Bridge disease detection positioning and adsorption walking control method and system

Abstract

This invention discloses a method and system for bridge defect detection, positioning, and adsorption-based movement control, relating to the field of bridge inspection technology. The system includes a mobile carrier, a composite adsorption-based movement module, a defect identification and detection module, a positioning and mapping module, and a control module. The composite adsorption-based movement module includes a magnetically assisted adsorption unit and a negative pressure adsorption unit, and can be configured with a wind-assisted adsorption unit to adapt to bridge surfaces with different materials, curvatures, and crack conditions. The defect identification and detection module is used to screen candidate defect areas on the bridge surface. The positioning and mapping module is used to record the movement path and keyframe pose information. The control module is used to perform main and auxiliary adsorption switching, collaborative compensation control, and defect positioning matching. This invention enables integrated operation of stable adhesion to bridge surfaces, continuous inspection, defect identification, and spatial positioning, improving the completeness of detection, identification stability, positioning accuracy, and overall practical application effect in complex bridge deck environments.

Description

Technical Field

[0001] This invention belongs to the field of bridge inspection technology, specifically relating to a method and system for bridge defect detection, positioning, and adsorption-based movement control. Background Technology

[0002] With the advancement of technology and society, the number of infrastructure bridges is increasing. During long-term service, bridges are prone to defects such as cracks, peeling, voids, corrosion, bulges, and localized deformation on the bridge deck, pier sidewalls, bearing areas, and corner connections. These defects are often scattered and located in complex areas, sometimes accompanied by dampness, rough surfaces, localized cracks, material changes, and insufficient lighting. Many regions in my country have introduced corresponding bridge maintenance policies; however, manual identification of bridge defects still faces many challenges. Several bridge defect detection solutions have been proposed to address these technical issues.

[0003] However, these bridge disaster detection solutions still have some shortcomings. For example, most detection methods can only detect disasters at fixed points over long distances, or only have the function of image recognition, lacking the ability to scan and clean the entire area and identify damage at close range. Although they can obtain local appearance information, they are often insufficient in identifying small defects, obstructed areas, and curved corners, and it is difficult to directly give the accurate location of the defects on the bridge surface. Some intelligent bridge disaster detection methods can only identify bridge damage at fixed angles; however, in most cases, bridge disasters are a long-term, overall development process, so some disaster images cannot reflect the overall extent of damage to the bridge.

[0004] Therefore, there is a need for a bridge defect detection and positioning system and method that can stably attach and continuously travel in complex bridge deck environments, while taking into account defect identification and spatial positioning output. Summary of the Invention

[0005] To address the aforementioned shortcomings, this invention provides a method and system for bridge defect detection, location, and adsorption-based movement control. This application integrates a magnetically assisted adsorption unit, a negative pressure adsorption unit, a wind pressure assisted adsorption unit, a defect identification and detection module, a location mapping module, and a control module onto a mobile carrier. This allows for continuous completion of bridge surface movement, defect identification, and spatial positioning within the same system. Compared to detection methods that can only perform long-distance imaging or fixed-point identification, this application enables close-range inspection along the bridge surface. After identifying candidate defect areas, it continues to approach the target location to complete precise identification and simultaneously outputs the corresponding location of the defect on the bridge surface. Therefore, it is more effective in reflecting the true distribution and development range of bridge defects.

[0006] This invention provides the following technical solution: a method for bridge defect detection, positioning, and adsorption-based movement control, applied to a bridge defect detection and positioning system with magnetic-assisted adsorption movement, negative pressure adsorption movement, and wind pressure-assisted adsorption movement, the method comprising:

[0007] The surface material of the bridge surface that the robot is currently traveling on is identified to determine whether the bridge surface is made of stone or steel.

[0008] When the bridge surface is determined to be stone, the system detects whether there are gaps in the corresponding travel area. If there are no gaps, the control system enters the normal negative pressure travel mode, which uses air pressure sensors on several suction cups to detect the air pressure in the cavity of each suction cup in real time, and performs closed-loop control of negative pressure adsorption based on the detection results. If there are gaps, the control system enters the multi-system collaborative work mode, which simultaneously activates the wind pressure-assisted adsorption travel mode and the magnetic force-assisted adsorption travel mode, and works in conjunction with the negative pressure adsorption travel mode to pass through the stone bridge surface area with gaps.

[0009] When the bridge surface is determined to be steel, the steel ratio of the corresponding travel area is detected. If the steel ratio is higher than a preset threshold, the control system enters the normal magnetic walking mode, and the robot moves by providing vertical adsorption force through the permanent magnet module. If the steel ratio is lower than or equal to the preset threshold, the control system enters the multi-system collaborative working mode, and simultaneously starts the wind pressure-assisted adsorption walking mode and the negative pressure adsorption walking mode, and works in conjunction with the magnetic adsorption walking mode to pass through areas with low steel ratio.

[0010] Regardless of whether the bridge deck is made of stone or steel, as long as the system enters the multi-system collaborative working mode, the wind pressure-assisted adsorption walking mode is activated to provide additional adsorption force and improve the robot's stable walking ability in abnormal bridge deck areas. When walking in normal walking and / or multi-system collaborative working mode, real-time image information of the bridge deck is collected, and then the bridge deck is identified by the defect identification and detection method to determine whether the bridge deck has defects and the type of defects.

[0011] Furthermore, when the bridge deck is a seamless stone bridge deck, the normal negative pressure walking mode includes the following steps:

[0012] S21: Acquire real-time air pressure data collected by air pressure sensors corresponding to several suction cups;

[0013] S22: Input the real-time air pressure data into the control unit and compare it with the target negative pressure value to obtain the air pressure deviation corresponding to each suction cup;

[0014] S23: Based on the air pressure deviation, perform PID closed-loop control and output the corresponding air pump drive signal to adjust the negative pressure in each suction cup cavity;

[0015] S24: During the robot's movement, the negative pressure of each suction cup is kept within a preset stable range to maintain normal negative pressure adsorption and walking state.

[0016] Furthermore, when the bridge deck is a stone bridge deck with gaps, the multi-system collaborative working mode includes the following steps:

[0017] S31: Based on the gap detection results, the current stone bridge surface area is determined to be an abnormal travel area;

[0018] S32: Activate the wind pressure-assisted adsorption walking mode, where an additional adsorption force is generated by a centrifugal fan;

[0019] S33: Activate the magnetic-assisted adsorption walking mode, with the permanent magnet module providing vertical auxiliary adsorption force;

[0020] S34: Maintain negative pressure adsorption walking mode and dynamically adjust the negative pressure adsorption capacity based on the real-time air pressure detection results of several suction cups;

[0021] S35: After the robot passes through the abnormal travel area, it resumes the corresponding normal walking mode according to the coordinated output results of wind pressure-assisted adsorption force, magnetic force-assisted adsorption force and negative pressure adsorption force.

[0022] Furthermore, when the bridge deck is a steel bridge deck with a high steel content, the normal magnetic travel mode includes the following steps:

[0023] S41: Detect the steel ratio of the steel bridge deck and compare the detection result with the preset steel ratio threshold;

[0024] S42: When the steel ratio is higher than the preset steel ratio threshold, control the permanent magnet module to enter the normal adsorption working state to provide magnetic adsorption force perpendicular to the bridge surface;

[0025] S43: During the robot's movement, adjust the direction or position of magnetic adsorption according to the local undulations of the bridge surface to maintain stable magnetic adsorption;

[0026] S44: Utilizes magnetic adsorption as the primary adsorption method to maintain the robot's normal movement.

[0027] Furthermore, when the bridge deck is a steel bridge deck with a relatively low steel content, the multi-system collaborative working mode includes the following steps:

[0028] S51: When the steel ratio detection result is not higher than the preset steel ratio threshold, the current steel bridge deck area is determined to be a low steel ratio area;

[0029] S52: Activate the wind pressure-assisted adsorption walking mode, where an additional adsorption force is generated by a centrifugal fan;

[0030] S53: Activate the negative pressure adsorption walking mode, forming a distributed negative pressure adsorption through several suction cups;

[0031] S54: Maintain the magnetic-assisted adsorption walking mode so that the permanent magnet module continues to output magnetic adsorption force;

[0032] S55: Based on the synergistic output results of wind pressure adsorption force, negative pressure adsorption force and magnetic adsorption force, the robot is controlled to stably pass through the low steel ratio area and then return to the normal magnetic walking mode or other matching normal walking mode.

