Rail transit structure operation health state monitoring system and method

Abstract

The present application relates to the technical field of track transportation structure health monitoring, in particular to a track transportation structure operation health state monitoring system and method, comprising a multi-source vibration perception module, a data acquisition and transmission module, a data processing and analysis module, a health state evaluation module, a warning and display module, the multi-source vibration perception module is used for collecting multi-source engineering vibration signals based on the coupling frequency response characteristics of the subway structure and multi-source vibration sources, and sensors are differentially arranged at the tunnel lining ring joint, the sleeper and the track bed interface, and the trackside equipment, the data acquisition and transmission module is used for pre-processing, multi-channel synchronous acquisition of the vibration signals, and completing wired or wireless data transmission under the condition of strong electromagnetic interference of the subway. The present application can realize accurate separation of subway multi-source construction vibration and quantitative evaluation of structure health, and perform intelligent grading warning, so as to ensure driving safety.

Description

Technical Field

[0001] This invention relates to the field of rail transit structural health monitoring technology, specifically to a rail transit structural operation health status monitoring system and method. Background Technology

[0002] Urban rail transit is an important municipal infrastructure in the country. As the network becomes denser, existing operating lines will inevitably be affected by vibrations from nearby construction projects. Typical vibration sources include blasting, ground impact (dynamic compaction, building demolition and collapse, etc.), and mechanical construction (rotary drilling, pile driving, etc.). If these vibrations are not properly controlled, they can easily lead to deformation, misalignment, cracking and leakage of tunnel lining segments, and even cause trackside equipment to fall off and encroach on the limit, seriously threatening train operation safety.

[0003] However, traditional solutions have significant shortcomings when dealing with the vibration impact of nearby construction projects. On the one hand, existing monitoring systems and standards are usually designed for single vibration sources (such as blasting), without considering the differences in the spectral characteristics and mechanisms of multi-source vibrations such as blasting, impact, and mechanical construction. When vibration signals are highly mixed, it is difficult to accurately separate the components of each vibration source, leading to inaccurate vibration impact assessments. On the other hand, traditional evaluation methods only use vibration characteristics such as vibration velocity and frequency as control indicators, without considering the current operational health status of the rail transit structure (such as existing cracks in the lining, cumulative damage to the track bed, etc.) for differentiated evaluation. Applying a uniform threshold to structures with existing damage can easily create safety hazards. Summary of the Invention

[0004] The purpose of this invention is to provide a system and method for monitoring the operational health status of rail transit structures, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for monitoring the operational health status of rail transit structures, comprising the following steps: based on the coupling frequency response characteristics of subway structures to multi-source construction vibration sources, sensors are differentially deployed at tunnel lining annular joints, sleeper-track interface, and trackside equipment to collect multi-source engineering vibration signals caused by blasting, ground impact, and mechanical construction; the collected multi-source vibration signals are preprocessed by DC removal, bandpass filtering, outlier removal, and multi-channel synchronous calibration; an adaptive variational mode decomposition and blind source separation cascade algorithm is used to decompose and purify the preprocessed aliased low signal-to-noise ratio vibration signals, wherein the adaptive variational mode decomposition automatically determines the number of decomposition layers K and the penalty factor. The aliased signal is decomposed into several intrinsic mode function components. After removing noise components with energy below a preset threshold, the remaining components are separated by independent component analysis to obtain independent pure vibration source components with clear physical meaning. The vibration source type is identified based on a pre-built vibration source feature library, and vibration feature parameters are extracted. The separated pure vibration source feature parameters are input into a physical information deep neural network damage assessment model trained by virtual and real methods. The loss function of the model includes data fitting terms and mechanical constraint terms. The virtual and real joint training specifically involves generating structural response simulation data under different construction vibration source types, amplitudes, and distances using finite element simulation, aligning the feature distribution with the field measured data using a deep domain adversarial adaptive method, and training the model with pure vibration source features as input. The model outputs quantitative health indicators such as tunnel segment joint opening and track bed cumulative damage index. The quantitative health indicators are compared with the subway structure damage tolerance standard, and graded early warnings are triggered according to dynamically adjusted early warning thresholds, generating a health assessment report. The monitoring data, pure vibration source components, quantitative health indicators, and early warning information are uploaded to the cloud to achieve remote monitoring and data sharing.

[0006] Furthermore, the method for automatically determining the number of decomposition layers K in the adaptive variational mode decomposition is as follows: variational mode decomposition is performed incrementally starting from K=2, the center frequency of each intrinsic mode function component is calculated, and when the ratio of adjacent center frequencies is lower than a preset threshold, it is determined that over-decomposition has occurred, and the K value before the occurrence is taken as the optimal number of decomposition layers.

[0007] Furthermore, the preset threshold is 1.2 to 1.5, and is specifically set according to the dominant frequency band range of the subway tunnel structure: the threshold is 1.3 when the dominant frequency band of blasting vibration is 50 to 200 Hz, the threshold is 1.2 when the dominant frequency band of mechanical impact is 10 to 50 Hz, and the threshold is 1.5 when there is multi-source aliasing.

[0008] Furthermore, the adaptive variational mode decomposition automatically determines the penalty factor. The method is as follows: After determining the number of decomposition layers K, the objective is to minimize the average envelope entropy of all intrinsic mode function components within a preset search range. Inner traversal penalty factor Choose the one that minimizes the average envelope entropy. The value is the optimal penalty factor.

