A concrete surface automatic inspection robot and method in a narrow space tunnel

Abstract

The application belongs to the technical field of tunnel engineering, and particularly relates to a kind of concrete surface automatic inspection robot and method in narrow space tunnel, the scheme completely abandons the traditional method of "uniform scanning, full collection", adopts the two-stage collaborative mechanism of "preliminary identification" and "key diagnosis" driven by intelligence, realizes structural reduction at data source, and theoretically can release more than 90% of scanning resources and data processing bandwidth from "invalid area", and redistribute them to high-value areas; at the same time, through dynamic calculation of interest threshold, the system can adapt to the differences of inspection environment and the slight differences of disease characteristics, not only solve the fundamental problem of data redundancy and storage and transmission pressure, but also ensure the attention to subtle and early diseases, fundamentally improve the detection probability of hidden dangers and the overall efficiency of the perception subsystem.

Description

Technical Field

[0001] This invention belongs to the field of tunnel engineering technology, specifically relating to an automatic inspection robot and method for concrete surfaces in narrow tunnels. Background Technology

[0002] Regular inspection of tunnel concrete surfaces is crucial for ensuring tunnel structural safety. Currently, both domestically and internationally, the inspection of concrete surfaces in narrow tunnel spaces mainly employs robotic inspection methods. While these methods have achieved initial autonomous navigation and inspection, the following problems still exist:

[0003] Currently, inspection robots generally adopt a fixed perception mode of "uniform scanning and full collection", which means that the same amount and precision of data are collected on the tunnel surface regardless of its condition or integrity. This data collection strategy will generate a lot of structural data redundancy. About 95% of the perception resources are consumed on the intact surface, while the high-value areas with real hidden dangers can only be allocated a limited share of resources. This not only directly reduces the detection rate of minor and early defects, but also generates a huge amount of invalid data storage and transmission pressure, resulting in an efficiency dilemma at the data perception level.

[0004] Traditional inspection robots adopt an open-loop architecture of "perception-transmission-cloud processing". Typically, the detection data is transmitted from the on-site collection point to a remote server through a communication link with limited bandwidth (especially in the narrow space of a tunnel). It takes several hours or even days of offline analysis to produce results, resulting in extremely poor rapid early warning capability of automatic inspection.

[0005] Existing technical solutions typically treat mobile, sensing, computing, and communication modules as independent subsystems, performing static planning before the task, which results in their behavior patterns being unable to adjust to the actual situation found on-site during inspections. Summary of the Invention

[0006] This invention addresses the problems existing in the prior art by providing an automatic inspection robot and method for concrete surfaces in narrow tunnels, effectively solving the problems of the prior art.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] An automated inspection robot for concrete surfaces in narrow tunnels includes a mobile platform, sensing devices mounted on the mobile platform, a computing unit, and a communication unit, wherein:

[0009] The sensing device includes a wide-area scanning unit, a fine detection unit, and a positioning unit; the computing unit adopts a local heterogeneous computing processing mechanism.

[0010] The computing unit specifically performs the following processing steps: During the inspection robot's movement along the tunnel axis, a sparse 3D map of the tunnel environment is constructed online based on synchronously acquired wide-area scanning data and positioning data. Real-time preliminary feature calculations are performed based on the acquired wide-area scanning data within a preset range to obtain the basic features of each analysis unit within the preset range. The wide-area scanning unit continues to operate during the inspection robot's movement along the tunnel axis. The real-time interest degree of each analysis unit is calculated based on the basic features, and a set of high-interest regions is determined based on the real-time interest degree. For the analysis units within the high-interest region set, local refined real-time detection is performed. In this local refined real-time detection, for each analysis unit, one or more matching refined sensors are activated to collect data based on the matching degree between the analysis unit and the sensors within the refined detection unit. Multi-scale feature extraction is performed on the collected data, and the detection result of the current analysis unit is obtained through a pre-built lightweight multi-task anomaly detection model based on deep learning.

[0011] Preferably, the calculation of the real-time interest degree of each analysis unit based on the basic features is specifically represented as follows:

[0012]

[0013] in, To analyze the real-time interest level of unit c, The sigmoid function is used to normalize numerical values. For geometric anomaly, For texture anomaly degree, , For learnable weights, , This is a scale parameter used to normalize different feature dimensions.

