Steel structure strain data monitoring method based on distributed optical fiber sensing technology

Abstract

The application provides a steel structure strain data monitoring method based on a distributed optical fiber sensing technology, and belongs to the technical field of steel structure damage monitoring. A pre-play type layout design is performed based on a digital twin model, distributed optical fiber stress sensors are arranged at key positions of a steel structure, a partition backbone network architecture is constructed, the sensing network is divided into multiple monitoring sub-areas, each monitoring sub-area is connected to an edge computing gateway for data preprocessing and feature extraction, a hierarchical AI model is constructed based on obtained feature data, the hierarchical AI model is verified and optimized through virtual-real interaction between the hierarchical AI model and the digital twin model, damage positioning and type discrimination of the steel structure are performed by using the optimized hierarchical AI model, and steel structure strain data monitoring is completed.

Description

Technical Field

[0001] This invention belongs to the field of steel structure damage monitoring technology, specifically a method for monitoring steel structure strain data based on distributed optical fiber sensing technology. Background Technology

[0002] In the field of steel structure engineering health monitoring, traditional strain monitoring methods mainly rely on point sensors such as resistance strain gauges or fiber Bragg gratings, which suffer from problems such as low deployment density, limited spatial coverage, and difficulty in comprehensively capturing structural stress concentration and fatigue damage evolution. Furthermore, existing monitoring systems mostly employ centralized data processing architectures, resulting in large data transmission delays and poor real-time performance, failing to meet the needs of long-term, dynamic, and full-life-cycle monitoring of large steel structures. Although distributed fiber optic sensing technology can achieve continuous spatial measurement, it still faces challenges in practical applications, including temperature cross-sensitivity, large data volumes making real-time processing difficult, damage identification relying on human experience, and a lack of deep integration with physical models. Simultaneously, existing methods for damage early warning and life prediction are mostly based on simple thresholds or statistical models, failing to effectively combine artificial intelligence and digital twin technologies, resulting in low early warning accuracy, high false alarm rates, and difficulty in accurately identifying damage types and simulating their evolution. Therefore, there is an urgent need for a comprehensive monitoring method that integrates sensor deployment optimization, intelligent data processing, virtual-real interactive verification, and dynamic early warning decision-making to improve the reliability, real-time performance, and intelligence level of steel structure health monitoring. Summary of the Invention

[0003] To address the aforementioned problems, this invention proposes a method for monitoring strain data of steel structures based on distributed optical fiber sensing technology, comprising:

[0004] Based on the digital twin model of the steel structure, a pre-deployment design is carried out. Distributed fiber optic stress sensors are deployed at the key stress concentration points of the digital twin model of the steel structure. A partitioned backbone network architecture consisting of multiple monitoring sub-regions is constructed, and each monitoring sub-region independently performs data preprocessing and feature extraction.

[0005] The feature data extracted from each monitoring sub-region are input into the hierarchical AI model, which includes a local damage identification model and a global fatigue prediction model.

[0006] The predictive performance of the hierarchical AI model is optimized through a virtual-real interaction verification mechanism between the hierarchical AI model and the digital twin model. The optimized hierarchical AI model is then used to locate and identify damage types in actual steel structures, thereby enabling the monitoring of strain data of actual steel structures.

[0007] Furthermore, the key stress concentration areas include node areas, bolted connection areas, and weld seam areas;

[0008] In the node region, a gridded or surrounding fiber optic stress sensor deployment method is adopted. The fiber optic stress sensors are attached to the node region in a parallel manner to form a sensing array to capture the strain field.

[0009] In the bolted connection area, critical bolts bearing load ≥ 70% of the load threshold are arranged in a figure-eight cross-shaped layout, while ordinary bolts bearing load < 70% of the load threshold are arranged with fiber optic stress sensors along the center line of the bolt holes, and are grouped and connected in series according to adjacent areas.

[0010] In the weld seam area, multiple fiber optic stress sensors are arranged in parallel at fixed intervals along both sides of the weld seam. One fiber optic stress sensor is close to the weld seam, while the other fiber optic stress sensors are located in the heat-affected zone.

