Lstm-driven dynamic prediction method and system for operational resilience of a utility tunnel

Abstract

This invention discloses an LSTM-driven dynamic prediction method and system for the operational resilience of integrated utility tunnels, relating to the field of deep learning. The method includes: acquiring vibration acceleration time histories, extracting peak acceleration and vibration energy accumulation time history curves, and generating a dynamic angle sequence; correcting the baseline vulnerability curve to obtain a dynamic damage probability curve, substituting the peak acceleration and vibration energy accumulation values ​​of the cross-section into the curve, and outputting a time-varying damage index sequence; identifying critical cross-sections with abrupt changes in the damage index, calculating the average damage index using the critical cross-section as a dividing point, and determining functional bottleneck sections; obtaining the average damage index and the redundancy of subsequent alternative paths, calculating the overall functional degradation rate of the utility tunnel, and obtaining the functional recovery multiple; calculating the initial recovery time, inputting it into a pre-trained dynamic time-series prediction model, and outputting the corrected final recovery time. This invention solves the problem of low prediction accuracy and high misjudgment rate in the prediction of the operational resilience of integrated utility tunnels due to the reliance on static experience for rough estimation in existing technologies.

Description

Technical Field

[0001] This invention relates to the field of deep learning, specifically to an LSTM-driven method and system for dynamic prediction of the operational resilience of integrated utility tunnels. Background Technology

[0002] Integrated utility tunnels centrally house various pipelines such as water supply, power supply, and communications within underground tunnels, serving as a crucial infrastructure for ensuring the normal operation of a city. Extreme natural disasters such as earthquakes pose a serious threat to the structural safety and functional integrity of integrated utility tunnels. Damage to these tunnels during an earthquake can lead to widespread water supply disruptions, power outages, and communication failures, with significant impacts and considerable difficulties in recovery.

[0003] Currently, traditional methods for predicting seismic damage to utility tunnels mainly rely on static parameters for rough estimation, lacking dynamic quantitative analysis of the structural response, damage evolution, and functional recovery process of the tunnels. This leads to technical problems such as low accuracy and high misjudgment rate in predicting the operational resilience of integrated utility tunnels. Summary of the Invention

[0004] This invention provides an LSTM-driven dynamic prediction method and system for the operational resilience of integrated utility tunnels, which solves the problem of low prediction accuracy and high misjudgment rate in the existing technology that relies on static experience for rough estimation.

[0005] In view of the above problems, the present invention provides an LSTM-driven method and system for dynamic prediction of the operational resilience of integrated utility tunnels.

[0006] In a first aspect, the present invention provides an LSTM-driven dynamic prediction method for the operational resilience of integrated utility tunnels, the method comprising: The vibration acceleration time history of each monitoring section along the integrated utility tunnel during strong earthquake is continuously collected. The peak acceleration and cumulative vibration energy time history curves of each section are extracted. The time difference sequence of vibration waveform arrival between adjacent sections is calculated, and the dynamic angle sequence between the direction of seismic wave propagation and the axis of the utility tunnel is generated. The baseline vulnerability curve of each section is corrected according to the dynamic angle sequence to obtain the dynamic damage probability curve. Then, the peak acceleration and cumulative vibration energy of each section are substituted into the dynamic damage probability curve to output the time-varying damage index sequence of each section. Scan the time-varying damage index sequence along the pipe gallery route, identify the critical section where the damage index changes abruptly, and calculate the average damage index of each independent section with the critical section as the dividing point. The section corresponding to the maximum average damage index is determined as the functional bottleneck section. Obtain the average damage index of the functional bottleneck section and the traffic redundancy of the subsequent alternative paths, calculate the overall functional degradation rate of the utility tunnel and obtain the functional recovery multiple; Based on the functional recovery factor and the unit time repair capability parameter, the initial recovery time is calculated. Then, the time-varying damage index sequence and the initial recovery time are input into the pre-trained dynamic time series prediction model, and the final recovery time is output after post-earthquake time recursion correction.

[0007] Secondly, this invention provides an LSTM-driven dynamic prediction system for the operational resilience of integrated utility tunnels, the system comprising: The angle sequence generation module is used to obtain the vibration acceleration time history of each monitoring section along the integrated utility tunnel during strong earthquake, extract the peak acceleration and vibration energy accumulation time history curve of each section, calculate the time difference sequence of vibration waveform arrival between adjacent sections, and generate a dynamic angle sequence between the direction of seismic wave propagation and the axis of the utility tunnel. The damage probability curve construction module is used to correct the reference vulnerability curve of each section according to the dynamic angle sequence to obtain the dynamic damage probability curve. Then, the peak acceleration and cumulative vibration energy of each section are substituted into the dynamic damage probability curve to output the time-varying damage index sequence of each section. The functional bottleneck section acquisition module is used to scan the time-varying damage index sequence along the pipe gallery, identify the critical section where the damage index changes abruptly, and calculate the average damage index of each independent section with the critical section as the dividing point, and determine the section corresponding to the maximum average damage index as the functional bottleneck section. The function recovery multiple calculation module is used to obtain the average damage index of the functional bottleneck section and the traffic redundancy of the subsequent alternative path, calculate the overall functional degradation rate of the utility tunnel and obtain the function recovery multiple. The recovery time generation module is used to calculate the initial recovery time based on the functional recovery multiple and the unit time repair capability parameter. Then, the time-varying damage index sequence and the initial recovery time are input into a pre-trained dynamic time series prediction model, and the final recovery time is output after post-earthquake time recursion correction. The final recovery time serves as a key decision-making basis, synchronously driving the priority ranking of repair resource scheduling and the dynamic adjustment of emergency response level, ensuring that limited resources are accurately allocated to the bottleneck sections with the most significant functional degradation and the highest requirements for repair timeliness.

[0008] One or more technical solutions provided in this invention have at least the following technical effects or advantages: This invention provides an LSTM-driven dynamic prediction method and system for the operational resilience of integrated utility tunnels. The system constructs a dynamic prediction process for operational resilience through modules for angle sequence generation, damage probability curve construction, functional bottleneck section acquisition, functional recovery multiple calculation, and recovery time generation. First, it acquires the vibration acceleration time history and extracts the peak acceleration and vibration energy accumulation time history curves, while simultaneously generating a dynamic angle sequence to capture time-domain accumulation characteristics in real time. Second, it corrects the baseline vulnerability curve based on the dynamic angle sequence to obtain a dynamic damage probability curve and outputs a time-varying damage index sequence, achieving multi-parameter damage assessment and improving the accuracy and adaptability of structural damage prediction. Third, it scans the time-varying damage index sequence to identify critical sections and divide them into independent segments, locating the most severely damaged functional bottleneck segments. Finally, it organically combines local bottleneck damage with system topological redundancy to quantify the system's functional retention capability under damage conditions. Finally, the initial recovery time is calculated, and the time-varying damage index sequence and the initial recovery time are input into the pre-trained LSTM dynamic time series prediction model. The final recovery time, which is corrected by post-earthquake time extrapolation, is output. This model can utilize the nonlinear learning ability of deep learning models to learn about the spatial distribution pattern of damage, providing reliable technical support for post-earthquake resilience assessment and rapid recovery decision-making of integrated utility tunnels. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating the LSTM-driven dynamic prediction method for the operational resilience of integrated utility tunnels provided by this invention. Figure 2 This is a logical schematic diagram of the LSTM-driven dynamic prediction method for the operational resilience of integrated utility tunnels provided by the present invention.

[0010] Figure 3 This is a schematic diagram of the structure of the LSTM-driven integrated utility tunnel operation resilience dynamic prediction system provided by the present invention.

[0011] In the attached diagram, the components represented by each number are as follows: Module 11 for generating angle sequence, module 12 for constructing damage probability curve, module 13 for obtaining functional bottleneck section, module 14 for calculating functional recovery multiple, and module 15 for generating recovery time. Detailed Implementation

[0012] This invention provides an LSTM-driven dynamic prediction method and system for the operational resilience of integrated utility tunnels, which solves the problem of low prediction accuracy and high misjudgment rate in the existing technology that relies on static experience for rough estimation.

[0013] The present invention will now be described in detail with reference to the accompanying drawings.

[0014] Example 1, as Figure 1 , Figure 2 As shown, this invention provides an LSTM-driven dynamic prediction method for the operational resilience of integrated utility tunnels, the method comprising: S10: Obtain the vibration acceleration time history of each monitoring section along the integrated utility tunnel during strong earthquake, extract the peak acceleration and vibration energy accumulation time history curve of each section, calculate the time difference sequence of vibration waveform arrival between adjacent sections, and generate the dynamic angle sequence between the direction of seismic wave propagation and the axis of the utility tunnel. In this embodiment of the invention, peak ground acceleration (PGA) is the maximum instantaneous value of the synthetic acceleration modulus reached in the seismic acceleration time history, representing the peak intensity of the seismic action; the seismic energy accumulation time history curve is the curve of the cumulative increase of seismic ground energy over time obtained by integrating the square of the instantaneous synthetic acceleration from the moment of seismic onset, reflecting the continuous energy input of the seismic action; the waveform arrival time difference is the difference in time between the propagation of the same seismic wavefront to two adjacent monitoring sections, calculated through waveform cross-correlation analysis; the dynamic angle sequence is the instantaneous value or short-time average value of the angle between the seismic wave propagation direction vector and the direction vector of the central axis of the pipe gallery, arranged in the order of the sections, reflecting the spatiotemporal changes of the seismic wave incident direction.

[0015] Specifically, under strong earthquake conditions, continuously recorded vibration acceleration data are read from the monitoring sections. The data from each section constitutes a vibration acceleration time history. The peak acceleration of each monitoring section is generated based on the vibration acceleration time history. Starting from the moment of vibration onset, the square of the instantaneous composite acceleration is integrated over time to obtain the vibration energy accumulation time history curve. Subsequently, the time difference sequence of vibration waveform arrival between adjacent sections is calculated, and the angle between the seismic wave propagation direction and the tunnel axis is obtained. All angles are arranged in section order to generate a dynamic angle sequence.