[0009] Furthermore, the preset search range is (200, 5000), and the traversal step size is 50; the average envelope entropy is calculated as follows: the envelope signal is obtained by performing Hilbert transform on each intrinsic mode function component, normalized, and the envelope entropy of each component is calculated according to the definition of entropy, and then the arithmetic mean is taken.

[0010] Furthermore, the preset energy threshold is 20% to 30% of the average energy of all intrinsic mode function components, or is dynamically determined based on the measured background noise variance.

[0011] Furthermore, in the physical information deep neural network damage assessment model, the mechanical constraint term is a residual term constructed from the linear or nonlinear prior relationship between the segment joint opening amount and the vibration energy, or a fatigue damage physical equation between the cumulative damage of the track bed and the cumulative distribution of vibration amplitude.

[0012] Furthermore, the dynamically adjusted early warning threshold is corrected in real time using Kalman filtering or Bayesian online update methods: the upper and lower limits of the threshold are dynamically updated based on the subway structure type, service life, and damage accumulation rate in historical monitoring data, so that the early warning threshold adapts to the remaining life of the structure and the current damage state.

[0013] Furthermore, the differentiated sensor deployment specifically includes: deploying high-frequency response strain sensors at the tunnel lining circumferential joints to capture dynamic strain from blasting vibrations; deploying triaxial acceleration sensors at the interface between the sleeper and the track bed to monitor instantaneous vibrations from mechanical impacts; and deploying velocity sensors on trackside equipment to monitor overall low-frequency vibration velocity.

[0014] Furthermore, a system for monitoring the operational health status of rail transit structures includes: a multi-source vibration sensing module, used to collect multi-source engineering vibration signals by differentially deploying sensors at tunnel lining annular joints, sleeper-track interface, and trackside equipment based on the coupling frequency response characteristics of the metro structure and multiple vibration sources; a data acquisition and transmission module, used to preprocess the vibration signals, acquire them synchronously through multiple channels, and adapt to the strong electromagnetic interference conditions of the metro to complete wired or wireless data transmission; a data processing and analysis module, used to separate pure vibration source components through adaptive variational mode decomposition and blind source separation cascade algorithms, and extract structure-specific characteristic parameters such as vibration frequency, amplitude, and energy distribution; a health status assessment module, used to output quantitative health indicators based on a physical information deep neural network model trained in a virtual-real joint manner, and compare them with damage tolerance standards; and an early warning and display module, used to realize graded early warnings based on dynamic thresholds, and to visualize and upload monitoring data, health indicators, and early warning information to the cloud.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. The rail transit structure operation health status monitoring system and method can accurately separate the aliased vibration signals in complex rail transit scenarios by using a multi-source vibration source coupling frequency response differential layout and cascade separation algorithm. This solves the problems of signal distortion and fuzzy vibration source identification in traditional monitoring technologies, improves the accuracy of vibration monitoring and vibration source identification, and thus adapts to the multi-source vibration monitoring needs under complex subway operating conditions.

[0016] 2. The rail transit structure operation health status monitoring system and method realizes the quantitative assessment and dynamic calibration of the health status of rail transit structures through the evaluation model of virtual and real joint training, and builds an intelligent hierarchical early warning mechanism. It can accurately predict structural safety hazards in real time, reduce operation and maintenance inspection costs and the risk of delayed fault handling, thereby ensuring the safe and efficient operation of rail transit. Attached Figure Description

[0017] Figure 1 This is a diagram of the overall system architecture of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] This invention provides a technical solution: such as Figure 1 The method for monitoring the operational health status of rail transit structures includes the following steps: Based on the coupling frequency response characteristics of subway structures to multi-source construction vibration sources, sensors are differentially deployed at tunnel lining annular joints, sleeper-track interface, and trackside equipment to collect multi-source engineering vibration signals caused by blasting, ground impact, and mechanical construction; the collected multi-source vibration signals are preprocessed by DC removal, bandpass filtering, outlier removal, and multi-channel synchronous calibration; an adaptive variational mode decomposition and blind source separation cascade algorithm is used to decompose and purify the preprocessed aliased low signal-to-noise ratio vibration signals, and the adaptive variational mode decomposition automatically determines the number of decomposition layers K and the penalty factor. The aliased signal is decomposed into several intrinsic mode function components. After removing noise components with energy below a preset threshold, the remaining components are separated by independent component analysis to obtain independent pure vibration source components with clear physical meaning. The vibration source type is identified based on a pre-built vibration source feature library, and vibration feature parameters are extracted. The separated pure vibration source feature parameters are input into a physical information deep neural network damage assessment model trained by virtual and real methods. The loss function of the model includes data fitting terms and mechanical constraint terms. The virtual and real joint training is as follows: finite element simulation is used to generate structural response simulation data under different construction vibration source types, amplitudes, and distances. The feature distribution is aligned with the field measured data using a deep domain adversarial adaptive method. The pure vibration source features are used as input to train the model. The model outputs quantitative health indicators such as tunnel segment joint opening and cumulative track bed damage index. The quantitative health indicators are then compared with the ground... The system compares the allowable damage standards for iron structures, triggers graded early warnings based on dynamically adjusted warning thresholds, and generates health assessment reports. Monitoring data, pure vibration source components, quantitative health indicators, and early warning information are uploaded to the cloud for remote monitoring and data sharing. The adaptive variational mode decomposition (AMD) automatically determines the decomposition level K by incrementing K=2, calculating the center frequencies of each intrinsic mode function component, and determining over-decomposition when the ratio of adjacent center frequencies is lower than a preset threshold. The K value before this occurs is taken as the optimal decomposition level. The preset threshold is 1.2–1.5, specifically set according to the dominant frequency band of the subway tunnel structure: 1.3 for the dominant frequency band of blasting vibration (50–200Hz), 1.2 for the dominant frequency band of mechanical impact (10–50Hz), and 1.5 for multi-source aliasing. The adaptive variational mode decomposition automatically determines the penalty factor. The method is as follows: After determining the number of decomposition layers K, the objective is to minimize the average envelope entropy of all intrinsic mode function components within a preset search range. Inner traversal penalty factor Choose the one that minimizes the average envelope entropy. The value is the optimal penalty factor; the preset search range is (200, 5000), and the traversal step size is 50; the average envelope entropy is calculated as follows: perform Hilbert transform on each intrinsic mode function component to obtain the envelope signal, normalize it, calculate the envelope entropy of each component according to the definition of entropy, and then take the arithmetic mean; the preset energy threshold is 20% to 30% of the average energy of all intrinsic mode function components, or dynamically determined according to the measured background noise variance; in the physical information deep neural network damage assessment model, the mechanical constraint term adopts the residual term constructed from the linear or nonlinear prior relationship between the segment joint opening and vibration energy, or the cumulative damage of the track bed and the vibration amplitude. The physical equations of fatigue damage between cumulative distributions; the dynamically adjusted early warning thresholds are corrected in real time using Kalman filtering or Bayesian online updating methods: the upper and lower limits of the thresholds are dynamically updated based on the subway structure type, service life, and damage accumulation rate in historical monitoring data, so that the early warning thresholds adapt to the remaining life of the structure and the current damage state; differentiated sensor deployment, specifically: high-frequency response strain sensors are deployed at the tunnel lining ring joints to capture dynamic strain of blasting vibration; triaxial acceleration sensors are deployed at the interface between the sleeper and the track bed to monitor instantaneous vibration of mechanical impact; velocity sensors are deployed on trackside equipment to monitor overall low-frequency vibration velocity.

