Steel structure real-time detection and feedback system based on data prediction

By using permanent magnet adsorption tracks and improved YOLOv5 algorithm in hull detection, combined with ultrasonic non-destructive testing framework, the problem of detection carts being disengaged from the wall and data instability is solved, and high-precision steel structure detection and data safe transmission are achieved.

CN118961877APending Publication Date: 2024-11-15CHONGQING TIANYAN ENG QUALITY INSPECTION CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410976130.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-19
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The existing hull steel structure detection has problems such as overturning torque causing the detection car to break off the wall, low intelligence, unstable data source, and unencrypted data transmission.

Method used

The permanent magnet adsorption track and the improved YOLOv5 algorithm are adopted, combined with multi-scale feature fusion design and ultrasonic non-destructive detection framework to realize real-time detection and data encryption to ensure stable contact and high-precision detection of the detection car.

Benefits of technology

It enhances the adsorption force and stability of the detection trolley, improves the detection accuracy of cracks and weld defects at different scales, and ensures the effectiveness of detection information and the security of data transmission.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118961877A_ABST
    Figure CN118961877A_ABST
Patent Text Reader

Abstract

The invention provides a steel structure real-time detection and feedback system based on data prediction, and relates to the technical field of hull detection, the system is composed of a machine execution system, a detection transmission system, an information processing system and a detection feedback system, the machine execution system comprises a detection trolley body and a detection unit, a driving part at the bottom of a detection trolley body adopts a permanent magnet adsorption track, the permanent magnet adsorption track on one side of the robot forms a triangular chain transmission structure through front and rear groups of double-row driven chain wheels and a middle group of double-row driving chain wheels, and the double-row driving chain wheels are driven by a gear motor, so that the whole permanent magnet adsorption track is driven to move. According to the detection trolley, the permanent magnet adsorption crawler belt is adopted as a bottom driving part, double rows of driving chain wheels are driven through the gear motor, stable motion control is achieved, the stability and durability of the crawler belt are enhanced through a triangular chain transmission structure, and the detection trolley is suitable for movement in complex terrains and curved surface environments and has excellent curved surface adaptability.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of hull detection, and in particular to a real-time detection and feedback system for steel structures based on data prediction. Background Art

[0002] Ship hull steel structure inspection is a comprehensive and meticulous process involving multiple links and multiple inspection items. Through strict inspection, the quality and safety of the ship hull steel structure can be ensured, providing strong protection for the navigation of the ship, which focuses on welds, cracks and surface defects.

[0003] The prior art has the following technical problems when used:

[0004] Problem 1: Many existing ship hull steel structure inspections are carried out by sticking to the hull without any structural restrictions, which can easily cause the overturning moment to cause the inspection vehicle to detach from the wall. Secondly, the intelligence level is not high, resulting in an unstable data source for the inspection.

[0005] Problem 2: The collected information is not processed based on the data model. Without the support of data prediction, the detection process may only be based on traditional physical detection and manual judgment.

[0006] The second problem is that the transmission channel between the detector and the receiver is not encrypted, which can easily lead to data loss.

[0007] Therefore, a real-time detection and feedback system for steel structures based on data prediction is needed to solve the above problems. Summary of the invention

[0008] Technical issues solved

[0009] In view of the shortcomings of the prior art, the present invention provides a real-time detection and feedback system for steel structures based on data prediction, which solves the following problems:

[0010] 1. Ensure the maximum contact between the hull and the crawler, thereby enhancing the adsorption force and stability, and detect the adsorption force of the trolley in real time. By adjusting the working state of the adsorption structure (such as the yoke and the excitation coil), ensure that the trolley can still maintain stable contact with the wall when subjected to the overturning moment, prevent it from detaching from the wall, and ensure that the detection information is valid;

[0011] 2. By improving the YOLOv5 algorithm and adopting multi-scale feature fusion design, the model can more effectively extract and identify cracks of different scales, especially small cracks.

[0012] 3. The ultrasonic nondestructive testing framework designed in the present invention realizes the automation of weld detection, simplifies the detection process, and improves the detection efficiency. According to the position and shape of the ultrasonic reflection pulse, it can quickly determine whether there are defects in the weld and determine the position, size and shape of the defect. Through algorithm improvement and data enhancement, the model's detection accuracy of cracks and weld defects has been significantly improved. The model can handle targets of different scales and can also effectively detect small cracks and subtle weld defects.

[0013] 4. Added RSA public key encryption to encrypt the transmission channel between the detector and the receiver to prevent data loss.

