Bridge steel structure defect nondestructive testing system based on infrared thermal imaging

Abstract

The invention relates to the field of physical testing technologies and systems, and discloses a bridge steel structure defect nondestructive testing system based on infrared thermal imaging, which comprises an unmanned aerial vehicle integrated with an infrared thermal imaging detection device and a ground computing station. The unmanned aerial vehicle flies along a planned path, transient thermal excitation is applied through a thermal excitation device, and dynamic thermal response data of the surface of a pre-coated infrared coating is collected by synchronously utilizing an infrared thermal imager. And the ground computing station performs space-time fusion preprocessing on the returned data to obtain a standardized panoramic temperature field image, and inputs the standardized panoramic temperature field image into a convolutional neural network model which fuses physical mechanism constraints and has an online self-optimization function for intelligent analysis. According to the method, signals are enhanced through the coating, efficient inspection is achieved through the unmanned aerial vehicle, the recognition precision and generalization ability are guaranteed through the physical information fusion model, continuous evolution is achieved in combination with online self-optimization, and a high-sensitivity, full-automatic and quantitative bridge steel structure defect intelligent detection solution is formed.

Description

Technical Field

[0001] This invention relates to the field of physical testing technology and systems, and in particular to a non-destructive testing system for defects in bridge steel structures based on infrared thermal imaging. Background Technology

[0002] As a core component of modern transportation infrastructure, bridges' steel structures are subjected to multiple factors during long-term service, including vehicle loads, environmental corrosion, material aging, and extreme weather conditions. This makes them susceptible to damage such as fatigue cracks, localized corrosion, loose connections, and welding defects. Failure to detect and address these defects in a timely manner will significantly weaken the structure's load-bearing capacity and may even lead to catastrophic accidents, seriously threatening public safety. Currently, non-destructive testing of bridge steel structures mainly relies on manual inspections combined with conventional testing methods such as ultrasonic, magnetic particle, and radiographic testing. While these methods can identify defects under certain conditions, they generally suffer from the following shortcomings: 1) low testing efficiency, making it difficult to achieve rapid coverage of large areas of the structure; 2) requiring personnel to work at close range or at heights, posing a high risk to personal safety; 3) the accuracy of the test results largely depends on the operator's experience and subjective judgment, lacking unified and objective quantitative standards.

[0003] To overcome the aforementioned shortcomings, infrared thermal imaging technology, as a non-contact, full-field detection method, has been introduced into the field of bridge steel structure inspection. However, traditional passive infrared thermal imaging technology faces three main problems in practical applications: First, steel structures have high thermal conductivity, and surface temperature differences caused by defects in natural environments are often extremely weak, making them difficult to capture by conventional infrared equipment; second, large bridge structures are large in scale and complex in construction, limiting the coverage of fixed platforms or manually held equipment, and the risks of working at heights remain; furthermore, existing infrared thermal imaging results largely rely on manual interpretation, lacking automated and quantitative defect identification and assessment capabilities. In recent years, some studies have attempted to enhance the thermal response of defects using active thermal excitation methods, but in large bridge structures, problems such as the difficulty in uniformly applying thermal excitation and extracting thermal signals still exist. Although some studies have introduced image processing or machine learning algorithms for defect identification, due to the lack of in-depth integration of the physical mechanisms of heat conduction, the models have insufficient generalization ability and a high misjudgment rate, making it difficult to meet the accuracy and reliability requirements of practical engineering.

[0004] In summary, existing technologies still suffer from limitations such as insufficient sensitivity in detecting early damage to bridge steel structures, low efficiency and safety in inspecting large structures, and strong subjectivity in result interpretation, making quantification difficult. Therefore, there is an urgent need to develop a new non-destructive testing method for bridge steel structures that can achieve highly sensitive, efficient, automated, and quantitative diagnosis. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] In view of the above-mentioned shortcomings and deficiencies of the existing technology, this application provides a non-destructive testing system for bridge steel structure defects based on infrared thermal imaging, which overcomes the three major technical problems existing in the detection of defects in existing bridge steel structures: First, traditional manual inspection and contact testing are inefficient, have incomplete coverage, and pose safety risks of high-altitude operations; second, the high thermal conductivity of steel structures results in extremely weak surface temperature differences caused by defects, making it difficult for conventional infrared thermal imaging technology to detect early or minor damage; third, the detection results heavily rely on human experience for interpretation, lacking automated and objective means for defect identification and quantitative assessment of severity.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the main technical solutions adopted in this application include:

[0009] This application provides a non-destructive testing system for bridge steel structure defects based on infrared thermal imaging, specifically including:

[0010] Unmanned aerial vehicles and ground computing stations integrated with infrared thermal imaging detection devices;

[0011] The UAV flies along a designated detection path according to the infrared image acquisition command sent by the ground computing station, and collects infrared thermal image data of the bridge steel structure within the area of ​​the designated detection path in real time.

[0012] The ground computing station receives infrared thermal image data of all test areas of the bridge transmitted by the UAV, and preprocesses the infrared thermal image data to obtain normalized image data of the entire detection area that eliminates environmental radiation interference and is aligned with the time and space dimensions.

[0013] The ground computing station uses a convolutional neural network model that incorporates physical mechanism constraints and has online self-optimization to identify and evaluate the image data, and simultaneously outputs quantitative evaluation results of defect type, defect area and defect severity.

[0014] The entire inspection area of ​​the bridge is pre-coated with an infrared functional coating, which is a multi-layer composite structure made of high infrared emissivity material, with an infrared emissivity greater than or equal to 0.9.

[0015] Real-time acquisition of infrared thermal image data of bridge steel structures within the area of ​​the specified detection path includes: applying transient thermal flux excitation to the area under test using an infrared radiation source, and simultaneously acquiring dynamic response information of the infrared functional coating surface after thermal flux excitation.

[0016] Optionally, the UAV integrates a thermal imaging detection device;

[0017] The thermal imaging detection device includes: a thermal excitation device, an infrared thermal imager, a temperature sensor, a positioning module, a data transmission module, and a flight control module;

[0018] The ground computing station transmits infrared image acquisition commands and detection paths via a data transmission module.

[0019] The flight control module flies based on the detection path;

[0020] The thermal excitation device is used to apply transient thermal flux excitation to the area under test based on infrared image acquisition commands;

[0021] The infrared thermal imager is used to continuously acquire the dynamic thermal response of the infrared functional coating surface at a frame rate greater than or equal to a preset frame rate after the thermal excitation device applies heat flow excitation, and to obtain an infrared thermal image sequence composed of multiple frames, wherein each frame contains temperature distribution information of the coating surface at that moment.

