Tunnel lining disease sonar detection device carried by underwater robot

Abstract

The invention relates to the technical field of tunnel detection, in particular to a tunnel lining disease sonar detection device carried by an underwater robot, which comprises a sonar detection system, a data processing and analyzing system and an underwater robot carrying and operating system. According to the invention, a tunnel does not need to be emptied, autonomous operation is realized through carrying the detection device by the underwater robot, high risks of manual detection and adverse effects on a tunnel structure are avoided, and the sonar detection system realizes all-directional coverage detection of a tunnel lining by virtue of a high-resolution probe and a three-dimensional scanning module; the self-adaptive signal adjusting module can dynamically adjust parameters according to the underwater environment and adapt to complex water quality and water flow conditions, the data processing system accurately recognizes cracks, spalling, void and other diseases through a deep learning algorithm and determines key information of the diseases, data are visually displayed and stored in real time, and the real-time performance of the system is improved. A stable carrying platform, intelligent navigation and a reliable energy communication system provide stable guarantee for detection.

Description

A sonar detection device for tunnel lining defects carried by an underwater robot Technical Field

[0001] This invention relates to the field of tunnel inspection technology, specifically to a sonar detection device for tunnel lining defects carried by an underwater robot. Background Technology

[0002] As a core infrastructure in water conservancy and other engineering projects, the integrity of the tunnel lining structure directly determines the safety and durability of the project's operation. During long-term service, the lining is affected by multiple factors such as water flow erosion, alternating water pressure, geological settlement, and chemical corrosion, which can easily lead to defects such as cracks, spalling, and voids. If these defects are not detected and addressed in a timely and accurate manner, they may cause major safety accidents such as lining collapse and water leakage.

[0003] Traditional tunnel lining inspection methods mainly rely on manual inspection and conventional instrument inspection. Manual inspection requires emptying the tunnel, which not only disrupts the normal operation of the tunnel and is inefficient, but also exposes the inspectors to high risks such as oxygen deficiency and collapse. Conventional ultrasonic inspection and electromagnetic induction inspection have shortcomings such as limited detection range (single-directional scanning), insufficient accuracy (difficult to identify tiny cracks smaller than 1 mm), and weak anti-interference ability (greatly affected by underwater turbidity and noise), which cannot meet the inspection needs of complex underwater environments.

[0004] With the development of underwater robot technology, its application in tunnel inspection has gradually become more widespread, but the existing inspection devices carried by underwater robots still have significant shortcomings:

[0005] Sonar detection systems lack effective signal noise reduction mechanisms, and underwater environmental noise can easily cause detection signal distortion, affecting the identification of minor defects;

[0006] Data processing only focuses on identifying the "current state" of the disease, without predicting its development trend, and cannot provide forward-looking guidance for engineering maintenance;

[0007] The underwater robot lacks active obstacle avoidance capabilities. Protrusions and falling objects inside the tunnel can easily cause collisions and damage to the robot, affecting the continuity of detection.

[0008] Sonar signal modulation largely relies on fixed thresholds, which are not well adapted to complex and variable underwater environments (such as local high turbidity and turbulent currents).

[0009] Therefore, developing a sonar detection device for tunnel lining defects that combines high-precision detection, multi-dimensional analysis, high-safety operation, and strong environmental adaptability has become the key to solving the current technological bottlenecks. Summary of the Invention

[0010] The purpose of this invention is to provide a sonar detection device for tunnel lining defects mounted on an underwater robot, so as to solve the problems mentioned in the background art.

[0011] To achieve the above objectives, the present invention provides the following technical solution: a sonar detection device for tunnel lining defects carried by an underwater robot, comprising a sonar detection system, a data processing and analysis system, and an underwater robot carrying and operation system;

[0012] The sonar detection system includes a high-resolution phased array sonar probe, a three-dimensional sonar scanning module, an adaptive sonar signal conditioning module, and a sonar signal noise reduction module.

[0013] The data processing and analysis system includes a high-performance data acquisition card, a deep learning data processing algorithm module, a real-time data display and storage module, and a disease trend prediction module.

[0014] The underwater robot mounting and operation system includes a stable underwater robot, an intelligent navigation and positioning system, an energy supply and communication system, and an emergency obstacle avoidance module.

[0015] Preferably, the frequency of the high-resolution phased array sonar probe is flexibly adjustable within the range of 100kHz-1MHz to detect tunnel lining cracks and thin voids of 0.8 mm or more.

[0016] Preferably, the three-dimensional sonar scanning module is installed at the front end of the underwater robot and is driven by a motor to rotate 360 ​​degrees horizontally and 180 degrees vertically, and the sonar probe moves according to a preset scanning path and step size.

[0017] Preferably, the adaptive sonar signal adjustment module automatically adjusts the intensity, frequency, and pulse width parameters of the sonar transmission signal in real time based on factors such as water turbidity, water flow velocity, and water temperature in the underwater environment, combined with a preset K-means clustering environment-parameter matching library.

