Structural health monitoring method based on distributed strain dynamic test

Abstract

The invention relates to the technical field of engineering structure monitoring, in particular to a structure health monitoring method based on a distributed strain dynamic test. According to the technical scheme, the method comprises the steps of arrangement of a composite sensor array, acquisition of strain data, self-adaptive data transmission, operation of a data processing and analysis center and operation of a man-machine interaction interface, and the composite sensor array composed of multiple types of distributed strain sensors is arranged at an engineering structure part. The strain sensors comprise a fiber bragg grating strain sensor, a resistance strain gauge and a piezoelectric film strain sensor. According to the invention, various types of distributed strain sensors are adopted to form a composite sensor array, and comprehensive, real-time and accurate health monitoring and damage prediction of an engineering structure are realized in combination with a self-adaptive data transmission network and an advanced data processing and analysis technology; and more effective technical support is provided for safety guarantee of civil engineering structures.

Description

Technical Field

[0001] The present invention relates to the technical field of engineering structure monitoring, and in particular to a structural health monitoring method based on distributed strain dynamic testing. Background Art

[0002] In the field of civil engineering, health monitoring of engineering structures is crucial to ensure the safety, reliability and service life of the structures. As engineering structures become increasingly large, complex and diverse, traditional structural health monitoring methods have gradually exposed many limitations.

[0003] Traditional structural health monitoring mainly relies on regular manual inspections and local non-destructive testing technologies, such as ultrasonic testing and magnetic particle testing. These methods have many problems. First, manual inspections are greatly affected by human factors. The accuracy and reliability of the results depend on the experience and skill level of the inspectors. Moreover, they can only be performed at the time of inspection, and it is impossible to achieve real-time monitoring of the structure. It is difficult to capture the instantaneous changes of the structure under different working conditions, and some potential damage or degradation of structural performance may be missed. Secondly, local non-destructive testing technology can usually only detect local areas of the structure, and it is difficult to conduct comprehensive and continuous monitoring of large and complex structures. Moreover, its operation often requires interrupting the normal use of the structure or performing special treatment on the surface of the structure, which is not convenient for long-term and continuous structural health monitoring.

[0004] Some existing structural health monitoring systems based on a single sensor, such as systems that only use fiber Bragg grating strain sensors or resistance strain gauges, can provide certain strain data, but due to the limited measurement range and characteristics of a single sensor, they cannot adapt to complex environments and structural stress conditions. For example, although fiber Bragg grating strain sensors have the advantages of high precision and anti-electromagnetic interference, they may not be sensitive enough to monitor the local micro-deformation and dynamic strain response of certain structures; and resistance strain gauges are easily affected by environmental factors such as temperature, and their measurement results may produce large errors in complex environments. At the same time, existing monitoring systems usually use a single transmission method for data transmission, which lacks adaptability to sensors in different locations and of different types, resulting in unstable data transmission, transmission delays or data loss in complex engineering environments, such as large bridges, high-rise buildings or underground structures.

[0005] In terms of data processing, traditional structural health monitoring systems often only perform simple analysis on the collected data, such as simple threshold comparison or trend analysis, and lack the deep integration of multi-source data and the use of complex algorithms, making it difficult to accurately evaluate the health status of the structure and accurately predict potential damage. Moreover, the assessment of the health status of the structure is often based only on current data, lacking analysis and prediction of the long-term performance evolution of the structure, and cannot provide sufficient decision support for preventive maintenance of the structure.

[0006] Most existing structural health monitoring systems can only provide relatively rough information in terms of anomaly detection and damage location, and it is difficult to quickly and accurately locate the damage position of the structure. In particular, the detailed monitoring and evaluation capabilities after structural anomalies are insufficient, and cannot meet the high requirements of complex engineering structures for structural safety assurance.

[0007] Therefore, in order to overcome the above-mentioned deficiencies of the prior art, the present application proposes a structural health monitoring method based on distributed strain dynamic testing. Summary of the invention

[0008] The purpose of the present invention is to address the problem of low monitoring accuracy of existing structural monitoring methods in the background technology and to propose a structural health monitoring method based on distributed strain dynamic testing.

[0009] The technical solution of the present invention is a structural health monitoring method based on distributed strain dynamic testing, comprising the following steps:

[0010] Arrange a composite sensor array composed of various types of distributed strain sensors at the engineering structure, including fiber Bragg grating strain sensors, resistance strain gauges and piezoelectric film strain sensors; use the composite sensor array to collect strain data of the engineering structure under different working conditions, including static load, dynamic load, temperature change, humidity change and other environmental and load factors in real time and synchronously;

[0011] The collected strain data is transmitted to the data processing and analysis center through an adaptive data transmission network;

[0012] In the data processing and analysis center, multimodal data fusion algorithms are used to comprehensively process strain data and supplementary data obtained from other auxiliary monitoring equipment to assess the health status of engineering structures and predict potential damage.

[0013] Optionally, the fiber grating strain sensors in the composite sensor array are arranged according to a predetermined topological structure to form a three-dimensional sensing network, which is distributed on the surface and inside of the structural component. The spacing d1 between adjacent fiber grating strain sensors on the surface is [5cm, 30cm], and the spacing d2 inside the component is [15cm, 45cm]. The arrangement of sensors at different depths follows an optimized layout based on structural finite element analysis to sense the strain field and cover the structural area.