[0011] Furthermore, when deploying distributed fiber optic stress sensors, fiber optic temperature sensors are laid in parallel along the same path, and temperature compensation is achieved through the following steps.

[0012] set up Brillouin frequency shift for fiber optic temperature sensors: Calculate the current temperature change based on the Brillouin frequency shift of the fiber optic stress sensor. :

[0013] ;

[0014] in, The Brillouin frequency shift of the fiber optic temperature sensor at the reference temperature. This is the temperature sensitivity coefficient of the fiber optic temperature sensor.

[0015] Calculate the effect of temperature on the Brillouin frequency shift of the fiber optic stress sensor. :

[0016] ;in, The temperature sensitivity coefficient of the fiber optic stress sensor;

[0017] By removing the influencing factors from the total frequency shift of the fiber optic stress sensor, the strain value caused by pure mechanical strain is obtained. :

[0018] ;

[0019] in, The Brillouin frequency shift of the fiber optic stress sensor under reference conditions; This represents the strain sensitivity coefficient of the fiber optic stress sensor.

[0020] Furthermore, the data preprocessing and feature extraction include: calculating the principal stress direction and magnitude from the sensor data of the node region; extracting the strain mutation coefficient, load distribution uniformity, and strain peak ratio from the sensor data of the bolted connection region; and extracting the strain gradient difference, micro-strain accumulation, and strain fluctuation frequency features from the sensor data of the welded seam region to construct a time series feature matrix.

[0021] Furthermore, the time series feature matrix is ​​a two-dimensional matrix of dimension T×F, where T represents the number of time steps and F represents the number of features extracted from all monitoring sub-regions.

[0022] Furthermore, the hierarchical AI model employs absolute position encoding. :

[0023] ;

[0024] Where pos is the load cycle number corresponding to the current time step, and N max The absolute position code and the time series feature matrix together serve as the input to the global fatigue prediction model in the hierarchical AI model, representing the total fatigue life cycle count.

[0025] Furthermore, the local damage identification model is based on a graph convolutional network, using the sensing points within the monitoring sub-region as graph nodes and the fiber optic stress sensor path and physical connection as edges for damage identification; the global fatigue prediction model is based on a Transformer encoder, fusing absolute position encoding for fatigue life prediction.

[0026] Furthermore, the optimization of the predictive performance of the hierarchical AI model through the virtual-real interaction verification mechanism of the hierarchical AI model and the digital twin model includes: using the digital twin model to simulate the mechanical response of the steel structure under various load and damage conditions, generating virtual training data with damage labels; and inputting the virtual training data into the hierarchical AI model for enhanced training.

[0027] Furthermore, damage localization includes: node area, bolted connection area, and weld seam area; classification results include: fatigue cracks, loose bolts, and welding defects.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] By employing a pre-planned deployment design based on a digital twin model, the deployment of distributed fiber optic stress sensors at key stress concentration points in the steel structure was optimized to ensure that the sensors could effectively capture local strain fields and stress distributions. Combined with a temperature compensation mechanism, the interference of ambient temperature changes on strain measurements was eliminated, thereby significantly improving the accuracy and reliability of strain data and reducing the risk of false alarms and missed alarms.

[0030] A partitioned backbone network architecture is adopted, with each monitoring sub-region undergoing independent data preprocessing and feature extraction. Combined with a hierarchical AI model, multi-scale analysis is achieved. The local damage identification model is based on a graph convolutional network and utilizes the sensor network topology to accurately identify the location and type of damage, improving the sensitivity and specificity of damage detection.

[0031] The global fatigue prediction model, based on a Transformer encoder and incorporating absolute position encoding, effectively captures strain time-series characteristics under long-term loading, thereby accurately predicting the remaining fatigue life of steel structures. This method utilizes a virtual-real interaction verification mechanism with a digital twin model, leveraging simulated data to enhance AI model training, thus optimizing the prediction model's generalization ability and robustness, and reducing reliance on large amounts of on-site monitoring data.