[0022] The temperature sensor is used to measure and record the ambient background temperature before the start of each thermal excitation cycle;

[0023] The positioning module is used to record the spatial coordinates and absolute timestamp of each frame of infrared thermal image when it is acquired in real time.

[0024] The data transmission module is used to transmit the infrared thermal image sequence and its corresponding spatial location coordinates, absolute timestamps and ambient background temperature to the ground computing station.

[0025] The infrared thermal image data includes: the original sequence of infrared thermal images consisting of multiple frames acquired by the infrared thermal imager; the ambient background temperature value measured by the temperature sensor and corresponding to it; and the spatial location coordinates and absolute timestamp of each frame of image synchronously recorded by the positioning module.

[0026] Optionally, the ground computing station preprocesses the infrared thermal image data by including:

[0027] S301. Extract the original infrared thermal image sequence from the infrared thermal image data; based on the known infrared emissivity of the infrared functional coating and the corresponding ambient background temperature, calibrate the grayscale value of each frame in the infrared thermal image sequence and convert it into an absolute temperature value image that characterizes the real thermal radiation to obtain a multi-frame absolute temperature image sequence.

[0028] S302. Using the trigger signal of each thermal excitation as the unified time origin, perform inter-frame alignment on the multi-frame absolute temperature image sequence obtained in S301 to generate a transient temperature field sequence that is strictly synchronized in the time dimension.

[0029] S303. Based on the precise spatial position and attitude information recorded by the positioning module for each frame of the image, perform geometric correction and registration on each frame of the transient temperature field sequence obtained in S302, and fuse them to generate a seamless two-dimensional panoramic temperature field image covering the entire detection area.

[0030] S304. Perform contrast stretching on the two-dimensional panoramic temperature field image generated in S303, and linearly normalize its pixel values ​​to the [0,1] interval to form standardized input data for use by the convolutional neural network model.

[0031] Optionally, the convolutional neural network model with online self-optimization function is a pre-trained model using a multi-task learning architecture;

[0032] The training process of the pre-trained model uses the total loss function. Loss by Defect Type Defect region segmentation loss Defect severity assessment regression loss And a physical constraint loss Weighted summation constitutes the result;

[0033] The physical constraint loss The temperature field evolution data predicted by the convolutional neural network model for training samples is compared with the baseline temperature field data generated by simulation based on physical laws. This is used to measure the consistency between the model prediction and physical laws.

[0034] The reference temperature field data is obtained by establishing a finite element model that is completely consistent with the experimental specimen used in the training samples in terms of geometry, material physical properties and applied thermal excitation parameters, and generating it by numerical simulation through solving the heat conduction control equation;

[0035] The physical constraint loss The mean square error is the difference between the temperature field data predicted by the model and the reference temperature field data.

[0036] Optionally, the physical constraint loss Calculated using the following formula:

[0037] ;

[0038] Where N is the number of samples in the training batch, and t is the number of time frames for a single sample. The model predicts the temperature field for the i-th sample at the j-th time step. This is the reference temperature field of the i-th sample at the j-th time step obtained through the numerical simulation.

[0039] Optionally, the infrared functional coating comprises, from the inside out, an adhesive base layer, a functional layer, and a weather-resistant protective layer;

[0040] The functional layer contains a high infrared emissivity filler prepared by a nanocomposite process. The high infrared emissivity filler is composed of graphene, carbon nanotubes and alumina, and its mass fraction in the functional layer is 30% to 60%.

[0041] The substrate material of the bonding base layer is modified epoxy resin, wherein 10% by mass of metallic aluminum powder is dispersed.

[0042] Optionally, before the ground computing station uses the convolutional neural network model to identify and evaluate the image data, it further includes a step of online adaptive optimization of the model, including:

[0043] Based on the defect identification and evaluation results output in real time by the convolutional neural network model, high-confidence diagnostic samples with a confidence level higher than a preset threshold are automatically selected.

[0044] When the number of high-confidence diagnostic samples accumulates to a preset scale, the incremental learning process is automatically triggered.

[0045] In the incremental learning process, the parameters of the encoder of the convolutional neural network model are fixed, while the parameters of the decoder and output layer are fine-tuned.

[0046] The high-confidence diagnostic samples and their corresponding complete test data are archived into a defect knowledge base that is categorized and indexed by bridge, component, and defect type.

[0047] Optionally, the preset threshold is 0.95; the preset size is 100 diagnostic samples; and the learning rate used when fine-tuning the decoder and output layer parameters is 0.00001.

[0048] Optionally, the thermal excitation device is a high-power, short-pulse infrared radiation source with an output power range of 3kW to 5kW and a single pulse duration range of 100 milliseconds to 500 milliseconds; the infrared thermal imager operates at a frame rate ≥50Hz.

[0049] Optionally, a training dataset for training the convolutional neural network model can be constructed by combining experimental simulation and numerical simulation.

[0050] The experimental simulation is achieved by collecting thermal response data of specimens coated with the infrared functional coating and prefabricated with defects of different types and sizes under thermal excitation; the numerical simulation is based on finite element analysis to simulate the transient temperature field of the infrared functional coating under thermal excitation consistent with experimental conditions and coupled with different defects.

[0051] (III) Beneficial Effects

[0052] This application presents a non-destructive testing system for bridge steel structure defects based on infrared thermal imaging. By applying a high infrared emissivity functional coating, the system significantly amplifies the weak thermal signals caused by minute defects inside the steel structure, solving the technical problem of passive thermal imaging's insensitivity to early damage. It employs an unmanned aerial vehicle (UAV) integrated with an active thermal excitation and imaging system, achieving non-contact, fully automated, and full-coverage inspection, greatly improving inspection efficiency and operational safety, and overcoming the risks and limitations of manual inspection. Furthermore, it introduces a multi-task convolutional neural network that integrates physical mechanisms to achieve end-to-end intelligent analysis of infrared thermal images, simultaneously completing automatic defect type identification, precise region segmentation, and quantitative severity assessment, outputting objective and quantifiable diagnostic results. This completely eliminates reliance on human experience, forming a highly sensitive, efficient, automated, and intelligent non-destructive testing solution. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating the workflow of a non-destructive testing system for bridge steel structure defects based on infrared thermal imaging, as described in this application.

[0054] Figure 2 This is a schematic diagram of the infrared functional coating of the bridge steel structure defect non-destructive testing system based on infrared thermal imaging, according to an embodiment of this application.

[0055] Figure 3 This is a module connection diagram of the thermal imaging detection device of the non-destructive testing system for bridge steel structure defects based on infrared thermal imaging, according to an embodiment of this application.