[0014] Preferably, determining the set of high-interest regions based on the real-time interest level specifically involves:

[0015]

[0016] in, A set of regions of high interest. Indicates the analysis unit, The interest threshold;

[0017] The interest threshold Dynamic settings are used, as shown below:

[0018]

[0019] in, All within the current sliding window The mean, All within the current sliding window standard deviation This is the sensitivity coefficient.

[0020] Preferably, the step of activating one or more matching fine sensors to acquire data based on the matching degree between the unit to be analyzed and the sensors in the fine detection unit specifically includes the following processing steps:

[0021] For each unit to be analyzed in the set of regions of high interest, calculate the expected utility, quality factor and resource consumption of each sensor in the fine detection unit for the range of interest suggested by the unit to be analyzed.

[0022] Priority calculation is performed based on the expected utility, quality factor, and resource consumption.

[0023] Based on the obtained priority, each unit to be analyzed is assigned one or a group of sensors with the highest priority for its refined real-time detection.

[0024] Preferably, the priority is calculated as follows:

[0025]

[0026] )

[0027] in, For the expected utility, Let be the quality factor, representing the calibration value of the j-th sensor's detection capability for typical diseases in the current environment. Resource consumption indicates that the j-th sensor is activated and the unit to be analyzed is started. The normalized value of the overall cost of conducting the test. Let i be the interest level of the i-th unit to be analyzed. Let be the area of ​​the region of the i-th unit to be analyzed. Let be the estimated time taken for the j-th sensor to collect data from the i-th unit under analysis in a single acquisition. Let be the matching degree between the historical disease type of the i-th unit to be analyzed and the j-th sensor. To adjust the parameters.

[0028] Preferably, the lightweight multi-task anomaly detection model is deployed on the local edge computing unit of the inspection robot and has the ability to co-evolve with the cloud. The inspection robot uploads the local low-confidence detection sample data to the cloud. The cloud processes the samples through a high-precision teacher model to obtain high-quality labels and periodically updates the optimized model parameters to the edge computing unit via OTA through knowledge distillation technology.

[0029] Preferably, the wide-area scanning unit includes a line lidar and a panoramic vision camera; the fine detection unit includes a high-definition zoom camera, an infrared thermal imager, and an acoustic sensor; the computing unit adopts a local heterogeneous computing processing mechanism, specifically including a GPU-based edge computing module and a real-time microcontroller-based control module; and the positioning unit adopts an IMU + odometry + visual SLAM structure design.

[0030] An automated inspection method for concrete surfaces in narrow-space tunnels, based on the aforementioned automated inspection robot for concrete surfaces in narrow-space tunnels, includes:

[0031] As the inspection robot travels along the tunnel axis, a sparse 3D map of the tunnel environment is constructed online based on the synchronously acquired wide-area scanning data and positioning data.

[0032] Based on the wide-area scanning data obtained within the preset range, real-time preliminary feature calculations are performed to obtain the basic features of each analysis unit within the preset range. The wide-area scanning unit works continuously as the inspection robot travels along the tunnel axis.

[0033] The real-time interest degree of each analysis unit is calculated based on the aforementioned basic features, and a set of high-interest regions is determined based on the real-time interest degree.

[0034] For the units to be analyzed in the set of regions of high interest, local refined real-time detection is performed. In the local refined real-time detection, for each unit to be analyzed, one or more refined sensors that match the unit to be analyzed are activated to collect data based on the matching degree between the unit to be analyzed and the sensors in the refined detection unit. Multi-scale feature extraction is performed on the collected data, and the detection result of the current unit is obtained through a pre-built lightweight multi-task anomaly detection model based on deep learning.

[0035] An electronic device includes a memory, a processor, and a computer program stored in the memory and running on it, characterized in that the processor executes the program to implement an automatic inspection method for concrete surfaces in a narrow space tunnel.

[0036] A non-transitory computer-readable storage medium storing a computer program that, when executed by a processor, implements an automatic inspection method for concrete surfaces in a narrow-space tunnel.