[0033] Furthermore, the PID closed-loop control in step S23 includes: performing proportional regulation, integral regulation and derivative regulation based on the deviation between the target negative pressure value and the real-time air pressure value to generate a PWM control signal, and adjusting the air pump output power corresponding to each suction cup according to the PWM control signal, so that the negative pressure enters a fine maintenance state when it is close to the target negative pressure value, and enters a rapid compensation state when it deviates from the target negative pressure value.

[0034] Furthermore, in the multi-system collaborative working mode, the control unit performs graded control of the wind pressure-assisted adsorption walking mode according to the degree of bridge deck anomaly, including basic auxiliary mode, medium auxiliary mode and enhanced auxiliary mode; wherein, the degree of bridge deck anomaly is determined at least according to one or more of the following: gap size, number of gaps, steel ratio deviation, and adsorption stability index.

[0035] Furthermore, the step of maintaining the negative pressure adsorption walking mode also includes:

[0036] The real-time air pressure of several suction cups is compared and analyzed to determine the deviation between the air pressure of each suction cup and the average air pressure.

[0037] Suction cups with deviations exceeding a preset threshold are identified as abnormal suction cups.

[0038] Adjust the target negative pressure value or suction power of the remaining suction cups according to the number and distribution of abnormal suction cups in order to maintain the overall negative pressure adsorption capacity basically stable.

[0039] Furthermore, the disease identification and detection method includes the following steps:

[0040] S71: Normalize the collected bridge damage images and perform wavelet denoising to obtain the denoised reconstructed image;

[0041] S72: Extract grayscale structure map, morphological gradient map and local binary pattern texture map from the denoised image, and construct a multi-channel fused input tensor;

[0042] S73: Generate a salient weight map of cracks based on morphological gradient information and local texture information, and perform salient weight guidance on the fused input tensor;

[0043] S74: Input the weighted fused image into the improved VGG16 network, which passes through convolutional layers, activation layers, pooling layers and fully connected layers in sequence, and outputs the predicted probability of each disease category.

[0044] S75: Construct classification loss, enhanced consistency loss, and significant guidance loss, and use a joint loss function to iteratively optimize the network parameters;

[0045] S76: Perform forward inference on the image of the bridge under test and output the defect category with the highest probability as the final identification result.

[0046] The present invention also provides a bridge defect detection and positioning system, including a bridge deck property detection module, a composite adsorption walking module, a defect identification and detection module, a positioning and mapping module, and a control module; the bridge deck property detection module includes a gap detection module and a steel ratio detection module; the composite adsorption walking module includes a magnetic assisted adsorption unit, a negative pressure adsorption unit, and a wind pressure assisted adsorption unit;

[0047] The bridge surface attribute detection module is used to identify the surface material of the bridge surface that the robot is currently traveling on, so as to determine whether the bridge surface is a stone bridge surface or a steel bridge surface.

[0048] The gap detection module is used to detect whether there are gaps in the corresponding travel area of ​​the stone bridge surface when the bridge surface is a stone bridge surface;

[0049] The steel ratio detection module is used to detect the steel ratio of the corresponding travel area of ​​the steel bridge deck when the bridge deck is made of steel.

[0050] The magnetically assisted adsorption unit includes a permanent magnet module for providing vertical adsorption force;

[0051] The negative pressure adsorption unit includes a plurality of suction cups and air pressure sensors respectively arranged corresponding to the plurality of suction cups;

[0052] The wind pressure-assisted adsorption unit includes a centrifugal fan for generating additional adsorption force;

[0053] The defect identification and detection module is used to collect real-time image information of the bridge deck and then identify whether the bridge deck has defects and the type of defects through defect identification and detection methods.

[0054] The positioning and mapping module is used to acquire path information of the bridge surface and record key frame pose information during the movement of the mobile carrier along the bridge surface, and match the defect identification result output by the defect identification and detection module with the corresponding key frame pose information, and obtain the positioning result of the defect on the bridge surface through coordinate transformation.

[0055] The control module is connected to the bridge deck attribute detection module, the gap detection module, the steel ratio detection module, the composite adsorption walking module, the defect identification and detection module, and the positioning and mapping module, and is configured to execute the adsorption walking control method described above.

[0056] The beneficial effects of this invention are as follows:

[0057] 1. This application sets up a front-end rapid recognition unit and a close-range detection unit in the defect identification and detection module. First, the bridge surface image is processed by grayscale conversion, filtering, and threshold segmentation to quickly screen out candidate areas of obstacles or defects. Then, the target area is subjected to close-range fine recognition, thus balancing detection speed and recognition accuracy. This setup can reduce the time consumption caused by the whole machine blindly stopping on a large area of ​​bridge surface, and improve the targeting of defects such as cracks, peeling, voids, corrosion, bulges, and deformation.

[0058] 2. The disease identification and detection method in this application, based on the original identification network, introduces wavelet denoising, fusion input of grayscale structure map, morphological gradient map and local binary pattern texture map, crack saliency weight guidance, and enhanced consistency constraints, so that the network can retain edge and texture information related to the disease as much as possible before entering the classification judgment. Experimental results show that Example 7 achieves recognition rates of 93%, 91%, 90%, 93%, 91%, and 91% for six types of diseases: cracks, peeling, voids, corrosion, bulging, and deformation, respectively. Compared with the traditional VGG16, these rates are improved by 9, 10, 10, 8, 10, and 9 percentage points, respectively, indicating that this application has better distinguishing ability in the identification of fine cracks, weak texture diseases, and diseases with unclear boundaries.

[0059] 3. The supplementary lighting component in this application can automatically adjust the luminous intensity according to the ambient illuminance, thereby improving the image acquisition quality under bridges, backlit surfaces, nighttime operations, and low-illuminance areas, and reducing the impact of shadows, reflections, and insufficient brightness on the recognition results. Combined with the experimental results of Example 7, under conditions of low illumination, noise, blur, reflection, weak texture, and small sample size, the indicators of Example 7 reached 0.90, 0.88, 0.87, 0.85, 0.86, and 0.89, all significantly higher than traditional methods. This indicates that this application not only improves the recognition accuracy under ideal conditions but also enhances the recognition stability under complex on-site conditions.

[0060] 4. This application matches the defect identification results with the keyframe pose information recorded by the positioning and mapping module, and obtains the location result of the defect on the bridge surface through coordinate transformation. This allows the output result to go beyond a simple image recognition conclusion and further provide the spatial location of the defect. This not only facilitates subsequent maintenance personnel to quickly reach the target area, but also facilitates the formation of bridge defect distribution maps and inspection records, thereby improving the pertinence and continuity of bridge operation and maintenance work.

[0061] 5. This application organically combines stable adhesion to bridge surfaces, movement in complex environments, defect identification, and spatial positioning, which not only improves the automation level of bridge defect detection but also enhances the completeness, stability, and traceability of the detection results. This application has significant practical value for scenarios requiring long-term inspections of the bridge underside, pier sidewalls, corner connection areas, and areas with a high incidence of local defects. Attached Figure Description

[0062] The invention will now be described in more detail with reference to embodiments and the accompanying drawings.

[0063] Figure 1 This is a general system block diagram of the bridge defect detection and location system of this application;

[0064] Figure 2 This is a flowchart of the adsorption and walking control process under the working conditions of a stone bridge deck in this application;

[0065] Figure 3 This is a flowchart illustrating the adsorption and travel control process under steel bridge deck conditions as described in this application.

[0066] Figure 4 This is a schematic diagram of the suction cup structure in the negative pressure adsorption unit of this application;

[0067] Figure 5 This is a schematic diagram showing the pressure change over time during the negative pressure adsorption process of this application;

[0068] Figure 6 This is a schematic diagram showing the change of gas pressure over time during the negative pressure adsorption process of this application;

[0069] Figure 7 This is a schematic diagram of the negative pressure adsorption control system of this application.

[0070] Figure 8 This is a schematic diagram of the vehicle body structure of the mobile carrier of this application;

[0071] Figure 9 This is a schematic diagram of the internal structure of the magnetic wheel shaft in this application;

[0072] Figure 10 This is a schematic diagram of the magnetic adsorption and movement of the mobile carrier of this application under different bridge deck postures;

[0073] Figure 11 This is a schematic diagram of the magnetization curve and permeability curve of the magnetic material used in this application;

[0074] Figure 12 This is a flowchart illustrating the linkage control of the negative pressure adsorption stagnation and magnetic wall climbing system in this application.

[0075] Figure 13 This is a flowchart illustrating the steps for identifying surface cracks on bridges in this application.

