High-speed rail station building railway passing area structure health monitoring system and method and medium

Abstract

The invention discloses a high-speed rail station building railway passing area structure health monitoring system and method and a medium, and the system comprises a concrete structure self-sensing monitoring subsystem which is used for collecting sensing data; the image acquisition and recognition subsystem is used for acquiring surface images of the concrete structure and the steel structure, inputting a surface defect recognition model and outputting a structured surface defect recognition result; the data acquisition and preprocessing unit subsystem is used for receiving the sensing data and the structured surface defect recognition result, and respectively performing preprocessing and feature extraction to obtain a sensing data feature vector and an image feature matrix; and the data fusion and analysis subsystem is used for receiving the sensing data feature vector and the image feature matrix, inputting the sensing data feature vector and the image feature matrix into a structure defect recognition model and outputting a structure defect recognition result. According to the invention, all-around, accurate and real-time monitoring of the concrete structure and the steel structure of the railway passing area of the high-speed rail station building from inside to outside is realized.

Description

Technical Field

[0001] This invention relates to the field of structural health monitoring technology for infrastructure, and in particular to a structural health monitoring system, method and medium for railway crossing areas of high-speed railway stations. Background Technology

[0002] In the field of structural health monitoring in railway crossing areas of high-speed railway stations, existing technologies mainly rely on discrete sensor deployment and regular manual inspections. These technologies suffer from limitations such as a single sensing dimension, inability to simultaneously capture internal concrete damage and surface defects in steel structures, fragmented and independent multimodal data lacking deep fusion and intelligent correlation analysis, and insufficient system intelligence leading to a break in the chain from damage identification to maintenance decisions. Furthermore, traditional methods have poor environmental adaptability and struggle to cope with vibration, temperature, humidity, and light interference in complex operating environments, resulting in low monitoring efficiency, high false negative rates, and an inability to meet the demand for comprehensive, accurate, and real-time safety early warning for critical infrastructure. Summary of the Invention

[0003] This invention provides a structural health monitoring system, method, and medium for railway crossing areas of high-speed railway stations, in order to solve the problems of low monitoring efficiency and high false negative rate in existing structural health monitoring technologies for railway crossing areas of high-speed railway stations.

[0004] Firstly, a structural health monitoring system for railway passage areas in high-speed railway stations is provided, including:

[0005] The self-sensing monitoring subsystem for concrete structures includes a piezoelectric sensor array and a fiber optic sensor array deployed in the self-sensing concrete load-bearing structure in the railway crossing area of ​​the high-speed railway station, for collecting sensor data.

[0006] The image acquisition and recognition subsystem is used to acquire surface images of concrete and steel structures in the railway crossing area of ​​high-speed railway stations, and input the acquired surface images into a deep learning-based surface defect recognition model to output the structured surface defect recognition results of concrete and steel structures.

[0007] The data acquisition and preprocessing unit subsystem is used to receive the sensor data acquired by the concrete structure self-sensing monitoring subsystem and the structured surface defect identification results output by the image acquisition and recognition subsystem, and perform preprocessing and feature extraction respectively to obtain the sensor data feature vector and the image feature matrix.

[0008] The data fusion and analysis subsystem receives the sensor data feature vectors and image feature matrices output by the data acquisition and preprocessing unit subsystem, inputs them into the deep learning-based structural defect recognition model, and outputs the structural defect recognition results.

[0009] Furthermore, the self-sensing monitoring subsystem for concrete structures includes:

[0010] A piezoelectric sensor array is deployed in the self-sensing concrete load-bearing structure of the railway passage area of ​​the high-speed railway station. The piezoelectric sensor array is configured to transmit and receive stress waves in active detection mode and collect structural vibration response signals in passive monitoring mode.

[0011] A fiber grating sensor array is deployed in the self-sensing concrete load-bearing structure of the railway passage area of ​​the high-speed railway station and is connected to a fiber grating demodulator.

[0012] A synchronization control unit is used to be electrically connected to the piezoelectric sensor array and send a synchronization trigger signal to it, and to be communicatively connected to the fiber Bragg grating demodulator and send a synchronization trigger command to it; the fiber Bragg grating demodulator is optically connected to the fiber Bragg grating sensor array.

[0013] The central control unit, electrically connected to the synchronization control unit, is configured to execute one or more of the following triggering strategies: periodically triggering the active detection mode of the piezoelectric sensor array; triggering the passive monitoring mode of the piezoelectric sensor array by a train proximity sensor signal; or triggering the active detection mode or passive monitoring mode of the piezoelectric sensor array by a strain threshold monitored by the fiber Bragg grating sensor array.

[0014] Furthermore, the image acquisition and recognition subsystem includes:

[0015] Cameras are used to capture surface images of concrete and steel structures in the railway crossing area of ​​high-speed railway stations at preset time intervals or triggered commands.

[0016] The image recognition processing unit, which is communicatively connected to the camera, is configured as follows:

[0017] The received surface image is preprocessed, including gamma grayscale correction and size normalization, and an optical flow motion compensation algorithm is used to eliminate image blur.

[0018] The preprocessed surface image is input into the trained deep learning-based surface defect recognition model to identify surface defects in concrete and steel structures, and outputs structured surface defect recognition results containing surface defect type, location information and quantification parameters.