[0037] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0038] (1) This invention provides an automatic inspection robot and method for concrete surfaces in narrow tunnels. The solution completely abandons the traditional approach of "uniform scanning and full data acquisition" and adopts a collaborative mechanism of "preliminary identification" and "key diagnosis" driven by intelligence. First, the inspection robot uses a wide-area scanning unit and real-time preliminary feature calculation to screen out potential abnormal areas in real time, transforming the "area-like" perception pressure of the entire inspection surface into a "point-like" fine diagnosis task for a small number of suspicious areas. The above strategy achieves structural reduction at the data source, which can theoretically release more than 90% of the scanning resources and data processing bandwidth from the "invalid area" and redistribute them to high-value areas. At the same time, through the dynamic calculation of the interest threshold, the system can adapt to the differences in the inspection environment and the slight differences in the characteristics of the defects. This not only solves the fundamental problems of data redundancy and storage and transmission pressure, but also ensures attention to subtle and early defects, fundamentally improving the detection probability of hidden dangers and the overall efficiency of the perception subsystem.

[0039] (2) In the solution described in this invention, the forward inference of the multi-scale feature extraction and lightweight multi-task anomaly detection model is directly executed by the local heterogeneous computing unit carried by the inspection robot. This architecture enables the entire decision-making process from sensor signal acquisition to anomaly judgment (classification, localization and parameter quantization) to be completed within seconds or even sub-seconds. Once a high-risk disease is detected, an alarm can be issued quickly on-site without waiting for data to be transmitted remotely or back to the cloud. This reduces the early warning delay from "hours / days" to "real-time", significantly enhancing the system's ability to quickly warn and respond to emergencies.

[0040] (3) The solution described in this invention successfully solves the problem of rigid allocation of sensing resources and inability to dynamically adjust with the environment in traditional inspections through an innovative sensor scheduling strategy. The system no longer pre-allocates the same sensor combination to all areas, but dynamically schedules the most suitable sensor combination according to the specific characteristics of each candidate high-interest area. This "area-adaptive" sensor scheduling concept enables limited sensor resources to be accurately configured according to the unique needs of different areas.

[0041] (4) The solution described in this invention uploads local low-confidence samples to the cloud, generates high-quality labels using a high-precision “teacher model”, and regularly sends the optimized model parameters to the entire inspection cluster via OTA through knowledge distillation technology. This not only improves the diagnostic capabilities of a single inspection robot, but also achieves continuous iteration of diagnostic accuracy and experience sharing of implicit knowledge. Attached Figure Description

[0042] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below:

[0043] Figure 1 This is a schematic diagram of the basic structure of an automatic inspection robot for concrete surfaces in a narrow tunnel, as described in an embodiment of the present invention.

[0044] Figure 2 This is a schematic diagram of the startup process of the precision sensor described in an embodiment of the present invention;

[0045] Figure 3 This is a flowchart of an automatic inspection method for concrete surfaces in a narrow space tunnel, as described in an embodiment of the present invention.

[0046] Figure 4 This is a schematic diagram of an electronic device structure according to an embodiment of the present invention. Detailed Implementation

[0047] To better understand the above-mentioned objectives, features and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0048] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0049] Example 1

[0050] The following is a detailed description of an automatic inspection robot for concrete surfaces in a narrow tunnel, as described in Embodiment 1, with reference to the accompanying drawings.

[0051] like Figure 1 As shown, an automated inspection robot for concrete surfaces in a narrow tunnel includes a mobile platform, sensing devices mounted on the mobile platform, a computing unit, and a communication unit, wherein:

[0052] The sensing device includes a wide-area scanning unit, a fine detection unit, and a positioning unit; the computing unit adopts a local heterogeneous computing processing mechanism.

[0053] The computing unit specifically performs the following processing steps: During the inspection robot's movement along the tunnel axis, a sparse 3D map of the tunnel environment is constructed online based on synchronously acquired wide-area scanning data and positioning data. Real-time preliminary feature calculations are performed based on the acquired wide-area scanning data within a preset range to obtain the basic features of each analysis unit within the preset range. The wide-area scanning unit continues to operate during the inspection robot's movement along the tunnel axis. The real-time interest degree of each analysis unit is calculated based on the basic features, and a set of high-interest regions is determined based on the real-time interest degree. For the analysis units within the high-interest region set, local refined real-time detection is performed. In this local refined real-time detection, for each analysis unit, one or more matching refined sensors are activated to collect data based on the matching degree between the analysis unit and the sensors within the refined detection unit. Multi-scale feature extraction is performed on the collected data, and the detection result of the current analysis unit is obtained through a pre-built lightweight multi-task anomaly detection model based on deep learning.

