions represent spatial location and the third dimension represents temporal evolution. A smoothing filter is applied to the three-dimensional spatiotemporal data cube to eliminate high-frequency spatial jitter caused by interpolation errors, generating a continuous and smooth spatiotemporal map of decoupling intensity. Through spatial interpolation and temporal cumulative mapping, the discrete step events from the previous step are transformed into continuous spatiotemporal map data of decoupling intensity, enabling visualization of the degree of interface decoupling and tracking of its evolution.

[0117] For example, a 12×8 fiber optic grating sensor array is deployed on a fireproof window glass with dimensions of 1.2m × 0.8m. When a phase step is detected at t=50s between three adjacent sensor nodes with amplitudes of 0.05rad, 0.06rad, and 0.04rad respectively, the Kriging interpolation method is used, with a range of 0.1m and a nugget effect of 0.01, to calculate the estimated debonding intensity of the surrounding grid nodes. A forgetting factor of 0.95 is introduced to attenuate the historical cumulative intensity before t=50s, and the interpolation result at the current moment is superimposed on the cumulative field. After normalization, the debonding intensity index of the grid cells in the central region increases from 0.2 to 0.65, forming a distinct local bright area. The final generated spatiotemporal map of debonding intensity clearly shows the trend of the debonding region expanding from the edge to the center, with a spatial resolution of 0.1m and a time update frequency of 10Hz, significantly improving the accuracy and continuity of interface damage localization.

[0118] Step S4: The residual phase sequence is recursively reconstructed to extract three nonlinear dynamic indices to construct a three-dimensional phase chaos exponential vector. Specifically, this includes:

[0119] S4.1: Perform phase space reconstruction processing on the residual phase sequence, determine the optimal embedding dimension and time delay based on the delay coordinate method, and map the residual phase sequence into a high-dimensional phase space trajectory vector to construct a phase space state matrix that can completely preserve the dynamic characteristics of the system.

[0120] The residual phase sequence after the intrinsic creep effect has been removed by step S3 is received. This sequence characterizes the dynamic evolution of interface damage and nonlinear disturbance.

[0121] The delayed coordinate method is used to reconstruct the phase space topology of the one-dimensional residual phase time series, mapping the scalar time series into trajectory vectors in the high-dimensional state space, so as to recover the implicit dynamic attractor structure of the system.

[0122] The statistical dependence between the residual phase sequence and its own lagged version is calculated using the mutual information method. The mutual information function curve is plotted, and the time step corresponding to the first local minimum of the mutual information is selected as the optimal time delay to ensure that the independence between the reconstructed components is maximized.

[0123] Based on the analysis of the proportion of false neighbor points in the phase space trajectory under different embedding dimensions, the dimension is determined to be the optimal embedding dimension when the proportion of false neighbor points drops below a preset threshold and no longer decreases significantly with the increase of dimension, so as to ensure the sufficiency of phase space unfolding.

[0124] Based on the determined optimal time delay and optimal embedding dimension, a phase space state matrix is ​​constructed. Each row of the matrix represents the high-dimensional state vector of the system at a certain moment, and the column vectors correspond to the phase observation values ​​at different delay times.

[0125] Through the phase space reconstruction process described above, the one-dimensional residual phase sequence from the previous step is transformed into a high-dimensional phase space state matrix that retains complete dynamic characteristics, thereby achieving a geometrical characterization of the nonlinear features of microcrack evolution and providing a structured data foundation for subsequent recursive graph quantitative analysis.

[0126] S4.2: Based on the phase space state matrix, perform distance threshold determination processing, calculate the Euclidean distance between any two points in the high-dimensional phase space trajectory vector and compare it with a preset recursive threshold to generate a binary recursive adjacency matrix representing the probability of repeated occurrence of the system state.

[0127] S4.3: Perform topological structure quantization on the binary recursive adjacency matrix, statistically analyze the density distribution of recursive points in the matrix to calculate the recursion rate index, and analyze the continuous length of the diagonal structure to calculate the deterministic index, thereby obtaining a set of linear dynamic features characterizing the orderliness of microcrack propagation.

