Concrete surface intelligent construction method and system based on multi-modal sensing

Abstract

The invention discloses a concrete surface intelligent construction method and system based on multi-modal sensing. The method comprises the following steps: firstly, receiving multi-source sensing data, and generating standardized feature data and defect feature vectors through standardization processing and feature extraction; performing time sequence alignment processing on the sensing data, and generating time sequence alignment data in combination with multi-scale analysis; constructing a state evaluation model based on the dynamic Bayesian network, and generating an environment state vector and a construction difficulty graph; carrying out region segmentation and optimizing task allocation by adopting a spectral clustering algorithm, and generating a task allocation scheme and a collaboration strategy; then, construction parameters are determined through multi-objective optimization, tool configuration is planned, and an optimization parameter set and a tool configuration scheme are generated; and finally, trajectory planning is carried out based on an improved A * algorithm, and a multi-robot cooperative construction trajectory is generated through comprehensive optimization. Full-process control of intelligent construction of the concrete surface is achieved, and the construction quality and efficiency are improved.

Description

Technical Field

[0001] The invention relates to intelligent construction, in particular to an intelligent construction method based on a leveling, trowelling and finishing robot. Background Art

[0002] The quality of concrete surface construction directly affects the appearance quality and service life of the building, and is a key link in the construction of construction projects. With the increase in large-scale construction projects and the development of intelligent construction technology, traditional manual operations have been unable to meet the construction requirements of high efficiency, high quality and high consistency. The use of robots for intelligent construction of concrete surfaces can not only improve construction efficiency and quality stability, but also reduce the labor intensity of workers, which has important engineering application value and social benefits.

[0003] At present, concrete surface construction robots mainly use a single sensor for environmental perception and construction control. For example, the surface defect detection system based on visual sensing uses image processing algorithms to identify surface defects such as bubbles and cracks; the surface flatness detection system based on laser scanning evaluates the surface undulation through point cloud data analysis; the construction force control system based on pressure sensing adjusts the construction parameters through real-time pressure feedback. In terms of multi-robot collaboration, existing research mainly uses preset task allocation schemes and fixed construction parameters, lacking the ability to dynamically perceive and adaptively adjust the construction environment. In terms of trajectory planning, the path optimization of a single robot is mainly considered, and the collaborative obstacle avoidance and tool switching of multiple robots are not considered enough.

[0004] The current intelligent construction technology for concrete surfaces has the following problems: First, the perception ability of a single sensor is limited, and it is difficult to obtain surface state information comprehensively and accurately, especially in harsh environments such as complex lighting and high humidity, where the reliability of sensor data is poor; second, the time inconsistency problem of multi-source heterogeneous sensor data is prominent, and the sampling rates, response delays and data formats of different sensors are different, making it difficult to achieve accurate alignment and fusion of data; third, the state assessment method lacks an effective processing mechanism for parameter uncertainty, making it difficult to accurately assess the difficulty level and environmental change trend of the construction area; fourth, task planning lacks sufficient consideration of the robot's ability characteristics, and the preset allocation scheme is difficult to adapt to the complex and changeable construction environment; fifth, construction parameter optimization lacks overall consideration of multi-objective constraints, making it difficult to balance efficiency and energy consumption while ensuring construction quality; sixth, trajectory planning fails to effectively handle the coupling relationship between tool switching and path optimization, making it difficult to ensure the continuity and stability of the construction process. These technical problems have seriously restricted the engineering application of intelligent construction technology for concrete surfaces. Summary of the invention

[0005] The purpose of the invention is to provide a method and system for intelligent construction of concrete surface based on multimodal perception, in order to solve one of the above-mentioned problems existing in the prior art.

[0006] The technical solution provides a method for intelligent construction of concrete surface based on multimodal perception, comprising the following steps:

[0007] Step S1, receiving original sensor data sets, including original pressure data of the pressure sensor array, original spectral data of the near-infrared spectral sensor, original point cloud data of the 3D laser scanner, original temperature data of the temperature sensor, original ultrasonic data of the ultrasonic detector, and original environmental parameter data of the environmental parameter sensor, performing standardization processing and feature extraction on various types of data, and generating standardized feature data and defect feature vectors;

[0008] Step S2, reading standardized feature data, performing time series segmentation and time alignment processing on the sensor data, performing multi-scale analysis in combination with a pre-stored time series processing parameter library, and generating time series alignment data;

[0009] Step S3, reading the time alignment data and defect feature vector, combining with the pre-stored state assessment parameter library, building a Bayesian network model, and generating an environmental state vector and a construction difficulty map;

[0010] Step S4, read the construction difficulty map and the environment state vector, combine with the pre-stored task planning parameter library, divide the construction area into subtasks and optimize the allocation, generate the task allocation plan and coordination strategy;

[0011] Step S5, reading the environment state vector, the task allocation scheme and the defect feature vector, optimizing the construction parameters and tool configuration in combination with the pre-stored process parameter library, and generating an optimized parameter set and tool configuration scheme;

[0012] Step S6: read the construction difficulty map, task allocation plan, collaborative strategy and optimization parameter set, and generate a collaborative construction trajectory of multiple robots in combination with a pre-stored path planning parameter library.

[0013] According to another aspect of the present application, a concrete surface intelligent construction system based on multimodal perception is also provided, comprising:

[0014] at least one processor; and,

[0015] a memory communicatively connected to at least one of the processors; wherein,

[0016] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the intelligent construction method of concrete surface based on multimodal perception described in any of the above technical solutions.

[0017] Beneficial effect: it realizes the full-process control of intelligent construction of concrete surface and improves the construction quality and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a flow chart of the method of the present invention.

[0019] Figure 2 is a flow chart of step S1 of the method of the present invention.

[0020] Figure 3 It is a flow chart of step S2 of the method of the present invention.

[0021] Figure 4 It is a flow chart of step S3 of the method of the present invention.

[0022] Figure 5 It is a flow chart of step S4 of the method of the present invention.

[0023] Figure 6 It is a flow chart of step S5 of the method of the present invention.

[0024] Figure 7 It is a flow chart of step S6 of the method of the present invention. DETAILED DESCRIPTION

[0025] like Figure 1 As shown, a method for intelligent construction of concrete surface based on multimodal perception includes the following steps:

[0026] Step S1, receiving original sensor data sets, including original pressure data of the pressure sensor array, original spectral data of the near-infrared spectral sensor, original point cloud data of the 3D laser scanner, original temperature data of the temperature sensor, original ultrasonic data of the ultrasonic detector, and original environmental parameter data of the environmental parameter sensor, performing standardization processing and feature extraction on various types of data, and generating standardized feature data and defect feature vectors;

[0027] Step S2, reading standardized feature data, performing time series segmentation and time alignment processing on the sensor data, performing multi-scale analysis in combination with a pre-stored time series processing parameter library, and generating time series alignment data;

[0028] Step S3, reading the time alignment data and defect feature vector, combining with the pre-stored state assessment parameter library, building a Bayesian network model, and generating an environmental state vector and a construction difficulty map;

[0029] Step S4, read the construction difficulty map and the environment state vector, combine with the pre-stored task planning parameter library, divide the construction area into subtasks and optimize the allocation, generate the task allocation plan and coordination strategy;

[0030] Step S5, reading the environment state vector, the task allocation scheme and the defect feature vector, optimizing the construction parameters and tool configuration in combination with the pre-stored process parameter library, and generating an optimized parameter set and tool configuration scheme;

[0031] Step S6: read the construction difficulty map, task allocation plan, collaborative strategy and optimization parameter set, and generate a collaborative construction trajectory of multiple robots in combination with a pre-stored path planning parameter library.

[0032] Through the organic combination of the six core links of multimodal perception, time series processing, state evaluation, task planning, parameter optimization and trajectory control, a full-process solution for intelligent construction of concrete surfaces is realized. Specifically, firstly, a unified perception of the construction environment is established through the standardized processing and feature extraction of multi-source sensor data; then, the temporal consistency of the perception data is achieved through time series alignment and multi-scale analysis; then, state evaluation and construction difficulty analysis are performed based on the dynamic Bayesian network; on this basis, the collaborative planning of multiple robots is realized through regional segmentation and task optimization; at the same time, the construction effect is guaranteed through parameter optimization and tool configuration; finally, safe and efficient collaborative control is achieved through comprehensive trajectory optimization. The entire technical solution forms a closed-loop processing flow from perception to control. Through multi-level optimization strategies and a complete constraint processing mechanism, it significantly improves the intelligence level and construction quality of concrete surface treatment, and has important engineering application value.