[0054] In specific implementation, the step of performing real-time preliminary feature calculations based on the wide-area scanning data within a preset range to obtain the basic features of each analysis unit within the preset range specifically involves:

[0055] Based on the wide-area scanning data of the current sliding window (e.g., within 10 meters in front of the inspection robot), the data is divided into several analysis units through grid partitioning.

[0056] Based on the wide-area scan data of each analysis unit, the calculation of basic features is performed, including:

[0057] First, define the analysis unit c: divide the surface within the current sliding window into M x N units, with each unit c corresponding to a local region;

[0058] Secondly, the geometric anomaly and texture anomaly in the basic features are calculated, where:

[0059] LiDAR point cloud and panoramic camera images As a data source;

[0060] The geometric anomaly is specifically represented as follows:

[0061]

[0062] in, For surface curvature, For deviations from the Dharma aspect, For point density variation, , , These are the weighting coefficients for different items, which are set to 0.4, 0.3, and 0.3 respectively in this embodiment.

[0063] Specifically:

[0064]

[0065]

[0066]

[0067] in, For the subset of laser point cloud within analysis unit c, Let i be the average curvature of point i. Let be the first principal curvature at point i (i.e., the curvature in the direction of maximum bending). Let be the second principal curvature at point i (i.e., the curvature in the direction of minimum bending). , The unit weighted average normal vector For the ideal tunnel wall normal vector, The pitch angle of the analysis unit is given; by subdividing the analysis unit c into several sub-cells, the average point density is obtained based on the obtained sub-cell point density. and point density standard deviation .

[0068] Texture anomaly:

[0069] Texture anomaly This is used to measure the degree of deviation between the regional surface texture pattern and the expected behavior of healthy barrier concrete, specifically expressed as follows:

[0070]

[0071] in, For gradient consistency analysis terms, For texture structure entropy, This is an LBP mode exception. , , These are the weighting coefficients for different items. In this embodiment, , , Set them to 0.45, 0.35, and 0.2 respectively;

[0072] Gradient consistency analysis terms:

[0073]

[0074] in, To analyze the grayscale image intensity at position (u, v) in image unit c, Represents the image intensity gradient vector. Indicates the gradient magnitude. This represents the average gradient magnitude of analysis unit c. H represents the width and height of the image; Let represent the effective pixel mask at position (u, v), where:

[0075] Condition 1: in, The gradient noise threshold;

[0076] Condition 2: Judgment of Illumination Uniformity: in, The threshold for uneven illumination. To analyze the intensity variance within the image block of unit c;

[0077] When the above two conditions are met It is 1 if it is true, otherwise it is 0.

[0078] Texture structure entropy term:

[0079]

[0080] in, is the number of buckets in the histogram (e.g., dividing the grayscale range 0-255 into 32 equal-width buckets), and b is the bucket index. Let be the probability mass function. , This represents the number of pixels that fall into bucket b. () is the base-2 logarithm, used to calculate discrete entropy. This is a numerical stability parameter.

[0081] LBP mode anomalies:

[0082]

[0083] in, Let k be the number of sub-blocks in the image of analysis unit c, and k be the sub-block index. The LBP pattern histogram for the k-th sub-block; This is an LBP histogram template; For reference, a collection of histogram templates, , For healthy concrete histogram templates, For noise template, This is a shadow template.

[0084] In specific implementation, the calculation of the real-time interest degree of each analysis unit based on the basic features is as follows:

[0085]

[0086] in, To analyze the real-time interest level of unit c, The sigmoid function is used to normalize numerical values. , For learnable weights, , This is a scale parameter used to normalize different feature dimensions.

[0087] In one or more implementations, a set of high-interest regions is generated based on the real-time interest level of each analysis unit. Specifically:

[0088]

[0089] in, Indicates the analysis unit, The interest threshold is defined as follows: Dynamic settings are used, as shown below:

[0090]

[0091] in, All within the current sliding window The mean, All within the current sliding window standard deviation This is the sensitivity coefficient.

[0092] In practical implementation, for sets of high-interest regions Each cell to be analyzed in the dataset is associated with the following attributes:

[0093] (1) Level of interest ;

[0094] (2) Three-dimensional spatial position: precise coordinates determined based on SLAM positioning;

[0095] (3) Recommended scope of focus: Define a cubic region with a preset side length (e.g., 0.1m) centered on the unit to be analyzed. This scope will guide the determination of the sensor’s focusing and scanning range during fine detection.