[0016] Step S10 in the method provided in this embodiment of the invention includes: At each monitoring section, acceleration values ​​in three orthogonal directions are continuously recorded at a fixed sampling frequency. The three orthogonal directions are the horizontal direction parallel to the extension direction of the pipe gallery, the vertical direction perpendicular to the extension direction of the pipe gallery, and the vertical direction. The three acceleration values ​​at the same moment are combined into one data point, and all data points are arranged in chronological order to generate the original vibration acceleration sequence corresponding to each monitoring section. Baseline drift correction is performed on the original vibration acceleration sequence to obtain the corrected vibration acceleration time history; From the corrected vibration acceleration time history, the composite modulus of the acceleration values ​​in three orthogonal directions at each sampling moment is calculated to obtain the instantaneous composite acceleration value at the corresponding moment. Iterate through all the instantaneous composite acceleration values ​​in the vibration acceleration time history and take the value with the largest value as the peak acceleration of the cross section; Starting from the moment the vibration begins, the instantaneous synthetic acceleration value is accumulated over time, and the accumulated results are recorded in chronological order to generate the vibration energy accumulation time history curve of the cross section.

[0017] In this embodiment of the invention, firstly, at each monitoring section, acceleration values ​​in three orthogonal directions, such as the X, Y, and Z directions, are continuously recorded at a fixed sampling frequency to capture the components of ground motion in different directions. The three orthogonal directions are a horizontal direction parallel to the extension direction of the utility tunnel, a horizontal direction perpendicular to the extension direction of the utility tunnel, and a vertical direction.

[0018] Secondly, the three acceleration values ​​at the same sampling time are combined into a single data point (x, y, z), and all data points are arranged in chronological order to generate the original vibration acceleration sequence corresponding to each monitoring section.

[0019] Baseline drift correction is performed on the original vibration acceleration sequence to obtain the corrected vibration acceleration time history. Baseline drift is a phenomenon in which the zero line in the acceleration time history slowly shifts over time, usually caused by sensor thermal noise, analog-to-digital conversion error, or permanent ground displacement.

[0020] Specifically, firstly, a stable segment of data prior to the strong earthquake is identified, and the average acceleration in each direction within this segment is used as the static zero-point offset. Then, the offset in the corresponding direction is subtracted from each acceleration value in the entire sequence to obtain a preliminary debiased sequence. Next, a high-pass filter is used to extract and subtract the slowly changing trend drift component from the preliminary debiased sequence to obtain the corrected earthquake acceleration time history.

[0021] Next, for the corrected vibration acceleration time history, the composite modulus of the acceleration values ​​in the three orthogonal directions at each sampling time is calculated to obtain the instantaneous composite acceleration value at the corresponding time. Specifically, for each sampling time, the corrected acceleration values ​​in the three directions are squared and summed, and the square root of the sum is taken to obtain the composite modulus.

[0022] Furthermore, all instantaneous composite acceleration values ​​in the vibration acceleration time history are traversed, and the value with the largest value is taken as the peak acceleration of the cross section.

[0023] Finally, starting from the moment the vibration begins, the instantaneous composite acceleration value is accumulated over time, that is, the integral of the instantaneous composite acceleration value over time from the moment the vibration begins to the moment it ends is calculated. For discrete data, the trapezoidal rule can be used to integrate the instantaneous composite acceleration value. The trapezoidal rule is a numerical integration method used to estimate the value of the definite integral. Then, the accumulated results are sorted and recorded in chronological order, and plotted as a curve to generate a cross-sectional vibration energy accumulation time history curve, where the horizontal axis represents time, the vertical axis represents the instantaneous composite acceleration value, and the area represents the vibration energy accumulation at the corresponding time.

[0024] The method provided in this embodiment of the invention includes baseline drift correction of the original vibration acceleration sequence to obtain the corrected vibration acceleration time history, comprising: The stable recording segment before the start of the strong earthquake event is identified from the original vibration acceleration sequence. The static zero-point offset in each orthogonal direction is determined based on the stable recording segment. The strong earthquake event is defined by the instantaneous synthetic acceleration value of any monitoring section exceeding the trigger threshold. The trigger threshold is determined based on the background environmental vibration level of the site where the utility tunnel is located. Subtract the static zero-point offset in the corresponding direction from each acceleration value in the original vibration acceleration sequence to obtain a preliminary de-biasing sequence; The preliminary debiased sequence is subjected to trend term separation processing to extract the trend drift component that changes slowly over time; The trend drift component is subtracted from the initial debiased sequence to obtain the corrected acceleration value at each sampling time. All corrected acceleration values ​​are arranged in chronological order of sampling time to generate the corrected vibration acceleration time history.

[0025] In this embodiment of the invention, firstly, a stable recording segment before the onset of a strong earthquake event is identified from the original vibration acceleration sequence, such as the first 200 sampling points. Based on this stable recording segment, the static zero-point offset in each orthogonal direction is determined; that is, the static zero-point offset in three directions is obtained by calculating the arithmetic mean of the values ​​on the stable recording segment. The static zero-point offset is the fluctuation value of the original vibration acceleration in the orthogonal direction before the strong earthquake event occurs. A strong earthquake event is defined by the instantaneous synthetic acceleration value at any monitoring section exceeding a trigger threshold. The trigger threshold is a preset acceleration amplitude determined based on the background environmental vibration level of the site where the utility tunnel is located.

[0026] Secondly, the static zero-point offset in the corresponding direction is subtracted from each acceleration value in the original vibration acceleration sequence to obtain the preliminary debiased sequence. The value in the preliminary debiased sequence is equal to each acceleration value in the original vibration acceleration sequence minus the static zero-point offset.

[0027] Next, the preliminary debiased sequence undergoes trend term separation processing to extract the slowly changing trend drift component over time. Specifically, a low-pass filter can be used to remove high-frequency vibration components from the preliminary debiased sequence, retaining the low-frequency trend component as the trend drift component. The trend drift component is a non-periodic, slowly changing low-frequency error, typically caused by permanent ground displacement or the residual effect of instrument zero drift.

[0028] Further, the trend drift component is subtracted from the initial debiased sequence to obtain the corrected acceleration value at each sampling time. Corrected acceleration value = Initial debiased sequence value - Trend drift component. The corrected acceleration value is used to eliminate systematic baseline errors from all known sources.

[0029] Finally, all corrected acceleration values ​​are arranged in chronological order of sampling time to generate the corrected vibration acceleration time history.

[0030] Step S10 in the method provided in this embodiment of the invention further includes: From the vibration acceleration time history of each monitoring section, the instantaneous synthetic acceleration values ​​are extracted and arranged in order of change with the sampling time, which serves as the vibration waveform time series data corresponding to the section. Select the vibration waveform time series data of two adjacent monitoring sections, compare the waveform of one section at different time offsets with the waveform of the other section, and record the offset corresponding to the highest waveform matching degree as the time difference of the vibration waveform from the near source section to the far source section. According to the extension direction of the utility tunnel, the time differences calculated for each pair of adjacent sections are arranged sequentially to generate a time difference sequence; The propagation velocity of seismic waves in the surrounding soil and rock medium of the utility tunnel and the time difference sequence are obtained. The propagation distance of seismic waves between adjacent sections is determined. Then, based on the geometric correspondence between the propagation distance and the straight-line distance between adjacent sections, the angle between the propagation direction of the seismic waves and the axis of the utility tunnel is calculated. Connect the included angles corresponding to all adjacent cross sections according to the cross section arrangement order to generate the dynamic included angle sequence.

[0031] In this embodiment of the invention, firstly, the instantaneous synthetic acceleration value is extracted from the vibration acceleration time history of each monitoring section and regarded as the vibration waveform time series data corresponding to that section. The values ​​are arranged in the order of the changes in the sampling time as the vibration waveform time series data corresponding to the section.

[0032] Secondly, the seismic waveform time series data of two adjacent monitoring sections are selected. The waveform of one section at different time offsets is compared with the waveform of the other section. The offset corresponding to the highest degree of waveform matching is recorded as the time difference of the seismic waveform propagating from the near-source section to the far-source section, that is, the time taken for the seismic wave to travel the distance between the two sections. Here, the time offset is the length of time by which the seismic waveform of a section is shifted forward or backward along the time axis.

[0033] Next, according to the extension direction of the utility tunnel, the time differences calculated for each pair of adjacent cross sections are arranged sequentially to generate a time difference sequence.

[0034] Furthermore, the propagation velocity and time difference sequence of seismic waves in the surrounding rock and soil medium of the pipe gallery are obtained. The propagation distance of seismic waves between adjacent sections is calculated based on the product of the propagation velocity and time difference sequence of seismic waves in the surrounding rock and soil medium of the pipe gallery. The propagation distance of seismic waves between adjacent sections = propagation velocity × each time difference value in the time difference sequence.

[0035] Based on the geometric correspondence between the propagation distance and the straight-line distance between adjacent sections, the angle between the seismic wave propagation direction and the tunnel axis is obtained. Dividing the straight-line distance S1 between adjacent sections by the propagation distance S2 and taking the inverse cosine value yields the angle θ between the seismic wave propagation direction and the tunnel axis. The propagation velocity of seismic waves in the surrounding rock and soil medium of the utility tunnel was obtained through wave velocity tests conducted during the site geological survey prior to the construction of the utility tunnel.

[0036] Finally, the included angles corresponding to all adjacent cross-sections are connected according to the cross-section arrangement order to generate a dynamic included angle sequence.