[0128] The binary recursive adjacency matrix is ​​subjected to topological structure quantization processing. The density distribution of recursive points in the matrix is ​​statistically analyzed to calculate the recursion rate index, and the continuous length of the diagonal structure is analyzed to calculate the deterministic index, thereby obtaining a set of linear dynamic features characterizing the orderliness of microcrack propagation.

[0129] Read the binary recursive adjacency matrix generated in the previous steps. The element value of this matrix is ​​1, which means that the distance between the trajectory points in the phase space is less than a preset threshold, and the value is 0, which means that the distance exceeds the threshold. This forms a discrete topology network that describes the reproducibility characteristics of the system state.

[0130] The system iterates through all elements of the matrix and accumulates the count to determine the total number of recursive points with a value of 1. It then combines this count with the matrix dimensions to calculate the proportion of recursive points in the total number of points, generating a global density index that characterizes the stability of the system state.

[0131] Identify the line segments with continuous values ​​of 1 along the main diagonal of the identification matrix and in its parallel directions. These diagonal structures correspond to parallel motion segments of the phase space trajectory during the time evolution process, reflecting the degree of determinism of the system's dynamic behavior.

[0132] The number and length of all diagonal segments with a length greater than the preset minimum threshold of 4 are counted. The probability density function of the length distribution is calculated, and the average diagonal length is extracted as the quantitative basis for the deterministic indicator.

[0133] The determinism is calculated using the following formula. :

[0134]

[0135] in, For length is The probability of the diagonal segment appearing. The minimum diagonal length threshold set to eliminate noise interference. The maximum diagonal length threshold. A value close to 1 indicates that the microcrack propagation has a highly predictable linear pattern, while a decrease in the value indicates an increased tendency towards chaos.

[0136] The calculated recurrence rate RR is combined with the deterministic DET to form a two-dimensional linear dynamic eigenvector. This vector is stripped of nonlinear complexity interference and focuses on characterizing the orderliness and stability of microcrack group evolution.

[0137] By using the above-mentioned topological statistics and probability distribution calculation methods, the binary adjacency matrix of the previous step is transformed into a set of linear dynamic features that characterize the orderliness of microcrack propagation, thus achieving accurate quantitative differentiation of glass damage evolution stages from random noise to structured failure.

[0138] S4.4: Perform shortest path search processing based on the binarized recursive adjacency matrix, use graph theory algorithms to calculate the minimum number of hops between any two recursive state nodes in the recursive graph, and extract the average shortest path length of all node pairs as a nonlinear dynamic scalar to characterize the complexity of the system state transition.

[0139] Based on the binary recursive adjacency matrix generated in the previous steps, an undirected weighted graph structure with phase space trajectory points as nodes is constructed. Elements with a value of 1 in the matrix are mapped to connected edges between nodes in the graph, while elements with a value of 0 indicate that there is no direct connection between nodes.

[0140] Initialize the shortest path distance matrix for all sources, setting the distance between directly connected node pairs to 1, the distance between non-connected node pairs that are not themselves to infinity, and the distance between nodes themselves to themselves to 0.

[0141] The Floyd-Warshall algorithm (a method for finding the shortest path between any two points) is used to iteratively update the shortest hop distance between nodes. For any three nodes i, j, and k, if the path length through node k is less than the currently recorded distance, then the shortest path length from node i to node j is updated.

[0142] Traverse the updated distance matrix and extract the shortest path length values ​​for all valid node pairs. Remove isolated node pairs with infinite distances to ensure the validity of the statistical sample.

[0143] The arithmetic mean of the shortest path lengths for all valid nodes is calculated and used as a nonlinear dynamic scalar to characterize the complexity of the system's state transitions. This scalar reflects the ergodicity and chaos of the phase space trajectory during microcrack evolution.

[0144] By using graph theory shortest path search, the binary topology from the previous step is transformed into an average path length index that quantifies the global connectivity of the system, thus achieving an accurate characterization of the disorder of the evolution of microcrack groups inside the glass.

[0145] S4.5: Perform multidimensional vector encapsulation processing on the linear dynamic feature set and nonlinear dynamic scalar, and arrange and combine the recursion rate index, deterministic index and the average shortest path length in a fixed order to generate a three-dimensional phase chaos index vector characterizing the evolution state of the microcrack group.