[0014] Technical Solution

[0015] To achieve the above objectives, the present invention is implemented through the following technical solutions: a real-time detection and feedback system for steel structures based on data prediction, the system is composed of a machine execution system, a detection transmission system, an information processing system and a detection feedback system, wherein:

[0016] The machine execution system includes a detection trolley body and a detection unit. The bottom driving part of the detection trolley body adopts a permanent magnetic adsorption track. The permanent magnetic adsorption track on one side of the robot is composed of a front and rear double-row driven sprocket and a middle double-row driving sprocket to form a triangular chain transmission structure. The double-row driving sprocket is driven by a reduction motor, thereby driving the entire permanent magnetic adsorption track to move;

[0017] The track support part of the permanent magnetic adsorption track is equipped with a damping shock absorber, and the track is tightened to better fit the curved surface of the hull. One end of the permanent magnetic adsorption track is fixedly connected to the frame chassis, and the other end is hinged to one end of the spring damping shock absorber, and the other end of the spring damping shock absorber is hinged to the frame chassis to ensure maximum contact between the hull and the track.

[0018] Because the detection car is affected by gravity and is subjected to the overturning moment, the detection car has a tendency to overturn and break away from the wall at the contact point A. In order to overcome the situation where the robot is caused to break away from the wall due to the overturning moment, it is necessary to detect the adsorption force of the car itself in real time to overcome the condition that the detection car does not break away from the wall and overturn.

[0019] Preferably, a driving motor, a motor driving board and a relay are installed in the frame of the detection vehicle, the adsorption structure in the vehicle track is composed of a yoke and an excitation coil, the control chips of the detection vehicle are respectively an ESP8266 module and an MPU-6050 chip, and the detection vehicle receives instructions issued by the control platform through ESP8266, and converts the instructions received by ESP826 into electrical signals of the driving motor through L298N, so as to remotely drive and control the robot.

[0020] Preferably, the detection unit includes a crack inspection probe, a weld ultrasonic detection head and a surface defect detection probe.

[0021] Preferably, a magnetic beacon is pasted on the surface of the detection hull, and the front of the beacon is marked with a high-brightness color. The interior of the detection vehicle contains a CPU, a digital input and output module, a CANopen module, an RS485 communication module, and a Profibus communication module. The bottom side of the detection vehicle is externally connected to an obstacle sensor, which scans a semicircular area within 1m in front of it and detects obstacles in the driving direction. The obstacles here mainly include seaweed and marine biological attachments. If there is an obstacle in front of the vehicle, the vehicle will detour.

[0022] Preferably, the detection hull is equipped with a HuskyLens visual sensor, which is installed on the front of the detection vehicle frame to check the road signs pasted on the path in real time for identifying the guide route and correcting the vehicle body. The detection vehicle uses the visual sensor to perform object recognition, color recognition and label recognition on the pointer. The visual sensor is externally connected to a 2.0-inch IPS display (320×240 resolution), which can directly display the debugging process and recognition results. Interface: onboard UART / I2C interface, which can be connected to mainstream controllers such as Arduino, control board, micro:bit, Raspberry Pi, etc. Programming support: supports programming languages ​​such as C / C++ or Python for communication and interaction, and also supports graphical programming software such as Mind+, MakeCode, etc.

[0023] Preferably, the crack detection probe adopts the YOLOv5 network model algorithm. The scale of crack samples in the main hull structure external crack data set varies greatly, and there are many small cracks. The model trained on this data set is difficult to ensure the recognition accuracy of targets of each scale. By improving the backbone network of the YOLOv5 algorithm, the feature extraction capability is improved; at the same time, the FPN-PAN design is integrated to ensure the full integration of multi-scale features, especially to ensure that the features of small targets can be extracted; the input end uses the MOSAIC algorithm for data enhancement to increase the frequency of small targets. Therefore, the present invention adopts YOLOv5 as the basic model for the external crack detection task of the main hull structure. The information processing system receives the crack image collected by the crack detection probe through the detection transmission system, and organizes and constructs the external crack data set of the main hull structure. The training of the model is first pre-trained on the source data set, and the backbone and neck parts of the obtained pre-trained model are parameter-frozen, and then transfer learning training is performed on the external crack data set of the main hull structure to complete knowledge migration. Finally, a variety of optimization algorithms are introduced to improve the detection accuracy.

[0024] Preferably, after obtaining the pre-trained model, the parameters of the backbone and neck parts are frozen, 260 samples of the main hull structure external crack data set are randomly selected for training, and the rest are used for testing. The number of training iterations is 700 times, the iteration batch size is set to 32, the optimizer selects SGD, uses the quaternary data loader, the learning rate uses the cosine annealing algorithm, and the basic learning rate is set to 0.01. The present invention takes the deployment of the model on a hardware platform equipped with a 2080T graphics card as an example, tests the inference speed of the model, performs crack detection on 50 test samples, and the average inference time for each sample is 14.2ms, and the processing frame rate reaches 70 frames / s.

