Tank-oriented twin model construction method, system, device and medium for tank decommissioning

Abstract

The present application relates to the technical field of data processing, and more particularly to a twin model construction method, system, device and medium for tank decommissioning, the construction method comprising the following steps: acquiring point cloud data, reconstructing a point cloud network, extracting size information, and constructing a twin model; real-time monitoring of sensor health status through time series feature analysis, capturing device degradation rules using a gated recurrent network, generating a hidden vector reflecting the degree of performance degradation, constructing a Poisson equation to describe the implicit surface, and using the finite element method to solve to obtain a continuous mesh model; and collaborative mapping of three-dimensional mesh and two-dimensional projection data, which not only preserves the overall topological form of the tank, but also extracts key size parameters, providing dual precision protection for path planning, and solving the problem of data drift in the model construction method of the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a twin model construction method, system, device and medium for tank decommissioning. BACKGROUND

[0002] The safe and efficient disposal of radioactive waste liquid tanks is a highly challenging task. The existing technology mainly relies on remote control of mechanical arms and other equipment by operators, which is highly dependent on the experience and psychological quality of the operators, and is not only inefficient, but also prone to safety accidents such as tool jamming, collision, and radioactive material leakage in complex tank environments due to lack of accurate environmental perception and intelligent decision support. Therefore, the industry has introduced digital twin technology to provide full-process visual monitoring, operation simulation and intelligent path planning for decommissioning operations by constructing a high-fidelity virtual model that is real-time mapped with the physical tank and interacts between the virtual and real worlds, thereby improving the safety and efficiency of operations.

[0003] However, when applying digital twin technology to the decommissioning of tanks with strong radioactivity and high airtightness, the existing model construction methods have the following problems:

[0004] First, sensor irradiation aging and data drift. The accuracy of the digital twin is highly dependent on real-time data provided by various sensors such as radiation detectors and temperature sensors. In the strong radiation environment inside the tank, the core components of these sensors will undergo irreversible performance degradation, i.e., irradiation aging, resulting in slow and difficult-to-detect systematic bias or drift in their measurement data over time. The digital twin model constructed and driven based on this continuously distorted data will gradually become a distorted mirror that deviates from physical reality, and any decisions made based on this model, such as assessing the cumulative radiation dose received by equipment or judging hot spots, may have significant errors, thereby causing serious safety hazards and fundamentally undermining the credibility of the digital twin system.

[0005] Second, geometric distortion under aerosol and medium interference. The geometric skeleton of the digital twin is usually constructed using optical measurement methods such as three-dimensional laser scanning to accurately reproduce the internal status of the tank. However, in actual operations, the tank is often filled with radioactive aerosols, water vapor and suspended particles generated by the evaporation of waste liquid, chemical reactions or flushing operations. These media can scatter and absorb laser beams, resulting in a large number of noise points and incorrect measurement values in the collected point cloud data, and further causing severe spatial distortion in the reconstructed three-dimensional geometric model. SUMMARY

[0006] The main purpose of the present application is to provide a twin model construction method for tank decommissioning, which aims to solve the problem of data drift in the existing model construction method.

[0007] To achieve the above object, the application provides a twin model construction method for tank decommissioning, which is used for constructing a twin model of a tank and comprises the following steps:

[0008] Point cloud data in the tank is acquired, and the acquired point cloud data is preprocessed to remove noise points and realize uniform sampling of the point cloud data;

[0009] Normal vector estimation, creation of a Poisson equation, numerical method solution of the Poisson equation and surface reconstruction are sequentially performed to obtain a mesh model of a continuous surface from the point cloud data;

[0010] Points in the three-dimensional point cloud in the mesh model are projected into a two-dimensional plane to generate a two-dimensional point set to represent the distribution of the point cloud in the plane;

[0011] The mesh model is taken as morphological information reference, and the two-dimensional point set is taken as size information reference, and a twin model of the tank is constructed in a twin layer;

[0012] The method further comprises:

[0013] The twin model is used to plan an operation path and a cutting sequence of the tank.

[0014] Further, before the point cloud data in the tank is acquired, the method further comprises:

[0015] Historical data in an ideal environment are acquired to perform supervised training through a loss function and adaptive calibration through an online loss function.

[0016] Further, the preprocessing of the acquired point cloud data specifically comprises: using a radial outlier removal algorithm to remove spatial isolated noise points in a single frame of point cloud caused by medium scattering; and then comparing point clouds of consecutive time frames to identify and remove instantaneous point clusters.

[0017] Further, the preprocessing of the acquired point cloud data further comprises: outputting a hidden state vector through a gated recurrent unit and outputting a corrected point cloud data through a multi-layer perception.

[0018] Further, the radial outlier removal algorithm for removing spatial isolated noise points in a single frame of point cloud caused by medium scattering specifically comprises the following steps: a spherical neighborhood is defined for each point, and the number of points in the neighborhood is calculated; if the number of points in the neighborhood of a point is less than a preset threshold, the point is determined as an outlier and removed.

[0019] Further, the identifying and removing the instantaneous point cluster specifically comprises the following steps: extracting multi-scale geometric features by using a graph convolution network, and approximating an implicit signed distance function by using a multi-layer perception, and adding a constraint to the signed distance function to obtain a signed distance function field satisfying the constraint, and an output value of the signed distance function represents a distance from a spatial point in the point cloud data to a nearest surface.

[0020] Further, the re-comparing the point clouds of continuous time frames specifically comprises the following steps: aligning continuous multi-frame point clouds in a short time window, then dividing space into a three-dimensional voxel grid, and finally distinguishing real surfaces and transient noise by analyzing the stability of each voxel in the time dimension.

[0021] To achieve the above object, the present application further provides a twin model construction system for tank decommissioning, which comprises:

[0022] The acquisition module is configured to acquire point cloud data in the tank, and pre-process the acquired point cloud data to uniformly sample the point cloud data while removing noise points;

[0023] The reconstruction module is configured to sequentially perform normal vector estimation, creation of a Poisson equation, numerical method solving of the Poisson equation, and surface reconstruction to obtain a mesh model of a continuous surface from the point cloud data;

[0024] The extraction module is configured to project points in the three-dimensional point cloud in the mesh model to a two-dimensional plane to generate a two-dimensional point set to represent the distribution of the point cloud in the plane;

[0025] The construction module is configured to take the mesh model as morphological information reference, and take the two-dimensional point set as dimensional information reference, to construct a twin model of the tank in a twin layer.

