Archaeological site intelligent detection robot system based on multi-mode perception

Abstract

The invention provides an archaeological site intelligent detection robot system based on multi-modal perception, and relates to the technical field of archaeological exploration, and the system comprises a multi-modal sensor array which is used for synchronously obtaining underground target information, the multi-mode sensor array comprises an optical camera, a geological radar, an ultra-wideband radar, a sonar sensor, an electronic magnetic induction detector and a micro-seismic sensor, and the sensor array has the maximum detection burial depth capacity not smaller than 15 meters and the minimum spatial resolution capacity not larger than 1 centimeter. According to the archaeological site intelligent detection robot system based on multi-modal sensing, the balance of the depth, the resolution ratio and the coverage range of an underground structure is achieved through the organic combination of a multi-modal sensor array and self-adaptive parameter adjustment. Therefore, the false alarm rate and the omission ratio are obviously reduced, and the high-precision detection capability under the complex soil texture and the changeable environment is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of archaeological exploration, and in particular to an archaeological site intelligent detection robot system based on multimodal perception. Background Art

[0002] Currently, archaeological site exploration relies heavily on a combination of manual surveys and a single or small number of sensors. During field operations, traditional archaeological teams obtain stratigraphic and archaeological information through manual excavation and cross-sectioning, while simultaneously scanning the target area point by point using equipment such as ground-penetrating radar (GPR) and metal detectors. Aerial photography using drones, equipped with optical or multispectral cameras, enables large-scale terrain and vegetation monitoring. Some research, in laboratory or small, controlled environments, is experimenting with the combined deployment of multiple sensors, including sonar, geological radar, and inertial measurement units (IMUs), aiming to simultaneously detect and collect data from archaeological targets at varying depths and types.

[0003] The above technologies still have obvious shortcomings in practical applications: single or simply combined sensors find it difficult to strike a balance between detection depth, spatial resolution and coverage, and often find it difficult to fully and accurately reveal underground structures. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides an intelligent archaeological site detection robot system based on multimodal perception. The technical problem to be solved by this invention is: how to achieve high-precision, fully autonomous, and long-endurance detection of underground relics through multi-sensor synchronous acquisition and adaptive parameter adjustment, edge deep fusion and real-time anomaly detection, implicit field digital twin three-dimensional reconstruction and interactive visualization.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: an archaeological site intelligent detection robot system based on multimodal perception, comprising: A multimodal sensor array for synchronously acquiring underground target information, comprising an optical camera, a geological radar, an ultra-wideband radar, a sonar sensor, an electronic magnetic induction detector, and a microseismic sensor. The sensor array has a maximum detection depth capability of no less than 15 meters and a minimum spatial resolution capability of no more than 1 centimeter. An adaptive sensor parameter adjustment module, which dynamically adjusts the operating parameters of the multimodal sensor array based on real-time environmental feedback, including but not limited to transmit frequency, pulse width, scan rate, and gain, to maintain optimal detection performance under varying soil quality, moisture content, and burial depth conditions; Edge intelligent data fusion and self-learning anomaly detection module, which uses a deep neural network model to self-learn and identify abnormal archaeological targets in real time; A digital twin 3D reconstruction and visualization module, which provides an interactive visualization interface for archaeological experts; The energy-adaptive mobile platform and autonomous navigation module are used to integrate solar panels and kinetic energy recovery devices to continuously power the robot.