[0037] In this embodiment of the invention, acceleration data in three orthogonal directions of the pipe gallery cross-section are collected and instantaneous synthetic acceleration is calculated to fully capture the vibration vector characteristics; then, peak acceleration is extracted to characterize the ground motion intensity, and a vibration energy accumulation time history curve is constructed to provide basic data for damage probability mapping and cumulative effect assessment, ensuring the accuracy of subsequent time-varying damage index calculation.

[0038] S20: Correct the reference vulnerability curve of each section according to the dynamic angle sequence to obtain the dynamic damage probability curve, and then substitute the peak acceleration and vibration energy accumulation value of each section into the dynamic damage probability curve to output the time-varying damage index sequence of each section. In this embodiment of the invention, the baseline vulnerability curve is a function curve describing the relationship between the probability of a structure reaching or exceeding a certain limit damage state and the seismic intensity parameter under a standard reference incident angle; the dynamic damage probability curve is the damage probability curve in the current seismic wave propagation direction after correction by the actual incident angle; the time-varying damage index is a dimensionless quantitative index with a value range of 0 to 1, used to comprehensively characterize the degree of physical damage accumulated in the cross-section of the utility tunnel during the entire process of a strong earthquake, where 0 represents no damage and 1 represents complete failure.

[0039] Specifically, the incident angle of the cross-section is extracted from the corresponding dynamic angle sequence. Then, vulnerability curves at multiple discrete angles are accessed from the reference vulnerability curve library. A linear interpolation method is used to interpolate between the reference curves at adjacent angles based on the incident angle, generating a dynamic damage probability curve. Next, the extracted peak acceleration of the cross-section is substituted into the dynamic damage probability curve to calculate the reference damage probability. Simultaneously, the total accumulated energy at the endpoint is extracted from the vibration energy accumulation time history curve to calculate the energy accumulation factor. Finally, the time-varying damage index is calculated to obtain the time-varying damage index sequence.

[0040] Step S20 in the method provided in this embodiment of the invention includes: For each monitoring section, the included angle value corresponding to each section is extracted from the dynamic included angle sequence and used as the incident angle parameter; Based on the incident angle parameter, the pre-calibrated baseline vulnerability curve is subjected to angle interpolation to generate dynamic damage probability curves for each section under the incident angle, wherein the dynamic damage probability curve represents a damage probability mapping relationship with peak acceleration as the independent variable. Substitute the peak acceleration of each section into the dynamic damage probability curve, read the corresponding probability value, and use it as the baseline damage probability. Extract the cumulative energy value at the end point from the cumulative vibration energy time history curve of the cross section, and calculate the ratio of the cumulative energy value to the unit reference energy as the energy accumulation factor. The unit reference energy is the reference vibration energy value of the pipe gallery cross section under the elastic limit state, which is obtained by vibration table test calibration. Based on the baseline damage probability and the energy accumulation factor, the time-varying damage index of each section under the current strong earthquake is determined, wherein the value of the time-varying damage index ranges from 0 to 1, and the larger the value, the more severe the damage. According to the arrangement order of the monitoring sections along the pipe gallery, the time-varying damage indices of all sections are arranged sequentially to generate the time-varying damage index sequence.

[0041] In this embodiment of the invention, firstly, for each monitoring section, the included angle value corresponding to each section is extracted from the dynamic included angle sequence as the incident angle parameter.

[0042] Secondly, based on the incident angle parameter, the pre-calibrated baseline vulnerability curve is interpolated to generate dynamic damage probability curves for each section at the incident angle. These dynamic damage probability curves represent the damage probability mapping relationship with peak acceleration as the independent variable. The pre-calibrated baseline vulnerability curves are vulnerability curves pre-established based on the incident angle of each section in the historical records of the integrated utility tunnel, through numerical simulation or shaking table tests, and are used to reflect the vulnerability degree of the sections.

[0043] Specifically, the incident angle in the baseline vulnerability curve is interpolated. After the incident angle interpolation is completed, the interpolated incident angle is aligned with the baseline vulnerability curve according to the correspondence between peak acceleration and incident angle to obtain the dynamic damage probability curve under the incident angle.

[0044] Next, the peak acceleration of each section is substituted into the dynamic damage probability curve, and the corresponding probability value is read as the baseline damage probability.

[0045] Furthermore, the cumulative energy value at the endpoint is extracted from the vibration energy accumulation time history curve of the cross-section, and the ratio of the cumulative energy value to the unit reference energy is calculated as the energy accumulation factor. The energy accumulation factor = cumulative energy value / unit reference energy, indicating the degree of energy accumulation relative to the unit reference energy. The unit reference energy is the reference vibration energy value of the pipe gallery cross-section under the elastic limit state, obtained through vibration table testing.

[0046] Furthermore, based on the baseline damage probability and the energy accumulation factor, the time-varying damage index of each section under the current strong earthquake is determined. The time-varying damage index is calculated as min(baseline damage probability × energy accumulation factor, 1), whereby the time-varying damage index equals the baseline damage probability multiplied by the energy accumulation factor, and then the smaller value between the product and 1 is taken. The value ranges from 0 to 1; a larger value indicates more severe damage, and a value less than 0 is taken as 0.

[0047] Finally, according to the arrangement order of the monitoring sections along the pipe gallery, the time-varying damage indices of all sections are arranged in sequence to generate a time-varying damage index sequence.

[0048] The method provided in this embodiment of the invention includes, based on the incident angle parameter, performing angle interpolation processing on a pre-calibrated reference vulnerability curve to generate dynamic damage probability curves for each cross-section at the incident angle, including: Obtain the baseline vulnerability curves corresponding to multiple pre-calibrated discrete angle values. Each discrete angle value corresponds to a damage probability mapping curve with peak acceleration as the independent variable. From the plurality of discrete angle values, find two angle values ​​that are adjacent to the incident angle parameter and record them as the smaller angle and the larger angle, respectively. Extract the damage probability value of the reference vulnerability curve corresponding to the smaller angle under any peak acceleration, and the damage probability value of the reference vulnerability curve corresponding to the larger angle under the same peak acceleration; Based on the difference between the incident angle parameter and the smaller angle, and combined with the difference between the larger angle and the smaller angle, the angle interpolation weight is obtained; For any given peak acceleration, the two damage probability values ​​are weighted and summed according to the angle interpolation weight to obtain the damage probability value under the incident angle parameter. The weighted summation results corresponding to all possible peak accelerations are collected to form a curve with peak acceleration as the independent variable, which serves as the dynamic damage probability curve corresponding to the cross section.

[0049] In this embodiment of the invention, firstly, reference vulnerability curves corresponding to multiple pre-calibrated discrete angle values ​​are obtained. Each discrete angle value corresponds to a damage probability mapping curve with peak ground acceleration as the independent variable, typically obtained using a log-normal cumulative distribution function. The pre-calibrated discrete angle values ​​are typical seismic wave incident directions pre-selected through numerical simulation or experiment in structural vulnerability analysis, such as 30°.

[0050] Next, from the multiple discrete angle values, find the two angle values ​​adjacent to the incident angle parameter, and denote them as the smaller angle and the larger angle, respectively. The smaller angle is the angle value that is less than the incident angle among the two adjacent angle values, and the larger angle is the angle value that is greater than the incident angle among the two adjacent angle values.

[0051] Next, the damage probability value of the reference vulnerability curve corresponding to the smaller angle under any peak acceleration is extracted, and the damage probability value of the reference vulnerability curve corresponding to the larger angle under the same peak acceleration is obtained to obtain the two endpoint values ​​of the interpolation calculation.

[0052] Furthermore, based on the difference between the incident angle parameter and the smaller angle, and combined with the difference between the larger angle and the smaller angle, the angle interpolation weight is obtained. Angle interpolation weight = (incident angle parameter - smaller angle) / (larger angle - smaller angle). The angle interpolation weight is a value between 0 and 1, used to represent the relative position of the incident angle between the smaller and larger angles, and determines the proportion of the larger angle curve in the interpolation result.

[0053] Furthermore, for any given peak acceleration, the two damage probability values ​​are weighted and summed according to the angle interpolation weights to obtain the damage probability value under the incident angle parameter. Specifically, the angle interpolation weight for the larger angle = (incident angle parameter - smaller angle) / (larger angle - smaller angle). The angle interpolation weight for the smaller angle = 1 - the angle interpolation weight for the larger angle, and the sum of the two is 1.

[0054] Finally, the weighted summation results corresponding to all possible peak accelerations are collected to form a curve with peak acceleration as the independent variable and damage probability value as the dependent variable, which serves as the dynamic damage probability curve corresponding to the cross section.

[0055] In this embodiment of the invention, a dynamic damage probability curve is obtained through angle interpolation based on a calibrated baseline vulnerability curve. Then, the corresponding baseline damage probability is read based on the peak acceleration. Subsequently, an energy accumulation factor is obtained based on the ratio of the accumulated energy value to the unit reference energy, and a time-varying damage index is calculated. This comprehensive characterization of damage from both intensity and energy consumption dimensions results in higher accuracy and physical consistency.

[0056] S30: Scan the time-varying damage index sequence along the pipe gallery, identify the critical section where the damage index changes abruptly, and calculate the average damage index of each independent section with the critical section as the dividing point, and determine the section corresponding to the maximum average damage index as the functional bottleneck section. In this embodiment of the invention, a sudden change in the damage index is a statistically non-gradual jump in the damage index values ​​of adjacent sections along the extension direction of the pipe gallery; a critical section is the location of the section where the damage index sudden change occurs, which can be used as the dividing point for dividing sections with different damage levels; an independent section is a continuous pipe gallery section with relatively uniform internal damage distribution, defined by adjacent critical sections; and a functional bottleneck section is an independent section with the highest average damage index.