[0014] Optionally, the resistance strain gauge, piezoelectric film strain sensor and fiber Bragg grating strain sensor are staggered, and the pasting position of the resistance strain gauge is determined according to the results of structural stress concentration theory analysis, and the piezoelectric film strain sensor is arranged at a position prone to vibration deformation according to the results of structural vibration modal analysis. The three work together to form a complementary strain monitoring system. The pasting angle θ of the resistance strain gauge is adjusted according to the direction of the principal stress, and the range is [0°, 90°], so as to obtain structural strain information to the greatest extent.

[0015] Optionally, the adaptive data transmission network adopts a hybrid transmission mode, including wired transmission and wireless transmission, wherein the wired transmission part uses industrial Ethernet, and the wireless transmission part adopts a combination of 5G and LoRa; the fiber grating strain sensor uses industrial Ethernet transmission, and the piezoelectric film strain sensor and the resistance strain gauge use 5G or LoRa transmission. Adaptive modulation and coding technology is used in the data transmission process to dynamically adjust the transmission rate and modulation mode M according to the signal strength and channel quality, where the range of R is [1Mbps, 100Mbps], and the modulation mode M can be adaptively selected from QPSK, 16QAM and 64QAM according to the channel status.

[0016] Optionally, the multimodal data fusion algorithm adopts a deep learning-based method to construct a deep neural network model DNN, and uses the strain data from different sensors and the supplementary data obtained by other auxiliary monitoring devices as auxiliary data S aux As input, strain data includes fiber Bragg grating strain sensor data S FBG , resistance strain gauge data S RS , Piezoelectric film strain sensor data S PF ,Other auxiliary monitoring equipment includes temperature sensors, humidity sensors, and acceleration sensors, and the output is the structural health status H and damage prediction D. The loss function L of the network model adopts the combination of mean square error and cross entropy, as follows:

[0017] L = αMSE(H,H pred )+(1-α)CE(D,D pred )

[0018] Among them, MSE is the mean square error, CE is the cross entropy, α is the weight coefficient, H pred and D pred They are the predicted values ​​of structural health status and damage prediction, respectively. They are trained through a large amount of historical data and integrate data from different modes.

[0019] Optionally, during the training process of the deep neural network model DNN, data enhancement technology is used to rotate, translate, scale, and add noise to the original data to increase the diversity and robustness of the training samples. FBG , the added noise N follows the normal distribution N(μ,σ 2 ), where μ ranges from [-0.1, 0.1] and σ ranges from [0.01, 0.1] to simulate different actual working conditions and noise environments.

[0020] Optionally, the data processing and analysis center performs real-time anomaly detection on the collected strain data, using a statistical analysis method based on a sliding window, assuming that the length of the sliding window is W and the mean of the strain data in the window is The standard deviation is σ S , for a newly collected data point x, if It is judged as abnormal data, where k is the set threshold coefficient, and the value range is [2,5]. At the same time, a rapid traceability analysis of the source of the abnormal data is initiated. According to the sensor position and time information corresponding to the abnormal data, the damage location is analyzed in combination with the force model of the structure.

[0021] Optionally, when the data processing and analysis center determines that there is potential damage to the structure, the local refined monitoring mode is activated, and a micro-electromechanical system sensor array is deployed in the suspected damage area to perform high-resolution measurement of the local strain gradient G, which is calculated by the following formula:

[0022]

[0023] Among them, ΔS is the strain difference measured by adjacent MEMS sensors, Δl is the distance between adjacent MEMS sensors, and the distance range is [1mm, 5mm]. The strain gradient measurement can assist in determining the location and extent of the damage.

[0024] Optionally, the historical structural health status and damage prediction data are subjected to time series analysis, and the time series data T = {H1, D1, H2, D2, …, H n ,D n} as input, predict the structural health status H for the next m time steps future and D future Damage trends, long-term trend prediction of structural health and maintenance decision support.

[0025] Optionally, the data processing and analysis center is provided with a human-computer interaction interface, which displays the real-time health status of the structure, damage prediction results, trend analysis of historical data, and status information of the sensor, and provides a remote control function. The sampling frequency f (range is [0.1Hz, 10Hz]), data transmission parameters and monitoring mode of the sensor are adjusted according to user needs, and the structural maintenance and reinforcement information input by the user is received through the human-computer interaction interface, the structural model information is updated, and closed-loop management of monitoring and maintenance is performed.

[0026] Compared with the prior art, the present invention has at least one of the following beneficial technical effects:

[0027] The present invention combines fiber Bragg grating strain sensors, resistance strain gauges and piezoelectric film strain sensors into a composite sensor array. The sensors work based on different physical principles and achieve functional complementarity in structural health monitoring. Fiber Bragg grating strain sensors are sensitive to both temperature and strain, resistance strain gauges are sensitive to small strain changes, and piezoelectric film strain sensors can quickly respond to dynamic strain. The coordinated work enables the monitoring system to not only perceive conventional strain information, but also deeply detect the strain characteristics of structures in different dimensions under complex working conditions, thereby improving the depth perception capability.

[0028] Fiber Bragg grating strain sensors construct a three-dimensional sensing network and optimize the layout based on structural finite element analysis. At the same time, resistance strain gauges and piezoelectric film strain sensors are staggered according to stress concentration theory and vibration modal analysis results, respectively, so that the sensors can accurately capture the strain changes of key parts of the structure under different stress modes, providing rich and unique data dimensions for structural health status assessment, far surpassing traditional methods in the scientificity and innovation of sensor layout.