[0032] A differentiated deployment strategy was adopted for different key components, achieving efficient utilization of sensor resources and avoiding over-deployment. Meanwhile, the partitioned network architecture allowed for parallel data processing, reducing system latency and enabling real-time or near-real-time monitoring, making it suitable for long-term health monitoring of large steel structures. Attached Figure Description

[0033] Figure 1 This is a flowchart of the monitoring sub-region division process of the present invention;

[0034] Figure 2 A schematic diagram illustrating the virtual planning of the deployment path for distributed fiber optic stress sensors in a digital twin model;

[0035] Figure 3 A schematic diagram of the layout of a surround-type fiber optic stress sensor;

[0036] Figure 4 A schematic diagram of the gridded fiber optic stress sensor layout;

[0037] Figure 5 This is a flowchart of the hierarchical AI model optimization process of the present invention;

[0038] Figure 6 A schematic diagram of deploying distributed fiber optic stress sensors at stress concentration points in an actual steel structure;

[0039] Figure 7 This is a schematic diagram of the digital twin model of the steel frame in a specific embodiment;

[0040] Figure 8 This is a diagram illustrating a situation where the bolts have become loose. Detailed Implementation

[0041] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0042] Step 1: Based on the digital twin model of the steel structure, a pre-deployment design is carried out. Distributed fiber optic stress sensors are deployed at key stress concentration points in the digital twin model of the steel structure to construct a partitioned backbone network architecture composed of multiple monitoring sub-regions. For example... Figure 1 As shown, it includes the following steps:

[0043] Step 1.1: Conduct a pre-determined layout design based on the digital twin model of the steel structure, and deploy distributed fiber optic stress sensors at key stress concentration points in the digital twin model of the steel structure.

[0044] Before physical deployment, a digital twin model is created based on the physical characteristics of the actual steel structure. Finite element analysis is used to simulate the structural stress, identifying stress concentration areas and fatigue hotspots. Based on this, the deployment path of distributed fiber optic stress sensors is virtually planned within the digital twin model. Figure 2 As shown. This digital twin technology is mature in structural engineering, and technicians can implement it based on finite element analysis.

[0045] Step 1.2: Construct a partitioned backbone network architecture consisting of multiple monitoring sub-regions.

[0046] a. In the node region, a gridded or surrounding fiber optic stress sensor deployment method is adopted. The fiber optic stress sensors are attached to the node region in a parallel manner to form a sensing array to capture the strain field.

[0047] Preferably, in complex beam-column joint areas, fiber optic stress sensors are deployed in a grid-like or wraparound pattern. The fiber optic stress sensors are attached parallel to each other at specific intervals on the joint area, forming a miniature distributed sensing array. This method can capture the strain field of the joint and identify the direction and magnitude of the maximum principal stress, which is crucial for assessing the complex stress state of the joint.

[0048] like Figure 3 The diagram shows a surrounding layout, with distributed fiber optic stress sensors 1#, 2#, and 3# deployed around the beam-column joint area. Figure 4 The diagram shows a gridded layout, with sensors at each point on the steel structure ( Figure 4 (A, B, C, D) are fixed fiber optic stress sensors to capture the strain field.

[0049] b. In the bolted connection area, a combination of key bolts crossing and surrounding, ordinary bolts being connected in lines, and graded series connection is adopted.

[0050] (1) Identification of critical bolts: Through finite element stress analysis of digital twin models, bolts bearing load ≥ 70% of the load threshold are screened as critical bolts. Critical bolts include, for example, the connecting bolts at the support nodes of a steel box girder cable-stayed bridge, and the bolts connecting the lower chord of the truss to the strut. These types of bolts require special monitoring of preload loss and local compressive strain.

[0051] (2) Cross-circle layout: For critical bolts, the distributed fiber optic stress sensors are arranged in a figure-eight cross-circle with the bolt axis as the center. Preferably, the number of circles is 2. The distance between the fiber optic stress sensor and the edge of the bolt head is controlled at 2-3mm to avoid damage to the fiber optic during bolt disassembly and assembly, while ensuring that the local strain field caused by the change of bolt preload can be captured.

[0052] (3) Connection layout: For ordinary bolt connection points, the load is less than 70% of the load threshold. Fiber optic stress sensors are laid out along the connection line of the bolt hole center. The distance between the fiber optic stress sensor and the edge of the bolt hole is uniformly 3-5mm. This distance can avoid the interference of stress concentration at the edge of the hole on the fiber optic signal and can accurately monitor the uniformity of load transfer of the bolt group.