[0056] Figure 4 This is a flowchart of a non-destructive testing method for bridge steel structure defects based on infrared thermal imaging, as proposed in this application.

[0057] [Explanation of Labels in the Attached Image]

[0058] 11-Adhesive base layer, 12-Functional layer, 13-Weather-resistant protective layer, 14-Bridge steel structure substrate, 21-Thermal excitation device, 22-Infrared thermal imager, 23-Temperature sensor, 24-Positioning module, 25-Data transmission module, 26-Flight control module. Detailed Implementation

[0059] To better explain and facilitate understanding of this application, the following detailed description of the application is provided in conjunction with the accompanying drawings and specific embodiments.

[0060] In existing technologies, methods for detecting defects in bridge steel structures can be mainly categorized into the following two types:

[0061] The first category is traditional contact / close-range manual inspection methods. These methods rely on visual inspection supplemented by conventional non-destructive testing techniques such as ultrasonic testing, magnetic particle testing, radiographic testing, and penetrant testing. They depend on personnel working at close range or at heights, and the inspection process is conducted in a point-to-point or line-to-line manner, heavily reliant on personnel experience. This method requires inspectors to carry ultrasonic, magnetic particle, or radiographic equipment to inspect and assess the steel structure point-by-point. Essentially, it is a highly experience-dependent "sampling inspection," with extremely low efficiency, making it difficult to achieve rapid, full coverage of long-span bridge structures. Furthermore, working at heights poses significant personal safety risks. Simultaneously, the inspection results are highly subjective, lacking objective and unified quantitative standards, and cannot provide accurate support for the digital health management and life prediction of structures.

[0062] The second category is detection methods based on passive or simple active infrared thermal imaging. This method uses infrared equipment to non-contactly acquire surface temperature images of structures to identify anomalies. However, due to the high thermal conductivity of bridge steel structures, surface temperature differences caused by internal defects are often extremely weak, resulting in insufficient sensitivity for detecting hidden damage such as microcracks and early corrosion. Although some studies have attempted to introduce active thermal excitation, challenges remain in large structures, including uneven excitation and difficulties in signal extraction. Furthermore, the analysis process often relies on manual interpretation or simple image processing algorithms, lacking integration with the physical mechanisms of heat conduction, leading to low automation, high misidentification rates, and an inability to reliably quantitatively assess the severity of defects.

[0063] To address this, this application provides a non-destructive testing system for bridge steel structure defects based on infrared thermal imaging. By integrating a specially designed infrared functional coating, an unmanned aerial vehicle (UAV) active excitation platform, spatiotemporal fusion preprocessing, and an intelligent model based on physical mechanism constraints and online self-optimization, this system systematically solves the core problems of low detection sensitivity, contradiction between efficiency and safety, strong subjective dependence on results, and insufficient model generalization ability in the prior art. Ultimately, it achieves a fundamental leap from inefficient, subjective, and qualitative traditional detection to efficient, automatic, and quantitative intelligent diagnosis.

[0064] To better explain and facilitate understanding of this application, a detailed description of its embodiments is provided below in conjunction with the accompanying drawings. While exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a clearer and more thorough understanding of this application and to fully convey the scope of this application to those skilled in the art.

[0065] Example 1

[0066] Figure 1The following is a flowchart illustrating the workflow of a non-destructive testing system for bridge steel structure defects based on infrared thermal imaging, according to this application. Figure 1 As shown, the non-destructive testing system for bridge steel structure defects based on infrared thermal imaging includes:

[0067] Unmanned aerial vehicles and ground computing stations integrated with infrared thermal imaging detection devices;

[0068] The UAV flies along a designated detection path according to the infrared image acquisition command sent by the ground computing station, and collects infrared thermal image data of the bridge steel structure within the area of ​​the designated detection path in real time.

[0069] The ground computing station receives infrared thermal image data of all test areas of the bridge transmitted by the UAV, and preprocesses the infrared thermal image data to obtain normalized image data of the entire detection area that eliminates environmental radiation interference and is aligned with the time and space dimensions.

[0070] The ground computing station uses a convolutional neural network model that incorporates physical mechanism constraints and has online self-optimization to identify and evaluate the image data, and simultaneously outputs quantitative evaluation results of defect type, defect area and defect severity.

[0071] The entire inspection area of ​​the bridge is pre-coated with an infrared functional coating, which is a multi-layer composite structure made of high infrared emissivity material, with an infrared emissivity greater than or equal to 0.9.

[0072] Real-time acquisition of infrared thermal image data of bridge steel structures within the area of ​​the specified detection path includes: applying transient thermal flux excitation to the area under test using an infrared radiation source, and simultaneously acquiring dynamic response information of the infrared functional coating surface after thermal flux excitation.

[0073] This embodiment achieves highly sensitive, efficient, fully automated, and quantitative non-destructive testing of defects in bridge steel structures by integrating a specially designed infrared functional coating, UAV active excitation detection, spatiotemporal fusion preprocessing, and physical mechanism fusion into a self-optimizing intelligent model. Its beneficial effects are mainly reflected in the following four aspects:

[0074] By coating with a special functional coating with high infrared emissivity (≥0.9), the weak thermal conduction anomalies caused by defects inside the steel structure are effectively amplified into significant temperature difference anomalies on the coating surface. This fundamentally overcomes the technical bottleneck of weak thermal signals in high thermal conductivity steel structures, which are difficult to detect with traditional infrared imaging. This allows hidden damage such as micro-cracks and early corrosion to be captured stably and clearly.

[0075] The use of an integrated active thermal excitation UAV platform for automated scanning overcomes the problems of low efficiency, incomplete coverage, and high risk of high-altitude operations that exist in manual inspection and fixed platform testing. It achieves non-contact, rapid, and full-area in-service inspection, greatly improving operational safety and inspection efficiency.

[0076] By introducing a convolutional neural network model that incorporates physical mechanism constraints, the system can perform end-to-end intelligent analysis on the pre-processed panoramic temperature field data, and synchronously and automatically output quantitative assessment results of the defect type, precise spatial location and range (segmentation), and severity, completely eliminating the reliance on human experience and providing objective, unified, and traceable quantitative diagnostic reports.

[0077] The physical constraint loss term introduced in the model enables the network to learn in accordance with the fundamental laws of heat conduction, rather than merely statistical laws at the data level. This significantly improves its generalization ability to unknown defects and different environmental conditions. Combined with online self-optimization capabilities, the system can autonomously iterate and optimize using high-confidence results over long-term use, and accumulate a structured defect knowledge base. This gives the entire system continuous self-improvement and adaptability, achieving a leap from "one-time training" to "lifelong learning."