[0019] Furthermore, the data acquisition and preprocessing unit subsystem is configured as follows:

[0020] It receives sensor data acquired from the self-sensing monitoring subsystem of concrete structures, as well as structured surface defect identification results from the image acquisition and recognition subsystem; the sensor data includes micro-strain and temperature data from the fiber optic grating sensor array and stress wave voltage signals from the piezoelectric sensor array.

[0021] The received sensor data and the structured surface defect identification results are preprocessed separately through independent preprocessing pipelines;

[0022] Temporal features are extracted from the preprocessed sensor data as sensor data feature vectors, and spatial features are extracted from the preprocessed structured surface defect identification results as image feature matrices.

[0023] Furthermore, the process of preprocessing and feature extraction of sensor data includes:

[0024] For micro-strain and temperature data: a temperature compensation algorithm is used to eliminate the influence of temperature changes on strain readings; a sliding window mean filter is applied to the compensated micro-strain data; the filtered data is normalized; the time-domain statistical features and frequency-domain dominant frequency features of the normalized data are extracted to form a feature vector; the extracted feature vector is up-dimensional to a preset dimension through a fully connected neural network layer to obtain the micro-strain feature vector.

[0025] For stress wave voltage signals: the stress wave voltage signals are first subjected to bandpass filtering; the filtered stress wave voltage signals are then enveloped by Hilbert transform; the extracted signals are normalized and time-domain features are calculated, and the calculated time-domain features are upscaled to a preset dimension through a fully connected neural network layer to obtain the stress wave feature vector;

[0026] The micro-strain feature vector and the stress wave feature vector are concatenated to obtain the sensor data feature vector.

[0027] Furthermore, the preprocessing and feature extraction process for the structured surface defect identification results includes:

[0028] Analyze the surface defect type, location information, and quantification parameters in the structured surface defect identification results;

[0029] Surface defect types are mapped to numerical labels, and the location information and quantization parameters are normalized.

[0030] The processed surface defect types, location information, and quantization parameters are combined in a preset order to generate a two-dimensional image feature matrix of a preset size.

[0031] Furthermore, the structural defect identification model in the data fusion and analysis subsystem includes a feature encoding unit, an attention fusion unit, and a multi-task decision-making unit;

[0032] The feature encoding unit includes a sensing feature encoding subunit and an image feature encoding subunit, which are used to encode the sensing data feature vector and the image feature matrix, respectively, to obtain sensing depth features and image depth features.

[0033] The attention fusion unit uses an attention mechanism to perform weighted fusion of sensor depth features and image depth features to generate a weighted fusion feature vector.

[0034] The multi-task decision unit analyzes the weighted fusion feature vector and simultaneously outputs the type of structural defect, three-dimensional location coordinates, and damage severity index.

[0035] Furthermore, it also includes:

[0036] The early warning and decision support subsystem receives the structural defect identification results output by the data fusion and analysis subsystem and issues early warnings in conjunction with predefined multi-level early warning strategies; and generates decision recommendations including suggested maintenance measures based on the structural defect identification results according to a predefined maintenance strategy knowledge base.

[0037] Secondly, a method for monitoring the structural health of railway passageways in high-speed railway station buildings is provided, including the following steps:

[0038] Acquire sensing data of the self-sensing concrete load-bearing structure in the railway passage area of ​​the high-speed railway station building, collected by the pre-embedded piezoelectric sensing array and fiber optic grating sensing array.

[0039] Acquire surface images of concrete and steel structures in the railway passage area of ​​high-speed railway stations, input the acquired surface images into a deep learning-based surface defect recognition model, and output the structured surface defect recognition results of concrete and steel structures.

[0040] The sensor data and the structured surface defect identification results are preprocessed and feature extracted respectively to obtain the sensor data feature vector and the image feature matrix;

[0041] The sensor data feature vector and image feature matrix are input into the deep learning-based structural defect recognition model, and the structural defect recognition result is output.

[0042] Thirdly, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the method for monitoring the structural health of railway passage areas in high-speed railway stations as described above.

[0043] This invention proposes a structural health monitoring system, method, and medium for railway passageways in high-speed railway stations. Through multimodal sensing and intelligent fusion technology, it achieves comprehensive, accurate, and real-time monitoring and intelligent operation and maintenance of critical infrastructure from the inside out. Its core beneficial effects can be summarized as follows:

[0044] (1) By integrating internal sensors such as piezoelectric sensors and fiber optic grating sensors with external machine vision, the surface defects of concrete and steel structures and internal damage of concrete structures can be perceived simultaneously, thus solving the limitations of traditional single-point monitoring.

[0045] (2) Based on the fusion model of attention mechanism and deep learning, it can intelligently weight multi-source data, automatically output the structural defect identification results, realize the accurate identification and quantitative evaluation of sub-millimeter level damage, and significantly improve the defect detection accuracy;

[0046] (3) It realizes full-process automation from data collection and analysis to early warning decision-making. It can automatically trigger multi-channel alarms and generate maintenance suggestions based on real-time risks, upgrading the traditional "manual inspection" mode to "intelligent predictive maintenance", improving monitoring efficiency and significantly reducing operation and maintenance costs and safety risks. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a schematic diagram of a structural health monitoring system for railway passage areas in high-speed railway stations, provided by an embodiment of the present invention.

