Space-time adaptive fusion health monitoring method and system for concrete filled steel tube arch bridge

Abstract

The invention discloses a space-time adaptive fusion health monitoring method and system for a concrete-filled steel tube arch bridge. The method comprises the following steps: periodically collecting data through an ultrasonic sensor, an infrared thermal imaging sensor and a sonic sensor, denoising and compressing edge nodes, and calculating a signal-to-noise ratio and an environmental adaptation factor; after standardization and exception elimination, time evolution modeling and spatial evolution modeling are respectively carried out by using a long and short term memory network and a graph convolution network, and then self-adaptive weighted fusion is carried out on space-time results by taking a signal-to-noise ratio and an environmental adaptation factor as weights; sequentially executing weighted average initial estimation value, particle swarm optimization and genetic algorithm three-level optimization, and inputting a final optimization result into the CNN-LSTM model to obtain a damage type, a damage grade and a damage development trend; and finally, calculating a health score and a repair priority, generating a quantitative evaluation report and outputting a personalized repair scheme. The system is intelligent and self-adaptive, and can realize full-life-cycle high-precision health monitoring and scientific maintenance of the bridge.

Description

Technical Field

[0001] This invention relates to health monitoring technology for steel-concrete composite arch bridges, specifically to a spatiotemporal adaptive fusion health monitoring method and system for steel-concrete composite arch bridges. Background Technology

[0002] With the rapid development of modern bridge engineering, steel-concrete composite arch bridges have become one of the important bridge structural forms due to their high structural strength, superior seismic resistance, and durability. While steel-concrete composite arch bridges possess excellent structural performance, they are susceptible to damage over long-term use due to various factors, including environmental changes, load variations, and material aging. This damage can lead to varying degrees of injury (such as cracks, corrosion, and voids), affecting the structural safety. Therefore, the application of Structural Health Monitoring (SHM) technology is crucial to ensuring the long-term stability and safety of steel-concrete composite arch bridges.

[0003] Currently, health monitoring technology for steel-concrete composite arch bridges mainly relies on the acquisition of data from multiple sensors, including ultrasonic sensors, infrared thermal imaging technology, acoustic sensors, and lidar. These sensors can monitor different structural characteristics of the bridge, such as cracks, corrosion, and crack propagation. However, existing bridge health monitoring technologies and systems suffer from the following technical bottlenecks and shortcomings:

[0004] (1) The accuracy and stability of traditional data fusion methods are insufficient.

[0005] In existing technologies, data from multiple sensors are often fused using weighted averaging or simple superposition. Such fusion methods are poorly adaptable to spatiotemporal variations and cannot adequately account for sensor performance fluctuations under different environmental conditions. For example, factors such as temperature, humidity, and construction environment can affect sensor accuracy and real-time performance, but existing technologies often fail to provide effective compensation and correction during data fusion.

[0006] Existing data fusion methods are prone to data errors and distortion of health monitoring results when faced with complex monitoring environments (such as bridges under different workloads or temperature conditions). For example, infrared thermal imaging is greatly affected by temperature changes, ultrasonic sensors may be interfered with by environmental noise, and sound wave sensors may cause signal aliasing under multipath propagation. These problems have not been effectively solved by existing traditional data fusion methods.

[0007] (2) Existing systems lack dynamic adaptability and intelligence.

[0008] Current bridge health monitoring systems typically employ preset monitoring modes and parameters, failing to dynamically adjust in real time according to changes in environmental conditions and bridge status during structural health monitoring. This makes existing technologies unable to adapt to the operating characteristics of different sensors in complex environments, leading to a significant decrease in the accuracy of monitoring data. For example, some sensors may degrade in performance under high temperatures or extreme weather conditions, and existing static data fusion methods cannot adaptively adjust sensor weights or optimize data acquisition strategies.

[0009] Furthermore, existing technologies lack effective intelligent diagnostic and prediction mechanisms. Even if sensors collect a large amount of data, it is difficult to achieve in-depth data analysis, making it impossible to perform real-time diagnosis and long-term prediction of structural damage. Most traditional methods rely on manually set thresholds or rules based on simple algorithms, which cannot automatically optimize and adjust in changing environments, thus affecting the system's intelligence level and emergency response capabilities.

[0010] (3) Lack of multi-scale damage assessment and comprehensive analysis

[0011] Existing technologies for damage assessment typically rely on single-type sensor data (e.g., using only ultrasonic or infrared technology), enabling quantitative analysis of specific damage types such as cracks and corrosion. However, structural damage in steel-concrete composite arch bridges exhibits multi-scale and multi-dimensional characteristics. Damage at different scales (such as micro-cracks and macro-corrosion) has varying impacts on the bridge structure, but current technologies have failed to effectively integrate multi-source sensor data for comprehensive damage analysis from micro to macro levels.

[0012] Currently, most systems focus on detecting surface damage, neglecting the assessment of internal structural damage or potential risks, leading to incomplete detection results. For example, ultrasonic sensors can effectively detect microcracks, but are ineffective in diagnosing surface corrosion or coating damage; while infrared thermal imaging technology can identify surface cracks and thermal changes, it struggles to detect deep-seated corrosion or other types of structural damage. Therefore, existing technologies have significant limitations in multi-scale damage assessment, failing to provide a comprehensive and detailed analysis of the structural health status.

[0013] (4) Lack of long-term health prediction and intelligent repair decision support

[0014] Most existing bridge health monitoring systems can only provide an assessment of the current health status and lack the ability to predict long-term health trends. Most systems provide short-term damage diagnoses based on real-time monitoring data, but they are unable to predict the evolution of damage or long-term structural changes, and cannot provide early warnings of potential structural failure risks.

[0015] Furthermore, existing systems have shortcomings in repair decision-making. Typically, maintenance personnel manually determine whether repair is necessary based on real-time monitoring data. However, this manual judgment often lacks accurate evidence, leading to delayed or excessive repair decisions, resulting in wasted resources and increased maintenance costs. Existing systems fail to provide automated repair decision support based on health monitoring data and damage prediction, and lack intelligent long-term maintenance solutions.

[0016] (5) The system is complex and difficult to integrate.

[0017] Current bridge health monitoring systems typically rely on multiple independent sensors and monitoring devices. The lack of effective integration and optimization between these monitoring modules leads to high system complexity. Existing technologies face challenges in sensor network coordination and data transmission, resulting in data loss, transmission delays, or equipment failures during long-term operation, thus affecting system reliability and stability.

[0018] The main problems with existing technologies are:

[0019] (1) The data fusion accuracy is insufficient and cannot effectively cope with the impact of environmental changes on sensor data.

[0020] (2) It lacks dynamic adaptability and intelligent mechanisms, and cannot be optimized and adjusted according to the actual monitoring environment.

[0021] (3) It is impossible to achieve multi-scale structural damage assessment and it is difficult to comprehensively diagnose structural health problems at different levels.

[0022] (4) Insufficient support for long-term health prediction and repair decision-making; existing systems cannot provide intelligent long-term structural prediction and maintenance recommendations.