[0076] Figure 14 This is a schematic diagram of the simple mechanical obstacle removal system of this application;

[0077] Figure 15 This is a flowchart of the supplemental lighting adjustment control process for this application;

[0078] Figure 16 This is a schematic diagram comparing the image acquisition of the three-channel camera and the single-phase camera in this application;

[0079] Figure 17 This is a schematic diagram of the distance measuring module of this application;

[0080] Figure 18 This is a schematic diagram of the object distance and camera imaging model in the adsorption detection platform of this application;

[0081] Figure 19 This is a flowchart of the disease location method in this application.

[0082] Figure 20 This is a schematic diagram of the Raspberry Pi hardware system architecture for this application;

[0083] Figure 21 This is a schematic diagram of the YOLO MobileNet disease identification network structure in this application;

[0084] Figure 22 This is a schematic diagram illustrating the process of establishing the bridge underside crack dataset for this application.

[0085] Figure 23 This is a schematic diagram showing the labels of the dataset in this application;

[0086] Figure 24 This is a schematic diagram illustrating the image data enhancement effect of this application;

[0087] Figure 25 This is a distribution matrix diagram showing the recognition improvement of Embodiment 7 of this application compared to traditional recognition methods;

[0088] Figure 26 This is a comparison diagram of the feature space distribution of Embodiment 7 of this application and traditional recognition methods. Detailed Implementation

[0089] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0090] Example 1

[0091] This embodiment provides a bridge defect detection and positioning system, including a mobile carrier, a composite adsorption walking module, a defect identification and detection module, a positioning and mapping module, and a control module. The composite adsorption walking module is mounted on the mobile carrier and is used to enable the mobile carrier to travel along the bridge surface, overcome obstacles, and stop at fixed points. The composite adsorption walking module includes at least a magnetically assisted adsorption unit and a negative pressure adsorption unit. The defect identification and detection module is used to acquire images of bridge surface defects and output defect identification results. The positioning and mapping module is used to acquire the travel path information of the mobile carrier on the bridge surface and record keyframe pose information. The control module is connected to the composite adsorption walking module, the defect identification and detection module, and the positioning and mapping module, respectively. It is used to perform master-slave switching control between the magnetically assisted adsorption unit and the negative pressure adsorption unit based on the material characteristics, surface smoothness, adsorption state, and operating state of the bridge surface, and to match the defect identification results with the keyframe pose information to output the location result of the defect on the bridge surface. See the overall system configuration for details. Figure 1 For the operational procedures on different bridge surfaces, please refer to the traffic procedures for stone bridges. Figure 2 For the travel procedures on the iron bridge, please refer to [link / reference]. Figure 3 .

[0092] In this embodiment, the mobile carrier preferably adopts a four-wheeled vehicle structure, with the vehicle body used to carry the adsorption actuator, image acquisition components, positioning and mapping components, and control hardware. The vehicle body can be designed as a low-center-of-gravity frame structure to improve overall stability when operating on the bridge underside, bridge web sides, and pier facades. A detection component mounting bracket can be installed at the front of the mobile carrier to mount a front-end rapid recognition camera, a close-range precision inspection camera, and auxiliary lighting components; a composite adsorption and movement module is arranged at the bottom or circumferentially to achieve stable attachment and movement in complex bridge environments. The vehicle body shape and installation method can be combined with... Figure 1 The structure shown is arranged as shown.

[0093] In this embodiment, the system is suitable for close-range inspection and location output of bridge surface defects. Compared with detection methods that only perform long-distance fixed-point photography, the mobile carrier in this embodiment can continuously travel along the bridge surface, perform scanning detection in different areas of the bridge, and continue to approach the target location for fine identification after discovering candidate defect areas, thereby improving the ability to grasp the overall development status of bridge defects. In addition to visible light image recognition, the defect identification and detection module can also be equipped with ultrasonic detection components to supplement the detection of cracks, voids, or internal anomalies. This application adopts a composite adsorption walking method, which switches between magnetic adsorption and negative pressure adsorption as the main and auxiliary methods according to the material characteristics, surface flatness, adsorption state, and working state of the bridge surface, and introduces wind-assisted adsorption when necessary, so that the system can adapt to iron bridge decks, stone bridge decks, and shallow reinforced concrete bridge decks. With this setting, the mobile carrier can maintain good obstacle crossing ability in the traveling state and maintain high adhesion stability in the fixed-point detection or long-term residence state, avoiding the problem of limited applicability of a single adsorption method in complex bridge deck environments.

[0094] Example 2

[0095] Based on Example 1, this embodiment, when the bridge deck is a seamless stone bridge deck, includes the following steps in the normal negative pressure walking mode:

[0096] S21: Acquire real-time air pressure data collected by air pressure sensors corresponding to several suction cups;

[0097] S22: Input the real-time air pressure data into the control unit and compare it with the target negative pressure value to obtain the air pressure deviation corresponding to each suction cup;

[0098] S23: Based on the air pressure deviation, perform PID closed-loop control and output the corresponding air pump drive signal to adjust the negative pressure in each suction cup cavity;

[0099] S24: During the robot's movement, the negative pressure of each suction cup is kept within a preset stable range to maintain normal negative pressure adsorption and walking state.

[0100] The structure and control method of the negative pressure adsorption unit are further described below. The negative pressure adsorption unit includes multiple independent suction cup chambers distributed circumferentially along the moving carrier. Each independent suction cup chamber is equipped with a pressure sensor and a suction device. Preferably, multiple suction cup units are distributed around the moving carrier; for example, sixteen independent suction cup chambers can be used to form a compartmentalized adsorption structure. See [link to suction cup structure] for details. Figure 4 In some embodiments, the number of suction cups can be 14, 16, etc., and several suction cups can be arranged horizontally in a straight line, or in a matrix form with several horizontal and vertical equidistant spacing.

[0101] In this embodiment, the stability of negative pressure adsorption is achieved through real-time closed-loop adjustment of the air pressure inside the suction cup cavity. The pressure variation over time and the air pressure variation over time can be found in [references to be inserted here]. Figure 5 and Figure 6 During control, the target negative pressure value and control upper and lower limits are first set via the control panel. A high-precision air pressure sensor collects real-time air pressure data within the sealed cavity of the suction cup at millisecond-level frequency and transmits the data to the control module. The control module calculates the deviation between the target value and the real-time feedback value based on a proportional-integral-derivative control algorithm and outputs corresponding adjustment signals to regulate the power of the suction device. When insufficient negative pressure is detected, the suction power is increased; when the detected value approaches the target value, it enters a fine maintenance state; when the air pressure is within the set range, the suction device is controlled to maintain low power operation, thus balancing adsorption stability and energy consumption control. The process is described in [link to process details]. Figure 7 .

[0102] In this embodiment, when any independent suction cup chamber leaks air, the control module cuts off the corresponding air path and increases the pumping capacity of the remaining independent suction cup chambers to maintain the overall adsorption force above a preset safety threshold. Preferably, the overall adsorption force is maintained at more than twice the weight of the mobile carrier to improve the safety of the stationary operation. For situations where the bridge surface has roughness, cracks, dampness, or localized air permeability, the multi-chamber multi-pump structure can prevent the entire machine from detaching due to the failure of a single chamber.

[0103] Furthermore, the negative pressure adsorption unit also includes a flexible seal and a pre-tightening mechanism. The flexible seal is located at the contact edge between each independent suction cup chamber and the bridge surface, and is preferably made of aging-resistant silicone material. The pre-tightening mechanism is used to drive the flexible seal to dynamically adhere to the bridge surface, thereby reducing the risk of local air leakage. Through this structure, even if there are slight undulations, crack edges, or damp adhesion layers on the bridge surface, the adhesion ability of the suction cup edges can be improved, thereby improving the actual adsorption effect. The negative pressure adsorption unit in this application adopts a structure that combines multiple independent suction cup chambers, pressure sensors, and air extraction devices, and combines crack recognition and pressure compensation control methods. When individual suction cups leak air, the failed air path can be cut off in time, and the working capacity of the remaining effective adsorption units can be improved, so that the total adsorption capacity is kept within a safe range. For bridge surfaces with cracks, potholes, dampness, or local rough areas, this application can still maintain a relatively stable adhesion state, thereby improving the operational reliability of the whole machine on aging bridge bodies and complex bridge surfaces.