[0078] In summary, this solution constitutes a complete closed-loop detection system with physical laws at its core and data intelligence as its driving force, providing a high-precision, automated diagnostic tool for the safe operation and maintenance of bridge steel structures.

[0079] Example 2

[0080] Figure 1 This is a flowchart illustrating the workflow of a non-destructive testing system for bridge steel structure defects based on infrared thermal imaging, according to this application. Figure 2 This is a schematic diagram of the infrared functional coating shown in the embodiments of this application. Figure 3 This is a module connection diagram of the thermal imaging detection device according to an embodiment of this application.

[0081] like Figure 1 and Figure 3 As shown, the non-destructive testing system for bridge steel structure defects based on infrared thermal imaging includes:

[0082] Unmanned aerial vehicles and ground computing stations integrated with infrared thermal imaging detection devices;

[0083] The UAV flies along the designated detection path according to the infrared image acquisition instructions sent by the ground computing station, and collects infrared thermal image data of the bridge steel structure in the area where the designated detection path is located in real time.

[0084] The ground computing station receives infrared thermal image data of all test areas of the bridge transmitted by the UAV, and preprocesses the infrared thermal image data to obtain normalized image data of the entire detection area that eliminates environmental radiation interference and is aligned with the time and space dimensions.

[0085] The ground computing station uses a convolutional neural network model that incorporates physical mechanism constraints and has online self-optimization to identify and evaluate the image data, and simultaneously outputs quantitative evaluation results of defect type, defect area and defect severity.

[0086] The entire inspection area of ​​the bridge is pre-coated with an infrared functional coating, which is a multi-layer composite structure made of high infrared emissivity material, with an infrared emissivity greater than or equal to 0.9.

[0087] Real-time acquisition of infrared thermal image data of bridge steel structures within the area of ​​the specified detection path includes: applying transient thermal flux excitation to the area under test using an infrared radiation source, and simultaneously acquiring dynamic response information of the infrared functional coating surface after thermal flux excitation.

[0088] Furthermore, the drone integrates a thermal imaging detection device;

[0089] Specifically, such as Figure 3 As shown, the thermal imaging detection device includes: a thermal excitation device 21, an infrared thermal imager 22, a temperature sensor 23, a positioning module 24, a data transmission module 25, and a flight control module 26;

[0090] The thermal excitation device 21 is equipped with an adaptive adjustment mechanism, which can dynamically adjust the focusing range and intensity of the thermal excitation according to the relative distance between the UAV and the surface of the bridge steel structure and the incident angle of the light source, thereby overcoming environmental interference and effectively stimulating thermal anomalies in the defect area. The adaptive adjustment mechanism integrates a laser rangefinder and a two-dimensional servo motor. The laser rangefinder monitors the relative distance between the UAV and the surface of the bridge steel structure and the incident angle of the light source in real time, and drives the two-dimensional servo motor to dynamically adjust the pitch and yaw angles of the thermal excitation device 21, and adjusts the focal length of the focusing lens in front of the light source in conjunction with the adjustment, so that the output thermal excitation spot can uniformly cover the target area with a preset size and intensity, overcoming the problem of uneven excitation caused by fluctuations in the flight attitude of the UAV or unevenness of the structural surface.

[0091] An infrared thermal imager 22 is used to collect thermal response signals from the surface of the bridge steel structure. A cooled mid-wave infrared focal plane detector is preferred, with a working wavelength of 8 to 15 micrometers. The infrared thermal imager 22 is synchronously triggered with the thermal excitation device 21. Immediately after the thermal excitation pulse ends, it continuously collects the dynamic thermal response process of the infrared functional coating surface at a frame rate of ≥50Hz, acquiring infrared thermal image data containing time and temperature evolution information. A temperature sensor 23 measures and records the ambient background temperature before each thermal excitation cycle begins. A positioning module 24 records the time and spatial location corresponding to the infrared thermal image data in real time. A data transmission module 25 transmits the infrared thermal image data and its corresponding time, spatial location, and ambient background temperature to a ground computing station. A flight control module 26 drives the UAV to automatically fly along the bridge steel structure surface at a preset constant distance according to a pre-planned detection path, ensuring the relative distance and incident angle of the light source between the thermal excitation device 21 and the infrared thermal imager 22, thereby guaranteeing the continuity and stability of data acquisition.

[0092] The ground computing station transmits infrared image acquisition commands and detection paths via a data transmission module;

[0093] The flight control module flies based on the detection path; it drives the UAV to fly automatically at a preset constant distance along the surface of the bridge steel structure, ensuring the relative distance and incident angle of the light source between the thermal excitation device and the infrared thermal imager, thereby ensuring the continuity and stability of data acquisition.

[0094] The thermal excitation device is used to apply transient thermal flux excitation to the area under test based on infrared image acquisition commands;

[0095] The infrared thermal imager is used to continuously acquire the dynamic thermal response of the infrared functional coating surface at a frame rate greater than or equal to a preset frame rate after the thermal excitation device applies heat flow excitation, and to obtain an infrared thermal image sequence composed of multiple frames, wherein each frame contains temperature distribution information of the coating surface at that moment.

[0096] The temperature sensor is used to measure and record the ambient background temperature before the start of each thermal excitation cycle;

[0097] The positioning module is used to record the spatial coordinates and absolute timestamp of each frame of infrared thermal image when it is acquired in real time.

[0098] The data transmission module is used to transmit the infrared thermal image sequence and its corresponding spatial location coordinates, absolute timestamps and ambient background temperature to the ground computing station.

[0099] The infrared thermal image data includes: the original sequence of infrared thermal images consisting of multiple frames acquired by the infrared thermal imager; the ambient background temperature value measured by the temperature sensor and corresponding to it; and the spatial location coordinates and absolute timestamp of each frame of image synchronously recorded by the positioning module.

[0100] Specifically, the thermal excitation device is a high-power, short-pulse infrared radiation source with an output power range of 3kW to 5kW and a single pulse duration range of 100 milliseconds to 500 milliseconds; the infrared thermal imager operates at a frame rate ≥50Hz.

[0101] The power and pulse time range are determined through analysis of the coating's thermophysical properties and experimental calibration. This enables the formation of a significant and measurable transient temperature response on the surface of the infrared functional coating, while avoiding coating damage or thermal saturation due to overheating. This ensures sufficient temperature contrast between defective and healthy areas, facilitating infrared thermal imager capture and subsequent identification.

[0102] Before the testing begins, a special coating must be applied to the steel structure of the bridge to be tested (such as the outer surface of a steel box girder).