[0049] Figure 2 This is a flowchart of a method for monitoring the structural health of a high-speed railway station building in a railway crossing area, provided by an embodiment of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0051] In the description of this invention, it should be understood that the terms "upper," "lower," "top," "bottom," "inner," "outer," "center," "longitudinal," "lateral," "vertical," and "horizontal," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. When an element is considered to be "connected" to another element, it can be directly connected to the other element or there may be an intervening element present. Furthermore, in the description of this invention, unless otherwise stated, "a plurality of" means at least two.

[0052] like Figure 1 As shown, this invention discloses a structural health monitoring system for railway passageways in high-speed railway stations, comprising a concrete structure self-sensing monitoring subsystem 1, an image acquisition and recognition subsystem 2, a data acquisition and preprocessing unit subsystem 3, and a data fusion and analysis subsystem 4. Each subsystem is described in detail below.

[0053] The self-sensing monitoring subsystem 1 for concrete structures is used to collect sensor data, specifically including:

[0054] A piezoelectric sensor array is deployed within the self-sensing concrete load-bearing structure of the railway passage area of ​​a high-speed railway station. The piezoelectric sensor array is configured to transmit and receive stress waves in active detection mode and to collect structural vibration response signals in passive monitoring mode. In this embodiment, the piezoelectric sensor array may be composed of multiple piezoelectric smart aggregates.

[0055] A fiber optic grating sensor array is deployed within the self-sensing concrete load-bearing structure of the railway passage area in the high-speed railway station building and connected to a fiber optic grating demodulator (FBG demodulator). In this embodiment, the fiber optic grating sensor array consists of fiber optic grating sensors connected in series. The fiber optic grating demodulator has a built-in temperature compensation function, which is used to perform preliminary processing on the acquired raw wavelength data to eliminate most of the influence of temperature changes on the strain readings. This compensation aims to provide pre-purified strain data for subsequent transmission.

[0056] A synchronization control unit is used to be electrically connected to the piezoelectric sensing array and send a synchronization trigger signal to it, and to be communicatively connected to the fiber Bragg grating demodulator and send a synchronization trigger command to it; the fiber Bragg grating demodulator is optically connected to the fiber Bragg grating sensing array; the synchronization control unit directly receives data collected by the piezoelectric sensing array and receives sensing data collected by the fiber Bragg grating sensing array via the fiber Bragg grating demodulator.

[0057] The central control unit, electrically connected to the synchronization control unit, is configured to execute one or more of the following triggering strategies: periodically triggering the active detection mode of the piezoelectric sensor array; triggering the passive monitoring mode of the piezoelectric sensor array by the train proximity sensor signal; or triggering the active detection mode or passive monitoring mode of the piezoelectric sensor array by the strain threshold monitored by the fiber Bragg grating sensor array.

[0058] The self-sensing concrete load-bearing structure is made of concrete mixed with conductive reinforcing materials and is used to form the supporting columns, beams, or the base of the canopy above the track in the railway passage area of ​​the high-speed railway station building. In a preferred embodiment, the conductive reinforcing material in the self-sensing concrete load-bearing structure is copper-plated steel fiber with a diameter of 0.1 mm and a length of 12 mm, with a volume fraction of 1.2%. This material is uniformly mixed in C50 concrete, and after being poured and cured for 28 days, its resistivity can be reduced to below 35 Ωm, and its compressive strength is not less than 55 MPa, thus meeting both structural load-bearing and electrical characteristic requirements.

[0059] In some preferred embodiments, each piezoelectric smart aggregate core of the piezoelectric sensing array uses a PZT-5H piezoelectric ceramic sheet, measuring 10 mm × 10 mm × 0.5 mm, with a resonant frequency of 200 kHz, encapsulated in epoxy resin to form a sphere with a diameter of 20 mm to ensure durability. The array is arranged in a grid pattern, with the placement points determined based on the finite element mechanical analysis results of the concrete load-bearing structure. In one embodiment, at the location of the maximum bending moment in the supporting column or at the mid-span of the beam, the placement density is optimized to 4 aggregates per cubic meter of concrete to ensure coverage of stress concentration areas.

[0060] In some preferred embodiments, the fiber Bragg grating (FBG) sensing array comprises 16 FBG sensors with a center wavelength spacing of 2 nm, uniformly distributed in the 1528 nm to 1560 nm band. The sensors are deployed via pre-embedded microtubes, with a total series length not exceeding 50 meters to minimize optical signal attenuation. The FBG sensing array is connected to a Micron Optics SM130 demodulator, which has a scanning frequency of 250 Hz, a range of ±1500 microstrain, -20°C to 80°C, and an accuracy of ±0.5 microstrain, ±0.1°C, for real-time acquisition of microstrain and temperature data of the structure. The acquired data is processed using a preliminary temperature compensation algorithm to eliminate the influence of temperature changes on the strain readings. The specific compensation formula is as follows:

[0061]

[0062] In the formula, T is the raw micro-strain reading directly read by the FBG demodulator; T is the current temperature value collected in real time by the fiber optic grating sensor; T0 is the reference temperature, which is 20℃ in this embodiment; K is the temperature sensitivity coefficient of the fiber optic grating sensor, which is determined to be 9.6 με / ℃ in this embodiment after calibration in a laboratory chamber between 20℃ and 60℃. The true micro-strain values ​​of the structure after eliminating the influence of temperature.