[0006] Preferably, the optical camera is a dual-spectrum camera of visible light and near-infrared, with a resolution of not less than 40 million pixels and an adjustable focal length of 2 to 10 times, so as to achieve high-precision imaging of targets at different distances. The geological radar operates in the frequency band of 200 MHz to 1 GHz, is equipped with a dual-polarization antenna, and uses pulse compression technology to increase the distance resolution to no more than 5 cm. The ultra-wideband radar uses a spectrum of 3.1 GHz to 10.6 GHz, with a bandwidth of not less than 7.5 GHz, and achieves timing synchronization of transmission and reception through coaxial time multiplexing switching between the transmitting antenna and the receiving antenna. The sonar sensor is configured as a broadband piezoelectric transducer array with a frequency range of 20 kHz to 200 kHz and a unit spacing of no more than 5 cm to ensure the acoustic imaging accuracy of small-scale targets. The electronic magnetic induction detector includes a three-axis fluxgate sensor module. The microseismic sensor is a piezoelectric ceramic accelerometer with a range of not less than ±2 g and a bandwidth of 0.1 Hz to 500 Hz to capture weak geological vibration signals and assist in deep target identification.

[0007] Preferably, the adaptive sensor parameter adjustment module includes: The environmental perception submodule is used to collect on-site environmental parameters in real time through soil conductivity sensors, humidity sensors, and temperature sensors; A parameter decision submodule is used to calculate and output the optimal transmission frequency, pulse width, antenna gain and scanning rate of each sensor based on the environmental parameters using a pre-trained offline sparse Gaussian process regression model; The parameter execution submodule is used to send the optimal parameters to the geological radar, ultra-wideband radar and sonar unit through the on-site digital signal processor.

[0008] Preferably, the edge intelligent data fusion and self-learning anomaly detection module includes: The multi-source point cloud alignment submodule is used to align the pre-processed multi-modal point cloud dataset. Indicates the sensor type, executes the iterative closest point algorithm, and obtains a unified point cloud set in the global coordinate system; the three-layer convolutional neural network feature extraction submodule executes the following in sequence: Where "∗" represents a three-dimensional convolution operation; Online Mahalanobis distance anomaly detection submodule to calculate anomaly scores.

[0009] Preferably, the digital twin 3D reconstruction and visualization module includes: Implicit field reconstruction submodule, based on fused point cloud Construct an implicit field of radial basis functions: , : Implicit field numerical function, at the spatial point scalar value at ; : The total number of all sample points in the fused point cloud; : Any query point, three-dimensional space coordinate vector; : No. The three-dimensional coordinates of the fused point cloud samples; : Euclidean distance, indicating the query point With sample points The straight-line distance between : Gaussian kernel bandwidth, which controls the spatial diffusion range of the radial basis function. ; : No. The weight coefficient of the kernel function determines the contribution of the point to the implicit field.

[0010] Preferably, the unified point cloud set: , : A unified point cloud set fused into the global coordinate system; : No. The first Original point cloud coordinates; : No. The rigid body rotation correction matrix of the sensor class represents the rotation transformation from the sensor coordinate system to the global coordinate system; : No. The translation correction vector of the class sensor represents the translation transformation from the sensor coordinate system to the global coordinate system, in meters (m); : Total number of sensor categories; : No. The number of raw point cloud samples corresponding to the sensor class.

[0011] Preferably, the energy adaptive mobile platform and autonomous navigation module include: an energy collection submodule, an energy management submodule and an autonomous navigation and endurance monitoring submodule.

[0012] Preferably, the autonomous navigation and endurance monitoring submodule adopts the shortest path priority algorithm.

[0013] The present invention provides an intelligent archaeological site detection robot system based on multimodal perception. It has the following beneficial effects:

[0014] This intelligent archaeological site detection robot system, based on multimodal perception, achieves a balance between depth, resolution, and coverage of underground structures through the organic combination of a multimodal sensor array and adaptive parameter adjustment. On the one hand, it synchronously collects information from multiple sources, including optics, GPR, ultra-wideband radar, sonar, magnetic induction, and microseismic information. Through an online parameter optimization module based on a sparse Gaussian process, it adaptively adjusts key parameters such as transmission frequency, pulse width, and antenna gain in real time. On the other hand, a deep neural network and an online Mahalanobis distance anomaly detection algorithm deployed at the edge efficiently complete data fusion and self-learning identification of potential relics on a mobile platform. This significantly reduces false alarms and missed detection rates, and ensures high-precision detection capabilities in complex soils and changing environments.