[0023] (5) The system integration is complex, and the coordination of sensors and data processing face significant challenges. Summary of the Invention

[0024] To address the shortcomings of the existing technologies, this invention provides a spatiotemporal adaptive fusion health monitoring method and system for steel-concrete composite arch bridges. The aim is to improve the accuracy and efficiency of health monitoring of steel-concrete composite arch bridges by introducing innovative technologies such as spatiotemporal evolution analysis, adaptive weighted fusion, multi-level optimization algorithms, deep learning damage assessment, and intelligent repair decision support, thereby achieving the system's intelligence, adaptability, and sustainability.

[0025] To solve the above-mentioned technical problems, the specific solution of the present invention is as follows:

[0026] A spatiotemporal adaptive health monitoring data fusion method for steel-concrete composite arch bridges includes the following steps:

[0027] Step S1: Collect raw data of the bridge according to the set working cycle. The raw data is obtained by ultrasonic sensors, infrared thermal imaging sensors and acoustic sensors respectively deployed on the bridge. Noise removal and compression are completed at the edge nodes of each sensor, and the signal-to-noise ratio and environmental adaptability factor of each sensor are calculated in real time.

[0028] Step S2: Standardize and remove anomalies from the data uploaded in Step S1 to obtain standardized data;

[0029] Step S3: Input the standardized data described in step S2 into a long short-term memory network for time evolution modeling to obtain time prediction results. At the same time, input the standardized data of each measuring point at the same moment into a graph convolutional network to perform spatial evolution modeling of the bridge spatial nodes to obtain spatial fusion data.

[0030] Step S4: Using the signal-to-noise ratio and environmental adaptation factor of each sensor described in Step S1 as weights, perform adaptive fusion on the time prediction results and spatial fusion data described in Step S3 in a weighted average manner to generate fused data;

[0031] Step S5: Perform three levels of optimization on the data fused in step S4: weighted average initial estimate, particle swarm optimization for local optimization, and genetic algorithm for global optimization. Then input the final optimization result into the CNN-LSTM model to obtain the damage type, level and development trend.

[0032] Step S6: Calculate the health score and repair priority based on the damage type, grade and development trend described in Step S5, generate a quantitative assessment report and output a personalized repair plan.

[0033] Furthermore, the ultrasonic sensor described in step S1 is used to monitor the ultrasonic echo signals of microcracks, voids, and corrosion inside the bridge structure; the infrared thermal imaging sensor is used to detect the infrared thermal imaging images of cracks, coating damage, and hot spots on the surface of the bridge structure; and the acoustic wave sensor is used to detect the acoustic wave signals of macroscopic corrosion and deformation of the bridge structure. The ultrasonic echo signals, infrared thermal imaging images, and acoustic wave signals constitute the raw data.

[0034] The ultrasonic echo signal is denoised using a Kalman filter, and the formula is as follows:

[0035] (1),

[0036] In the formula: This represents the state estimate after the update at time k; This represents the state estimate at the previous time step (k-1); Indicates Kalman gain; Indicates the actual measured value; Represents the measurement matrix;

[0037] The infrared thermal imaging image is denoised using an adaptive threshold segmentation algorithm. The image denoising formula is as follows:

[0038] (2),

[0039] In the formula: The threshold representing the number of pixels in an image; Represents the pixel value of the image; This represents the average value of the image pixels. The standard deviation of image pixels; Indicates the adaptive adjustment coefficient;

[0040] The acoustic signal is denoised using wavelet transform, and the formula is as follows:

[0041] (3),

[0042] In the formula: This represents the signal s after wavelet transform; Represents wavelet coefficients; Describe the wavelet basis functions; Indicates a signal; Indicates scale index; Indicates the total number of scales;

[0043] The formulas for calculating the signal-to-noise ratio and environmental adaptability factor of each sensor are as follows:

[0044] (7),

[0045] In the formula: Let be the weight of the i-th sensor at time t; Let be the signal-to-noise ratio of the i-th sensor at time t; Let be the signal-to-noise ratio of the j-th sensor at time t; Let i be the environmental adaptation factor of the i-th sensor; Let j be the environmental adaptation factor of the j-th sensor; This represents the total number of sensors.

[0046] Furthermore, the standardization described in step 2 adopts z-score standardization, the formula of which is as follows;

[0047] (4),

[0048] In the formula: This represents the standardized data; Represents the original data; This represents the mean of the data; The standard deviation of the data.

[0049] Furthermore, the time evolution modeling described in step 3 uses a Long Short-Term Memory (LSTM) network model to process the standardized data. The hidden state update formula of the LTM network is as follows:

[0050] (5),

[0051] In the formula: Represents the hidden state at time t; This represents the hidden state at the previous time step t-1; This represents the sensor's input data at time t;

[0052] The spatial evolution modeling employs a graph convolutional network to model the spatial nodes of the bridge, and further uses a weighted spatial averaging algorithm to fuse the data from various sensors at the same time point, as shown in the following formula:

[0053] (6),

[0054] In the formula: Spatial fusion data representing time t; The spatial weighting factor for the i-th sensor at time t; Let be the data from the i-th sensor at time t.

[0055] Furthermore, the formula for adaptive fusion in step S4 is as follows:

[0056] (8),

[0057] In the formula: The merged data; This represents the data from the i-th sensor at time t. Let be the weight of the i-th sensor at time t; This indicates the total number of sensors.

[0058] Furthermore, the formula for the weighted average initial estimate described in step S5 is as follows:

[0059] (9),

[0060] In the formula: These are preliminary assessment results; Let be the weight of the i-th sensor at time t; This represents the data from the i-th sensor at time t. Indicates the total number of sensors;

[0061] The objective function for local optimization in particle swarm optimization is as follows:

[0062] (10)

[0063] In the formula: It is the objective function value, which represents the fitness value or error value in particle swarm optimization; This represents the data from the i-th sensor at time t. The merged data; Indicates the total number of sensors;

[0064] The formula for global optimization in the genetic algorithm is as follows:

[0065] (11),

[0066] In the formula: This represents the objective function to be optimized. This represents the data from the i-th sensor at time t. The merged data; Indicates the total number of sensors;

[0067] The loss function of the CNN-LSTM model using a convolutional neural network is:

[0068] (12)

[0069] In the formula: Represents the loss function; The predicted damage level; This represents the actual level of damage. This represents the total number of samples in the training set;

[0070] The development trend is obtained through the following method: using a CNN-LSTM model to process the current time data D1(t), D2(t), …D from each sensor. n (t) Perform long-term trend prediction and output the predicted value of the future structural health status at time t according to the damage prediction formula. To form the aforementioned development trend, the damage prediction formula is as follows:

[0071] (13)

[0072] In the formula: This is a prediction of the future structural health status at time t; This is the data from the first sensor at time t; This is the data from the second sensor at time t; This represents the data from the i-th sensor at time t. This represents the data from the nth sensor at time t.

[0073] Furthermore, the formula for the health score mentioned in step S6 is as follows:

[0074] (14)

[0075] In the formula: The structural health status score calculated at time t; Let be the weight of the i-th damage type at time t; This represents the evaluation result for the i-th damage type at time t; The total number of damage types;

[0076] The formula for the repair priority is as follows:

[0077] (15)

[0078] In the formula: The repair priority is calculated at time t; This is the repair priority coefficient for the i-th damage type; This represents the damage assessment result for the i-th damage type at time t. This represents the total number of damage types.