[0026] To achieve the above object, the present application further provides a computer device, which comprises a memory and a processor, the memory stores a computer program, and the processor executes the computer program.

[0027] To achieve the above object, the present application further provides a computer readable storage medium, which stores a computer program, and a processor executes the computer program.

[0028] The present application can achieve the following beneficial effects:

[0029] The application monitors the health state of a sensor in real time through time sequence feature analysis, uses a gated recurrent network to capture the device degradation law, generates a hidden vector reflecting the performance degradation degree, the vector guides a multi-layer perception network to dynamically compensate the original point cloud, eliminates the systematic error caused by irradiation aging, the preprocessing link adopts density clustering and space-time joint filtering, first removes spatial isolated points through spherical neighborhood density detection, and then combines time stability analysis of multiple frame point clouds to filter transient noise clusters, in the geometric reconstruction stage, the local geometric relationship of the point cloud is established through normal vector estimation, the Poisson equation is used to describe the implicit surface, the finite element method is used to solve to obtain a continuous grid model, the cooperative mapping of the three-dimensional grid and the two-dimensional projection data retains the overall shape topology of the tank and extracts key size parameters, provides double precision guarantee for path planning, and solves the problem of data drift in the model construction method in the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0030] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the drawings needed in the specific embodiments or prior art description will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, each element or part is not necessarily drawn according to the actual scale.

[0031] Figure 1 The flow block diagram of the construction method in embodiment 1 of the present application is shown in the figure.

[0032] Figure 2 The structure block diagram of the construction system in embodiment 2 of the present application is shown in the figure.

[0033] The implementation of the present application, functional features and advantages will be further described with reference to the drawings. DETAILED DESCRIPTION

[0034] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0035] If the description of "first", "second" and the like is involved in the embodiments of the present application, the description of "first", "second" and the like is only for the purpose of description, and cannot be understood as indicating or implying the relative importance of the indicated technical features or implicitly indicating the number of the indicated technical features. Therefore, the features limited by "first", "second" can be explicitly or implicitly included at least one of the features. In addition, the meaning of "and / or" appearing throughout the text includes three parallel schemes. For example, "A and / or B" includes A scheme, or B scheme, or A and B scheme. In addition, the technical solutions of each embodiment can be combined with each other, but it must be based on the realization of the ordinary skilled in the art. When the combination of technical solutions appears contradictory or unachievable, it should be considered that the combination of technical solutions does not exist, and is not within the protection scope required by the present application.

[0036] Embodiment 1:

[0037] As shown in the accompanying Figure 1 The present embodiment provides a twin model construction method for tank decommissioning, which is used for constructing a twin model of a tank, and comprises the following steps:

[0038] Obtaining point cloud data in the tank, and pre-processing the obtained point cloud data to realize uniform sampling of the point cloud data while eliminating noise points;

[0039] Sequentially performing normal vector estimation, creating a Poisson equation, solving the Poisson equation by a numerical method, and surface reconstruction to obtain a mesh model of a continuous surface from the point cloud data;

[0040] Projecting points in the three-dimensional point cloud in the mesh model onto a two-dimensional plane to generate a two-dimensional point set to represent the distribution of the point cloud in the plane;

[0041] Taking the mesh model as morphological information reference and taking the two-dimensional point set as size information reference, a twin model of the tank is constructed in the twin layer;

[0042] The method further comprises:

[0043] Planning the operation path and cutting sequence of the tank by using the twin model.

[0044] It should be noted that in the prior art, the safe disposal of radioactive waste liquid tanks relies on remote operation and digital twin technology, but there are two major problems of sensor irradiation aging leading to data drift and aerosol interference causing geometric distortion. Strong radiation environment causes the performance of sensor elements to decline, resulting in systematic deviation of measurement data, leading to distortion of the twin model; aerosols and suspended particles in the tank form scattering of laser scanning, and the point cloud data contains a large number of noise points, and the reconstructed three-dimensional model appears spatial distortion, affecting the accuracy of operation path planning.

[0045] To solve the above problems, a dynamic calibration mechanism is considered to be established for sensor performance degradation problems, time series data analysis is used to capture device state changes, and for medium interference, a multi-stage geometric reconstruction method is explored, and constraint conditions are introduced in the noise elimination and surface reconstruction link; a degradation state estimation model is constructed to compensate for data drift, combined with space-time joint filtering to eliminate transient noise, and finally a solution is formed that takes into account dynamic calibration and geometric accuracy.

[0046] Specifically, in the data acquisition stage, the health status of the sensor is monitored in real time through time series feature analysis, the device degradation rule is captured by using a gated recurrent network, and a hidden vector reflecting the degree of performance degradation is generated. The vector guides the multi-layer perception network to dynamically compensate the original point cloud, eliminating the systematic error caused by irradiation aging. In the preprocessing link, the spherical neighborhood density is used to detect and eliminate spatial isolated points, and then the time stability analysis of multiple frames of point clouds is combined to filter out transient noise clusters. In the geometric reconstruction stage, the local geometric relationship of the point cloud is established by estimating the normal vector, the Poisson equation is used to describe the implicit surface, and the finite element method is used to solve to obtain a continuous grid model. The cooperative mapping of three-dimensional grid and two-dimensional projection data not only preserves the overall topological form of the tank, but also extracts key size parameters, providing double precision guarantee for path planning.

[0047] Compared with the prior art, the traditional method uses fixed parameters to calibrate the sensor, which cannot adapt to the gradual performance degradation in the radiation environment. The embodiment realizes dynamic error compensation through time series modeling, effectively suppresses data drift, and the existing point cloud filtering technology mainly relies on single-frame spatial analysis, which is difficult to distinguish between real surfaces and transient medium interference. The embodiment combines density clustering and multi-frame space-time correlation analysis to significantly improve the accuracy of noise identification; traditional surface reconstruction algorithms are prone to produce holes or distortion under medium interference, and the embodiment establishes a global continuous surface through numerical solving of the Poisson equation, enhancing the geometric integrity of the model.