[0020] In this invention, the sensors are arranged as follows: Tunnel lining circumferential joints: High-frequency response fiber optic strain sensors (range ±5000) are installed at the circumferential joints (4 measuring points per ring) and longitudinal joints (2 measuring points per ring) of each tunnel segment. With a response frequency of 0-200Hz and a wavelength accuracy of ±1pm, the sensor is surface-mounted or embedded and the sampling frequency is set to 500Hz to capture dynamic strain waves generated by blasting vibrations.

[0021] Sleeper-to-ballast interface: At the connection point between the sleeper end and the ballast slab (one three-dimensional measuring point is set up for each sleeper, and a monitoring section is set up every 10 meters along the line), a three-dimensional IEPE accelerometer (frequency response 0.5-5000Hz, range ±50g, sensitivity 100mV / g) is installed. The sensor is fixed by magnetic base or adhesive, and the sampling frequency is set to 2000Hz to monitor instantaneous three-dimensional vibration due to mechanical impact.

[0022] Trackside equipment: Install a magnetoelectric velocity sensor (frequency response 5-200Hz, range 0-20mm / s, sensitivity 300mV / mm / s) on the foundation or top of the cabinet of key trackside equipment such as signal lights and switch machines (one measuring point for each device), with a sampling frequency of 200Hz, to monitor the overall low-frequency vibration velocity.

[0023] All sensors are connected to the data acquisition unit via shielded twisted-pair cables, and a GPS / BeiDou time synchronization module is used to achieve time synchronization across all measurement points, with a synchronization accuracy better than ±1. .

[0024] Furthermore, the acquired raw multi-channel vibration signals are first preprocessed, specifically including: DC removal: Calculate the arithmetic mean of the signal for each channel, subtract the mean from the original signal, and eliminate the DC component caused by sensor zero drift or ambient temperature.

[0025] Bandpass filtering: Based on the energy distribution range of the vibration source (the main energy of blasting vibration is in 10-200Hz, mechanical impact is in 5-100Hz, and low-frequency vibration is in 0.5-20Hz), a second-order Butterworth bandpass filter is adopted, with the passband range set to 0.5-200Hz and the stopband attenuation ≥40dB, to eliminate high-frequency noise and ultra-low frequency trend terms.

[0026] Outlier removal: using 3 The criterion (Raida criterion) calculates the standard deviation of the signal for each channel. The amplitude exceeds the mean by ±3 The sampled points were identified as outliers and replaced with linear interpolations of adjacent normal values.

[0027] Multi-channel synchronous calibration: The start time of each data packet is marked using the pulse of second (PPS) provided by the GPS / BeiDou timing module or the timestamp of the IEEE1588 protocol to ensure that all channel signals are strictly aligned on the time axis; for micro-offsets (<1ms) caused by transmission delay, cubic spline interpolation resampling is used for compensation.

[0028] The adaptive VMD and blind source separation cascade algorithm is the core signal processing link, which is used to extract each independent vibration source component from the low signal-to-noise ratio aliased signal. Step 1: Adaptive Variational Mode Decomposition (VMD) Parameter Determination and Decomposition First, the number of decomposition levels K is automatically determined. This embodiment uses the center frequency observation method, and the specific process is as follows: (1) Initialize K=2 and set the temporary penalty factor. =2000.

[0029] (2) Perform VMD decomposition on the preprocessed signal x(t) to obtain K intrinsic mode function (IMF) components.