[0146] Step S5: Input the debonding intensity spatiotemporal map and the three-dimensional phase chaotic exponential vector into the lightweight graph attention network, and output the global rupture trend probability distribution. Specifically, this includes:

[0147] S5.1: Discretize the monitoring area on the surface of the fireproof window glass to generate a set of grid cells containing multiple independent monitoring units, and map each grid cell to a basic node in the graph structure. At the same time, calculate the Euclidean distance between adjacent nodes based on the spatial geometric position of the grid cells to construct an initial geometric adjacency matrix.

[0148] The S5.1 sub-step aims to discretize the continuous glass monitoring area into basic units for graph computation, construct an initial adjacency matrix that reflects the topological relationships in the physical space, and provide a geometric basis for subsequent fusion of phase synergy.

[0149] The physical dimensional parameters of the fireproof window glass surface and the spatial coordinate distribution data of the miniature fiber optic array sensor are obtained to determine the length and width boundary conditions of the monitoring area.

[0150] A two-dimensional Cartesian coordinate system is established based on the geometric center of the monitoring area. The glass surface is divided into uniform square grid units, and the grid side length is set to an integer multiple of the sensor node spacing to ensure coverage integrity.

[0151] Map the geometric center point of each grid cell to an independent node in the graph structure, assign a unique node identifier, and form a node set V containing A nodes.

[0152] Iterate through any two nodes in the node set V, calculate their Euclidean distance in the two-dimensional plane, and generate an A×A distance matrix D.

[0153] Set a geometric adjacency threshold, which is determined based on the glass stress wave attenuation characteristics and the effective sensing radius of the sensor. It is usually set to 1.5 times the length of the grid diagonal.

[0154] Binarize the distance matrix D. If the distance between nodes is less than or equal to the geometric adjacency threshold and the nodes do not overlap, they are considered geometrically adjacent, and the corresponding adjacency matrix element is set to 1; otherwise, it is set to 0.

[0155] Through the above processing method, the physical space information of the previous step is transformed into an initial geometric adjacency matrix that represents the geometric connection relationship between nodes, realizing the mapping of the monitoring area from the continuous physical domain to the discrete graph topological domain, laying the structural foundation for the subsequent introduction of phase dynamic features.

[0156] For example, a fiber optic grating array with a spacing of 50mm is deployed on a 1000mm × 800mm fireproof window. The glass surface is divided into 320 grid cells (20 × 16), with each cell corresponding to a graph node. The geometric adjacency threshold is set to 70mm (approximately 1.4 times the grid edge length). The Euclidean distance between any two nodes is calculated; if the distance is less than 70mm, it is recorded as 1 in the adjacency matrix; otherwise, it is recorded as 0. This results in a 320 × 320 sparse binary matrix, with an average of 8 geometric neighbors per node. This accurately reflects the local spatial connectivity of the glass surface, ensuring that subsequent message passing mechanisms only occur between physically adjacent nodes, avoiding interference from invalid information over long distances.

[0157] S5.2: Based on the initial geometric adjacency matrix and the corrected instantaneous phase trajectory generated in the previous steps, calculate the dynamic time warping distance of the historical instantaneous phase trajectory sequence between adjacent nodes to quantify the phase coherence, and perform weighted fusion processing on the phase coherence and the initial geometric adjacency matrix to generate the final edge weight set characterizing the spatial topology and state association coupling characteristics.

[0158] Obtain the initial geometric adjacency matrix generated by S5.1 and the corrected instantaneous phase trajectory sequence output by S2 as the input data source for edge weight calculation.

[0159] For each pair of adjacent nodes in the discretized grid on the glass surface, extract the time series fragment of the historical instantaneous phase trajectory at the corresponding position.

[0160] The dynamic time warping method is used to calculate the nonlinear similarity distance of the phase trajectory between adjacent nodes in order to eliminate the phase time axis misalignment interference caused by thermal conduction delay.

[0161] A local cumulative distance matrix is ​​constructed, and the synchronization degree of the phase evolution modes of two nodes is quantified by recursively searching for the path with the minimum cumulative cost.

[0162] Calculate the nodes using the following formula With nodes Dynamic time-warped distance between :

[0163]

[0164] in, , They are nodes and nodes The instantaneous phase trajectory, It is a Euclidean distance metric. To accumulate the minimum cost.