[0048] Through the above technical solutions, the embodiment realizes online compensation of sensor data drift, controls the point cloud acquisition error within an acceptable range, and guarantees the data input accuracy of the twin model. Through the multi-stage geometric reconstruction technology, the surface distortion caused by aerosol scattering is effectively eliminated, and a high-fidelity three-dimensional grid model is generated; the cooperative mapping mechanism of three-dimensional form and two-dimensional projection provides accurate spatial constraints and size references for robot path planning, reducing the risk of tool collision and radioactive leakage during operation.

[0049] In the embodiment, before the point cloud data is acquired, the method further includes:

[0050] The acquisition module obtains historical data in an ideal environment before operation, to perform supervised training through a loss function, and to perform self-adaptive calibration through an online loss function. It can be understood that the object of obtaining the historical data is the acquisition module, that is, obtaining the historical data of the acquisition module in an ideal environment.

[0051] It can be understood that, in the offline stage, the acquisition module obtains multiple groups of point cloud data without interference in a controlled environment as a historical data set, an initial calibration model is established through supervised training, and the response characteristics of the sensor in a normal state are learned. In the online operation stage, the real-time collected point cloud data is compared with the historical reference, the deviation between the current measurement value and the expected value is calculated by using an online loss function, the model parameters are dynamically optimized by using a gradient descent algorithm, and the sensitivity decline or zero point drift caused by irradiation aging is compensated in real time. For example, when the photoelectric conversion efficiency of the sensor is reduced due to long-term radiation, the online calibration module reversely compensates the signal attenuation by increasing the gain coefficient, so that the output data is restored to a reliable range.

[0052] It can also be understood that, through the above technical solutions, the embodiment effectively suppresses the systematic measurement error of the sensor caused by irradiation aging, ensures the long-term stability of the point cloud data acquisition, provides initial robustness for the calibration model through offline training, and corrects the deviation caused by environmental interference in real time through an online adaptive mechanism. The double mechanisms work together to enable the acquisition module to still output high-precision point cloud data under strong radiation conditions, thereby laying a reliable data foundation for subsequent geometric reconstruction and twin model construction.

[0053] In the embodiment, the preprocessing of the obtained point cloud data specifically includes: removing spatial isolated noise points caused by medium scattering in a single frame of point cloud by using a radial outlier removal algorithm; and then comparing the point clouds of consecutive time frames to identify and remove transient point clusters.

[0054] It can be understood that the preprocessing process is implemented in two stages: first, in the single frame dimension, the spatial isolated noise points formed by aerosol scattering are quickly filtered out through a spherical neighborhood point number statistical mechanism. Such noise points are usually far away from the main point cloud and have a neighborhood density significantly lower than that of the real surface points. Then, in the time dimension, the transient point clusters appearing only in a single frame or a few frames are identified by analyzing the stability of the three-dimensional voxel grid in consecutive scanning frames. Such noise is generated by dynamic medium interference and presents spatial position randomness in the time sequence. The two stages are respectively directed to the spatial isolation and time transient characteristics of the two types of noise, forming a complementary noise reduction mechanism. The spatial filtering eliminates the abnormal values of single scanning data, and the time filtering suppresses the dynamic interference signal, which together ensures that the preprocessed point cloud data conforms to the real surface distribution rule.

[0055] In the embodiment, the preprocessing of the acquired point cloud data further includes: outputting a hidden state vector through a gated recurrent unit, and outputting corrected point cloud data through a multi-layer perception.

[0056] It can be understood that the essence of the above preprocessing process is that the feature vector is input into a degradation state estimator to output a hidden state vector reflecting the health state of the acquisition module through a gated recurrent unit, and the hidden state vector is input into a data corrector to output corrected point cloud data of the acquisition module through a multi-layer perception.

[0057] It can also be understood that the feature vector construction stage normalizes the sensor operating parameters and splices them into a multi-dimensional input vector, for example, linearly combining the log-scaled compressed radiation dose cumulative value and the temperature drift. The degradation state estimator filters historical information through a gating mechanism, for example, using a forget gate to control the proportion of old state retention and an update gate to adjust the fusion strength of new features, thereby dynamically tracking the sensor performance degradation trajectory. The hidden state vector is input into the data corrector as a health state quantitative indicator, for example, mapping the hidden vector to a three-dimensional space transformation matrix through a fully connected layer to perform affine transformation compensation on the original point cloud coordinates. The multi-layer perception further introduces a nonlinear activation function, for example, using a ReLU function to construct a residual connection to eliminate local distortions caused by sensor noise.

[0058] In the embodiment, the radial outlier removal algorithm for removing spatial isolated noise points caused by medium scattering in a single frame of point cloud data specifically includes the following steps: defining a spherical neighborhood for each point and calculating the number of points in the neighborhood, if the number of points in the neighborhood of a point is less than a predetermined threshold, the point is determined to be an outlier and removed.

[0059] It should be noted that in a strong scattering medium interference scene, real surface points usually exhibit spatial continuity characteristics, while noise points generated by aerosol scattering often exhibit random isolated distribution; by calculating the number of adjacent points of each point in its spherical neighborhood, the local density of the region can be quantified. When the neighborhood density of a point is lower than a predetermined threshold, it indicates that the point lacks sufficient adjacent support points, which meets the spatial distribution characteristics of isolated noise points. The density screening mechanism established in this way can effectively distinguish real surface points from scattering noise and accurately remove noise points

[0060] In the embodiment, the identification and removal of instantaneous point clusters specifically include the following steps: extracting multi-scale geometric features using a graph convolution network, and approximating an implicit signed distance function through a multi-layer perception, and then adding constraints to the signed distance function to obtain a signed distance function field that satisfies the constraints, the output value of the signed distance function representing the distance from a spatial point in the point cloud data to the nearest surface.