[0103] First, the area to be coated is sandblasted and then cleaned with anhydrous ethanol to ensure that the surface is rough, clean, and dry.

[0104] Then, automated equipment is used to apply a layered coating to the entire bridge under test:

[0105] In specific implementation, the infrared functional coating adopts a three-layer composite structure design with a thickness controlled within the range of 50 micrometers to 200 micrometers. As shown in the figure, from the bridge steel structure substrate 14 outwards, there are bonding base layer 11, functional layer 12 and weather-resistant protective layer 13.

[0106] The bonding substrate 11 is directly coated onto the surface of the bridge steel structure substrate 14, with a dry film thickness of 10 to 30 micrometers. It uses modified epoxy resin as the matrix material and disperses 10% by mass of aluminum powder. The bonding substrate 11 forms a strong bond with the bridge steel structure substrate 14 through chemical bonding and physical anchoring, which helps reduce corrosion of the substrate and improves the uniformity of surface heat conduction. The bonding substrate 11 is prepared by placing modified epoxy resin, aluminum powder, curing agent, and diluent in a mixer at a mass ratio of 100:15:25:10 and stirring at room temperature until uniformly mixed. The aluminum powder significantly improves the thermal conductivity of the bonding substrate, ensuring uniform and continuous interfacial heat conduction between the bridge steel structure substrate and the functional layer, effectively avoiding additional thermal resistance caused by poor bonding or material thermal mismatch.

[0107] The functional layer 12 is coated on the bonding base layer 11, using heat-resistant acrylic resin as the matrix material and uniformly incorporating 30% to 60% by mass of high infrared emissivity filler. The high infrared emissivity filler is prepared by a nanocomposite process of graphene, carbon nanotubes, and alumina, and has stable high infrared emission characteristics, ensuring that the overall infrared emissivity of the infrared functional coating is stable above 0.90. When there are defects in the bridge steel structure substrate 14, the slight changes in local thermal conductivity will be transmitted to the surface through the functional layer 12, forming temperature anomalies that can be clearly captured by an infrared thermal imager. The functional layer 12 is prepared by placing heat-resistant acrylic resin, high infrared emissivity filler, and dispersant in a high-speed shear emulsifier at a mass ratio of 100:45:2 for thorough dispersion, followed by grinding with a sand mill to a fineness of no more than 30 micrometers to ensure uniform distribution of the filler and smoothness of the coating. In this composite filler system, graphene and carbon nanotubes form a three-dimensional thermally conductive network through nanocomposite technology, which is responsible for providing an efficient heat transfer path; while alumina nanoparticles mainly ensure the high infrared emissivity of the coating.

[0108] The weather-resistant protective layer 13 is the outermost layer, using polyurethane resin as the base material and adding an anti-ultraviolet agent to prevent the inner coating from being eroded by environmental factors including rainwater erosion, ultraviolet radiation, temperature difference changes and dust pollution, ensuring that the infrared functional coating can maintain the stability of optical and thermophysical properties for a long time in the harsh service environment of the bridge; the weather-resistant protective layer 13 is prepared by mixing polyurethane resin, anti-ultraviolet agent and leveling agent in a mass ratio of 100:3:0.5.

[0109] During infrared thermal imaging, the infrared functional coating can effectively convert the minute changes in local thermal conductivity caused by defects in the bridge steel structure into significant temperature differences on the surface that can be detected by the infrared thermal imager, thereby enhancing the contrast and consistency of the thermal signal. The infrared functional coating is firmly bonded to the bridge steel structure substrate through chemical bonding or physical adsorption to form a continuous and dense covering layer.

[0110] Optionally, the timing and method of coating application can be flexibly arranged according to project requirements. Depending on the actual situation, critical parts can be recoated to ensure the accuracy of measurements of critical parts. To improve efficiency, local coating can be applied to critical parts instead of overall coating. The operation is carried out using automated equipment.

[0111] Further preprocessing of the infrared thermal image data includes:

[0112] S301. Extract the original infrared thermal image sequence from the infrared thermal image data; based on the known infrared emissivity of the infrared functional coating and the corresponding ambient background temperature, calibrate the grayscale value of each frame in the infrared thermal image sequence and convert it into an absolute temperature value image that characterizes the real thermal radiation to obtain a multi-frame absolute temperature image sequence.

[0113] S302. Using the trigger signal of each thermal excitation as the unified time origin (t=0), perform inter-frame alignment on the multi-frame absolute temperature image sequence obtained in S301 to generate a transient temperature field sequence that is strictly synchronized in the time dimension.

[0114] For example, unify the timeline of all images to ensure that the nth frame represents the nth frame after the excitation ends in all sequences. At 0.02 seconds, a transient temperature field sequence is generated.

[0115] S303. Based on the precise spatial position and attitude information recorded by the positioning module for each frame of the image, perform geometric correction and registration on each frame of the transient temperature field sequence obtained in S302, and fuse them to generate a seamless two-dimensional panoramic temperature field image covering the entire detection area.

[0116] Specifically, the attitude information includes the shape and angle information of the bridge components.

[0117] For example, this is like taking many partial photos taken from different angles, correcting them, and stitching them together into a complete, undistorted two-dimensional panoramic temperature field image of the bridge surface (including the time dimension).

[0118] S304. Perform contrast stretching on the two-dimensional panoramic temperature field image generated in S303, and linearly normalize its pixel values ​​to the [0,1] interval to form standardized input data for use by the convolutional neural network model.

[0119] Specifically, the convolutional neural network model with online self-optimization function is a pre-trained model using a multi-task learning architecture;

[0120] Using finite element analysis software, a transient heat conduction model was established that perfectly matches the geometry, material properties, and thermal excitation parameters of the experimental specimen. The governing equations of the model are:

[0121] ;

[0122] in, For density, Let ρ be the specific heat capacity, k be the thermal conductivity, Q be the thermal excitation term, T be the temperature, and t be the time. Let be the partial derivative of temperature with respect to time, representing the rate of change of temperature over time; in models with the same defect configuration, The Nabla operator is used to represent the gradient. For temperature gradient, Let be the divergence of heat flow, i.e., the diffusion of heat in space. Solving the above equation yields simulated temperature field data. ;

[0123] The training process of the pre-trained model uses the total loss function. Loss by Defect Type Defect region segmentation loss Defect severity assessment regression loss And a physical constraint loss Weighted summation constitutes the result;

[0124] The physical constraint loss The temperature field evolution data predicted by the convolutional neural network model for training samples is compared with the baseline temperature field data generated by simulation based on physical laws. This is used to measure the consistency between the model prediction and physical laws. The loss function of the neural network model takes into account the physical constraint loss, which makes the output of the neural network model more consistent with the physical laws of heat conduction during the training process.