[0018] Preferably, the sonar signal noise reduction module uses a db4 wavelet basis for 3-level decomposition, applies hard thresholding to high-frequency noise components, preserves and reconstructs low-frequency signal components, and outputs a signal-to-noise ratio ≥45dB.

[0019] Preferably, the deep learning data processing algorithm module is based on the improved YOLOv5 model. An attention mechanism module is built in the improved YOLOv5 model and it forms a collaborative working link with the original feature extraction network and detection head architecture of the model. It is trained with a training dataset containing a large number of tunnel lining defects, and can identify cracks, spalling and voids, and accurately measure the size, location and severity level of the defects, with an identification accuracy of ≥95%.

[0020] Preferably, the intelligent navigation and positioning system integrates underwater acoustic positioning, inertial navigation and visual positioning technologies. The underwater robot automatically travels according to the preset detection path and adjusts its speed and direction in real time, and has a dynamic path correction function.

[0021] Preferably, the disease trend prediction module is based on the LSTM time series analysis model. It takes historical detection data and real-time environmental parameters as input, predicts the spread rate, development direction and risk level of the disease in the next 6-12 months, and outputs trend prediction curves and maintenance suggestions.

[0022] Preferably, the emergency obstacle avoidance module consists of a lidar sensor, a vision camera, and an obstacle avoidance algorithm. It monitors obstacles in front of the robot in real time, and achieves emergency obstacle avoidance by adjusting the driving direction or slowing down. After obstacle avoidance is completed, it automatically restores the original detection path.

[0023] Compared with the prior art, the beneficial effects of the present invention are:

[0024] 1. This invention eliminates the need to empty the tunnel. It enables autonomous underwater operation by using an underwater robot equipped with a detection device, completely avoiding the high risks of manual inspection. It also avoids the adverse effects of emptying the tunnel on the project's operational status and structure, significantly improving the convenience of inspection. The sonar detection system, with its high-resolution phased array sonar probe and three-dimensional scanning module, achieves omnidirectional, blind-spot-free scanning of the tunnel lining, solving the problem of limited detection range of traditional instruments. The adaptive sonar signal adjustment module can dynamically adjust signal parameters according to changes in underwater water turbidity, flow velocity, water temperature, and other environmental factors, enabling the device to adapt to complex and ever-changing underwater environments and ensuring detection stability.

[0025] 2. This invention also utilizes a deep learning data processing algorithm module to accurately identify various defects such as cracks, spalling, and voids, and clearly define the type, location, and severity of these defects. This overcomes the weakness in data processing capabilities of existing detection devices. A stable platform and intelligent navigation and positioning system provide a reliable operating foundation for the detection device, ensuring a smooth and orderly detection process. The sonar signal noise reduction module effectively filters out environmental interference, the emergency obstacle avoidance module prevents robot collision damage, and the defect trend prediction module provides forward-looking guidance for engineering maintenance. This invention comprehensively overcomes the shortcomings of existing technologies in terms of accuracy, safety, and functionality, meeting the practical needs for efficient and accurate detection of tunnel lining defects. Attached Figure Description

[0026] Figure 1 is a block diagram of the overall structure of a sonar detection device for tunnel lining defects carried by an underwater robot according to the present invention. Detailed Implementation

[0027] 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.

[0028] Please refer to Figure 1. The present invention provides a technical solution: a sonar detection device for tunnel lining defects carried by an underwater robot, including a sonar detection system, a data processing and analysis system, and an underwater robot carrying and operation system. The above systems are modularly connected and work together to complete the task of detecting tunnel lining defects.

[0029] Sonar detection system:

[0030] It includes a high-resolution phased array sonar probe, a three-dimensional sonar scanning module, an adaptive sonar signal conditioning module, and a sonar signal noise reduction module;

[0031] High-resolution phased array sonar probe: The frequency can be flexibly adjusted within 100kHz-1MHz, and it can scan the tunnel lining with a 360-degree angular resolution, and detect cracks and thin voids of 0.8 mm and above.

[0032] 3D sonar scanning module: Installed at the front of the underwater robot, it can rotate 360 ​​degrees horizontally and 180 degrees vertically through motor drive. The sonar probe can move according to the preset scanning path and step size to achieve all-round scanning without blind spots.

[0033] Adaptive sonar signal adjustment module: Based on factors such as water turbidity, water flow velocity, and water temperature in the underwater environment, and using the K-means clustering environment-parameter matching algorithm, it can dynamically adjust the signal strength, frequency, and pulse width in real time according to water turbidity, water flow velocity, and water temperature (adjustment period ≤ 0.05s). It can work stably in the range of turbidity 0-50NTU, water flow 0-3m / s, and water temperature 5-35℃, with significantly enhanced environmental adaptability compared to existing devices. It also has a preset optimal parameter library for multiple scenarios to improve adjustment response speed and accuracy.

[0034] Sonar signal noise reduction module: It adopts wavelet transform noise reduction algorithm to perform multi-scale decomposition and threshold processing on the sonar received signal, filter out interference signals such as water flow noise and equipment vibration noise, retain the disease characteristic signal, improve the output signal-to-noise ratio, effectively filter out environmental noise interference, and thus improve the detection accuracy.