[0053] (4) Series path planning: The fiber optic stress sensor path is constructed using a hierarchical series principle:

[0054] ① The number of connection points in series for a single fiber optic stress sensor shall not exceed 8. If there are many bolt connection points, they shall be grouped according to adjacent areas. The distance between adjacent connection points in a group shall be ≤1.5m. Each group shall be connected in series with an independent fiber optic stress sensor.

[0055] ② The serial path is laid along the main force direction of the bolted connection of the steel structure, such as the axial direction of the steel truss or the transverse direction of the steel box girder, avoiding areas such as weld seams and chamfers that are prone to bending of the fiber optic stress sensor. The curvature radius of the fiber optic stress sensor is ≥30mm, which meets the minimum bending radius requirement of the fiber optic stress sensor and avoids signal attenuation.

[0056] ③ A margin of 1-2m is reserved at both ends of each series path to facilitate the connection and maintenance of subsequent demodulation equipment.

[0057] c. In the weld seam area, multiple fiber optic stress sensors are arranged in parallel at fixed intervals along both sides of the weld seam, with one fiber optic stress sensor close to the weld seam and the rest located in the heat-affected zone.

[0058] Along both sides of the weld seam, two to three fiber optic stress sensors are arranged in parallel at minute intervals of 5-10 mm, one close to the weld seam and the other within the heat-affected zone. This allows for high-resolution monitoring of strain gradient changes in the weld seam itself and its heat-affected zone.

[0059] Two single-mode fiber optic stress sensors are laid in parallel at the same point or on adjacent paths: one is a fiber optic stress sensor, sensitive to both temperature and strain; the other is a fiber optic temperature sensor, sensitive only to temperature. The two fiber optic stress sensors acquire single-ended or double-ended Brillouin frequency shift signals through the same BOTDA / BOTDR to achieve real-time temperature compensation and output pure mechanical strain.

[0060] Using a BOTDA or BOTDR system, simultaneously read the Brillouin frequency shift of two fiber optic sensors at a spatial resolution of 1 meter and a sampling frequency of at least 10 Hz:

[0061] Brillouin frequency shift for fiber optic temperature sensors: Calculate the current temperature change based on the Brillouin frequency shift of the fiber optic stress sensor. :

[0062] ;

[0063] in, The Brillouin frequency shift of the fiber optic temperature sensor at the reference temperature. This represents the temperature sensitivity coefficient of the fiber optic temperature sensor.

[0064] Based on the obtained temperature changes Calculate the effect of temperature on the Brillouin frequency shift of the fiber optic stress sensor. :

[0065] ;in, This represents the temperature sensitivity coefficient of the fiber optic stress sensor.

[0066] By removing the influencing factors from the total frequency shift of the fiber optic stress sensor, the strain value caused by pure mechanical strain is obtained. :

[0067] ;

[0068] in, The Brillouin frequency shift of the fiber optic stress sensor under reference conditions; This represents the strain sensitivity coefficient of the fiber optic stress sensor.

[0069] The calculated pure mechanical strain value The data is pushed to the monitoring platform in real time via a standard communication protocol for analysis or display. This algorithm effectively eliminates the cross-influence of temperature on strain measurements, making it suitable for applications such as structural health monitoring.

[0070] Step 2: Each monitoring sub-region undergoes independent data preprocessing and feature extraction to construct a hierarchical AI model. The feature data extracted from each monitoring sub-region is input into the hierarchical AI model, which includes a local damage identification model and a global fatigue prediction model. The predictive performance of the hierarchical AI model is optimized through a virtual-real interaction verification mechanism between the hierarchical AI model and the digital twin model. Figure 5 As shown.

[0071] Step 2.1: Extract strain characteristics of three key areas—nodes, bolted joints, and weld seams—from fiber optic stress sensor data to construct a time-series feature matrix for the hierarchical AI model.