[0096] In specific implementation, for the units to be analyzed within the set of regions of high interest, local refined real-time detection is performed. In this local refined real-time detection, for each unit to be analyzed, based on the matching degree between the unit to be analyzed and the sensors within the refined detection unit, one or more matching refined sensors are activated to collect data. Multi-scale feature extraction is performed on the collected data, and the detection result of the current unit is obtained through a pre-built lightweight multi-task anomaly detection model based on deep learning. Specifically:

[0097] like Figure 2 As shown, the process of activating one or more matching fine sensors to acquire data based on the matching degree between the unit to be analyzed and the sensors in the fine detection unit specifically includes the following steps:

[0098] Step 1: For each unit to be analyzed in the set of regions of high interest The calculation of each sensor (including a high-definition zoom camera, an infrared thermal imager, and an acoustic sensor) within the fine detection unit is performed on the unit to be analyzed. The expected utility of the suggested area of ​​focus quality factor and resource consumption ;in:

[0099] Expected utility Specifically, it is expressed as follows:

[0100] )

[0101] in, Unit to be analyzed interest Unit to be analyzed The area of ​​the region, For sensors Unit of analysis Estimated time for a single data collection session. Unit to be analyzed Historical disease types and sensors The degree of matching.

[0102] quality factor For sensors The calibration value for the detection capability of typical diseases in the current environment is in the range of [0, 1], and can be set according to expert evaluation.

[0103] Resource consumption To activate the sensor and the unit to be analyzed The normalized value of the total cost of conducting the test (including but not limited to estimates of energy consumption, CPU or GPU usage, and time spent);

[0104] Step 2: Based on the expected utility obtained quality factor and resource consumption Priority calculation is performed, as shown below:

[0105]

[0106] Step 3: Based on the obtained priority, analyze each unit to be analyzed. distribute The largest one or a group of sensors is used for its refined real-time detection.

[0107] In one or more implementations, based on the obtained set of regions of high interest, each unit to be analyzed is sequentially executed according to its positional relationship and longitudinal position in the direction of travel and longitudinal position of the inspection robot, in ascending order of both. Refined real-time detection, wherein when any unit to be analyzed is detected... When, select based on its priority calculation result. The largest sensor or a group of sensors acquires data, and multi-scale feature extraction is performed on the acquired data. Then, the detection results of the current analysis unit are obtained through a pre-built lightweight multi-task anomaly detection model based on deep learning.

[0108] In specific implementation, the multi-scale feature extraction based on the collected data specifically includes:

[0109] Based on one or more selected sensors, high-resolution, multimodal data acquisition is performed on the unit to be analyzed;

[0110] Based on the collected data, geometric anomalies, texture anomalies, and thermal radiation anomalies were calculated at different scales; among which:

[0111] The geometric anomalies at different scales are specifically as follows:

[0112] Based on high-quality point clouds of multi-scale neighborhoods of the unit to be analyzed (e.g., 0.05m, 0.1m, and 0.2m ranges centered on the unit to be analyzed), the geometric anomalies are calculated separately, and the combined anomaly after fusion is obtained by weighted summation; specifically, the calculation of the geometric anomalies adopts the calculation formula in the real-time preliminary feature calculation process.

[0113] The texture anomaly degree is calculated based on the high-resolution image, specifically using the calculation formula in the real-time preliminary feature calculation process;

[0114] The thermal radiation anomaly was obtained based on measurements taken with an infrared thermal imager.

[0115] It should be noted that the above description is based on the case where the unit to be analyzed has selected a high-definition zoom camera and an infrared thermal imager, and is combined with the point cloud data obtained by the wide-area scanning unit. Other options can be added, removed or reduced according to the actual situation.

[0116] In specific implementation, the detection result of the current analysis unit is obtained by using a pre-constructed lightweight multi-task anomaly detection model based on deep learning. Specifically, the obtained multi-scale features are used as input, and the MobileNet-SSD model is used to obtain the detection result.