[0063] In some preferred embodiments, the synchronization control unit employs an FPGA-based hardware design and is electrically connected to the piezoelectric sensor array and fiber Bragg grating sensor array to generate a synchronization trigger signal with an accuracy of ±100 nanoseconds, ensuring the time consistency of multi-sensor data acquisition. The central control unit executes an intelligent triggering strategy, including: periodic triggering of an active detection mode for a full scan every 30 minutes; triggering a passive monitoring mode when the lidar (train proximity sensor) detects a train approaching within 100 meters; or setting a strain threshold of ±800 micro-strain, with the strain threshold monitored by the fiber Bragg grating sensor array triggering either active or passive monitoring modes. All acquired sensor data, including the voltage signal from the piezoelectric sensor and the wavelength offset data from the fiber Bragg grating sensor, are uploaded to the central control unit via the synchronization control unit for subsequent processing.

[0064] In active detection mode, at least one piezoelectric smart aggregate in the piezoelectric sensor array acts as an actuator, receiving a 5-cycle sinusoidal wave excitation with a peak-to-peak amplitude of 80 volts and a frequency of 150 kHz, emitting a stress wave. Other piezoelectric smart aggregates in the array act as receivers, synchronously receiving the response signal at a sampling rate of 1 MHz. Internal microcracks are identified by analyzing wave velocity changes, and damage is marked when wave velocity attenuation exceeds 5%. In passive monitoring mode, the piezoelectric sensor array continuously collects vibration data at a sampling rate of 50 kHz to monitor the dynamic response caused by train passage.

[0065] Image acquisition and recognition subsystem 2 is used to acquire surface images of concrete and steel structures in the railway crossing area of ​​high-speed railway stations, and input the acquired surface images into a surface defect recognition model based on deep learning, and output the structured surface defect recognition results of concrete and steel structures.

[0066] Specifically, the image acquisition and recognition subsystem includes:

[0067] An industrial-grade high-definition camera is used to capture surface images of the concrete and steel structures in the railway crossing area of ​​a high-speed railway station at preset time intervals or triggered by commands. As an example, in one embodiment, the camera uses a Basler ace 2 series 20-megapixel CMOS sensor, model acA2440-75um, with a dynamic range of 120dB, equipped with a 6x optical zoom lens with a focal length of 8-48mm and autofocus. It is deployed at key locations around the railway crossing area of ​​the high-speed railway station and mounted on vibration-damping supports to effectively suppress the low-frequency vibrations caused by train passage and ensure image acquisition stability. The camera is powered and communicates via PoE, deployed at a height of 3 meters above the ground at a 45-degree angle, covering the surface of supporting columns, the bottom of beams, and the ceiling base above the track. The acquisition triggering strategy is managed by a central control unit, supporting periodic acquisition at 30-minute intervals or instant acquisition triggered by train proximity sensor signals. The acquisition frame rate is set to 5fps to meet the requirements of high-resolution image processing.

[0068] The image recognition processing unit communicates with the camera via a wired or wireless network. This unit is mounted on an edge computing device based on the NVIDIA Jetson AGX Xavier platform. The image recognition processing unit is configured as follows:

[0069] The received surface images are preprocessed. First, gamma grayscale correction is performed, with the gamma value set to 2.2 to enhance details in low-contrast areas. Then, size normalization is performed, scaling all images to a fixed resolution of 224x224 pixels, and a bicubic interpolation algorithm is used to preserve geometric features. To eliminate image blur caused by camera vibration or target movement, a motion compensation algorithm based on the Lucas-Kanade optical flow method is integrated to calculate the pixel displacement vector field between consecutive frames, and image reconstruction is performed through inverse mapping, reducing the image blur index to below 0.05.

[0070] The preprocessed surface images are input into a trained deep learning-based surface defect recognition model to identify surface defects in concrete and steel structures. The model outputs structured surface defect recognition results containing the defect type, location information, and quantification parameters. In some preferred embodiments, the surface defect recognition model is built on the Mask R-CNN framework, with a ResNet-101 backbone network, and is trained on a custom dataset containing over 80,000 labeled samples, covering typical defects such as concrete cracks, spalling, voids, and steel structure weld defects, member buckling, and corrosion.

[0071] The output layer of the surface defect recognition model is designed with a multi-task head. The segmentation head generates a pixel-level defect segmentation mask with a crack width quantization accuracy of 0.1 mm; the detection head outputs the bounding box coordinates and defect classification confidence; and the regression head calculates quantized parameters (such as the percentage of peeling area and corrosion depth estimation). The inference accuracy, i.e., the average accuracy, of the surface defect recognition model reaches 96.5%, and the inference time per frame is less than 200 milliseconds, meeting the requirements for real-time processing.

[0072] Surface defect identification results are encapsulated into structured identification results and output as a JSON data packet. The data packet includes a timestamp, camera identifier, defect classification confidence score, bounding box coordinates, pixel-level contour point sequence, and quantization parameters. The timestamp uses ISO 8601 format with millisecond precision; the camera identifier is a unique device ID; the defect classification confidence score is a floating-point number ranging from 0 to 1; the bounding box coordinates are normalized to the 0-1 range; the pixel-level contour point sequence is a list of coordinates describing the defect shape; and the quantization parameters include the maximum crack width and the number of pixels in the defect area.