[0033] like Figure 2 As shown, according to one aspect of the present application, step S1 is specifically:

[0034] Step S11, receiving the original sensor data set, normalizing various types of data to the interval [-1, 1], and generating a standardized sensor data set; reading the calibration coefficient from the pre-stored sensor calibration parameters, calibrating the standardized sensor data set, and generating a calibrated standardized data set;

[0035] Step S12, reading the original point cloud data, original ultrasonic data and original spectral data in the calibrated standardized data set, extracting local geometric features and material features; calculating the curvature, normal vector, depth information and material parameters of each local area, and constructing a feature description vector; performing feature matching in combination with a pre-stored feature template library to generate surface feature data;

[0036] Step S13, reading the surface feature data, using the region growing algorithm to identify connected regions with similar features, and generating defect candidate region data; matching the defect candidate region data with the pre-stored defect feature library, extracting feature parameters, and generating a defect feature vector;

[0037] Step S14, reading the calibrated standardized data set, combining the pre-stored sensor performance parameters, calculating the signal-to-noise ratio, integrity and reliability index of various types of data, and generating a data reliability matrix;

[0038] Step S15: input the calibrated standardized data set, defect feature vector and data reliability matrix into the feature fusion algorithm, assign weights to different features based on the data reliability matrix, and fuse them to generate standardized feature data.

[0039] The unified representation and fusion of multi-source heterogeneous data are achieved through standardized processing and feature extraction of multimodal sensor data. Specifically, the sensor data of different dimensions and ranges are first mapped to a unified interval through normalization processing, eliminating the impact of dimensional differences; then the calibration processing is performed based on the sensor calibration parameters to effectively compensate for the nonlinear error and zero drift of the sensor; on this basis, the local geometric feature and material feature extraction method is used to achieve the deep feature expression of the 3D point cloud, ultrasonic and spectral data of the concrete surface; further, the defect candidate area is adaptively identified through the regional growing algorithm, and matched with the defect feature library to accurately extract the surface defect features; finally, the weighted fusion of different features is performed based on the data reliability matrix, which not only ensures the integrity of feature extraction, but also reflects the confidence differences of different sensor data, thereby providing a reliable feature data basis for subsequent status assessment and construction planning. The entire data processing process forms a complete chain from raw data to standardized features, which significantly improves the utilization efficiency of multimodal perception data and the accuracy of feature extraction.

[0040] like Figure 3 As shown, according to one aspect of the present application, step S2 is specifically:

[0041] Step S21, read the standardized feature data, perform sliding time window segmentation according to the pre-stored time window parameters, and generate a time series segment set; perform data integrity check on the time series segment set, supplement missing data, and generate a complete time series set;

[0042] Step S22, read the complete time series set, combine the pre-stored timing alignment parameters, use the fast dynamic time warping algorithm to calculate the time distortion measure between different sensor data; construct the distortion path matrix, solve the optimal alignment path, and generate the aligned time series;

[0043] Among them, the dynamic time warping algorithm DTW is: cumulative distance function D(i,j) = d(i,j) + min{D(i-1,j-1), D(i-1,j), D(i,j-1)}; where: d(i,j) is the local distance between sequences; adaptive bandwidth function W(t) = Wbase + α•exp(-β•V(t)); where: Wbase is the basic bandwidth value; V(t) is the sequence change rate at time t; α and β are adjustment parameters; sequence change rate V(t) = ||S(t) - S(t-1)|| / Δt; S(t) is the sequence value at time t; local similarity function Sim(i,j) = exp(-||F(i) - F(j)||2 / σ2); where: F(i) is the feature vector of point i; σ is the scale parameter; cumulative cost matrix M(i,j) = Sim(i,j)•D(i,j), when |ij| ≤ When W(t), M(i,j) is calculated.

[0044] Step S23, performing wavelet decomposition on the aligned time series, reading pre-stored wavelet decomposition parameters, extracting time-frequency features at different scales; matching each scale feature with a pre-stored feature template library to generate a multi-scale feature matrix;

[0045] Step S24, reading the multi-scale feature matrix, combining the pre-stored feature weight parameters, calculating the information entropy and importance index of each feature, and generating a feature weight vector;

[0046] Step S25: input the multi-scale feature matrix and the feature weight vector into the feature fusion algorithm for weighted combination to generate time series alignment data.

[0047] Through time series segmentation and multi-scale time series alignment processing, the problem of asynchrony of multi-source sensor data in the time dimension is solved. Specifically, the data stream is first segmented by a sliding time window, and the regular expression of the time series is achieved through data integrity check and supplementation; then the time distortion degree between different sensor data is calculated based on the fast dynamic time warping algorithm, and the precise alignment of sensor data with different sampling rates and delays is achieved by constructing a distortion path matrix and solving the optimal alignment path; then the aligned time series data is subjected to wavelet decomposition, and the time-frequency features are extracted at multiple scales and matched with the feature template library to obtain a multi-scale feature expression; further, the feature weights are calculated based on information entropy and importance indicators, and a unified time series expression is obtained through weighted fusion. This multi-scale time series processing method not only solves the problem of time series inconsistency of sensor data, but also captures the feature changes at different time scales through wavelet decomposition, so that the subsequent state evaluation can pay attention to both fast-changing and slowly evolving feature patterns.

[0048] like Figure 4As shown, according to one aspect of the present application, step S3 is specifically:

[0049] Step S31, mapping the timing alignment data and the defect feature vector to a high-dimensional feature space, reading pre-stored feature mapping parameters, extracting state feature parameters, and generating a state feature matrix;

[0050] Step S32, read the state feature matrix and the pre-stored Bayesian network model parameters to construct a dynamic Bayesian network; calculate the conditional probability table between nodes and generate a probability dependency graph;

[0051] Step S33: input the state feature matrix into the probability dependency graph, combine it with the pre-stored difficulty assessment parameters, calculate the difficulty coefficient of the construction area, and generate a construction difficulty graph;

[0052] Step S34: Based on the dynamic Bayesian network constructed based on the probability dependency graph, the state feature matrix is ​​input, state reasoning is performed, and an environment state vector is generated.

[0053] By constructing a dynamic Bayesian network model, the probabilistic reasoning of the concrete surface state and the quantitative evaluation of the construction difficulty are realized. Specifically, the time-series alignment data and defect feature vectors are first mapped to a high-dimensional feature space, and the state feature parameters are extracted to form a complete state representation; then the probability dependency relationship is constructed based on the Bayesian network model parameters, and the causal relationship between each state variable is accurately described by calculating the conditional probability table between nodes; then the difficulty coefficient of the construction area is calculated based on the probability dependency graph and the difficulty evaluation parameters, and a construction difficulty graph reflecting the spatial distribution characteristics is formed; finally, the state reasoning is performed through the dynamic Bayesian network to achieve an accurate evaluation of the environmental state. This state evaluation method based on the probabilistic graphical model not only takes into account the complex dependency relationship between state variables, but also incorporates the time-series evolution characteristics. It can effectively handle the uncertainty of the perception data and provide a reliable basis for state judgment for subsequent construction planning.

[0054] like Figure 5 As shown, according to one aspect of the present application, step S4 is specifically:

[0055] Step S41, reading the construction difficulty map and the environmental state vector, combining the pre-stored area segmentation parameters, and using the spectral clustering algorithm to segment the construction area; merging and evaluating adjacent areas to generate a task area set;

[0056] Among them, the spectral clustering algorithm is: similarity matrix W(i,j) = exp(-||fi - fj||2 / σ2)•exp(-||xi- xj||2 / d2); where: fi is the eigenvector; xi is the spatial coordinate; σ, d are scale parameters; Laplace matrix L = D-1 / 2(DW)D-1 / 2; where: D is the degree matrix, D(i,i) = ∑W(i,j); eigenvalue decomposition L = UΛU^T; cluster number adaptive function k(λ) = argmax{(λi - λi+1) / (λi+1 - λi+2)}; where: λi is the i-th eigenvalue; region merging criterion M(Ri,Rj) = ω1•S(i,j) + ω2•H(i,j) + ω3•B(i,j); where: S(i,j) is the region similarity; H(i,j) is the homogeneity measure; B(i,j) is the boundary complexity.

[0057] Step S42: read the task area set and the construction difficulty map, calculate the spatiotemporal position coefficient and the construction difficulty coefficient of each area in combination with the pre-stored priority evaluation parameters; integrate multiple evaluation indicators to generate a task priority vector;

[0058] Step S43, reading the environment state vector and the pre-stored robot performance parameter library, calculating the operating efficiency, accuracy and stability index of each robot in the current environment; comprehensively evaluating various performance parameters to generate a capability evaluation matrix;

[0059] Step S44: input the task area set, the task priority vector and the capability evaluation matrix into the task allocation algorithm, combine the pre-stored task allocation constraints, solve the optimal allocation plan, and generate a task allocation plan;

[0060] Step S45, read the task allocation plan and the environment state vector, combine with the pre-stored collaborative rule library, formulate the communication plan and obstacle avoidance rules between robots, and generate a collaborative strategy.