[0165] The calculated dynamic time warp distance is normalized and its reciprocal is mapped to the phase coordination coefficient. The larger the value, the more synchronized the strain response of the two nodes.

[0166] Read the geometric distance weight of the corresponding position in the initial geometric adjacency matrix. This weight is determined by the inverse ratio of the physical Euclidean distance between nodes.

[0167] The fusion ratio coefficient between geometric topology weights and phase synergy weights is set, and is usually calibrated based on the thermal diffusivity of the glass material.

[0168] Perform a weighted fusion operation, multiply the geometric distance weights by the phase synergy coefficients and normalize them to generate the final set of edge weights.

[0169] By using dynamic time warping distance calculation and weighted fusion processing, the geometric adjacency relationship of the previous step is transformed into a final set of edge weights that characterize the coupling properties of spatial topology and state association, thus achieving the expected technical effect of accurate modeling of the propagation path of fracture stress waves.

[0170] S5.3: Map the debonding intensity spatiotemporal map and the three-dimensional phase chaotic exponential vector generated in the previous step to the feature attributes of the basic node, respectively, to construct the node initial feature vector containing interface damage features and nonlinear dynamic features, and load the node initial feature vector and the final edge weight set as input data source into the lightweight graph attention network.

[0171] The debonding intensity spatiotemporal map data matrix and the three-dimensional phase chaos index vector set generated in the previous steps are obtained as the basic input source for node feature construction. For each grid cell node after discretization of the fireproof window glass surface, the local debonding intensity value at its corresponding spatial coordinate position in the debonding intensity spatiotemporal map is extracted. This local debonding intensity value is normalized and mapped to a scalar feature component representing the degree of damage to the interface bonding state. Simultaneously, the three-dimensional phase chaos index vector corresponding to the grid cell node is extracted. This vector contains three nonlinear dynamic indicators: recursion rate, determinism, and shortest path length. Z-score normalization is performed on the three-dimensional phase chaos index vector to eliminate numerical differences between indicators of different dimensions and ensure the consistency of feature distribution. The normalized three-dimensional phase chaos index vector and the normalized debonding intensity scalar are concatenated. A node initial feature vector containing four dimensions is constructed, where the first dimension is the debonding intensity, and the latter three dimensions are the recursion rate, determinism, and shortest path length, respectively. All grid cell nodes are traversed to generate a complete node initial feature matrix.

[0172] The initial node feature matrix is ​​aligned with the final edge weight set generated in the previous step. The aligned node features and edge weights are then encapsulated into a graph data structure object. This graph data structure object is loaded into the input layer interface of the lightweight graph attention network. Through multi-dimensional feature fusion processing, the discrete physical indices from the previous step are transformed into initial node feature vectors containing interface damage and nonlinear dynamic information, achieving a unified representation of multi-source heterogeneous features in the graph topology and providing high-fidelity input for subsequent spatial diffusion modeling.

[0173] For example, a fireproof window glass with dimensions of 1000mm × 1000mm is divided into 100 grid cells of 10 × 10. For the (5,5)th grid cell node, its current debonding intensity value is read from the debonding intensity spatiotemporal graph as 0.75, and maintained at 0.75 after Min-Max normalization. The corresponding three-dimensional phase chaos index vector of this node is read as [0.82, 0.65, 4.2], and converted to [1.2, -0.5, 0.8] after Z-Score normalization. The initial feature vector of the node is generated by concatenation [0.75, 1.2, -0.5, 0.8]. The same operation is performed on all 100 nodes to generate a 100 × 4 initial feature matrix of the nodes. This matrix and the 100 × 100 edge weight matrix are loaded together into the graph attention network. This processing method significantly improves the integrity of the feature input, ensuring that the network can simultaneously perceive the interface debonding and microcrack evolution state.

[0174] S5.4: The multi-head self-attention mechanism in the lightweight graph attention network is used to perform feature transformation processing on the initial feature vector of the node to generate query vector, key vector and value vector, and the similarity between the key vector and query vector of the adjacent node is calculated based on the final edge weight set to generate a normalized attention coefficient matrix.