[0079] Furthermore, step S6 further includes: predicting the health status at future times based on the damage prediction model, obtaining the prediction results, and writing them into the quantitative assessment report to form long-term maintenance early warning information. The prediction formula is as follows:

[0080] (16)

[0081] In the formula: For the future Health assessment prediction results; This represents the data from the i-th sensor at time t. This represents the total number of sensors.

[0082] The spatiotemporal adaptive fusion health monitoring system for steel-concrete composite arch bridges implementing the method includes:

[0083] The sensor network consists of ultrasonic sensors, infrared thermal imaging sensors, and acoustic sensors deployed at key stress-bearing parts, surfaces, and arch rib-beam-node areas of the bridge. It is used to collect raw data from each sensor according to a set working cycle and upload it to the data preprocessing module via a wireless communication module. It is also used to perform noise reduction and compression at edge nodes and calculate the signal-to-noise ratio and environmental adaptation factor in real time, and dynamically adjust the sampling frequency according to the calculation results before uploading.

[0084] The data preprocessing module is wirelessly connected to the sensor network. It sequentially uses Kalman filtering, adaptive threshold segmentation, and wavelet transform to denoise the raw data, and then uses z-score standardization and 3σ outlier removal to output standardized data.

[0085] The spatiotemporal evolution analysis module is connected to the data preprocessing module. It uses a long short-term memory network for temporal evolution modeling and a graph convolutional network and a weighted spatial averaging algorithm for spatial evolution modeling, and outputs temporal prediction results and spatial fusion data respectively.

[0086] The adaptive data fusion module, connected to the spatiotemporal evolution analysis module, is used to adaptively weight and fuse the temporal prediction results and spatial fusion data by using the signal-to-noise ratio and environmental adaptation factor of each sensor calculated by the sensor network as weights, and output the fused data.

[0087] The damage assessment module, connected to the adaptive data fusion module, is used to perform three levels of optimization on the fused data: weighted average initial estimate, particle swarm optimization local optimization, and genetic algorithm global optimization. The final optimization result is then input into the CNN-LSTM model to output the damage type, level, and development trend.

[0088] The repair decision support module, connected to the damage assessment module, is used to calculate health scores and repair priorities based on damage type, level, and development trend, generate quantitative assessment reports, and output personalized repair plans.

[0089] A computer program product comprising a storage medium and computer-readable instructions stored on the medium, which, when executed by a computer, implement the method described.

[0090] Advantages of the present invention

[0091] 1. The spatiotemporal adaptive fusion health monitoring method and system for steel-concrete composite arch bridges of this invention introduces a spatiotemporal evolution analysis model to capture the temporal and spatial variation patterns of raw data from ultrasonic sensors, infrared thermal imaging sensors, and acoustic sensors. During data fusion, the working environment, measurement characteristics, and data fluctuations of different sensors are fully considered. This spatiotemporal evolution-based analysis method makes the fused data more accurate and reliable, especially in complex environments, effectively eliminating interference and improving data accuracy. This addresses the problem that traditional multi-source data fusion methods lack consideration for the spatiotemporal variation characteristics of sensor data.

[0092] 2. This invention employs an adaptive weighted fusion algorithm, dynamically adjusting the weight of each sensor based on its signal-to-noise ratio (SNR) and environmental adaptation factors (such as temperature and humidity). This allows for dynamic optimization of sensor weights during real-time data acquisition based on data quality, ensuring high accuracy and robustness of the final fusion result. This innovation significantly improves the system's adaptability, enabling it to automatically optimize the fusion strategy under different environmental conditions, thus solving the problem of traditional methods being unable to adapt to complex environmental changes.

[0093] 3. This invention introduces a multi-level optimization algorithm. Through progressive refinement processing of primary, intermediate, and advanced optimization, it not only improves data accuracy but also comprehensively optimizes data at multiple levels, ensuring that the evaluation results approximate the true state as closely as possible. In particular, the use of Particle Swarm Optimization (PSO) and Genetic Algorithm (GA) gives the evaluation results strong global optimality, providing more accurate support for bridge health monitoring. This solves the problem that traditional damage assessment methods cannot accurately reflect the health status and future development trends of bridges.

[0094] 4. This invention combines techniques such as Convolutional Neural Networks (CNN) and Long Short-Term Memory Networks (LSTM) to deeply analyze the spatiotemporal characteristics of sensor data, enabling accurate assessment of different types of structural damage (such as cracks, corrosion, and voids). Compared with traditional methods, the damage assessment model of this invention can automatically identify and predict potential structural damage and accurately quantify the type, location, and severity of damage, significantly improving the accuracy and reliability of damage assessment. It also solves the problem that traditional methods are ill-suited to the diversity and complexity of structural damage.

[0095] 5. This invention introduces an intelligent repair decision support system. Based on multi-dimensional health assessment results and damage prediction, it calculates the repair priority for each damage point and recommends the most suitable repair methods, materials, and timing based on the assessment results. Simultaneously, the generated repair plan considers resource optimization, ensuring efficient and economical repair work. This intelligent decision support system provides maintenance personnel with scientific repair plans, significantly improving repair efficiency and reducing maintenance costs. It addresses the issue of traditional bridge health monitoring systems lacking intelligent decision support, leading to delays or over-repair in repair work.

[0096] 6. This invention provides a systematic method based on full life-cycle health monitoring. Through long-term accumulation and analysis of multi-source data, combined with a damage prediction model, the system can track and predict the health status of structures over the long term, providing a scientific basis for the long-term maintenance of bridges. By predicting future damage development, the system can provide early warnings of potential risks and offer decision support for long-term maintenance planning, thereby improving the health management level of bridges and extending the service life of the structures.

[0097] 7. The innovation of this invention also lies in its high degree of intelligence and adaptability. The system can automatically adjust its data acquisition strategy, preprocessing method, data fusion process, damage assessment model, and repair decision support according to different environments, sensor performance fluctuations, and changes in the bridge's health status. Through real-time adaptive optimization, the system can automatically respond to challenges under different working conditions, ensuring stable and accurate health monitoring services in any environment.

[0098] In summary, compared with existing technologies, this invention, by introducing innovative technologies such as spatiotemporal evolution analysis, adaptive weighted fusion, multi-level optimization algorithms, deep learning damage assessment, and intelligent repair decision support, not only significantly improves the accuracy and efficiency of health monitoring for steel-concrete composite arch bridges, but also achieves system intelligence, adaptability, and sustainability. Through these innovations, this invention provides strong technical support for the long-term health management of bridges, yielding significant social and economic benefits. Attached Figure Description

[0099] Figure 1 This is a flowchart of the spatiotemporal adaptive health monitoring data fusion method for steel-concrete composite arch bridges according to the present invention. Detailed Implementation

[0100] The present invention will be further explained and described below with reference to the accompanying drawings and specific embodiments. It should be noted that the specific embodiments are not intended to limit the scope of the present invention.