[0057] Specifically, a damage gradient sequence is obtained by calculating the time-varying damage index sequence. Then, a mutation threshold is set. The damage gradient sequence is traversed, and all cross-sectional locations exceeding the mutation threshold are marked. Candidate mutation points are spatially merged, and then the critical cross-sections are obtained through merging. The average damage index of the segment is calculated. Finally, the segment corresponding to the maximum average damage index is designated as the functional bottleneck segment.

[0058] Step S30 in the method provided in this embodiment of the invention includes: According to the extension direction of the utility tunnel, the cross-section order in the time-varying damage index sequence is used as the scanning order, and the absolute value of the time-varying damage index difference between two adjacent cross-sections is calculated in turn to generate a damage gradient sequence. Traverse the damage gradient sequence and mark the cross-sectional locations where the average gradient value of all gradient values ​​in the sequence exceeds a specified multiple as mutation candidate locations; The candidate mutation locations are merged by combining multiple candidate locations whose spatial distance is less than the distance between adjacent sections into one location, which is then used as the critical section. Using the critical section as the dividing boundary, the utility tunnel is divided into multiple independent sections. Each independent section consists of a continuous section between two adjacent critical sections. The sections at both ends are respectively from the starting point of the utility tunnel to the first critical section and from the last critical section to the end point of the utility tunnel. For each independent segment, the arithmetic mean of the time-varying damage indices of all sections within the segment is calculated as the average damage index of the segment. Compare the average damage index of all independent segments, and extract the segment with the highest average damage index as the functional bottleneck segment.

[0059] In this embodiment of the invention, firstly, according to the extension direction of the pipe gallery, the cross-section order in the time-varying damage index sequence is used as the scanning order, and the absolute value of the time-varying damage index difference between two adjacent cross-sections is calculated in turn to generate a damage gradient sequence, wherein the damage gradient is the absolute value of the time-varying damage index difference.

[0060] Secondly, the damage gradient sequence is traversed, and the cross-sectional locations where the average gradient value of all gradient values ​​in the sequence exceeds a specified multiple are marked as mutation candidate locations. Here, the specified multiple is a preset hyperparameter used to control the sensitivity of mutation detection, which can be taken as 2 to 3 times the average gradient value, as the mutation determination threshold.

[0061] Specifically, the average gradient value of each damage gradient sequence is calculated, the average gradient value is compared with a specified multiple, and the cross-sectional locations corresponding to the average gradient value being greater than the specified multiple are marked.

[0062] Next, candidate locations for abrupt changes that are too close in space are merged. Multiple candidate locations with a spatial distance less than the distance between adjacent cross sections are combined into one location as the critical cross section, typically the midpoint between the mileage coordinates of two cross sections. The spatial distance is the Euclidean distance between the mileage coordinates of the two cross sections, which is the square root of the sum of the squares of the differences between the two coordinates. The distance between adjacent cross sections is the actual distance between the two cross sections.

[0063] Furthermore, the utility tunnel is divided into multiple independent sections using critical cross-sections as dividing boundaries. Each independent section consists of a continuous cross-section between two adjacent critical cross-sections. The two end sections extend from the starting point of the utility tunnel to the first critical cross-section, and from the last critical cross-section to the end point of the utility tunnel, respectively. This division ensures that the degree of damage within each independent section remains consistent. For example, if the critical cross-sections are cross-sections 4, 7, and 10, then the independent sections are: Section A (cross-sections 1-3), Section B (cross-sections 5-6), Section C (cross-sections 8-9), and Section D (cross-section 11).

[0064] Furthermore, for each independent segment, the arithmetic mean of the time-varying damage indices of all cross sections within the segment is calculated as the average damage index of the segment. For example, segment B contains cross sections 5 and 6, with damage indices of 0.75 and 0.72 respectively, so the average damage index = (0.75 + 0.72) / 2 = 0.735.

[0065] Finally, the average damage index of all independent sections is compared, and the section with the highest average damage index is extracted as the functional bottleneck section. The resulting functional bottleneck section is the section in the utility tunnel system that is most severely damaged and most likely to limit the overall functional recovery.

[0066] In this embodiment of the invention, a damage gradient sequence is generated according to the extension direction of the utility tunnel, and the spatial distribution of the damage index is converted into a gradient distribution. Then, damage mutation judgment is performed according to a specified multiple to obtain mutation candidate positions, so that mutation detection can adapt to the overall level of damage gradient under different utility tunnels and different earthquake cases, avoiding missed detections or false detections. Then, the mutation candidate positions are merged to obtain the critical section. Then, the utility tunnel is divided into multiple independent sections according to the critical section to avoid over-segmentation due to local noise. For each independent section, the average damage index is calculated. Finally, the section corresponding to the largest average damage index is taken as the functional bottleneck section, laying the foundation for subsequent functional bottleneck section identification.

[0067] S40: Obtain the average damage index of the functional bottleneck section and the traffic redundancy of the alternative paths behind it, calculate the overall functional degradation rate of the utility tunnel and obtain the functional recovery multiple. In this embodiment of the invention, the alternative path is a branch tunnel or parallel passage upstream of the functional bottleneck section that can bypass the bottleneck section to achieve the function; the traffic redundancy is a measure of the additional traffic capacity of the alternative path relative to the original traffic capacity of the main line; the overall functional degradation rate is the proportion of the overall function that the tunnel system can maintain relative to the complete function, considering the damage to the bottleneck section and the existence of alternative paths; the functional recovery factor is an amplification factor used to correct the recovery time estimated based on ideal conditions.

[0068] Specifically, by combining functional bottleneck sections with subsequent alternative paths, the overall functional status is assessed through traffic redundancy. The overall functional degradation rate and functional recovery multiple are calculated, providing crucial data for subsequent recovery time calculations.

[0069] Step S40 in the method provided in this embodiment of the invention includes: Identify the branch pipe corridors or bypass passages connected to the functional bottleneck section, and count the number of all alternative paths and the available cross-sectional area of ​​each alternative path. Calculate the sum of the available cross-sectional areas of all alternative paths, and combine it with the original cross-sectional area of ​​the main line where the functional bottleneck section is located to obtain the traffic redundancy. Read the average damage index of the functional bottleneck section and calculate the proportion of irreplaceable loss. The overall functional degradation rate of the utility tunnel is calculated based on the average damage index of the functional bottleneck section and the proportion of irreplaceable losses. The overall functional retention rate is determined based on the overall functional degradation rate of the utility tunnel, and the functional recovery multiple is calculated based on the overall functional retention rate.

[0070] In this embodiment of the invention, firstly, the branch pipe corridors or bypass channels connected to the functional bottleneck section are identified, and the number of all alternative paths and the available cross-sectional area of ​​each alternative path are counted. Here, "behind the functional bottleneck section" refers to the location along the functional flow direction of the pipe corridor, after the bottleneck section, where an alternative path can be connected. All branches or bypass channels connected to the upstream node of the functional bottleneck section are obtained. For each path, the available cross-sectional area is read.

[0071] Secondly, calculate the sum of the available cross-sectional areas of all alternative routes, and combine this with the original cross-sectional area of ​​the main line where the functional bottleneck section is located to obtain the traffic redundancy. Traffic redundancy = total available cross-sectional area of ​​alternative routes / original cross-sectional area.

[0072] Next, read the average damage index of the functional bottleneck section and calculate the non-replaceable loss ratio, where the non-replaceable loss ratio = 1 - traffic redundancy.

[0073] Furthermore, based on the average damage index and the proportion of irreplaceable losses in the functional bottleneck section, the overall functional degradation rate of the utility tunnel is calculated: Overall functional degradation rate = Average damage index of functional bottleneck section × Proportion of irreplaceable losses.

[0074] Finally, the overall functional retention rate of the utility tunnel is determined based on the overall functional degradation rate, and the functional recovery multiple is calculated based on the overall functional retention rate. Overall functional retention rate = 1 - overall functional degradation rate; Functional recovery multiple = 1 / overall functional retention rate - 1.

[0075] In this embodiment of the invention, the number of alternative paths and their available cross-sectional areas are statistically analyzed. Then, based on the available cross-sectional areas of the alternative paths, the redundancy of the utility tunnel system is quantified. The proportion of irreplaceable losses is calculated based on the average damage index of the functional bottleneck sections to avoid overestimating actual functional losses. The overall functional degradation rate of the utility tunnel is obtained based on the average damage index of the functional bottleneck sections, and the functional recovery multiple is calculated. The overall functional retention rate is converted into a functional recovery multiple, which serves as analytical data connecting damage assessment and recovery time prediction.

[0076] S50: Calculate the initial recovery time based on the functional recovery multiple and the unit time repair capability parameter, then input the time-varying damage index sequence and the initial recovery time into the pre-trained dynamic time series prediction model, and output the final recovery time after post-earthquake time recursion correction.

[0077] In this embodiment of the invention, the unit time repair capability parameter is the number of standard damage units that can be repaired by all available emergency repair resources within a unit time; the initial recovery time is obtained based on the ratio of the overall damage amount to the repair capability, without considering the spatiotemporal coupling of damage distribution; the dynamic time series prediction model is a deep learning model built based on LSTM; the post-earthquake time recursive correction uses the time-varying damage index sequence and the initial recovery time as initial conditions, and outputs the final recovery time through the recursive update of its internal state.

[0078] Specifically, the function recovery factor and unit time repair capability parameters are read to calculate the initial recovery time. The time-varying damage index is combined with the initial recovery time to form the input feature vector. This vector is then input into a pre-trained LSTM dynamic temporal prediction model. The dynamic temporal prediction model dynamically recursively corrects the initial recovery time through its internal control mechanism, ultimately yielding the model-corrected final recovery time.

