Post-blasting rock mass collapse volume prediction method based on point cloud deep learning

Abstract

The invention discloses a post-blasting rock mass collapse volume prediction method based on point cloud deep learning. The method comprises the following steps: acquiring initial sparse point cloud data of a post-blasting rock mass and carrying out filtering preprocessing; acquiring a low-dimensional global feature vector by adopting a multi-layer perceptron encoder; photon signal features are generated by using the step-by-step autoregression prediction network and the neuromorphic photon feedforward matrix; high-density photon characteristics are formed through a coherent interference mechanism; high-density electric signal point cloud data are obtained through photoelectric conversion; and calculating the rock mass collapse volume through an Alpha Shape algorithm and outputting the rock mass collapse volume in real time. According to the invention, accurate and efficient prediction of the rock mass collapse volume is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of geotechnical engineering disaster prediction, and particularly relates to a blasting rock mass collapse volume prediction method based on point cloud deep learning. BACKGROUND

[0002] Rock mass blasting is widely used in mining, tunnel engineering and slope treatment, etc. After the implementation of blasting operation, the determination of the collapse rock mass volume is directly related to the construction safety, construction period control and cost management. At present, the prediction of the collapse volume after rock mass blasting is mainly realized through traditional measurement and point cloud data analysis. Among them, the traditional measurement method relies on manual field surveying and geometric calculation, which has the problems of high operation intensity, low measurement efficiency, insufficient precision and inability to measure and calculate in real time. In recent years, the rock mass volume prediction technology based on point cloud data analysis has become the mainstream technical path. The main principle is to obtain the spatial point cloud data of the collapsed rock mass by using three-dimensional laser scanning technology, and to obtain the rock mass volume value through data preprocessing, feature extraction and geometric reconstruction calculation.

[0003] The existing point cloud data analysis method generally uses conventional deep learning technologies such as multilayer perception network to extract rock mass point cloud features, and then predicts the collapse volume through a geometric calculation model. Such method generally removes noise through simple filtering processing in the data preprocessing stage, and relies on traditional deep network for spatial coding in the feature extraction stage, while the geometric calculation model generally adopts a spatial convex hull algorithm. However, the existing method has the problems of high feature dimension and insufficient global spatial information expression ability in the process of extracting point cloud data features, which leads to the inability to fully reflect the detailed features such as rock mass cracks in subsequent geometric calculation. In addition, the spatial convex hull algorithm cannot accurately describe the complex geometric boundary of the rock mass, which makes it difficult to effectively improve the prediction accuracy. The existing technology has the problems of insufficient feature extraction accuracy, inaccurate geometric boundary recognition and poor real-time performance, which cannot meet the needs of high-precision and real-time monitoring and management of current rock mass engineering.

[0004] Therefore, how to provide a blasting rock mass collapse volume prediction method based on point cloud deep learning is a problem that needs to be solved by those skilled in the art. SUMMARY

[0005] One object of the present application is to provide a blasting rock mass collapse volume prediction method based on point cloud deep learning. In view of the problem of insufficient point cloud feature extraction accuracy in the existing technology for predicting the collapse volume of rock mass blasting, a technical scheme is proposed, which uses a multilayer perception encoder and a step-by-step autoregressive prediction network in combination with a neuromorphic photon feedforward matrix to process features, and uses a coherent interference mechanism and an Alpha Shape algorithm to realize accurate prediction of the collapse volume. The present application has the effects of high prediction accuracy, good real-time performance and fine spatial geometric boundary recognition.

[0006] The point cloud deep learning-based post-blasting rock mass collapse volume prediction method according to the embodiment of the present application comprises: Collecting initial sparse point cloud data of a rock mass after blasting, and performing preprocessing to obtain sparse point cloud data; Encoding and globally pooling the sparse point cloud data through a multi-layer perception encoder to obtain an initial low-dimensional global feature vector; Constructing a step-by-step autoregressive prediction network based on a PointARU algorithm, inputting the initial low-dimensional global feature vector into a first-level prediction network, and performing step-by-step autoregressive prediction processing in each prediction network node by using a neuromorphic photon feedforward matrix to obtain photon signal features at each level; Inputting the photon signal features at each level into the neuromorphic photon feedforward matrix, performing prediction calculation through a coherent interference mechanism of the photon signal to obtain high-density photon signal features; Decoding the high-density photon signal features into high-density point cloud data through photoelectric conversion; Performing calculation and processing on the high-density point cloud data by using a volume calculation model based on an Alpha Shape algorithm to obtain a rock mass collapse volume value; Real-time outputting the rock mass collapse volume value to a field monitoring management system.

[0007] Optionally, the collecting of the initial sparse point cloud data of the rock mass after blasting and the preprocessing to obtain the sparse point cloud data are specifically as follows: Obtaining initial sparse point cloud data at different viewing angles by cooperatively scanning a rock mass surface through multiple three-dimensional laser scanners arranged around the blasted rock mass; Fusing the initial sparse point cloud data at different viewing angles to a unified spatial coordinate system, deleting isolated noise points that do not simultaneously appear in scanning results at at least two adjacent viewing angles to obtain fused point cloud data; Based on a local density-adaptive non-fixed radius filtering rule, expanding a neighborhood radius centered at each point to a point density that is stable, removing local outlier noise points whose point number variance within an adaptive radius exceeds a threshold, and obtaining filtered sparse point cloud data; Performing normalization processing on the filtered sparse point cloud data, and adjusting scale proportions of different coordinate dimensions with a center of point cloud coordinate extreme values as an origin to obtain multi-scale nonlinear scale unified point cloud data; Mapping the multi-scale nonlinear scale unified point cloud data to a preset spatial reference plane coordinate system, and correcting spatial positions point by point according to a coordinate transformation relationship of spatial reference plane feature points to obtain preprocessed sparse point cloud data.