[0025] Preferably, there are four groups of weld ultrasonic detection heads, and the four groups of ultrasonic probes are located at different positions. The hardware facilities required for detecting the welds of the hull structure are calculated by a function generator, and the power of the detection head is amplified by a power amplifier. After collecting data, the collected results are displayed on a digital oscilloscope and dynamically updated through an information processing system.

[0026] Preferably, the weld detection step is to first construct an ultrasonic probe array, simulate the working state of the probe in the computer, construct the array according to the working state of each probe, set the probe drive circuit, the probe drive circuit includes multiple lines, each corresponding to a probe, the computer tracks in real time, records the working state of the probe, uses a digital display to receive signals, and displays the detection results on the LED screen.

[0027] Preferably, the surface defect detection probe collects images of the hull surface, and the information processing system creates a NEU data set, which contains 1,800 grayscale images, divided into five types of defects: scratches, folded heavy skin, ears, oxidized iron scale and pitting. The data set has two versions, NEU-CLS and NEU-DET, the former is used for classification, and the latter is used for target detection.

[0028] Beneficial Effects

[0029] The present invention provides a real-time detection and feedback system for steel structures based on data prediction. It has the following beneficial effects:

[0030] 1. The detection trolley of the present invention adopts a permanent magnetic adsorption track as the bottom driving part, and drives a double-row active sprocket through a reduction motor to achieve stable motion control. The triangular chain transmission structure enhances the stability and durability of the track, is suitable for movement in complex terrain and curved surface environments, and has excellent curved surface adaptability. The damping shock absorber installed at the track support part helps to tighten the track, better fit the hull surface, and improve the machine's mobility on surfaces with different curvatures. The spring damping shock absorber ensures maximum contact between the hull and the track by adjusting the angle between the track and the frame chassis, thereby enhancing the adsorption force and stability, and detecting the adsorption force of the trolley in real time. By adjusting the working state of the adsorption structure (such as the yoke and the excitation coil), it is ensured that the trolley can still maintain stable contact with the wall when subjected to an overturning moment, prevents it from detaching from the wall, and ensures that the detection information is valid.

[0031] 2. By improving the YOLOv5 algorithm and adopting a multi-scale feature fusion design, the model can more effectively extract and identify cracks of different scales, especially small cracks.

[0032] 3. The present invention uses MOSAIC algorithm for data enhancement to increase the frequency of small targets and improve the model's ability to identify small targets. Through transfer learning, the model parameters pre-trained by the source data set are used and then fine-tuned on the main hull structure external crack data set, which improves the generalization ability and detection accuracy of the model.

[0033] 4. The ultrasonic nondestructive testing framework designed in the present invention realizes the automation of weld detection, simplifies the detection process, and improves the detection efficiency. According to the position and shape of the ultrasonic reflection pulse, it can quickly determine whether there are defects in the weld and determine the position, size and shape of the defect. Through algorithm improvement and data enhancement, the model's detection accuracy of cracks and weld defects has been significantly improved. The model can handle targets of different scales and can also effectively detect small cracks and subtle weld defects. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a diagram of the detection system of the present invention;

[0035] Figure 2 This is a control framework diagram of the detection vehicle of the present invention;

[0036] Figure 3 It is the circuit diagram of the detection vehicle of the present invention;

[0037] Figure 4 It is the parameter diagram of the detection vehicle of the present invention;

[0038] Figure 5 It is a schematic diagram of the ultrasonic nondestructive testing principle of the present invention;

[0039] Figure 6 This is a schematic diagram of the weld detection principle of the present invention;

[0040] Figure 7 This is a weld detection control flow chart of the present invention;

[0041] Figure 8 It is a schematic diagram of a direction indicator of the present invention. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. Specific embodiment one:

[0044] like Figure 1-8 As shown in the figure, the real-time detection and feedback system of steel structure based on data prediction consists of a machine execution system, a detection transmission system, an information processing system and a detection feedback system, among which:

[0045] The machine execution system includes the detection trolley body and the detection unit. The bottom driving part of the detection trolley body adopts a permanent magnetic adsorption track. The permanent magnetic adsorption track on one side of the robot is composed of a front and rear double-row driven sprocket and a middle double-row driving sprocket to form a triangular chain transmission structure. The double-row driving sprocket is driven by a reduction motor, thereby driving the entire permanent magnetic adsorption track to move;

[0046] The track support part of the permanent magnetic adsorption track is equipped with a damping shock absorber, which can better fit the curved surface of the hull through track tensioning. One end of the permanent magnetic adsorption track is fixedly connected to the frame chassis, and the other end is hinged to one end of the spring damping shock absorber, and the other end of the spring damping shock absorber is hinged to the frame chassis to ensure maximum contact between the hull and the track.