[0029] When the amount of data increases suddenly and the channel quality allows, it quickly switches to high modulation order and high transmission rate mode to ensure the rapid transmission of large amounts of critical data.

[0030] By building a deep neural network model and introducing the attention mechanism, different sensor data and auxiliary data are fused. The attention mechanism can intelligently assign weights to different data, allowing the model to focus on key information when processing data and mine the complex potential relationships between data. The data fusion method improves the depth and accuracy of mining data value.

[0031] Multi-dimensional evaluation of refined monitoring: In the local refined monitoring mode, the strain gradient measurement of the MEMS sensor array is innovatively combined with the full-field deformation measurement of the digital image correlation (DIC) technology, and a comprehensive evaluation is performed through a fusion algorithm. This multi-dimensional evaluation method can obtain damage information from two levels: microscopic strain gradient and macroscopic full-field deformation, accurately determine the location, degree and development trend of the damage, and provide a more comprehensive and detailed perspective for structural damage assessment. It is an innovative breakthrough in the traditional single damage assessment method.

[0032] The Bayesian optimization algorithm is used to optimize the hyperparameters of the LSTM model. By defining the hyperparameter space, constructing the objective function, initializing the optimization process, and iteratively optimizing a series of fine steps, the complex interactions between the hyperparameters are taken into account to improve the accuracy of time series analysis.

[0033] The present invention adopts multiple types of distributed strain sensors to form a composite sensor array, combined with an adaptive data transmission network and advanced data processing and analysis technology, to achieve comprehensive, real-time and accurate health monitoring and damage prediction of engineering structures, providing more effective technical support for the safety of civil engineering structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 The figure is a flow chart of a structural health monitoring method based on distributed strain dynamic testing. DETAILED DESCRIPTION

[0035] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0036] Example 1

[0037] like Figure 1 As shown, the present invention proposes a structural health monitoring method based on distributed strain dynamic testing, including the arrangement of a composite sensor array, the collection of strain data, adaptive data transmission, the operation of a data processing and analysis center, and the operation of a human-computer interaction interface. Each step is described in detail below.

[0038] 1. Arrangement of composite sensor array

[0039] A composite sensor array consisting of multiple types of distributed strain sensors is arranged at key locations of the engineering structure. Specifically, for a large bridge structure, sensors are arranged at key locations such as piers, main beams, and bridge decks.

[0040] Fiber Bragg Grating Strain Sensors:

[0041] Arranged according to a predetermined topological structure, a three-dimensional sensing network is formed. On the surface of the pier, the spacing d1 between adjacent fiber Bragg grating strain sensors is set to 20cm, while inside the pier, the spacing d2 between adjacent sensors is 30cm. According to the results of the structural finite element analysis, the layout is optimized in key areas such as the bottom, middle and top of the pier so that the sensors can accurately sense the strain field. At the bottom of the pier, considering that it is under great pressure, the arrangement of sensors is encrypted to capture tiny strain changes. These fiber Bragg grating strain sensors are firmly attached to the surface of the structure or buried inside the structure through a special fiber Bragg grating sensor installation tool to ensure coordinated deformation with the structure.

[0042] Resistance strain gauge:

[0043] According to the results of the structural stress concentration theory analysis, the resistance strain gauge is pasted on the stress concentration area of ​​the structure, such as the support part of the bridge main beam and the cross-section change. The pasting angle θ is adjusted according to the direction of the principal stress. In the area where the principal stress direction is known, the pasting angle of the resistance strain gauge is set to be consistent with the principal stress direction to ensure that the structural strain information is obtained to the greatest extent. The resistance strain gauge is pasted with high-performance glue to ensure that it is tightly combined with the surface of the structure and connected to the data acquisition circuit through wires.

[0044] Piezoelectric Film Strain Sensors:

[0045] According to the results of structural vibration modal analysis, piezoelectric film strain sensors are arranged at locations prone to vibration deformation, such as the mid-span position of the bridge deck or the cantilever end of the main beam. Since these locations are prone to vibration under vehicle driving or wind load, piezoelectric film strain sensors can effectively sense these dynamic deformations. The piezoelectric film strain sensor is fixed to the surface of the structure through a special pasting process to ensure good contact with the structure.

[0046] 2. Collection of strain data

[0047] The composite sensor array is used to collect strain data of engineering structures under different working conditions in real time and synchronously.

[0048] For static load conditions, including the weight of the bridge and the weight of fixed equipment, the sensors will continuously collect strain data to reflect the deformation state of the structure under stable load.

[0049] For dynamic load conditions, including when a vehicle passes over a bridge, the sensor can capture the dynamic strain changes caused by the vehicle's dynamic load, including the strain amplitude, frequency and phase information.

[0050] At the same time, considering the influence of environmental factors, when the temperature and humidity change, different sensors will also generate strain signals accordingly, and these signals will be collected together. All sensors collect strain data of different parts of the structure at the same timestamp to ensure data synchronization.

[0051] 3. Adaptive Data Transmission

[0052] The collected strain data is transmitted to the data processing and analysis center through an adaptive data transmission network.