[0015] Digital twin 3D reconstruction and visualization technology is introduced, and millimeter-level subdivision grids are constructed based on the implicit field RBF model to meet archaeologists' needs for dynamic analysis of underground structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is to realize the structural intent of the invention; Figure 2 It is a flowchart for implementing the invention. DETAILED DESCRIPTION

[0017] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0018] Example 1 like Figure 1-2As shown, an embodiment of the present invention provides an archaeological site intelligent detection robot system based on multimodal perception, including a multimodal sensor array for synchronously acquiring underground target information. The multimodal sensor array includes: an optical camera, a geological radar, an ultra-wideband radar, a sonar sensor, an electronic magnetic induction detector and a microseismic sensor. The sensor array has a maximum detection depth of not less than 15 meters and a minimum spatial resolution of not more than 1 centimeter. The optical camera is a visible light and near-infrared dual-spectrum camera with a resolution of not less than 40 million pixels and an adjustable focal length of 2 to 10 times to achieve high-precision imaging of targets at different distances. The geological radar operates in the frequency band of 200 MHz to 1 GHz, is equipped with a dual-polarization antenna, and is connected to the ground. The pulse compression technology increases the distance resolution to no more than 5cm. The ultra-wideband radar uses a spectrum of 3.1GHz to 10.6GHz with a bandwidth of no less than 7.5GHz, and achieves timing synchronization of transmission and reception through coaxial time multiplexing switching between the transmitting antenna and the receiving antenna. The sonar sensor is configured as a broadband piezoelectric transducer array with a frequency range of 20kHz to 200kHz and a unit spacing of no more than 5cm to ensure the acoustic imaging accuracy of small-scale targets. The electronic magnetic induction detector includes a three-axis fluxgate sensor module, and the microseismic sensor is a piezoelectric ceramic accelerometer with a range of no less than ±2g and a bandwidth of 0.1Hz to 500Hz to capture weak geological vibration signals and assist in deep target identification.

[0019] The adaptive sensor parameter adjustment module is used to dynamically adjust the operating parameters of the multimodal sensor array based on real-time environmental feedback. The operating parameters include but are not limited to the transmission frequency, pulse width, scan rate, and gain to maintain optimal detection performance under different soil conditions, humidity, and burial depth. The adaptive sensor parameter adjustment module includes: The environmental perception submodule is used to collect on-site environmental parameters in real time through soil conductivity sensors, humidity sensors and temperature sensors.

[0020] The parameter decision submodule is used to calculate and output the optimal transmission frequency, pulse width, antenna gain and scanning rate of each sensor based on environmental parameters using a pre-trained offline sparse Gaussian process regression model.

[0021] The parameter execution submodule is used to send the optimal parameters to the geological radar, ultra-wideband radar and sonar unit through the on-site digital signal processor to ensure that under the conditions that the soil conductivity is not greater than 300mS / m and the soil moisture content is not greater than 25%, the effective penetration depth of the geological radar is not less than 12m and the spatial resolution of the ultra-wideband radar is not greater than 1cm.

[0022] The edge intelligent data fusion and self-learning anomaly detection module uses a deep neural network model to self-learn and identify abnormal archaeological targets in real time. The edge intelligent data fusion and self-learning anomaly detection module includes: The multi-source point cloud alignment submodule is used to align the pre-processed multi-modal point cloud dataset. Indicates the sensor type, executes the iterative closest point algorithm, and obtains a unified point cloud set in the global coordinate system; the three-layer convolutional neural network feature extraction submodule executes the following in sequence: Where “∗” represents a three-dimensional convolution operation.

[0023] Online Mahalanobis distance anomaly detection submodule to calculate anomaly scores.