[0047] Since the detection trolley is affected by gravity and is subjected to the overturning moment, the detection trolley will have a tendency to overturn from the wall at the contact point A. In order to overcome the situation where the robot is separated from the wall due to the overturning moment, the adsorption force of the trolley itself needs to be detected in real time to overcome it. The condition that the detection trolley does not overturn from the wall is:

[0048] ∑M A ≥0

[0049] By calculating the distance to point A, where the distance refers to calculating the sum of all moments about point A;

[0050] M A Represents the moment about point A; ∑MA ≥0 means that the sum of all moments about point A is greater than zero, which means that the system is unstable at point A or has a tendency to overturn in the clockwise direction;

[0051] The calculation formula for the system to reach equilibrium at point A is as follows:

[0052]

[0053] where F m1 is the force of the first adsorption unit, F N1 is the force of the first adsorption unit in the critical state, L is the distance from the adsorption unit to point A, G represents the weight of the object, H is the height of the object, ∑M A =0 means that the sum of all torques about point A is zero, and the system reaches equilibrium at point A;

[0054] When the first adsorption unit is in critical state, the following formula is obtained;

[0055] F N1 =0

[0056] The critical state is when the adsorption unit just leaves the surface of the object, and the adsorption force is zero. N1 =0 means that the adsorption force of the first adsorption unit is zero;

[0057] All adsorption units should meet the following conditions:

[0058] ∑F m ≥GH / 2L

[0059] where ∑F m is the sum of the adsorption forces of all adsorption units, GH / 2L is the torque generated by the object, which means the weight (G) multiplied by half the height (H / 2) divided by the distance from the adsorption unit to point A (L).

[0060] ∑F m ≥GH / 2L means the sum of the adsorption forces of all adsorption units must be greater than or equal to the torque generated by the object to ensure the stability of the system.

[0061] The detection car frame is equipped with a drive motor, a motor drive board and a relay. The adsorption structure in the car track is composed of a yoke and an excitation coil. The control chips of the detection car are ESP8266 module and MPU-6050 chip respectively. The detection car receives instructions from the control platform through ESP8266, and converts the instructions received by ESP8266 into electrical signals of the drive motor through L298N, so as to remotely drive and control the robot.

[0062] The main body of the inspection vehicle adopts a permanent magnetic adsorption crawler as the bottom drive part, and drives the double-row active sprocket through a reduction motor to achieve stable motion control.

[0063] The triangular chain drive structure enhances the stability and durability of the crawler, making it suitable for movement in complex terrain and curved environments, and has excellent curved surface adaptability. The damping shock absorber installed at the crawler support helps to tighten the crawler, better fit the hull surface, and improve the machine's mobility on surfaces with different curvatures. The spring damping shock absorber adjusts the angle between the crawler and the frame chassis to ensure maximum contact between the hull and the crawler, thereby enhancing the adsorption force and stability, and real-time detection of the trolley's adsorption force, by adjusting the working state of the adsorption structure (such as the yoke and the excitation coil) to ensure that the trolley is under the influence of The overturning moment can still maintain stable contact with the wall to prevent it from falling off the wall. The combination of the ESP8266 module and the MPU-6050 chip enables the detection car to receive remote commands and perform corresponding actions. The ESP8266 module is used to receive commands from the control platform, while the MPU-6050 chip may be used for posture detection (such as tilt angle, etc.) to provide data support for stable control. The L298N motor driver board converts the received commands into electrical signals for driving the motor, realizing remote drive control of the robot and improving the convenience and flexibility of operation.

[0064] At the same time, the system adopts a modular design, making each part (such as the drive system, control system, detection system, etc.) relatively independent, which is convenient for maintenance and upgrading. Through the optimized mechanical structure and control algorithm, the working reliability and stability of the detection vehicle in complex environments are improved.

[0065] The detection unit includes a crack inspection probe, a weld ultrasonic detection head and a surface defect detection probe, which respectively detect cracks, welds and surface defects on the hull surface, and transmit them to the information processing system for processing through the detection transmission system.

[0066] The detection feedback system feeds back the detection results of the detection unit to the remote terminal. A pair of public and private keys are generated between the feedback system and the remote terminal. The public key is public and used to encrypt data, and the private key is private and used to decrypt data to ensure the security of data encryption transmission. The public key password is generated as follows:

[0067] S=Sig(M)

[0068] Where S is the digital signature of data M obtained from user A;

[0069] Attach S to M and send it to user B. User B verifies S. The formula is as follows:

[0070] M′=S′modn

[0071] M' is the digital signature M; S' is the digital signature of the transmission end. Verify whether M' is consistent with M. When M'=M, S=S, which proves that the signature S' comes from user A and the public key password is correct. When M'≠M, S'≠S, which proves that the signature S' is a forged signature and the public key password is wrong.