[0125] The reference temperature field data is obtained by establishing a finite element model that is completely consistent with the experimental specimen used in the training samples in terms of geometry, material physical properties and applied thermal excitation parameters, and generating it by numerical simulation through solving the heat conduction control equation;

[0126] The physical constraint loss The mean square error is the difference between the temperature field data predicted by the model and the reference temperature field data.

[0127] Physical constraint loss Calculated using the following formula:

[0128] ;

[0129] Where N is the number of samples in the training batch, and t is the number of time frames for a single sample. The model predicts the temperature field for the i-th sample at the j-th time step. This is the reference temperature field of the i-th sample at the j-th time step obtained through the numerical simulation.

[0130] During model training, not only was a large amount of experimental data used (measured on specimens with pre-fabricated defects), but a physical constraint loss was also introduced. Specifically, for each experimental sample, a finite element model with identical geometry, materials, and thermal excitation will be built for simulation to obtain a theoretically absolutely accurate "reference temperature field." During training, the mean square error between the model's predicted temperature field and this simulation reference will be included in the total loss, enabling the model to learn the real physical laws of heat conduction, rather than simply memorizing image features.

[0131] The architecture of a convolutional neural network model includes an input layer, an encoder, a decoder, and an output layer;

[0132] The input layer is used to receive preprocessed infrared thermal image data, and its dimension is h×w×t, where h and w are the height and width of a single frame of infrared thermal image, respectively, and t is the number of time frames acquired after the thermal excitation process; the value of each pixel is the calibrated absolute temperature value; specifically, the dimension of the input layer is set to H×W×t=256×256×10, indicating that the input is a 256×256 pixel infrared thermal image of 10 consecutive time frames; the value of each pixel is the calibrated absolute temperature value.

[0133] The encoder consists of a spatial feature extraction backbone and a temporal feature fusion module connected in series, used to extract multi-level spatial and temporal features from the input data. The spatial feature extraction backbone uses a ResNet-34 network pre-trained on the ImageNet dataset as the backbone network and is adapted to multi-temporal frame input. To adapt to multi-temporal frame input, the first convolutional kernel of ResNet-34 is modified to 7×7×10 to fuse temporal dimension information in the initial stage. The four residual stages of ResNet-34 downsample the input image to generate feature maps of different scales, with the spatial size decreasing and the number of channels increasing step by step, thereby capturing multi-level spatial features from local texture to global semantics. The temporal feature fusion module feeds the output feature map of the last residual stage of ResNet-34 into a bidirectional convolutional long short-term memory network layer. By introducing a gating mechanism and recurrent connections in the convolution operation, it learns the temporal dependencies of the feature maps, thereby accurately capturing the differences in dynamic thermal behavior between defective and healthy regions during the thermal excitation and cooling process.

[0134] The decoder employs a symmetrical structure based on U-Net, gradually recovering spatial details through upsampling and skip connections. The upsampling process begins with the feature map output from the encoder end, and through a series of deconvolutional layers or bilinear interpolation upsampling operations, the spatial size of the feature map is gradually enlarged to the size of the original input image. The skip connections refer to concatenating the feature map of the current layer of the decoder with the feature map of the same spatial scale in the encoder path at each upsampling stage, thereby fusing shallow detail features and deep semantic features and improving the segmentation accuracy of defect boundaries.

[0135] The output layer consists of three branches: a defect type classification branch, a defect region segmentation branch, and a defect severity assessment branch. The defect type classification branch first performs global average pooling on the feature map output by the decoder, compressing the spatial features into a global feature vector. Then, it outputs a five-dimensional probability vector P=(P1,P2,P3,P4,P5) through two fully connected layers and a Softmax activation function, where P1, P2, P3, P4, and P5 represent the probability distributions of cracks, debonding, corrosion, loose connections, and welding defects in the input image, respectively. The defect region segmentation branch generates a probability map of the same size as the input image, and then outputs the location and area information of the defect region. This branch directly applies a 1×1 convolutional layer to the feature map output by the decoder, reducing the number of channels to 1, and then uses a sigmoid activation function to generate a probability map of the same size as the input image. The value of each pixel in the probability map represents the probability that the location belongs to a defect region. By setting a threshold of 0.5, the probability is... Figure 2 The defect is quantified to obtain a defect segmentation mask, thereby outputting the location and area information of the defect region. The defect severity assessment branch adopts a regression network, whose input is the defect region features provided by the defect region segmentation branch. A fixed-size feature vector is extracted through regional interest pooling, and then passed through two fully connected layers. Finally, a scalar between 0 and 1 is output, representing the quantitative assessment result of the defect severity. The closer the value is to 1, the more severe the defect.

[0136] Specifically, the online self-optimization function includes:

[0137] Based on the defect identification and evaluation results output in real time by the convolutional neural network model, high-confidence diagnostic samples with a confidence level higher than a preset threshold are automatically selected.

[0138] Specifically, for each prediction the model makes (such as "this is a crack"), it automatically generates a "confidence score" between 0 and 1 based on internal calculations. This is the confidence level.

[0139] Comparison with a preset standard: A preset threshold is set in advance, such as 0.95. When a prediction's confidence score is higher than this threshold (e.g., 0.98), it is judged as a "high confidence" prediction, meaning that the model is extremely confident about this result and the quality is very high.

[0140] When the number of high-confidence diagnostic samples accumulates to a preset scale, the incremental learning process is automatically triggered.

[0141] In the incremental learning process, the parameters of the encoder of the convolutional neural network model are fixed, while the parameters of the decoder and output layer are fine-tuned.

[0142] The high-confidence diagnostic samples and their corresponding complete test data are archived into a defect knowledge base that is categorized and indexed by bridge, component, and defect type.

[0143] The preset threshold is 0.95; the preset size is 100 diagnostic samples; and the learning rate used when fine-tuning the decoder and output layer parameters is 0.00001.

[0144] For the identification results with extremely high model confidence (e.g., >0.95) in this detection, the system marks them as "high confidence diagnostic samples". When 100 such samples are accumulated, incremental learning is automatically triggered.

[0145] During learning, the model encoder is fixed (preserving the learned general physical features), and only the decoder and output layer are fine-tuned (learning rate 0.00001) to adapt to the characteristics of new data while avoiding "catastrophic forgetting".

[0146] These high-confidence samples and their complete data (original images, locations, and results) are archived in a defect knowledge base and indexed by methods such as "XX Bridge - Main Beam - Corrosion" to provide valuable information for the continuous evolution of the model and historical tracing.