[0008] Optionally, the encoding and globally pooling of the sparse point cloud data through the multi-layer perception encoder to obtain the initial low-dimensional global feature vector are specifically as follows: The pre-processed sparse point cloud data is input into the input layer of the multi-layer perception encoder point by point, and the spatial three-dimensional coordinates of each point are linearly mapped to obtain an initial encoding vector point by point; The initial encoding vector point by point is input into the first level hidden layer node of the multi-layer perception encoder, linear combination operation is performed through a dynamic weighting factor, and a nonlinear activation function is used for processing to obtain a first level hidden layer feature vector; The first level hidden layer feature vector is input into the second level hidden layer node of the multi-layer perception encoder, the correlation weight is calculated according to the spatial position correlation relationship of adjacent nodes, the first level hidden layer feature vector is subjected to non-uniform linear combination and nonlinear activation processing, and a second level hidden layer feature vector is obtained. The second level hidden layer feature vector is transmitted to the last level hidden layer node of the multi-layer perception encoder, a feature attention weight matrix is generated according to the spatial position feature difference, and the second level hidden layer feature vector is subjected to feature selection and adaptive combination processing to obtain a last level hidden layer feature vector. A maximum value pooling operation is performed on each feature component of the last level hidden layer feature vector, and the maximum value of the corresponding component of all points in the spatial point cloud is selected dimension by dimension to form a single global feature vector after pooling. The single global feature vector after pooling is mapped to a low-dimensional feature space through a pre-trained linear dimension reduction mapping matrix to obtain an initial low-dimensional global feature vector.

[0009] Optionally, the step-by-step autoregressive prediction network comprises a first level feature initialization layer, a bidirectional feature interaction transmission layer, a dynamic memory unit update layer, a multi-scale feature self-enhancement layer and a feature state fusion output layer: The first level feature initialization layer is configured to receive the initial low-dimensional global feature vector and generate an initial input feature vector of the first level prediction network in a linear mapping manner. The bidirectional feature interaction transmission layer is configured to perform forward feature vector prediction and feedback feature vector transmission step by step from the first level prediction network: In the forward feature vector prediction path, each level node receives the feature vector of the previous node through a neuromorphic photon feedforward matrix and outputs a photon signal feature of the current node. In the feedback feature vector transmission path, each level node receives the photon signal feature fed back by the next node through a neuromorphic photon feedforward matrix and generates a feedback enhanced photon signal feature. The dynamic memory unit update layer is configured to fuse the photon signal features output by the forward path and the feedback path to generate a node feature state updated step by step in the dynamic memory unit. The multi-scale feature self-enhancement layer is configured to decompose the node feature state in the dynamic memory unit into multiple scale sub-feature spaces, and perform self-enhancement reconstruction on the features in a step-by-step manner to obtain the scale-enhanced feature state. The feature state fusion output layer is configured to fuse the final feature states of the feedforward path and the feedback path, and output high-density photon signal features.

[0010] Optionally, the initial low-dimensional global feature vector is input into the first-level prediction network, and a neuromorphic photon feedforward matrix is used in each prediction network node to perform step-by-step autoregressive prediction processing, thereby obtaining photon signal features at each level, specifically as follows. The initial low-dimensional global feature vector is input into the first-level prediction network node, and the neuromorphic photon feedforward matrix is used to perform light intensity and phase modulation, thereby obtaining a photon input feature vector. The photon input feature vector is transmitted to the silicon-based optical waveguide structure in the neuromorphic photon feedforward matrix to perform weighted linear combination operation, thereby obtaining an intermediate photon signal feature at the first level. The intermediate photon signal feature at the first level is transmitted to the micro-ring resonator structure in the neuromorphic photon feedforward matrix to perform nonlinear light intensity modulation, thereby obtaining a photon signal feature after autoregressive prediction at the first level. The photon signal feature after autoregressive prediction at the first level is transmitted to the next prediction network node in a step-by-step manner, and the light intensity and phase modulation, the weighted linear combination operation, and the nonlinear light intensity modulation are repeatedly performed in a step-by-step manner, thereby obtaining a photon signal feature after autoregressive prediction at the current node at each level. The dynamic memory unit is used to memorize the photon signal features output by the step-by-step prediction network nodes, thereby forming a step-by-step updated node feature state. The step-by-step updated node feature state is divided into multiple scale sub-feature spaces, scale attention weights are calculated, and the feature states in the scale sub-feature spaces are reconstructed by the scale attention weights, thereby obtaining photon signal features after autoregressive processing at each level.

[0011] Optionally, the step-by-step prediction calculation is performed on the photon signal features at each level by using the coherent interference mechanism of the photon signal, thereby obtaining high-density photon signal features, specifically as follows. The photon signal features after autoregressive processing at each level are input into the silicon-based optical waveguide structure in a step-by-step manner, and the phase parameters of the photon signal features are modulated in a spatial wave division multiplexing manner, thereby obtaining input photon features after phase modulation. The directional coupler array in the silicon-based optical waveguide structure is used to perform step-by-step spatial coupling operation on the input photon features, thereby generating intermediate photon signals after spatial coupling. The intermediate photon signal after spatial coupling is introduced into a multi-stage cascaded micro-ring resonant cavity structure, so that the intermediate photon signal is subjected to optical wave coherent interference in the multi-stage cascaded micro-ring resonant cavity; The optical path difference of the intermediate photon signal is adjusted through the cavity optical path delay mechanism of the multi-stage cascaded micro-ring resonant cavity structure, so that spatial coherent superposition or destructive interference is realized, and interference photon signals are generated step by step; The light intensity of the interference photon signals generated step by step is detected by a photon intensity detector, and the light intensity spatial distribution characteristics are fused to obtain high-density photon signal characteristics.

[0012] Optionally, the high-density photon signal characteristics are decoded into high-density point cloud data through photoelectric conversion, specifically: The high-density photon signal characteristics are input into a plurality of photoelectric conversion units of an array photoelectric detector, and photoelectric conversion is performed point by point; The light intensity is measured point by point by the array photoelectric detector to obtain a light intensity voltage value; The light intensity voltage value is normalized point by point to convert into a feature intensity scalar corresponding to a point cloud spatial coordinate, and a high-density point cloud feature scalar with a spatial position index is generated; Based on the high-density point cloud feature scalar, the spatial coordinate information of the preprocessed sparse point cloud data is used for interpolation calculation to supplement and reconstruct the high-density electrical signal point cloud spatial coordinate data; Based on a spatial neighborhood feature consistency determination rule, a consistency threshold between the feature scalar and the adjacent spatial position feature scalar is calculated point by point, outliers with a spatial feature difference exceeding a set threshold are removed, and filtered high-density electrical signal point cloud data are obtained; Through fusion mapping of the spatial position coordinates and the feature intensity scalar, the filtered high-density electrical signal point cloud data are converted into high-density point cloud data with fused spatial coordinates and detail crack features.