[0030] (3) Perform Hilbert transform on each IMF component, calculate its instantaneous frequency, and then take the time average to obtain the center frequencies f1, f2, ..., f of each component. K (Arranged in ascending order of frequency).

[0031] (4) Calculate the ratio of adjacent center frequencies. If all All are greater than the preset threshold (This embodiment is set according to the dominant frequency band: when the monitoring interval is mainly used for blasting construction) =1.3, when mechanical impact is the primary factor. =1.2, during multi-source aliasing =1.5), then let K=K+1, and repeat the above steps (2)-(4); if there exists a certain ≤ If the current K value leads to over-decomposition (two spurious components with excessively close center frequencies appear), then the optimal decomposition level is determined. =K-1.

[0032] (5) If K increases to 15 and still does not appear ≤ Then take =15 is the upper limit.

[0033] For example, in a practical application of this embodiment, a signal containing overlapping blasting and rotary drilling operations is processed. When K increases from 2 to 5, the center frequencies are 12Hz, 45Hz, 78Hz, 110Hz, and 118Hz, respectively. The calculated values ​​are... =110 / 78≈1.41>1.3, =118 / 110≈1.07<1.3, therefore, it is determined that... =4.

[0034] Sure Then, determine the penalty factor. This embodiment adopts the criterion of minimizing the average envelope entropy: (1) Fix K= ,set up The search range is [200, 5000], with a step size of [missing value]. =50.

[0035] (2) For each The value is decomposed using VMD to obtain One IMF component.

[0036] (3) For each IMF component The envelope signal A is calculated using the Hilbert transform. i(t) =|c i(t) +j·H[c i(t) Then, the envelope signal is normalized to p. i(t) =A i(t) / ∑A i(t) Then calculate the envelope entropy E of this component.i =-∑p i(t) ·log2p i(t) (Add a small amount when taking the logarithm) =1e-12 (avoiding infinity).

[0037] (4) Calculate the current The average envelope entropy E of all IMF components mean ( )=(1 / K opt )·∑E i .

[0038] (5) Traverse all Value, select to make E mean ( The smallest As the optimal penalty factor .

[0039] In this embodiment, when K opt When =4, The average envelope entropy curve calculated by iterating from 200 to 5000 is shown in... The minimum value is obtained at =1850, therefore take =1850.

[0040] Finally, using parameter combinations ( , The original signal is then subjected to final VMD decomposition to obtain... One IMF component.

[0041] Furthermore, when removing noise components from the energy threshold, the energy E of each IMF component is calculated. IMFi =∑|c i(t) |². Set the energy threshold T E =λ·E total / K opt E total The total energy of all IMF components. Take 20%-30% (25% in this embodiment). The energy is lower than T. E The IMF components were determined to be noise-dominant components and were removed. The remaining components constituted a new observation matrix X. IMF .

[0042] Furthermore, in the process of blind source separation and vibration source identification, For observation matrix X IMF Independent component analysis (ICA) is used for blind source separation. This embodiment uses the FastICA algorithm, with the following specific steps: (1) For X IMF After centering and whitening, the whitening matrix Z is obtained.

[0043] (2) Randomly initialize the unmixing matrix W and iteratively update it. ,in It is a nonlinear function (such as tanh).

[0044] (3) Repeat the iteration until convergence, obtain the separation matrix W, and calculate the independent components S= .

[0045] (4) Output M independent pure vibration source components (M≤ number of remaining IMFs).

[0046] For each pure vibration source component, its time-frequency characteristics are extracted: peak acceleration (in g), dominant frequency (Hz), vibration duration (s), 1 / 3 octave band spectrum (15 frequency bands from center frequency 1.6Hz to 200Hz), and wavelet packet energy spectrum (db4 wavelet, decomposition level 5). These characteristics are then matched with a pre-built vibration source feature library (containing sample features of typical vibration sources such as blasting, dynamic compaction, rotary drilling, pile driving, and building collapse) using Euclidean distance to identify the vibration source type of each component. For example, if a component has a dominant frequency of 80-120Hz, a duration of <0.1s, and energy concentrated in the high-frequency band, it is identified as blasting vibration.

[0047] Furthermore, for the physical information deep neural network damage assessment using virtual-real joint training, this step constructs a deep neural network (DNN) model, inputting the extracted pure vibration source feature parameters (forming feature vector X), and outputting two quantitative health indicators: tunnel segment joint opening. (unit mm) and the cumulative damage index D of the track bed (dimensionless, 0-1).

[0048] The model structure employs a multi-layer fully connected network: the number of nodes in the input layer equals the feature dimension (64 dimensions in this embodiment, including peak value, dominant frequency, energy, etc. of each vibration source), the hidden layers have 4 layers with 128, 256, 128, and 64 nodes respectively, and the output layer has 2 nodes; the activation function is ReLU, and the output layer uses linear activation; the loss function is L=L data +β·L physics ,in: L data Mean squared error (MSE) is the mean squared error between the model's predicted value and the training label. L physics As a physical constraint, this embodiment uses the segment joint opening amount. With vibrational energy E vib Linear relationship between residuals: L physics =[ pred -(a·E vib+b)]², where a and b are coefficients pre-calculated based on structural mechanics theory (e.g., determined through finite element static analysis); or for cumulative damage to the track bed, the residual term of the Miner linear accumulation criterion is used.

[0049] The weighting coefficient is 0.1-0.5 (0.2 in this embodiment).