[0053] For fiber Bragg grating strain sensors, since the amount of data collected is relatively large and the transmission stability is high, industrial Ethernet is preferred for transmission. The fiber Bragg grating strain sensor is connected to the Ethernet switch and the data is sent to the data processing and analysis center via industrial Ethernet.

[0054] For resistance strain gauges and piezoelectric film strain sensors, especially those arranged in highly mobile or remote locations, such as the cantilever end of a bridge or high up on a bridge tower, 5G or LoRa transmission methods are used. For example, for a piezoelectric film strain sensor arranged high up on a bridge tower, data is sent to a nearby 5G base station through a built-in 5G module, and then forwarded to a data processing and analysis center. During data transmission, adaptive modulation and coding technology is used to dynamically adjust the transmission rate and modulation method based on signal strength and channel quality. For example, when the signal strength is strong and the channel quality is good, 64QAM modulation is selected, and the transmission rate can be adjusted to 80Mbps; when the signal is weak, switch to QPSK modulation, and the transmission rate is reduced to 1Mbps to ensure efficient and reliable data transmission.

[0055] IV. Operation of the Data Processing and Analysis Center

[0056] Multimodal data fusion algorithm:

[0057] In the data processing and analysis center, the deep neural network model DNN based on deep learning is used to process the data. The strain data from different sensors and the supplementary data obtained by other auxiliary monitoring equipment are used as auxiliary data S aux As input, strain data includes fiber Bragg grating strain sensor data S FBG , resistance strain gauge data S RS , Piezoelectric film strain sensor data S PF ,Other auxiliary monitoring devices include temperature sensors, humidity sensors, and acceleration sensors.

[0058] Construction of deep neural network model DNN:

[0059] First, determine the network structure, including the input layer, multiple hidden layers, and the output layer. The input layer receives data from various sensors, and the hidden layer can use a multi-layer neuron structure, such as using the ReLU activation function for nonlinear transformation. The output layer outputs the structural health status H and damage prediction D.

[0060] Training network: Collect a large amount of historical data, including the strain data of the structure under different working conditions and the actual health status and damage of the structure, and divide these data into training set, validation set and test set. During the training process, the combination of mean square error and cross entropy is used as the loss function L, that is:

[0061] L=0.7MSE(H,H pred )+(1-0.7)CE(D,D pred )

[0062] At the same time, data enhancement technology is used to perform operations such as rotation, translation, scaling, and noise addition on the original data. FBG , the added noise N follows the normal distribution N(0.05,0.05 2 ) to simulate different actual working conditions and noise environments. Through multiple iterative training, the weights and parameters of the network are continuously adjusted until the network converges to obtain a trained deep neural network model.

[0063] Real-time anomaly detection:

[0064] Real-time anomaly detection is performed on the collected strain data using a statistical analysis method based on a sliding window. The length of the sliding window is set to 100 data points. For the newly collected data point x, the mean of the strain data in the window is calculated. and standard deviation σ S .like It is determined to be abnormal data. Where k is the set threshold coefficient, and the value range is [2,5]. For example, in bridge monitoring, when the data point of a sensor exceeds this range, the rapid traceability analysis of the source of the abnormal data is immediately started. According to the sensor position and time information corresponding to the data point, combined with the force model of the structure, the possible damage location is analyzed.

[0065] Local refined monitoring:

[0066] When the data processing and analysis center determines that there is potential damage to the structure, the local refined monitoring mode is activated. A micro-electromechanical system (MEMS) sensor array is deployed in the suspected damage area to perform high-resolution measurement of the local strain gradient. For example, when it is found that there may be damage in a certain area of ​​the bridge main beam, MEMS sensors are arranged in this area with a spacing of 2mm, and the strain gradient is calculated by the following formula:

[0067]

[0068] Where ΔS is the strain difference measured by adjacent MEMS sensors, and Δl is the distance between adjacent MEMS sensors. High-resolution strain gradient measurement can help determine the exact location and extent of damage.

[0069] Time Series Analysis:

[0070] The historical structural health status and damage prediction data are analyzed in time series using the long short-term memory network (LSTM) model. The time series data T = {H1, D1, H2, D2, …, H n ,D n} as input, predict the structural health status and damage trend for the next 5 time steps. Based on the structural health data of the past month, predict the structural health status H in the next week future and D future Damage trend provides a basis for structural maintenance decisions.

[0071] 5. Operation of the human-computer interaction interface

[0072] A human-computer interaction interface is set up in the data processing and analysis center, and the following functions are realized through the display and operating software.

[0073] Intuitively display the real-time health status of the structure: The structural health status of different parts is displayed on the interface with different colors and icons. Green indicates health, yellow indicates potential risks, and red indicates damage.

[0074] Display damage prediction results: The future damage trend of the structure is presented in the form of a chart, which allows users to intuitively understand the long-term performance evolution of the structure.

[0075] Provide trend analysis of historical data: Users can view the trend curves of the health status and damage of the structure over a period of time, which is convenient for reviewing and analyzing the performance changes of the structure.

[0076] Display the status information of the sensor: including the working status of the sensor (normal, abnormal), data transmission status, etc., to facilitate users to monitor the operation of the monitoring system.

[0077] Remote control function:

[0078] The user can adjust the sampling frequency f of the sensor according to needs, for example, adjust the sampling frequency from 5 Hz to 2 Hz to meet different monitoring needs.