[0013] Optionally, the high-density point cloud data are calculated and processed by using a volume calculation model based on an Alpha Shape algorithm to obtain a rock mass collapse volume value, specifically: The high-density electrical signal point cloud data are input into the volume calculation model based on the Alpha Shape algorithm, and a corresponding relationship between the spatial position index and the three-dimensional spatial coordinate is established point by point; Based on the spatial position index, the Euclidean distance between each spatial point and other spatial points is calculated to generate a complete distance matrix between the spatial points; According to the distance matrix, the spatial scale parameter is gradually increased, and when the connection relationship between the spatial points is stable, the current spatial scale parameter is taken as a final determined parameter value of the Alpha Shape algorithm; According to the finally determined spatial scale parameter, the effective spatial neighborhood relationship is recognized point by point for the high-density electric signal point cloud data, and the adjacent spatial points meeting the conditions are connected into a spatial closed geometric grid structure; For each effective triangular face element of the spatial closed geometric grid structure, a spatial tetrahedron structure is formed by connecting the spatial coordinate origin and the three vertices of the triangular face element, and the element volume value of the spatial tetrahedron structure is calculated one by one; The element volume values of all spatial tetrahedron structures are added one by one to obtain the rock mass collapse volume value corresponding to the high-density electric signal point cloud data.

[0014] The beneficial effects of the present application are: (1) The present application realizes fine intelligent extraction of rock mass collapse point cloud features after blasting by adopting a multi-layer perception machine encoder and a step-by-step autoregressive prediction network based on a neuromorphic photonic feedforward matrix, effectively improves the precision and stability of feature extraction, and enhances the expression ability of detail features such as rock mass surface fissures.

[0015] (2) The present application realizes fine demarcation of the spatial geometric boundary of the rock mass collapse area by the Alpha Shape algorithm and the dynamic scale parameter optimization mechanism, significantly improves the spatial accuracy and real-time performance of rock mass volume prediction, and shows better adaptability and reliability in the scene of real-time monitoring and evaluation of rock mass engineering disasters.

[0016] (3) In the aspect of real-time and accurate prediction of rock mass collapse volume, the present application effectively solves the problems of low feature expression precision and insufficient calculation real-time performance in the prior art by using the step-by-step coherent interference mechanism of neuromorphic photonic signals and the high-efficiency photoelectric conversion structure, breaks through the bottleneck that traditional methods cannot accurately express complex rock mass detail features, realizes specific and significant progress in the precision and calculation real-time performance of rock mass collapse volume prediction, and effectively improves the application level in the field of rock mass engineering monitoring management and disaster evaluation. BRIEF DESCRIPTION OF DRAWINGS

[0017] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, illustrate the present application together with the embodiments thereof, and explain the present application, and do not constitute a limitation of the present application. In the drawings: Figure 1 The flowchart of the rock mass collapse volume prediction method after blasting based on point cloud deep learning proposed in the present application. DETAILED DESCRIPTION

[0018] The present application will now be further described in detail in conjunction with the drawings. These drawings are all simplified schematic diagrams, and only illustrate the basic structure of the present application in a schematic manner, and therefore only show the components related to the present application.

[0019] REFERENCE Figure 1, a blasting rock mass collapse volume prediction method based on point cloud deep learning, comprising: Collecting initial sparse point cloud data of the rock mass after blasting, removing isolated noise points by using a statistical filtering algorithm, removing local outlier noise points by using a radius filtering algorithm, and performing scale unification processing on the filtered data by using a normalization method to obtain preprocessed sparse point cloud data; Performing point-by-point spatial mapping encoding on the preprocessed sparse point cloud data by using a multi-layer perception encoder, and performing global pooling processing on the local feature vectors of each point after encoding to obtain an initial low-dimensional global feature vector of the point cloud data; Constructing a step-by-step autoregressive prediction network based on the PointARU algorithm, and taking the initial low-dimensional global feature vector as the input feature of the first-level prediction network. In each prediction network node, a neuromorphic photon feedforward matrix is used to complete step-by-step autoregressive prediction processing of the feature vector, and photon signal features after each autoregressive processing are obtained. Inputting the photon signal features of each level into the neuromorphic photon feedforward matrix, performing step-by-step prediction calculation on the input photon signal features through the coherent interference mechanism of the photon signal, and obtaining high-density photon signal features after coherent interference processing; Through photoelectric conversion, the high-density photon signal features are decoded into high-density point cloud data in the form of electrical signals, and high-density point cloud data fused with surface crack detail features of the rock mass are obtained; Using a volume calculation model based on the Alpha Shape algorithm to calculate and process the spatial geometric relationship between each point in the high-density point cloud data, obtaining the spatial geometric boundary of the rock mass collapse area, and obtaining the rock mass collapse volume value accordingly; Outputting the rock mass collapse volume value to the field monitoring management system in real time for use in the blasting construction decision-making process.

[0020] In the embodiment, the initial sparse point cloud data of the rock mass after blasting is collected, isolated noise points are removed by using a statistical filtering algorithm, local outlier noise points are removed by using a radius filtering algorithm, and scale unification processing is performed on the filtered data by using a normalization method to obtain preprocessed sparse point cloud data, specifically: A plurality of three-dimensional laser scanners arranged at different spatial positions around the blasted rock mass are used to cooperatively scan the surface of the blasted rock mass, and initial sparse point cloud data at different viewing angles are obtained; The initial sparse point cloud data of each view angle is fused into a unified spatial coordinate system, and a spatial topological consistency method of the multi-view overlapping area is used to identify and delete isolated noise points that do not appear in the scanning results of at least two adjacent view angles at the same time, to obtain the fused point cloud data, and the spatial topological consistency method of the multi-view overlapping area is as follows: in the unified spatial coordinate system, the spatial topological structure is established according to the relative position relationship between points in the point cloud data obtained by scanning different view angles, the topological structure similarity of the point cloud data of each view angle in the overlapping area is calculated, and only the points that meet the consistency condition are retained and the isolated noise points that do not meet the consistency condition are removed, with the topological structure similarity reaching a preset consistency threshold as the judgment condition; For the fused point cloud data, a non-fixed radius filtering rule based on local density adaptation is established, the neighborhood radius is automatically expanded to a local point density reaching a stable threshold centering on each point, and the variance of the number of points in the adaptive radius is used as the filtering judgment standard to remove local outlier points exceeding the set variance threshold, to obtain the filtered sparse point cloud data, and the non-fixed radius filtering rule based on local density adaptation is as follows: taking each point in the fused point cloud data as the center, the initial neighborhood radius is the minimum radius value, and the radius is gradually increased until the density change rate of the points in the neighborhood is less than a preset density change rate threshold, to determine the adaptive neighborhood radius of the corresponding point, and within the determined adaptive neighborhood radius, the number of points in the neighborhood of each point is counted and the variance of the number of points in the neighborhood is calculated, and if the variance exceeds the preset variance threshold, the corresponding point is determined as a local outlier point and is removed; The filtered sparse point cloud data is subjected to scale normalization processing, the scale proportion of different coordinate dimensions is nonlinearly adjusted according to the spatial distribution characteristics of the point cloud, and multi-scale nonlinear scale unified point cloud data is obtained; The multi-scale nonlinear scale unified point cloud data is mapped into a preset spatial reference plane coordinate system, the spatial position is corrected point by point according to the coordinate transformation relationship of the spatial reference plane feature points, and pre-processed sparse point cloud data with spatial reference plane feature matching is obtained.