[0104] When the bridge deck is a stone bridge deck with gaps, the multi-system collaborative working mode includes the following steps:

[0105] S31: Based on the gap detection results, the current stone bridge surface area is determined to be an abnormal travel area;

[0106] S32: Activate the wind pressure-assisted adsorption walking mode, where an additional adsorption force is generated by a centrifugal fan;

[0107] S33: Activate the magnetic-assisted adsorption walking mode, with the permanent magnet module providing vertical auxiliary adsorption force;

[0108] S34: Maintain negative pressure adsorption walking mode and dynamically adjust the negative pressure adsorption capacity based on the real-time air pressure detection results of several suction cups;

[0109] S35: After the robot passes through the abnormal travel area, it resumes the corresponding normal walking mode according to the coordinated output results of wind pressure-assisted adsorption force, magnetic force-assisted adsorption force and negative pressure adsorption force.

[0110] Example 3

[0111] This embodiment, based on Embodiment 2, further explains the pressure compensation control and wind-assisted adsorption control methods in a fractured bridge deck environment. When the bridge deck is a steel bridge deck with a high steel content, the normal magnetic travel mode includes the following steps:

[0112] S41: Detect the steel ratio of the steel bridge deck and compare the detection result with the preset steel ratio threshold;

[0113] S42: When the steel ratio is higher than the preset steel ratio threshold, control the permanent magnet module to enter the normal adsorption working state to provide magnetic adsorption force perpendicular to the bridge surface;

[0114] S43: During the robot's movement, adjust the direction or position of magnetic adsorption according to the local undulations of the bridge surface to maintain stable magnetic adsorption;

[0115] S44: Utilizes magnetic adsorption as the primary adsorption method to maintain the robot's normal movement.

[0116] For cases where the surface of some aged bridge bodies has numerous cracks, pits, or localized uneven areas, the control module calculates the average air pressure based on the real-time air pressure values ​​of each independent suction cup chamber, and identifies crack adsorption units based on the deviation between the real-time air pressure values ​​of each independent suction cup chamber and the average air pressure. Let the real-time air pressure of the nth suction cup be pn, and the average air pressure of all suction cups be P. The determination can be made based on the absolute value of the difference between P and pn, where the deviation ΔPn equals the difference between P and pn. When the deviation exceeds a preset threshold, the corresponding suction cup is determined to be located in a crack, pit, or poorly sealed area. For identified crack adsorption units, the control module removes them from the effective adsorption units and recalculates the target air pressure or target adsorption force based on the number of remaining effective adsorption units to maintain a relatively constant total adsorption capacity. The crack pressure compensation method is similar to... Figure 2 , Figure 3The different bridge surface driving logics shown are compatible. The magnetically assisted adsorption unit in this application adopts a magnetically driven wheel structure with adjustable magnetic adsorption direction. The magnetic adsorption module can rotate relative to the wheel to adjust the adsorption direction in a timely manner in undulating areas, corner areas, or locations with large local curvature changes on the bridge surface. With this configuration, the adhesion and obstacle-crossing ability of the mobile carrier on vertical surfaces, corner surfaces, and irregular bridge surfaces are improved, and slippage, detachment, or difficulty in passing caused by local adsorption direction mismatch are reduced.

[0117] In this embodiment, the composite adsorption walking module further includes a wind pressure-assisted adsorption unit. The wind pressure-assisted adsorption unit is preferably a centrifugal fan device, used to create additional compressed airflow between the moving carrier and the bridge surface when the magnetic adsorption unit fails to adsorb sufficiently or the negative pressure adsorption unit partially fails. This wind-assisted adsorption method does not rely entirely on a sealed cavity to provide additional adsorption force, thus it is suitable for transitional travel in areas with large-area air leakage, localized high curvature, or low steel ratio. The control module controls the wind pressure-assisted adsorption unit to operate at at least two different power levels according to the degree of adsorption insufficiency. Preferably, a low-power compensation mode and a high-power compensation mode can be set, and if necessary, a higher power level can be set to meet the safety adhesion requirements under different working conditions.

[0118] This embodiment can also be configured with a master-slave collaborative control mode for stone bridge applications. In this mode, the negative pressure adsorption unit provides the primary adsorption force, the wind pressure-assisted adsorption unit provides additional clamping force when some suction cups fail, and, if necessary, the magnetic-assisted adsorption unit provides supplementary support in the shallow steel reinforcement area of ​​the bridge structure. This allows the mobile carrier to move and remain stably on bridge surfaces with numerous cracks or complex local structures. The process flow for stone bridges can be combined with… Figure 2 understand.

[0119] Example 4

[0120] This embodiment, based on Embodiment 1, further explains the magnetically assisted adsorption unit and its movement control method. When the bridge deck is a steel bridge deck with a low steel content (such as severely corroded areas, non-ferrous repair areas, or areas with a thick concrete cover), the adsorption force of the permanent magnet module may not be sufficient to ensure stable robot movement. Therefore, the system is designed with an auxiliary wind-powered pressurization system to provide non-contact additional adsorption force in areas with weak magnetic force. In this case, the multi-system collaborative working mode includes the following steps:

[0121] S51: When the steel ratio detection result is not higher than the preset steel ratio threshold, the current steel bridge deck area is determined to be a low steel ratio area;

[0122] S52: Activate the wind pressure-assisted adsorption walking mode, where an additional adsorption force is generated by a centrifugal fan;

[0123] S53: Activate the negative pressure adsorption walking mode, forming a distributed negative pressure adsorption through several suction cups;

[0124] S54: Maintain the magnetic-assisted adsorption walking mode so that the permanent magnet module continues to output magnetic adsorption force;

[0125] S55: Based on the synergistic output results of wind pressure adsorption force, negative pressure adsorption force and magnetic adsorption force, the robot is controlled to stably pass through the low steel ratio area and then return to the normal magnetic walking mode or other matching normal walking mode.

[0126] The magnetically assisted adsorption unit includes multiple magnetic wheels, each containing a magnetic adsorption module. Each magnetic wheel is rotatably mounted relative to its corresponding wheel. Preferably, the magnetic adsorption module is mounted inside the wheel via bearings, and a magnetic adjustment drive unit rotates the magnetic adsorption module via a coupling to adjust the magnetic adsorption direction. See the vehicle body structure diagram for details. Figure 8 For the internal structure of the magnetic wheel, please refer to [link / reference]. Figure 9 .

[0127] In this embodiment, when the mobile carrier travels normally along the bridge surface, the magnetic module can automatically adjust its direction according to the undulations of the wall, keeping the adsorption direction pointing towards the bridge surface as much as possible, thereby providing a stable climbing force for the mobile carrier. The states of the vehicle traveling on a vertical wall and on a right-angled wall can be seen respectively. Figure 10 (a) and Figure 10 (b) Through the rotatable magnetic module, the mobile carrier can still maintain its adsorption and obstacle-crossing ability in undulating or corner areas of the bridge surface.

[0128] Preferably, in step S52, the wind pressurization adopts a staged start-up strategy:

[0129] First-level start: When the steel ratio is at a medium to low level, the magnetic adsorption is slightly insufficient but has not yet affected driving stability. The centrifugal fan runs at a low speed to provide basic additional adsorption force to assist the magnetic system in working.

[0130] Level 2 Start-up: When the steel ratio remains low, or when the robot needs to cross a large non-ferrous area, the fan speed is increased to a medium-high setting to generate a strong airflow field, ensuring that the total adsorption force is always maintained above the safe threshold.

[0131] Level 3 Start-up: In extreme cases, such as continuous large areas with low steel ratios or sudden magnetic attenuation, the fan operates at full power to provide maximum additional adsorption force, while the robot slows down to ensure safety.

[0132] In this embodiment, the selection of magnetic materials can be combined with... Figure 11The magnetization and permeability curves are used to determine the magnetic properties. To balance high attraction force and fast dynamic response, the magnetic material should ideally possess high resistivity, high Curie temperature, high initial permeability, high maximum permeability, high saturation magnetic induction, and low coercivity and remanent magnetic induction. Preferably, the magnetic attraction module uses neodymium iron boron (N35) permanent magnet material to adapt to the crawling requirements of bridge surfaces with different curvatures and shallow reinforced concrete bridge surfaces.

[0133] In this embodiment, the control module controls the magnetically assisted adsorption unit as the main adsorption unit when the mobile carrier is in motion, and controls the negative pressure adsorption unit as the main adsorption unit when the mobile carrier is in a fixed-point detection, hovering operation, or long-term residence state. When an increase in local curvature of the bridge surface, a decrease in material magnetic permeability, a decrease in steel content, or an air leakage in the suction cup chamber exceeding a threshold is detected, the control module activates main-auxiliary coordinated compensation control, enabling magnetic adsorption, negative pressure adsorption, and wind-assisted adsorption to switch or work in coordination according to the current working conditions. The travel process of the iron bridge can be combined with... Figure 3 and Figure 12 understand.