[0079] Data transmission parameters can be adjusted, such as changing the power of the 5G module, to optimize data transmission performance.

[0080] The monitoring mode can be switched, such as switching from regular monitoring mode to intensive monitoring mode, to improve the accuracy and frequency of monitoring.

[0081] Receive user input of structural repair, reinforcement and other information, and update the structural model information. For example, when a bridge has completed partial repair, the user can enter the repair information on the interface, and the system will update the corresponding structural model to ensure the accuracy of subsequent monitoring and evaluation.

[0082] In this embodiment, the arrangement of the composite sensor array details how to arrange different types of sensors, including fiber Bragg grating strain sensors, resistance strain gauges, and piezoelectric film strain sensors, at key locations of specific engineering structures (taking large bridges as an example), based on their respective arrangement principles and processes to ensure that the sensors can effectively sense the strain of the structure. The collection of strain data explains how sensors collect strain data under different working conditions (static load, dynamic load, environmental changes), emphasizing the importance of synchronization. Adaptive data transmission: explains how to use different data transmission methods based on the characteristics of different sensors, and how to use adaptive modulation and coding technology to ensure the effectiveness of data transmission.

[0083] It is worth noting that the operations of the Data Processing and Analysis Center:

[0084] For the multimodal data fusion algorithm, the construction of the deep neural network model, the training process and the specific application of data enhancement technology are described. The specific implementation of real-time anomaly detection, including the setting of sliding windows and the process of anomaly determination. The operation of local refined monitoring, including the deployment of MEMS sensors and the calculation of strain gradients. The implementation of time series analysis, using the LSTM model to predict the future structural health status and damage trend. Through the operation of the human-computer interaction interface, the various functions of the human-computer interaction interface are demonstrated, including the specific operation of information display and remote control functions, realizing the comprehensive management and operation of the monitoring system.

[0085] Example 2

[0086] The difference between this embodiment and Embodiment 1 lies in the operation of the data processing and analysis center. The operation of the data processing and analysis center of this embodiment is described in detail below.

[0087] Multimodal data fusion algorithm:

[0088] In the data processing and analysis center, the deep neural network model DNN based on deep learning is used to process the data. The strain data from different sensors and the supplementary data obtained by other auxiliary monitoring equipment are used as auxiliary data S aux As input, strain data includes fiber Bragg grating strain sensor data S FBG , resistance strain gauge data S RS, Piezoelectric film strain sensor data S PF ,Other auxiliary monitoring devices include temperature sensors, humidity sensors, and acceleration sensors.

[0089] Construction of deep neural network model DNN:

[0090] First, determine the network structure, including the input layer, multiple hidden layers, and the output layer. The input layer receives data from various sensors, and the hidden layer can use a multi-layer neuron structure, such as using the ReLU activation function for nonlinear transformation. The output layer outputs the structural health status H and damage prediction D.

[0091] Training network: Collect a large amount of historical data, including the strain data of the structure under different working conditions and the actual health status and damage of the structure, and divide these data into training set, validation set and test set. During the training process, the combination of mean square error and cross entropy is used as the loss function L, that is:

[0092] L=0.7MSE(H,H pred )+(1-0.7)CE(D,D pred )

[0093] At the same time, data enhancement technology is used to perform operations such as rotation, translation, scaling, and noise addition on the original data. FBG , the added noise N follows the normal distribution N(0.05,0.05 2 ) to simulate different actual working conditions and noise environments. Through multiple iterative training, the weights and parameters of the network are continuously adjusted until the network converges to obtain a trained deep neural network model.

[0094] In order to further improve the effect of data fusion, the attention mechanism is introduced. In the hidden layer of the deep neural network, different attention weights are assigned to each input data, so that the network can pay different degrees of attention to the data according to its importance when processing the data. For example, higher attention weights are assigned to strain data collected at key structural parts or data collected during high-risk periods, while lower attention weights are assigned to data in a relatively stable state. The calculation of attention weights can be obtained by the following formula:

[0095]

[0096] Among them, v i It is an intermediate value calculated based on the data features, which can be calculated through linear transformation and activation function. j represents different input data. By introducing the attention mechanism, the network can focus more on important data, improve the accuracy of data fusion and the reliability of structural health status assessment.

[0097] Real-time anomaly detection:

[0098] Real-time anomaly detection is performed on the collected strain data using a statistical analysis method based on a sliding window. The length of the sliding window is set to 100 data points. For the newly collected data point x, the mean of the strain data in the window is calculated. and standard deviation σ S .like It is determined to be abnormal data. Where k is the set threshold coefficient, and the value range is [2,5]. For example, in bridge monitoring, when the data point of a sensor exceeds this range, the rapid traceability analysis of the source of the abnormal data is immediately started. According to the sensor position and time information corresponding to the data point, combined with the force model of the structure, the possible damage location is analyzed.

[0099] In order to solve the limitation of traditional anomaly detection that only considers a single indicator, in addition to statistical anomaly detection, an anomaly detection method based on physical models is introduced. A physical model is established based on the mechanical principles and material properties of the structure. For the collected strain data, the theoretical strain value S under the current working condition is predicted by the physical model. theory and compare it with the actual measured strain value S actual For comparison.