[0021] In the embodiment, the pre-processed sparse point cloud data is encoded by the multi-layer perception encoder, and the local feature vectors of each point after encoding are subjected to global pooling processing, to obtain the initial low-dimensional global feature vectors of the point cloud data, specifically: The pre-processed sparse point cloud data is input into the input layer of the multi-layer perception encoder point by point in the order of spatial coordinates, and the spatial three-dimensional coordinates of each point are independently subjected to initial linear mapping calculation to obtain the initial encoding vector of each point; The point-by-point initial encoding vector is input to the first-level hidden layer node of the multi-layer perceptron encoder, each node respectively performing a linear combination operation based on a dynamic weighting factor on the initial encoding vector of all input layer nodes, the dynamic weighting factor being obtained by corresponding calculation of the input features and the preset weight matrix of the hidden layer node, and being point-by-point nonlinearly mapped by using a nonlinear activation function to obtain a first-level hidden layer feature vector; The first-level hidden layer feature vector is input to the second-level hidden layer node of the multi-layer perceptron encoder, and the second-level hidden layer node calculates the correlation weight of the features of the adjacent spatial nodes on the basis of inheriting the spatial position correlation of the first-level hidden layer feature vector, and performs non-uniform linear combination and nonlinear activation processing on each node feature vector according to the correlation weight to obtain a second-level hidden layer feature vector with enhanced spatial position features point by point; The second-level hidden layer feature vector is transmitted to the last-level hidden layer node of the multi-layer perceptron encoder in a step-by-step transmission manner, and the last-level hidden layer node generates a feature attention weight matrix through a preset feature self-attention mechanism according to the spatial position feature difference of the hidden layer node, and performs feature selection and adaptive combination processing on the second-level hidden layer feature vector based on the feature attention weight matrix to obtain a last-level hidden layer feature vector, wherein the preset feature self-attention mechanism is that, for the point-by-point second-level hidden layer feature vector input to the last-level hidden layer node, the correlation strength between each feature component and the corresponding feature components of other points in the spatial neighborhood is calculated point by point to form a point-by-point feature correlation strength matrix, and a point-by-point feature attention weight matrix is generated according to the point-by-point feature correlation strength matrix, and the point-by-point second-level hidden layer feature vector is weighted and combined by using the point-by-point feature attention weight matrix to obtain a point-by-point last-level hidden layer feature vector; A maximum value pooling operation is performed on each feature component of the last-level hidden layer feature vector, and the maximum value of the corresponding component of all points in the spatial point cloud is selected for each feature component of the last-level hidden layer feature vector to form a single global feature vector after pooling dimension by dimension; The single global feature vector after pooling is mapped to a low-dimensional feature space by using a pre-trained linear dimension reduction mapping matrix to obtain an initial low-dimensional global feature vector representing the overall spatial structure features of the point cloud, wherein the pre-trained linear dimension reduction mapping matrix is a feature mapping matrix obtained by matrix parameter optimization training through a linear regression method, using a historical blasting rock mass point cloud data to construct a training sample set, using the single global feature vector after pooling as input, and using the true measurement value of the corresponding rock mass collapse volume as target output, and is used to map the single global feature vector after pooling to a low-dimensional feature space to obtain an initial low-dimensional global feature vector.

[0022] In the embodiment, the step-by-step autoregressive prediction network comprises a first-level feature initialization layer, a bidirectional feature interaction transmission layer, a dynamic memory cell updating layer, a multi-scale feature self-enhancement layer, and a feature state fusion output layer. The first-level feature initialization layer is configured to receive an initial low-dimensional global feature vector, map the initial low-dimensional global feature vector to an initial state feature space of a first prediction node by using a linear mapping manner, and generate an initial input feature vector of the first-level prediction network. The bidirectional feature interaction transmission layer is configured to perform a feedforward feature vector prediction from the first-level prediction network in a step-by-step manner, and simultaneously establish a feedback feature vector transmission path in a step-by-step manner, to form a bidirectional dynamic feature transmission channel based on the initial input feature vector. In the feedforward feature vector prediction path, each node receives a feature vector output by a previous node in a step-by-step autoregressive manner through a neuromorphic photonic feedforward matrix, and outputs a photonic signal feature predicted by the current node. In the feedback feature vector transmission path, each node receives a high-level photonic signal feature fed back by a next node, and performs feature reverse mapping processing through a neuromorphic photonic feedforward matrix to generate a low-level node photonic signal feature enhanced by feedback. The neuromorphic photonic feedforward matrix is composed of a plurality of cascaded photonic neuron arrays, each photonic neuron array comprising an input photonic signal coupling unit, a photonic weighted interconnection unit, and an output photonic nonlinear activation unit. The input photonic signal coupling unit is configured to receive an input photonic signal feature transmitted from a previous node or a next node, and modulate the input photonic signal feature in terms of initial light intensity and phase. The photonic weighted interconnection unit is configured to implement weighted coupling and linear combination of the input photonic signal feature modulated at the input end by using an integrated silicon-based optical waveguide structure, and output a weighted intermediate photonic signal feature. The output photonic nonlinear activation unit is configured to receive the weighted intermediate photonic signal feature, and implement nonlinear activation modulation of the photonic signal based on an integrated micro-ring resonant cavity structure, to output a photonic signal feature activated by nonlinearity. The dynamic memory unit updating layer is configured to dynamically memorize the photon signal features output by the feedforward path and the feedback path of each level node, and to fuse the photon signal features in the bidirectional feature paths by using an adaptive gating mechanism, so as to determine the feature fusion ratio by using a gating state weight matrix, and to form the node feature state updated in the dynamic memory unit. The adaptive gating mechanism is configured to calculate the feature correlation coefficient between the feedforward features and the feedback features for the photon signal features output by the feedforward path and the feedback path, to construct the gating state weight matrix by using the feature correlation coefficient, to determine the feature fusion ratio of the feedforward path and the feedback path according to the corresponding weight ratio in the gating state weight matrix, and to fuse the photon signal features of the feedforward path and the feedback path according to the ratio, so as to update the node feature state in the dynamic memory unit. The multi-scale feature self-enhancement layer is configured to decompose the node feature state updated in the dynamic memory unit into a plurality of sub-feature spaces of different scales in the spatial scale, to calculate the scale attention weight for each sub-feature space, and to perform a self-enhancement reconstruction operation on the features by using the scale attention weight, so as to obtain the feature state of each node after scale enhancement. The feature state fusion output layer is configured to receive the final feature states of the feedforward path and the feedback path based on the adaptive feature fusion matrix after the feedforward path and the feedback path complete the step-by-step prediction transmission, to perform fusion calculation on the final feature states of the two paths according to the weight distribution ratio corresponding to the fusion matrix, and to output the high-density photon signal features with the prediction features of the feedforward path and the enhanced features of the feedback path.