[0079] Step S50 in the method provided in this embodiment of the invention includes: The initial recovery time is calculated by reading the function recovery multiple and the unit time repair capacity parameter. The unit time repair capacity parameter is determined comprehensively based on the number of emergency repair personnel available in the utility tunnel, the number of emergency repair equipment sets, and the reserve of spare structural components. The time-varying damage index sequence is combined with the initial recovery time to form an input feature set; The input feature set is fed into a pre-trained dynamic time-series prediction model, and the output is the final recovery time after post-earthquake time recursion correction. The pre-training steps of the dynamic time-series prediction model include: For each historical earthquake sample, the peak acceleration and cumulative vibration energy time history curves of each monitoring section are extracted based on the vibration acceleration time history. The dynamic angle sequence between adjacent sections is calculated, and the time-varying damage index sequence of each section arranged along the extension direction of the pipe gallery is obtained. Based on the emergency repair resource records of the corresponding earthquake case, the initial recovery time is calculated, and the actual functional recovery time of the corresponding earthquake case is used as the output monitoring value; The time-varying damage index sequence and initial recovery time corresponding to each historical earthquake case are used as input features, and the corresponding actual functional recovery time is used as the output label to construct training sample pairs; Construct a cyclic gated memory network, wherein the cyclic gated memory network includes a forget gate, an input gate, an output gate, and a memory unit vector and a hidden state vector that are updated with each processing step; The time-varying damage index sequence in each training sample pair is used as the input for each processing step in the order of cross-section arrangement, and the initial recovery time is used as a constant input feature in all processing steps, and then fed into the recurrent gated memory network in sequence. At each processing step, the recurrent gated memory network updates the memory vector of the current step through the forget gate and the input gate based on the current step input, the hidden state vector of the previous step and the memory unit vector, and generates the hidden state vector of the current step through the output gate and the updated memory unit vector. After all the data from all sections in the time-varying damage index sequence have been processed, the hidden state vector of the last processing step is read through a fully connected output layer and mapped to the final predicted recovery time. The model input is taken from the input features in the training sample pairs, and the output label is taken as the supervision signal. The internal weight parameters of the forget gate, input gate, output gate and fully connected output layer are updated iteratively using the backpropagation algorithm, and the mean absolute percentage error is taken as the loss function. Training stops when the loss function no longer decreases after multiple iterations on the independent validation set, and the trained recurrent gated memory network is used as a dynamic time-series prediction model.

[0080] In this embodiment of the invention, firstly, the functional recovery multiple and the unit time repair capability parameter are read to calculate the initial recovery time, which is calculated as: Initial recovery time = Functional recovery multiple / Unit time repair capability parameter. The unit time repair capability parameter reflects the comprehensive allocation capability of emergency repair resources. It is determined based on the number of available emergency repair personnel, the number of emergency repair equipment sets, and the reserve of spare structural components within the utility tunnel. It is typically calibrated using historical earthquake data, for example, at 0.1 / hour.

[0081] Secondly, the time-varying damage index sequence is merged with the initial recovery time, so that each time-varying damage index sequence and the initial recovery time constitute a binary data pair of time-varying damage index sequence-initial recovery time, forming an input feature set, which consists of multiple binary data pairs.

[0082] Next, the input feature set is fed into a pre-trained dynamic temporal prediction model, and the output is the final recovery time after post-earthquake time recursion correction. The dynamic temporal prediction model is a deep learning model pre-trained on an LSTM. LSTM is a special type of recurrent neural network that can effectively learn long-term dependencies in long sequence data and alleviate the gradient vanishing or exploding problems in traditional recurrent neural networks.

[0083] Specifically, the pre-training steps for dynamic time series prediction models include: First, for each historical earthquake sample, following the calculation method described above, the peak ground acceleration and cumulative ground acceleration time history curves of each monitoring section are extracted based on the ground acceleration time history. The dynamic angle sequence between adjacent sections is then calculated, and the time-varying damage index sequence for each section arranged along the extension direction of the utility tunnel is obtained. The historical earthquake samples are datasets of real strong earthquake events that occurred in history, with each sample containing complete monitoring data records and subsequent actual functional recovery time.

[0084] Secondly, the repair capacity parameter per unit time is calculated by combining the emergency repair resource records of corresponding historical earthquake cases, and the initial recovery time is calculated by combining the functional recovery multiple of that earthquake case. Simultaneously, the actual functional recovery time of historical earthquake cases is read as the output supervision value. Output labels required for training are constructed so that the model can learn the mapping relationship from input features to actual recovery time. Here, the emergency repair resource records are the data on the personnel and equipment actually mobilized after historical earthquakes. The actual functional recovery time is the real time experienced by the utility tunnel from the time of the earthquake to the restoration of its intended function in historical earthquake cases.

[0085] Next, the time-varying damage index sequence of each earthquake case is combined with the initial recovery time to form the input, and the actual recovery time is used as the supervision label to form training sample pairs.

[0086] Furthermore, a recurrent gated memory network is constructed, which includes a forget gate, an input gate, an output gate, and a memory cell vector and a hidden state vector that are updated with each processing step. The forget gate determines the proportion of information discarded from the memory cells at the previous time step; the input gate determines the number of new information in the current input that should be written into the memory cells; and the output gate determines the number of information in the current memory cells that should be output to the hidden state. The memory cell vector is responsible for long-term information storage; the hidden state vector is used for prediction in the current step and input for the next step.

[0087] Specifically, the input gate receives the feature vector from each processing step, with a feature dimension of 2 and 2 nodes. Following this is the first LSTM hidden layer, containing 128 LSTM neurons. Each LSTM unit contains three gate structures: a forget gate, an input gate, and an output gate. The activation functions are uniformly the hyperbolic tangent (tanh) as the state activation function for the memory unit and the sigmoid function as the activation function for the gate units, extracting local temporal patterns from the input two-dimensional features and conveying long-term dependency information. The second LSTM hidden layer contains 64 neurons, also using tanh and sigmoid activation functions, abstracting and reducing the dimensionality of the high-level features output from the first layer. The third layer is the output gate, containing one neuron with a linear activation function, mapping the hidden state vector of the last processing step to the final predicted recovery time.

[0088] Furthermore, the time-varying damage index sequence in each training sample pair is used as the input for each processing step in the order of cross-section arrangement, and the initial recovery time is used as a constant input feature in all processing steps, and then fed into the recurrent gated memory network.

[0089] Furthermore, the initial learning rate was set to 0.001, the Adam optimizer was used for parameter updates, the batch size was set to 16, and the maximum number of iterations was 200. During training, for each historical earthquake sample, its time-varying damage index sequence and initial recovery time were first extracted and combined into an input feature set. The actual functional recovery time of the earthquake was used as the output label to construct training sample pairs. Through forward propagation, the cross-sectional data of each sample were sequentially fed into the LSTM network step by step: at a certain time step, the input vector interacted with the hidden state and memory units of the previous time step. The forget gate determined the proportion of old memories to be discarded, the memory units were updated through the input gate and candidate memories, and then the current hidden state was obtained through the output gate. After traversing all cross-sections, the final hidden state was fed into the fully connected output layer to obtain the predicted recovery time.

[0090] Furthermore, using the input features from the training sample pairs as model input and the output labels as supervision signals, the backpropagation algorithm is used to iteratively update the internal weight parameters of the forget gate, input gate, output gate, and fully connected output layer. The gradient is calculated layer by layer from the output layer to the input layer using the chain rule. After each iteration, the mean absolute percentage error between the predicted value and the true label is calculated on the validation set. A very small positive constant is added to the actual function recovery time during loss function calculation to avoid the risk of division by zero.

[0091] Finally, training stops when the loss function no longer decreases after multiple iterations on the independent validation set. The trained recurrent gated memory network is then used as a dynamic temporal prediction model for dynamic prediction of the final recovery time. The dynamic temporal prediction model uses its internal gated mechanism to dynamically and recursively correct the initial recovery time.

[0092] In this embodiment of the invention, a deep learning network skeleton with long-term memory capability is constructed to lay the foundation for subsequent training. Then, historical earthquake examples are acquired, and time-varying damage index sequences and functional recovery multiples are calculated. The time-varying damage index sequences and functional recovery multiples are combined to calculate the initial recovery time, and output labels required for training are constructed, enabling the model to learn the mapping relationship from input features to the actual recovery time. The recovery time is recursively corrected after the earthquake using a pre-trained LSTM dynamic temporal prediction model, and finally, a high-precision final recovery time is output, realizing high-precision dynamic prediction of the operational resilience of integrated utility tunnels under strong earthquakes.

[0093] Through the above specific implementation methods, the embodiments of the present invention achieve the following technical effects: In this embodiment of the invention, firstly, acceleration data in three orthogonal directions of the pipe gallery cross-section are collected and instantaneous composite acceleration is calculated to fully capture the vibration vector characteristics. Then, peak ground acceleration is extracted to characterize the seismic intensity, and a vibration energy accumulation time history curve is constructed to characterize the energy time-domain evolution. This provides fundamental characteristic parameters for damage probability mapping and cumulative effect assessment, ensuring the accuracy of subsequent time-varying damage index calculations.

[0094] Secondly, based on the calibrated baseline vulnerability curve, a dynamic damage probability curve is obtained through angle interpolation. Then, the baseline damage probability is read based on the peak acceleration. Subsequently, the energy accumulation factor is obtained based on the ratio of the accumulated energy value to the unit reference energy, and the time-varying damage index is calculated. This comprehensive characterization of damage from both intensity and energy consumption dimensions results in higher accuracy and physical consistency.

[0095] Next, a damage gradient sequence is generated according to the extension direction of the utility tunnel, converting the spatial distribution of the damage index into a gradient distribution. Then, damage abrupt changes are judged according to a specified multiple to obtain candidate locations for abrupt changes, enabling abrupt change detection to adapt to the overall level of damage gradient under different utility tunnels and different earthquake cases, avoiding missed or false detections. Subsequently, the candidate locations for abrupt changes are merged to obtain the critical section. Then, the utility tunnel is divided into multiple independent segments according to the critical section to avoid over-segmentation due to local noise. For each independent segment, the average damage index is calculated. Finally, the segment corresponding to the largest average damage index is taken as the functional bottleneck segment, laying the foundation for subsequent identification of functional bottleneck segments.