[0072] (1) Node region feature extraction

[0073] A sensor array, formed by a gridded layout in the nodal region, is used to capture the strain field and then calculate the principal stresses. Within the nodal region, a local coordinate system (x, y) is established, and strain values ​​in orthogonal directions can be simultaneously measured using two sets of mutually perpendicular optical fibers within the grid. Using strain values ​​in orthogonal directions Calculate the magnitude and direction of the principal stresses.

[0074] First, through shear strain Calculate the direction of principal stress Reflecting the direction of stress concentration:

[0075] .

[0076] In mesh layout, the linear strain in the diagonal direction of the mesh cells is calculated. Solve for shear strain :

[0077] .

[0078] Further calculation of the principal stress magnitude :

[0079] ;

[0080] Where E is the elastic modulus. It is Poisson's ratio.

[0081] (2) Feature extraction of bolted connection area

[0082] In the bolted connection area, local strain distribution was monitored. Feature extraction included:

[0083] (a) Strain mutation coefficient:

[0084] For a complete fiber optic sensing path, the segment between adjacent bolt holes is used as a calculation unit. For each calculation unit, the strain difference between two adjacent sensing points on the fiber optic path is calculated. The strain abrupt change coefficient in this section for:

[0085] ;

[0086] in The strain difference between adjacent points i Its mean value, N is the number of points, and the strain mutation coefficient is used to detect bolt loosening or preload loss.

[0087] (b) Load distribution uniformity: quantify the strain variance of all bolted connection areas. The larger the strain variance, the more uneven the distribution.

[0088] (c) Strain peak ratio: As an auxiliary indicator, the ratio of maximum strain to average strain is calculated to enhance the ability to identify abnormal bolts.

[0089] (3) Feature extraction of weld seam area

[0090] In the weld seam and its heat-affected zone, the following features are extracted using strain data from multiple fiber optic sensors:

[0091] (a) Strain gradient difference: Calculate the strain gradient of the optical fibers on both sides of the weld. and absolute difference The formula is:

[0092] ;

[0093] Used to capture asymmetric deformation caused by cracks.

[0094] (b) Accumulated micro-strain: by integrating small-amplitude strain The curve obtained over time reflects the accumulation of fatigue damage. :

[0095] ;

[0096] (c) Strain fluctuation frequency characteristics f 主导 The dominant frequency is extracted through frequency domain analysis to distinguish between creep and crack propagation in the heat-affected zone.

[0097] (4) Constructing the time series feature matrix

[0098] A time-series feature matrix is ​​constructed by integrating the strain characteristics of all monitored sub-regions. This matrix is ​​a two-dimensional matrix with dimensions T×F, where T represents the number of time steps. If the system collects data and calculates features every minute, then T = 1440 for one day. In practice, a sliding window method is often used to generate this matrix, with a window size of 60 seconds and a sliding interval of 30 seconds to form a continuous time-series matrix. F represents the number of features. These features are a collection of various strain characteristics extracted from all monitored sub-regions, all nodes, bolts, and key weld seams, including: principal stress direction θ, principal stress magnitude, and strain mutation coefficient C. 突变 Load distribution uniformity, peak strain ratio, strain gradient difference Micro-strain accumulation Features such as strain fluctuation frequency are included. This time series feature matrix serves as the input to the local damage identification model of the hierarchical AI model, supporting local damage identification.

[0099] Step 2.2: Perform global fatigue prediction of the steel structure based on the time series feature matrix and absolute position encoding to predict the overall damage location and type of the steel structure.

[0100] Local damage identification has been achieved using the time-series feature matrix. To further locate and classify damage in the overall steel structure and adapt to its fatigue accumulation characteristics under cyclic loading, this step introduces a global fatigue prediction model that integrates physical mechanisms and data-driven approaches. This global fatigue prediction model uses the time-series feature matrix as input and integrates absolute position codes directly related to the fatigue process.

[0101] First, to endow the global fatigue prediction model with the ability to perceive the fatigue process, an absolute position encoding based on the number of load cycles is adopted. This absolute position encoding provides the global fatigue prediction model with the relative position signal of each time-series feature matrix within the structure. The formula is as follows:

[0102] ;

[0103] Where pos is the load cycle number corresponding to the current time step, and N max This is used to predict the total fatigue life cycles based on structural design and material properties. The absolute position encoding normalizes the position information to the [0,1] interval, enabling the global fatigue prediction model to clearly perceive the stage of the time series feature matrix within the structure.