[0101] like Figure 1 As shown in the figure, this specific embodiment provides a spatiotemporal adaptive health monitoring data fusion method for steel-concrete composite arch bridges, including the following steps:

[0102] Step S1: Collect raw bridge data according to a set working cycle. The raw data is obtained by ultrasonic sensors, infrared thermal imaging sensors, and acoustic sensors respectively deployed on the bridge. Noise reduction and compression are performed at the edge nodes of each sensor, and the signal-to-noise ratio and environmental adaptability factor of each sensor are calculated in real time. The specific steps are as follows:

[0103] 1.1 Sensor Selection and Placement

[0104] This embodiment introduces an innovative adaptive data quality optimization method during sensor data acquisition and preprocessing to ensure the fusion accuracy of multi-source data. Especially when sensor performance fluctuates in complex environments, it can intelligently adjust and optimize data quality, thereby improving the accuracy of the final health assessment.

[0105] The selection and arrangement of sensors in this embodiment are crucial to ensuring comprehensive monitoring of the health status of the steel-concrete composite arch bridge.

[0106] The ultrasonic sensor is used to monitor the ultrasonic echo signals of microcracks, voids, and corrosion inside the bridge structure. The ultrasonic sensor obtains detailed internal information of the steel-concrete composite arch bridge structure by transmitting and receiving sound wave signals, and is a common method for detecting cracks and voids.

[0107] Infrared thermal imaging sensors are used to detect cracks, coating damage, and hot spots on the surface of bridge structures. Infrared thermal imaging technology can reveal surface damage to structures by detecting temperature distribution.

[0108] Acoustic sensors are used to detect acoustic signals of macroscopic corrosion and deformation in bridge structures; by analyzing the propagation characteristics of sound waves, acoustic sensors can detect the overall health status of the structure.

[0109] The aforementioned ultrasonic echo signal, infrared thermal imaging image, and acoustic signal constitute the raw data described in step S1.

[0110] 1.2 Data Acquisition

[0111] Ultrasonic sensors, infrared thermal imaging sensors, and acoustic sensors periodically collect data according to a set working cycle (e.g., every minute or hour) and upload the data in real time to a central processing system for storage and processing via a wireless communication module (such as Wi-Fi or 5G). The central processing system includes at least a data preprocessing module. The ultrasonic sensors, infrared thermal imaging sensors, and acoustic sensors are rationally arranged according to the key stress-bearing parts, surfaces, and arch rib-beam-node areas of the steel-concrete composite bridge structure to achieve comprehensive monitoring of all parts of the structure.

[0112] After the ultrasonic sensors, infrared thermal imaging sensors, and acoustic sensors are deployed, data is collected periodically according to a set working cycle and transmitted to the data processing system for further analysis and processing. During the data acquisition process, the ultrasonic sensors, infrared thermal imaging sensors, and acoustic sensors will each generate raw data (…). The data includes a large amount of information, where i represents the i-th sensor and t is the acquisition time.

[0113] 1.3 Data Preprocessing

[0114] Raw data often contains noise, outliers, and inconsistencies, and must be transformed into effective information that can be used for further analysis and evaluation through a series of preprocessing steps.

[0115] 1.3.1 Noise Removal and Signal Smoothing

[0116] (1) Ultrasonic signal denoising: Since ultrasonic signals are usually affected by environmental noise and equipment errors, this embodiment uses a Kalman filter to denoise the ultrasonic signal to ensure signal accuracy. The Kalman filter dynamically adjusts the noise effect by estimating the signal state, as shown in the following formula:

[0117] (1),

[0118] In the formula: This represents the state estimate after the update at time k; This represents the state estimate at the previous time step (k-1); Indicates Kalman gain; Indicates the actual measured value; This represents the measurement matrix.

[0119] Kalman filtering can effectively reduce noise interference with signal quality and improve the accuracy of ultrasonic signals.

[0120] (2) Infrared Image Denoising and Enhancement: Infrared imaging sensors may be affected by environmental changes (such as temperature fluctuations) and background noise, leading to a decrease in image quality. In this embodiment, an adaptive threshold segmentation algorithm is used to remove noise from the infrared thermal imaging image according to formula (2), and image enhancement technology is used to improve the image quality to ensure that the hot spot area is clearer and to improve the visibility of cracks and defects in the infrared image. Formula (2) is as follows:

[0121] (2),

[0122] In the formula: The threshold representing the number of pixels in an image; Represents the pixel value of the image; This represents the average value of the image pixels. The standard deviation of image pixels; This represents the adaptive adjustment coefficient.

[0123] (3) Acoustic signal denoising: Acoustic signals are often affected by multipath propagation and background noise, leading to inaccurate signals. Therefore, in this embodiment, wavelet transform is used to denoise the acoustic signal. Wavelet transform can extract useful frequency components and retain important features of the signal while removing unnecessary noise, thus improving the quality of the acoustic signal. The wavelet transform formula is as follows:

[0124] (3),

[0125] In the formula: This represents the signal s after wavelet transform; Represents wavelet coefficients; Describe the wavelet basis functions; Indicates a signal; Indicates scale index; Indicates the total number of scales.

[0126] 1.3.2 Calculate the signal-to-noise ratio and environmental adaptation factor of each sensor as weights.

[0127] The core of adaptive weighted fusion lies in adjusting the weights based on the data quality of each sensor. The quality of each sensor is influenced by its signal-to-noise ratio (SNR) and environmental adaptation factor (FOR). (Effected by factors such as temperature and humidity). The calculation formulas for the signal-to-noise ratio and environmental adaptability factor of each sensor are as follows:

[0128] (7),

[0129] In the formula: Let be the weight of the i-th sensor at time t; Let be the signal-to-noise ratio of the i-th sensor at time t; Let be the signal-to-noise ratio of the j-th sensor at time t; Let i be the environmental adaptation factor of the i-th sensor; Let j be the environmental adaptation factor of the j-th sensor; This represents the total number of sensors.

[0130] Formula (7) can adaptively adjust the weights of each sensor according to the signal-to-noise ratio and environmental adaptability of the sensor, thereby improving the accuracy and robustness of data fusion.

[0131] Step S2: Standardize and remove outliers from the data uploaded in Step S1 to obtain standardized data; the specific steps are as follows:

[0132] 2.1 Data Standardization

[0133] Because data generated by different types of sensors have different dimensions and ranges, they must be standardized for subsequent fusion and analysis. To ensure data compatibility on the same platform, the data uploaded in step S1 is standardized using the z-score method:

[0134] (4),

[0135] In the formula: This represents the standardized data; Represents the original data; This represents the mean of the data; The standard deviation of the data.

[0136] This standardized method can convert data collected by different sensors to the same dimension, ensuring the consistency and comparability of data during subsequent fusion and analysis.

[0137] 2.2 Outlier Detection

[0138] During data acquisition, sensor malfunctions, environmental interference, or other abnormal factors may cause data anomalies. To ensure data reliability, this embodiment sets a reasonable threshold range to automatically identify and remove data that does not conform to the normal range in the raw data described in step S1, thereby further improving data accuracy.

[0139] 2.3 Data Quality Assessment and Optimization

[0140] An adaptive data quality optimization mechanism is introduced to automatically optimize data based on the quality and environmental adaptability of each sensor during data acquisition and preprocessing. In the data preprocessing stage, sensor quality assessment is primarily based on the signal-to-noise ratio (SNR) and environmental adaptability factor. The data acquisition strategy is dynamically adjusted to ensure the high quality of the final fused data. Environmental adaptation factors refer to the impact of environmental changes such as temperature and humidity on sensor performance.