[0035] Data processing and analysis system:

[0036] It includes a high-performance data acquisition card, a deep learning data processing algorithm module, a real-time data display and storage module, and a disease trend prediction module;

[0037] High-performance data acquisition card: sampling rate meets 400kHz, has high-speed cache function, cache capacity of not less than 64GB, to ensure continuous data acquisition without loss;

[0038] Deep learning data processing algorithm module: By constructing a training dataset containing a large number of tunnel lining defects (cracks, spalling, voids, etc.), a deep learning model based on improved YOLOv5 is trained. An attention mechanism module is built in the improved YOLOv5 model and forms a collaborative working link with the model's original feature extraction network and detection head architecture. This can quickly and accurately identify defects such as cracks, spalling, and voids, and accurately measure the three-dimensional location, physical size (error ≤ ±0.1mm) and severity level of the defects, thereby improving data processing efficiency and providing accurate data support for defect treatment.

[0039] Real-time data display and storage module: The processed data is displayed in real time on the control terminal in the form of 3D images of tunnel lining, disease location markings, disease parameter lists, and detection path trajectories; it adopts a distributed storage architecture, supports local data caching and cloud backup, and the storage format is compatible with commonly used engineering data formats (such as CAD and GIS).

[0040] Disease trend prediction module: Based on the LSTM (Long Short-Term Memory Network) time series analysis model, it takes historical and real-time detection data as input, and combines environmental factors (water flow velocity, water temperature, water pressure) to predict the expansion rate, development direction and risk level of diseases in the next 6-12 months. It outputs trend prediction curves and maintenance suggestions, which breaks through the limitations of traditional post-event detection and realizes pre-event prediction. It provides a scientific basis for engineering preventive maintenance and reduces the probability of accidents.

[0041] Underwater robot mounting and operation system:

[0042] It includes a stable underwater robot, an intelligent navigation and positioning system, an energy supply and communication system, and an emergency obstacle avoidance module;

[0043] Stable underwater robot: Made of high-strength corrosion-resistant alloy materials (such as titanium alloy and fiberglass composite material), it can maintain stable operation in an environment with a water flow speed of 3m / s. The platform is equipped with a high-precision sonar mounting bracket, which can be adjusted by motor drive to achieve ±5° angle fine adjustment and ±10cm position fine adjustment of the sonar detection device.

[0044] Intelligent navigation and positioning system: It integrates underwater acoustic positioning, inertial navigation and visual positioning technologies, with a positioning accuracy error of ≤±5cm; the underwater robot can drive automatically according to the preset detection path, and adjust the driving speed (adjustable from 0.1-1m / s) and direction in real time. It optimizes the scanning path by combining real-time detection data and performs dynamic path correction to avoid repeated detection or missed areas.

[0045] Energy supply and communication system: It adopts a large-capacity lithium battery pack, with a single-charge endurance of no less than 8 hours; the communication system adopts a dual-mode switching method of fiber optic communication and underwater acoustic communication. Fiber optic communication is used for high-speed data transmission over short distances (≤500m) (rate ≥1Gbps), and underwater acoustic communication is used for data transmission over long distances (≤5km) (rate ≥100kbps), ensuring continuous and stable data transmission;

[0046] Emergency obstacle avoidance module: Composed of a LiDAR sensor, a vision camera, and an obstacle avoidance algorithm, it monitors obstacles (such as tunnel protrusions or falling rocks) within a range of 0.5-5m in front of the robot in real time. When a collision risk is detected, it automatically triggers an obstacle avoidance command, adjusting the robot's direction or decelerating to achieve emergency obstacle avoidance. The obstacle avoidance response time is ≤0.1s.

[0047] Software initialization:

[0048] Preset scanning path: Based on the tunnel cross-sectional dimensions (e.g., diameter 3-10m), a spiral scanning path can be set via the control terminal, with an adjustable step size of 0.1-0.5m;

[0049] Training dataset construction: Collect 10,000+ sonar images of tunnel lining defects of different types, label the defect type, size, location and other information, and divide them into training set, validation set and test set in a 7:2:1 ratio. Train the improved YOLOv5 model. After 100 iterations, the model recognition accuracy stabilized at over 95%.

[0050] Environment-Parameter Matching Library Establishment: Optimal sonar signal parameters (intensity, frequency, pulse width) under different turbidity, water flow velocity, and water temperature were collected, and 10 typical scenarios were divided using the K-means clustering algorithm to establish a parameter matching library.