[0061] It can be understood that the embodiment extracts multi-scale features of the original point cloud through the graph convolution network, establishes geometric correlation description at the point level and the neighborhood level respectively, and enhances the recognition ability of the non-uniform noise distribution caused by the aerosol medium. Subsequently, an implicit signed distance function is constructed by using a multilayer perceptron to map the three-dimensional space coordinates to the corresponding signed distance values. The process approximates the real surface distance field through supervised learning, so that the abnormal distance values corresponding to the noise points are automatically suppressed in the implicit field. Further, by introducing gradient constraint and smoothing constraint, it is ensured that the change trend of the signed distance function in space conforms to the geometric characteristics of the real object surface, and the local distortion caused by the transient point cluster is eliminated, and finally a high-fidelity continuous distance field is generated for surface reconstruction.

[0062] In the embodiment, the point cloud of the re-contrast continuous time frame specifically includes the following steps: aligning the continuous multi-frame point cloud in a short time window, then dividing the space into a three-dimensional voxel grid, and finally distinguishing the real surface and the transient noise by analyzing the stability of each voxel in the time dimension.

[0063] It should be noted that the multi-frame point cloud data collected in the continuous time window is first aligned by registration to eliminate the coordinate deviation caused by the movement or vibration of the acquisition device. Then the three-dimensional space is divided into voxel units, and each voxel records its occupancy state in each time frame. The voxel corresponding to the real surface has spatial persistence and will exist stably in most time frames. The noise point cloud formed by the transient medium such as aerosol will only appear in individual frames. By counting the frequency of each voxel in the time window, the threshold is set to filter out the transient voxels with low frequency, and the stable voxels with high frequency are retained as real surface data. This spatio-temporal analysis method breaks through the limitation that the traditional single-frame denoising method cannot distinguish between transient noise and real structure.

[0064] Among them, the feature vector refers to a multi-dimensional data set formed by integrating sensor original readings, time stamps and operation logs, and preferably a sliding time window is used to intercept time sequence segments, and statistical quantities, frequency domain features and operation mode codes are extracted through feature engineering to realize this. The feature vector is used to represent the real-time working state of the acquisition module, and provides an input basis for subsequent degradation state estimation.

[0065] Among them, the degradation state estimator refers to a time sequence model constructed based on a gated recurrent unit, and preferably a bidirectional GRU network is used to capture the progressive degradation law of the sensor performance, and the device health degree is quantified through the hidden state vector. This module converts the unobservable sensor aging process into a calculable hidden variable to provide a state basis for data correction.

[0066] The data corrector is a nonlinear mapping network composed of multiple layers of perception mechanisms, preferably adopts a residual connection structure to learn the deviation mode of the original point cloud and the real space coordinates, and dynamically adjusts the correction parameters according to the hidden state vector, thereby realizing online compensation of the sensor aging error and suppressing the influence of data drift on subsequent modeling.

[0067] The radial outlier removal algorithm in the preprocessing refers to a point cloud filtering method based on spatial density distribution, preferably defines a spherical neighborhood for each point and counts the number of neighborhood points, and identifies isolated noise points by setting a density threshold; the algorithm effectively eliminates spatial abnormal points caused by medium scattering and retains the characteristics of the real surface structure.

[0068] The continuous time frame comparison technique refers to noise identification using the spatiotemporal correlation of multiple frames of point clouds, preferably uses a three-dimensional voxel grid to divide the spatial region and analyze the temporal stability of the point cloud in each voxel; the motion consistency principle is used to distinguish between static surfaces and transient interference, thereby improving the accuracy of noise removal.

[0069] Embodiment 2:

[0070] As shown in the accompanying Figure 2 The embodiment provides a twin model construction system for tank decommissioning, which comprises:

[0071] The acquisition module is configured to obtain point cloud data in the tank, and pre-process the obtained point cloud data to uniformly sample the point cloud data while removing noise points;

[0072] The reconstruction module is configured to sequentially perform normal vector estimation, creation of a Poisson equation, numerical method solution of the Poisson equation, and surface reconstruction to obtain a mesh model of a continuous surface from the point cloud data;

[0073] The extraction module is configured to project points in the three-dimensional point cloud in the mesh model to a two-dimensional plane to generate a two-dimensional point set to represent the distribution of the point cloud in the plane;

[0074] The construction module is configured to use the mesh model as morphological information reference and the two-dimensional point set as size information reference to construct a twin model of the tank in the twin layer.

[0075] In some embodiments, the system further comprises a correction module configured to construct a feature vector at each time point to output corrected point cloud data of the acquisition module through the degradation state estimator and the data corrector.

[0076] It can be understood that the acquisition module establishes a complete data stream containing time dimension information by synchronously recording original sensor readings and operation logs, the correction module aligns sensor data and device state based on timestamps, and the degradation state estimator dynamically compensates for drift caused by irradiation aging. The reconstruction module establishes local geometric features of point cloud through normal vector estimation, reconstructs discrete point cloud into implicit surface using Poisson equation, and generates topologically continuous mesh model through numerical solution, effectively eliminating geometric distortion caused by medium scattering. The extraction module projects the three-dimensional mesh model to the two-dimensional plane along the main normal direction, separates size information and spatial noise through dimension reduction processing. The construction module fuses morphological features expressed by continuous surface and plane distribution features of two-dimensional projection to form a twin model with geometric consistency and parameter analysis capability.

[0077] Embodiment 3:

[0078] Overall, the construction method in the present application includes four stages, one is a multi-modal data acquisition and drift online correction stage; two is a point cloud preprocessing and depth denoising stage, three is a high-fidelity three-dimensional geometric model reconstruction stage; four is a multi-dimensional information fusion and twin application stage,

[0079] For the first stage, the correction module includes a self-encoder and a self-constrained unit, the acquisition module includes a sensing unit, and the acquisition module is preferably a survey modeling robot equipped with a sensing unit, and the identifying and removing instantaneous point clusters specifically includes the following steps: using a graph convolution network to extract multi-scale geometric features, and approximating an implicit signed distance function through a multi-layer perception, and then adding constraints to the signed distance function to obtain a signed distance function field that satisfies the constraints, wherein the output value of the signed distance function represents the distance from a spatial point in the point cloud data to the nearest surface.

[0080] In some embodiments, in the step of adding constraints to the signed distance function, the added constraints include that the gradient norm of the signed distance function is 1.