[0073] The normalization transformation formula is as follows:

[0074]

[0075] In the formula, x n The normalized parameter values ​​range from [-1, 1]; x represents the original quantization parameter values; x min This represents the minimum value of the parameter; x max This indicates the maximum value of the parameter.

[0076] The JSON packet is published to the data acquisition and preprocessing unit subsystem via the MQTT protocol for subsequent multimodal fusion. The entire system is deployed using Docker containers, supporting remote updates and monitoring to ensure long-term operational stability and maintainability.

[0077] The data acquisition and preprocessing unit subsystem 3 is used to receive the sensor data acquired by the concrete structure self-sensing monitoring subsystem 1 and the structured surface defect identification results output by the image acquisition and recognition subsystem 2, and perform preprocessing and feature extraction respectively to obtain the sensor data feature vector and the image feature matrix.

[0078] Specifically, the data acquisition and preprocessing unit subsystem 3 is configured as follows:

[0079] The system receives sensor data acquired by the self-sensing monitoring subsystem 1 for concrete structures, as well as structured surface defect identification results from the image acquisition and recognition subsystem 2. The sensor data includes micro-strain and temperature data from a fiber optic grating sensor array and stress wave voltage signals from a piezoelectric sensor array. Data input is achieved via a gigabit Ethernet interface or wireless link, using the MQTT protocol to ensure real-time performance and reliability. Micro-strain and temperature data are transmitted in binary stream format with a sampling rate of 250 Hz; the stress wave voltage signal is an analog voltage signal, converted to digital form by a 16-bit ADC; the structured data packet is in JSON format, containing timestamps, defect types, and quantization parameters.

[0080] The received sensor data and the structured surface defect identification results are preprocessed separately through independent preprocessing pipelines;

[0081] Temporal features are extracted from the preprocessed sensor data as sensor data feature vectors, and spatial features are extracted from the preprocessed structured surface defect identification results as image feature matrices.

[0082] More specifically, the process of preprocessing and feature extraction of sensor data includes:

[0083] For micro-strain and temperature data, a temperature compensation algorithm is used to eliminate the influence of temperature changes on strain readings. The specific compensation formula is the same as the aforementioned temperature compensation formula.

[0084] The compensated strain data were subjected to sliding window mean filtering with a window width of 100 sample points to suppress high-frequency noise.

[0085] The filtered data is then normalized using the same formula as described above.

[0086] The time-domain statistical features (such as mean and variance) and frequency-domain dominant frequency features (calculated through fast Fourier transform) of the normalized data are extracted to form a feature vector;

[0087] The extracted feature vectors are upscaled to a preset high dimension (e.g., 256 dimensions) by using a fully connected neural network layer to obtain micro-strain feature vectors.

[0088] The preprocessing and feature extraction process for stress wave voltage signals includes: the stress wave voltage signal is first subjected to bandpass filtering with a cutoff frequency set to 1000 Hz to 10000 Hz to remove low-frequency vibrations and high-frequency noise; the filtered signal is then enveloped using Hilbert transform; the extracted signal is normalized and time-domain features (such as peak value and root mean square value) are calculated using the same normalization formula as described above; and the calculated time-domain features are then upscaled to a preset dimension (such as 256 dimensions) through a fully connected neural network layer to obtain the stress wave feature vector.

[0089] The micro-strain feature vector and the stress wave feature vector are concatenated to obtain the sensor data feature vector.

[0090] The process of preprocessing and feature extraction of structured surface defect identification results includes:

[0091] Analyze the surface defect type, location information, and quantification parameters in the structured surface defect identification results;

[0092] Surface defect types are mapped to numerical labels (e.g., cracks are labeled as 1, corrosion as 2), and all quantification parameters (including crack width and defect area) and location information are normalized.

[0093] The processed surface defect types, location information, and quantization parameters are combined in a preset order to generate a two-dimensional image feature matrix of a preset size (e.g., 32×32).

[0094] All processed data is encapsulated into standardized feature vectors and feature matrices. Sensor data feature vectors, including micro-strain and stress wave features, are extracted from sensor data; image feature matrices, representing the geometric and quantitative information of defects, are extracted from image data. Data is transmitted to the data fusion and analysis subsystem via a PCIe interface with a transmission latency of less than 1 millisecond. The system runs on a Linux real-time kernel and employs a multi-threaded architecture to process each data stream in parallel.

[0095] The data fusion and analysis subsystem 4 is used to receive the sensor data feature vector and image feature matrix output by the data acquisition and preprocessing unit subsystem 3, input them into the deep learning-based structural defect recognition model, and output the structural defect recognition result.

[0096] The structural defect identification model in the data fusion and analysis subsystem 4 includes a feature encoding unit, an attention fusion unit, and a multi-task decision unit, which are used to receive sensor data feature vectors and image feature matrices, and output structural defect type, three-dimensional location coordinates, and damage severity index.