[0051] Component selection and assembly:

[0052] Installation and debugging of high-resolution phased array sonar probe: Install a high-resolution phased array sonar probe on the front-end 3D sonar scanning module of the underwater robot. Select a phased array sonar probe with a frequency range of 100kHz-1MHz (such as ResonSeabatT20P), and match it with a 360-degree rotating motor. The sonar signal noise reduction module uses an FPGA chip (such as Xilinx Zynq-7000) to implement the wavelet transform algorithm. After installation, the sonar probe is debugged using professional debugging equipment, including calibrating the transmission and reception performance of the sonar transducer unit and testing the working status of the sonar probe at different frequencies. This ensures that the sonar probe can transmit and receive sound signals normally and meet the resolution requirements. The operation and control of the 3D sonar scanning mechanism: The 3D sonar scanning mechanism is connected to the control system of the underwater robot. Parameters such as the rotation angle range, rotation speed, and scanning step size of the scanning mechanism are set through programming. During the detection process, the control system drives the motor of the scanning mechanism according to the preset parameters, causing it to rotate and scan the sonar probe in both horizontal and vertical directions, achieving a comprehensive 3D scan of the tunnel lining. Simultaneously, the operating status of the scanning mechanism is monitored in real time to ensure its stable and reliable operation.

[0053] The workflow of the adaptive sonar signal conditioning module is as follows: Before the underwater robot is deployed, the adaptive sonar signal conditioning module is connected to the sonar probe and underwater environmental sensors (such as turbidity sensors, current velocity sensors, and temperature sensors). When the underwater robot enters the tunnel to begin detection, the environmental sensors collect underwater environmental parameters in real time and transmit the data to the adaptive sonar signal conditioning module. Based on the received environmental parameters, the conditioning module automatically adjusts the intensity, frequency, and pulse width of the sonar transmission signal using a preset algorithm, and then sends the adjusted signal to the sonar probe for transmission. During the detection process, changes in environmental parameters are continuously monitored, and the sonar signal parameters are dynamically adjusted in real time to ensure the accuracy of sonar detection.

[0054] Specifically, before the underwater robot is deployed, the adaptive sonar signal conditioning module undergoes initialization. First, the module's hardware connections are checked for proper functioning, including connections to the sonar probe, underwater environment sensors (turbidity sensor, flow velocity sensor, temperature sensor, etc.), and the underwater robot's control system. All connections are ensured to be secure and unsecured, and signal transmission interfaces are clean and undamaged. Next, the module's parameters are initialized. Initial parameters such as the sonar signal's initial intensity, frequency, and pulse width are preset; these initial parameters can be set based on the general environmental characteristics of the tunnel and past detection experience. Simultaneously, the data storage area is initialized to provide storage space for subsequently collected environmental and conditioning parameters. Finally, a module self-test is performed, initiating the self-test program to check if the core components, such as the processor, memory, and signal processor, are functioning correctly. After passing the self-test, the module enters standby mode, awaiting the underwater robot's start command. To further improve the accuracy of sonar detection, the core processor of the adaptive sonar signal conditioning module reads pre-processed environmental parameter data from the buffer and, combined with preset conditioning rules, makes signal parameter adjustment decisions.

[0055] When the turbidity of the water exceeds a preset threshold, the processor determines that the intensity of the sonar signal needs to be increased to ensure that the sound waves can penetrate the turbid water and be effectively reflected back. Based on the specific value of turbidity and the cluster category, an appropriate intensity increase value is selected from a preset intensity adjustment table, and the target signal intensity value is calculated.

[0056] When the water flow velocity exceeds a preset threshold, the processor analyzes the impact of the water flow velocity on the propagation of the sonar signal and decides to adjust the pulse width of the sonar signal. Based on the magnitude of the water flow velocity and the clustering results, a suitable pulse width adjustment scheme is selected, and the target pulse width value is determined.

[0057] When the water temperature changes significantly, considering the impact of water temperature on the speed of sound wave propagation, the processor appropriately adjusts the frequency of the sonar signal. Based on the magnitude of the water temperature change and the clustering category, the processor retrieves the corresponding frequency adjustment parameters from the frequency adjustment database and calculates the target frequency value. The processor then converts the calculated target signal strength, frequency, and pulse width into corresponding control signals and transmits them to the sonar signal drive unit. The sonar signal drive unit adjusts the sonar probe's transmitting circuit according to the control signals, causing the sonar probe to transmit sound wave signals according to the new parameters. During the signal parameter adjustment process, the module monitors the sonar probe's operating status in real time, collects the actual output signal parameters of the sonar probe through the feedback circuit, and compares them with the target parameters. If there is a deviation between the actual and target parameters, the processor will readjust until the actual parameters meet the target parameter requirements, ensuring that the sonar signal is transmitted stably according to the adjusted parameters.

[0058] The underwater environmental data, including turbidity, water flow velocity, water temperature, and corresponding optimal sonar signal parameters (intensity, frequency, pulse width), are collected. The specific processing method is as follows:

[0059] 1. Preprocess the original dataset, first removing missing values ​​and obvious outliers. For missing values, imputation can be done using the mean, median, or interpolation based on neighboring data; for outliers, they can be identified and removed using box plots, Z-scores, etc. Then, standardize the preprocessed data, transforming environmental and signal parameters of different magnitudes to the same magnitude range (e.g., [0,1] interval) to eliminate the impact of data magnitude differences on clustering results. The standardization formula is: x1=(x-min(x)) / max(x)-min(x); where x is the original data, min(x) and max(x) are the minimum and maximum values ​​of the feature data, respectively, and x1 is the standardized data.