[0104] Secondly, in the training and decision interpretation stages of the global fatigue prediction model, a discriminative model with a Transformer encoder as its backbone is adopted. The time-series feature matrix incorporating location encoding is input into the global fatigue prediction model for training, with the goal of outputting damage localization and type discrimination.

[0105] Damage localization includes: node areas, bolted connection areas, and weld seam areas.

[0106] The classification results include: fatigue cracks, loose bolts, and welding defects.

[0107] Step 2.3: Optimize the predictive performance of the hierarchical AI model through a virtual-real interaction verification mechanism between the hierarchical AI model and the digital twin model.

[0108] (1) Virtual scene generation and adversarial training

[0109] Digital twin models are used to simulate the mechanical response of steel structures under various extreme loads such as typhoons and earthquakes, long-term fatigue loads, and different damage conditions such as crack initiation and bolt loosening.

[0110] These simulated high-fidelity strain data, along with known real damage locations and type labels, are used as an augmented training dataset and input into a hierarchical AI model for offline training, enabling it to identify damage patterns that are rare or have not yet occurred in the real world.

[0111] (2) Comparison of virtual and real data and confidence assessment

[0112] In actual monitoring, the hierarchical AI model analyzes real sensor data and outputs preliminary damage identification and fatigue prediction results.

[0113] At the same time, the same actual load and environmental conditions are applied to the digital twin model, and the digital twin model outputs the theoretically expected structural response and damage state under this working condition.

[0114] The system performs real-time comparative analysis on the two results. When the two judgments are highly consistent, the confidence of the AI ​​prediction result is enhanced; when a significant deviation occurs, the data is marked as an anomaly to be verified.

[0115] (3) Bias-based model parameter fine-tuning and knowledge distillation

[0116] For the aforementioned abnormal cases to be verified, a large amount of virtual data with precise labels that are similar to the operating conditions of the abnormal cases are generated using a digital twin model.

[0117] By leveraging this high-quality virtual data, the global fatigue prediction model within the hierarchical AI model is fine-tuned to correct biased decision paths, thereby continuously bringing its prediction logic closer to high-fidelity physical mechanisms. Knowledge distillation is employed to infuse the physical principles inherent in the digital twin into the data-driven AI model.

[0118] (4) Adaptive optimization of dynamic threshold and early warning rules

[0119] Digital twin models can predict the range of feature value changes in a structure under different health states. Based on this, the system can dynamically adjust the feature thresholds used for damage warning in the hierarchical AI model, rather than using fixed empirical values.

[0120] For example, when a digital twin simulation shows that the overall strain level of a structure will increase at a specific temperature, the AI ​​model will adaptively increase the alarm threshold for the strain peak, thereby effectively reducing the false alarm rate and improving the accuracy of the warning.

[0121] It should be noted that the specific training algorithms, hyperparameter settings, or data labeling processes of the hierarchical AI model all adopt existing mature model training methods. Technicians can determine these parameters through routine experiments, so they will not be described in detail here.

[0122] Step 3: Use the optimized hierarchical AI model to locate and identify the type of damage to the steel structure, and realize the monitoring of steel structure strain data.

[0123] First, based on the key area deployment plan simulated in the digital twin model, distributed fiber optic stress sensors are precisely deployed at stress concentration points in the actual steel structure, such as... Figure 6 As shown, the strain field is captured, and fiber optic temperature sensors are laid in parallel. The sensor network is divided into multiple monitoring sub-regions, and each monitoring sub-region is connected to an edge computing gateway.

[0124] The extracted features, such as actual strain gradient and load distribution, are integrated and input into the optimized hierarchical AI model to output damage location and classification results, thus completing damage identification.

[0125] Example 2

[0126] This embodiment uses a 128-meter span tensioned truss steel roof of a large sports center as a specific application example to explain in detail the implementation process of the monitoring method described in this invention. The roof structure consists of 12 main trusses with complex node configurations and is subjected to long-term cyclical effects of wind-induced vibration and temperature changes, resulting in a significant risk of fatigue damage.