[0072] In the above formula, S is the signature result, Sig(M) is the digital signature of message M, M′ is the value calculated during the verification process, S′ is the transmitted signature value, n is the modulus, and M is the original message;

[0073] The data is encrypted in two ways during the encryption process. The specific combined algorithms are as follows:

[0074] C=AESk AES (M)

[0075] Where C is the encrypted data, AESk AES (M) is the AES algorithm and key k AES Encrypted data M. Specific embodiment 2:

[0077] like Figure 1-8 As shown, a magnetic beacon is pasted on the surface of the inspection hull, and the front of the beacon is marked with a high-brightness color. The interior of the inspection vehicle contains a CPU, a digital input and output module, a CANopen module, an RS485 communication module, and a Profibus communication module. An obstacle sensor is connected to the side of the bottom of the inspection vehicle to scan the semicircular area within 1m in front of it and detect obstacles in the driving direction. The obstacles here mainly include seaweed and marine biological attachments. If there is an obstacle in front of the vehicle, the vehicle will detour.

[0078] The inspection hull is equipped with a HuskyLens visual sensor, which is installed on the front of the inspection vehicle frame. It checks the road signs pasted on the path in real time to identify the guide route and correct the vehicle body. The inspection vehicle uses the visual sensor to identify the object, color and label of the sign. The visual sensor is connected to a 2.0-inch IPS display (320×240 resolution) that can directly display the debugging process and recognition results. Interface: Onboard UART / I2C interface, can be connected to mainstream controllers such as Arduino, control board, micro:bit, Raspberry Pi, etc. Programming support: Supports programming languages ​​such as C / C++ or Python for communication and interaction, and also supports graphical programming software such as Mind+, MakeCode, etc.

[0079] The specific direction indicator pattern is: Figure 8 As shown, the color of the pointer pattern is red, and you can choose to add it according to the actual situation.

[0080] The solution of the inspection vehicle with magnetic beacons attached to the surface of the hull has achieved good results in road sign recognition and guidance, obstacle detection and detour, communication interface and programming support, display and debugging, modular design and scalability, as well as intelligence and autonomy, providing a strong guarantee for the automation and accuracy of hull inspection tasks. Specific embodiment three:

[0082] like Figure 1-8 As shown, the crack detection probe adopts the YOLOv5 network model algorithm. The scale of crack samples in the main hull structure external crack data set varies greatly, and there are many small cracks. The model trained on this data set is difficult to ensure the recognition accuracy of targets of each scale. By improving the backbone network of the YOLOv5 algorithm, the feature extraction capability is improved; at the same time, the FPN-PAN design is integrated to ensure the full integration of multi-scale features, especially to ensure that the features of small targets can be extracted; the input end uses the MOSAIC algorithm for data enhancement to increase the frequency of small targets. Therefore, the present invention adopts YOLOv5 as the basic model for the external crack detection task of the main hull structure. The information processing system receives the crack image collected by the crack detection probe through the detection transmission system, and organizes and constructs the external crack data set of the main hull structure. The training of the model is first pre-trained on the source data set, and the backbone and neck parts of the obtained pre-trained model are parameter-frozen, and then transfer learning training is performed on the external crack data set of the main hull structure to complete knowledge migration. Finally, a variety of optimization algorithms are introduced to improve the detection accuracy.

[0083] After obtaining the pre-trained model, the parameters of the backbone and neck parts are frozen, 260 samples of the main hull structure external crack data set are randomly selected for training, and the rest are used for testing. The number of training iterations is 700, the iteration batch size is set to 32, the optimizer selects SGD, uses the quaternary data loader, the learning rate uses the cosine annealing algorithm, and the basic learning rate is set to 0.01. The present invention takes the deployment of the model on a hardware platform equipped with a 2080T graphics card as an example, tests the inference speed of the model, and performs crack detection on 50 test samples. The average inference time for each sample is 14.2ms, and the processing frame rate reaches 70 frames / s.

[0084] The system uses the Label Img tool to annotate samples and annotate the crack areas in the sample images in the form of rectangular frames. The annotated areas try to include as little background as possible. Based on five open source crack datasets, including CFD, CRACK500, Deep-CRACK, and GAPs384, the system organizes and re-annotates a crack dataset with 2,798 images as the source dataset for model migration training.