[0060] 2. Analyze the preprocessed sonar signal parameters, select K data points as cluster centers from the acquired data, and then calculate the Euclidean distance between each data point and each cluster center. Next, assign each data point to the cluster containing the nearest cluster center. Based on the data points in each cluster, recalculate the cluster center of that cluster and obtain the mean of the data points. Repeat the above process until the change in cluster centers is less than a preset threshold or the maximum number of iterations is reached. At this point, the clustering results converge, yielding K stable clusters. Each cluster represents a specific underwater environmental characteristic (turbidity, water flow velocity, water temperature) and the corresponding optimal sonar signal parameters (intensity, frequency, pulse width).

[0061] 3. Calculate the Euclidean distance between the real-time environmental parameter data and each cluster center, find the nearest cluster center, and assign the real-time environmental parameter data to the corresponding cluster. Determine the cluster category to which the current underwater environment belongs. The optimal sonar signal parameters corresponding to this cluster category are the reference adjustment parameters for the current environment. Based on the degree of difference between the real-time environmental parameters and the cluster center, fine-tune the reference adjustment parameters to obtain the final sonar signal adjustment parameters. Obtaining the final sonar signal adjustment parameters through the above steps improves the accuracy of the sonar signal data, thereby improving the accuracy of sonar signal adjustment.

[0062] Implementation of the Data Processing and Analysis System: Installation and Configuration of the High-Performance Data Acquisition Card: The high-performance data acquisition card is installed in the data processing unit of the underwater robot and connected to the signal output port of the sonar probe. The high-performance data acquisition card selected is the NIPCIe-6363 (sampling rate 400kHz, cache 16GB). The deep learning model is trained based on the PyTorch framework, and the disease trend prediction module uses TensorFlow to build an LSTM network. Based on the characteristics of the sonar signal and data processing requirements, the data acquisition card is configured, setting parameters such as sampling rate and cache size. This ensures that the data acquisition card can quickly and accurately acquire the sonar reflected signal and convert it into a digital signal for transmission to the data processing unit.

[0063] Training and Application of Deep Learning Data Processing Algorithms: A large number of sonar data samples containing different types of tunnel lining defects were collected and labeled to establish a training dataset. A Convolutional Neural Network (CNN) model was used as a deep learning framework. The model was trained using the training dataset, and its parameters were continuously adjusted to enable accurate identification and analysis of defects. During actual detection, the collected sonar data was input into the trained deep learning model, and the model output the defect identification results and related parameters, such as defect type, size, and location.

[0064] Specifically, for disease types, clear category labels such as cracks, spalling, and voids are used to ensure that the disease type label for each sample is accurate and unique. For disease location, bounding box annotation is used, with the pixel coordinates of the sonar image as the reference, recording the coordinates (x1, y1, x2, y2) of the upper left and lower right corners of the bounding box to accurately define the disease area. For disease size, the pixel size is converted to the actual physical size (e.g., millimeters and above) according to the scale of the sonar image, and specific parameters such as the length, width, and depth of cracks, and the area and thickness of voids are labeled. Professional annotation tools (such as LabelImg, VGGImageAnnotator, etc.) are used during the annotation process to ensure the accuracy and consistency of the annotation. At the same time, the quality of the annotated data is checked, and a certain proportion of samples are randomly selected for cross-verification by multiple professionals to remove samples with incorrect or blurry annotations. For sonar image data, the size is uniformly adjusted to a fixed size (e.g., 416×416 pixels), and letterbox scaling is used to maintain the aspect ratio of the image to avoid image distortion affecting the model's extraction of disease features. In the data augmentation stage, new enhancement methods targeting disease characteristics were added, such as random-angle rotation enhancement of crack samples to simulate cracks in different directions; and brightness and contrast adjustments to void samples to simulate sonar images under different underwater lighting conditions. When partitioning the data, it was ensured that the distribution ratio of various diseases in the training, validation, and test sets was consistent to avoid model training bias towards certain diseases due to uneven data distribution. Stratified sampling can be used for data partitioning to ensure that each subset represents the characteristics of the overall dataset.

[0065] YOLOv5 was chosen as the base network architecture and customized to meet the needs of tunnel lining defect detection. The YOLOv5 backbone feature extraction network, such as CSPDarknet, was retained, which effectively extracts multi-scale defect features from sonar images through a cross-stage local network structure. In the neck network, the depth of feature fusion was increased, and a PANet structure was adopted to enhance the information interaction between feature maps of different scales, enabling the model to better identify defects of different sizes (especially micro-cracks). In the head network, a multi-task output layer was designed to output the defect type probability, bounding box coordinate offset, and defect size parameters, respectively. Specifically, the type output layer uses a softmax activation function to output the probability distribution of each category; the bounding box output layer predicts the coordinate offset of the bounding box relative to the anchor box; and the size output layer uses a linear activation function to directly predict the actual physical size parameters of the defect. Based on the size distribution of defect bounding boxes in the dataset, a K-means clustering algorithm was used to generate suitable anchor box sizes. The width and height of all bounding boxes in the training set are statistically analyzed. Using the intersection-over-union (IoU) ratio of the bounding boxes as the distance metric, nine anchor boxes of different sizes (e.g., [10,13], [16,30], [33,23], [30,61], [62,45], [59,119], [116,90], [156,198], [373,326]) are clustered and assigned to different feature scale layers to improve the model's localization accuracy for diseases of different sizes.