[0061] Through intelligent segmentation of construction areas and optimal task allocation, efficient collaborative work in complex construction scenarios is achieved. Specifically, based on the construction difficulty map and the environmental state vector, the spectral clustering algorithm is used to adaptively segment the construction area. The reasonable division of the task area is ensured by analyzing the regional boundary characteristics and merging evaluation. Then, the task priority is calculated by combining the spatiotemporal position coefficient and the construction difficulty coefficient, and a priority evaluation system reflecting the urgency of the task is constructed. At the same time, the robot's operating ability is evaluated based on the environmental state vector and the data reliability matrix. By analyzing the operating efficiency, accuracy and stability indicators, a complete ability evaluation system is established. Then, the task requirements are matched and optimized with the robot's capabilities, and the optimal allocation scheme is solved by mixed integer programming to achieve balanced distribution of tasks. Finally, the communication scheme and obstacle avoidance rules between robots are formulated to form a complete collaborative strategy. This task allocation method takes into account both the difficulty differences in the construction area and the ability characteristics of the robot. By optimizing the allocation and formulating the collaborative strategy, the construction efficiency of the multi-robot system is significantly improved.

[0062] like Figure 6 As shown, according to one aspect of the present application, step S5 is specifically:

[0063] Step S51, read the environment state vector and the task allocation plan, combine with the pre-stored parameter constraint library, establish the value range and constraint relationship of the construction parameter; integrate multiple constraint conditions to generate a parameter constraint set;

[0064] Step S52: read the environmental state vector and the defect feature vector, combine with the pre-stored tool performance parameter library, evaluate the construction effect and adaptability index of various tools, and generate a tool evaluation matrix;

[0065] Step S53: input the parameter constraint set and the tool evaluation matrix into the multi-objective optimization algorithm, combine the pre-stored optimization target weights, solve the optimal parameter combination, and generate an optimization parameter set;

[0066] Step S54: read the optimization parameter set and the environmental state vector, combine with the pre-stored tool configuration rule library, plan the tool use strategy for different construction stages, and generate a tool configuration plan.

[0067] Precise control of concrete surface treatment is achieved through construction parameter optimization and tool configuration planning. Specifically, the parameter constraint relationship is first established based on the environmental state vector and the task allocation scheme to form a complete constraint condition system; then, the tool performance evaluation system is established by analyzing the construction effect of the tool under different environments and defect characteristics; then, the optimal parameter combination is solved based on the multi-objective optimization algorithm, and the construction effect is optimized while satisfying the constraints; finally, the tool use strategy and switching sequence are formulated according to the optimized parameters to form a complete tool configuration plan. This parameter optimization method takes environmental status, defect characteristics and tool performance into consideration, and realizes precise control of construction parameters through multi-objective optimization, providing reliable parameter support for achieving high-quality surface treatment.

[0068] like Figure 7 As shown, according to one aspect of the present application, step S6 is specifically:

[0069] Step S61, reading the construction difficulty map and the task allocation plan, combining the pre-stored environmental modeling parameters, converting the spatial information into a three-dimensional grid representation; performing terrain feature extraction and feasibility analysis on the grid data, and generating an environmental model;

[0070] Step S62: read the environment model and the coordination strategy, and calculate the comprehensive cost including the difficulty of passage, the difficulty of construction and the coordination constraint in combination with the pre-stored cost function parameter library; perform cost evaluation on each spatial position and generate a path cost map;

[0071] Step S63, reading the path cost map and the task allocation plan, combining the pre-stored path planning parameters, using the improved A* algorithm to plan the initial path for each robot; verifying the feasibility of the path, and generating an initial path set;

[0072] Step S64, reading the initial path set and the coordination strategy, and calculating the space-time intersection of the robot path in combination with the pre-stored collision detection parameters; analyzing the potential collision risk and generating a collision constraint set;

[0073] Step S65: input the initial path set, the collision constraint set and the optimization parameter set into the trajectory optimization algorithm, and perform collision avoidance optimization on the path in combination with the pre-stored trajectory optimization parameters to generate a collaborative construction trajectory.

[0074] Through multi-robot collaborative trajectory planning, safe and efficient construction in complex environments is achieved. Specifically, three-dimensional space modeling is first performed based on the construction difficulty map and environmental model, and the feasibility of trajectory planning is ensured through terrain feature extraction and tool accessibility analysis; then a path cost map that comprehensively considers the difficulty of passage, construction difficulty and collaborative constraints is constructed, providing an evaluation basis for trajectory optimization; then the improved A* algorithm is used for initial path planning, and collision detection and avoidance constraint generation are used to ensure the safe operation of the multi-robot system; finally, comprehensive optimization is performed based on the trajectory optimization algorithm, and the tool switching timing and environmental adaptation strategy are integrated into the trajectory planning, achieving the continuity and stability of the construction process. This collaborative trajectory planning method not only ensures the optimality of the single robot trajectory, but also realizes the collaborative operation of the multi-robot system, significantly improving the safety and efficiency of the construction process.

[0075] According to one aspect of the present application, step S11 is specifically:

[0076] Step S111, read the original sensor data, combine with the pre-stored data format parameter library, unify the format and convert the dimensions of the data of different sensors; extract the timestamp and sensor identifier, and generate unified format data and data attribute table;

[0077] Step S112, reading the unified format data, combining with the pre-stored normalization parameter library, analyzing the value distribution and range of various sensor data; calculating the normalization coefficient, and generating a normalization parameter set;

[0078] Step S113: read the unified format data and the normalization parameter set, perform minimum-maximum normalization processing on various sensor data; perform numerical mapping and scale conversion, and generate a normalized data set and a normalized mapping relationship;

[0079] Step S114, read the normalized data set, combine with the pre-stored sensor calibration parameter library, extract the calibration curve and zero point parameters of each sensor; analyze the sensor characteristic deviation, and generate a calibration coefficient matrix;

[0080] Step S115, read the normalized data set and the calibration coefficient matrix, perform zero point correction and sensitivity calibration; compensate for nonlinear errors, and generate a calibrated standardized data set.

[0081] Through the standardized processing of multi-source sensor data, the unified expression of heterogeneous data is achieved. Specifically, the format of the original sensor data is unified and the dimension is converted. The correlation between the data is established by extracting the timestamp and sensor identification. Then, the numerical distribution characteristics of various sensor data are analyzed, and the minimum-maximum normalization method is used to map the data to a unified interval. Then, zero point correction and sensitivity calibration are performed based on the sensor calibration parameters, effectively compensating for the nonlinear error of the sensor. This data standardization processing method establishes a complete conversion chain from raw data to standardized expression, laying the foundation for subsequent feature extraction and data fusion.

[0082] According to one aspect of the present application, step S13 is specifically:

[0083] S131, read the surface feature data, and perform temporal segmentation and spatial block division on the feature data in combination with the pre-stored data segmentation parameter library; calculate the statistical features of each data block, and generate a block feature set and a data block index table.

[0084] S132, reading the block feature set, combining with the pre-stored seed selection parameter library, calculating the local variance and mutation degree; determining the potential defect area based on the significance analysis, and generating the seed feature point set and the area marking table.

[0085] S133, read the block feature set, seed feature point set and region marking table, and build a feature space similarity matrix in combination with the pre-stored growth control parameter library; calculate the neighborhood weight coefficient and dynamic threshold, perform adaptive region growing operation, perform feature clustering based on local statistical characteristics and directional consistency, and generate region growing results and growth path records.

[0086] The adaptive region growing operation is implemented by the adaptive region growing algorithm, and the similarity measurement function D(p,q) = ω1•||I(p) - I(q)||2 + ω2•||G(p) - G(q)||2 + ω3•||T(p) - T(q)||2; where: I(p) is the intensity value of pixel p; G(p) is the gradient value vector of pixel p; T(p) is the texture feature vector of pixel p; ω1, ω2, ω3 are feature weight coefficients, and ω1 + ω2 + ω3 = 1; dynamic threshold function T(R) = μR + k•σR; where: μR is the average similarity of the current region R; σR is the standard deviation of the similarity of the current region R; k is the adaptive coefficient, k = α•exp(-β•SR); SR is the area of ​​region R; α and β are control parameters; growth direction weight W(d) = γ•cos(θd) +(1-γ)•exp(-λ•Ld); where: θd is the angle between direction d and the main direction of the region; Ld is the boundary curvature in direction d; γ and λ are weight parameters; comprehensive scoring function Score(p) = D(p,q)•W(d), when Score(p) < T(R), pixel point p is added to region R.