[0097] Specifically, the feature encoding unit comprises a sensing feature encoding subunit and an image feature encoding subunit, which process in parallel to encode the sensing data feature vector and the image feature matrix respectively, obtaining sensing depth features and image depth features. The sensing feature encoding subunit employs a 5-layer one-dimensional convolutional neural network architecture, with each layer followed by a ReLU activation function and a batch normalization layer, ultimately outputting a 256-dimensional sensing depth feature vector h. sThe image feature encoding subunit integrates an intelligent data routing module. When the input is a 32x32 image feature matrix, it automatically activates the ResNet-18 backbone network for processing. When the input data contains deformation parameters or erosion features, it activates a 50-layer residual network such as ResNet-50 to capture more complex feature patterns. The image feature encoding subunit ultimately outputs a 256-dimensional image depth feature vector h. i .

[0098] The attention fusion unit employs an attention mechanism to weightedly fuse sensor depth features and image depth features, generating a weighted fused feature vector. Specifically, it is configured as follows:

[0099] Receive 256-dimensional sensing depth features h from the feature coding unit s With 256-dimensional image depth features h i ;

[0100] The dynamic contribution weight of the sensed depth features in the fusion is calculated based on the trainable weight matrix, and the calculation formula is as follows:

[0101]

[0102] In the formula, h fusion For weighted fusion feature vectors, The contribution weight of sensing features in the fusion is defined as:

[0103]

[0104] In the formula, W s W i This is a trainable weight matrix. When ResNet-18 is used for image feature encoding, W... i The weight matrix is ​​256×256; when using the ResNet-50 deep residual network, W i The weight matrix is ​​512×256. This design ensures that the weight values ​​are in the range of 0 to 1 and sum to 1 through exponential operations and normalization, enabling the model to adaptively adjust the contribution of different modal features.

[0105] The entire fusion computation is implemented on an FPGA hardware accelerator, using a parallel pipeline architecture with a processing latency of less than 5 milliseconds; the fused 512-dimensional weighted fusion feature vector is directly transferred to the memory space of the multi-task decision unit via DMA (Direct Memory Access).

[0106] The multi-task decision unit analyzes the weighted fusion feature vector and simultaneously outputs the type of structural defect, three-dimensional location coordinates, and damage severity index; the multi-task decision unit includes a defect type identification subunit, a location regression subunit, and a severity assessment subunit.

[0107] The defect type identification subunit receives a 512-dimensional weighted fused feature vector and processes it through a three-layer fully connected network. The first layer is 512×256 dimensions, the second layer is 256×128 dimensions, and the third layer is 128×N dimensions, where N is the total number of defect categories. Each layer is followed by a ReLU activation function and a dropout layer with a dropout rate of 0.2. Finally, the probability distribution of defect types is output through a softmax function, calculated using the following formula:

[0108]

[0109] In the formula, Given a weighted fusion feature vector h fusion The conditional probability that the defect type is category c; c and k are category indices, N is the total number of defect categories; w c and b c These are the weight vector and bias term corresponding to category c, respectively;

[0110] The output includes confidence scores for categories such as cracks, spalling, holes, weld defects, and corrosion, with an accuracy of 0.001.

[0111] The location regression sub-unit employs a center point offset-based regression strategy. This sub-unit contains four fully connected layers, which sequentially map the 512-dimensional input to 256, 128, and 64 dimensions, ultimately outputting a 6-dimensional coordinate vector. The first three dimensions represent the absolute coordinates of the defect center point, while the latter three dimensions represent the extended dimensions in the three dimensions. The coordinate output uses a linear activation function, achieving a numerical accuracy of 0.1 mm.

[0112] The severity assessment subunit maps the fused feature vector into a damage severity index using a deep neural network. This network employs a hierarchical processing structure: the input layer processes 512-dimensional features, the first hidden layer contains 64 neurons, the second hidden layer contains 16 neurons, and the final output layer uses one neuron and a sigmoid function to compress the output to the 0-1 range, generating a damage severity index. This index is defined in segments: 0 to 0.3 represents mild damage, 0.3 to 0.7 represents moderate damage, and 0.7 to 1 represents severe damage.

[0113] All output results are encapsulated into structured JSON data packets, including fields such as timestamp, device ID, defect type confidence level, 3D coordinates, and severity index. The system uploads data to the cloud monitoring platform in real time via a gigabit Ethernet interface, with a sampling interval of 100 milliseconds. It should be noted that the sensor array's deployment locations are known and possess inherent spatial location attributes. When a sensor detects an anomaly, its physical coordinates are known. When the image acquisition and recognition subsystem identifies a defect in the vicinity of that sensor, it can fuse the sensor's spatial location information with the visual information of the image, thereby assigning a high-precision 3D coordinate to the defect.

[0114] In some preferred embodiments, it further includes:

[0115] The early warning and decision support subsystem 5 is used to receive the structural defect identification results output by the data fusion and analysis subsystem 4, and to issue early warnings in combination with predefined multi-level early warning strategies; and to generate decision recommendations containing suggested maintenance measures based on the structural defect identification results according to the predefined maintenance strategy knowledge base.