[0100] |S actual -S theory |>∈

[0101] Among them, ∈ is the set error threshold, which is determined according to the design parameters and accuracy requirements of the structure, and is also judged as abnormal data. In this way, the rationality of the data can be verified from different angles and the accuracy of abnormal detection can be improved. For example, for a bridge structure, the theoretical strain value is calculated according to the bending theory of the beam, and it is compared with the strain value measured by the sensor. When the difference between the two exceeds the set threshold, it is marked as abnormal data.

[0102] Local refined monitoring:

[0103] When the data processing and analysis center determines that there is potential damage to the structure, the local refined monitoring mode is activated. A micro-electromechanical system (MEMS) sensor array is deployed in the suspected damage area to perform high-resolution measurement of the local strain gradient. For example, when it is found that there may be damage in a certain area of ​​the bridge main beam, MEMS sensors are arranged in this area with a spacing of 2mm, and the strain gradient is calculated by the following formula:

[0104]

[0105] Where ΔS is the strain difference measured by adjacent MEMS sensors, and Δl is the distance between adjacent MEMS sensors. High-resolution strain gradient measurement can help determine the exact location and extent of damage.

[0106] On this basis, in order to more accurately assess the degree of damage, digital image correlation (DIC) technology is used to perform full-field deformation measurement of the suspected damaged area. Multiple DIC markers are arranged around the damaged area, and the images of the markers at different times are collected by the camera. The displacement field u(x, y) and strain field ∈(x, y) between the markers are calculated using the image correlation algorithm to obtain detailed deformation information of the damaged area. The strain gradient information measured by the MEMS sensor is combined with the full-field strain field information obtained by the DIC technology, and a comprehensive evaluation is performed using the following fusion algorithm:

[0107] F=w1G+w2∈ avg

[0108] Among them, w1 and w2 are weight coefficients, ∈ avg It is the average strain value obtained by DIC measurement. By optimizing the weight coefficient, the damage degree and location can be evaluated more comprehensively.

[0109] Time Series Analysis:

[0110] The historical structural health status and damage prediction data are analyzed in time series using the long short-term memory network (LSTM) model. The time series data T = {H1, D1, H2, D2, …, H n ,D n} as input, predict the structural health status and damage trend for the next 5 time steps. Based on the structural health data of the past month, predict the structural health status H in the next week future and D future Damage trend provides a basis for structural maintenance decisions.

[0111] In order to enhance the accuracy of time series analysis, the Bayesian optimization algorithm is introduced to optimize the hyperparameters of the LSTM model. The specific steps include:

[0112] Determine hyperparameters and value ranges

[0113] Learning rate: Set its value range to [10 -5 ,10 -1 ]. The learning rate controls the size of the step size of each parameter update of the model, which has a significant impact on the convergence speed and final performance of the model. Within this range, sampling is performed on a logarithmic scale to ensure that the impact of learning rates of different orders of magnitude on the model can be explored.

[0114] Number of hidden units: Considering the balance between computing resources and model complexity, the value range of the number of hidden units is set to [32, 256], with a discrete value of 16 steps. The number of hidden units determines the complexity of the feature representation that the model can learn. Different tasks and data sizes may require different numbers of hidden units to effectively capture data features.

[0115] Number of LSTM layers: The range is set to [1,3]. The number of LSTM layers determines the depth of the model. More layers can theoretically learn more complex time series features, but it is also easy to cause overfitting and increase training time.

[0116] Dropout rate: The value range is [0.1, 0.5], which is used to evenly distribute the samples. Dropout is a regularization technique to prevent overfitting. It reduces the co-adaptation between neurons by randomly dropping neurons. In practice, the value within this range can usually effectively balance the generalization ability and training efficiency of the model.

[0117] Constructing the objective function

[0118] Data preprocessing: The historical structural health status and damage prediction data are divided into input sequence X and target sequence y according to the time step t. For a time series of data with a length of N, t time steps are grouped as a group, the data of the first t-1 time steps are used as input, and the data of the tth time step is used as the target, and multiple such data pairs are constructed. At the same time, the data is normalized and mapped to the [0,1] interval to accelerate model convergence.

[0119] Model construction and training: For a given set of hyperparameters θ = {learning rate, number of hidden layer units, number of LSTM layers, dropout rate}, the LSTM model is constructed and trained as follows.

[0120] Build a model of corresponding depth according to the number of LSTM layers. If the number of LSTM layers is 1, the model structure is:

[0121] from keras.models import Sequential

[0122] from keras.layers import LSTM,Dense,Dropout

[0123] model=Sequential()

[0124] model.add(LSTM(units=number of hidden layer units, input_shape=(time step, number of features)))

[0125] model.add(Dropout(Dropout rate))

[0126] model.add(Dense(1))

[0127] model.compile(optimizer = keras.optimizers.Adam(lr = learning rate), loss = 'mse')

[0128] If the number of LSTM layers is 2, the model structure is:

[0129] model=Sequential()

[0130] model.add(LSTM(units = number of hidden layer units, input_shape = (time step, number of features), return_sequences = True))

[0131] model.add(Dropout(Dropout rate))

[0132] model.add(LSTM(units=number of hidden layer units))

[0133] model.add(Dropout(Dropout rate))

[0134] model.add(Dense(1))

[0135] model.compile(optimizer = keras.optimizers.Adam(lr = learning rate), loss = 'mse')

[0136] A multi-layer LSTM model is constructed in this way.