[0175] The initial feature vectors and final edge weight sets of nodes, received from the lightweight graph attention network, serve as the computational basis for the multi-head self-attention mechanism. A linear projection transformation is performed on the initial feature vector of each grid cell node, generating query vectors, key vectors, and value vectors using a learnable weight matrix, achieving decoupled mapping of the feature space. Based on the adjacency topology defined by the final edge weight set, the key vectors and query vectors of the current node and its first-order neighbors are extracted to construct a local attention computation subgraph. The scaling dot product attention method is used to calculate the similarity score between the query vector and the key vector, resulting in the original attention score. A masking process is performed on the original attention score, forcing the attention weights of non-adjacent nodes to zero to ensure that information transmission strictly follows the physical topology of the glass surface. Softmax normalization is performed on the masked attention score, converting the score into attention coefficients in probability distribution form, representing the influence weights of adjacent nodes on the current node's state.

[0176] A multi-head parallel computing mechanism is introduced to segment the feature vector into multiple independent subspaces, and the aforementioned attention coefficient calculation is performed separately to capture the correlation patterns between the phase chaos index and decoupling intensity in different semantic subspaces. The attention coefficients output by each head are concatenated and linearly fused to generate the final normalized attention coefficient matrix, fully preserving multi-scale feature interaction information. Through feature transformation and weighted calculation using the multi-head self-attention mechanism, the node features from the previous step are transformed into a normalized attention coefficient matrix containing spatial diffusion patterns, achieving efficient aggregation of local phase anomalies into global rupture trends.

[0177] S5.5: Based on the normalized attention coefficient matrix, a weighted aggregation operation is performed on the value vectors of adjacent nodes to realize the spatial diffusion and information fusion of local phase anomaly features along the graph topology, thereby outputting a global breakage trend probability distribution that characterizes the overall breakage trend of the glass.

[0178] Receive the normalized attention coefficient matrix and the value vector of the neighboring nodes generated by the previous step. The value vector contains the interface debonding intensity and the three-dimensional phase chaotic exponential component after feature transformation.

[0179] Based on the weight distribution defined in the normalized attention coefficient matrix, a weighted summation operation is performed on the value vectors of the neighboring nodes of each target node to construct an aggregated representation of local anomalies, resulting in an aggregated feature vector. The aggregated local feature vector is then nonlinearly mapped through a fully connected layer to eliminate dimensional differences and enhance feature representation capabilities.

[0180] A global average pooling operation is performed on the mapped feature vector to fuse the feature information of discrete grid cells into a global breakage trend probability distribution vector that characterizes the overall state of the glass.

[0181] By using weighted aggregation and global pooling, the local attention features from the previous step are transformed into global fracture trend probability distribution data, enabling a system-level quantitative assessment of the evolution from local microcracks to overall fracture risk.

[0182] For example, a 1.5m × 1.0m fireproof window glass was divided into 600 grid nodes (30 × 20). The graph attention network was set to have 2 layers, with a hidden layer dimension of 64. Under heated conditions, the central region nodes experienced a phase chaos exponential jump due to the temperature gradient. The normalized attention coefficients of the central node and its 8 neighboring nodes were calculated to be 0.15, while those of the edge nodes were 0.05. After weighted aggregation, the feature vector magnitude of the central node increased significantly, reflecting a local high-risk state. After global average pooling, the probability value corresponding to the high-risk area in the output fracture trend probability distribution vector reached above 0.85, while that of the low-risk area was below 0.1. This distribution accurately mapped the microcrack propagation path caused by stress concentration inside the glass, verifying the effectiveness of spatial diffusion modeling and significantly improving the sensitivity of identifying early fracture signs.

[0183] Step S6: Utilizing the global breakage trend probability distribution for mapping calculation, a final warning command is generated, and a warning is issued for fireproof window glass breakage based on the final warning command. Specifically, this includes:

[0184] S6.1: Obtain the global fracture trend probability distribution output by the lightweight graph attention network, extract the three-dimensional phase chaos index vector representing the evolution state of the microcrack group, calculate the instantaneous value of the three-dimensional phase chaos index vector at the current time step as the current value of the phase chaos index, and perform time difference operation on the current value of the phase chaos index to obtain the phase chaos index change acceleration representing the evolution rate.