[0087] S134, reading the regional growth results and growth path records, combining with the pre-stored regional optimization parameter library, analyzing the feature correlation of adjacent regions; performing regional merging and splitting optimization, and generating an optimized region set and a region description table.

[0088] S135, reading the optimized region set and the region description table, and calculating the regional morphological features and statistical features in combination with the pre-stored feature extraction parameter library; constructing a feature vector space, and generating a feature vector group and a feature importance table.

[0089] S136, read the feature vector group and the feature importance table, and perform feature screening and dimensionality reduction processing in combination with the pre-stored feature dimension reduction parameter library; calculate the main feature components, and generate the defect feature vector and feature evaluation report.

[0090] The adaptive region growing algorithm is specifically:

[0091] An 8-neighborhood dynamic weight model is established to assign different weights to neighborhood pixels according to the gradient direction;

[0092] Calculate the similarity feature vector (grayscale difference, texture feature, direction consistency) between the candidate pixel and the current area;

[0093] Design an adaptive growth criterion: dynamically adjust the threshold according to the statistical characteristics of pixels in the region;

[0094] Use hierarchical queues to store candidate pixels and prioritize pixels with high similarity;

[0095] Apply morphological constraints to control the growth direction of the region shape.

[0096] Accurate identification of concrete surface defects is achieved through adaptive region growing algorithm. Specifically, the feature data is first segmented in time and space, and the basic unit of data processing is determined through statistical feature analysis; then the potential defect area is determined based on local variance and mutation degree, and an effective seed point selection mechanism is established; then the adaptive growth of the region is achieved by constructing a feature space similarity matrix and a dynamic threshold mechanism; finally, the growth results are optimized and features are extracted to form a complete defect feature expression. This defect recognition method not only takes into account the similarity of local features, but also incorporates global constraints, and improves the robustness of recognition through an adaptive threshold mechanism.

[0097] According to one aspect of the present application, S22 is specifically:

[0098] S221, read the complete time series set, combine with the pre-stored sampling characteristic parameter library, analyze the sampling characteristics and timing structure of each sensor; calculate the sampling consistency index, and generate the sampling characteristic matrix and sequence attribute table.

[0099] S222, read the complete time series set and the sampling characteristic matrix, and calculate the local similarity between the sequences in combination with the pre-stored time series alignment parameters; construct a similarity accumulation matrix, and generate a similarity distance matrix and a local constraint table.

[0100] S223, read the similarity distance matrix and the local constraint table, combine with the pre-stored regularization algorithm parameter library, and use the constrained cumulative distance calculation model to initialize the regularization matrix; construct a local search window with a Sakoe-Chiba band constraint bandwidth, and calculate the forward cumulative distance value based on the monotonicity and continuity constraints; calculate the regularization cost at different time scales through a parallel blocking strategy, and construct a multi-layer warp path candidate set based on the minimum regularization cost principle, and screen the effective warp path in combination with the speed constraint and the slope constraint; generate the warp path matrix and the regularization cost table.

[0101] S224. Read the twisted path matrix and the regularized cost table, combine with the pre-stored path optimization parameter library, and use a hierarchical iterative backtracking search strategy to calculate the local optimal path; use the improved BeamSearch algorithm to maintain a fixed-width candidate path set, and dynamically score the candidate paths based on a multi-objective evaluation function of path smoothness, path length, and cumulative cost; eliminate path jumps and intersections by combining local path correction and global path optimization, and determine the global optimal alignment path using a dynamic programming method based on the Bellman principle; generate an alignment path set and a path evaluation table.

[0102] S225, read the alignment path set and the complete time series set, combine with the pre-stored interpolation parameter library, perform time series resampling and alignment transformation; verify the alignment accuracy, and generate an alignment time series and alignment quality report.

[0103] The accurate alignment of multi-source sensor data is achieved through an improved dynamic time warping algorithm. Specifically, the sampling characteristics and timing structure of the sensor are first analyzed, and a complete sampling characteristic description system is established; then, the local similarity between sequences is calculated and the similarity accumulation matrix is ​​constructed to form the evaluation basis of sequence alignment; then, the DTW algorithm with Sakoe-Chiba band constraint is used for path search, and the adaptability of the algorithm is improved through the adaptive bandwidth adjustment mechanism; at the same time, the improved BeamSearch algorithm is introduced for path optimization, and the search width is dynamically adjusted based on complexity evaluation, achieving a balance between computational efficiency and optimization effect; finally, a unified time expression is obtained through time resampling and alignment transformation. This time alignment method not only ensures the accuracy of alignment, but also significantly improves the computational efficiency through the improved algorithm, providing a consistent time data basis for subsequent feature extraction.

[0104] In a further embodiment, step S22 may also be:

[0105] Step S221, receiving a complete time series set, reading the sampling rate and time distribution characteristics from the pre-stored sampling characteristic parameter library, calculating the sequence sampling consistency index and time domain characteristic distribution, and generating a sampling characteristic matrix and a sequence attribute table.

[0106] Step S222: extract similarity calculation parameters from pre-stored timing alignment parameters according to the sampling characteristic matrix and the complete time series set, calculate the Euclidean distance of the sequence segments and the local similarity value in the dynamic time window, and generate a similarity distance matrix and a local constraint table.

[0107] Step S223, based on the similarity distance matrix and the local constraint table, obtain the initial bandwidth coefficient and the adjustment step size from the pre-stored regularization algorithm parameter library, calculate the time domain fluctuation characteristics and frequency domain energy distribution of the sequence, and generate the time-frequency feature vector; use the time-frequency feature vector to calculate the local change rate and global trend characteristics of the sequence, and generate the sequence change characteristics; according to the sequence change characteristics, dynamically adjust the bandwidth parameters of the Sakoe-Chiba band to generate an adaptive bandwidth sequence; construct a local search window based on the adaptive bandwidth sequence, calculate the cumulative distance matrix, and generate the distorted path matrix and the regularization cost table.

[0108] Step S224, receiving the twisted path matrix and the regularized cost table, obtaining the beam search parameter from the pre-stored path optimization parameter library, calculating the complexity index of the search space according to the cumulative cost value of the path, and generating a complexity evaluation value; based on the complexity evaluation value, using an adaptive strategy to calculate the beam width parameter, and generating a search width sequence; using the search width sequence, executing the improved BeamSearch algorithm, dynamically maintaining a fixed-width candidate path set at each time step, and generating a candidate path set; performing multi-objective scoring on the paths in the candidate path set, calculating the smoothness, length and cumulative cost of the path, and generating a path scoring matrix; based on the path scoring matrix, using a dynamic programming method to determine the global optimal alignment path, and generating an alignment path set and a path evaluation table.

[0109] The improved BeamSearch algorithm is specifically as follows: path scoring function Score(P) = ω1•L(P) + ω2•S(P)+ ω3•C(P); where: L(P) is the length term of path P, L(P) = ∑||pi+1 - pi||; S(P) is the path smoothness term, S(P) = ∑(1 - cos(θi)); θi is the path turning angle; C(P) is the cumulative cost term, C(P) = ∑M(i,j); search width function B(t) = Bmin + (Bmax - Bmin)•exp(-λ•E(t)); where: E(t) is the search space entropy value at time t; Bmin and Bmax are the search width ranges; λ is the attenuation coefficient; the candidate path set update probability P(k) = exp(-Score(k) / T) / ∑exp(-Score(i) / T); where: T is the temperature parameter.

[0110] Step S225: read interpolation algorithm parameters from a pre-stored interpolation parameter library according to the alignment path set and the complete time series set, perform sequence resampling and time axis mapping transformation, and generate an alignment time series and alignment quality report.

[0111] By reading the sampling rate and time distribution characteristics from the pre-stored sampling characteristic parameter library, the system can quickly identify sampling anomalies and time domain distribution unevenness in the sequence. When calculating the sampling consistency index, an adaptive time window analysis method is used to accurately capture the changing characteristics of the local sampling density. The generated sampling characteristic matrix can quantitatively describe the sampling characteristics of the sequence at different time scales, providing an important reference for subsequent sequence alignment. Compared with the traditional single sampling rate detection, this sampling analysis method based on multidimensional features can more comprehensively characterize the temporal structure characteristics of the sequence and improve the accuracy and robustness of the subsequent alignment process. The innovation of this step is to parameterize the sampling characteristics and realize rapid feature extraction through the pre-stored parameter library, which greatly improves the processing efficiency while ensuring the accuracy and consistency of feature extraction.