[0137] The model is trained using the divided training data, with the number of training epochs set to E = 50 and the batch size to B = 32. During the training process, early stopping is used to prevent overfitting, that is, if the loss on the validation set does not decrease within P = 10 consecutive epochs, the training is stopped.

[0138] Model evaluation: Use the validation set to evaluate the trained model, and use the root mean square error (RMSE) as the performance indicator to construct the objective function. RMSE can intuitively reflect the average error between the model's predicted value and the true value. Its calculation formula is:

[0139]

[0140] Where n is the number of samples in the validation set, and y i is the true value, is the predicted value of the model. The goal is to minimize J(θ).

[0141] Bayesian Optimization Process

[0142] Initialization: Randomly select M = 5 sets of hyperparameter combinations in the hyperparameter space and calculate the objective function value J(θ) corresponding to each set of hyperparameters. These M sets of initial samples are used to provide initial information for Bayesian optimization and help establish the initial relationship between hyperparameters and objective functions.

[0143] Constructing a proxy model: Based on the initial sample, a surrogate model is constructed using Gaussian Process Regression. Gaussian Process Regression can make probability predictions on the objective function values ​​of other unknown points in the hyperparameter space based on the existing sample data, providing a basis for selecting a new hyperparameter combination in the next step.

[0144] Select the next hyperparameter combination: Use the acquisition function to select the next hyperparameter combination to be evaluated from the results predicted by the surrogate model. Here, the expected improvement is used as the acquisition function, and its calculation formula is:

[0145] EI(θ)=E[max(0,f(θ best )-f(θ))]

[0146] Among them, f(θ) is the objective function value, f(θ best ) is the optimal objective function value among the currently evaluated hyperparameter combinations. The expected improvement function measures the expected improvement in the objective function value that may be brought about by selecting a certain hyperparameter combination under the current knowledge, and selects the hyperparameter combination that maximizes EI(θ) as the next evaluation point.

[0147] Iterative optimization: Calculate the objective function value corresponding to the newly selected hyperparameter combination, add it to the sample set, update the proxy model and acquisition function, and repeat the above steps until the preset number of iterations N = 20 is reached or the improvement of the objective function is less than a certain threshold ∈ = 10 -3 .

[0148] Determining the optimal hyperparameters

[0149] After the iteration, the set of hyperparameters that minimizes the objective function J(θ) is selected as the optimal hyperparameter from all evaluated hyperparameter combinations. This set of optimal hyperparameters is used to rebuild and train the LSTM model to predict the future structural health status and damage trend.

[0150] Through the above detailed Bayesian optimization process, it is possible to more effectively search the hyperparameter space of the LSTM model and find a relatively optimal set of hyperparameters, thereby improving the accuracy of time series analysis and providing a more reliable basis for health monitoring and maintenance decisions of engineering structures.

[0151] In this embodiment, traditional hyperparameter selection is based on experience or grid search, and may not be able to fully explore the hyperparameter space. Bayesian optimization uses Gaussian process regression to build a proxy model, combined with an acquisition function (such as an expected improvement function), which can more intelligently select the next hyperparameter combination for evaluation, effectively avoiding blind search, greatly improving search efficiency, and reducing unnecessary waste of computing resources. Bayesian optimization can implicitly consider the interaction between different hyperparameters when building a proxy model. For example, the learning rate and the number of hidden layer units do not affect the model performance in isolation. Bayesian optimization can capture these complex relationships and find the optimal combination of hyperparameters, rather than focusing only on the optimal value of a single hyperparameter, thereby improving the overall model performance.

[0152] This embodiment uses continuous iterative optimization and, based on the principle of minimizing the objective function (such as RMSE), the Bayesian optimization algorithm can accurately find the hyperparameter settings that optimize the model performance within the preset hyperparameter value range. This enables the LSTM model to more accurately learn data features and patterns when processing time series data such as structural health status and damage prediction, thereby improving the accuracy of predictions.

[0153] In the process of determining hyperparameters, such as optimizing the number of hidden layer units, the number of LSTM layers, and the Dropout rate, the complexity and generalization ability of the model can be effectively balanced. This avoids overfitting due to overly complex models, or models that are too simple to fully learn data features, so that the model can maintain good performance in different data sets and actual application scenarios.

[0154] In addition, the solution sets flexible and reasonable value ranges for different hyperparameters, which can adapt to a variety of time series analysis tasks and data characteristics. Whether the data scale is small or large, and the structural health monitoring data has different feature complexity, the Bayesian optimization process can find suitable hyperparameters to improve the versatility and adaptability of the model.

[0155] The above specific embodiments are only several optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant inspirations of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A structural health monitoring method based on distributed strain dynamic testing, characterized in that: The following steps are involved: Arrange a composite sensor array consisting of various types of distributed strain sensors at the engineering structure, including fiber grating strain sensors, resistance strain gauges and piezoelectric film strain sensors; The composite sensor array is used to collect strain data of engineering structures under different working conditions, including static load, dynamic load, temperature change, humidity change and other environmental and load factors in real time and synchronously; The collected strain data is transmitted to the data processing and analysis center through an adaptive data transmission network; In the data processing and analysis center, multimodal data fusion algorithms are used to comprehensively process strain data and supplementary data obtained from other auxiliary monitoring equipment to assess the health status of engineering structures and predict potential damage.