[0096] Furthermore, the number of alternative paths and their available cross-sectional areas are statistically analyzed. Then, based on the available cross-sectional areas of the alternative paths, the redundancy of the utility tunnel system is quantified. The proportion of irreplaceable losses is calculated based on the average damage index of the functional bottleneck sections to avoid overestimating actual functional losses. The overall functional degradation rate of the utility tunnel is obtained based on the average damage index of the functional bottleneck sections, and the functional recovery multiple is calculated. The overall functional retention rate is converted into a functional recovery multiple, which serves as analytical data connecting damage assessment and recovery time prediction.

[0097] Finally, a deep learning network skeleton with long-term memory capability was constructed, laying the foundation for subsequent training. Then, historical earthquake examples were acquired, and time-varying damage index sequences and functional recovery multiples were calculated. The time-varying damage index sequences and functional recovery multiples were combined to calculate the initial recovery time, and the output labels required for training were constructed, enabling the model to learn the mapping relationship from input features to the actual recovery time. The recovery time was recursively corrected after the earthquake using a pre-trained LSTM dynamic temporal prediction model, and finally, a high-precision final recovery time was output, realizing high-precision dynamic prediction of the operational resilience of integrated utility tunnels under strong earthquakes.

[0098] Example 2, as Figure 3 As shown, based on the same inventive concept as the LSTM-driven dynamic prediction method for the operational resilience of integrated utility tunnels provided in Embodiment 1, this embodiment of the invention also provides an LSTM-driven dynamic prediction system for the operational resilience of integrated utility tunnels, the system comprising: The angle sequence generation module 11 is used to obtain the vibration acceleration time history of each monitoring section along the integrated utility tunnel during strong earthquake, extract the peak acceleration and vibration energy accumulation time history curve of each section, calculate the time difference sequence of vibration waveform arrival between adjacent sections, and generate a dynamic angle sequence between the direction of seismic wave propagation and the axis of the utility tunnel. The damage probability curve construction module 12 is used to correct the reference vulnerability curve of each section according to the dynamic angle sequence, obtain the dynamic damage probability curve, and then substitute the peak acceleration and vibration energy accumulation value of each section into the dynamic damage probability curve to output the time-varying damage index sequence of each section. The functional bottleneck section acquisition module 13 is used to scan the time-varying damage index sequence along the pipe gallery, identify the critical section where the damage index changes abruptly, and calculate the average damage index of each independent section with the critical section as the dividing point, and determine the section corresponding to the maximum average damage index as the functional bottleneck section. The function recovery multiple calculation module 14 is used to obtain the average damage index of the functional bottleneck section and the traffic redundancy of the subsequent alternative path, calculate the overall functional degradation rate of the utility tunnel and obtain the function recovery multiple. The recovery time generation module 15 is used to calculate the initial recovery time based on the functional recovery multiple and the unit time repair capability parameter, and then input the time-varying damage index sequence and the initial recovery time into the pre-trained dynamic time series prediction model to output the final recovery time after post-earthquake time recursion correction. The final recovery time serves as a key decision-making basis, synchronously driving the priority ranking of repair resource scheduling and the dynamic adjustment of emergency response level, ensuring that limited resources are accurately allocated to the bottleneck section with the most significant functional degradation and the highest requirements for repair timeliness.

[0099] In one embodiment, the included angle sequence generation module 11 is used for: At each monitoring section, acceleration values ​​in three orthogonal directions are continuously recorded at a fixed sampling frequency. The three orthogonal directions are the horizontal direction parallel to the extension direction of the pipe gallery, the vertical direction perpendicular to the extension direction of the pipe gallery, and the vertical direction. The three acceleration values ​​at the same moment are combined into one data point, and all data points are arranged in chronological order to generate the original vibration acceleration sequence corresponding to each monitoring section. Baseline drift correction is performed on the original vibration acceleration sequence to obtain the corrected vibration acceleration time history; From the corrected vibration acceleration time history, the composite modulus of the acceleration values ​​in three orthogonal directions at each sampling moment is calculated to obtain the instantaneous composite acceleration value at the corresponding moment. Iterate through all the instantaneous composite acceleration values ​​in the vibration acceleration time history and take the value with the largest value as the peak acceleration of the cross section; Starting from the moment the vibration begins, the instantaneous synthetic acceleration value is accumulated over time, and the accumulated results are recorded in chronological order to generate the vibration energy accumulation time history curve of the cross section.

[0100] The process of performing baseline drift correction on the original vibration acceleration sequence to obtain the corrected vibration acceleration time history includes: The stable recording segment before the start of the strong earthquake event is identified from the original vibration acceleration sequence. The static zero-point offset in each orthogonal direction is determined based on the stable recording segment. The strong earthquake event is defined by the instantaneous synthetic acceleration value of any monitoring section exceeding the trigger threshold. The trigger threshold is determined based on the background environmental vibration level of the site where the utility tunnel is located. Subtract the static zero-point offset in the corresponding direction from each acceleration value in the original vibration acceleration sequence to obtain a preliminary de-biasing sequence; The preliminary debiased sequence is subjected to trend term separation processing to extract the trend drift component that changes slowly over time; The trend drift component is subtracted from the initial debiased sequence to obtain the corrected acceleration value at each sampling time. All corrected acceleration values ​​are arranged in chronological order of sampling time to generate the corrected vibration acceleration time history.

[0101] In one embodiment, the included angle sequence generation module 11 is further configured to: From the vibration acceleration time history of each monitoring section, the instantaneous synthetic acceleration values ​​are extracted and arranged in order of change with the sampling time, which serves as the vibration waveform time series data corresponding to the section. Select the vibration waveform time series data of two adjacent monitoring sections, compare the waveform of one section at different time offsets with the waveform of the other section, and record the offset corresponding to the highest waveform matching degree as the time difference of the vibration waveform from the near source section to the far source section. According to the extension direction of the utility tunnel, the time differences calculated for each pair of adjacent sections are arranged sequentially to generate a time difference sequence; The propagation velocity of seismic waves in the surrounding soil and rock medium of the utility tunnel and the time difference sequence are obtained. The propagation distance of seismic waves between adjacent sections is determined. Then, based on the geometric correspondence between the propagation distance and the straight-line distance between adjacent sections, the angle between the propagation direction of the seismic waves and the axis of the utility tunnel is calculated. Connect the included angles corresponding to all adjacent cross sections according to the cross section arrangement order to generate the dynamic included angle sequence.

[0102] In one embodiment, the damage probability curve construction module 12 is used for: For each monitoring section, the included angle value corresponding to each section is extracted from the dynamic included angle sequence and used as the incident angle parameter; Based on the incident angle parameter, the pre-calibrated baseline vulnerability curve is subjected to angle interpolation to generate dynamic damage probability curves for each section under the incident angle, wherein the dynamic damage probability curve represents a damage probability mapping relationship with peak acceleration as the independent variable. Substitute the peak acceleration of each section into the dynamic damage probability curve, read the corresponding probability value, and use it as the baseline damage probability. Extract the cumulative energy value at the end point from the cumulative vibration energy time history curve of the cross section, and calculate the ratio of the cumulative energy value to the unit reference energy as the energy accumulation factor. The unit reference energy is the reference vibration energy value of the pipe gallery cross section under the elastic limit state, which is obtained by vibration table test calibration. Based on the baseline damage probability and the energy accumulation factor, the time-varying damage index of each section under the current strong earthquake is determined, wherein the value of the time-varying damage index ranges from 0 to 1, and the larger the value, the more severe the damage. According to the arrangement order of the monitoring sections along the pipe gallery, the time-varying damage indices of all sections are arranged sequentially to generate the time-varying damage index sequence.

[0103] Specifically, based on the incident angle parameter, the pre-calibrated baseline vulnerability curve is subjected to angle interpolation processing to generate dynamic damage probability curves for each cross-section at the incident angle, including: Obtain the baseline vulnerability curves corresponding to multiple pre-calibrated discrete angle values. Each discrete angle value corresponds to a damage probability mapping curve with peak acceleration as the independent variable. From the plurality of discrete angle values, find two angle values ​​that are adjacent to the incident angle parameter and record them as the smaller angle and the larger angle, respectively. Extract the damage probability value of the reference vulnerability curve corresponding to the smaller angle under any peak acceleration, and the damage probability value of the reference vulnerability curve corresponding to the larger angle under the same peak acceleration; Based on the difference between the incident angle parameter and the smaller angle, and combined with the difference between the larger angle and the smaller angle, the angle interpolation weight is obtained; For any given peak acceleration, the two damage probability values ​​are weighted and summed according to the angle interpolation weight to obtain the damage probability value under the incident angle parameter. The weighted summation results corresponding to all possible peak accelerations are collected to form a curve with peak acceleration as the independent variable, which serves as the dynamic damage probability curve corresponding to the cross section.

[0104] In one embodiment, the functional bottleneck segment acquisition module 13 is used for: According to the extension direction of the utility tunnel, the cross-section order in the time-varying damage index sequence is used as the scanning order, and the absolute value of the time-varying damage index difference between two adjacent cross-sections is calculated in turn to generate a damage gradient sequence. Traverse the damage gradient sequence and mark the cross-sectional locations where the average gradient value of all gradient values ​​in the sequence exceeds a specified multiple as mutation candidate locations; The candidate mutation locations are merged by combining multiple candidate locations whose spatial distance is less than the distance between adjacent sections into one location, which is then used as the critical section. Using the critical section as the dividing boundary, the utility tunnel is divided into multiple independent sections. Each independent section consists of a continuous section between two adjacent critical sections. The sections at both ends are respectively from the starting point of the utility tunnel to the first critical section and from the last critical section to the end point of the utility tunnel. For each independent segment, the arithmetic mean of the time-varying damage indices of all sections within the segment is calculated as the average damage index of the segment. Compare the average damage index of all independent segments, and extract the segment with the highest average damage index as the functional bottleneck segment.