[0127] 1. Digital twin-driven sensor network pre-simulation and deployment

[0128] First, based on the design drawings and measured data, a high-fidelity digital twin model of the steel frame is established, such as... Figure 7 As shown in Table 1: The parameters of the digital twin model are as follows:

[0129] Table 1 Parameters of Digital Twin Model

[0130]

[0131] Through finite element analysis, the support nodes, the connection nodes between the lower chord and the strut, and the welded joints of some members were identified as stress concentration areas. Based on this, the deployment path of the distributed fiber optic stress sensors was planned, resulting in the deployment scheme shown in Table 2.

[0132] Table 2 Deployment Plan

[0133]

[0134] The entire roof is divided into 6 monitoring sub-zones, each covering 2 main trusses and connected to an edge computing gateway.

[0135] 2. Data acquisition, temperature compensation, and feature extraction

[0136] The system employs a BOTDR sensor with a spatial resolution of 1 meter and a sampling frequency of 10Hz. A temperature compensation algorithm is executed in real-time at the edge gateway of each monitoring sub-area to eliminate the effects of temperature variations and output pure mechanical strain values. .

[0137] Based on the compensated strain data, a time series feature matrix is ​​calculated and constructed in real time with a sliding window of 60 seconds (M=60).

[0138] 3. Construction, training, and deployment of hierarchical AI models

[0139] (1) Local damage identification

[0140] A lightweight local damage identification model is deployed at the edge gateway of each monitoring sub-region. The local damage identification model uses all sensor points within the monitoring sub-region as graph nodes and fiber optic paths and physical connections as edges, performing graph convolution operations. The training objective is to quickly classify and identify fatigue cracks, loose bolts, and welding defects that may occur in node areas, bolted connections, and welded seams. After training the model on over 100,000 sets of historical data in the cloud, the weights are then fixed and deployed at the edge. Figure 8 As shown, this indicates that the bolts are loose, and this method can quickly identify such situations.

[0141] (2) Global fatigue life prediction

[0142] The cloud platform integrates the time-series feature matrices of all monitored sub-regions and deploys a Transformer-based global fatigue prediction model. The global fatigue prediction model input incorporates absolute position encoding. , where N max Based on the SN curve, the estimated value is 2×10. 6 The loop continues.

[0143] 4. Comparative analysis with traditional resistance strain gauge monitoring methods

[0144] Resistance strain gauges were discretely deployed at typical nodes of the steel structure, with 8-10 measuring points per truss, resulting in a monitoring coverage of less than 20%. Data was centrally transmitted to a cloud server via wired connection, with a transmission delay of 2-3 seconds, leading to poor real-time performance. Strain measurements were significantly affected by ambient temperature, with measurement errors exceeding ±50με without temperature compensation. Damage identification relied on manual inspections and fixed threshold alarms, failing to distinguish damage types and resulting in a false alarm rate exceeding 35%. Due to the lack of spatially continuous data, damage location accuracy was only at the meter level, with an error >2 meters, and it could not capture the initiation of microcracks in the heat-affected zone of welds. The statically fixed model could not adaptively optimize according to the evolution of the structural state, leading to a significant decrease in prediction accuracy over long-term monitoring. The centralized architecture resulted in a heavy cloud computing load under massive data volumes, limiting system scalability and robustness. Table 3 shows the performance comparison.

[0145] Table 3 Comparison of Effects

[0146]