[0023] In the embodiment, the initial low-dimensional global feature vector is taken as the input feature of the first level prediction network, and in each level prediction network node, a neuromorphic photon feedforward matrix is used to complete the step-by-step autoregressive prediction processing of the feature vector, so as to obtain the photon signal features after the autoregressive processing of each level. Specifically, The initial low-dimensional global feature vector is input into the first level prediction network node, and a neuromorphic photon feedforward matrix is used to implement the light intensity and phase modulation processing on the initial low-dimensional global feature vector, so as to obtain the photon input feature vector after the initial modulation; The photon input feature vector after the initial modulation is transmitted to the silicon-based optical waveguide structure in the neuromorphic photon feedforward matrix, a weighted linear combination operation is performed on each photon signal component in the input feature vector, and the first level intermediate photon signal feature is obtained; The first level intermediate photon signal feature is transmitted to the micro-ring resonator structure in the neuromorphic photon feedforward matrix, and the micro-ring resonator structure is used to implement the nonlinear light intensity modulation on the intermediate photon signal feature, so as to obtain the photon signal feature after the autoregressive prediction of the first level; The first-order autoregressive predicted photon signal features are transmitted to the next prediction network node, and the light intensity and phase modulation of the neuromorphic photon feedforward matrix, the weighted linear combination operation of the silicon-based optical waveguide structure, and the nonlinear light intensity modulation processing of the micro-ring resonant cavity structure are repeatedly performed at the next prediction network node to generate the current node's autoregressive predicted photon signal features step by step. The dynamic memory unit is used to store the state of the autoregressive predicted photon signal features output by the prediction network node step by step, and the node feature state is updated step by step. The node feature state updated step by step is divided into multiple scale feature spaces in the spatial dimension, and the scale attention weight is calculated for each scale component of the scale feature space. The scale attention weight is used to reconstruct the feature state of the scale feature space step by step, and the scale-enhanced photon signal features after each level of autoregressive processing are obtained.

[0024] In this embodiment, the photon signal features of each level are input into the neuromorphic photon feedforward matrix, and the input photon signal features are predicted and calculated step by step through the coherent interference mechanism of the photon signal to obtain high-density photon signal features after coherent interference processing, specifically: The photon signal features after each level of autoregressive processing are input into the silicon-based optical waveguide structure of the neuromorphic photon feedforward matrix step by step to modulate the initial phase parameters of each photon signal feature in a spatial wave division multiplexing manner, and obtain the phase-modulated input photon features of each level. The spatial wave division multiplexing manner is to modulate each autoregressive processed photon signal feature into a different optical waveguide transmission channel, so that each input photon feature propagates independently along parallel and non-interfering optical transmission paths in the same silicon-based optical waveguide structure, to realize the spatial multiplexing of multiple photon signal features. The directional coupler array arranged inside the silicon-based optical waveguide structure is used to perform step-by-step photon signal spatial coupling operation on each level of input photon features, so that the input photon features from different nodes are superimposed along the spatial dimension to generate intermediate photon signals after spatial coupling of each level. The directional coupler array is a plurality of directional coupler units arranged in an array in the silicon-based optical waveguide structure. Each directional coupler unit is composed of two silicon-based optical waveguides arranged in close proximity. By adjusting the coupling distance and length parameters between the two parallel optical waveguides, the photon signal can be controlled to exchange energy between adjacent waveguides according to the preset coupling ratio, thereby realizing the spatial superposition coupling of input photon features from different nodes. The intermediate photon signals after spatial coupling of each level are introduced into the multi-stage cascaded micro-ring resonant cavity structure in the silicon-based optical waveguide structure, so that the superimposed intermediate photon signals propagate in the multi-stage cascaded micro-ring resonant cavity and undergo light wave coherent interference phenomenon in the micro-ring resonant cavity. The optical path difference of each intermediate photon signal on different propagation paths is adjusted by a cavity optical path delay mechanism of a multi-stage cascaded micro-ring resonant cavity structure, and the spatial coherent superposition or destructive interference of the intermediate photon signals is realized by the optical path difference, and the interference photon signals after coherent interference processing are generated step by step, and the cavity optical path delay mechanism is that: by adjusting the refractive index and structure size parameters of the ring-shaped silicon-based optical waveguide in the micro-ring resonant cavity unit, the photon signals propagating in different paths in the micro-ring resonant cavity unit produce a preset optical path difference, so as to realize the controllable phase delay relationship of the photon signals on the output end; The point-by-point light intensity detection of the interference photon signals generated step by step is performed by a photon intensity detector, the light intensity spatial distribution characteristics corresponding to the interference photon signals are obtained, and the light intensity spatial distribution characteristics of the interference photon signals at each stage are fused to obtain high-density photon signal characteristics.