[0066] For sonar image feature and disease detection tasks, the network parameter initialization method was adjusted. For the convolutional layers of the backbone network, pre-trained weight parameters (based on YOLOv5 weights trained on the COCO dataset) were used for initialization, leveraging the general feature extraction capabilities learned by the pre-trained model to accelerate model convergence. For the newly added feature fusion layer and head output layer, the He initialization method was adopted to ensure a reasonable distribution of initial parameter values ​​and avoid gradient vanishing or exploding problems. When setting the initial hyperparameter values, the batch size was set to 16 or 32 based on the GPU memory capacity, the initial number of training epochs was set to 100, the initial learning rate was set to 0.01, the momentum parameter was set to 0.937, and the weight decay coefficient was set to 0.0005.

[0067] The specific training process is as follows: A multi-task loss function is constructed to comprehensively optimize the tasks of disease type identification, location localization, and size measurement. The type loss uses the cross-entropy loss function:

[0068] ;

[0069] Where yi is the true class label and pi is the class probability predicted by the model. The location loss uses the CIoU loss function, considering the overlap of the bounding boxes, the distance between the center points, and the aspect ratio:

[0070] ;

[0071] Improve positioning accuracy. Size loss is calculated using mean square error loss.

[0072] ;

[0073] Where sj is the actual size parameter. j represents the predicted size parameters of the model, and M represents the number of size parameters.

[0074] The total loss function is: ;

[0075] Formula for adjusting weighting coefficients: Balance the training priorities of each task.

[0076] Five warm-up training rounds are conducted in the initial training phase, with the learning rate linearly increasing from 1 / 1000 of the initial value to the initial learning rate to avoid model parameter oscillations caused by an excessively large initial learning rate. A small batch size (e.g., 8) is used during the warm-up phase to allow the model to gradually adapt to the data distribution. After the warm-up training, the formal iterative training phase begins. In each training round, the training set is input into the model in batches, forward propagation is used to calculate the predicted output, and the total loss is calculated using the loss function. The backpropagation algorithm and the Adam optimizer are used to update the network parameters, with the optimizer's β1 parameter set to 0.9, β2 parameter set to 0.999, and epsilon parameter set to 1e-8. Dynamic learning rate adjustment: A cosine annealing learning rate scheduling strategy is used to automatically adjust the learning rate during training. When the validation set loss decreases slowly, the learning rate decays according to a cosine function, and the learning rate is reduced in the later stages of training to fine-tune the model parameters. Simultaneously, a minimum learning rate of 1e-6 is set to prevent the model from stagnating due to an excessively low learning rate.

[0077] After each training round, the model is evaluated on the validation set, calculating the recognition accuracy, recall, mAP@0.5 (average precision at an IoU threshold of 0.5), and mean absolute error (MAE) for dimensional measurements for various diseases. Loss curves for the training and validation sets are plotted to monitor the model's training status in real time. If mAP@0.5 on the validation set does not improve for 15 consecutive rounds, an early stopping strategy is triggered, halting training and saving the current optimal model parameters (i.e., the model weights at which mAP@0.5 on the validation set is highest). The saved model file includes the network structure configuration and weight parameters for subsequent deployment and inference.

[0078] Operation of the real-time data display and storage module: Connect the real-time data display and storage module to the data processing unit, and install the corresponding display software on the control terminal. After the data processing unit processes the sonar data, it transmits the processing results to the real-time data display and storage module. The display module displays the data in real-time on the control terminal in the form of 3D images, annotation information, etc., allowing operators to view the detection status in real time through the control terminal. Simultaneously, the storage module stores all detection data according to preset storage formats and paths for subsequent retrieval and analysis.

[0079] Implementation of the Underwater Robot Mounting and Operation System: Construction and Testing of a Stable Underwater Robot: Based on the underwater robot's design requirements, a stable underwater robot was constructed using high-strength, corrosion-resistant materials. The underwater robot employs a composite structure of a titanium alloy frame and fiberglass panels. During construction, the platform's dimensional accuracy and structural strength were strictly controlled to ensure it met the installation and operational requirements of the sonar detection device. After construction, the underwater robot underwent testing, including stability tests under different water flow velocities and underwater environments, and adjustments to the sonar detection device's installation position and angle. This ensured the underwater robot provided a stable and reliable working environment for the sonar detection device. By mounting the detection device on the underwater robot, detection could be completed underwater without emptying the tunnel, perfectly matching the actual operating conditions of the tunnel, avoiding structural damage caused by emptying, and ensuring the continuity of the project operation.