[0024] in and the covariance matrix Updated in real time by exponential smoothing algorithm: And the initial threshold and update rules , determination For abnormal targets, ; The digital twin 3D reconstruction and visualization module is used to provide an interactive visualization interface for archaeological experts. The digital twin 3D reconstruction and visualization module includes: Implicit field reconstruction submodule, based on fused point cloud Construct an implicit field of radial basis functions: , : Implicit field numerical function, at the spatial point scalar value at ; : The total number of all sample points in the fused point cloud; : Any query point, three-dimensional space coordinate vector, unit meter (m); : No. The three-dimensional coordinates of the fused point cloud samples, in meters (m); : Euclidean distance, indicating the query point With sample points The straight-line distance between them, in meters (m); : Gaussian kernel bandwidth, which controls the spatial diffusion range of the radial basis function. ; : No. The weight coefficient of the kernel function determines the contribution of the point to the implicit field and is obtained by solving the following linear equations: , weight coefficient By linear system The solution is: , : Regularization parameter, used to improve the numerical stability of the equation system, here we take ; : identity matrix, and Same dimension, used for regularization; : the weight vector to be found; : Target vector, where all components are set to , which is used to ensure that the implicit field is approximately constant at the sample points and unify the point cloud set: , : A unified point cloud set fused into the global coordinate system; : No. The first Original point cloud coordinates; : No. The rigid body rotation correction matrix of the sensor class represents the rotation transformation from the sensor coordinate system to the global coordinate system; : No. The translation correction vector of the class sensor represents the translation transformation from the sensor coordinate system to the global coordinate system, in meters (m); : Total number of sensor categories; : No. The number of original point cloud samples corresponding to the sensor class; The energy-adaptive mobile platform and autonomous navigation module integrates solar panels and a kinetic energy recovery device to continuously power the robot. The module includes an energy collection submodule, an energy management submodule, and an autonomous navigation and endurance monitoring submodule. The latter uses a shortest path first algorithm. The energy collection submodule.

[0025] Solar panel components: Use monocrystalline silicon solar panels, the total area of ​​the components ; Standard test conditions ( ) output maximum power 160W; Average field measured light intensity When the power is on, the output power is about 96W.

[0026] Kinetic energy recovery device The robot has a curb weight of 25kg and brake deceleration ( ) The process takes 0.5s; Recycling efficiency Single braking can recover electric energy (approximately 0.047Wh).

[0027] Battery pack status monitoring: Equipped with lithium battery pack, rated capacity ; The measured SOC is down to It takes 3.5 hours, and edge monitoring updates SOC and SOH once a minute.

[0028] Autonomous navigation and endurance monitoring submodule: Path planning parameters, in the shortest path first (SPF) algorithm, the maximum plannable path length is set as: Lmax , weight ; In a typical mission, the target point is 150m away from the starting point, and the estimated energy consumption Ec , currently available Ea ,cost .

[0029] Obstacle avoidance and return strategy When radar or visual SLAM detects an obstacle, , automatically recalculate the path and decelerate to ; When SOC or Ea When the vehicle is in a state of emergency, priority is given to planning the return path and triggering kinetic energy recovery auxiliary braking.

[0030] Combining solar energy, kinetic energy recovery and battery output, it can operate continuously for 4.2 hours without external charging; The path update frequency is 0.5Hz, ensuring smooth navigation without significantly increasing energy consumption.

[0031] As verified by the above data, the mobile platform integrating solar panels and kinetic energy recovery devices in this embodiment can continuously supply energy, accurately navigate under typical archaeological site conditions, and intelligently switch operating modes at critical energy moments, fully demonstrating the technical advantages of the present invention in terms of battery life and autonomy.

[0032] Example 2 The edge intelligent data fusion and self-learning anomaly detection module is described to help understand the technical solution of the present invention, but the present invention is not limited to this embodiment.

[0033] 1. Multi-source point cloud alignment submodule Original data: Preprocessed point clouds from three types of sensors: optical camera, geological radar, and ultra-wideband radar: Sensor 1 (optical camera): points Sensor 2 (geological radar): points Sensor 3 (Ultra-Wideband Radar): points.