[0105] In one embodiment, the function recovery factor calculation module 14 is used for: Identify the branch pipe corridors or bypass passages connected to the functional bottleneck section, and count the number of all alternative paths and the available cross-sectional area of ​​each alternative path. Calculate the sum of the available cross-sectional areas of all alternative paths, and combine it with the original cross-sectional area of ​​the main line where the functional bottleneck section is located to obtain the traffic redundancy. Read the average damage index of the functional bottleneck section and calculate the proportion of irreplaceable loss. The overall functional degradation rate of the utility tunnel is calculated based on the average damage index of the functional bottleneck section and the proportion of irreplaceable losses. The overall functional retention rate is determined based on the overall functional degradation rate of the utility tunnel, and the functional recovery multiple is calculated based on the overall functional retention rate.

[0106] In one embodiment, the recovery time generation module 15 is used for: The initial recovery time is calculated by reading the function recovery multiple and the unit time repair capacity parameter. The unit time repair capacity parameter is determined comprehensively based on the number of emergency repair personnel available in the utility tunnel, the number of emergency repair equipment sets, and the reserve of spare structural components. The time-varying damage index sequence is combined with the initial recovery time to form an input feature set; The input feature set is fed into a pre-trained dynamic time-series prediction model, and the output is the final recovery time after post-earthquake time recursion correction. The pre-training steps of the dynamic time-series prediction model include: For each historical earthquake sample, the peak acceleration and cumulative vibration energy time history curves of each monitoring section are extracted based on the vibration acceleration time history. The dynamic angle sequence between adjacent sections is calculated, and the time-varying damage index sequence of each section arranged along the extension direction of the pipe gallery is obtained. Based on the emergency repair resource records of the corresponding earthquake case, the initial recovery time is calculated, and the actual functional recovery time of the corresponding earthquake case is used as the output monitoring value; The time-varying damage index sequence and initial recovery time corresponding to each historical earthquake case are used as input features, and the corresponding actual functional recovery time is used as the output label to construct training sample pairs; Construct a cyclic gated memory network, wherein the cyclic gated memory network includes a forget gate, an input gate, an output gate, and a memory unit vector and a hidden state vector that are updated with each processing step; The time-varying damage index sequence in each training sample pair is used as the input for each processing step in the order of cross-section arrangement, and the initial recovery time is used as a constant input feature in all processing steps, and then fed into the recurrent gated memory network in sequence. At each processing step, the recurrent gated memory network updates the memory vector of the current step through the forget gate and the input gate based on the current step input, the hidden state vector of the previous step and the memory unit vector, and generates the hidden state vector of the current step through the output gate and the updated memory unit vector. After all the data from all sections in the time-varying damage index sequence have been processed, the hidden state vector of the last processing step is read through a fully connected output layer and mapped to the final predicted recovery time. The model input is taken from the input features in the training sample pairs, and the output label is taken as the supervision signal. The internal weight parameters of the forget gate, input gate, output gate and fully connected output layer are updated iteratively using the backpropagation algorithm, and the mean absolute percentage error is taken as the loss function. Training stops when the loss function no longer decreases after multiple iterations on the independent validation set, and the trained recurrent gated memory network is used as a dynamic time-series prediction model.

[0107] Compared to existing technologies, this invention provides an LSTM-driven dynamic prediction method for the operational resilience of integrated utility tunnels. First, it corrects the baseline vulnerability curve by using a dynamic angle sequence between the seismic wave propagation direction and the tunnel axis, and calculates a time-varying damage index by combining peak ground acceleration and cumulative seismic energy. This overcomes the shortcomings of traditional methods, such as the static incident angle assumption and single-parameter evaluation, improving the adaptability and physical accuracy of structural damage assessment. Second, by scanning the time-varying damage index sequence along the tunnel's alignment, it identifies critical sections and divides them into independent segments based on abrupt changes in damage gradients, pinpointing functional bottleneck segments and achieving zonal localization of the spatial distribution of post-earthquake damage in the tunnel. Finally, the initial recovery time, calculated using the functional recovery factor and the repair capacity per unit time parameter, is input into a pre-trained LSTM dynamic temporal prediction model along with the time-varying damage index sequence. Utilizing the model's nonlinear learning capability for damage spatial distribution patterns, it achieves post-earthquake recursive correction of the initial recovery time. The output recovery time exhibits higher prediction accuracy, effectively supporting the post-earthquake resilience assessment of integrated utility tunnels.

Claims

1. An LSTM-driven dynamic prediction method for the operational resilience of integrated utility tunnels, characterized in that, The method includes: The vibration acceleration time history of each monitoring section along the integrated utility tunnel during strong earthquake is continuously collected. The peak acceleration and cumulative vibration energy time history curves of each section are extracted. The time difference sequence of vibration waveform arrival between adjacent sections is calculated, and the dynamic angle sequence between the direction of seismic wave propagation and the axis of the utility tunnel is generated. The baseline vulnerability curve of each section is corrected according to the dynamic angle sequence to obtain the dynamic damage probability curve. Then, the peak acceleration and cumulative vibration energy of each section are substituted into the dynamic damage probability curve to output the time-varying damage index sequence of each section. Scan the time-varying damage index sequence along the pipe gallery route, identify the critical section where the damage index changes abruptly, and calculate the average damage index of each independent section with the critical section as the dividing point. The section corresponding to the maximum average damage index is determined as the functional bottleneck section. Obtain the average damage index of the functional bottleneck section and the traffic redundancy of the subsequent alternative paths, calculate the overall functional degradation rate of the utility tunnel and obtain the functional recovery multiple; Based on the functional recovery factor and the unit time repair capability parameter, the initial recovery time is calculated. Then, the time-varying damage index sequence and the initial recovery time are input into the pre-trained dynamic time series prediction model, and the final recovery time is output after post-earthquake time recursion correction.

2. The LSTM-driven dynamic prediction method for the operational resilience of integrated utility tunnels according to claim 1, characterized in that, The vibration acceleration time histories of each monitoring section along the integrated utility tunnel during a strong earthquake were continuously collected. The peak acceleration versus cumulative vibration energy time histories curves for each section were extracted, including: At each monitoring section, acceleration values ​​in three orthogonal directions are continuously recorded at a fixed sampling frequency. The three orthogonal directions are the horizontal direction parallel to the extension direction of the pipe gallery, the vertical direction perpendicular to the extension direction of the pipe gallery, and the vertical direction. The three acceleration values ​​at the same moment are combined into one data point, and all data points are arranged in chronological order to generate the original vibration acceleration sequence corresponding to each monitoring section. Baseline drift correction is performed on the original vibration acceleration sequence to obtain the corrected vibration acceleration time history; From the corrected vibration acceleration time history, the composite modulus of the acceleration values ​​in three orthogonal directions at each sampling moment is calculated to obtain the instantaneous composite acceleration value at the corresponding moment. Iterate through all the instantaneous composite acceleration values ​​in the vibration acceleration time history and take the value with the largest value as the peak acceleration of the cross section; Starting from the moment the vibration begins, the instantaneous synthetic acceleration value is accumulated over time, and the accumulated results are recorded in chronological order to generate the vibration energy accumulation time history curve of the cross section.

3. The LSTM-driven dynamic prediction method for the operational resilience of integrated utility tunnels according to claim 2, characterized in that, The original vibration acceleration sequence is baseline drift corrected to obtain the corrected vibration acceleration time history, including: The stable recording segment before the start of the strong earthquake event is identified from the original vibration acceleration sequence. The static zero-point offset in each orthogonal direction is determined based on the stable recording segment. The strong earthquake event is defined by the instantaneous synthetic acceleration value of any monitoring section exceeding the trigger threshold. The trigger threshold is determined based on the background environmental vibration level of the site where the utility tunnel is located. Subtract the static zero-point offset in the corresponding direction from each acceleration value in the original vibration acceleration sequence to obtain a preliminary de-biasing sequence; The preliminary debiased sequence is subjected to trend term separation processing to extract the trend drift component that changes slowly over time; The trend drift component is subtracted from the initial debiased sequence to obtain the corrected acceleration value at each sampling time. All corrected acceleration values ​​are arranged in chronological order of sampling time to generate the corrected vibration acceleration time history.

4. The LSTM-driven dynamic prediction method for the operational resilience of integrated utility tunnels according to claim 1, characterized in that, Calculate the time difference sequence of vibration waveform arrival between adjacent sections to generate a dynamic angle sequence between the seismic wave propagation direction and the tunnel axis, including: From the vibration acceleration time history of each monitoring section, the instantaneous synthetic acceleration values ​​are extracted and arranged in order of change with the sampling time, which serves as the vibration waveform time series data corresponding to the section. Select the vibration waveform time series data of two adjacent monitoring sections, compare the waveform of one section at different time offsets with the waveform of the other section, and record the offset corresponding to the highest waveform matching degree as the time difference of the vibration waveform from the near source section to the far source section. According to the extension direction of the utility tunnel, the time differences calculated for each pair of adjacent sections are arranged sequentially to generate a time difference sequence; The propagation velocity of seismic waves in the surrounding soil and rock medium of the utility tunnel and the time difference sequence are obtained. The propagation distance of seismic waves between adjacent sections is determined. Then, based on the geometric correspondence between the propagation distance and the straight-line distance between adjacent sections, the angle between the propagation direction of the seismic waves and the axis of the utility tunnel is calculated. Connect the included angles corresponding to all adjacent cross sections according to the cross section arrangement order to generate the dynamic included angle sequence.