[0147] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for monitoring strain data of a steel structure based on distributed optical fiber sensing technology, characterized in that, include: Based on the digital twin model of the steel structure, a pre-deployment design is carried out. Distributed fiber optic stress sensors are deployed at the key stress concentration points of the digital twin model of the steel structure. A partitioned backbone network architecture consisting of multiple monitoring sub-regions is constructed, and each monitoring sub-region independently performs data preprocessing and feature extraction. The data preprocessing and feature extraction include: calculating the principal stress direction and magnitude from the sensor data of the node region; extracting the strain mutation coefficient, load distribution uniformity, and strain peak ratio from the sensor data of the bolt connection region; and extracting the strain gradient difference, micro-strain accumulation, and strain fluctuation frequency features from the sensor data of the weld seam region to construct a time series feature matrix. The feature data extracted from each monitoring sub-region are input into a hierarchical AI model, which uses absolute position encoding. : ； wherein pos is the load cycle number corresponding to the current time step, N max is the total fatigue life cycle number, and the absolute position encoding and the time series feature matrix are jointly used as the input of a global fatigue prediction model in a hierarchical AI model. The hierarchical AI model includes a local damage identification model and a global fatigue prediction model; The predictive performance of the hierarchical AI model is optimized through a virtual-real interaction verification mechanism between the hierarchical AI model and the digital twin model. The optimized hierarchical AI model is then used to locate and identify damage types in actual steel structures, thereby enabling the monitoring of strain data of actual steel structures.

2. The method for monitoring steel structure strain data based on distributed optical fiber sensing technology according to claim 1, characterized in that, The key stress concentration areas include the node area, the bolted connection area, and the weld seam area; In the node region, a gridded or surrounding fiber optic stress sensor deployment method is adopted. The fiber optic stress sensors are attached to the node region in a parallel manner to form a sensing array to capture the strain field. In the bolted connection area, critical bolts bearing load ≥ 70% of the load threshold are arranged in a figure-eight cross-shaped layout, while ordinary bolts bearing load < 70% of the load threshold are arranged with fiber optic stress sensors along the center line of the bolt holes, and are grouped and connected in series according to adjacent areas. In the weld seam area, multiple fiber optic stress sensors are arranged in parallel at fixed intervals along both sides of the weld seam. One fiber optic stress sensor is close to the weld seam, while the other fiber optic stress sensors are located in the heat-affected zone.

3. The method for monitoring steel structure strain data based on distributed optical fiber sensing technology according to claim 2, characterized in that, When deploying distributed fiber optic stress sensors, fiber optic temperature sensors are laid in parallel along the same path, and temperature compensation is achieved through the following steps. set up Brillouin frequency shift for fiber optic temperature sensors: Calculate the current temperature change based on the Brillouin frequency shift of the fiber optic stress sensor. : ； in, The Brillouin frequency shift of the fiber optic temperature sensor at the reference temperature. This is the temperature sensitivity coefficient of the fiber optic temperature sensor. Calculate the effect of temperature on the Brillouin frequency shift of the fiber optic stress sensor. : ;in, The temperature sensitivity coefficient of the fiber optic stress sensor; By removing the influencing factors from the total frequency shift of the fiber optic stress sensor, the strain value caused by pure mechanical strain is obtained. : ； in, The Brillouin frequency shift of the fiber optic stress sensor under reference conditions; This represents the strain sensitivity coefficient of the fiber optic stress sensor.

4. The method for monitoring steel structure strain data based on distributed optical fiber sensing technology according to claim 1, characterized in that, The time series feature matrix is ​​a two-dimensional matrix of dimension T × F, where T represents the number of time steps and F represents the number of features extracted from all monitoring sub-regions.

5. The method for monitoring steel structure strain data based on distributed optical fiber sensing technology according to claim 1, characterized in that, The local damage identification model is based on a graph convolutional network, using the sensing points within the monitoring sub-region as graph nodes and the fiber optic stress sensor path and physical connection as edges for damage identification; the global fatigue prediction model is based on a Transformer encoder, fusing absolute position encoding for fatigue life prediction.

6. The method for monitoring steel structure strain data based on distributed optical fiber sensing technology according to claim 1, characterized in that, The predictive performance of the hierarchical AI model is optimized through a virtual-real interaction verification mechanism between the hierarchical AI model and the digital twin model, including: using the digital twin model to simulate the mechanical response of the steel structure under various load and damage conditions, generating virtual training data with damage labels; and inputting the virtual training data into the hierarchical AI model for enhanced training.

7. The method for monitoring steel structure strain data based on distributed optical fiber sensing technology according to claim 1, characterized in that, Damage localization includes: node area, bolted connection area, and weld seam area; classification results include: fatigue cracks, loose bolts, and welding defects.

Images