[0134] Example 5

[0135] This embodiment, based on Embodiment 1, further explains the structure and recognition process of the disease identification and detection module. The disease identification and detection module includes a front-end rapid identification unit and a near-field detection unit. The front-end rapid identification unit performs grayscale conversion, filtering, and threshold segmentation processing on the acquired images to identify obstacle or disease candidate areas and controls the moving vehicle to approach the target area. The front-end rapid identification camera can be installed at the front of the vehicle body; its installation location is described in [reference needed]. Figure 8 The surface crack identification steps are described in [reference needed]. Figure 13 .

[0136] In this embodiment, the front-end rapid recognition unit preferably performs grayscale processing on the acquired image first, followed by mean filtering or wavelet denoising to reduce the impact of environmental noise. Then, it extracts the edges of defects and the contours of obstacles through threshold segmentation and binarization. This method enables rapid extraction of candidate defect areas while ensuring processing speed. For bridge surface areas with attached debris, dust, or small obstacles, a rolling obstacle-clearing axle can be installed at the front of the vehicle. When a target obstacle is detected, the motor is activated for simple cleaning; its structure is described in [reference needed]. Figure 14 .

[0137] In this embodiment, the disease identification and detection module further includes a supplementary lighting component. The supplementary lighting component includes a light sensor and an adjustable light source. The control module adjusts the luminous intensity of the adjustable light source according to the ambient illuminance to improve the contrast and stability of the disease image and reduce interference from shadows, reflections, or low-light environments on the identification results. See the supplementary lighting adjustment process below. Figure 15 Preferably, the adjustable light source employs a ring-shaped light-emitting array to reduce the shadowing effect caused by the probe body obstructing the light. More preferably, red light illumination can be used to enhance the signal at the crack edge and suppress local reflections. The image acquisition device may also refer to... Figure 16 The three-channel and single-channel acquisition methods shown demonstrate how to select an appropriate image acquisition strategy based on the object being identified and the ambient lighting conditions.

[0138] In this embodiment, the close-range detection unit is used to identify defects in the target area image and output at least one defect category information, including cracks, peeling, voids, corrosion, bulging, or deformation. For crack defects, they can be further distinguished into transverse cracks, longitudinal cracks, blocky cracks, and network cracks. During the close-range fine inspection process, traditional feature recognition can be performed using grayscale features, texture features, and local binary pattern features, or a deep learning model can be combined for category determination.

[0139] Furthermore, the front-end rapid identification unit can automatically approximate the candidate region through the target ranging module. For the principle of the ranging module, please refer to [link / reference needed]. Figure 17 After receiving the pixel size and center position data of the target in the image, the control module adjusts the vehicle's forward, backward, and steering movements to keep the area under test within a range suitable for close-range precision inspection, thereby completing the automatic tracking and positioning of the target defect area.

[0140] Example 6

[0141] This embodiment further describes the positioning and mapping module and the bridge defect detection and positioning method based on any one of embodiments 1 to 5. The positioning and mapping module is used to acquire path information and record keyframe pose information during the movement of the mobile carrier along the bridge surface. Preferably, the positioning and mapping module is implemented based on visual positioning data and a synchronous positioning and mapping algorithm. It can use a forward-looking camera or an airborne camera to acquire continuous image sequences and employ a synchronous positioning and mapping algorithm to depict the movement path of the mobile carrier on the bridge surface. Figure 18 This illustrates the relationship between object distance and camera imaging. Figure 19 The algorithm flow of the disease location method is shown.

[0142] In this embodiment, the near-field detection unit can be equipped with a micro-airborne computing platform such as a Raspberry Pi to perform edge computing tasks. See the Raspberry Pi hardware system section. Figure 20To balance detection accuracy and real-time performance, a lightweight deep learning recognition network can be constructed, for example, using... Figure 21 The lightweight detection network architecture shown performs disease feature recognition and category determination at the edge. Pre-trained weights can be incorporated into the recognition phase to improve model convergence speed and recognition accuracy, and a classification network can be used to further analyze suspected disease types.

[0143] In this embodiment, a method for detecting and locating bridge defects based on the above-mentioned bridge defect detection and location system includes the following steps.

[0144] S1: Control the moving carrier to travel along the bridge surface and collect the material state, adsorption state and image data of the bridge surface.

[0145] S2: Based on the material state, adsorption state, and operation state, perform main / auxiliary switching or collaborative compensation control between magnetic adsorption, negative pressure adsorption, and wind-assisted adsorption.

[0146] S3: Identify candidate areas of defects in bridge surface images and control the moving vehicle to approach the target area.

[0147] S4: Perform close-range disease identification on the target area to obtain disease category information.

[0148] S5: Establish bridge surface path information and record key frame pose information during the journey.

[0149] S6: Match the defect identification results with the keyframe pose information, and obtain the location result of the defect on the bridge surface through coordinate transformation. Steps S5 and S6 can be combined. Figure 18 and Figure 19 The process shown is implemented.

[0150] To improve the adaptability of the defect identification module in actual bridge environments, this embodiment can also establish a defect image dataset and perform annotation and enhancement on the dataset. The dataset establishment process is as follows:

[0151] Each photo in the dataset was labeled using a labeling tool. The labels and details for each type of support defect are shown below. Figure 23 Most defects in rubber bearings are rectangular; therefore, in this dataset, rectangular boxes are used to label various defects. The dataset labeling process consists of the following three steps:

[0152] ① Use rectangular frames to mark various defects on the supports;

[0153] ② To improve the quality of marking, check and fine-tune after marking is completed to avoid marking errors and randomness in the selection range;

[0154] ③ The labeled data is in XML format. Python is used to convert it into a TXT file format suitable for YOLO. Finally, the dataset is randomly divided.

[0155] The label names and number of the dataset after labeling are shown in Table 3.1. It should be noted that support defects often occur simultaneously, so each image to be tested may contain several defects of the same type or several different types.

[0156] Table 1 Dataset Label Details

[0157] See some disease image data Figure 22 For a label display, please refer to Table 1 and... Figure 23 .

[0158] To avoid overfitting, image enhancement was used to enhance the collected disease images. This enhancement employed one or a combination of methods, including horizontal mirror flipping, Gaussian blur, mean blur, image convolution, brightness and contrast adjustment, grayscale processing, and noise addition, to simulate data under different lighting and interference conditions. After enhancement, the images were manually screened to remove those with poor simulation results, resulting in 1614 valid images. The dataset is... The model is divided into training and validation sets. The test set uses photographs of the supports taken during field testing to evaluate the model's generalization ability. The training set is used to train the weight parameters of the YOLO v8 detection model, the validation set is used to adjust the model parameters to obtain the optimal model, and the test set is used to predict and evaluate the final model. For image enhancement effects, see [link to image enhancement effects]. Figure 24 By performing enhancement operations such as flipping, blurring, brightness adjustment, contrast enhancement, grayscale processing, and noise superposition on the original image, the generalization ability of the recognition model to bridge defect images under different lighting, angles, and surface conditions can be improved, thereby enhancing the recognition stability in field applications.

[0159] In the above embodiments, the composite adsorption walking module, the defect identification and detection module, and the positioning and mapping module can be centrally installed on the same mobile carrier, or they can be modularly adjusted according to the bridge structure and the requirements of the detection task. For the bridge bottom surface, bridge web, pier sidewalls, and local corner areas, the control module can dynamically select a single mode or a combined mode of magnetic adsorption, negative pressure adsorption, and wind-assisted adsorption based on the surface material, local curvature, adsorption state, and identification task, and output the spatial location of bridge surface defects by combining visual recognition and positioning and mapping results.

[0160] Example 7

[0161] This embodiment, based on Embodiments 5 and 6, further explains the specific calculation process of the disease identification and detection method. In this embodiment, the disease identification and detection method uses an improved VGG16 identification network as the classification backbone, and combines it with... Figures 22 to 24 The dataset construction, labeling, and image enhancement processes shown herein are used to identify and classify bridge surface defects. Unlike the direct single-input classification of raw images using standard VGG16, this embodiment introduces a fusion representation of grayscale structure information, texture structure information, and local crack saliency information on the input side. It also introduces salient region weight modulation during feature extraction and enhanced consistency constraints during training, thereby improving the recognition stability against complex backgrounds such as fine cracks, weakly textured cracks, low-light cracks, and defects accompanied by corrosion, bulging, and peeling. The defect categories can include aging cracking, excessive shear deformation, debris blockage, voids, steel plate corrosion, and bulging, and can be expanded to include cracks, peeling, voids, corrosion, bulging, and deformation, depending on the actual bridge surface defect type.