[0080] Calibration and Operation of the Intelligent Navigation and Positioning System: Before the underwater robot is launched, the intelligent navigation and positioning system is calibrated, including the calibration of the underwater acoustic positioning system (such as Teledyne RDI WorkHorse), the initial parameter setting of the inertial navigation system (such as VectorNav VN-100), and the feature point recognition training of the visual positioning system. After calibration, the navigation and positioning system is connected to the underwater robot's control system, and the underwater robot's travel path and navigation parameters are preset according to the actual conditions of the tunnel and the inspection requirements. During the inspection process, the navigation and positioning system monitors the underwater robot's position and attitude in real time, automatically adjusting its speed and direction according to the preset path to ensure that the sonar inspection device can efficiently inspect the tunnel lining. Simultaneously, the underwater robot's position and attitude information is transmitted to the control terminal in real time, facilitating monitoring and intervention by operators. The underwater robot autonomously completes omnidirectional scanning according to the preset path, resulting in high inspection efficiency. No personnel are required to enter the tunnel for operation, completely avoiding safety risks such as height, lack of oxygen, and collapse, significantly reducing operational safety costs.

[0081] Implementation of the emergency obstacle avoidance module: The emergency obstacle avoidance module uses a lidar (such as Velodyne VLP-16) and a high-definition underwater camera (such as SeaLife Micro 3.0).

[0082] Maintenance and management of the energy supply and communication systems: Before each operation of the underwater robot, check the lithium battery pack power of the energy supply system to ensure sufficient power. Simultaneously, check the fiber optic and underwater acoustic communication equipment of the communication system to ensure they are functioning properly, including testing parameters such as signal strength and transmission rate. During operation, monitor the power consumption of the energy supply system in real time. When the power level falls below a preset threshold, promptly notify the operator to take appropriate measures, such as replacing the battery or returning to the charging station. For the communication system, continuously monitor the stability and reliability of data transmission. In the event of a communication failure, automatically switch to the backup communication link and promptly troubleshoot and repair the problem to ensure uninterrupted communication between the underwater robot and the control terminal.

[0083] Sonar detection system workflow:

[0084] After the underwater robot is launched, the sonar signal noise reduction module first starts a self-test, samples and analyzes the received environmental noise, and sets an initial noise reduction threshold.

[0085] The three-dimensional sonar scanning module rotates 360 degrees horizontally and 180 degrees vertically according to the preset path and is driven by a motor. The sonar probe moves in 0.2m steps to complete the all-round scanning of the tunnel lining.

[0086] The adaptive sonar signal conditioning module collects environmental parameters in real time through turbidity sensors, flow rate sensors, and temperature sensors, compares them with a parameter matching library, quickly determines the initial signal parameters, and then dynamically fine-tunes them based on signal feedback, with an adjustment cycle of ≤0.05s.

[0087] The sonar received signal is processed by the noise reduction module: it uses a db4 wavelet basis for 3-level decomposition, applies hard thresholding to high-frequency noise components, retains low-frequency signal components, and outputs a clean signal after reconstruction, improving the signal-to-noise ratio to over 45dB.

[0088] Data processing and analysis system workflow:

[0089] The high-performance data acquisition card acquires the noise-reduced sonar signal at a sampling rate of 400kHz, stores it in the local cache, and transmits it to the data processing unit at the same time.

[0090] The deep learning data processing algorithm module receives digital signals, inputs them into the trained improved YOLOv5 model, identifies the type of disease (cracks, spalling, voids), calculates the size of the disease (e.g., crack length is calculated using pixel conversion + geometric correction, with an error ≤ ±0.1mm), the three-dimensional position coordinates (combined with navigation system positioning data), and classifies the severity level (Level I-IV) according to industry standards.

[0091] The real-time data display and storage module displays the processing results in the form of 3D images (using the VTK visualization library), disease annotations (different colors distinguish disease types), and parameter lists (size, location, grade), etc., on the control terminal, and stores them to the local hard drive and cloud server, supporting real-time data export and historical query;

[0092] The disease trend prediction module calls the LSTM model, inputs the data from the last three inspections (if it is the first inspection, historical data under similar working conditions is used) and the current environmental parameters, predicts the disease expansion in the next 6-12 months, and outputs trend curves (such as crack length growth curves) and maintenance suggestions (such as re-inspecting Class I diseases every 6 months and repairing Class III diseases immediately).

[0093] Underwater robot mounting and operation system workflow:

[0094] In environments with water flow speeds ≤3m / s, the stable underwater robot maintains a horizontal position through a center of gravity adjustment device. The sonar mounting bracket electrically fine-tunes the angle (±3°) and position (±5cm) of the sonar probe according to the detection requirements to ensure the optimal detection angle.

[0095] The intelligent navigation and positioning system integrates underwater acoustic positioning, inertial navigation, and visual positioning data to calculate the robot's position in real time with an error of ≤±5cm. It automatically travels according to a preset path at a speed of 0.5m / s. If an uncovered area is detected, the path is dynamically corrected.

[0096] The energy supply system monitors the lithium battery level in real time. When the level drops below 20%, it sends a warning signal to the control terminal, prompting the robot to return. The communication system automatically switches communication modes based on the distance between the robot and the control terminal: fiber optic communication at a rate of 1Gbps is used when the distance is ≤500m; underwater acoustic communication at a rate of 150kbps is used when the distance is >500m to ensure continuous data transmission.