[0025] In the embodiment, the high-density photon signal characteristics are decoded into high-density point cloud data in the form of electrical signals by photoelectric conversion, high-density point cloud data fused with the fracture detail characteristics of the rock mass surface is obtained, and the high-density point cloud data fused with the fracture detail characteristics of the rock mass surface is obtained, and the high-density point cloud data fused with the fracture detail characteristics of the rock mass surface is obtained. The high-density photon signal characteristics are input into an arrayed photodetector in the photoelectric conversion structure step by step, and the arrayed photodetector is composed of a plurality of photoelectric conversion units, each photoelectric conversion unit receives corresponding photon signal characteristics and performs point-by-point photoelectric conversion; The point-by-point light intensity measurement of the received photon signal characteristics at each stage is performed by the arrayed photodetector, and the light intensity voltage value corresponding to each photoelectric conversion unit is obtained; The point-by-point normalization processing of the light intensity voltage value is performed, the normalized light intensity voltage value is converted into a feature intensity scalar corresponding to the spatial coordinate position of the point cloud, and a high-density point cloud feature scalar in the form of electrical signals with spatial position index is generated; A spatial feature interpolation reconstruction method is adopted, the high-density point cloud feature scalar is taken as the basis, the interpolation calculation of the missing points between the spatial positions is performed by using the spatial coordinate information of the known preprocessed sparse point cloud data, and the high-density electrical signal point cloud spatial coordinate data is supplemented and reconstructed; The spatial neighborhood feature consistency determination rule is used to calculate the consistency threshold value between the feature scalar of each spatial position and the feature scalar of the adjacent spatial position based on the reconstructed high-density electrical signal point cloud spatial coordinate data, the outlier points with the spatial feature difference exceeding the set threshold value are removed according to the consistency threshold value, and the filtered high-density electrical signal point cloud data is generated. The filtered high-density electrical signal point cloud data is converted into high-density point cloud data in the form of electrical signal of high-density spatial coordinates and detail fissure feature fusion representation through the fusion mapping of the spatial position coordinates and the feature intensity scalar.

[0026] In the embodiment, the volume calculation model based on the Alpha Shape algorithm is used to calculate and process the spatial geometric relationship between each point in the high-density point cloud data, obtain the spatial geometric boundary of the rock mass collapse area, and obtain the rock mass collapse volume value accordingly, specifically as follows: The high-density electrical signal point cloud data is input into the volume calculation model based on the Alpha Shape algorithm, the point cloud data spatial position index is constructed point by point according to the three-dimensional spatial coordinates of each spatial point in the point cloud data, and the corresponding relationship between the spatial position index and the three-dimensional spatial coordinates is established point by point; Based on the relationship between the spatial position index and the corresponding three-dimensional spatial coordinates, the Euclidean distance between the spatial point and each spatial point in the point cloud data set is calculated with the spatial point as the center, a complete distance matrix between the spatial points is generated, and each element of the distance matrix is the Euclidean distance value between each two spatial points; The spatial scale parameter threshold value of the Alpha Shape algorithm is determined according to the distance matrix, the spatial scale parameter threshold value is determined by gradually increasing the spatial scale parameter and continuously judging the effective neighborhood formation of the spatial point, and the specific manner is as follows: the minimum spatial scale parameter is initially set, the spatial scale parameter threshold value is gradually increased, when the spatial scale parameter threshold value is increased to make the spatial points in the point cloud have stable and no obvious new spatial connection relationship, the current spatial scale parameter threshold value is selected as the final determined parameter value of the Alpha Shape algorithm; According to the spatial scale parameter threshold, the spatial effective neighborhood of each spatial point in the high-density electric signal point cloud data is identified and determined, if the distance between any two spatial points satisfies the condition of the spatial scale parameter threshold, it is determined that the two spatial points constitute an effective spatial connection relationship, and the spatial points constituting the effective connection relationship are connected point by point to generate a spatial closed geometric grid structure composed of a large number of effective triangular face elements connected to each other; For each effective triangular face element in the spatial closed geometric grid structure, the spatial coordinate origin and the three vertices constituting the effective triangular face element are connected respectively to form independent spatial tetrahedral structures one by one, and the unit volume of each spatial tetrahedral structure is calculated one by one: the spatial included angle between each edge in the tetrahedral space unit is calculated, and the unit volume value of the corresponding tetrahedral structure is obtained by calculating the volume value according to the edge length and the included angle; The unit volume values of all tetrahedral space units are calculated one by one, and the unit volume values of all spatial tetrahedral structures are superimposed to obtain the total volume value of the rock mass collapse region corresponding to the high-density electric signal point cloud data, and the rock mass collapse volume value which can fully represent the geometric contour characteristics of the rock mass collapse region is obtained.

[0027] Embodiment 1: In order to verify the feasibility of the application in implementation, the application is applied to a practical rock mass blasting construction scene in a mountainous area to accurately predict the rock mass collapse volume after blasting in real time, and to ensure the safety and efficiency of blasting construction. In this application scenario, the traditional method mainly uses field manual measurement or simple point cloud data analysis technology. The manual measurement method has long operation period, low efficiency and insufficient accuracy. Although the current point cloud analysis method improves the efficiency, it usually uses single filtering method for denoising and simple deep network feature extraction, which leads to incomplete point cloud feature extraction and not detailed data expression, and it is difficult to accurately reflect the detailed fracture characteristics of the rock mass, so the gap between the prediction result and the actual volume value is large, which affects the accuracy and timeliness of the construction decision.

[0028] In the actual operation process, first, multiple three-dimensional laser scanners are deployed around the blasting rock mass to cooperatively scan the rock mass surface and obtain initial sparse point cloud data from different angles. Then, the multi-angle point cloud data is fused into a unified coordinate system, and the spatial topological consistency method is used to accurately remove the isolated noise points that do not appear repeatedly in the multi-angle, and based on the local density adaptive non-fixed radius filtering algorithm, the local outlier noise points are effectively removed, and finally the scale normalization processing is performed to obtain high-quality preprocessed sparse point cloud data. Next, through the multilayer perception machine encoder, the preprocessed point cloud is mapped and the hierarchical hidden layer features are extracted point by point to obtain the initial low-dimensional global feature vector representing the overall spatial structure of the point cloud.