[0162] The disease identification and detection method in this embodiment can be executed according to the following steps. In some embodiments, the disease identification and detection method is executed and implemented by a disease identification and detection module;

[0163] S71: Normalize the collected bridge damage images and perform wavelet denoising to obtain the denoised reconstructed image;

[0164] S72: Extract grayscale structure map, morphological gradient map and local binary pattern texture map from the denoised image, and construct a multi-channel fused input tensor;

[0165] S73: Generate a salient weight map of cracks based on morphological gradient information and local texture information, and perform salient weight guidance on the fused input tensor;

[0166] S74: Input the weighted fused image into the improved VGG16 network, which passes through convolutional layers, activation layers, pooling layers and fully connected layers in sequence, and outputs the predicted probability of each disease category.

[0167] S75: Construct classification loss, enhanced consistency loss, and significant guidance loss, and use a joint loss function to iteratively optimize the network parameters;

[0168] S76: Perform forward inference on the image of the bridge under test and output the defect category with the highest probability as the final identification result.

[0169] In this embodiment, a disease identification image dataset is first established. Images of diseases on the bridge's underside, elevation, sides of bridge supports, and other areas to be tested are collected and labeled according to preset disease categories. Preferably, the disease areas are first marked using bounding boxes, and then the labeling results are checked and fine-tuned to reduce labeling bias. After labeling, the dataset is divided into a training set, a validation set, and a test set. To improve the model's adaptability to different lighting conditions, pollution levels, and shooting states, enhancement processing is performed on the original images. Enhancement methods include at least one or more of the following: horizontal flipping, Gaussian blur, mean blur, brightness adjustment, contrast adjustment, grayscale processing, and noise addition. The dataset establishment process is detailed in Table 1. Figure 22 See tag examples. Figure 23 See image enhancement effects. Figure 24 .

[0170] In step S71, the acquired virus image is normalized before being input into the network. Let the input virus image be... ,in and Let each represent the pixel coordinates of the image. Then the normalized image... for:

[0171] ;

[0172] After normalization, wavelet denoising is performed on the image to reduce interference from dust, shadows, noise, and surface contaminants in the field detection environment on virus edge features. The normalized image is then processed. Wavelet decomposition is performed to obtain the low-frequency approximate components. and high-frequency detail components of each layer The reconstructed image after denoising for:

[0173] ;

[0174] in, The wavelet decomposition level is denoted as . For the first The threshold shrinkage result of the high-frequency components of the layer. Preferably, the threshold shrinkage function is a soft threshold function, the expression of which is:

[0175] ;

[0176] in, For the first Layer wavelet thresholding. Through the above processing, high-frequency noise can be suppressed while preserving crack edges and local virus texture, thus improving the stability of subsequent feature extraction.

[0177] After obtaining the denoised image Subsequently, to enhance the ability of standard VGG16 to represent fine cracks and weak texture defects, this embodiment does not directly feed a single color image into the network. Instead, in step S72, a multi-feature fusion input consisting of a grayscale structure map, a morphological gradient map, and a local binary pattern texture map is constructed. Morphological processing is then performed on the denoised image to obtain the morphological gradient map. :

[0178] ;

[0179] in, This represents the expansion operation. This represents the erosion operation. This represents the structural element. The morphological gradient map is primarily used to highlight areas of abrupt gray-level changes in crack edges, peeling edges, and other disease contours. Simultaneously, local binary pattern textures are calculated on the denoised image. Its expression is:

[0180] ;

[0181] in, Indicates the grayscale value of the center pixel. Indicates the first grayscale values ​​of neighboring pixels Indicates the number of neighboring sampling points. For symbolic functions:

[0182] ;

[0183] Local binary mode features can enhance the model's ability to express local texture differences, and are especially suitable for areas with large variations in surface roughness, such as cracks, bulges, and rust.

[0184] Furthermore, the grayscale structure diagram Morphological gradient map and local binary pattern texture map Stack them along the channel direction to form a multi-feature fusion input tensor. :

[0185] ;

[0186] in, This indicates a channel splicing operation. In this way, the input to the network is no longer the original image along a single path, but a fused representation that simultaneously contains grayscale structure information, edge variation information, and local texture information, thereby enhancing the model's ability to represent complex bridge damage images.

[0187] To further improve the network's ability to focus on salient crack regions, this embodiment constructs a crack salient weight map based on the morphological gradient map and local texture map in step S73. Its expression is:

[0188] ;

[0189] in, and Let be the weighting coefficients, and satisfy:

[0190] ;

[0191] Preferably, Pick , Pick Since the morphological gradient map responds more directly to crack edges, it is given a higher weight; the local binary pattern map is mainly sensitive to texture changes, so it is used as an auxiliary term to jointly construct salient guiding information. Then, the salient weight map is applied to the fused input tensor to obtain the weighted input tensor. :

[0192] ;

[0193] in, This indicates element-wise multiplication. This significant guidance method enhances the network's responsiveness to crack edges, broken edges, and areas with anomalous texture clusters, while reducing the interference of irrelevant background areas on classification results.

[0194] In terms of network structure, this embodiment still adopts the VGG16 multi-layer convolutional and pooling backbone structure, but in step S74, the weighted fused input tensor is... As network input. Let the first... The input feature map of each convolutional layer is The convolution kernel is , bias is Then the first layer Output feature map of each channel for:

[0195] ;

[0196] in, This represents the convolution operation. This represents a nonlinear activation function. Preferably, the activation function is a modified linear unit function, whose expression is:

[0197] ;

[0198] After each convolutional layer, max pooling is performed on the feature map. Let the pooling window region be... The pooling result for:

[0199] ;

[0200] After multiple convolutional and pooling processes, the network obtains the roof layer features, which are then input into the fully connected layer for classification output. Let the final fully connected layer output vector be... ,in, Let be the number of disease categories. After processing by the Softmax function, the th... Predicted probability of similar diseases for:

[0201] ;

[0202] During training, to ensure classification accuracy, the classification cross-entropy loss function is first constructed in step S75. Let the true label be... The predicted probability is Then classification loss for:

[0203] ;

[0204] Building upon this, to improve the network's ability to consistently distinguish between samples before and after image enhancement, this embodiment introduces an enhancement consistency constraint loss. Let the output probability vector of the original image after inputting it into the network be... The output probability vector after enhancing the image input network is: This would increase the consistency loss. for:

[0205] ;

[0206] This consistency constraint ensures that the network maintains relatively stable classification outputs for the same disease image even when faced with changes in brightness, noise interference, blurring disturbances, and contrast variations. To further concentrate the network's high-response regions on cracks or areas with significant disease, this embodiment also introduces a significant guiding loss. Let the feature response map of the last convolutional layer of the network be... This significantly guides the loss. for:

[0207] ;

[0208] in, This is the aforementioned crack saliency weight map. This loss term further constrains the network attention to crack edges, lesion edges, and texture anomaly regions, improving the consistency between the model's feature responses and the actual lesion regions. A joint total loss function is constructed by combining the classification loss, consistency enhancement loss, and saliency guidance loss. :

[0209] ;

[0210] in, and These are the loss weight coefficients, used to adjust the proportion of each loss term in the overall training objective. During training, gradient descent is used to adjust the network parameters. Updated, number The parameter update formula for the next iteration is:

[0211] Where η is the learning rate.

[0212] After the network training is completed, in step S76, the images of bridge defects to be tested are input into the improved VGG16 recognition network to obtain the predicted probability of each defect category, and the category with the highest predicted probability is taken as the final recognition result.

[0213] ;

[0214] This enables automatic classification and recognition of bridge surface defects. Because this embodiment adds wavelet denoising, grayscale structure and texture feature fusion, weighted guidance for significant crack regions, and an enhanced consistency constraint training mechanism to the original VGG16 recognition framework, it can more effectively improve the stability and accuracy of identifying fine cracks, low-contrast cracks, and defects in complex backgrounds compared to directly using standard VGG16 to classify the original image.