[0085] There are four groups of weld ultrasonic detection heads. The four groups of ultrasonic probes are located in different positions. The function generator calculates the hardware facilities required to detect the welds of the hull structure. The power amplifier amplifies the power of the detection head. After collecting data, the collected results are displayed on the digital oscilloscope and dynamically updated through the information processing system.

[0086] The steps for weld detection are as follows: first, build an ultrasonic probe array, simulate the working status of the probe in the computer, build the array according to the working status of each probe, set the probe drive circuit, the probe drive circuit includes multiple lines, each corresponding to a probe, the computer tracks in real time, records the working status of the probe, uses a digital display to receive signals, and displays the detection results on the LED screen.

[0087] Before establishing the ultrasonic nondestructive testing method, it is necessary to analyze the working principle of the ultrasonic nondestructive testing method. The schematic diagram of the ultrasonic nondestructive testing principle is as follows: Figure 5 As shown, observe Figure 5 , where A, B, C, and D represent different positions of the ultrasonic probe. Points 1 to 6 represent the connection points between the metal and the ultrasonic wave. The ultrasonic wave will form different reflection arcs on the bottom and the surface. After multiple collisions, the reflection arc will form a secondary reflection point. The pulse between the bottom reflection pulse and the surface reflection pulse is a secondary pulse. If a secondary pulse appears, it proves that there are welding defects in the weld of the hull structure. Welding defects can also be detected based on the location of the pulse point. If the pulse point is located at point A and point C, it means that there are welding defects in the weld; if the pulse point is located at point B and point D, it means that there are no welding defects in the weld. Detecting welding defects is an important part of the ultrasonic non-destructive testing method. Usually, the staff will determine the location of the weld hull defects based on the position of the reflected pulse probe; judge the size and shape of the defects based on the shape of the reflected pulse; and quickly detect the defects of the weld based on whether the ultrasonic wave has a continuous direction of movement. The ultrasonic probe is moved in the parallel and perpendicular directions of the weld, and the movement pattern is zigzag.

[0088] Design the working framework according to the above working principle, such as Figure 6 As shown by Figure 6 It can be seen that the ultrasonic nondestructive testing process is controlled by remote control. The hardware facilities required for detecting the hull structure welds are calculated in the function generator, the power amplifier amplifies the power of the detector, and after collecting data, the collected results are displayed in the digital oscilloscope and dynamically updated through the real-time data analysis system. Figure 6The working framework is an automated framework, which has a certain simplification effect on the ultrasonic detection process, and the motion state of the probe can also be simplified. The ultrasonic detection probe here is installed on a special detector, which will automatically arrange the detection array to make the probe detection easier. The probe array keeps a uniform motion in the matrix, and the computer system records the pulses reflected from the surface, the pulses emitted from the bottom surface, and the pulses between the bottom surface and the surface. When the reflected pulse appears, the probe does not need to move in the section, but only needs to move at a uniform speed in the vertical direction, which can analyze the nature, size, quantity, shape and location of the weld. At the same time, the synchronous recorder should record the movement process.

[0089] The surface defect detection probe collects images of the hull surface, and the information processing system creates the NEU dataset, which contains 1,800 grayscale images, divided into five types of defects: scratches, folded heavy skin, ears, oxidized iron scale, and pitting. The dataset has two versions, NEU-CLS and NEU-DET. The former is used for classification, and the latter is used for target detection.

[0090] In summary:

[0091] 1. The main body of the inspection vehicle adopts a permanent magnetic adsorption crawler as the bottom driving part, and drives the double-row active sprocket through a reduction motor to achieve stable motion control.

[0092] The triangular chain drive structure enhances the stability and durability of the track, making it suitable for movement in complex terrain and curved environments, and has excellent curved surface adaptability. The damping shock absorber installed at the track support helps to tighten the track, better fit the hull surface, and improve the machine's mobility on surfaces with different curvatures. The spring damping shock absorber adjusts the angle between the track and the frame chassis to ensure maximum contact between the hull and the track, thereby enhancing the adsorption force and stability, and detecting the adsorption force of the trolley in real time. By adjusting the working state of the adsorption structure (such as the yoke and the excitation coil), it ensures that the trolley can still maintain stable contact with the wall when subjected to an overturning moment to prevent it from detaching from the wall. The ESP8266 module and MPU-6050 chip The combination of the two chips enables the detection car to have the ability to receive remote commands and perform corresponding actions. The ESP8266 module is used to receive commands from the control platform, while the MPU-6050 chip may be used for posture detection (such as tilt angle, etc.) to provide data support for stable control. The L298N motor driver board converts the received commands into electrical signals for driving the motor, realizing remote drive control of the robot and improving the convenience and flexibility of operation. At the same time, the system adopts a modular design, which makes each part (such as drive system, control system, detection system, etc.) relatively independent, easy to maintain and upgrade, and improves the working reliability and stability of the detection car in complex environments through optimized mechanical structure and control algorithm.