[0117] The MobileNet-SSD model uses MobileNet as the backbone network and SSD as the detection head. By taking geometric anomalies, texture anomalies, and thermal radiation anomalies as inputs to the MobileNet-SSD model, it outputs six feature maps at different scales. The final detection results are obtained by taking these feature maps at different scales as inputs to the SSD model.

[0118] The MobileNet-SSD model described above is a simple combination of two existing models. MobileNet is from the paper: Howard AG, Zhu M, Chen B, et al. MobileNets: Efficient Convolutional NeuralNetworks for Mobile Vision Applications[J]. 2017.DOI:10.48550 / arXiv.1704.04861; SSD model is from the paper: [1] Wei L, Dragomir A, Dumitru E, et al. SSD: Single Shot MultiBox Detector[J]. Springer, Cham, 2016.DOI:10.1007 / 978-3-319-46448-0_2. It will not be elaborated here.

[0119] In one or more embodiments, the lightweight multi-task anomaly detection model of the inspection robot has continuous self-evolution capabilities, and its core evolution mechanism adopts a collaborative framework of edge computing-cloud training-secure deployment. When the robot's detection confidence of a certain unit to be detected is lower than a preset threshold, it will package the original multimodal data, context information, and intermediate inference features of the existing model for that area and temporarily cache them locally. These uncertain samples will be asynchronously uploaded to the cloud analysis center when communication conditions permit (e.g., when the robot is near the inspection base station or during a rest period between tasks).

[0120] At the cloud-based analytics center, these uncertain samples are refined using more sophisticated models. For example, more complex "teacher models" with larger parameter scales are deployed in the cloud (e.g., a hybrid architecture based on EfficientNet-B4 and Point-Transformer), or manual review and annotation are performed with the assistance of experts. This generates a highly reliable "real" diagnostic label (including disease category, precise boundaries, and quantitative parameters) for each ambiguous sample. At the same time, a continuous and incremental sample-expert label database can be established in the cloud.

[0121] In practice, the cloud-based analytics center periodically (e.g., monthly or quarterly) utilizes the aforementioned database to incrementally optimize the edge model through knowledge distillation technology. For example, using a cloud-based teacher model as a knowledge source, a softened loss function with a training temperature greater than 1 is employed to guide the training of the student model. This maximizes the absorption of the discriminative capabilities of complex models while strictly maintaining the model's computational load and latency constraints. After offline verification of the new version of the edge model, it will be deployed to all inspection robots via OTA upgrades, thereby achieving the co-evolution of the detection capabilities of all inspection robots without downtime or manual intervention.

[0122] In practice, the mobile platform adopts a tracked chassis design and is equipped with several sensing devices. These sensing devices are integrated into one unit and connected to the mobile platform via telescopic poles and a gimbal. The sensing devices can achieve multi-degree-of-freedom movement through the telescopic poles combined with the gimbal.

[0123] In one or more embodiments, the sensing device includes:

[0124] Wide-area scanning unit: line LiDAR and panoramic vision camera;

[0125] Fine detection unit: high-definition zoom camera, infrared thermal imager, and acoustic sensor.

[0126] The computing unit adopts a local heterogeneous computing processing mechanism, specifically including an edge computing module and a control module, wherein:

[0127] The edge computing module uses an embedded GPU, and its main deployment algorithm model and task processing are shown below:

[0128] (1) Active sensing algorithm stack:

[0129] Multi-scale feature entropy calculation model: the core algorithm for real-time calculation of geometric and texture anomalies during the fast scanning stage.

[0130] Attention mechanism network (or weight calculation module): Integrates multimodal information to generate an interest score map.

[0131] (2) Lightweight multi-task disease identification model:

[0132] Used for synchronous multi-task inference on data collected by the fine detection unit: disease type classification, bounding box regression, semantic segmentation, and geometric parameter measurement.

[0133] (3) Dynamic scheduling strategy solver:

[0134] This is used to solve sensor scheduling priorities and optimal detection dwell point planning in real time based on interest graphs.

[0135] (4) Data fusion and 3D reconstruction:

[0136] It works in conjunction with the front-end for point cloud processing and visual SLAM to perform temporal information fusion and 3D reconstruction.

[0137] The control module employs a real-time microcontroller, and its main algorithm model and task processing are as follows:

[0138] (1) Robot underlying motion control algorithm:

[0139] Motor PID control and chassis trajectory tracking control.

[0140] High-frequency low-level path tracing is performed based on path points issued by the embedded GPU.