[0215] In this embodiment, the defect identification and detection method can be executed according to the following process: First, the collected bridge defect images are normalized and denoised using wavelet denoising to obtain a denoised reconstructed image; then, grayscale structure map, morphological gradient map, and local binary pattern texture map are extracted from the denoised image, and a multi-channel fusion input tensor is constructed; then, a crack saliency weight map is generated based on the morphological gradient information and local texture information, and saliency weight guidance is performed on the fusion input tensor; then, the weighted fusion image is input into an improved VGG16 network, which passes through convolutional layers, activation layers, pooling layers, and fully connected layers in sequence to output the predicted probability of each defect category; then, classification loss, enhanced consistency loss, and saliency guidance loss are constructed, and the network parameters are iteratively optimized using a joint loss function; finally, forward inference is performed on the bridge image to be tested, and the defect category with the highest probability is output as the final identification result.

[0216] Compared with the existing method of directly classifying and recognizing a single original image using the standard VGG16, this embodiment improves the network's ability to distinguish fine cracks, low-contrast cracks, and complex background defects by introducing wavelet denoising, morphological gradient enhancement, local binary pattern texture representation, crack saliency weight guidance, and enhanced consistency constraint training mechanism, thereby improving the stability and accuracy of bridge defect recognition results.

[0217] Test Example 1

[0218] To compare the defect identification and detection method used in Embodiment 7 of this application with the traditional VGG16 method in terms of efficiency and misjudgment in bridge surface defect detection, Test Example 1 was conducted using a standard defect sample library for comparative identification. Six types of bridge surface defects were selected as test subjects: cracks, peeling, voids, corrosion, bulges, and deformation. Each defect type corresponded to a set of pre-prepared and numbered standard bridge test specimens or standard defect areas. Before collection, a fixed correspondence was established between the specimen number and the defect type for each set of samples. After image acquisition, the defect type information corresponding to the specimen number was directly inherited, thus avoiding reliance on subsequent manual annotation of each image. 500 images were collected for each defect type, resulting in a total of 3000 images. The training set consisted of 300 images per type (1800 images total), the validation set consisted of 100 images per type (600 images total), and the test set consisted of 100 images per type (600 images total). During testing, 600 test images were input into both the traditional VGG16 model and the model from Example 7. The true class and the model output class for each image were recorded, and statistics were performed with the true class as the vertical dimension and the predicted class as the horizontal dimension.

[0219] Taking the test results of Example 7 as an example, there were 100 test samples of the crack type. Among them, 93 samples were identified as cracks, 2 samples were identified as peeling, 1 sample was identified as voiding, 2 samples were identified as corrosion, 1 sample was identified as bulging, and 1 sample was identified as deformation. Among the peeling test samples, 91 samples were correctly identified; 90 samples were identified as voiding; 93 samples were identified as corrosion; 91 samples were identified as bulging; and 91 samples were identified as deformation.

[0220] Based on this, a difference analysis experiment was conducted. First, the 6x6 category discrimination distribution matrices of the traditional VGG16 model and the Example 7 model were obtained on the same set of 600 test images. Then, the recognition ratio at the corresponding position in Example 7 was subtracted from the recognition ratio of the traditional VGG16 model at the same position to obtain the recognition improvement distribution matrix. Since each class in the test set has 100 images, each value in the matrix can be directly represented as the difference in the number of samples or the percentage difference. Specifically, the main diagonal position of the crack class improved from 84 to 93, an improvement of 9; the peeling class improved from 81 to 91, an improvement of 10; the voiding class improved from 80 to 90, an improvement of 10; the corrosion class improved from 85 to 93, an improvement of 8; the bulge class improved from 81 to 91, an improvement of 10; and the deformation class improved from 82 to 91, an improvement of 9. Correspondingly, the proportion of misjudgments at non-main diagonal positions generally decreased. For example, the proportion of cracks misjudged as peeling decreased from 5 to 2, peeling was misjudged as cracks from 6 to 2, and voids were misjudged as peeling from 6 to 3.

[0221] Depend on Figure 25 It can be seen that, under the same dataset, the same training rounds, and the same testing conditions, Example 7 significantly improves the recognition results of various bridge defects compared to the traditional VGG16. Simultaneously, the misclassification values ​​corresponding to non-primary discriminant regions are reduced overall, with most misclassified channels decreasing to the range of 1-3. This indicates that Example 7 not only improves the correct recognition capability but also reduces the mutual confusion between different defects. These results demonstrate that this application does not achieve performance changes solely by increasing network complexity. Instead, it achieves this by introducing a fusion representation of grayscale structure information, morphological gradient information, and local texture information at the input, combined with salient region guidance and consistency constraints. This allows the network to more fully distinguish discriminative image features such as crack edges, peeling boundaries, rust textures, and bulge contours before classification, thus exhibiting relatively stable recognition gains across all defect categories.

[0222] Test Example 2

[0223] This test case uses 600 test images from Test Case 1, with 100 images for each disease category. After training, both the traditional VGG16 and Example 7 models are switched to inference mode. High-level feature vectors for each image are extracted from before the last fully connected classification layer, with a feature dimension of 4096. Then, the tSNE dimensionality reduction method with a fixed random seed of 42 is used to compress the 4096-dimensional features into a two-dimensional plane, where perplexity is 30, learning rate is 200, and the number of iterations is 1000. After 2D dimensionality reduction, feature points are distinguished and drawn according to the actual disease category, thus forming... Figure 26The feature distribution results are shown. To establish a correlation between the illustrated results and classification performance, the average intra-class distance and inter-class center distance can be further calculated. Example 7 provides a more concentrated high-level feature representation of the disease image, and the distribution boundaries between different categories are clearer. The average Euclidean distance between category centers is greater than that of the traditional VGG16, while the average intra-class dispersion is less than that of the traditional VGG16. Although the feature points corresponding to the traditional VGG16 in the figure have formed a certain cluster, there are still obvious overlapping areas between categories, especially in the middle feature area, where different disease samples are still easy to approach or even partially overlap. It can be seen that the distribution of feature points of each category obtained by the disease identification and detection method in steps S71-S76 of Example 7 of this application is more concentrated in two-dimensional space, which indicates that the high-level semantic features extracted in Example 7 are more conducive to the separation of different diseases. This test case does not directly re-measure the bridge structure, but rather achieves repeated testing and comparison of the model's representation ability by reading the high-level features inside the model and performing fixed-parameter dimensionality reduction visualization.

[0224] Specifically, the recognition rate for cracks has been improved to 9, peeling to 10, voids to 10, corrosion to 8, bulges to 10, and deformation to 9.

[0225] Combined Figure 26 It can be further observed that the feature points corresponding to Example 7 exhibit a more compact intra-class distribution and a clearer inter-class separation trend. This indicates that after multi-feature fusion input and significant weight modulation, the disease representation extracted by the network is closer to the structural differences of various diseases themselves, rather than being excessively affected by background grayscale, surface noise, or local illumination changes. In other words, Example 7 has further separated the previously easily confused disease classes in the feature space, which is consistent with... Figure 25 The improvements in the identification of various diseases and the reduction in misjudgments are mutually reinforcing.

[0226] comprehensive Figure 25 and Figure 26It can be explained that the improvement in Example 7 is not merely a numerical change in the final identification result, but rather an improvement in the representation of different defects from the feature formation stage. The defect identification and detection method of this application improves the identification and detection of various bridge defect categories by 8 to 10%, indicating that the method is more accurate in judging defects such as cracks, peeling, voids, corrosion, bulges, and deformation. Furthermore, the feature distribution of each category of feature points obtained by the defect identification and detection method of this application in two-dimensional space is more concentrated within categories and more distinct between categories compared to the relatively loose and locally overlapping feature point distribution in two-dimensional space detected by the traditional VGG16 method. This demonstrates that the defect identification method proposed in this application can more effectively extract key discriminative information from bridge defect images. Therefore, this application not only improves the identification accuracy in bridge defect identification tasks but also enhances the ability to distinguish between different defects, adapting to the detection of bridge deck defects of different materials while being more accurate in detecting defects in complex bridge deck environments.