[0097] The emergency obstacle avoidance module monitors the environment ahead in real time: the lidar scans a range of 0.5-5m, and the visual camera assists in identifying obstacles. When an obstacle is detected and the distance between the robot and the obstacle is ≤1m, the obstacle avoidance algorithm calculates the optimal obstacle avoidance path. The robot's driving direction is adjusted by the control system (offset angle ≤15°) or decelerated to 0.1m / s. After obstacle avoidance is completed, the original path is restored. The obstacle avoidance response time is ≤0.1s.

[0098] Implementation verification

[0099] On-site verification was conducted in a hydraulic tunnel (5m in diameter, 3m in depth, 1.5m / s in flow velocity, and 25 NTU in turbidity).

[0100] Test results: Three 0.6mm cracks, two delamination layers (0.3cm thick), and one peeling area (0.2m²) were successfully identified. The accuracy of disease type identification was 100%. The size measurement error was ≤±0.1mm and the location positioning error was ≤±3cm.

[0101] Environmental adaptability: Under conditions where the turbidity of the water body suddenly increases to 40 NTU and the water flow velocity instantly reaches 2.8 m / s, the device still works stably, the signal is not distorted, and the detection results are reliable.

[0102] Operational safety: During the detection process, a 0.3m × 0.2m falling rock was encountered. The emergency obstacle avoidance module successfully identified and avoided the rock, and the robot did not collide with it.

[0103] Data usability: The real-time display is clear and intuitive, and the trend prediction curve matches the follow-up test results for the next 3 months with a consistency of ≥90%, providing effective guidance for engineering maintenance.

[0104] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0105] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A sonar detection device for tunnel lining defects mounted on an underwater robot, characterized in that: It includes a sonar detection system, a data processing and analysis system, and an underwater robot mounting and operation system; the sonar detection system includes a high-resolution phased array sonar probe, a three-dimensional sonar scanning module, an adaptive sonar signal conditioning module, and a sonar signal noise reduction module; the data processing and analysis system includes a high-performance data acquisition card, a deep learning data processing algorithm module, a real-time data display and storage module, and a disease trend prediction module; the underwater robot mounting and operation system includes a stable underwater robot, an intelligent navigation and positioning system, an energy supply and communication system, and an emergency obstacle avoidance module.

2. The sonar detection device for tunnel lining defects mounted on an underwater robot according to claim 1, characterized in that: The frequency of the high-resolution phased array sonar probe can be flexibly adjusted within the range of 100kHz-1MHz to detect tunnel lining cracks and thin voids of 0.8 mm or larger.

3. The sonar detection device for tunnel lining defects mounted on an underwater robot according to claim 2, characterized in that: The three-dimensional sonar scanning module is installed at the front end of the underwater robot and is driven by a motor to rotate 360 ​​degrees horizontally and 180 degrees vertically. The sonar probe moves according to a preset scanning path and step size.

4. The sonar detection device for tunnel lining defects mounted on an underwater robot according to claim 3, characterized in that: The adaptive sonar signal adjustment module automatically adjusts the intensity, frequency, and pulse width parameters of the sonar transmission signal in real time based on factors such as water turbidity, water flow velocity, and water temperature in the underwater environment, combined with a preset K-means clustering environment-parameter matching library.

5. The sonar detection device for tunnel lining defects mounted on an underwater robot according to claim 4, characterized in that: The sonar signal noise reduction module uses a db4 wavelet basis for 3-level decomposition, applies hard thresholding to high-frequency noise components, preserves and reconstructs low-frequency signal components, and outputs a signal-to-noise ratio ≥45dB.

6. The sonar detection device for tunnel lining defects mounted on an underwater robot according to claim 5, characterized in that: The deep learning data processing algorithm module is based on the improved YOLOv5 model. An attention mechanism module is built in the improved YOLOv5 model and it forms a collaborative working link with the original feature extraction network and detection head architecture of the model.

7. The sonar detection device for tunnel lining defects mounted on an underwater robot according to claim 6, characterized in that: The intelligent navigation and positioning system integrates underwater acoustic positioning, inertial navigation and visual positioning technologies. The underwater robot automatically travels according to the preset detection path and adjusts its speed and direction in real time, and has a dynamic path correction function.

8. The sonar detection device for tunnel lining defects mounted on an underwater robot according to claim 7, characterized in that: The disease trend prediction module is based on the LSTM time series analysis model. It takes historical detection data and real-time environmental parameters as input, predicts the spread rate, development direction and risk level of diseases in the next 6-12 months, and outputs trend prediction curves and maintenance suggestions.

9. A sonar detection device for tunnel lining defects mounted on an underwater robot according to claim 8, characterized in that: The emergency obstacle avoidance module consists of a lidar sensor, a vision camera, and an obstacle avoidance algorithm. It monitors obstacles in front of the robot in real time and achieves emergency obstacle avoidance by adjusting the driving direction or slowing down. After the obstacle avoidance is completed, it automatically restores the original detection path.

Images