[0029] Then, based on the PointARU step-by-step autoregressive prediction network, the initial low-dimensional feature vector is input to the first prediction node, and the neuromorphic photon feedforward matrix is used to generate photon signal features step by step. Through the coherent interference mechanism of the photon signal, high-density photon signal features are predicted and calculated step by step. Then, the high-density photon signal features are converted into high-density electrical signal point cloud data point by point through the photoelectric conversion structure, and the point cloud coordinates are accurately reconstructed using the spatial feature interpolation method to obtain high-density point cloud data that fuse the details of the rock mass surface fissure. Subsequently, using the improved Alpha Shape algorithm, according to the spatial scale parameter threshold, the effective spatial connection relationship is identified, and the accurate rock mass collapse area geometric grid structure is gradually constructed, the overall volume value of the rock mass collapse area is calculated, and is output to the field monitoring management system in real time to form the basis for blasting construction decision.

[0030] The specific parameter settings in the implementation process are as follows: the spatial topological consistency threshold is 0.85, the local density stability threshold is set to 15 neighborhood points, the initial low-dimensional global feature vector dimension is set to 64; the step-by-step autoregressive prediction network has a total of 3 nodes, the refractive index adjustment range of the micro-ring resonant cavity unit is 1.45 to 1.55, and the photoelectric conversion unit light intensity measurement accuracy is 0.01 mW; the spatial scale parameter threshold in the Alpha Shape algorithm is determined to be 0.65 meters through 27 step-by-step iterations.

[0031] To verify the accuracy of the prediction method, five different rock mass blasting experimental sites were selected, and the predicted volume and the actual manual accurate measurement volume were recorded respectively, and the relevant data are listed in the following table.

[0032] Table 1 Comparison of rock mass collapse volume prediction value and measured value As can be seen from Table 1, the prediction method proposed by the present application has a very high consistency with the actual measured value, and the relative error of volume prediction is stably controlled within 1%. For example, in experimental site I, the predicted volume value is 1248.5 m³, the measured volume value is 1254.7 m³, the error is only 6.2 m³, and the relative error is only 0.49%. In the largest experimental site III, the point cloud scale is close to 300,000 points, the predicted volume value is 1616.9 m³, the measured volume value is 1625.4 m³, the error is 8.5 m³, and the relative error is only 0.52%. The above results show that the method of the present application has very high prediction accuracy and stability in actual blasting construction scene, effectively improves the timeliness and reliability of the construction site decision, and is significantly better than the prior art scheme.

[0033] In summary, the results of the embodiments fully verify the effectiveness and practicality of the method, not only effectively improving the accuracy and real-time of the rock mass collapse volume prediction, but also significantly reducing the labor input of field manual measurement and the risk of construction decision, having important practical application value and wide popularization prospect.

[0034] The above merely describes the preferred embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can make equivalent substitutions or changes within the technical scope disclosed by the present application and according to the technical solutions and inventive concept of the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for predicting the volume of rock mass collapse after blasting based on point cloud deep learning, characterized in that, include: Initial sparse point cloud data of the rock mass after blasting was collected and preprocessed to obtain sparse point cloud data. The sparse point cloud data is encoded and globally pooled by a multilayer perceptron encoder to obtain an initial low-dimensional global feature vector. A stepwise autoregressive prediction network is constructed based on the PointARU algorithm. The initial low-dimensional global feature vector is input into the first-level prediction network, and the neuromorphic photon feedforward matrix is ​​used to perform stepwise autoregressive prediction processing in the nodes of each level of the prediction network to obtain the photon signal features of each level. The photon signal features at each level are input into the neuromorphic photon feedforward matrix, and prediction calculations are performed through the coherent interference mechanism of the photon signals to obtain high-density photon signal features. High-density photonic signal features are decoded into high-density point cloud data through photoelectric conversion; The volume calculation model based on the Alpha Shape algorithm is used to calculate and process high-density point cloud data to obtain the volume value of rock mass collapse. The volume value of rock mass collapse is output to the on-site monitoring and management system in real time.

2. The method for predicting the volume of rock mass collapse after blasting based on point cloud deep learning according to claim 1, characterized in that, The process of collecting initial sparse point cloud data of the rock mass after blasting and preprocessing it to obtain sparse point cloud data specifically involves: Multiple 3D laser scanners positioned around the blasted rock mass were used to collaboratively scan the rock mass surface to obtain initial sparse point cloud data from different perspectives. The initial sparse point cloud data from different perspectives are fused into a unified spatial coordinate system, and isolated noise points that do not appear simultaneously in the scan results of at least two adjacent perspectives are deleted to obtain the fused point cloud data. For the fused point cloud data, based on the non-fixed radius filtering rule of local density adaptation, the neighborhood radius is expanded to the center of each point until the point density is stable, and local outlier noise points with the variance of the number of points within the adaptive radius exceeding the threshold are removed to obtain the filtered sparse point cloud data. The filtered sparse point cloud data is normalized, and the scale ratio of different coordinate dimensions is adjusted with the extreme center of the point cloud coordinates as the origin to obtain point cloud data after multi-scale nonlinear scale unification. The point cloud data after multi-scale nonlinear scaling is mapped to a preset spatial reference plane coordinate system. The spatial position is corrected point by point according to the coordinate transformation relationship of the feature points of the spatial reference plane to obtain the preprocessed sparse point cloud data.

3. The method for predicting the volume of rock mass collapse after blasting based on point cloud deep learning according to claim 1, characterized in that, The process of encoding and global pooling the sparse point cloud data using a multilayer perceptron encoder to obtain an initial low-dimensional global feature vector is as follows: The preprocessed sparse point cloud data is input point by point into the input layer of the multilayer perceptron encoder. The spatial three-dimensional coordinates of each point are linearly mapped and calculated to obtain the point-by-point initial encoding vector. The point-by-point initial encoding vector is input into the first hidden layer node of the multilayer perceptron encoder, and linear combination operation is performed through dynamic weighting factors. Then, a nonlinear activation function is used to process it to obtain the first hidden layer feature vector. The first-level hidden layer feature vector is input into the second-level hidden layer node of the multilayer perceptron encoder. The correlation weight is calculated according to the spatial position relationship between adjacent nodes. The first-level hidden layer feature vector is subjected to non-uniform linear combination and non-linear activation processing to obtain the second-level hidden layer feature vector. The second-level hidden layer feature vector is transmitted step by step to the last hidden layer node of the multilayer perceptron encoder. A feature attention weight matrix is ​​generated based on the differences in spatial location features. Feature selection and adaptive combination processing are performed on the second-level hidden layer feature vector to obtain the last hidden layer feature vector. Max pooling is performed on each feature component of the feature vector of the last hidden layer. The maximum value of the corresponding component of all points in the spatial point cloud is selected dimension by dimension to form a single global feature vector after pooling. By using a pre-trained linear dimensionality reduction mapping matrix, the pooled single global feature vector is mapped to a low-dimensional feature space to obtain an initial low-dimensional global feature vector.