[0141] 2.4 Sensor data acquisition and preprocessing followed by data optimization and transmission

[0142] After data acquisition and preprocessing, all sensor data is ready for the next stage of data fusion. At this point, all sensor data has been aligned to the same standard, noise has been removed, and the data has been standardized. This provides a solid foundation for spatiotemporal evolution analysis and adaptive weighted fusion, ensuring that the fusion of multi-source data can be carried out efficiently and accurately.

[0143] The interaction between the sensors and the data processing unit: Each sensor uploads real-time data to a cloud server for processing via a wireless communication module (such as Wi-Fi, 5G, LoRa, etc.). Each sensor has independent data storage and processing capabilities, and performs preliminary local preprocessing (such as noise reduction and data format conversion) before uploading data to reduce the amount of data transmitted and improve transmission efficiency.

[0144] Real-time data processing: After data is uploaded, the cloud system analyzes the transmitted data in real time. Through designed analysis algorithms, it can perform instant data processing, generate health assessments, and provide real-time feedback. During data transmission, all data is encrypted to ensure security and prevent leakage or tampering. A data quality assessment mechanism also dynamically adjusts the sensor sampling rate and data transmission frequency based on environmental conditions (such as temperature and humidity). For example, the sampling frequency can be increased in noisy environments to ensure more accurate data collection.

[0145] Step S3: Input the standardized data from Step S2 into a Long Short-Term Memory (LSTM) network for time evolution modeling to obtain time prediction results. Simultaneously, input the standardized data from each measurement point at the same time into a graph convolutional network to perform spatial evolution modeling of the bridge's spatial nodes, obtaining spatial fusion data. The specific steps are as follows:

[0146] 3.1 Spatiotemporal Evolution Analysis Model

[0147] Spatiotemporal evolution analysis is used to model the spatiotemporal characteristics of sensor data. Data from different sensors exhibit variability in time and space; spatiotemporal evolution analysis captures these variations and utilizes them in the data fusion process. This approach elevates data fusion beyond simple weighted averaging; it models the spatiotemporal changes of data, considers the characteristics of different sensors under varying environments, and automatically adjusts their weights, thereby improving the accuracy of data fusion.

[0148] 3.1.1 Temporal Evolution Modeling

[0149] Temporal evolution analysis primarily involves time-series modeling of sensor data to predict trends in data changes over time. To model time-series data more accurately, the standardized data obtained in step S2 is input into a Long Short-Term Memory (LSTM) network for time evolution modeling to obtain time prediction results.

[0150] LSTM is a type of recurrent neural network (RNN) capable of capturing long-term dependencies, making it suitable for processing sensor data with time-series characteristics. The hidden state update formula for an LSTM network is as follows:

[0151] (5),

[0152] In the formula: Represents the hidden state at time t; This represents the hidden state at the previous time step t-1; This represents the sensor's input data at time t;

[0153] for A unit refers to a single unit in a long short-term memory network. It is a special type of recurrent neural network ( It has a structure with input gates, forget gates, and output gates, and is used to control the transmission of information. These gate mechanisms enable the learning and memorization of long-term dependencies in time-series data. The goal is to determine the hidden state based on time t-1. and input at time t To calculate the hidden state at the current time t In complex environments (such as varying temperatures, humidity, and structural loads), LSTM units can dynamically adjust network weights based on historical and real-time data, and adaptively optimize for fluctuations in sensor performance to ensure prediction accuracy.

[0154] LSTM units, through training, can capture long-term trends in data, providing dynamic predictive support for data fusion. This method allows for prediction of time-series data from different sensors, providing more temporal information for subsequent fusion.

[0155] 3.1.2 Spatial Evolution Modeling

[0156] Besides temporal variations, the spatial distribution characteristics of sensor data are also crucial, especially since sensor measurements can differ across different parts of the bridge structure and under varying environmental conditions. Considering factors such as sensor location, local bridge structural characteristics, and construction environment, while performing temporal evolution modeling, standardized data from various measurement points at the same time are input into a graph convolutional network to model the spatial evolution of the bridge's spatial nodes, resulting in spatially fused data.

[0157] The spatial evolution modeling employs graph convolutional networks to model the spatial nodes of the bridge, and further uses a weighted spatial averaging algorithm to fuse the data from various sensors at the same time. The formula for the weighted spatial averaging algorithm is as follows:

[0158] (6),

[0159] In the formula: Spatial fusion data representing time t; The spatial weighting factor for the i-th sensor at time t; Let be the data from the i-th sensor at time t.

[0160] By using the above spatiotemporal evolution modeling, we can more accurately reflect the changes in sensor data under different time and space conditions, thus laying the foundation for the next step of data fusion.

[0161] Step S4: Using the signal-to-noise ratio and environmental adaptation factor of each sensor described in Step S1 as weights, perform adaptive fusion on the time prediction results and spatial fusion data described in Step S3 in a weighted average manner to generate fused data;

[0162] The formula for the adaptive fusion is as follows:

[0163] (8),

[0164] In the formula: The merged data; This represents the data from the i-th sensor at time t. Let be the weight of the i-th sensor at time t; This indicates the total number of sensors.

[0165] Merged data It can more accurately reflect the health status of bridge structures, especially under complex environmental conditions, and can improve the adaptability and accuracy of the system by dynamically adjusting weights. The results obtained after data fusion... This data will serve as input for subsequent steps, allowing the multi-level optimization and damage assessment modules to process it further. At this stage, the data has undergone spatiotemporal evolution analysis and adaptive weighted fusion, resulting in high accuracy and reliability, providing a solid foundation for damage assessment and structural health prediction.

[0166] Traditional data fusion methods typically fuse data from different sensors using fixed weighting coefficients. However, traditional static weighted fusion methods often fail to adapt to real-world conditions. Therefore, this embodiment captures the temporal and spatial changes in data through spatiotemporal evolution analysis and dynamically adjusts the fusion strategy based on sensor quality and environmental changes using an adaptive weighted fusion method, thereby significantly improving the accuracy of data fusion. The combination of temporal and spatial evolution modeling effectively addresses issues such as data distortion, signal interference, and poor environmental adaptability in traditional techniques, providing accurate data support for subsequent health assessments and damage predictions.

[0167] Step S5: The data fused in Step S4 undergoes a three-stage optimization process: weighted average initial estimate, particle swarm optimization for local optimization, and genetic algorithm for global optimization. The final optimization result is then input into the CNN-LSTM model to obtain the damage type, level, and development trend. This step is crucial for ensuring data accuracy and assessing damage type and severity. By introducing multi-stage optimization algorithms and deep learning damage assessment models, accurate assessment of structural damage is achieved, and long-term predictions are provided.

[0168] 5.1 Level 3 Optimization

[0169] The purpose of introducing optimization algorithms is to further improve the accuracy and reliability of data processing results. By using multi-level optimization algorithms to progressively adjust and optimize data at different levels, the final output evaluation results are ensured to be as close as possible to the actual situation.

[0170] 5.1.1 Basic Optimization

[0171] The initial optimization is a weighted average preliminary estimate: the fused data obtained in step S4 is used to calculate the weight of each sensor and perform a weighted average of the data to obtain a preliminary structural health assessment result, which is used as the initial input for subsequent particle swarm optimization.