[0112] Combining Euclidean distance calculation with dynamic time window analysis, the precise quantification of sequence fragment similarity is achieved. Through the pre-stored time alignment parameters, the system can adaptively adjust the parameters of similarity calculation so that the calculation results are more consistent with the characteristics of the actual sequence. When calculating the Euclidean distance of sequence fragments, the sliding window technology is used to effectively capture local similarity changes. Through the design of dynamic time windows, the system can adaptively adjust the size of the comparison interval, improving efficiency while ensuring calculation accuracy. The generated similarity distance matrix not only contains global similarity information, but also records fine-grained similarity change characteristics through local constraint tables. This multi-scale similarity analysis method provides richer reference information for subsequent sequence alignment. The innovativeness of the technical effect is reflected in: the rapid adaptive adjustment of similarity calculation is achieved by pre-storing parameters, and the accurate characterization of local features of the sequence is achieved through the design of local constraint tables, which improves the accuracy of the subsequent alignment process.

[0113] By combining time-frequency domain analysis with adaptive bandwidth adjustment, a comprehensive characterization and precise alignment of sequence features are achieved. First, through the calculation of time-frequency feature vectors, the system can simultaneously capture the time-domain fluctuation characteristics and frequency-domain energy distribution of the sequence. This dual-domain analysis method provides a more comprehensive description of sequence features. The dynamic bandwidth adjustment mechanism based on sequence change characteristics enables the Sakoe-Chiba band to adaptively change with sequence characteristics. This adaptive mechanism significantly improves the accuracy and robustness of alignment. When constructing the local search window, the system makes full use of the information of the adaptive bandwidth sequence to ensure that the search space can cover the potential optimal alignment path without causing excessive waste of computing resources. Through the calculation of the cumulative distance matrix, the system can effectively evaluate the cost of different alignment paths, providing a reliable basis for subsequent path optimization. This adaptive bandwidth adjustment method based on multi-dimensional features not only improves the accuracy of alignment, but also significantly reduces the computational complexity. The technical innovation is mainly reflected in: the rapid adaptive adjustment of bandwidth parameters is achieved through the pre-stored regularization algorithm parameter library, and the precise characterization of sequence features is achieved through time-frequency feature analysis.

[0114] By improving the BeamSearch algorithm, efficient search and optimization of alignment paths are achieved. Specifically, by introducing an adaptive beam width parameter calculation mechanism, the system can dynamically adjust the search width according to the complexity of the search space. This adaptive mechanism significantly improves the search efficiency. In terms of path scoring, a multi-objective scoring strategy is adopted to comprehensively consider the smoothness, length and cumulative cost of the path, so that the final selected alignment path has better global characteristics. The improved BeamSearch algorithm realizes the probabilistic update of the candidate path set by introducing the temperature parameter, effectively avoiding the local optimal solution. The global optimal alignment path is determined by the dynamic programming method, which not only ensures the optimality of the result, but also improves the convergence speed of the algorithm. The technical innovation of this step is mainly reflected in: through adaptive search width and multi-objective scoring mechanism, efficient optimization of alignment path search is achieved; through probabilistic update strategy, the robustness and convergence of the algorithm are improved; through the pre-stored path optimization parameter library, rapid adaptive adjustment of parameters is achieved.

[0115] High-quality sequence alignment is achieved through sequence resampling and timeline mapping transformation methods. Through the pre-stored interpolation parameter library, the system can quickly select interpolation algorithm parameters suitable for the current sequence characteristics to ensure the accuracy of the resampling process. When performing the timeline mapping transformation, the system makes full use of the information of the alignment path set to achieve accurate correspondence mapping of sequence time points. By generating an alignment quality report, the system can quantitatively evaluate the quality of the alignment results and provide reliable quality assurance for subsequent applications. The technical innovation of this step is mainly reflected in: the rapid adaptive adjustment of the interpolation algorithm is achieved through the pre-stored interpolation parameter library; the high quality of the alignment results is ensured through the precise timeline mapping transformation; and the reliability evaluation of the alignment results is achieved through the generation of the alignment quality report. This alignment method based on pre-stored parameters and quality assessment not only improves the accuracy of the alignment, but also ensures the reliability of the results.

[0116] According to one aspect of the present application, step S32 is specifically:

[0117] S321, read the state feature matrix, combine the pre-stored data discretization parameter library, perform equal-frequency binning and conditional segmentation on the continuous state features; calculate the optimal segmentation point, and generate a discrete state set and binning mapping table.

[0118] S322, read the discrete state set and the bin mapping table, combine with the pre-stored variable association parameter library, use mutual information and chi-square test to calculate the association strength between variable pairs; build an initial network skeleton based on the maximum weight spanning tree algorithm, and generate a network structure diagram and an association strength matrix.

[0119] S323, read the network structure diagram and the association strength matrix, combine with the pre-stored Bayesian network model parameters, and apply the local scoring search algorithm to optimize the network structure; determine the optimal network structure by adding, deleting, and reversing edges and the BIC scoring criterion, and verify the conditional independence based on the directed separation criterion; generate an optimized network diagram and an independence test table.

[0120] S324, read the optimized network diagram and discrete state set, combine with the pre-stored parameter learning parameter library, use the expectation maximization algorithm to estimate the conditional probability distribution; process missing data and apply parameter smoothing technology to calculate the conditional probability table and marginal probability distribution of each node; generate a conditional probability table set and a node distribution table.

[0121] S325, read the conditional probability table set and the node distribution table, and build the dynamic transfer probability across time slices in combination with the pre-stored timing dependency parameter library; calculate the timing dependency strength between nodes, and establish the time expansion structure of the dynamic Bayesian network; generate a probability dependency graph and a network evaluation report.

[0122] The dynamic Bayesian network is: node association strength I(X,Y) = ∑P(x,y)•log(P(x,y) / (P(x)P(y))); where: P(x,y) is the joint probability; P(x) and P(y) are marginal probabilities; network structure score BIC(G) = L(D|G) -0.5•log(N)•|G|; where: L(D|G) is the data likelihood; N is the number of samples; |G| is the number of network parameters; time series transition probability P(Xt+1|Xt) = softmax(W•Xt + b); where: W is the weight matrix; b is the bias vector; conditional probability table update ΔP = η•(Pt - Pt-1)•(1 - |Et| / Emax); where: Et is the prediction error; η is the learning rate.

[0123] Accurate reasoning of the concrete surface state is achieved through the construction of a dynamic Bayesian network. Specifically, the state characteristics are first discretized, and the reasonable division of the feature space is achieved by determining the optimal segmentation point; then the correlation strength between variables is analyzed based on mutual information and chi-square test, and the initial network structure is constructed using the maximum weight spanning tree algorithm; then the network structure is optimized through a local scoring search algorithm, and the optimization effect of the network structure is ensured by combining the BIC scoring criterion; finally, the time expansion structure of the dynamic Bayesian network is established to achieve a dynamic description of state evolution. This state reasoning method based on a probabilistic graphical model not only captures the static dependency relationship between state variables, but also describes the dynamic evolution characteristics of the state, significantly improving the accuracy of state assessment.

[0124] According to one aspect of the present application, step S41 is specifically:

[0125] S411, read the construction difficulty map and the environmental state vector, and build the spatial adjacency relationship between regions in combination with the pre-stored similarity calculation parameter library; calculate the local feature similarity, and generate a similarity matrix and an adjacency relationship table.

[0126] S412, read the similarity matrix and the adjacency table, and construct the Laplace matrix in combination with the pre-stored spectral decomposition parameter library; calculate the eigenvalues ​​and eigenvectors, extract the main eigencomponents, and generate the eigenspace matrix and spectral decomposition report.

[0127] S413, read the feature space matrix and spectral decomposition report, and perform K-means clustering operation in combination with the pre-stored clustering parameter library; optimize the number of clusters based on the silhouette coefficient, and generate an initial segmentation result and a clustering evaluation table.

[0128] S414, reading the initial segmentation result and the adjacency relationship table, combining the pre-stored region segmentation parameters, analyzing the region boundary characteristics; calculating the boundary smoothness and region connectivity, and generating a region segmentation map and a boundary feature table.

[0129] S415, read the region segmentation map and boundary feature table, and calculate the merging cost of adjacent regions in combination with the pre-stored region merging parameter library; evaluate the feasibility of merging based on the uniformity of construction difficulty and consistency of environmental status, and generate a task area set and a region attribute report.

[0130] Adaptive segmentation of construction areas is achieved through spectral clustering algorithm. Specifically, the spatial adjacency relationship between regions is first constructed, and the metric standard of regional association is established by calculating the local feature similarity; then the main feature components are extracted through feature space construction and spectral decomposition, providing a dimensionality reduction expression for clustering analysis; then regional clustering is performed based on the K-means algorithm, and the optimal number of clusters is determined by optimizing the silhouette coefficient; finally, the boundary of the segmentation results is optimized and the regions are merged to achieve a reasonable division of the task area. This regional segmentation method not only takes into account the spatial continuity constraint, but also incorporates the feature similarity measurement, and achieves a reasonable division of the region through adaptive clustering, providing basic support for subsequent task allocation.