[0185] S6.2: Call the accelerated aging test calibration dataset pre-stored in the edge computing unit, construct a bivariate input space based on the current value of the phase chaos index and the acceleration of the phase chaos index change, and use the piecewise linear interpolation method to fit the discrete sample points in the accelerated aging test calibration dataset to generate an empirical mapping function describing the complete evolution cycle from the initial microcrack to the macroscopic fracture.

[0186] The accelerated aging test calibration dataset, which is pre-stored in the edge computing unit storage medium, is invoked. This dataset contains phase chaos index sample points and their corresponding timestamps for the entire process of glass from initial microcracks to macroscopic fracture under different temperature gradients and stress loading conditions.

[0187] A two-dimensional input feature space is constructed based on the current value of the phase chaos index and the acceleration of the phase chaos index change. The current value of the phase chaos index is used as the variable on the horizontal axis, and the acceleration of the phase chaos index change is used as the variable on the vertical axis, forming a discrete feature distribution matrix.

[0188] The discrete feature distribution matrix is ​​subjected to Delaunay triangulation. Based on the nearest neighbor principle, three adjacent sample points are connected to form a set of non-overlapping triangular meshes, ensuring that each triangle does not contain other sample points, so as to establish the topological basis for local linear interpolation.

[0189] For the position of the current state point to be predicted in the two-dimensional input feature space, the triangular mesh set is traversed, and the target triangular region to which the point belongs is determined by the centroid coordinate method. The phase chaos index value, change acceleration value and corresponding remaining lifetime true label of the three vertices of the target triangle are obtained.

[0190] Within the target triangular region, a piecewise linear interpolation method is applied. Based on the Euclidean distance weights from the point to be predicted to the three vertices, the remaining lifetime labels of the three vertices are weighted and linearly combined to calculate the theoretical remaining lifetime estimate corresponding to the current state point.

[0191] Repeat the above interpolation process to cover all key evolution stage nodes recorded in the accelerated aging test, generating a set of mapping relationship sequences composed of discrete estimates.

[0192] The least squares method is used to perform curve fitting on the mapping relationship sequence to construct a continuous empirical mapping function that describes the nonlinear relationship between the phase chaos index and its changing acceleration and the remaining lifetime.

[0193] By using piecewise linear interpolation and curve fitting, the discrete calibration data from the previous step is transformed into a continuously differentiable empirical mapping function, achieving an accurate mapping from multidimensional phase characteristics to the rupture time domain, thus solving the problem that traditional single-threshold early warning cannot quantify the remaining safe time.

[0194] For example, the edge computing unit loads an accelerated aging calibration dataset containing 500 samples, setting the phase chaos index threshold to 0.85. When the real-time monitoring shows the current value of the phase chaos index is 0.72 and the acceleration is 0.03 / s², the system locates the target triangle in two-dimensional space formed by sample points A (0.70, 0.02, 120s), B (0.75, 0.04, 90s), and C (0.71, 0.05, 100s). Using the centroid coordinate method, weight coefficients of 0.5, 0.3, and 0.2 are calculated, and the theoretical remaining lifetime is determined to be 111 seconds. After fitting the full dataset, the root mean square error of the generated empirical mapping function on the test set is reduced to less than 3.5 seconds, significantly improving the accuracy and reliability of the failure countdown prediction.

[0195] S6.3: Input the current value of the phase chaos index and the acceleration of the phase chaos index change into the empirical mapping function to perform a nonlinear mapping transformation, calculate the fracture probability scalar field characterizing the failure probability of each grid unit on the glass surface, and map the fracture probability scalar field into a fracture probability heat map in a visualization format to intuitively display the spatial distribution of high-risk areas.

[0196] The system receives the global rupture trend probability distribution output by the lightweight graph attention network, and extracts the three-dimensional phase chaotic exponential vector corresponding to each grid node and the changing acceleration obtained by its time difference.

[0197] A bivariate nonlinear mapping function is constructed, which takes the current value of the phase chaos exponent and the changing acceleration as input variables and maps them to the scalar space of the rupture probability.

[0198] The instantaneous fracture probability of the x-th grid cell is calculated using the following empirical mapping formula:

[0199]

[0200] in, Let x be the breakage probability of the x-th grid cell. The current value of the phase chaos exponent. For the phase chaotic exponential change acceleration, , These are the feature weight coefficients. This is a bias term, and its coefficients are determined by fitting accelerated aging test data.