2. The structural health monitoring method based on distributed strain dynamic testing according to claim 1 is characterized in that: The fiber Bragg grating strain sensors in the composite sensor array are arranged according to a predetermined topological structure to form a three-dimensional sensing network, which is distributed on the surface and inside of the structural component. The spacing d1 between adjacent fiber Bragg grating strain sensors on the surface is [5cm, 30cm], and the spacing d2 inside the component is [15cm, 45cm]. The arrangement of sensors at different depths follows an optimized layout based on structural finite element analysis to sense the strain field and cover the structural area.

3. The structural health monitoring method based on distributed strain dynamic testing according to claim 1 is characterized in that: The resistance strain gauge, piezoelectric film strain sensor and fiber Bragg grating strain sensor are staggered, and the pasting position of the resistance strain gauge is determined according to the results of the structural stress concentration theory analysis, and the piezoelectric film strain sensor is arranged at the position prone to vibration deformation according to the results of the structural vibration modal analysis. The three work together to form a complementary strain monitoring system. The pasting angle θ of the resistance strain gauge is adjusted according to the principal stress direction, and the range is [0°, 90°], so as to obtain structural strain information.

4. The structural health monitoring method based on distributed strain dynamic testing according to claim 1 is characterized in that: The adaptive data transmission network adopts a hybrid transmission mode, including wired transmission and wireless transmission, wherein the wired transmission part uses industrial Ethernet, and the wireless transmission part adopts a combination of 5G and LoRa; the fiber grating strain sensor adopts industrial Ethernet transmission, and the piezoelectric film strain sensor and the resistance strain gauge use 5G or LoRa transmission. Adaptive modulation and coding technology is used in the data transmission process to dynamically adjust the transmission rate and modulation mode M according to the signal strength and channel quality, wherein the range of R is [1Mbps, 100Mbps], and the modulation mode M can be adaptively selected from QPSK, 16QAM and 64QAM according to the channel status.

5. The structural health monitoring method based on distributed strain dynamic testing according to claim 1 is characterized in that: The multimodal data fusion algorithm adopts a deep learning-based method to construct a deep neural network model DNN, and uses the strain data from different sensors and supplementary data obtained by other auxiliary monitoring devices as auxiliary data S aux As input, strain data includes fiber Bragg grating strain sensor data S FBG , resistance strain gauge data S RS , Piezoelectric film strain sensor data S PF ,Other auxiliary monitoring equipment includes temperature sensor, humidity sensor, acceleration sensor, and the output is structural health status H and damage prediction D. The loss function L of the network model adopts the combination of mean square error and cross entropy, as follows: L=αMSE(H,H pred )+(1-α)CE(D,D pred ) Among them, MSE is the mean square error, CE is the cross entropy, α is the weight coefficient, H pred and D pred They are the predicted values ​​of structural health status and damage prediction, respectively. They are trained through a large amount of historical data and integrate data from different modes.

6. The structural health monitoring method based on distributed strain dynamic testing according to claim 5 is characterized in that: In the training process of the deep neural network model DNN, data enhancement technology is used to rotate, translate, scale, and add noise to the original data to increase the diversity and robustness of the training samples. FBG , the added noise N follows the normal distribution N(μ,σ 2 ), where μ ranges from [-0.1, 0.1] and σ ranges from [0.01, 0.1] to simulate different actual working conditions and noise environments.

7. The structural health monitoring method based on distributed strain dynamic testing according to claim 1 is characterized in that: The data processing and analysis center performs real-time anomaly detection on the collected strain data, using a statistical analysis method based on a sliding window. Assume that the length of the sliding window is W, and the mean of the strain data in the window is The standard deviation is σ S , for a newly collected data point x, if It is judged as abnormal data, where k is the set threshold coefficient, and the value range is [2,5]. At the same time, a rapid traceability analysis of the source of the abnormal data is initiated. According to the sensor position and time information corresponding to the abnormal data, the damage location is analyzed in combination with the force model of the structure.

8. The structural health monitoring method based on distributed strain dynamic testing according to claim 1 is characterized in that: When the data processing and analysis center determines that there is potential damage to the structure, the local refined monitoring mode is activated, and the MEMS sensor array is deployed in the suspected damage area to perform high-resolution measurement of the local strain gradient G, which is calculated using the following formula: Among them, ΔS is the strain difference measured by adjacent MEMS sensors, Δl is the distance between adjacent MEMS sensors, and the distance range is [1mm, 5mm]. The strain gradient measurement can assist in determining the location and extent of the damage.

9. The structural health monitoring method based on distributed strain dynamic testing according to claim 1 is characterized in that: The historical structural health status and damage prediction data are analyzed in time series. The long short-term memory network model is used to transform the time series data T = {H1, D1, H2, D2, …, H n ,D n } as input, predict the structural health status H for the next m time steps future and D future Damage trends, long-term trend prediction of structural health and maintenance decision support.

10. The structural health monitoring method based on distributed strain dynamic testing according to claim 1, characterized in that: The data processing and analysis center is provided with a human-computer interaction interface, which displays the real-time health status of the structure, damage prediction results, trend analysis of historical data and status information of the sensor, and provides a remote control function. The sampling frequency f, data transmission parameters and monitoring mode of the sensor are adjusted according to user needs, and the structural maintenance and reinforcement information input by the user is received through the human-computer interaction interface, the structural model information is updated, and closed-loop management of monitoring and maintenance is performed.

Images