[0131] According to one aspect of the present application, step S44 is specifically:

[0132] S441, read the task area set and task priority vector, combine with the pre-stored task feature parameter library, analyze the task time and space constraints and resource requirements; build a task dependency graph, generate a task constraint matrix and resource requirement table.

[0133] S442, read the capability evaluation matrix and resource requirement table, and calculate the degree of adaptation between the execution unit and the task in combination with the pre-stored matching calculation parameter library; evaluate the task completion time and resource consumption, and generate the matching matrix and execution cost table.

[0134] S443, read the task constraint matrix and the matching degree matrix, and build a mixed integer programming model in combination with the pre-stored task allocation constraint conditions; set the objective function and constraint conditions, and generate an optimization model and a constraint condition set.

[0135] Among them, the mixed integer programming is specifically as follows: objective function min F = ∑∑cijxij + α•∑|Li - Lj|•yij + β•∑ti; where: cij is the cost assigned to robot j for task i; xij is a 0-1 decision variable; Li is the load of robot i; yij is the collaborative mark of robots i and j; ti is the task completion time; constraint condition ∑xij = 1; ∑tijxij≤ Tmax; load balance degree B = max(Li) / avg(Li); resource matching degree M(i,j) = exp(-||Ri - Tj||2 / σ2); where: Ri is the resource capability vector; Tj is the task requirement vector; Lagrangian relaxation function L(x,λ) = F + ∑λk•gk(x); where: gk(x) is the constraint function; λk is the Lagrangian multiplier.

[0136] S444, read the optimization model and constraint condition set, combine with the pre-stored solver parameter library, use the branch and bound algorithm to solve the optimal allocation plan; optimize the search efficiency through Lagrangian relaxation, and generate an initial allocation plan and optimization process report.

[0137] S445, read the initial allocation plan and execution cost table, combine with the pre-stored plan evaluation parameter library, perform plan feasibility verification; calculate the load balance and task completion efficiency, and generate a task allocation plan and plan evaluation report.

[0138] The optimal allocation of tasks is achieved through mixed integer programming. Specifically, the time and space constraints and resource requirements of the tasks are first analyzed, and a complete task dependency graph is constructed; then the degree of adaptation between the execution unit and the task is calculated, and an evaluation system for task completion efficiency is established; then a mixed integer programming model is constructed, and a complete optimization problem is formed by setting the objective function and constraints; finally, the branch and bound algorithm is used to solve the optimal allocation solution, and the solution efficiency is improved through Lagrangian relaxation. This task allocation method not only meets multiple constraints, but also achieves the global optimal resource allocation, significantly improving the operating efficiency of the multi-robot system.

[0139] According to one aspect of the present application, step S53 is specifically:

[0140] S531. Read the parameter constraint set and tool evaluation matrix, analyze the coupling relationship between parameters in combination with the pre-stored parameter sensitivity parameter library; calculate parameter sensitivity and interaction effects, and generate a parameter association matrix and sensitivity report.

[0141] S532, read the parameter association matrix and sensitivity report, and build a multi-objective optimization model in combination with the pre-stored objective function parameter library; set the objective function expression and constraint boundaries, and generate the optimization objective set and constraint boundary table.

[0142] S533, read the optimization target set and constraint boundary table, and calculate the weighted objective function in combination with the pre-stored optimization target weights; construct the Pareto optimal frontier, and generate the objective function matrix and weight influence table.

[0143] S534, read the objective function matrix and weight influence table, combine with the pre-stored evolutionary algorithm parameter library, perform multi-objective genetic algorithm optimization; use non-dominated sorting and crowding calculation to perform population evolution, generate parameter candidate sets and evolution process reports.

[0144] Among them, the multi-objective genetic algorithm is: fitness function F(x) = ω1•Q(x) + ω2•E(x) + ω3•T(x); where: Q(x) is the quality target; E(x) is the energy consumption target; T(x) is the time target; parameter encoding structure C = {p1,p2,...,pn,t1,t2,...,tm}; where: pi is the process parameter; ti is the tool switching time; crossover operator Pc(t) =Pc1 - (Pc1 - Pc2)•(1 - r / rmax)2; where: r is the current generation; mutation operator Pm(t) = Pm1 + (Pm2 -Pm1)•exp(-λ•f / fmax); where: f is the individual fitness value; non-dominated sorting level R(i) = 1 + |{j|∀k,fk(j) ≤ fk(i)}|; congestion calculation D(i) = ∑(fk(i+1) - fk(i-1)) / fkmax.

[0145] S535, read the parameter candidate set and tool evaluation matrix, and perform parameter combination verification in combination with the pre-stored verification parameter library; evaluate the feasibility and robustness of the scheme, and generate an optimized parameter set and an optimization evaluation report.

[0146] The optimal configuration of construction parameters was achieved through a multi-objective optimization algorithm. Specifically, the coupling relationship and sensitivity characteristics between parameters were first analyzed, and a quantitative evaluation system for parameter influence was established; then a multi-objective optimization model was constructed, and a unified expression of multiple optimization objectives was achieved through a weighted objective function; then a multi-objective genetic algorithm was used for parameter optimization, and the diversity of solutions was ensured through non-dominated sorting and congestion calculation; finally, the feasibility of the optimization results was verified to ensure the practicality of the parameter configuration. This parameter optimization method not only takes into account multiple optimization objectives, but also ensures the feasibility of the optimization results, providing the optimal parameter configuration for concrete surface treatment.

[0147] According to one aspect of the present application, step S63 is specifically:

[0148] S631, read the path cost map and task allocation plan, combine with the pre-stored map processing parameter library, build a rasterized environment model; calculate the obstacle expansion area and safe channel, and generate a raster map and passability matrix.

[0149] S632, read the grid map and the passability matrix, and design a dynamic heuristic function in combination with the pre-stored heuristic function parameter library; consider terrain changes and robot motion characteristics, and generate a heuristic function set and a cost evaluation table.

[0150] S633, read the heuristic function set and the cost evaluation table, and execute the improved A* algorithm search in combination with the pre-stored path planning parameters; adopt the hierarchical search strategy and dynamic weight adjustment to generate the search path set and node expansion table.

[0151] The improved A* algorithm is specifically as follows: dynamic heuristic function h(n) = (1 + ε•C(n))•(ω1•d(n) + ω2•r(n)); where: C(n) is the terrain complexity of node n; d(n) is the target distance; r(n) is the posture cost; ε is the adaptive coefficient; node expansion priority P(n) = g(n) + h(n) + μ•v(n); where: g(n) is the actual cost; v(n) is the speed constraint term; path smoothing function S(p) = ∑wi•||pi+1 - 2pi + pi-1||2; where: wi is the smoothing weight; motion constraint verification K(p) = max(|κ(s)|) ≤ κmax; where: κ(s) is the path curvature.

[0152] S634, reading the search path set and the node expansion table, and performing path smoothing optimization in combination with the pre-stored smoothing parameter library; applying cubic spline interpolation and local adjustment to generate a smooth path set and a curvature analysis table.

[0153] S635, read the smooth path set and passability matrix, and perform path feasibility test in combination with the pre-stored verification parameter library; evaluate the motion constraint satisfaction and collision risk, and generate an initial path set and feasibility report.

[0154] The robot's initial path planning is realized by the improved A algorithm. Specifically, the path cost map is firstly rasterized, and the safety of path planning is ensured by obstacle expansion and safe channel construction; then the dynamic heuristic function is designed based on the robot's motion characteristics and terrain changes to realize the intelligent guidance of path search; then the improved A algorithm is implemented by adopting a hierarchical search strategy and a dynamic weight adjustment mechanism, which significantly improves the efficiency of path search; finally, the path is smoothed by cubic spline interpolation and local adjustment, and motion constraint verification is performed to ensure the feasibility of the path. This path planning method not only ensures the optimality of the path, but also ensures the real-time requirements through improved heuristic search and multi-layer optimization strategies, and at the same time improves the algorithm's adaptability to complex environments through dynamic weight adjustment.

[0155] According to one aspect of the present application, step S64 is specifically:

[0156] S641, read the initial path set and the coordination strategy, combine with the pre-stored trajectory sampling parameter library, perform path time-space discretization; calculate the position sequence of key time points, and generate the time-space trajectory matrix and sampling point set.