5. The LSTM-driven dynamic prediction method for the operational resilience of integrated utility tunnels according to claim 1, characterized in that, The baseline vulnerability curves of each section are corrected based on the dynamic angle sequence to obtain dynamic damage probability curves. Then, the peak acceleration and cumulative vibration energy of each section are substituted into the dynamic damage probability curves to output the time-varying damage index sequence of each section, including: For each monitoring section, the included angle value corresponding to each section is extracted from the dynamic included angle sequence and used as the incident angle parameter; Based on the incident angle parameter, the pre-calibrated baseline vulnerability curve is subjected to angle interpolation to generate dynamic damage probability curves for each section under the incident angle, wherein the dynamic damage probability curve represents a damage probability mapping relationship with peak acceleration as the independent variable. Substitute the peak acceleration of each section into the dynamic damage probability curve, read the corresponding probability value, and use it as the baseline damage probability. Extract the cumulative energy value at the end point from the cumulative vibration energy time history curve of the cross section, and calculate the ratio of the cumulative energy value to the unit reference energy as the energy accumulation factor. The unit reference energy is the reference vibration energy value of the pipe gallery cross section under the elastic limit state, which is obtained by vibration table test calibration. Based on the baseline damage probability and the energy accumulation factor, the time-varying damage index of each section under the current strong earthquake is determined, wherein the value of the time-varying damage index ranges from 0 to 1, and the larger the value, the more severe the damage. According to the arrangement order of the monitoring sections along the pipe gallery, the time-varying damage indices of all sections are arranged sequentially to generate the time-varying damage index sequence.

6. The LSTM-driven dynamic prediction method for the operational resilience of integrated utility tunnels according to claim 5, characterized in that, Based on the incident angle parameter, the pre-calibrated baseline vulnerability curve is interpolated to generate dynamic damage probability curves for each cross-section at the incident angle, including: Obtain the baseline vulnerability curves corresponding to multiple pre-calibrated discrete angle values. Each discrete angle value corresponds to a damage probability mapping curve with peak acceleration as the independent variable. From the plurality of discrete angle values, find two angle values ​​that are adjacent to the incident angle parameter and record them as the smaller angle and the larger angle, respectively. Extract the damage probability value of the reference vulnerability curve corresponding to the smaller angle under any peak acceleration, and the damage probability value of the reference vulnerability curve corresponding to the larger angle under the same peak acceleration; Based on the difference between the incident angle parameter and the smaller angle, and combined with the difference between the larger angle and the smaller angle, the angle interpolation weight is obtained; For any given peak acceleration, the two damage probability values ​​are weighted and summed according to the angle interpolation weight to obtain the damage probability value under the incident angle parameter. The weighted summation results corresponding to all possible peak accelerations are collected to form a curve with peak acceleration as the independent variable, which serves as the dynamic damage probability curve corresponding to the cross section.

7. The LSTM-driven dynamic prediction method for the operational resilience of integrated utility tunnels according to claim 1, characterized in that, The time-varying damage index sequence is scanned along the pipe gallery to identify critical sections where damage indices abruptly change. The average damage index of each independent segment is calculated using these critical sections as dividing points. The segment corresponding to the maximum average damage index is identified as the functional bottleneck segment, including: According to the extension direction of the utility tunnel, the cross-section order in the time-varying damage index sequence is used as the scanning order, and the absolute value of the time-varying damage index difference between two adjacent cross-sections is calculated in turn to generate a damage gradient sequence. Traverse the damage gradient sequence and mark the cross-sectional locations where the average gradient value of all gradient values ​​in the sequence exceeds a specified multiple as mutation candidate locations; The candidate mutation locations are merged by combining multiple candidate locations whose spatial distance is less than the distance between adjacent sections into one location, which is then used as the critical section. Using the critical section as the dividing boundary, the utility tunnel is divided into multiple independent sections. Each independent section consists of a continuous section between two adjacent critical sections. The sections at both ends are respectively from the starting point of the utility tunnel to the first critical section and from the last critical section to the end point of the utility tunnel. For each independent segment, the arithmetic mean of the time-varying damage indices of all sections within the segment is calculated as the average damage index of the segment. Compare the average damage index of all independent segments, and extract the segment with the highest average damage index as the functional bottleneck segment.

8. The LSTM-driven dynamic prediction method for the operational resilience of integrated utility tunnels according to claim 1, characterized in that, Obtain the average damage index of the functional bottleneck section and the traffic redundancy of the subsequent alternative paths, calculate the overall functional degradation rate of the utility tunnel, and obtain the functional recovery multiple, including: Identify the branch pipe corridors or bypass passages connected to the functional bottleneck section, and count the number of all alternative paths and the available cross-sectional area of ​​each alternative path. Calculate the sum of the available cross-sectional areas of all alternative paths, and combine it with the original cross-sectional area of ​​the main line where the functional bottleneck section is located to obtain the traffic redundancy. Read the average damage index of the functional bottleneck section and calculate the proportion of irreplaceable loss. The overall functional degradation rate of the utility tunnel is calculated based on the average damage index of the functional bottleneck section and the proportion of irreplaceable losses. The overall functional retention rate is determined based on the overall functional degradation rate of the utility tunnel, and the functional recovery multiple is calculated based on the overall functional retention rate.

9. The LSTM-driven dynamic prediction method for the operational resilience of integrated utility tunnels according to claim 1, characterized in that, Based on the functional recovery factor and the unit time repair capability parameter, the initial recovery time is calculated. Then, the time-varying damage index sequence and the initial recovery time are input into a pre-trained dynamic time-series prediction model, which outputs the final recovery time after post-earthquake time recursion correction, including: The initial recovery time is calculated by reading the function recovery multiple and the unit time repair capacity parameter. The unit time repair capacity parameter is determined comprehensively based on the number of emergency repair personnel available in the utility tunnel, the number of emergency repair equipment sets, and the reserve of spare structural components. The time-varying damage index sequence is combined with the initial recovery time to form an input feature set; The input feature set is fed into a pre-trained dynamic time-series prediction model, and the output is the final recovery time after post-earthquake time recursion correction. The pre-training steps of the dynamic time-series prediction model include: For each historical earthquake sample, the peak acceleration and cumulative vibration energy time history curves of each monitoring section are extracted based on the vibration acceleration time history. The dynamic angle sequence between adjacent sections is calculated, and the time-varying damage index sequence of each section arranged along the extension direction of the pipe gallery is obtained. Based on the emergency repair resource records of the corresponding earthquake case, the initial recovery time is calculated, and the actual functional recovery time of the corresponding earthquake case is used as the output monitoring value; The time-varying damage index sequence and initial recovery time corresponding to each historical earthquake case are used as input features, and the corresponding actual functional recovery time is used as the output label to construct training sample pairs; Construct a cyclic gated memory network, wherein the cyclic gated memory network includes a forget gate, an input gate, an output gate, and a memory unit vector and a hidden state vector that are updated with each processing step; The time-varying damage index sequence in each training sample pair is used as the input for each processing step in the order of cross-section arrangement, and the initial recovery time is used as a constant input feature in all processing steps, and then fed into the recurrent gated memory network in sequence. At each processing step, the recurrent gated memory network updates the memory vector of the current step through the forget gate and the input gate based on the current step input, the hidden state vector of the previous step and the memory unit vector, and generates the hidden state vector of the current step through the output gate and the updated memory unit vector. After all the data from all sections in the time-varying damage index sequence have been processed, the hidden state vector of the last processing step is read through a fully connected output layer and mapped to the final predicted recovery time. The model input is taken from the input features in the training sample pairs, and the output label is taken as the supervision signal. The internal weight parameters of the forget gate, input gate, output gate and fully connected output layer are updated iteratively using the backpropagation algorithm, and the mean absolute percentage error is taken as the loss function. Training stops when the loss function no longer decreases after multiple iterations on the independent validation set, and the trained recurrent gated memory network is used as a dynamic time-series prediction model.

10. An LSTM-driven dynamic prediction system for the operational resilience of integrated utility tunnels, characterized in that, The system is used to implement the LSTM-driven dynamic prediction method for the operational resilience of integrated utility tunnels according to any one of claims 1-9, the system comprising: The angle sequence generation module is used to obtain the vibration acceleration time history of each monitoring section along the integrated utility tunnel during strong earthquake, extract the peak acceleration and vibration energy accumulation time history curve of each section, calculate the time difference sequence of vibration waveform arrival between adjacent sections, and generate a dynamic angle sequence between the direction of seismic wave propagation and the axis of the utility tunnel. The damage probability curve construction module is used to correct the reference vulnerability curve of each section according to the dynamic angle sequence to obtain the dynamic damage probability curve. Then, the peak acceleration and cumulative vibration energy of each section are substituted into the dynamic damage probability curve to output the time-varying damage index sequence of each section. The functional bottleneck section acquisition module is used to scan the time-varying damage index sequence along the pipe gallery, identify the critical section where the damage index changes abruptly, and calculate the average damage index of each independent section with the critical section as the dividing point, and determine the section corresponding to the maximum average damage index as the functional bottleneck section. The function recovery multiple calculation module is used to obtain the average damage index of the functional bottleneck section and the traffic redundancy of the subsequent alternative path, calculate the overall functional degradation rate of the utility tunnel and obtain the function recovery multiple. The recovery time generation module is used to calculate the initial recovery time based on the functional recovery multiple and the unit time repair capability parameter. Then, the time-varying damage index sequence and the initial recovery time are input into a pre-trained dynamic time series prediction model, and the final recovery time is output after post-earthquake time recursion correction. The final recovery time serves as a key decision-making basis, synchronously driving the priority ranking of repair resource scheduling and the dynamic adjustment of emergency response level, ensuring that limited resources are accurately allocated to the bottleneck sections with the most significant functional degradation and the highest requirements for repair timeliness.

Images