[0147] Specifically, the training dataset used to train the convolutional neural network model is constructed through a combination of experimental simulation and numerical simulation.

[0148] The numerical simulation, based on finite element analysis, simulates the transient heat conduction process of the infrared functional coating under thermal excitation and couples the perturbation effect of defects on heat flow distribution to generate infrared thermal image data with precise physical labels. The experimental simulation involves artificially introducing defects of different types, sizes, and severity onto calibrated bridge steel structure specimens and collecting their infrared thermal image data. During training, a transfer learning strategy is adopted, using pre-trained ResNet-34 weights for initialization to accelerate convergence and improve recognition accuracy under small sample sizes.

[0149] This embodiment demonstrates a complete technical closed loop, from signal enhancement (coating), active excitation acquisition (UAV), data normalization (preprocessing), to intelligent diagnosis and evolution (physical constraint model). It is not simply a combination of existing technologies, but rather an innovative solution to the challenges of detecting, quantifying, and automating the inspection of minute defects in bridge steel structures through the deep coupling of coating materials science, thermophysics, UAV engineering, and physical information artificial intelligence. This achieves highly sensitive, fully automated, quantitative, and continuously learning-capable non-destructive testing.

[0150] Example 3

[0151] To better understand the non-destructive testing system for bridge steel structure defects based on infrared thermal imaging in the above embodiments, this embodiment describes a method and process based on the system. Figure 4 The flowchart of a non-destructive testing method for bridge steel structure defects based on infrared thermal imaging according to this application is shown below. Figure 4 As shown, the process of this non-destructive testing method for bridge steel structure defects based on infrared thermal imaging includes:

[0152] S1. Application of infrared functional coating: A special infrared functional coating is uniformly applied to the surface of the bridge steel structure to convert the defect information of the bridge steel structure into an easily detectable infrared thermal image signal; the infrared functional coating is composed of a high infrared emissivity material with an infrared emissivity greater than or equal to 0.9, and has uniform and controllable thermal conductivity and thermal diffusivity.

[0153] S2. Infrared thermal imaging inspection: Under normal service conditions, a thermal imaging inspection device is used to inspect the surface of the bridge steel structure that has been coated with the infrared functional coating.

[0154] S3. Preprocessing of Infrared Thermal Image Data: The raw infrared thermal image data acquired in step S2 is preprocessed, including temperature calibration, time-domain registration, spatial-domain stitching, data enhancement, and normalization, to obtain preprocessed infrared thermal image data. The data preprocessing includes the following steps:

[0155] S301, Temperature Calibration: Based on the known emissivity of the infrared functional coating and the ambient background temperature, the original image grayscale value is converted into an absolute temperature value to eliminate environmental radiation interference;

[0156] S302, Time Domain Registration: Align continuously acquired infrared thermal image data to ensure the accuracy of pixel positions in the time dimension;

[0157] S303, Spatial Domain Stitching: Using the positioning module data, multiple local infrared thermal images are spatially stitched and geometrically corrected to generate infrared thermal image data covering the entire detection area;

[0158] S304. Data Enhancement and Normalization: The infrared thermal image data obtained in step S303 is subjected to contrast enhancement processing to highlight the thermal anomaly signal, and the temperature data is normalized to the range of [0,1] to form preprocessed infrared thermal image data.

[0159] S4. Infrared thermal image recognition based on neural network model: The preprocessed infrared thermal image data obtained in step S3 is used as input and fed into a pre-trained convolutional neural network model for automatic identification and quantitative evaluation of defect features; the defect features include cracks, debonding, corrosion, loose connections and welding defects; the convolutional neural network model adopts a multi-task learning architecture, which can simultaneously realize defect type classification, defect region segmentation and defect severity evaluation.

[0160] S5. Self-optimization of the neural network model: The convolutional neural network model has online self-optimization capability, and incrementally learns the model parameters based on real-time collected data, specifically including the following steps:

[0161] S501, High Confidence Data Filtering: Automatically label samples with confidence levels higher than a preset threshold in the real-time identification results and store them in a temporary incremental dataset;

[0162] S502, Incremental Learning Trigger: When the temporary incremental dataset accumulates to a preset size, the incremental learning process is automatically triggered;

[0163] S503, Online parameter fine-tuning: Fix the encoder parameters of the model, and only fine-tune and optimize the decoder and output layer, using a preset learning rate for a small number of iterations;

[0164] S504, Defect Knowledge Base Update: After the incremental learning process is completed, samples with confidence scores higher than the preset threshold and their identification results are archived to a persistent defect knowledge base, categorized and indexed by bridge, component, and defect type, to support model traceability and retraining.

[0165] This embodiment presents a non-destructive testing method for bridge steel structure defects based on infrared thermal imaging. By applying a high infrared emissivity functional coating, the thermal signal of minute defects is significantly amplified. Combined with active thermal excitation by a UAV, efficient and safe non-contact full-coverage detection is achieved. High-quality temperature field data is obtained through spatiotemporal fusion preprocessing and input into a multi-task convolutional neural network model trained based on physical mechanisms. This enables automated and quantitative intelligent identification of defect type, location, and severity. Furthermore, the model has online self-optimization capabilities, continuously evolving and accumulating defect knowledge through incremental learning. Ultimately, this forms a complete solution for non-destructive testing of bridge steel structure defects that is highly sensitive, fully automated, quantifiable, and possesses continuous learning capabilities.

[0166] It should be noted that any reference numerals placed between parentheses in the claims should not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In claims that enumerate several means, several of these means may be embodied by the same hardware. The use of the terms first, second, third, etc., is merely for convenience of expression and does not indicate any order. These terms can be understood as part of the component names.

[0167] Furthermore, it should be noted that in the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0168] In this application, unless otherwise expressly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0169] In this application, unless otherwise expressly specified and limited, "above" or "below" the second feature can mean that the first and second features are in direct contact, or that they are in indirect contact through an intermediate medium. Furthermore, "above," "on top of," and "over" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0170] In the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0171] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A non-destructive testing system for bridge steel structure defects based on infrared thermal imaging, characterized in that, include: Unmanned aerial vehicles and ground computing stations integrated with infrared thermal imaging detection devices; The UAV flies along a designated detection path according to the infrared image acquisition command sent by the ground computing station, and collects infrared thermal image data of the bridge steel structure within the area of ​​the designated detection path in real time. The ground computing station receives infrared thermal image data of all test areas of the bridge transmitted by the UAV, and preprocesses the infrared thermal image data to obtain normalized image data of the entire detection area that eliminates environmental radiation interference and is aligned with the time and space dimensions. The ground computing station uses a convolutional neural network model that incorporates physical mechanism constraints and has online self-optimization to identify and evaluate the image data, and simultaneously outputs quantitative evaluation results of defect type, defect area and defect severity. The entire inspection area of ​​the bridge is pre-coated with an infrared functional coating, which is a multi-layer composite structure made of high infrared emissivity material, with an infrared emissivity greater than or equal to 0.