[0141] (2) Direct control and preprocessing of sensor hardware:

[0142] Core preprocessing: Reducing the burden on the GPU. For example, hardware-level image distortion correction and initial feature point extraction are performed on the video stream from the panoramic camera (PV) before being provided to the GPU for SLAM calculations.

[0143] Sensor synchronization triggering: Strict timestamp synchronization ensures spatiotemporal alignment of data from multiple sensors.

[0144] Simple filtering: Low-pass filtering and complementary filtering are applied to IMU and odometry data to provide smooth initial attitude values.

[0145] (3) Real-time security monitoring:

[0146] Emergency stop signal processing, collision detection (based on IMU mutation), and battery and communication status monitoring.

[0147] (4) Data collection and forwarding:

[0148] As the data flow management layer between the sensor and the GPU, it reliably collects and forwards all raw data.

[0149] The positioning unit adopts an IMU + odometry + visual SLAM structure design;

[0150] The communication unit adopts 5G / WiFi-6 dual-mode communication and supports Mesh networking.

[0151] Example 2

[0152] In one or more embodiments, such as Figure 3 As shown, this embodiment provides an automatic inspection method for concrete surfaces in narrow-space tunnels, which is based on the aforementioned automatic inspection robot for concrete surfaces in narrow-space tunnels, including:

[0153] As the inspection robot travels along the tunnel axis, a sparse 3D map of the tunnel environment is constructed online based on the synchronously acquired wide-area scanning data and positioning data.

[0154] Based on the wide-area scanning data obtained within the preset range, real-time preliminary feature calculations are performed to obtain the basic features of each analysis unit within the preset range. The wide-area scanning unit works continuously as the inspection robot travels along the tunnel axis.

[0155] The real-time interest degree of each analysis unit is calculated based on the aforementioned basic features, and a set of high-interest regions is determined based on the real-time interest degree.

[0156] For the units to be analyzed in the set of regions of high interest, local refined real-time detection is performed. In the local refined real-time detection, for each unit to be analyzed, one or more refined sensors that match the unit to be analyzed are activated to collect data based on the matching degree between the unit to be analyzed and the sensors in the refined detection unit. Multi-scale feature extraction is performed on the collected data, and the detection result of the current unit is obtained through a pre-built lightweight multi-task anomaly detection model based on deep learning.

[0157] In further embodiments, the following is also provided:

[0158] like Figure 4As shown, an electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When executed by the processor, the computer instructions perform the methods described in the above embodiments. For brevity, further details are omitted here.

[0159] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0160] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0161] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the methods described in the above embodiments.

[0162] The methods described in the above embodiments can be directly executed by a hardware processor, or executed by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, detailed descriptions are omitted here.

[0163] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An automated inspection robot for concrete surfaces in narrow tunnel spaces, comprising a mobile platform, sensing devices mounted on the mobile platform, a computing unit, and a communication unit, characterized in that, The sensing device includes a wide-area scanning unit, a fine detection unit, and a positioning unit; the computing unit adopts a local heterogeneous computing processing mechanism. The computing unit specifically performs the following processing steps: During the inspection robot's movement along the tunnel axis, a sparse 3D map of the tunnel environment is constructed online based on synchronously acquired wide-area scanning data and positioning data. Real-time preliminary feature calculations are performed based on the acquired wide-area scanning data within a preset range to obtain the basic features of each analysis unit within the preset range. The wide-area scanning unit continues to operate during the inspection robot's movement along the tunnel axis. The real-time interest degree of each analysis unit is calculated based on the basic features, and a set of high-interest regions is determined based on the real-time interest degree. For the analysis units within the high-interest region set, local refined real-time detection is performed. In this local refined real-time detection, for each analysis unit, one or more matching refined sensors are activated to collect data based on the matching degree between the analysis unit and the sensors within the refined detection unit. Multi-scale feature extraction is performed on the collected data, and the detection result of the current analysis unit is obtained through a pre-built lightweight multi-task anomaly detection model based on deep learning.

2. The automatic inspection robot for concrete surfaces in a narrow tunnel as described in claim 1, characterized in that, The calculation of the real-time interest degree of each analysis unit based on the aforementioned basic features is specifically represented as follows: in, To analyze the real-time interest level of unit c, The sigmoid function is used to normalize numerical values. For geometric anomaly, For texture anomaly degree, , For learnable weights, , This is a scale parameter used to normalize different feature dimensions.