[0050] The specific implementation of virtual-real joint training is as follows: Simulation data generation: Using a validated three-dimensional finite element model of "tunnel-soil-vibration source" (using ABAQUS software), 12 different vibration source types were simulated (blasting peak value 0.5 / 1 / 2MPa, dynamic compaction impact force 100 / 300 / 500kN, rotary drilling frequency 10 / 20 / 30Hz, etc.). Five distances (5m, 10m, 20m, 50m, 100m) were set for each vibration source, generating a total of 60 working conditions. For each working condition, vibration response and structural damage (joint opening amount, cumulative damage index of the track bed) were calculated, and corresponding 64-dimensional features were extracted. A total of 12,000 sets of simulation data pairs were generated.

[0051] Actual data acquisition: In high-risk sections of the operating subway line (e.g., approximately 200m long), in addition to the sensors in this system, high-precision fiber optic displacement gauges (accuracy 0.001mm) were installed to directly measure joint opening, and embedded strain gauges were used for track bed stress monitoring. During the 18-month monitoring period, 27 adjacent construction events were recorded (9 blasting operations, 12 dynamic compaction operations, and 6 rotary drilling operations), and a total of 324 sets of valid measured data samples were obtained.

[0052] Joint training process: First, the network is pre-trained using 12,000 sets of simulation data with a learning rate of 0.001, a batch size of 64, and 100 iterations to allow the model to initially grasp the vibration-damage mapping relationship. Then, 324 sets of real-world data are divided into a training set (260 sets) and a validation set (64 sets). Based on the pre-trained model, fine-tuning is performed, with the learning rate reduced to 0.0001 and 50 iterations. Early stopping is used to prevent overfitting.

[0053] On-site calibration and verification: In practical applications, the system compares the model output with the measured values ​​of a small number of on-site benchmarks (fixed displacement gauges) every 30 days or before and after each major construction project. If the deviation exceeds the preset threshold (e.g., joint opening deviation > 0.05mm), online adaptive calibration is triggered: the parameters of the first few layers of the network are frozen, and the last two layers are quickly fine-tuned using only the measured data from the most recent week (3-5 iterations), allowing the model to quickly adapt to changes in the current structural state.

[0054] In the dynamic early warning and health assessment report, the joint opening amount output by the model will be included. The cumulative damage index D of the track bed is compared with the allowable damage standard for subway structures. This embodiment refers to the "Subway Design Code" and tunnel maintenance standards: the allowable value of joint opening is 0.2mm under normal serviceability limit and 0.5mm under ultimate bearing capacity limit; the allowable value of the cumulative damage index of the track bed is: D<0.3 is safe, 0.3≤D<0.7 is a warning, and D≥0.7 is an alarm.

[0055] The dynamic early warning threshold adjustment adopts a Bayesian online update method: Let the initial warning threshold T0 be (e.g., D warning threshold 0.3); for each new set of monitoring data obtained, calculate the current damage index D. t Based on historical monitoring sequences D1, D2, ..., D t Kalman filtering is used to estimate the damage accumulation rate v. t Then update the warning threshold T. t =T0-k·v t •Δt, where k is the safety factor (taken as 0.8) and Δt is the time interval. If the damage accumulation rate is fast, the warning threshold is dynamically lowered to trigger the warning in advance; otherwise, the threshold is appropriately relaxed to avoid false alarms.

[0056] Specifically, the tiered early warning logic is as follows: Green (Safe): If the diameter is less than 0.15mm and D is less than 0.3, no warning is issued, and normal monitoring is maintained. Yellow (Warning): 0.15≤ If the value is less than 0.3mm or 0.3≤D<0.6, the system will issue a yellow warning and recommend increasing the monitoring frequency (from once a day to once every 4 hours) and notifying maintenance personnel to conduct on-site inspections. Orange (Severe Warning): 0.3≤ If the value is less than 0.5mm or 0.6≤D<0.8, the system will issue an orange warning and recommend speed limit measures (20km / h) and organize professional testing. Red (Alarm): If the diameter is ≥0.5mm or D≥0.8, the system will issue a red alarm and it is recommended to immediately stop operation and activate the emergency reinforcement plan.

[0057] Each time an alert is triggered, the system automatically generates a health assessment report, which includes: alert level, triggering indicators (…). The specific values ​​of D), the associated vibration source type (such as "the joint opening exceeds the limit due to adjacent blasting construction"), recommended measures, and the monitoring data curves for the most recent 7 days.

[0058] Furthermore, all raw monitoring data (after compression), separated pure vibration source components, extracted feature parameters, model-output health indicators, early warning information, and assessment reports are encrypted and uploaded to a cloud server via a 4G / 5G wireless network (using the MQTT protocol, data format is JSON). A web monitoring platform and a mobile APP are deployed on the cloud to achieve the following functions: 3D GIS visualization: Displays the location of each monitoring point on a map, using colors (green / yellow / orange / red) to represent health status, and allows users to view real-time data and historical curves by clicking. Vibration source tracking: Automatically identifies the location, type, and intensity of vibration sources for each construction operation and marks them on the timeline; Data sharing: Authorized users (such as subway operating companies, construction companies, and design institutes) can remotely access data and export reports in Excel / CSV format.