[0081] The goal of this stage is to construct an intelligent correction model that can perceive the health state changes of the sensing unit caused by cumulative damage due to irradiation in real time, and dynamically adjust its output data based on the state. This stage can be divided into the following steps:

[0082] Step 1: Synchronous data acquisition and dynamic feature construction, that is, obtaining point cloud data in the tank, and preprocessing the obtained point cloud data to achieve uniform sampling of the point cloud data while removing noise points;

[0083] This step not only collects original data, but more importantly, constructs a dynamic feature vector that can reflect the historical cumulative damage of the sensor, providing input for the subsequent degradation perception model.

[0084] Step 1.1: Multi-modal data synchronous acquisition: The acquisition module records the following data streams synchronously during the operation:

[0085] Raw sensor readings : Unmodified measurements of each sensor at time point t ;

[0086] Timestamp t: Records the length of time the sensor is in service;

[0087] Operation log: Records the specific actions of the robot, which are used to trigger subsequent virtual calibration.

[0088] Step 1.2: Cumulative dose feature construction: The direct cause of sensor performance degradation is the cumulative received radiation dose. Since it cannot be directly measured, it is dynamically estimated by integrating the historical readings of the sensor itself.

[0089] At discrete time steps, the estimated cumulative dose at time point

[0090] ;

[0091] is the estimated cumulative radiation dose at time point t k ;

[0092] is the raw sensor reading at time point t k , such as the raw radiation dose rate reading;

[0093] is the time interval between two consecutive sampling points.

[0094] Finally, at each time point, we construct a feature vector which not only contains the current measurement value, but also contains key historical information reflecting the life history of the sensor.

[0095] Step 2: Degradation-aware recurrent correction network modeling;

[0096] It can be understood that the degradation-aware recurrent correction network is a dual-module deep learning model that can judge the credibility of the sensor's readings and correct them according to the length of service and the degree of strain of the sensor, just like an experienced engineer. Specifically:

[0097] Degradation state estimator: A recurrent neural network based on gated recurrent unit (GRU) that processes time series features X 1, X 2, …, X kand output a hidden state vector that represents the current health status of the sensor h k .

[0098] The update process of the gated recurrent unit can be summarized as:

[0099] ;

[0100] where, X k is the input feature vector at time point k ;

[0101] h k-1 the hidden state vector at the previous time step, contains all the historical information from t 0 to t k-1 ;

[0102] h k the hidden state vector at the current time step, can be regarded as the degradation signature of the sensor unit at this time step.

[0103] is a complex nonlinear transformation function representing the recurrent neural network of the gated recurrent unit.

[0104] Data corrector: a relatively simple multi-layer perceptron (MLP) that takes as joint input the current raw reading and the degradation state vector generated by the GRU h k and outputs the final corrected data .

[0105] The expression satisfies:

[0106] ;

[0107] where, denotes the concatenation of the current raw reading and the degradation state vector, forming a more comprehensive input;

[0108] represents the nonlinear mapping function of the MLP network.

[0109] Step 3: Hybrid training strategy of offline pre-training and online adaptive fine-tuning:

[0110] Since ground truth is not available inside the tank for supervised learning, the sensors are subjected to accelerated aging experiments in the lab before deployment, where both sensor readings and high-precision instrument ground truth are available under controlled radiation sources to supervise the degradation-aware recurrent network.

[0111] where the loss function is:

[0112] ;

[0113] where, M is the total number of samples in the pre-training dataset;

[0114] is the modified output of the model for the i th sample;

[0115] is the lab ground truth for the i th sample;

[0116] For online self-adaptive fine-tuning: after deploying the pre-trained model to the tank, it needs to adjust itself according to the real environment, by utilizing physical constraints and sparse virtual calibration points as unsupervised or self-supervised signals.

[0117] Online loss function:

[0118] ;

[0119] Physical residual loss L phy Similar to the previous scheme, penalize those modified results that violate known physical laws (e.g. thermodynamic equations, radioactive decay laws).

[0120] ;

[0121] is the governing equation describing the physical law;

[0122] Virtual calibration loss L cal is an innovative loss term that can obtain a temporary, sparse ground truth when the robot performs certain operations with deterministic results (e.g. a tool contacts a new tool head surface known to be non-radioactive, at which time the radiation reading should theoretically be the environmental background value).

[0123] The expression satisfies:

[0124] ;

[0125] C is the virtual calibration eventc a set of the plurality of features;

[0126] t c for calibrating the event c a time point at which the event occurs;

[0127] for calibrating the event c a corresponding theoretical true value;

[0128] is a weight hyperparameter balancing different loss terms.

[0129] Embodiment 4:

[0130] In this embodiment, the constructing a feature vector specifically comprises the following steps: inputting the feature vector into a degradation state estimator to output a hidden state vector reflecting the health state of the acquisition module through a gated loop, and inputting the hidden state vector into a data corrector to output corrected point cloud data of the acquisition module through a multi-layer perception.

[0131] Before the point cloud data is acquired, the method further comprises:

[0132] Before the acquisition module works, the acquisition module acquires historical data of the acquisition module in an ideal environment, performs supervised training through a loss function, and performs self-adaptive calibration through an online loss function.

[0133] The radial outlier removal algorithm specifically comprises the following steps: defining a spherical neighborhood for each point and calculating the number of points in the neighborhood, and if the number of points in the neighborhood of a point is less than a preset threshold, the point is determined as an outlier and removed.

[0134] For the second stage: multi-stage mixed point cloud preprocessing and depth denoising;

[0135] The goal of this stage is to restore the most physically realistic geometric surface inside the storage tank from the original point cloud that is severely contaminated. Through this embodiment, three typical noises caused by medium scattering, suspended particles and dense aerosols can be systematically stripped: sparse outliers, transient noise clusters and boundary layer artifacts. This stage specifically comprises the following steps:

[0136] Step 2.1: Outlier coarse filtering based on neighborhood density;

[0137] This step is the first line of defense for preprocessing, aiming to quickly remove sparse outliers that are completely isolated in space and formed by single laser beam scattering. A radial outlier removal algorithm (ROR) is used, which is more robust than traditional statistical filtering (SOR) when the point cloud density is uneven.

[0138] The ROR algorithm defines a spherical neighborhood for each point and counts the number of points within the neighborhood. If the number of points within a point's neighborhood is less than a pre-set threshold, the point is determined to be an outlier and removed.