[0093] 2. The solution of the inspection vehicle with magnetic beacons attached to the surface of the hull has achieved good results in road sign recognition and guidance, obstacle detection and detour, communication interface and programming support, display and debugging, modular design and scalability, as well as intelligence and autonomy, providing a strong guarantee for the automation and accuracy of hull inspection tasks.

[0094] 3. The present invention adopts YOLOv5 as the basic model for the task of external crack detection of the main hull structure. The information processing system receives the crack images collected by the crack detection probe through the detection transmission system, and organizes and constructs the external crack data set of the main hull structure. The training of the model is first pre-trained on the source data set, and the parameters of the backbone and neck parts of the obtained pre-trained model are frozen. Then, transfer learning training is performed on the external crack data set of the main hull structure to complete knowledge transfer. Finally, a variety of optimization algorithms are introduced to improve the detection accuracy.

[0095] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprising a reference structure" do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0096] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. Real-time detection and feedback system for steel structures based on data prediction, characterized by: The system is composed of a machine execution system, a detection transmission system, an information processing system and a detection feedback system, wherein: The machine execution system includes a detection trolley body and a detection unit. The bottom driving part of the detection trolley body adopts a permanent magnetic adsorption track. The permanent magnetic adsorption track on one side of the robot is composed of a front and rear double-row driven sprocket and a middle double-row driving sprocket to form a triangular chain transmission structure. The double-row driving sprocket is driven by a reduction motor, thereby driving the entire permanent magnetic adsorption track to move; The track support part of the permanent magnetic adsorption track is equipped with a damping shock absorber, and the track is tightened to better fit the curved surface of the hull. One end of the permanent magnetic adsorption track is fixedly connected to the frame chassis, and the other end is hinged to one end of the spring damping shock absorber, and the other end of the spring damping shock absorber is hinged to the frame chassis to ensure maximum contact between the hull and the track. The detection trolley is affected by gravity, so that the detection trolley is subjected to the overturning torque. The detection trolley has a tendency to overturn away from the wall at the contact point A. In order to overcome the situation where the robot is separated from the wall due to the overturning torque, it is necessary to detect the adsorption force of the trolley in real time to overcome it. The condition that the detection trolley does not overturn away from the wall is: ∑M A ≥0 By calculating the distance to point A, where the distance refers to calculating the sum of all moments about point A; M A Represents the moment about point A; ∑M A ≥0 means that the sum of all moments about point A is greater than zero, which means that the system is unstable at point A or has a tendency to overturn in the clockwise direction; The calculation formula for the system to reach equilibrium at point A is as follows: where F m1 is the force of the first adsorption unit, F N1 is the force of the first adsorption unit in the critical state, L is the distance from the adsorption unit to point A, G represents the weight of the object, H is the height of the object, ∑M A =0 means that the sum of all torques about point A is zero, and the system reaches equilibrium at point A; When the first adsorption unit is in critical state, the following formula is obtained; F N1 =0 The critical state is when the adsorption unit just leaves the surface of the object, and the adsorption force is zero. N1 =0 means that the adsorption force of the first adsorption unit is zero; All adsorption units should meet the following conditions: ΣF m ≥GH / 2L where ∑F m is the sum of the adsorption forces of all adsorption units, GH / 2L is the torque generated by the object, which means the weight (G) multiplied by half the height (H / 2) divided by the distance from the adsorption unit to point A (L); ∑F m ≥GH / 2L means the sum of the adsorption forces of all adsorption units must be greater than or equal to the torque generated by the object to ensure the stability of the system. The detection feedback system feeds back the detection results of the detection unit to the remote terminal. A pair of public key and private key is generated between the feedback system and the remote terminal. The public key is public and used to encrypt data, and the private key is private and used to decrypt data to ensure the security of data encryption transmission. The public key password is generated as: S=Sig(M) Where S is the digital signature of data M obtained from user A; Attach S to M and send it to user B. User B verifies S. The formula is as follows: M′=S′modn M' is the digital signature M; S' is the digital signature of the transmission end. Verify whether M' is consistent with M. When M'=M, S=S, which proves that the signature S' comes from user A and the public key password is correct. When M'≠M, S'≠S, which proves that the signature S' is a forged signature and the public key password is wrong. In the above formula, S is the signature result, Sig(M) is the digital signature of message M, M′ is the value calculated during the verification process, S′ is the transmitted signature value, n is the modulus, and M is the original message; The data is encrypted in two ways during the encryption process. The specific combined algorithms are as follows: C=AESk AES (M) Where C is the encrypted data, AESk AES (M) is the AES algorithm and key k AES The encrypted data M is encrypted.