[0201] Traverse all discretized grid cells on the glass surface, substitute their respective chaos exponents and acceleration values, and perform point-by-point probability calculations.

[0202] The calculated scalar probability values ​​are mapped to the RGB color space, with low probability ranges designated as cool tones and high probability ranges as warm tones, generating a two-dimensional fracture probability heatmap.

[0203] Spatial smoothing filtering is applied to the heatmap to eliminate local probability spikes caused by sensor noise and preserve the spatial continuity of macroscopic risk areas.

[0204] By using a nonlinear mapping transformation process, the chaotic dynamic indicators from the previous step are converted into intuitive and visual heatmap data of rupture probability, achieving the expected technical effect of precise spatial positioning and visualization of high-risk areas.

[0205] S6.4: Based on the inversion logic of the empirical mapping function, the theoretical time span required for the phase chaos index to reach the critical rupture threshold is traced using the current value of the phase chaos index, and the theoretical time span is dynamically corrected by combining the acceleration of the phase chaos index change, so as to generate a rupture countdown interval estimate with confidence boundary to quantify the remaining safe duration.

[0206] The current value and acceleration of the phase chaos index calculated in S6.1 are obtained as the initial state variables for the fracture countdown estimation. The accelerated aging test calibration dataset pre-stored in the edge computing unit is invoked. This dataset records the time series trajectory of the phase chaos index evolving from its initial value to the critical fracture threshold under different temperature stress coupling conditions. Based on the initial state variables, historical sample clusters with similar evolution paths are retrieved from the calibration dataset, and the reference evolution curve is extracted using the weighted nearest neighbor method. A time-reversal mapping model is constructed, starting from the current phase chaos index value, tracing back along the reference evolution curve to the fracture threshold point to calculate the theoretical remaining time span. The acceleration of the phase chaos index is introduced as a dynamic correction factor to evaluate the deviation between the current damage evolution rate and the historical average rate. When the acceleration is positive and exceeds a preset fluctuation threshold, the damage is determined to be in an accelerated expansion stage, and a nonlinear compression coefficient is applied to the theoretical remaining time span; when the acceleration is negative or close to zero, the damage is determined to be in a stable or stagnant stage, and the theoretical remaining time span is kept unchanged or a small time extension compensation is applied. The corrected upper and lower limits of the rupture countdown interval are calculated using the following formula:

[0207]

[0208]

[0209] in, This is the upper limit of the countdown interval. This is the lower limit of the countdown interval. For the theoretical remaining time span, For the phase chaotic exponential change acceleration, and These are correction coefficients set based on the confidence boundary. The calculated... and The combined data forms a countdown interval estimate with confidence boundaries. Through the aforementioned dynamic correction mechanism, the static theoretical time span is transformed into a countdown interval that adapts to real-time operating conditions, enabling a quantitative output of the remaining safe time and significantly improving the accuracy of the time dimension of the early warning.

[0210] S6.5: Integrate the rupture probability heatmap with the rupture countdown interval estimate, encapsulate it into a final warning instruction that includes spatial risk location and time trend prediction dimensions, and send the final warning instruction to the fireproof window monitoring terminal to trigger a multi-level audible and visual alarm or linkage shutdown mechanism.

[0211] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0212] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0213] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for early warning of fire-resistant window glass breakage based on strain mutation, characterized in that, Specifically, it includes: S1: Dynamic response sequences are collected using sensors deployed on fireproof window glass; S2: Process the dynamic response sequence to obtain the intrinsic mode function component set, calculate and transform the intrinsic mode function component set to obtain the instantaneous phase trajectory sequence; S3: Based on the instantaneous phase trajectory sequence, a theoretical phase reference curve is constructed. The residual phase sequence is calculated by comparing the instantaneous phase trajectory sequence with the theoretical phase reference curve. The residual phase sequence is then processed to obtain a spatiotemporal diagram of debonding intensity. S4: Perform recursive graph reconstruction on the residual phase sequence and extract three nonlinear dynamic indices to construct a three-dimensional phase chaos exponential vector; S5: Input the debonding strength spatiotemporal map and the three-dimensional phase chaotic exponential vector into the lightweight graph attention network, and output the global fracture trend probability distribution; S6: Utilize the global breakage trend probability distribution to perform mapping calculations, generate a final warning command, and issue a warning for the breakage of the fireproof window glass based on the final warning command.