[0139] The core logic is:

[0140] For any point in the point cloud P p i The retention condition is:

[0141] ;

[0142] p i The current point to be judged is:

[0143] p j Other points in the point cloud are:

[0144] The Euclidean distance between the point p i and the point p j ;

[0145] r The defined neighborhood search radius is a key parameter that needs to be set according to the average density of the point cloud;

[0146] The radius r is denoted as ;

[0147] This method can effectively remove isolated noise points in the background and provide a relatively clean input for subsequent processing.

[0148] Step 2.2: Transient noise cluster filtering based on spatiotemporal consistency;

[0149] This step is specifically aimed at transient noise clusters caused by moving suspended particle groups in aerosols and water vapor. Although these noise points may have a relatively high local density in a single frame, they are unstable and discontinuous in the time series. A spatiotemporal voxel grid filtering algorithm is used to identify and remove them.

[0150] First, align the continuous multiple frames of point cloud within a short time window, then divide the space into a three-dimensional voxel grid, and finally distinguish between real surfaces and transient noise by analyzing the stability of each voxel in the time dimension.

[0151] Use the Iterative Closest Point (ICP) algorithm or its variants to align the time window T ={ t N ,…, t ,…, t+N 2 within} N +1 frame: All point clouds are aligned to the reference frame. t In the coordinate system;

[0152] For each voxel in space v Calculate its time window T Occupancy probability within P occ ( v );

[0153] The expression satisfies:

[0154] ;

[0155] Among them, indicator function Defined as:

[0156] ;

[0157] in, v It is a three-dimensional voxel unit;

[0158] k The frame index within the time window;

[0159] N The radius of the time window (e.g., N=2 means using 5 consecutive frames of data);

[0160] P occ ( v (voxel) v The frequency with which a content is occupied within a given time window reflects its temporal stability.

[0161] Then set another probability threshold τ (e.g., τ=0.8) in the reference frame. t In the middle, all those who fall into P occ ( v Points within voxels with a value less than τ will be treated as transient noise and removed.

[0162] Step 2.3: Boundary layer artifact correction and fine denoising based on graph neural hidden network;

[0163] This step is the core of the whole denoising process, aiming to solve the problem of boundary layer artifacts; dense aerosol will form a layer of false, continuous pseudo-surface near the surface of the real object, and traditional denoising methods are difficult to handle. For this purpose, a graph neural implicit network (GINN) is designed, which does not directly output points, but learns the continuous mathematical expression of the target object surface, i.e. the signed distance function (SDF), so as to fundamentally separate the real surface from the artifacts.

[0164] The model constructs a graph from the point cloud processed in the previous two steps, uses a graph convolution network (GCN) to extract multi-scale geometric features, and trains a multi-layer perceptron (MLP) to approximate an implicit SDF. SDF is a function f(x,y, z) whose output value represents the distance of a spatial point (x,y,z) to the nearest surface, and the sign indicates whether the point is inside (negative) or outside (positive) the object. The real surface is the zero level set of SDF f(x,y,z) =0.

[0165] The goal of network training is to learn an SDF approximation function defined by parameters θ , whose loss function L is composed of data items L data and regularization terms L eikonal , and satisfies the expression:

[0166] ;

[0167] For the data loss term :

[0168] This term forces the SDF value of all input (noisy) surface points p to approach zero;

[0169] ;

[0170] P s is the input point cloud set;

[0171] p is a point in the point cloud;

[0172] is the SDF prediction value of the neural network for point p ;

[0173] For the regularization term :

[0174] This term is a physical constraint to require the gradient norm (i.e. spatial variation rate) of SDF everywhere should be 1, which guarantees the learned function is a valid distance field, greatly improving the stability and generalization ability of the model.

[0175] The expression satisfies:

[0176] ;

[0177] is a set of query points randomly sampled in space;

[0178] q is a spatial query point;

[0179] is the gradient of the neural network predicted SDF at point q ;

[0180] is the L2 norm (i.e. module length) of the vector;

[0181] is a weight coefficient used to balance the data term and the regularization term.

[0182] After training is completed, a continuous and smooth SDF field is obtained, and finally, we use the marching cubes algorithm to extract the zero level set from the SDF field, i.e. to generate a high-quality, noise-free, and detailed final three-dimensional mesh model. This process not only removes all types of noise, but also reasonably repairs holes caused by occlusion or data sparsity.

[0183] Embodiment 5:

[0184] For the third stage, the preprocessing of the acquired point cloud data specifically includes: using a radial outlier removal algorithm to remove spatial isolated noise points in a single frame of point cloud caused by medium scattering; and then comparing the point clouds of continuous time frames to identify and remove transient point clusters.

[0185] The step of constructing the twin model of the storage tank in the twin layer specifically includes the following steps: applying all vertices in the size information reference to the mesh model in the shape information reference by using an inverse distance weighting method to construct the twin model of the storage tank. In addition, the attribute value of a point in the size information reference space is obtained by weighted average of the attribute values of the surrounding known points, and the size of the weight is inversely proportional to the distance from the known points.

[0186] In the step of planning the operation path and cutting sequence of the storage tank by using the twin model, an evaluation function is obtained by global path planning to find the optimal node for expansion;

[0187] The point cloud of the re-contrast continuous time frame specifically includes the following steps: aligning the continuous multi-frame point cloud in a short time window, then dividing the space into a three-dimensional voxel grid, and finally distinguishing the real surface and transient noise by analyzing the stability of each voxel in the time dimension.

[0188] Third stage: high-fidelity three-dimensional geometric model reconstruction;

[0189] For normal vector estimation, the premise of Poisson reconstruction is that the point cloud needs to be equipped with accurate normal vector information, so principal component analysis (PCA) is used to estimate the normal vector of each point. Specifically:

[0190] For each point p i , select its K nearest neighbor points to form a local neighborhood, calculate the covariance matrix C of the neighborhood points.