2. The steel structure real-time detection and feedback system based on data prediction according to claim 1 is characterized in that: A driving motor, a motor driving board and a relay are installed in the frame of the detection vehicle. The adsorption structure in the vehicle track is composed of a yoke and an excitation coil. The control chips of the detection vehicle are respectively an ESP8266 module and an MPU-6050 chip. The detection vehicle receives instructions issued by the control platform through ESP8266, and converts the instructions received by ESP826 into electrical signals of the driving motor through L298N, so as to remotely drive and control the robot.

3. The real-time detection and feedback system for steel structures based on data prediction according to claim 2 is characterized in that: The detection unit comprises a crack detection probe, a weld ultrasonic detection head and a surface defect detection probe.

4. The steel structure real-time detection and feedback system based on data prediction according to claim 1 is characterized in that: A magnetic beacon is pasted on the surface of the detection hull, and the front of the beacon is marked with a high-brightness color. The interior of the detection vehicle contains a CPU, a digital input and output module, a CANopen module, an RS485 communication module, and a Profibus communication module. The bottom side of the detection vehicle is externally connected to an obstacle sensor, which scans the semicircular area within 1m in front and detects obstacles in the driving direction. The obstacles here mainly include seaweed and marine biological attachments. If there is an obstacle in front of the vehicle, the vehicle will detour.

5. The steel structure real-time detection and feedback system based on data prediction according to claim 4 is characterized in that: The detection hull is equipped with a HuskyLens visual sensor, which is installed on the front of the detection vehicle frame. It checks the road signs pasted on the path in real time to identify the guide route and correct the vehicle body. The detection vehicle uses the visual sensor to perform object recognition, color recognition and label recognition on the pointer. The visual sensor is externally connected to a 2.0-inch IPS display (320×240 resolution), which can directly display the debugging process and recognition results. Interface: onboard UART / I2C interface, which can be connected to mainstream controllers such as Arduino, control board, micro:bit, RaspberryPi, etc.

6. The steel structure real-time detection and feedback system based on data prediction according to claim 3 is characterized in that: The crack detection probe adopts the YOLOv5 network model algorithm and integrates the FPN-PAN design to ensure the full integration of multi-scale features. The input end adopts the MOSAIC algorithm for data enhancement to increase the frequency of small targets. The information processing system receives the crack images collected by the crack detection probe through the detection transmission system, and organizes and constructs the external crack data set of the main hull structure. The training of the model is first pre-trained on the source data set, and the backbone and neck parts of the obtained pre-trained model are parameter-frozen. Then, transfer learning training is performed on the external crack data set of the main hull structure to complete knowledge transfer. Finally, a variety of optimization algorithms are introduced to improve the detection accuracy.

7. The steel structure real-time detection and feedback system based on data prediction according to claim 6 is characterized in that: After obtaining the pre-trained model, the parameters of the backbone and neck parts are frozen, 260 samples of the external crack data set of the main hull structure are randomly selected for training, and the rest are used for testing. The number of training iterations is 700, the iteration batch size is set to 32, the optimizer selects SGD, uses the quaternary data loader, the learning rate uses the cosine annealing algorithm, and the basic learning rate is set to 0.

01.

8. The steel structure real-time detection and feedback system based on data prediction according to claim 3 is characterized in that: There are four groups of weld ultrasonic detection heads, and the four groups of ultrasonic probes are at different positions. The function generator calculates the hardware facilities required for detecting the welds of the hull structure, and the power amplifier amplifies the power of the detection head. After collecting data, the collection results are displayed on the digital oscilloscope and dynamically updated through the information processing system.

9. The steel structure real-time detection and feedback system based on data prediction according to claim 8, characterized in that: The weld detection steps are as follows: first, an ultrasonic probe array is constructed, the working state of the probe is simulated in a computer, an array is constructed according to the working state of each probe, a probe driving circuit is set, the probe driving circuit includes multiple lines, each corresponding to a probe, the computer tracks in real time, records the working state of the probe, a digital display is used to receive signals, and the detection results are displayed on an LED screen.

10. The steel structure real-time detection and feedback system based on data prediction according to claim 3 is characterized in that: The surface defect detection probe collects images of the hull surface, and the information processing system creates the NEU data set, which contains 1,800 grayscale images, divided into five types of defects: scratches, folded heavy skin, ears, oxidized iron scale and pitting. The data set has two versions, NEU-CLS and NEU-DET, the former is used for classification, and the latter is used for target detection.