2. The method for early warning of fireproof window glass breakage based on strain mutation according to claim 1, characterized in that, S2 specifically includes: Extreme point detection and envelope fitting are performed on the dynamic response sequence. The upper and lower envelopes are constructed based on the cubic spline interpolation method. The mean sequence of the upper and lower envelopes is calculated and the mean sequence is subtracted from the dynamic response sequence to obtain the first-order residual strain component containing high-frequency oscillation components. The first-level residual strain components are subjected to iterative screening convergence judgment processing. Based on the preset screening termination criterion, the envelope mean removal operation is repeatedly performed until the preset condition is met, and the set of intrinsic mode function components is obtained. The complex analytic signal representation is obtained by transforming and combining the set of intrinsic mode function components. The complex analytic signal representation is subjected to arctangent phase calculation processing. The principal value of the phase angle is calculated based on the ratio of the imaginary part data to the real part data of the complex analytic signal representation. The phase jump ambiguity is eliminated by the phase expansion method, thereby extracting the instantaneous phase trajectory sequence.

3. The method for early warning of fireproof window glass breakage based on strain mutation according to claim 2, characterized in that, The transformation and combination of the intrinsic mode function component set to obtain the complex analytic signal representation includes: For each eigenmode function component in the set of eigenmode function components, perform Hilbert transform processing, construct an orthogonal conjugate signal sequence using Cauchy principal value integral, and combine the eigenmode function component with the orthogonal conjugate signal sequence to form an analytic signal vector in the complex domain, thereby obtaining a complex analytic signal representation.

4. The method for early warning of fireproof window glass breakage based on strain mutation according to claim 3, characterized in that, The complex analytic signal representation includes instantaneous amplitude and instantaneous phase information.

5. The method for early warning of fireproof window glass breakage based on strain mutation according to claim 1, characterized in that, S3 specifically includes: Based on the instantaneous phase trajectory sequence, a theoretical phase reference curve is constructed using the generalized rheological model parameter inversion method. Perform point-by-point difference operations on the instantaneous phase trajectory sequence and the theoretical phase reference curve to generate a residual phase sequence; The residual phase sequence is subjected to local statistical feature extraction processing based on the sliding window mechanism, and the standard deviation and skewness values ​​of the residual phase sequence within the sliding window are obtained respectively to generate a joint statistical feature vector. A set of local phase step events is obtained by performing a double threshold logic decision on the joint statistical feature vector; Based on the set of local phase step events, spatial interpolation and temporal cumulative mapping are performed to transform discrete local phase step events into a continuous spatial distribution matrix to generate a spatiotemporal map of debonding intensity.

6. The method for early warning of fireproof window glass breakage based on strain mutation according to claim 5, characterized in that, The theoretical phase reference curve characterizes the intrinsic creep behavior of the material in a damage-free state, and the residual phase sequence includes information on interface damage and nonlinear perturbation.

7. The method for early warning of fireproof window glass breakage based on strain mutation according to claim 5, characterized in that, The process of performing a double-threshold logic decision on the joint statistical feature vector yields a set of local phase step events, including: The joint statistical feature vector is subjected to dual-threshold logic decision processing. Based on the preset standard deviation threshold and skewness threshold, the timestamps and spatial coordinates of local phase step events that meet the mutation conditions are located to generate a set of local phase step events that mark the location of interface debonding.

8. The method for early warning of fireproof window glass breakage based on strain mutation according to claim 1, characterized in that, The three nonlinear dynamic metrics include recursion rate, determinism, and shortest path length.

9. The method for early warning of fireproof window glass breakage based on strain mutation according to claim 1, characterized in that, The three-dimensional phase chaotic exponential vector characterizes the quantitative features of the microcrack group evolution state.

10. The method for early warning of fireproof window glass breakage based on strain mutation according to claim 1, characterized in that, The final early warning command includes a rupture probability heatmap and a rupture countdown interval estimate.

Images