[0034] Alignment process 2. For each type of sensor, use the initial rigid body transformation matrix Approximately map a point cloud to global coordinates.

[0035] 1.2. Apply the iterative closest point algorithm for rigid body optimization, set the maximum number of iterations to 100 times, and the convergence threshold to the root mean square error .

[0036] 1.3. After the iteration, the alignment errors of the three point cloud sets are reduced to: Optical point cloud: root mean square error .

[0037] GPR point cloud: Root mean square error .

[0038] UWB point cloud: RMS error = 0.007m.

[0039] Output: , The total number of points is approximately point.

[0040] 2. Three-layer convolutional neural network feature extraction submodule Input representation: The aligned point cloud Convert to The voxel grid is set to 1, the blank value is set to 0, and the tensor size is obtained. .

[0041] Network structure: 2.1. First convolution layer , Convolution kernel, stride 1, padding 1 : Bias vector of length 64 Output size 2.2. Second convolution layer , , stride 2, padding 1 (downsampling) : Bias vector of length 128 Output size 2.3. Third convolution layer , , step size 2, padding 1; : bias vector of length 256; Output size .

[0042] 2.4. Feature Vectorization Will Global average pooling is performed to a feature vector of length 256 for subsequent anomaly detection.

[0043] 3. Online Mahalanobis distance anomaly detection submodule: Initial parameters: .

[0044] For the Frame feature vector 3.1. Update mean and covariance

[0045] 3.2. Calculating Mahalanobis distance , 3.3. Threshold Update and Decision Instance data; Frame 1: ,normal; Frame 2: , judged as abnormal and updated the threshold to .

[0046] Through the above-mentioned specific data and steps, this embodiment organically combines precise alignment of multi-source point clouds, deep feature extraction, and online anomaly detection, thereby achieving real-time and high-precision identification of potential abnormal targets at the archaeological site.

[0047] Example 3 The digital twin 3D reconstruction and visualization module is described. Taking the point cloud of as an example, the steps are as follows.

[0048] Implicit field reconstruction submodule: 1. Input data fusion point cloud set.

[0049] 2. Parameter settings: Gaussian kernel bandwidth Regularization coefficient Target vector .

[0050] 3. Weight solution Constructing the kernel matrix ,element: , Solve a linear system: , get .

[0051] 4. Implicit Field Calculation For any query point ,calculate: Mesh extraction and texture mapping submodule: 1. Isosurface extraction Selecting an isosurface threshold , in the cubic area The voxel spacing is 0.005m, and the surface mesh M is extracted using the MarchingCubes algorithm.

[0052] 2. Mesh refinement Subdivide the longest edge of the initially generated triangular mesh into , the total number of triangles increases to approximately .

[0053] 3. Texture Mapping For each mesh vertex , in the point cloud closest to Read the reflection intensity at , grayscale values are mapped to vertex colors through bilinear interpolation to achieve resolution Pixel texture map.

[0054] Interactive visualization submodule: 1. View control, users can enter the rotation axis parameters on the interface and angles , call the WebGL interface in real time to perform matrix transformation.

[0055] 2. Arbitrary sectioning, providing plane equation parameters , perform GPU sectioning on the mesh M, supporting section thickness of 1mm.

[0056] 3. Distance measurement tool, the user clicks two points , the interface automatically calculates: , And display the distance value with millimeter level accuracy.

[0057] Through the above embodiments, the system can efficiently reconstruct an implicit digital twin model based on a 300,000 point cloud, and provide visualization and section analysis functions with millimeter-level accuracy, meeting the archaeological experts' needs for refined interaction with underground structures.