[0172] The formula for the initial weighted average estimate is as follows:

[0173] (9),

[0174] In the formula: These are preliminary assessment results; Let be the weight of the i-th sensor at time t; This represents the data from the i-th sensor at time t. This indicates the total number of sensors.

[0175] The initial optimization results provide a basic health assessment of the structure, but these results are usually rather coarse and therefore require further refinement.

[0176] 5.1.2 Intermediate Optimization

[0177] Intermediate optimization employs the Particle Swarm Optimization (PSO) algorithm to locally optimize the initial assessment results, yielding intermediate optimization results for further refinement and optimization of the health assessment. The PSO algorithm is based on the cooperation and competition mechanisms among particles in nature. By simulating particle motion, it adjusts the velocity and position of each particle to minimize the objective function, thereby obtaining the global optimum. The objective function for the local optimization in the PSO algorithm is as follows:

[0178] (10)

[0179] In the formula: It is the objective function value, which represents the fitness value or error value in particle swarm optimization; This represents the data from the i-th sensor at time t. The merged data; This indicates the total number of sensors.

[0180] Intermediate optimization further improved the accuracy and stability of the data, providing more accurate input for advanced optimization.

[0181] 5.1.3 Advanced Optimization

[0182] Advanced optimization employs a genetic algorithm to globally optimize the intermediate optimization results, yielding the final optimized outcome. Building upon the primary and intermediate optimizations, advanced optimization algorithms such as Genetic Algorithm (GA) or Ant Colony Optimization (ACO) further refine the final health assessment result. Genetic algorithms mimic the process of natural selection, continuously optimizing parameters through selection, crossover, and mutation to reach the globally optimal solution. The optimization objective is to minimize assessment error and improve data accuracy.

[0183] The global optimization of the genetic algorithm is achieved through the following formula:

[0184] (11),

[0185] In the formula: This represents the objective function to be optimized. This represents the data from the i-th sensor at time t. The merged data; This indicates the total number of sensors.

[0186] Advanced optimizations ensure that the final output of health assessment results achieves optimal accuracy, maximizing data accuracy.

[0187] 5.2 Damage Assessment

[0188] This embodiment utilizes a deep learning model (CNN-LSTM model) to predict the damage type, severity, and development trend of a structure. The final optimization result obtained in step 5.1.3 is input into the CNN-LSTM model to output the damage type, level, and development trend. This process combines sensor data, physical models, and deep learning algorithms, making the assessment of structural health not merely a static process, but a dynamic evolutionary one.

[0189] 5.2.1 Damage Assessment

[0190] Convolutional Neural Networks (CNNs) are used to extract spatial features from sensor data, while Long Short-Term Memory (LSTM) networks are used to analyze time-series data. CNNs are responsible for extracting local features from image data, while LSTMs are responsible for capturing the trends in data over time. The combined use of both ensures accurate assessment of the dynamic health of structures.

[0191] The loss function of the CNN-LSTM model using a convolutional neural network is:

[0192] (12),

[0193] In the formula: Represents the loss function; The predicted damage level; This represents the actual level of damage. This represents the total number of samples in the training set.

[0194] Deep neural networks extract spatiotemporal features from sensor data by passing information layer by layer and output an assessment of the structure's health status. In particular, they extract local features through convolutional layers (CNN) and analyze the temporal characteristics of the data through LSTM networks to comprehensively assess the structure's health status.

[0195] 5.2.2 Damage Prediction and Assessment

[0196] Deep learning models, by combining historical data with the current state, can predict future damage trends when forecasting structural damage. For example, the system can predict crack propagation, increased corrosion, and even the risk of bridge failure. This provides decision support for the long-term health management of structures.

[0197] The development trend is obtained through the following method: using a CNN-LSTM model to process the current time data D1(t), D2(t), …D from each sensor. n (t) Perform long-term trend prediction and output the predicted value of the future structural health status at time t according to the damage prediction formula. This will help to shape the aforementioned development trend and provide maintenance personnel with more accurate long-term maintenance solutions through damage prediction.

[0198] The damage prediction formula is as follows:

[0199] (13)

[0200] In the formula: This is a prediction of the future structural health status at time t; This is the data from the first sensor at time t; This is the data from the second sensor at time t; This represents the data from the i-th sensor at time t. This represents the data from the nth sensor at time t.

[0201] Step S6: Calculate the health score and repair priority based on the damage type, grade and development trend described in Step S5, generate a quantitative assessment report and output a personalized repair plan.

[0202] 6.1 Calculate the health score

[0203] Health assessment is a crucial step in bridge structural health monitoring, accurately identifying the type and extent of damage and its impact on the overall health of the bridge. Based on optimized data and the output of deep learning models, this step enables precise health assessments and provides intelligent decision support for bridge repair based on the assessment results and damage predictions.

[0204] The multi-dimensional health assessment model is invoked, and the current level and development trend of the three types of damage—cracks, corrosion, and cavitation—are substituted into the health scoring formula as follows:

[0205] (14)

[0206] In the formula: The structural health status score calculated at time t; The weight of the i-th damage type (such as cracks, corrosion, etc.) at time t; The evaluation result of the i-th damage type at time t (output by the damage evaluation model); This represents the total number of damage types.

[0207] In health assessments, weighting The weights are determined based on the severity of the impact of different damage types on structural health, and the accuracy and reliability of the assessment results are ensured by flexibly adjusting these weights.

[0208] 6.2 Repair Priority

[0209] First, the repair priority of each injury site is assessed by combining the comprehensive health assessment results and the injury prediction model. The assessment of repair priority considers the following factors:

[0210] Damage severity: The severity of damage is determined by the damage level in the health assessment model, such as crack depth, degree of corrosion, and extent of voids.

[0211] Structural safety: Assess the impact of damage on the overall structural safety of the bridge, especially whether there are potential risks that could lead to structural failure.

[0212] Repair urgency: Based on the output of the damage prediction model, assess the future development trend of the damage and determine the urgency of repair.

[0213] Repair priority is calculated using the following formula:

[0214] (15)

[0215] In the formula: The repair priority is calculated at time t; This is the repair priority coefficient for the i-th damage type; This represents the damage assessment result for the i-th damage type at time t. This represents the total number of damage types.

[0216] High-priority damage types will be addressed first to ensure the safety and lifespan of the bridge.

[0217] 6.3 Quantitative Assessment Report

[0218] The quantitative assessment report includes the following:

[0219] (1) Crack grade: The number, size and development trend of cracks on the bridge surface are evaluated based on infrared thermal imaging data analysis. According to the crack width, depth and development trend, a grading standard (e.g. 0-5mm is a light crack, 5-10mm is a moderate crack and >10mm is a severe crack) is used for evaluation.

[0220] (2) Corrosion degree: Based on acoustic sensor data, the corrosion status of the bridge structure is analyzed, including the distribution, depth and severity of corrosion. The corrosion degree is quantified by the attenuation rate of the acoustic signal. When the attenuation rate is greater than a certain threshold (e.g., 15%), the structure is considered to have severe corrosion.

[0221] (3) Voiding situation: Voiding or voids inside the bridge are detected by ultrasonic sensors to assess the extent and impact of voiding. The extent of voiding is quantified based on the time difference of the echo signal from the ultrasonic sensor. When the echo signal deviation is greater than a certain threshold (e.g., 5ms), voiding is considered to exist.