[0157] S642, read the spatiotemporal trajectory matrix and the sampling point set, and build a spatial index structure in combination with the pre-stored spatial detection parameter library; apply the quadtree partitioning algorithm to quickly locate potential intersection areas, and generate an intersection area set and a spatial index table.

[0158] S643, read the intersection area set and the spatial index table, and calculate the minimum separation distance between robots in combination with the pre-stored collision detection parameters; consider the robot motion envelope and safety margin, and generate a distance matrix and a timing relationship table.

[0159] S644, read the distance matrix and the time series relationship table, and analyze the collision risk level in combination with the pre-stored risk assessment parameter library; evaluate the speed matching degree and avoidance feasibility, and generate a risk level table and an avoidance suggestion set.

[0160] S645, read the risk level table and the avoidance suggestion set, generate a parameter library in combination with the pre-stored constraints, and construct the spatiotemporal avoidance constraints; set the speed adjustment range and priority rules, and generate a collision constraint set and a constraint evaluation report.

[0161] The safe collaboration of multi-robot system is achieved through efficient collision detection. Specifically, the path is firstly processed in time and space, and the basic data structure of collision detection is established by sampling key time points; then the spatial index is constructed by quadtree partitioning algorithm, which realizes the rapid positioning of potential collision areas; then the minimum separation distance between robots is calculated, and a complete collision risk assessment system is established by considering the motion envelope and safety margin; finally, the avoidance strategy is formulated based on the risk level, forming a collision constraint set in time and space dimensions. This collision detection method significantly reduces the computational complexity through spatial indexing and hierarchical detection strategy, and ensures the operational safety of the multi-robot system through a complete risk assessment system.

[0162] According to one aspect of the present application, step S65 is specifically:

[0163] S651, read the initial path set and the collision constraint set, and build an optimization constraint model in combination with the pre-stored constraint conversion parameter library; convert the spatial constraints and timing constraints, and generate a constraint function set and a constraint weight table.

[0164] S652, read the constraint function set and the environmental state vector prediction, and build a dynamic environmental model in combination with the pre-stored prediction parameter library; evaluate the impact of environmental changes on the trajectory, and generate an environmental impact matrix and a change trend table.

[0165] S653, read the tool-trajectory configuration scheme and the optimization parameter set, plan the tool switching strategy in combination with the pre-stored tool parameter library; optimize the switching timing and transition trajectory, and generate the tool switching sequence and transition strategy table.

[0166] S654, read the constraint function set, environmental impact matrix and tool switching sequence, and build a multi-objective optimization model in combination with the pre-stored optimization parameter library; apply the improved particle swarm algorithm to solve and generate the optimization trajectory set and performance evaluation table.

[0167] S655, read the optimized trajectory set and transition strategy table, combine with the pre-stored adaptive parameter library, generate the environmental adaptation strategy; integrate the tool switching sequence, generate the collaborative construction trajectory and optimization report.

[0168] The improved particle swarm algorithm is specifically as follows: trajectory particle representation X(t) = {p1, p2, ..., pn, v1,v2, ..., vn, t1, t2, ..., tm}; where: pi is the path point position; vi is the path point speed; ti is the tool switching time; speed update equation v(t+1) = w(t)•v(t) + c1•r1•(pbest - x(t)) + c2•r2•(gbest - x(t)) + c3•r3•(E(t)•x(t)); where: w(t) is the adaptive inertia weight, w(t) = wmax - (wmax -wmin)•exp(-ρ•|ΔF(t)|); E(t) is the environment dynamic prediction matrix; trajectory evaluation function F(x) = α•S(x) + β•C(x) + γ•T(x) + η•W(x); where: S(x) is trajectory smoothness, S(x) = ∑||pi+1 - 2pi + pi-1||2 / ||pi+1 - pi-1||; C(x) is collision cost term, C(x) = ∑exp(-d(pi) / d0); T(x) is tool switching cost, T(x) = ∑|v(ti+)|•|v(ti-)|; W(x) is workspace constraint, W(x) = ∑max(0, g(pi))2; dynamic congestion calculation D(i) = ∑|Fi(x+Δx) - Fi(x-Δx)| / Δx; environmental adaptability factor R(x) = exp(-λ•||E(t+1)•x - E(t)•x||).

[0169] The collaborative construction control of multi-robot system is realized through trajectory comprehensive optimization. Specifically, an optimization model including spatial constraints and temporal constraints is first constructed, and a theoretical framework of trajectory optimization is established; then a dynamic environment model is constructed based on the prediction of the environment state vector to achieve early adaptation to environmental changes; then the tool switching strategy and transition trajectory are optimized to ensure the continuity of the construction process; finally, the trajectory is comprehensively optimized through the improved particle swarm algorithm to achieve the optimal trajectory planning that meets multiple constraints. This trajectory optimization method takes environmental prediction, tool switching and collision avoidance into account, and achieves the global optimality of the trajectory through multi-objective optimization, while ensuring the stability and safety of the construction process.

[0170] In another embodiment of the present application, step S43 may also be:

[0171] Read the environmental state vector, data reliability matrix and pre-stored robot performance parameter library, calculate the operating efficiency, accuracy and stability indicators of each robot in the current environment; evaluate various performance parameters in combination with the reliability of sensor data to generate a capability evaluation matrix.

[0172] Step S52 may also be: reading the environmental state vector, the defect feature vector and the multi-scale feature matrix, combining with the pre-stored tool performance parameter library, evaluating the construction effects and adaptability indicators of various tools under different scale characteristics, and generating a tool evaluation matrix.

[0173] Step S54 may also be: reading the optimization parameter set and the environmental state vector, combining with the pre-stored tool configuration rule library, planning the tool use strategy and switching sequence for different construction phases, and generating a tool-trajectory configuration plan.

[0174] Step S61 may also be: reading the construction difficulty map, task allocation plan and tool-trajectory configuration plan, combining the pre-stored environmental modeling parameters, converting the spatial information into a three-dimensional grid representation; performing terrain feature extraction and tool accessibility analysis on the grid data to generate an environmental model.

[0175] Step S65 can also be: inputting the initial path set, collision constraint set, optimization parameter set and tool-trajectory configuration plan into the trajectory optimization algorithm, combining the dynamic change prediction of the environmental state vector and the pre-stored trajectory optimization parameters, performing comprehensive trajectory optimization based on the dynamic prediction of the environment, including tool switching timing planning and collision constraint satisfaction, and generating a collaborative construction trajectory including tool switching timing and environmental adaptation strategy.

[0176] According to another aspect of the present application, a concrete surface intelligent construction system based on multimodal perception is also provided, comprising:

[0177] at least one processor; and,

[0178] a memory communicatively connected to at least one of the processors; wherein,

[0179] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the intelligent construction method of concrete surface based on multimodal perception described in any of the above technical solutions.

[0180] It should be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, the present invention will not further describe various possible combinations.

Claims

1. A method for intelligent construction of concrete surface based on multimodal perception, characterized in that: The steps include: Step S1, receiving original sensor data sets, including original pressure data of the pressure sensor array, original spectral data of the near-infrared spectral sensor, original point cloud data of the 3D laser scanner, original temperature data of the temperature sensor, original ultrasonic data of the ultrasonic detector, and original environmental parameter data of the environmental parameter sensor, performing standardization processing and feature extraction on various types of data, and generating standardized feature data and defect feature vectors; Step S2, reading standardized feature data, performing time series segmentation and time alignment processing on the sensor data, performing multi-scale analysis in combination with a pre-stored time series processing parameter library, and generating time series alignment data; Step S3, reading the time alignment data and defect feature vector, combining with the pre-stored state assessment parameter library, building a Bayesian network model, and generating an environmental state vector and a construction difficulty map; Step S4, read the construction difficulty map and the environment state vector, combine with the pre-stored task planning parameter library, divide the construction area into subtasks and optimize the allocation, generate the task allocation plan and coordination strategy; Step S5, reading the environment state vector, the task allocation scheme and the defect feature vector, optimizing the construction parameters and tool configuration in combination with the pre-stored process parameter library, and generating an optimized parameter set and tool configuration scheme; Step S6: read the construction difficulty map, task allocation plan, collaborative strategy and optimization parameter set, and generate a collaborative construction trajectory of multiple robots in combination with a pre-stored path planning parameter library.