3. The automatic inspection robot for concrete surfaces in a narrow tunnel as described in claim 1, characterized in that, The determination of the set of high-interest regions based on the real-time interest level specifically involves: in, A set of regions of high interest. Indicates the analysis unit, The interest threshold; The interest threshold Dynamic settings are used, as shown below: in, All within the current sliding window The mean, All within the current sliding window standard deviation This is the sensitivity coefficient.

4. The automatic inspection robot for concrete surfaces in a narrow tunnel as described in claim 1, characterized in that, Based on the matching degree between the sensors in the unit to be analyzed and the fine detection unit, one or more fine sensors that are matched with it are activated to collect data, specifically including the following processing steps: For each unit to be analyzed in the set of regions of high interest, calculate the expected utility, quality factor and resource consumption of each sensor in the fine detection unit for the range of interest suggested by the unit to be analyzed. Priority calculation is performed based on the expected utility, quality factor, and resource consumption. Based on the obtained priority, each unit to be analyzed is assigned one or a group of sensors with the highest priority for its refined real-time detection.

5. The automatic inspection robot for concrete surfaces in a narrow tunnel as described in claim 4, characterized in that, The priority is calculated as follows: ) in, For the expected utility, Let be the quality factor, representing the calibration value of the j-th sensor's detection capability for typical diseases in the current environment. Resource consumption indicates that the j-th sensor is activated and the unit to be analyzed is started. The normalized value of the overall cost of conducting the test. Let i be the interest level of the i-th unit to be analyzed. Let be the area of ​​the region of the i-th unit to be analyzed. Let be the estimated time taken for the j-th sensor to collect data from the i-th unit under analysis in a single acquisition. Let be the matching degree between the historical disease type of the i-th unit to be analyzed and the j-th sensor. To adjust the parameters.

6. The automatic inspection robot for concrete surfaces in a narrow tunnel as described in claim 1, characterized in that, The lightweight multi-task anomaly detection model is deployed on the local edge computing unit of the inspection robot and has the ability to co-evolve with the cloud. The inspection robot uploads the local low-confidence detection sample data to the cloud. The cloud processes the samples through a high-precision teacher model to obtain high-quality labels and periodically updates the optimized model parameters to the edge computing unit via OTA through knowledge distillation technology.

7. The automatic inspection robot for concrete surfaces in a narrow tunnel as described in claim 1, characterized in that, The wide-area scanning unit includes a line lidar and a panoramic vision camera; the fine detection unit includes a high-definition zoom camera, an infrared thermal imager, and an acoustic sensor; the computing unit adopts a local heterogeneous computing processing mechanism, specifically including a GPU-based edge computing module and a real-time microcontroller-based control module; the positioning unit adopts an IMU + odometry + visual SLAM structure design.

8. A method for automatic inspection of concrete surfaces in narrow-space tunnels, based on an automatic inspection robot for concrete surfaces in narrow-space tunnels as described in any one of claims 1-7, characterized in that, include: As the inspection robot travels along the tunnel axis, a sparse 3D map of the tunnel environment is constructed online based on the synchronously acquired wide-area scanning data and positioning data. Based on the wide-area scanning data obtained within the preset range, real-time preliminary feature calculations are performed to obtain the basic features of each analysis unit within the preset range. The wide-area scanning unit works continuously as the inspection robot travels along the tunnel axis. The real-time interest degree of each analysis unit is calculated based on the aforementioned basic features, and a set of high-interest regions is determined based on the real-time interest degree. For the units to be analyzed in the set of regions of high interest, local refined real-time detection is performed. In the local refined real-time detection, for each unit to be analyzed, one or more refined sensors that match the unit to be analyzed are activated to collect data based on the matching degree between the unit to be analyzed and the sensors in the refined detection unit. Multi-scale feature extraction is performed on the collected data, and the detection result of the current unit is obtained through a pre-built lightweight multi-task anomaly detection model based on deep learning.

9. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and running thereon, characterized in that, When the processor executes the program, it implements the automatic inspection method for concrete surfaces in a narrow space tunnel as described in claim 8.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the automatic inspection method for concrete surfaces in a narrow space tunnel as described in claim 8.

Images