[0058] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent archaeological site detection robot system based on multimodal perception, characterized by: include: A multimodal sensor array for synchronously acquiring underground target information, comprising: an optical camera, a geological radar, an ultra-wideband radar, a sonar sensor, an electronic magnetic induction detector, and a microseismic sensor; An adaptive sensor parameter adjustment module for dynamically adjusting the operating parameters of the multimodal sensor array based on real-time environmental feedback; Edge intelligent data fusion and self-learning anomaly detection module, which uses a deep neural network model to self-learn and identify abnormal archaeological targets in real time; A digital twin 3D reconstruction and visualization module, which provides an interactive visualization interface for archaeological experts; The energy-adaptive mobile platform and autonomous navigation module are used to integrate solar panels and kinetic energy recovery devices to continuously power the robot.

2. The multimodal sensing-based intelligent archaeological site detection robot system according to claim 1, characterized in that: The optical camera is a dual-spectrum camera of visible light and near-infrared, the geological radar operates in the frequency band of 200MHz to 1GHz, the ultra-wideband radar uses the spectrum of 3.1GHz to 10.6GHz, the sonar sensor is configured as a broadband piezoelectric transducer array with a frequency range of 20kHz to 200kHz, the electronic magnetic induction detector includes a three-axis fluxgate sensor module, and the microseismic sensor is a piezoelectric ceramic accelerometer with a range of not less than ±2g and a bandwidth of 0.1Hz to 500Hz.

3. The multimodal sensing-based intelligent archaeological site detection robot system according to claim 1, characterized in that: The adaptive sensor parameter adjustment module includes: The environmental perception submodule is used to collect on-site environmental parameters in real time through soil conductivity sensors, humidity sensors, and temperature sensors; A parameter decision submodule, for making a decision based on the environmental parameters; The parameter execution submodule is used to send the optimal parameters to the geological radar, ultra-wideband radar and sonar unit through the on-site digital signal processor.

4. The multimodal sensing-based intelligent archaeological site detection robot system according to claim 1, characterized in that: The edge intelligent data fusion and self-learning anomaly detection module includes: The multi-source point cloud alignment submodule is used to align the pre-processed multi-modal point cloud dataset. Execute the iterative closest point algorithm to obtain a unified point cloud set in the global coordinate system; the three-layer convolutional neural network feature extraction submodule executes the following in sequence: Where "∗" represents a three-dimensional convolution operation; Online Mahalanobis distance anomaly detection submodule to calculate anomaly scores.

5. The multimodal sensing-based intelligent archaeological site detection robot system according to claim 1, characterized in that: The digital twin 3D reconstruction and visualization module includes: Implicit field reconstruction submodule, based on fused point cloud Construct an implicit field of radial basis functions: ， : Implicit field numerical function, at the spatial point scalar value at ; : The total number of all sample points in the fused point cloud; : Any query point, three-dimensional space coordinate vector; : No. The three-dimensional coordinates of the fused point cloud samples; : Euclidean distance, indicating the query point With sample points The straight-line distance between : Gaussian kernel bandwidth, which controls the spatial diffusion range of the radial basis function. ; : No. The weight coefficient of the kernel function determines the contribution of the point to the implicit field.

6. The multimodal sensing-based intelligent archaeological site detection robot system according to claim 5, characterized in that: The unified point cloud collection: ， : A unified point cloud set fused into the global coordinate system; : No. The first Original point cloud coordinates; : No. The rigid body rotation correction matrix of the sensor class represents the rotation transformation from the sensor coordinate system to the global coordinate system; : No. The translation correction vector of the class sensor represents the translation transformation from the sensor coordinate system to the global coordinate system; : Total number of sensor categories; : No. The number of raw point cloud samples corresponding to the sensor class.

7. The multimodal sensing-based intelligent archaeological site detection robot system according to claim 1, characterized in that: The energy adaptive mobile platform and autonomous navigation module include: an energy collection submodule, an energy management submodule and an autonomous navigation and endurance monitoring submodule.

8. The multimodal sensing-based intelligent archaeological site detection robot system according to claim 7, characterized in that: The autonomous navigation and endurance monitoring submodule adopts the shortest path priority algorithm.

Images