[0227] Although the invention has been described with reference to preferred embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, the technical features mentioned in the various embodiments can be combined in any manner as long as there is no structural conflict. The invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for detecting, locating, and controlling the adsorption and movement of bridge defects, characterized in that, A bridge defect detection and positioning system applicable to magnetically assisted adsorption walking, negative pressure adsorption walking, and wind pressure assisted adsorption walking modes, the method comprising: The surface material of the bridge surface that the robot is currently traveling on is identified to determine whether the bridge surface is made of stone or steel. When the bridge surface is determined to be stone, the system detects whether there are gaps in the corresponding travel area. If there are no gaps, the control system enters the normal negative pressure travel mode, which uses air pressure sensors on several suction cups to detect the air pressure in the cavity of each suction cup in real time, and performs closed-loop control of negative pressure adsorption based on the detection results. If there are gaps, the control system enters the multi-system collaborative work mode, which simultaneously activates the wind pressure-assisted adsorption travel mode and the magnetic force-assisted adsorption travel mode, and works in conjunction with the negative pressure adsorption travel mode to pass through the stone bridge surface area with gaps. When the bridge surface is determined to be steel, the steel ratio of the corresponding travel area is detected. If the steel ratio is higher than a preset threshold, the control system enters the normal magnetic walking mode, and the robot moves by providing vertical adsorption force through the permanent magnet module. If the steel ratio is lower than or equal to the preset threshold, the control system enters the multi-system collaborative working mode, and simultaneously starts the wind pressure-assisted adsorption walking mode and the negative pressure adsorption walking mode, and works in conjunction with the magnetic adsorption walking mode to pass through areas with low steel ratio. Regardless of whether the bridge deck is made of stone or steel, as long as the system enters the multi-system collaborative working mode, the wind pressure-assisted adsorption walking mode is activated to provide additional adsorption force and improve the robot's stable walking ability in abnormal bridge deck areas. When walking in normal walking and / or multi-system collaborative working mode, real-time image information of the bridge deck is collected, and then the bridge deck is identified by the defect identification and detection method to determine whether the bridge deck has defects and the type of defects.

2. The adsorption-walking control method according to claim 1, characterized in that, When the bridge deck is a seamless stone bridge deck, the normal negative pressure walking mode includes the following steps: S21: Acquire real-time air pressure data collected by air pressure sensors corresponding to several suction cups; S22: Input the real-time air pressure data into the control unit and compare it with the target negative pressure value to obtain the air pressure deviation corresponding to each suction cup; S23: Based on the air pressure deviation, perform PID closed-loop control and output the corresponding air pump drive signal to adjust the negative pressure in each suction cup cavity; S24: During the robot's movement, the negative pressure of each suction cup is kept within a preset stable range to maintain normal negative pressure adsorption and walking state.

3. The adsorption-walking control method according to claim 1, characterized in that, When the bridge deck is a stone bridge deck with gaps, the multi-system collaborative working mode includes the following steps: S31: Based on the gap detection results, the current stone bridge surface area is determined to be an abnormal travel area; S32: Activate the wind pressure-assisted adsorption walking mode, where an additional adsorption force is generated by a centrifugal fan; S33: Activate the magnetic-assisted adsorption walking mode, with the permanent magnet module providing vertical auxiliary adsorption force; S34: Maintain negative pressure adsorption walking mode and dynamically adjust the negative pressure adsorption capacity based on the real-time air pressure detection results of several suction cups; S35: After the robot passes through the abnormal travel area, it resumes the corresponding normal walking mode according to the coordinated output results of wind pressure-assisted adsorption force, magnetic force-assisted adsorption force and negative pressure adsorption force.

4. The adsorption-walking control method according to claim 1, characterized in that, When the bridge deck is made of steel with a high steel content, the normal magnetic travel mode includes the following steps: S41: Detect the steel ratio of the steel bridge deck and compare the detection result with the preset steel ratio threshold; S42: When the steel ratio is higher than the preset steel ratio threshold, control the permanent magnet module to enter the normal adsorption working state to provide magnetic adsorption force perpendicular to the bridge surface; S43: During the robot's movement, adjust the direction or position of magnetic adsorption according to the local undulations of the bridge surface to maintain stable magnetic adsorption; S44: Utilizes magnetic adsorption as the primary adsorption method to maintain the robot's normal movement.

5. The adsorption-walking control method according to claim 1, characterized in that, When the bridge deck is a steel bridge deck with a low steel content, the multi-system collaborative working mode includes the following steps: S51: When the steel ratio detection result is not higher than the preset steel ratio threshold, the current steel bridge deck area is determined to be a low steel ratio area; S52: Activate the wind pressure-assisted adsorption walking mode, where an additional adsorption force is generated by a centrifugal fan; S53: Activate the negative pressure adsorption walking mode, forming a distributed negative pressure adsorption through several suction cups; S54: Maintain the magnetic-assisted adsorption walking mode so that the permanent magnet module continues to output magnetic adsorption force; S55: Based on the synergistic output results of wind pressure adsorption force, negative pressure adsorption force and magnetic adsorption force, the robot is controlled to stably pass through the low steel ratio area and then return to the normal magnetic walking mode or other matching normal walking mode.

6. The adsorption-walking control method according to claim 2, characterized in that, The PID closed-loop control in step S23 includes: performing proportional, integral, and derivative adjustments based on the deviation between the target negative pressure value and the real-time air pressure value to generate a PWM control signal, and adjusting the air pump output power corresponding to each suction cup according to the PWM control signal, so that the negative pressure enters a fine maintenance state when it is close to the target negative pressure value, and enters a rapid compensation state when it deviates from the target negative pressure value.

7. The adsorption-walking control method according to claim 1, characterized in that, In the multi-system collaborative working mode, the control unit performs graded control of the wind pressure-assisted adsorption walking mode according to the degree of bridge deck anomaly, including basic auxiliary mode, medium auxiliary mode and enhanced auxiliary mode; wherein, the degree of bridge deck anomaly is determined at least according to one or more of the following: gap size, number of gaps, steel ratio deviation, and adsorption stability index.

8. The adsorption-walking control method according to claim 3, characterized in that, The steps for maintaining the negative pressure adsorption walking mode also include: The real-time air pressure of several suction cups is compared and analyzed to determine the deviation between the air pressure of each suction cup and the average air pressure. Suction cups with deviations exceeding a preset threshold are identified as abnormal suction cups. Adjust the target negative pressure value or suction power of the remaining suction cups according to the number and distribution of abnormal suction cups in order to maintain the overall negative pressure adsorption capacity basically stable.

9. The adsorption-walking control method according to claim 1, characterized in that, The disease identification and detection method includes the following steps: S71: Normalize the collected bridge damage images and perform wavelet denoising to obtain the denoised reconstructed image; S72: Extract grayscale structure map, morphological gradient map and local binary pattern texture map from the denoised image, and construct a multi-channel fused input tensor; S73: Generate a salient weight map of cracks based on morphological gradient information and local texture information, and perform salient weight guidance on the fused input tensor; S74: Input the weighted fused image into the improved VGG16 network, which passes through convolutional layers, activation layers, pooling layers and fully connected layers in sequence, and outputs the predicted probability of each disease category. S75: Construct classification loss, enhanced consistency loss, and significant guidance loss, and use a joint loss function to iteratively optimize the network parameters; S76: Perform forward inference on the image of the bridge under test and output the defect category with the highest probability as the final identification result.

10. A bridge defect detection, positioning, and adsorption-based walking control system, characterized in that, It includes a bridge deck property detection module, a composite adsorption walking module, a defect identification and detection module, a positioning and mapping module, and a control module; the bridge deck property detection module includes a gap detection module and a steel ratio detection module; the composite adsorption walking module includes a magnetically assisted adsorption unit, a negative pressure adsorption unit, and a wind pressure assisted adsorption unit; The bridge surface attribute detection module is used to identify the surface material of the bridge surface that the robot is currently traveling on, so as to determine whether the bridge surface is a stone bridge surface or a steel bridge surface. The gap detection module is used to detect whether there are gaps in the corresponding travel area of ​​the stone bridge surface when the bridge surface is a stone bridge surface; The steel ratio detection module is used to detect the steel ratio of the corresponding travel area of ​​the steel bridge deck when the bridge deck is made of steel. The magnetically assisted adsorption unit includes a permanent magnet module for providing vertical adsorption force; The negative pressure adsorption unit includes a plurality of suction cups and air pressure sensors respectively arranged corresponding to the plurality of suction cups; The wind pressure-assisted adsorption unit includes a centrifugal fan for generating additional adsorption force; The defect identification and detection module is used to collect real-time image information of the bridge deck and then identify whether the bridge deck has defects and the type of defects through defect identification and detection methods. The positioning and mapping module is used to acquire path information of the bridge surface and record key frame pose information during the movement of the mobile carrier along the bridge surface, and match the defect identification result output by the defect identification and detection module with the corresponding key frame pose information, and obtain the positioning result of the defect on the bridge surface through coordinate transformation. The control module is connected to the bridge deck attribute detection module, the gap detection module, the steel ratio detection module, the composite adsorption walking module, the defect identification and detection module, and the positioning and mapping module, respectively, and is configured to execute the adsorption walking control method as described in any one of claims 1-9.

Images