[0191] The expression satisfies:

[0192] ;

[0193] Where, p j is p i the point in the neighborhood;

[0194] is the centroid of the neighborhood points;

[0195] The smallest eigenvalue of the covariance matrix C corresponding to the eigenvector is the normal vector of the local plane where the point p i is located n i 。

[0196] For the construction and solution of the Poisson equation; the core idea of Poisson surface reconstruction is to regard the directional point cloud (including points and their normal vectors) as a divergence sample of a three-dimensional vector field V, and then find an indicator function χ by solving a Poisson equation, whose gradient best approximates the vector field V. The isosurface of this indicator function defines the reconstructed object surface.

[0197] The expression satisfies: ;

[0198] is the Laplacian operator, representing the second derivative of the function;

[0199] χ is a scalar indicator function, whose value is 1 inside the object and 0 outside, i.e. the function to be solved;

[0200] is the divergence operator;

[0201] V is a smooth three-dimensional vector field constructed from the normal vectors of the input point cloud.

[0202] for isosurface extraction;

[0203] Finally, the isosurface (typically χ = 0.5) is extracted from the solved scalar field χ using the marching cubes algorithm, resulting in a final continuous, watertight triangular mesh model.

[0204] Embodiment 6:

[0205] For the fourth stage: multi-dimensional information fusion and twin application;

[0206] The goal of this stage is to transform the static three-dimensional geometric model into a dynamic, multi-physical field coupled digital twin, and based on this, to realize intelligent planning of the operation path and simulation deduction of the decommissioning process, which includes the following steps:

[0207] Step 4.1: Multi-physical field mapping based on inverse distance weighted interpolation;

[0208] This step aims to accurately and smoothly map the corrected discrete sensor data (such as radiation dose rate, temperature, etc.) from the first stage onto the high-fidelity triangular mesh model generated in the third stage, forming a continuous physical field distribution map.

[0209] Inverse distance weighting (IDW) is used to calculate the physical quantity values at any position on the grid model based on sparse sensor measurement points; in addition, the attribute value of a point in space can be obtained by weighted average of the attribute values of its surrounding known points, and the weight is inversely proportional to the distance from the known point.

[0210] For any vertex v or face center point p target on the grid model, the physical quantity value (e.g., radiation dose rate) to be solved can be calculated by the following formula:

[0211] ;

[0212] where the calculation method of the weight ω i is:

[0213] ;

[0214] is the point to be solved ptarget Estimates of physical quantities;

[0215] N is the number of nearest neighbor sensor data points used for interpolation;

[0216] S i For the first i The measured values ​​of physical quantities (corrected) at each sensor measurement point;

[0217] p i For the first i The three-dimensional spatial coordinates of each sensor;

[0218] Point to be sought p target A point with the sensing unit p i The three-dimensional Euclidean distance between them. This distance is calculated directly in three-dimensional space to ensure accuracy.

[0219] k The exponent is a positive real number used to adjust the degree to which the weight changes with distance. k The larger the value, the greater the weight of points that are closer together, and the more obvious the local features of the interpolation result. k The value is 2.

[0220] By applying the IDW algorithm to all vertices of the mesh model, a continuously changing visualization layer covering the entire inner wall of the tank, including radiation and temperature fields, can be generated, providing an intuitive data foundation for subsequent analysis and planning.

[0221] Example 7:

[0222] Step 4.2: Intelligent path planning for robot operations based on hybrid algorithms;

[0223] This step is one of the core applications of digital twins. By using a twin model that integrates multiphysics, it plans an optimal working path for operating robots (such as robotic arms) that balances collision-free operation, high efficiency, and low radiation.

[0224] First, the A* algorithm is used to search for an optimal path from the starting point to the ending point in the robot's configuration space. The configuration space is a multi-dimensional space composed of the angles of all the robot's joints as coordinate axes. The robot body is simplified to a point, while obstacles in the environment are expanded into no-go zones in this space.

[0225] The A* algorithm uses an evaluation function. f(n) To find the optimal node for expansion:

[0226] ;

[0227] To meet this special requirement of low irradiation, the traditional cost function is improved:

[0228] f(n) is the estimated total cost from the start point to the goal point via node n .

[0229] is the actual cumulative cost from the start point to node n, defined as the weighted sum of path length and cumulative radiation dose:

[0230] and satisfies:

[0231] ;

[0232] L i is the length of path segment i ;

[0233] is the radiation dose rate of the robot end effector at path point p i , which is directly queried from the twin model we built;

[0234] is the time required to pass path segment i ;

[0235] and are the weight coefficients for path length and radiation dose respectively, which can be adjusted according to task priority;

[0236] h(n) is the heuristic estimated cost from node n to the goal point (usually using Euclidean distance).

[0237] Embodiment 8:

[0238] For local path smoothing and real-time obstacle avoidance (artificial potential field method): the path planned by the A* algorithm is a broken line composed of a series of discrete nodes, which may not be smooth or may collide with unmodeled dynamic obstacles, therefore, the artificial potential field method is introduced to optimize and adjust the global path in real time.

[0239] The resultant force on the robot in the potential field F(q) is composed of attractive force F att (q) and repulsive force F rep (q) , which guides the robot to move in the direction of fastest potential energy descent.

[0240] Embodiment 9:

[0241] The embodiment provides a computer device, comprising a memory and a processor, the memory stores a computer program, and the processor executes the computer program.

[0242] The memory can comprise a non-permanent memory in a computer readable medium, random access memory (RAM) and / or non-volatile memory such as read-only memory (ROM) or flash memory (flash RAM), and the memory comprises at least one memory chip.

[0243] Embodiment 10:

[0244] The embodiment provides a computer readable storage medium, and the computer readable storage medium stores a computer program, and the processor executes the computer program.

[0245] The embodiment of the present application provides a processor, which is used for running a program, wherein the program is executed to perform the twin model construction method for the tank decommissioning.

[0246] The embodiment of the present application provides a device, which comprises a processor, a memory and a program stored in the memory and executable on the processor, and the processor executes the program to realize at least the steps of the twin model construction method for the tank decommissioning.

[0247] The device in the present application can be a server, a PC, a PAD, a mobile phone and the like.

[0248] The present application further provides a computer program product, which is adapted to execute the program initialized with at least the steps of the twin model construction method for the tank decommissioning when executed on a data processing device.