9. Real-time acquisition of infrared thermal image data of bridge steel structures within the area of ​​the specified detection path includes: applying transient thermal flux excitation to the area under test using an infrared radiation source, and simultaneously acquiring dynamic response information of the infrared functional coating surface after thermal flux excitation.

2. The system according to claim 1, characterized in that, The drone integrates a thermal imaging detection device. The thermal imaging detection device includes: a thermal excitation device, an infrared thermal imager, a temperature sensor, a positioning module, a data transmission module, and a flight control module; The ground computing station transmits infrared image acquisition commands and detection paths via a data transmission module. The flight control module flies based on the detection path; The thermal excitation device is used to apply transient thermal flux excitation to the area under test based on infrared image acquisition commands; The infrared thermal imager is used to continuously acquire the dynamic thermal response of the infrared functional coating surface at a frame rate greater than or equal to a preset frame rate after the thermal excitation device applies heat flow excitation, and to obtain an infrared thermal image sequence composed of multiple frames, wherein each frame contains temperature distribution information of the coating surface at that moment. The temperature sensor is used to measure and record the ambient background temperature before the start of each thermal excitation cycle; The positioning module is used to record the spatial coordinates and absolute timestamp of each frame of infrared thermal image when it is acquired in real time. The data transmission module is used to transmit the infrared thermal image sequence and its corresponding spatial location coordinates, absolute timestamps and ambient background temperature to the ground computing station. The infrared thermal image data includes: the original sequence of infrared thermal images consisting of multiple frames acquired by the infrared thermal imager; the ambient background temperature value measured by the temperature sensor and corresponding to it; and the spatial location coordinates and absolute timestamp of each frame of image synchronously recorded by the positioning module.

3. The system according to claim 1, characterized in that, The ground computing station preprocesses the infrared thermal image data, including: S301. Extract the original infrared thermal image sequence from the infrared thermal image data; based on the known infrared emissivity of the infrared functional coating and the corresponding ambient background temperature, calibrate the grayscale value of each frame in the infrared thermal image sequence and convert it into an absolute temperature value image that characterizes the real thermal radiation to obtain a multi-frame absolute temperature image sequence. S302. Using the trigger signal of each thermal excitation as the unified time origin, perform inter-frame alignment on the multi-frame absolute temperature image sequence obtained in S301 to generate a transient temperature field sequence that is strictly synchronized in the time dimension. S303. Based on the precise spatial position and attitude information recorded by the positioning module for each frame of the image, perform geometric correction and registration on each frame of the transient temperature field sequence obtained in S302, and fuse them to generate a seamless two-dimensional panoramic temperature field image covering the entire detection area. S304. Perform contrast stretching on the two-dimensional panoramic temperature field image generated in S303, and linearly normalize its pixel values ​​to the [0,1] interval to form standardized input data for use by the convolutional neural network model.

4. The system according to claim 1, characterized in that, The convolutional neural network model with online self-optimization function is a pre-trained model using a multi-task learning architecture; The training process of the pre-trained model uses the total loss function. Loss by Defect Type Defect region segmentation loss Defect severity assessment regression loss And a physical constraint loss Weighted summation constitutes the result; The physical constraint loss The temperature field evolution data predicted by the convolutional neural network model for training samples is compared with the baseline temperature field data generated by simulation based on physical laws. This is used to measure the consistency between the model prediction and physical laws. The reference temperature field data is obtained by establishing a finite element model that is completely consistent with the experimental specimen used in the training samples in terms of geometry, material physical properties and applied thermal excitation parameters, and generating it by numerical simulation through solving the heat conduction control equation; The physical constraint loss The mean square error is the difference between the temperature field data predicted by the model and the reference temperature field data.

5. The system according to claim 4, characterized in that, The physical constraint loss Calculated using the following formula: ； Where N is the number of samples in the training batch, and t is the number of time frames for a single sample. The model predicts the temperature field for the i-th sample at the j-th time step. This is the reference temperature field of the i-th sample at the j-th time step obtained through the numerical simulation.

6. The system according to claim 1, characterized in that, The infrared functional coating comprises, from the inside out, an adhesive base layer, a functional layer, and a weather-resistant protective layer. The functional layer contains a high infrared emissivity filler prepared by a nanocomposite process. The high infrared emissivity filler is composed of graphene, carbon nanotubes and alumina, and its mass fraction in the functional layer is 30% to 60%. The substrate material of the bonding base layer is modified epoxy resin, wherein 10% by mass of metallic aluminum powder is dispersed.

7. The system according to claim 1, characterized in that, Before the ground computing station uses the convolutional neural network model to identify and evaluate the image data, it also includes a step of online adaptive optimization of the model, including: Based on the defect identification and evaluation results output in real time by the convolutional neural network model, high-confidence diagnostic samples with a confidence level higher than a preset threshold are automatically selected. When the number of high-confidence diagnostic samples accumulates to a preset scale, the incremental learning process is automatically triggered. In the incremental learning process, the parameters of the encoder of the convolutional neural network model are fixed, while the parameters of the decoder and output layer are fine-tuned. The high-confidence diagnostic samples and their corresponding complete test data are archived into a defect knowledge base that is categorized and indexed by bridge, component, and defect type.

8. The system according to claim 7, characterized in that, The preset threshold is 0.95; the preset size is 100 diagnostic samples; and the learning rate used when fine-tuning the decoder and output layer parameters is 0.00001.

9. The system according to claim 2, characterized in that, The thermal excitation device is a high-power, short-pulse infrared radiation source with an output power range of 3kW to 5kW and a single pulse duration range of 100 milliseconds to 500 milliseconds; the infrared thermal imager operates at a frame rate ≥50Hz.

10. The system according to claim 4, characterized in that, A training dataset for training the convolutional neural network model is constructed by combining experimental simulation and numerical simulation. The experimental simulation was achieved by collecting thermal response data of specimens coated with the infrared functional coating and prefabricated with defects of different types and sizes under thermal excitation. The numerical simulation is based on finite element analysis and simulates the transient temperature field of the infrared functional coating under thermal excitation consistent with experimental conditions and coupled with different defects.

Images