[0059] refer to Figure 1 A system for monitoring the operational health status of rail transit structures includes: a multi-source vibration sensing module, used to collect multi-source engineering vibration signals by differentially deploying sensors at tunnel lining circumferential joints, sleeper-track interface, and trackside equipment based on the coupling frequency response characteristics of the metro structure and multiple vibration sources; a data acquisition and transmission module, used to preprocess the vibration signals, acquire them synchronously through multiple channels, and complete wired or wireless data transmission to adapt to the strong electromagnetic interference conditions of the metro; a data processing and analysis module, used to separate pure vibration source components through an adaptive variational mode decomposition and blind source separation cascade algorithm, and extract structure-specific characteristic parameters such as vibration frequency, amplitude, and energy distribution; a health status assessment module, used to output quantitative health indicators based on a physical information deep neural network model trained in a virtual-real joint manner, and compare them with damage tolerance standards; and an early warning and display module, used to implement graded early warnings based on dynamic thresholds, and to visualize and upload monitoring data, health indicators, and early warning information to the cloud.

[0060] In this invention, the core of the multi-source vibration sensing module lies in its layout basis and method. In order to accurately capture the subway structure response caused by different construction vibration sources, the layout of the multi-source vibration sensing module is not uniform or random, but is based on the differentiated design of the coupling frequency response characteristics of different types of construction vibration sources on the subway structure. In practice, a refined three-dimensional finite element model of the tunnel lining, track bed, sleepers, and surrounding soil is first established for the target subway section. Excitation loads from different vibration sources, such as simulated blasting, impact, and mechanical construction, are applied to the model. Through modal analysis and transient dynamic analysis, the frequency response function curves of different parts of the subway structure (such as lining annular joints, track bed mid-span, and sleeper ends) to vibrations of different frequencies are calculated. The analysis results show that the dynamic strain response of the tunnel lining annular joint is most sensitive to high-frequency (>50Hz) blasting vibration; the acceleration response amplitude of the sleeper-track bed interface is the largest for medium-to-high frequency (10-100Hz) mechanical impact vibration; while trackside equipment (such as signals and switch machines) shows a more significant velocity response to low-frequency (<10Hz) overall vibration.

[0061] Based on the above frequency response characteristic analysis results, the sensors in this module are deployed differently as follows: Tunnel lining circumferential joints: High-frequency response resistive or fiber optic strain sensors are installed at the circumferential and longitudinal joints of each ring segment; the selected strain sensor model should meet the following requirements: response frequency up to 200Hz, range of ±5000με, used to capture dynamic strain waves generated by blasting construction, and sampling frequency set to 500Hz. Sleeper-to-ballast interface: At the connection between the sleeper end and the ballast slab, a triaxial IEPE accelerometer is installed; the selected accelerometer model should meet the following requirements: frequency response range of 0.5-5000Hz, range of ±50g, used to monitor instantaneous triaxial vibrations caused by mechanical impacts (such as large machinery movement, material loading and unloading), and sampling frequency set to 2000Hz. Trackside equipment: Magnetoelectric velocity sensors are installed on the foundations or cabinets of key trackside equipment such as signals and switch machines. For example, the selected velocity sensor model should meet the following requirements: used to monitor the overall low-frequency vibration velocity caused by ground impact or heavy machinery construction, with a sampling frequency set to 200Hz; Through the differentiated deployment based on frequency response characteristics, a sensing network that precisely matches the vibration source type is formed, providing high-quality raw data for subsequent signal processing.

[0062] The data acquisition and transmission module is connected to all sensors. To address the harsh operating conditions of the subway underground space—enclosed, humid, and subject to strong electromagnetic interference from traction currents and power cables—the data acquisition and transmission module employs the following measures: Hardware protection: The data acquisition unit adopts an industrial-grade protective shell. All input channels have built-in signal isolators and anti-aliasing filters, and use shielded twisted-pair cables to transmit signals. The shielding layer is grounded at a single point at the acquisition end to effectively suppress common-mode interference.

[0063] Synchronous acquisition: Using GPS / BeiDou timing modules or IEEE1588 precise time protocol, high-precision synchronous acquisition of all channels is achieved to ensure the time consistency of vibration signals at different locations.

[0064] Multi-channel and transmission: The data acquisition unit has the ability to acquire data simultaneously through multiple channels such as 32 / 64 / 128, which can be flexibly configured according to the number of measurement points; in tunnels, wired transmission via industrial Ethernet is preferred to ensure data stability and bandwidth; in areas without wired access or construction areas, it can switch to 4G / 5G wireless transmission mode, and the data is uploaded after encryption.

[0065] The data processing and analysis module is the core processing unit of the system. Its core lies in the configuration of a variational mode decomposition and blind source separation cascade vibration source separation unit designed for the multi-source vibration aliasing and low signal-to-noise ratio conditions of subways.

[0066] The specific implementation process of this unit is as follows: First, preprocessing is performed. The raw vibration signal received from the acquisition module is first processed by DC removal, bandpass filtering (for example, setting a bandpass filter of 0.1-200Hz according to the energy distribution of the vibration source), and outlier removal. Then, the first step of cascade separation is performed: Variational Mode Decomposition (VMD). This step adaptively decomposes the complex, nonlinear, and non-stationary aliased signal into several intrinsic mode function (IMF) components with finite bandwidth, achieving preliminary signal separation and noise reduction. The specific parameter settings are as follows: the number of mode decompositions K=10 (adjustable based on experience and signal complexity), the penalty factor α=2000, and the fidelity coefficient τ=0. For a segment of acquired raw aliased signal s(t) containing blasting and mechanical construction data, after VMD processing, K IMF components are obtained, i.e., ...