[0222] Each assessment result in the report has quantitative indicators, providing bridge maintenance personnel with detailed and intuitive health status data.

[0223] 6.4 Generation and Optimization of Personalized Repair Solutions

[0224] Based on the repair priority assessment results, a personalized repair plan is provided for the bridge using a repair decision support algorithm. The generation of the repair plan considers the following factors:

[0225] Repair methods: Based on the damage type and repair priority, the system recommends the most suitable repair method. For example, for cracks, filling repair can be selected; for corrosion, anti-corrosion coating repair can be selected.

[0226] Repair materials: The system recommends the most suitable repair materials based on different damage types and repair methods.

[0227] Repair timing: Based on the damage prediction results, the system suggests the optimal repair time to avoid delays or over-repair.

[0228] Resource optimization: Based on the bridge's repair needs and available resources, the system optimizes the repair plan to ensure a balance between repair efficiency and cost.

[0229] The repair plan output provides maintenance personnel with detailed repair steps, including information such as repair methods, required materials, personnel and equipment requirements.

[0230] 6.5 Long-term maintenance and repair prediction

[0231] The method in this embodiment can also predict the health status at future times based on the damage prediction model, obtain the prediction results, and write them into the quantitative assessment report to form long-term maintenance early warning information. The prediction formula is as follows:

[0232] Post-repair health assessment is predicted using the following formula:

[0233] (16)

[0234] In the formula: For the future Health assessment prediction results; This represents the data from the i-th sensor at time t. This represents the total number of sensors.

[0235] Through this long-term forecasting, this embodiment not only supports current repairs but also provides early warnings for future maintenance, ensuring that the bridge remains in good health throughout its entire life cycle.

[0236] A spatiotemporal adaptive fusion health monitoring system for steel-concrete composite arch bridges, implementing the above method, is applicable to the full life-cycle health monitoring of steel-concrete composite arch bridges. This system enables accurate damage assessment, intelligent prediction, and repair decision support. The system includes:

[0237] The sensor network consists of ultrasonic sensors, infrared thermal imaging sensors, and acoustic sensors deployed at key stress-bearing parts, surfaces, and arch rib-beam-node areas of the bridge. It is used to collect raw data from each sensor according to a set working cycle and upload it to the data preprocessing module via a wireless communication module. It is also used to perform noise reduction and compression at edge nodes and calculate the signal-to-noise ratio and environmental adaptation factor in real time, and dynamically adjust the sampling frequency according to the calculation results before uploading.

[0238] The ultrasonic sensor used is the S1803 longitudinal wave DPC dry contact ultrasonic transducer manufactured by ACS-Solutions GmbH, Germany. The transducer has a nominal frequency of 100kHz and is suitable for dry contact ultrasonic testing in highly scattering materials such as concrete. It can be driven by concrete ultrasonic testing instruments such as the A1220 MONOLITH.

[0239] The infrared thermal imaging sensor selected is the T640 infrared thermal imager manufactured by FLIR Systems, Inc. (the T420 model from the same series can also be used). This model of infrared camera has been reported to be used in the infrared scanning inspection of concrete bridge decks. It can acquire infrared thermal imaging images of 640×480 pixels with a wavelength range of approximately 7.5–14μm for the identification of cracks, coating damage, and hot spots on bridge surfaces. Alternatively, the R550 series infrared thermal imager from NEC-Avio Corporation of Japan or the TI600 infrared thermal imager from Zhejiang Hongxiang Technology Co., Ltd., etc., can be used for engineering inspection.

[0240] Acoustic wave sensors employ acoustic emission, such as the R6a 60kHz general-purpose acoustic emission sensor manufactured by Physical Acoustics, Inc. in the United States. This sensor is a narrowband resonant, high-sensitivity structural acoustic wave sensor, suitable for acoustic emission monitoring of crack initiation and propagation in concrete structures; or the ACS-YD series acoustic emission sensors used in long-term bridge monitoring projects in China.

[0241] The sensor models mentioned above are all existing commercial products and can be purchased from the manufacturers. In practical applications, other similar sensors with similar performance can also be selected according to engineering requirements, and the specific models and manufacturers mentioned above are not limited.

[0242] The data preprocessing module is wirelessly connected to the sensor network. It sequentially uses Kalman filtering, adaptive threshold segmentation, and wavelet transform to denoise the raw data, and then uses z-score standardization and 3σ outlier removal to output standardized data.

[0243] The spatiotemporal evolution analysis module, connected to the data preprocessing module, employs a Long Short-Term Memory (LSTM) network for temporal evolution modeling, outputting temporal prediction results. It also uses a graph convolutional network and a weighted spatial averaging algorithm for spatial evolution modeling, outputting spatially fused data to more accurately reflect the changes in sensor data under different temporal and spatial conditions.

[0244] The adaptive data fusion module, connected to the spatiotemporal evolution analysis module, is used to adaptively weight and fuse the temporal prediction results and spatial fusion data by using the signal-to-noise ratio and environmental adaptation factor of each sensor calculated by the sensor network as weights. It dynamically adjusts the weights of each sensor to improve the accuracy and robustness of the fusion results and outputs the fused data.

[0245] The damage assessment module, connected to the adaptive data fusion module, is used to perform three levels of optimization on the fused data: weighted average initial estimate, particle swarm optimization local optimization, and genetic algorithm global optimization. The final optimization result is then input into the CNN-LSTM model to output the damage type, level, and development trend.

[0246] The repair decision support module, connected to the damage assessment module, is used to calculate health scores and repair priorities based on damage type, level, and development trend, generate quantitative assessment reports, and output personalized repair plans.

[0247] A computer program product comprising a storage medium and computer-readable instructions stored on the medium, which, when executed by a computer, implement the spatiotemporal adaptive fusion health monitoring method for steel-concrete composite arch bridges.

Claims

1. A spatiotemporal adaptive health monitoring data fusion method for steel-concrete composite arch bridges, characterized in that, Includes the following steps: Step S1: Collect raw data of the bridge according to the set working cycle. The raw data is obtained by ultrasonic sensors, infrared thermal imaging sensors and acoustic sensors respectively deployed on the bridge. Noise removal and compression are completed at the edge nodes of each sensor, and the signal-to-noise ratio and environmental adaptability factor of each sensor are calculated in real time. Step S2: Standardize and remove outliers from the data processed in Step S1 to obtain standardized data; Step S3: Input the standardized data described in step S2 into a long short-term memory network for time evolution modeling to obtain time prediction results. At the same time, input the standardized data of each measuring point at the same moment into a graph convolutional network to perform spatial evolution modeling of the bridge spatial nodes to obtain spatial fusion data. Step S4: Using the signal-to-noise ratio and environmental adaptation factor of each sensor described in Step S1 as weights, perform adaptive fusion on the time prediction results and spatial fusion data described in Step S3 in a weighted average manner to generate fused data; Step S5: Perform three levels of optimization on the data fused in step S4: weighted average initial estimate, particle swarm optimization for local optimization, and genetic algorithm for global optimization. Then input the final optimization result into the CNN-LSTM model to obtain the damage type, level and development trend. Step S6: Calculate the health score and repair priority based on the damage type, grade and development trend described in Step S5, generate a quantitative assessment report and output a personalized repair plan.