2. The method for intelligent construction of concrete surface based on multimodal perception according to claim 1 is characterized in that: Step S1 is specifically as follows: Step S11, receiving the original sensor data set, normalizing various types of data to the interval [-1, 1], and generating a standardized sensor data set; reading the calibration coefficient from the pre-stored sensor calibration parameters, calibrating the standardized sensor data set, and generating a calibrated standardized data set; Step S12, reading the original point cloud data, original ultrasonic data and original spectral data in the calibrated standardized data set, and extracting local geometric features and material features; Calculate the curvature, normal vector, depth information and material parameters of each local area to construct a feature description vector; perform feature matching with the pre-stored feature template library to generate surface feature data; Step S13, reading the surface feature data, using the region growing algorithm to identify connected regions with similar features, and generating defect candidate region data; matching the defect candidate region data with the pre-stored defect feature library, extracting feature parameters, and generating a defect feature vector; Step S14, reading the calibrated standardized data set, combining the pre-stored sensor performance parameters, calculating the signal-to-noise ratio, integrity and reliability index of various types of data, and generating a data reliability matrix; Step S15: input the calibrated standardized data set, defect feature vector and data reliability matrix into the feature fusion algorithm, assign weights to different features based on the data reliability matrix, and fuse them to generate standardized feature data.

3. The method for intelligent construction of concrete surface based on multimodal perception according to claim 1 is characterized in that: Step S2 is specifically as follows: Step S21, read the standardized feature data, perform sliding time window segmentation according to the pre-stored time window parameters, and generate a time series segment set; perform data integrity check on the time series segment set, supplement missing data, and generate a complete time series set; Step S22, reading the complete time series set, combining the pre-stored time series alignment parameters, and using a fast dynamic time warping algorithm to calculate the time distortion metric between different sensor data; Construct the warped path matrix, solve the optimal alignment path, and generate the aligned time series; Step S23, performing wavelet decomposition on the aligned time series, reading pre-stored wavelet decomposition parameters, and extracting time-frequency features at different scales; Match each scale feature with the pre-stored feature template library to generate a multi-scale feature matrix; Step S24, reading the multi-scale feature matrix, combining the pre-stored feature weight parameters, calculating the information entropy and importance index of each feature, and generating a feature weight vector; Step S25: input the multi-scale feature matrix and the feature weight vector into the feature fusion algorithm for weighted combination to generate time series alignment data.

4. The method for intelligent construction of concrete surface based on multimodal perception according to claim 1 is characterized in that: Step S3 is specifically as follows: Step S31, mapping the timing alignment data and the defect feature vector to a high-dimensional feature space, reading pre-stored feature mapping parameters, extracting state feature parameters, and generating a state feature matrix; Step S32, read the state feature matrix and the pre-stored Bayesian network model parameters to construct a dynamic Bayesian network; calculate the conditional probability table between nodes and generate a probability dependency graph; Step S33: input the state feature matrix into the probability dependency graph, combine it with the pre-stored difficulty assessment parameters, calculate the difficulty coefficient of the construction area, and generate a construction difficulty graph; Step S34: Based on the dynamic Bayesian network constructed based on the probability dependency graph, the state feature matrix is ​​input, state reasoning is performed, and an environment state vector is generated.

5. The method for intelligent construction of concrete surface based on multimodal perception according to claim 1 is characterized in that: Step S4 is specifically as follows: Step S41, reading the construction difficulty map and the environmental state vector, combining the pre-stored area segmentation parameters, and using the spectral clustering algorithm to segment the construction area; merging and evaluating adjacent areas to generate a task area set; Step S42: read the task area set and the construction difficulty map, calculate the spatiotemporal position coefficient and the construction difficulty coefficient of each area in combination with the pre-stored priority evaluation parameters; integrate multiple evaluation indicators to generate a task priority vector; Step S43, reading the environment state vector and the pre-stored robot performance parameter library, calculating the operating efficiency, accuracy and stability index of each robot in the current environment; comprehensively evaluating various performance parameters to generate a capability evaluation matrix; Step S44: input the task area set, the task priority vector and the capability evaluation matrix into the task allocation algorithm, combine the pre-stored task allocation constraints, solve the optimal allocation plan, and generate a task allocation plan; Step S45, read the task allocation plan and the environment state vector, combine with the pre-stored collaborative rule library, formulate the communication plan and obstacle avoidance rules between robots, and generate a collaborative strategy.

6. The method for intelligent construction of concrete surface based on multimodal perception according to claim 1 is characterized in that: Step S5 is specifically as follows: Step S51, read the environment state vector and the task allocation plan, combine with the pre-stored parameter constraint library, establish the value range and constraint relationship of the construction parameter; integrate multiple constraint conditions to generate a parameter constraint set; Step S52: read the environmental state vector and the defect feature vector, combine with the pre-stored tool performance parameter library, evaluate the construction effect and adaptability index of various tools, and generate a tool evaluation matrix; Step S53: input the parameter constraint set and the tool evaluation matrix into the multi-objective optimization algorithm, combine the pre-stored optimization target weights, solve the optimal parameter combination, and generate an optimization parameter set; Step S54: read the optimization parameter set and the environmental state vector, combine with the pre-stored tool configuration rule library, plan the tool use strategy for different construction stages, and generate a tool configuration plan.

7. The method for intelligent construction of concrete surface based on multimodal perception according to claim 1 is characterized in that: Step S6 is specifically as follows: Step S61, read the construction difficulty map and task allocation plan, and convert the spatial information into a three-dimensional grid representation in combination with the pre-stored environmental modeling parameters; perform terrain feature extraction and feasibility analysis on the grid data to generate an environmental model; Step S62: read the environment model and the coordination strategy, and calculate the comprehensive cost including the difficulty of passage, the difficulty of construction and the coordination constraint in combination with the pre-stored cost function parameter library; perform cost evaluation on each spatial position and generate a path cost map; Step S63, reading the path cost map and the task allocation plan, combining the pre-stored path planning parameters, using the improved A* algorithm to plan the initial path for each robot; verifying the feasibility of the path, and generating an initial path set; Step S64, reading the initial path set and the coordination strategy, and calculating the space-time intersection of the robot path in combination with the pre-stored collision detection parameters; analyzing the potential collision risk, and generating a collision constraint set; Step S65: input the initial path set, the collision constraint set and the optimization parameter set into the trajectory optimization algorithm, and optimize the path for collision avoidance in combination with the pre-stored trajectory optimization parameters to generate a collaborative construction trajectory.

8. The method for intelligent construction of concrete surface based on multimodal perception according to claim 1, characterized in that: Step S11 is specifically as follows: Step S111, read the original sensor data, combine with the pre-stored data format parameter library, unify the format and convert the dimensions of the data of different sensors; extract the timestamp and sensor identifier, and generate unified format data and data attribute table; Step S112, reading the unified format data, combining with the pre-stored normalization parameter library, analyzing the numerical distribution and range of various sensor data; calculating the normalization coefficient, and generating a normalization parameter set; Step S113, reading unified format data and normalization parameter set, and performing minimum-maximum normalization processing on various sensor data; Perform numerical mapping and scale conversion to generate normalized data sets and normalized mapping relationships; Step S114, reading the normalized data set, combining with the pre-stored sensor calibration parameter library, extracting the calibration curve and zero point parameters of each sensor; Analyze sensor characteristic deviations and generate calibration coefficient matrix; Step S115, reading the normalized data set and the calibration coefficient matrix, and performing zero point correction and sensitivity calibration; Compensate for nonlinear errors and generate a calibrated, standardized data set.

9. The method for intelligent construction of concrete surface based on multimodal perception according to claim 1, characterized in that: Step S41 is specifically as follows: Step S411: read the construction difficulty map and the environmental state vector, and build the spatial adjacency relationship between regions in combination with the pre-stored similarity calculation parameter library; calculate the local feature similarity, and generate a similarity matrix and an adjacency relationship table; Step S412: read the similarity matrix and the adjacency table, and construct the Laplace matrix in combination with the pre-stored spectral decomposition parameter library; calculate the eigenvalues ​​and eigenvectors, extract the main eigencomponents, and generate the eigenspace matrix and spectral decomposition report; Step S413, read the feature space matrix and spectral decomposition report, and perform K-means clustering operation in combination with the pre-stored clustering parameter library; optimize the number of clusters based on the silhouette coefficient, and generate an initial segmentation result and a clustering evaluation table; Step S414, reading the initial segmentation result and the adjacency table, combining the pre-stored region segmentation parameters, and analyzing the region boundary features; Calculate boundary smoothness and regional connectivity, generate region segmentation map and boundary feature table; Step S415, reading the region segmentation map and the boundary feature table, and calculating the merging cost of adjacent regions in combination with the pre-stored region merging parameter library; The feasibility of merging is assessed based on the uniformity of construction difficulty and consistency of environmental status, and task area collection and area attribute reports are generated.

10. A concrete surface intelligent construction system based on multimodal perception, characterized in that: include: at least one processor; as well as, a memory communicatively connected to at least one of the processors; wherein, The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the intelligent construction method for concrete surface based on multimodal perception as described in any one of claims 1 to 9.

Images