[0249] Obviously, those skilled in the art should understand that the modules or steps of the present application described above can be realized by general computing devices, which can be concentrated on a single computing device or distributed on a network composed of multiple computing devices, and they can be realized by program codes executable by the computing devices, so that they can be stored in storage devices and executed by the computing devices, and in some cases, the steps shown or described can be executed in different sequences, or they can be manufactured into individual integrated circuit modules or a single integrated circuit module. Thus, the present application is not limited to any specific combination of hardware and software.

[0250] Those skilled in the art will appreciate that embodiments of the application can be devised for a method, a system, or a computer program product. Accordingly, the present application can be embodied in the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer readable storage media (including, but not limited to, disk memory, CD-ROMs, optical storage devices, etc.) embodying computer readable program code.

[0251] The present application is described in reference to the flowchart and / or block diagrams of the method, apparatus (system) and computer program product according to embodiments of the application illustrated in the attached drawings, which are provided for purposes of explanation and should not be considered as limiting. Each flow and / or block in the flowchart and / or block diagrams, and combinations of flows and / or blocks in the flowchart and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0252] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0253] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0254] In one typical configuration, the computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0255] The memory can include non-persistent memory and / or volatile memory, such as random access memory (RAM) and / or cache memory. The memory can also include non-volatile memory, such as read only memory (ROM), electrically programmable read only memory (EPROM), or electrically erasable programmable read only memory (EEPROM), for example. Memory is an example of computer readable media.

[0256] Computer-readable media includes permanent and non-permanent, movable and non-movable media that can be implemented by any method or technology to store information; the information can be computer-readable instructions, data structures, program modules or other data.

[0257] Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD), or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible to computing devices. According to the definition herein, computer-readable media does not include transitory computer-readable media, such as modulated data signals and carrier waves.

[0258] The above is only the preferred embodiment of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent process transformation using the content of the specification and drawings of the present application, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A twin model construction method for a tank decommissioning, characterized by, The method is used for constructing a twin model of a storage tank, and comprises the following steps: Obtaining point cloud data in the storage tank, and pre-processing the obtained point cloud data to uniformly sample the point cloud data while removing noise points; Sequentially performing normal vector estimation, creating a Poisson equation, solving the Poisson equation by a numerical method, and surface reconstruction to obtain a mesh model of a continuous surface from the point cloud data; Projecting points in the three-dimensional point cloud in the mesh model onto a two-dimensional plane to generate a two-dimensional point set to represent the distribution of the point cloud in the plane; Taking the mesh model as morphological information reference and taking the two-dimensional point set as size information reference, a twin model of the storage tank is constructed in a twin layer; The method further comprises: Planning an operation path and a cutting sequence of the storage tank by using the twin model; The pre-processing of the obtained point cloud data specifically comprises: Removing spatial isolated noise points caused by medium scattering in a single frame of point cloud by using a radial outlier removal algorithm, including the following steps: defining a spherical neighborhood for each point, and calculating the number of points in the neighborhood, if the number of points in the neighborhood of a point is less than a preset threshold, the point is determined as an outlier and removed; Comparing point clouds of consecutive time frames to identify and remove transient point clusters; The identification and removal of the transient point clusters specifically comprise the following steps: extracting multi-scale geometric features by using a graph convolution network, and approximating an implicit signed distance function by using a multi-layer perception, and adding constraints to the signed distance function to obtain a signed distance function field satisfying the constraints, the output value of the signed distance function representing the distance from a spatial point in the point cloud data to the nearest surface; The pre-processing of the obtained point cloud data further comprises: inputting a time series feature vector into a degradation state estimator to output a hidden state vector reflecting the health status of the sensor through a gated loop, and inputting the hidden state vector into a data corrector to output corrected point cloud data of the sensor through a multi-layer perception.

2. The twin model construction method for tank decommissioning according to claim 1, wherein, Before the obtaining of the point cloud data in the storage tank, the method further comprises: Obtaining historical data in an ideal environment to perform supervised training by using a loss function, and performing adaptive calibration by using an online loss function.

3. The twin model construction method for tank decommissioning according to claim 1, wherein The comparison of the point clouds of consecutive time frames specifically comprises the following steps: aligning consecutive multiple frames of point clouds in a short time window, then dividing the space into a three-dimensional voxel grid, and finally distinguishing real surfaces and transient noise by analyzing the stability of each voxel in the time dimension.

4. A twin model construction system oriented to the decommissioning of a tank, characterized by, The system comprises: An acquisition module, which is configured to obtain point cloud data in the storage tank, and pre-process the obtained point cloud data to uniformly sample the point cloud data while removing noise points; A reconstruction module, which is configured to sequentially perform normal vector estimation, create a Poisson equation, solve the Poisson equation by a numerical method, and perform surface reconstruction to obtain a mesh model of a continuous surface from the point cloud data; An extraction module, which is configured to project points in the three-dimensional point cloud in the mesh model onto a two-dimensional plane to generate a two-dimensional point set to represent the distribution of the point cloud in the plane; The construction module is used to take the grid model as morphological information reference, take the two-dimensional point set as dimensional information reference, and construct a twin model of the tank in a twin layer; The acquisition module is further configured to remove spatial isolated noise points caused by medium scattering in the single-frame point cloud by using a radial outlier removal algorithm, including the following steps: defining a spherical neighborhood for each point, and calculating the number of points in the neighborhood; if the number of points in the neighborhood of a point is less than a preset threshold, the point is determined as an outlier and removed; The point clouds of the continuous time frames are compared again to identify and remove transient point clusters; The identification and removal of the transient point clusters specifically include the following steps: extracting multi-scale geometric features by using a graph convolution network, and approximating an implicit signed distance function by using a multi-layer perception; a constraint is added to the signed distance function to obtain a signed distance function field satisfying the constraint, and an output value of the signed distance function represents a distance from a spatial point in the point cloud data to a nearest surface; The acquisition module is further configured to input the time sequence feature vector into a degradation state estimator to output a hidden state vector reflecting a sensor health state through a gated recurrent unit, and input the hidden state vector into a data corrector to output corrected point cloud data of the sensor through a multi-layer perception.

5. A computer device, comprising: The computer device includes a memory and a processor, the memory stores a computer program, and the processor executes the computer program to implement the method in any one of claims 1-3.

6. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the processor executes the computer program to implement the method in any one of claims 1-3.

Images