[0116] The early warning and decision support subsystem 5 monitors the structural defect identification results output by the data fusion and analysis subsystem 4 in real time, focusing on tracking changes in the damage severity index. For example, in this embodiment, when the detected defect severity reaches a preset Level II or higher early warning threshold, a multi-channel early warning release process is immediately initiated. The early warning thresholds adopt a three-level classification system: Level I corresponds to an attention state threshold of 0.3, Level II corresponds to an early warning state threshold of 0.7, and Level III corresponds to a danger state threshold of 0.9.

[0117] The early warning system is activated simultaneously through three core channels. The audible and visual alarm channel activates the on-site industrial alarm, emitting an intermittent 105-decibel alarm sound and triggering a red rotating warning light; the SMS notification channel sends an early warning message via the operator's gateway to the pre-set mobile phone number of the responsible person, containing key information such as the location, type, and urgency of the defect; and the system pop-up channel displays a full-screen early warning window prominently on the monitoring software interface, requiring manual confirmation to receive.

[0118] The system has a built-in maintenance strategy knowledge base, containing standard handling procedures for typical structural defects. When an alert is triggered, the system automatically retrieves the matching maintenance solution from the knowledge base based on the identified defect type. The solution generation module quickly outputs a decision report, which includes practical information such as core construction steps, safety precautions, a material requirements list, and a preliminary cost estimate. The entire process is fully automated, completing from alert triggering to report generation within 5 seconds, ensuring a rapid response to serious structural defects.

[0119] The above-described embodiment discloses a structural health monitoring system for railway passageways in high-speed railway stations. Through multimodal perception and intelligent fusion technology, it achieves comprehensive, accurate, and real-time monitoring and intelligent operation and maintenance of critical infrastructure from the inside out. Its core beneficial effects can be summarized as follows: It integrates internal sensors such as piezoelectric sensors and fiber optic grating sensors with external machine vision to simultaneously perceive surface defects in concrete and steel structures as well as internal damage to concrete structures, overcoming the limitations of traditional single-point monitoring; Based on an attention mechanism and deep learning-based fusion model, it can intelligently weight multi-source data and automatically output structural defect identification results, achieving accurate identification and quantitative assessment of sub-millimeter level damage, significantly improving defect detection accuracy; It achieves full-process automation from data acquisition and analysis to early warning decision-making, automatically triggering multi-channel alarms and generating maintenance suggestions based on real-time risks, upgrading the traditional "manual inspection" mode to "intelligent predictive maintenance," improving monitoring efficiency, and significantly reducing operation and maintenance costs and safety risks.

[0120] This invention also provides a method for monitoring the structural health of railway passageways in high-speed railway station buildings, such as... Figure 2 As shown, it includes the following steps:

[0121] S1: Acquire sensing data of the self-sensing concrete load-bearing structure in the railway passage area of ​​the high-speed railway station building, collected by the pre-embedded piezoelectric sensing array and fiber optic grating sensing array.

[0122] S2: Acquire surface images of the concrete and steel structures in the railway passage area of ​​the high-speed railway station building, and input the acquired surface images into a deep learning-based surface defect recognition model to output the structured surface defect recognition results of the concrete and steel structures.

[0123] S3: Preprocess and extract features from the sensor data and the structured surface defect identification results to obtain the sensor data feature vector and the image feature matrix;

[0124] S4: Input the sensor data feature vector and image feature matrix into the deep learning-based structural defect recognition model, and output the structural defect recognition result.

[0125] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method for monitoring the structural health of railway passage areas in high-speed railway stations as described above.

[0126] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0127] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0128] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0129] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0130] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0131] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A structural health monitoring system for railway passage areas in high-speed railway stations, characterized in that, include: The self-sensing monitoring subsystem for concrete structures includes a piezoelectric sensor array and a fiber optic sensor array deployed in the self-sensing concrete load-bearing structure in the railway crossing area of ​​the high-speed railway station, for collecting sensor data. The image acquisition and recognition subsystem is used to acquire surface images of concrete and steel structures in the railway crossing area of ​​high-speed railway stations, and input the acquired surface images into a deep learning-based surface defect recognition model to output the structured surface defect recognition results of concrete and steel structures. The data acquisition and preprocessing unit subsystem is used to receive the sensor data acquired by the concrete structure self-sensing monitoring subsystem and the structured surface defect identification results output by the image acquisition and recognition subsystem, and perform preprocessing and feature extraction respectively to obtain the sensor data feature vector and the image feature matrix. The data fusion and analysis subsystem receives the sensor data feature vectors and image feature matrices output by the data acquisition and preprocessing unit subsystem, inputs them into the deep learning-based structural defect recognition model, and outputs the structural defect recognition results.

2. The structural health monitoring system for railway passage areas in high-speed railway stations according to claim 1, characterized in that, The self-sensing monitoring subsystem for concrete structures includes: A piezoelectric sensor array is deployed in the self-sensing concrete load-bearing structure of the railway passage area of ​​the high-speed railway station. The piezoelectric sensor array is configured to transmit and receive stress waves in active detection mode and collect structural vibration response signals in passive monitoring mode. A fiber grating sensor array is deployed in the self-sensing concrete load-bearing structure of the railway passage area of ​​the high-speed railway station and is connected to a fiber grating demodulator. A synchronization control unit is used to be electrically connected to the piezoelectric sensing array and send a synchronization trigger signal to it, and to be communicatively connected to the fiber Bragg grating demodulator and send a synchronization trigger command to it; the fiber Bragg grating demodulator is optically connected to the fiber Bragg grating sensing array. The central control unit, electrically connected to the synchronization control unit, is configured to execute one or more of the following triggering strategies: periodically triggering the active detection mode of the piezoelectric sensor array; triggering the passive monitoring mode of the piezoelectric sensor array by a train proximity sensor signal; or triggering the active detection mode or passive monitoring mode of the piezoelectric sensor array by a strain threshold monitored by the fiber Bragg grating sensor array.