[0067] The second step in the cascade separation process is blind source separation (BSS). Because each IMF component after VMD decomposition may still contain energy from multiple vibration sources, this step uses Independent Component Analysis (ICA) to further process the matrix formed by the K IMF components. The goal is to find a demixing matrix W such that the output... The components are statistically independent as much as possible to purify the independent vibration source components with clear physical meaning. The specific process is as follows: the 10 IMF components obtained from VMD decomposition are used as input to the ICA algorithm. Through iterative algorithms such as FastICA, the unmixing matrix W is calculated, and finally multiple (let's assume there are M, M≤K) independent vibration source components are output. By comparing the time-frequency characteristics of independent components with the on-site construction logs, it is possible to identify Corresponding to blasting vibration, Corresponding to mechanical impact and vibration, etc.; After the vibration source separation is completed, the vibration characteristic analysis unit extracts the features of each pure vibration source component, including: peak value, dominant frequency, vibration duration, 1 / 3 octave band spectrum, wavelet packet energy spectrum and other vibration characteristic parameters specific to the subway structure; then, the structural response analysis unit combines the finite element model of the subway structure to analyze the correlation between these characteristic parameters and the stress of tunnel segment joints, track bed acceleration response and so on.

[0068] The health status assessment module has a built-in database of subway structural damage tolerance standards (such as the "Subway Design Code" and the "Tunnel Maintenance Technical Specification"). Its core is to deploy a deep neural network damage assessment model that takes the characteristics of the separated pure vibration source as input and is trained by a virtual-real joint training of on-site measured data and finite element simulation data. The model was constructed and validated as follows: The first step is to input the model, and after processing by the data processing and analysis module, the pure vibration characteristic parameter vector X, which corresponds one-to-one with various vibration sources, is obtained. The second step is to output the model, with the goal of providing two quantitative health indicators: tunnel segment joint opening. and cumulative damage index of track bed Dimensionless, ranging from 0 to 1.

[0069] The third step is to conduct virtual-real joint training, which includes simulation data generation, real-world data acquisition, joint training strategy and on-site calibration and verification. Simulation data generation: Using the validated "tunnel-soil-vibration source" three-dimensional finite element model, the opening of the segment joints and the stress and strain history of the track bed under construction loads of different intensities and types of blasting and impact are simulated, and the corresponding cumulative damage index is calculated. At the same time, the vibration response data of the corresponding locations in the model calculation are extracted and its vibration characteristics are simulated. A total of 10,000 sets of simulation data pairs of "vibration characteristics-damage index" are generated.

[0070] Data Acquisition: In a specific risk section of a test section or an actual operating line, both the system's sensors and high-precision fiber optic displacement gauges (for directly measuring joint opening with an accuracy of 0.01 mm) were installed simultaneously. During the one-year monitoring period, vibration data and corresponding fiber optic displacement gauge measured data were recorded for each adjacent construction activity (blasting, dynamic compaction, etc.). At the same time, the track bed damage status was regularly assessed by combining on-site inspections and non-destructive testing methods such as impact echo. A total of 500 sets of high-quality "vibration characteristics - measured damage index" data pairs were obtained.

[0071] Joint training strategy: First, a deep neural network (e.g., a multi-layer fully connected network with a structure of 64-128-256-128-64-2) is pre-trained using 10,000 sets of simulation data to enable the network to initially learn the physical mapping relationship between vibration characteristics and damage response; then, the pre-trained network is fine-tuned using 500 sets of real-world data to enable the model to adapt to various complex factors in the real environment and improve its generalization ability.

[0072] On-site measurement and calibration verification: After the model is trained, its output results are not directly used in actual applications. In this embodiment, the system will periodically or during key construction stages compare and verify the predicted joint opening value output by the model with the measured values ​​of a small number of high-precision displacement gauges (reference points for calibration) on site. If a deviation is found, the system will trigger the online adaptive calibration mechanism of the model to fine-tune the model parameters and ensure the long-term accuracy of the evaluation results.

[0073] The early warning and display module provides graded early warnings based on the quantitative health indicators output by the health status assessment module. For example, a cumulative damage index D < 0.3 for the track bed is set as a green safe state; 0.3 ≤ D < 0.6 is a yellow warning state, requiring close attention; and D ≥ 0.6 is a red alarm state, requiring immediate speed limits or reinforcement measures. All monitoring data, separated vibration source components, health indicators, and early warning information are displayed in a 3D GIS visualization through web or mobile applications, intuitively presenting the health status of the subway structure.

[0074] In this invention, the early warning and display module divides the structural health status of the subway into three levels: safe, early warning, and alarm, based on the quantitative health indicators output by the deep neural network damage assessment model. The hierarchical early warning logic is in line with the actual operation and maintenance needs of the subway, and the early warning threshold can be dynamically adjusted. The visualization display function can intuitively present the location of monitoring points, vibration source type, quantitative damage indicators and early warning information, making it convenient for subway operation and maintenance personnel to quickly and clearly grasp the structural health status of the entire line.