2. The method according to claim 1, characterized in that, The ultrasonic sensor described in step S1 is used to monitor ultrasonic echo signals of microcracks, voids, and corrosion inside the bridge structure; the infrared thermal imaging sensor is used to detect infrared thermal imaging images of cracks, coating damage, and hot spots on the surface of the bridge structure; the acoustic sensor is used to detect acoustic signals of macroscopic corrosion and deformation of the bridge structure; the ultrasonic echo signals, infrared thermal imaging images, and acoustic signals constitute the raw data. The ultrasonic echo signal is denoised using a Kalman filter, and the formula is as follows: （1）， In the formula: This represents the state estimate after the update at time k; This represents the state estimate at the previous time step (k-1); Indicates Kalman gain; Indicates the actual measured value; Represents the measurement matrix; The infrared thermal imaging image is denoised using an adaptive threshold segmentation algorithm. The image denoising formula is as follows: （2）， In the formula: The threshold representing the number of pixels in an image; Represents the pixel value of the image; This represents the average value of the image pixels. The standard deviation of image pixels; Indicates the adaptive adjustment coefficient; The acoustic signal is denoised using wavelet transform, and the formula is as follows: （3）， In the formula: This represents the signal s after wavelet transform; Represents wavelet coefficients; Describe the wavelet basis functions; Indicates a signal; Indicates scale index; Indicates the total number of scales; The formulas for calculating the signal-to-noise ratio and environmental adaptability factor of each sensor are as follows: （7）， In the formula: Let be the weight of the i-th sensor at time t; Let be the signal-to-noise ratio of the i-th sensor at time t; Let be the signal-to-noise ratio of the j-th sensor at time t; Let i be the environmental adaptation factor of the i-th sensor; Let j be the environmental adaptation factor of the j-th sensor; This represents the total number of sensors.

3. The method according to claim 1, characterized in that, The standardization described in step 2 uses z-score standardization, and its formula is as follows; （4）， In the formula: This represents the standardized data; Represents the original data; This represents the mean of the data; The standard deviation of the data.

4. The method according to claim 1, characterized in that, The time evolution modeling described in step 3 uses a Long Short-Term Memory (LSTM) network model to process the standardized data. The hidden state update formula of the LTM network is as follows: （5）， In the formula: Represents the hidden state at time t; This represents the hidden state at the previous time step t-1; This represents the sensor's input data at time t; The spatial evolution modeling employs a graph convolutional network to model the spatial nodes of the bridge, and further uses a weighted spatial averaging algorithm to fuse the data from various sensors at the same time point, as shown in the following formula: （6）， In the formula: Spatial fusion data representing time t; The spatial weighting factor for the i-th sensor at time t; Let be the data from the i-th sensor at time t.

5. The method according to claim 1, characterized in that, The formula for adaptive fusion in step S4 is as follows: （8）， In the formula: The merged data; This represents the data from the i-th sensor at time t. Let be the weight of the i-th sensor at time t; This indicates the total number of sensors.

6. The method according to claim 1, characterized in that, The formula for the weighted average initial estimate mentioned in step S5 is as follows: （9）， In the formula: These are preliminary assessment results; Let be the weight of the i-th sensor at time t; This represents the data from the i-th sensor at time t. Indicates the total number of sensors; The objective function for local optimization in particle swarm optimization is as follows: （10）， In the formula: It is the objective function value, which represents the fitness value or error value in particle swarm optimization; This represents the data from the i-th sensor at time t. The merged data; Indicates the total number of sensors; The formula for global optimization in the genetic algorithm is as follows: （11）， In the formula: This represents the objective function to be optimized. This represents the data from the i-th sensor at time t. The merged data; Indicates the total number of sensors; The loss function of the CNN-LSTM model using a convolutional neural network is: （12）， In the formula: Represents the loss function; The predicted damage level; This represents the actual level of damage. This represents the total number of samples in the training set; The development trend is obtained through the following method: using a CNN-LSTM model to process the current time data D1(t), D2(t), …D from each sensor. n (t) Perform long-term trend prediction and output the predicted value of the future structural health status at time t according to the damage prediction formula. To form the aforementioned development trend, the damage prediction formula is as follows: （13）， In the formula: This is a prediction of the future structural health status at time t; This is the data from the first sensor at time t; This is the data from the second sensor at time t; This represents the data from the i-th sensor at time t. This represents the data from the nth sensor at time t.

7. The method according to claim 1, characterized in that, The formula for the health score mentioned in step S6 is as follows: （14）， In the formula: The structural health status score calculated at time t; Let be the weight of the i-th damage type at time t; This represents the evaluation result for the i-th damage type at time t; The total number of damage types; The formula for the repair priority is as follows: （15）， In the formula: The repair priority is calculated at time t; This is the repair priority coefficient for the i-th damage type; This represents the damage assessment result for the i-th damage type at time t. This represents the total number of damage types.

8. The method according to claim 1, characterized in that, Step S6 further includes: predicting the health status at future times based on the damage prediction model, obtaining the prediction results, and writing them into the quantitative assessment report to form long-term maintenance early warning information. The prediction formula is as follows: （16）， In the formula: For the future Health assessment prediction results; This represents the data from the i-th sensor at time t. This represents the total number of sensors.

9. A spatiotemporal adaptive fusion health monitoring system for steel-concrete composite arch bridges implementing the method of any one of claims 1 to 8, characterized in that, include: The sensor network consists of ultrasonic sensors, infrared thermal imaging sensors, and acoustic sensors deployed at key stress-bearing parts, surfaces, and arch rib-beam-node areas of the bridge. It is used to collect raw data from each sensor according to a set working cycle and upload it to the data preprocessing module via a wireless communication module. It is also used to perform noise reduction and compression at edge nodes and calculate the signal-to-noise ratio and environmental adaptation factor in real time, and dynamically adjust the sampling frequency according to the calculation results before uploading. The data preprocessing module is wirelessly connected to the sensor network. It sequentially uses Kalman filtering, adaptive threshold segmentation, and wavelet transform to denoise the raw data, and then uses z-score standardization and 3σ outlier removal to output standardized data. The spatiotemporal evolution analysis module is connected to the data preprocessing module. It uses a long short-term memory network for temporal evolution modeling and a graph convolutional network and a weighted spatial averaging algorithm for spatial evolution modeling, and outputs temporal prediction results and spatial fusion data respectively. The adaptive data fusion module, connected to the spatiotemporal evolution analysis module, is used to adaptively weight and fuse the temporal prediction results and spatial fusion data by using the signal-to-noise ratio and environmental adaptation factor of each sensor calculated by the sensor network as weights, and output the fused data. The damage assessment module, connected to the adaptive data fusion module, is used to perform three levels of optimization on the fused data: weighted average initial estimate, particle swarm optimization local optimization, and genetic algorithm global optimization. The final optimization result is then input into the CNN-LSTM model to output the damage type, level, and development trend. The repair decision support module, connected to the damage assessment module, is used to calculate health scores and repair priorities based on damage type, level, and development trend, generate quantitative assessment reports, and output personalized repair plans.

10. A computer program product, characterized in that, The product includes a storage medium and computer-readable instructions stored on the medium, which, when executed by a computer, implement the method of any one of claims 1 to 8.

Images