3. The structural health monitoring system for railway passage areas in high-speed railway stations according to claim 1, characterized in that, The image acquisition and recognition subsystem includes: Cameras are used to capture surface images of concrete and steel structures in the railway crossing area of ​​high-speed railway stations at preset time intervals or triggered commands. The image recognition processing unit, which is communicatively connected to the camera, is configured as follows: Preprocess the received surface image; The preprocessed surface image is input into the trained deep learning-based surface defect recognition model to identify surface defects in concrete and steel structures, and outputs structured surface defect recognition results containing surface defect type, location information and quantification parameters.

4. The structural health monitoring system for railway passage areas in high-speed railway stations according to claim 1, characterized in that, The data acquisition and preprocessing unit subsystem is configured as follows: It receives sensor data acquired from the self-sensing monitoring subsystem of concrete structures, as well as structured surface defect identification results from the image acquisition and recognition subsystem; the sensor data includes micro-strain and temperature data from the fiber optic grating sensor array and stress wave voltage signals from the piezoelectric sensor array. The received sensor data and the structured surface defect identification results are preprocessed separately through independent preprocessing pipelines; Temporal features are extracted from the preprocessed sensor data as sensor data feature vectors, and spatial features are extracted from the preprocessed structured surface defect identification results as image feature matrices.

5. The structural health monitoring system for railway passage areas in high-speed railway stations according to claim 4, characterized in that, The process of preprocessing and feature extraction of sensor data includes: For micro-strain and temperature data: a temperature compensation algorithm is used to eliminate the influence of temperature changes on strain readings; a sliding window mean filter is applied to the compensated micro-strain data; the filtered data is normalized; the time-domain statistical features and frequency-domain dominant frequency features of the normalized data are extracted to form a feature vector; the extracted feature vector is up-dimensional to a preset dimension through a fully connected neural network layer to obtain the micro-strain feature vector. For stress wave voltage signals: the stress wave voltage signals are first subjected to bandpass filtering; the filtered stress wave voltage signals are then enveloped by Hilbert transform; the extracted signals are normalized and time-domain features are calculated, and the calculated time-domain features are upscaled to a preset dimension through a fully connected neural network layer to obtain the stress wave feature vector; The micro-strain feature vector and the stress wave feature vector are concatenated to obtain the sensor data feature vector.

6. The structural health monitoring system for railway passage areas in high-speed railway stations according to claim 4, characterized in that, The process of preprocessing and feature extraction of structured surface defect identification results includes: Analyze the surface defect type, location information, and quantification parameters in the structured surface defect identification results; Surface defect types are mapped to numerical labels, and the location information and quantization parameters are normalized. The processed surface defect types, location information, and quantization parameters are combined in a preset order to generate a two-dimensional image feature matrix of a preset size.

7. The structural health monitoring system for railway passage areas in high-speed railway stations according to claim 1, characterized in that, The structural defect identification model in the data fusion and analysis subsystem includes a feature encoding unit, an attention fusion unit, and a multi-task decision-making unit. The feature encoding unit includes a sensing feature encoding subunit and an image feature encoding subunit, which are used to encode the sensing data feature vector and the image feature matrix, respectively, to obtain sensing depth features and image depth features. The attention fusion unit uses an attention mechanism to perform weighted fusion of sensor depth features and image depth features to generate a weighted fusion feature vector. The multi-task decision unit analyzes the weighted fusion feature vector and simultaneously outputs the type of structural defect, three-dimensional location coordinates, and damage severity index.

8. The structural health monitoring system for railway passage areas in high-speed railway stations according to claim 1, characterized in that, Also includes: The early warning and decision support subsystem is used to receive the structural defect identification results output by the data fusion and analysis subsystem, and to issue early warnings in combination with predefined multi-level early warning strategies; And based on a predefined maintenance strategy knowledge base, it generates decision recommendations including suggested maintenance measures based on the structural defect identification results.

9. A method for monitoring the structural health of railway passageways in high-speed railway station buildings, characterized in that, Includes the following steps: Acquire sensing data of the self-sensing concrete load-bearing structure in the railway passage area of ​​the high-speed railway station building, collected by the pre-embedded piezoelectric sensing array and fiber optic grating sensing array. Acquire surface images of concrete and steel structures in the railway passage area of ​​high-speed railway stations, input the acquired surface images into a deep learning-based surface defect recognition model, and output the structured surface defect recognition results of concrete and steel structures. The sensor data and the structured surface defect identification results are preprocessed and feature extracted respectively to obtain the sensor data feature vector and the image feature matrix; The sensor data feature vector and image feature matrix are input into the deep learning-based structural defect recognition model, and the structural defect recognition result is output.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for monitoring the structural health of railway passage areas in high-speed railway stations as described in claim 9.

Images