[0075] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-described technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for monitoring the operational health status of rail transit structures, characterized in that, Includes the following steps: Based on the coupling frequency response characteristics of subway structures to multi-source construction vibration sources, sensors are differentially deployed at tunnel lining annular joints, sleeper-to-track interface, and trackside equipment to collect multi-source engineering vibration signals caused by blasting, ground impact, and mechanical construction. The acquired multi-source vibration signals are preprocessed by DC removal, bandpass filtering, outlier removal, and multi-channel synchronous calibration. An adaptive variational mode decomposition and blind source separation cascade algorithm is used to decompose and purify the preprocessed aliased low signal-to-noise ratio vibration signal. The adaptive variational mode decomposition automatically determines the number of decomposition layers K and the penalty factor. The aliased signal is decomposed into several intrinsic mode function components. After removing the noise components with energy below a preset threshold, the remaining components are separated by independent component analysis to obtain independent pure vibration source components with clear physical meaning. The vibration source type is identified based on the pre-built vibration source feature library, and vibration feature parameters are extracted. The separated pure vibration source feature parameters are input into a physical information deep neural network damage assessment model trained with virtual and real methods. The loss function of the model includes a data fitting term and a mechanical constraint term. The virtual and real joint training specifically involves: generating structural response simulation data under different construction vibration source types, amplitudes, and distances using finite element simulation, and aligning the feature distribution with the field measured data using a deep domain adversarial adaptive method, using pure vibration source features as input to train the model; the model outputs quantitative health indicators such as tunnel segment joint opening and cumulative track bed damage index. The quantitative health indicators are compared with the allowable standards for subway structural damage. Based on the dynamically adjusted warning thresholds, a graded warning is triggered, and a health assessment report is generated. Monitoring data, pure vibration source components, quantitative health indicators, and early warning information are uploaded to the cloud to achieve remote monitoring and data sharing.

2. The method for monitoring the operational health status of rail transit structures according to claim 1, characterized in that, The method for automatically determining the number of decomposition levels K in adaptive variational mode decomposition is as follows: Variational mode decomposition is performed incrementally starting from K=2, and the center frequency of each intrinsic mode function component is calculated. When the ratio of adjacent center frequencies is lower than a preset threshold, it is determined that over-decomposition has occurred, and the K value before the occurrence is taken as the optimal decomposition level.

3. The method for monitoring the operational health status of rail transit structures according to claim 2, characterized in that, The preset threshold is 1.2 to 1.5, and is specifically set according to the dominant frequency band range of the subway tunnel structure: the threshold is 1.3 when the dominant frequency band of blasting vibration is 50 to 200 Hz, the threshold is 1.2 when the dominant frequency band of mechanical impact is 10 to 50 Hz, and the threshold is 1.5 when there is multi-source aliasing.

4. The method for monitoring the operational health status of rail transit structures according to claim 1, characterized in that, The adaptive variational mode decomposition automatically determines the penalty factor. The method is as follows: After determining the number of decomposition layers K, the objective is to minimize the average envelope entropy of all intrinsic mode function components within a preset search range. Inner traversal penalty factor Choose the one that minimizes the average envelope entropy. The value is the optimal penalty factor.

5. The method for monitoring the operational health status of rail transit structures according to claim 4, characterized in that, The preset search range is (200, 5000), and the traversal step size is 50. The average envelope entropy is calculated as follows: the envelope signal is obtained by performing Hilbert transform on each intrinsic mode function component, normalized, and then the envelope entropy of each component is calculated according to the definition of entropy, and finally the arithmetic mean is taken.

6. The method for monitoring the operational health status of rail transit structures according to claim 1, characterized in that, The preset energy threshold is 20% to 30% of the average energy of all intrinsic mode function components, or it can be dynamically determined based on the measured background noise variance.

7. The method for monitoring the operational health status of rail transit structures according to claim 1, characterized in that, In the physical information deep neural network damage assessment model, the mechanical constraint term is a residual term constructed from the linear or nonlinear prior relationship between the segment joint opening amount and the vibration energy, or a fatigue damage physical equation between the cumulative damage of the track bed and the cumulative distribution of vibration amplitude.

8. The method for monitoring the operational health status of rail transit structures according to claim 1, characterized in that, The dynamically adjusted early warning threshold is corrected in real time using Kalman filtering or Bayesian online update methods: the upper and lower limits of the threshold are dynamically updated based on the subway structure type, service life, and damage accumulation rate in historical monitoring data, so that the early warning threshold adapts to the remaining life of the structure and the current damage status.

9. The method for monitoring the operational health status of rail transit structures according to claim 1, characterized in that, The differentiated sensor deployment is specifically as follows: High-frequency response strain sensors are installed at the circumferential joints of the tunnel lining to capture the dynamic strain of blasting vibrations. A three-dimensional acceleration sensor is installed at the interface between the sleeper and the track bed to monitor instantaneous vibrations caused by mechanical impact. Speed ​​sensors are installed on trackside equipment to monitor overall low-frequency vibration speed.

10. A system using the rail transit structure operation health monitoring method as described in any one of claims 1-9, characterized in that, include: The multi-source vibration sensing module is used to collect multi-source engineering vibration signals by differentially deploying sensors in tunnel lining ring joints, sleeper-to-track interface, and trackside equipment based on the coupling frequency response characteristics of the subway structure and multiple vibration sources. The data acquisition and transmission module is used to preprocess vibration signals, acquire data synchronously through multiple channels, and adapt to the strong electromagnetic interference conditions of the subway to complete wired or wireless data transmission. The data processing and analysis module is used to separate pure vibration source components through an adaptive variational mode decomposition and blind source separation cascade algorithm, and extract structure-specific characteristic parameters such as vibration frequency, amplitude, and energy distribution. The health status assessment module is used to output quantitative health indicators based on a physical information deep neural network model trained jointly with virtual and real data, and compare them with the damage tolerance standard. The early warning and display module is used to implement tiered early warnings based on dynamic thresholds, and to visualize and upload monitoring data, health indicators, and early warning information to the cloud.

Images