4. The method for predicting the volume of rock mass collapse after blasting based on point cloud deep learning according to claim 1, characterized in that, The hierarchical autoregressive prediction network includes a first-level feature initialization layer, a bidirectional feature interaction and transmission layer, a dynamic memory unit update layer, a multi-scale feature self-enhancement layer, and a feature state fusion output layer. The first-level feature initialization layer is used to receive the initial low-dimensional global feature vector and generate the initial input feature vector of the first-level prediction network using a linear mapping method. The bidirectional feature interaction transmission layer is used to predict feedforward feature vectors step by step downwards from the first-level prediction network, while simultaneously transmitting feedback feature vectors step by step upwards. In the feedforward feature vector prediction path, each node receives the feature vector of the previous node through the neuromorphic photon feedforward matrix and outputs the photon signal features of the current node. In the feedback feature vector transmission path, each node receives the photon signal features fed back by the next node through the neuromorphic photon feedforward matrix, and generates the photon signal features after feedback enhancement. The dynamic memory unit update layer is used to fuse the photon signal features output from the feedforward path and the feedback path to generate the node feature state that is updated level by level within the dynamic memory unit. The multi-scale feature self-enhancement layer is used to decompose the node feature state in the dynamic memory unit into multiple scale sub-feature spaces, and perform self-enhancement reconstruction of the features step by step to obtain the scale-enhanced feature state. The feature state fusion output layer is used to fuse the final feature states of the feedforward path and the feedback path to output high-density photon signal features.

5. The method for predicting the volume of rock mass collapse after blasting based on point cloud deep learning according to claim 1, characterized in that, The process involves inputting the initial low-dimensional global feature vector into the first-level prediction network, and then performing stepwise autoregressive prediction processing using a neuromorphic photon feedforward matrix at each level of the prediction network to obtain the photon signal features at each level. The initial low-dimensional global feature vector is input into the first-level prediction network node, and the light intensity and phase are modulated through the neuromorphic photon feedforward matrix to obtain the photon input feature vector. The photon input feature vector is transmitted to the silicon-based optical waveguide structure in the neuromorphic photon feedforward matrix and subjected to weighted linear combination operation to obtain the first-level intermediate photon signal features. The first-level intermediate photon signal features are transmitted to the micro-ring resonator structure in the neuromorphic photon feedforward matrix for nonlinear light intensity modulation, and the photon signal features after the first-level autoregressive prediction are obtained. The photon signal features after the first-level autoregressive prediction are transmitted to the next prediction network node step by step. The light intensity and phase modulation, weighted linear combination operation and nonlinear light intensity modulation are repeated step by step to obtain the photon signal features after the autoregressive prediction of the current node. The dynamic memory unit is used to store the state of the photon signal features output by the nodes of the successive prediction network, forming the node feature state that is updated step by step. The node feature states that are updated level by level are divided into multiple scale sub-feature spaces. The scale attention weights are calculated, and the feature states of the scale sub-feature spaces are reconstructed by self-enhancing through the scale attention weights to obtain the photon signal features after autoregressive processing at each level.

6. The method for predicting the volume of rock mass collapse after blasting based on point cloud deep learning according to claim 1, characterized in that, The process involves inputting photon signal features at various levels into a neuromorphic photon feedforward matrix, and performing step-by-step prediction calculations through a coherent interference mechanism of the photon signals to obtain high-density photon signal features. Specifically: The photonic signal features after each stage of autoregressive processing are input into a silicon-based optical waveguide structure step by step, and the phase parameters of the photonic signal features are modulated by spatial wavelength division multiplexing to obtain the phase-modulated input photonic features. The input photon features are spatially coupled by a directional coupler array in a silicon-based optical waveguide structure to generate an intermediate photon signal after spatial coupling. The intermediate photon signal after spatial coupling is introduced into a multi-stage cascaded micro-ring resonant cavity structure, so that the intermediate photon signal undergoes coherent optical interference within the multi-stage cascaded micro-ring resonant cavity. By adjusting the optical path delay mechanism within the cavity of a multi-stage cascaded micro-ring resonant cavity structure, the optical path difference of the intermediate photon signal can be adjusted to achieve spatial coherent superposition or destructive interference, thereby generating interference photon signals step by step. The intensity of the progressively generated interference photon signals is detected by a photon intensity detector, and the spatial distribution characteristics of the intensity are fused to obtain high-density photon signal characteristics.

7. The method for predicting the volume of rock mass collapse after blasting based on point cloud deep learning according to claim 1, characterized in that, The process of decoding high-density photonic signal features into high-density point cloud data through photoelectric conversion specifically involves: The high-density photon signal characteristics are input into multiple photoelectric conversion units of the array photodetector, and photoelectric conversion is performed point by point; The light intensity voltage value is obtained by measuring the light intensity point by point using an array of photodetectors. The light intensity voltage values ​​are normalized point by point and transformed into feature intensity scalars corresponding to the spatial coordinates of the point cloud, generating high-density point cloud feature scalars with spatial location indexes. Based on the feature scalar of high-density point cloud, interpolation calculation is performed using the spatial coordinate information of preprocessed sparse point cloud data to supplement and reconstruct the spatial coordinate data of high-density electrical signal point cloud. Based on the spatial neighborhood feature consistency judgment rule, the consistency threshold between the feature scalar and the feature scalar of adjacent spatial locations is calculated point by point. Outliers with spatial feature differences exceeding the set threshold are removed to obtain filtered high-density electrical signal point cloud data. By fusing spatial coordinates and feature intensity scalars, the filtered high-density electrical signal point cloud data is transformed into high-density point cloud data that fuses spatial coordinates and detailed crack features.

8. The method for predicting the volume of rock mass collapse after blasting based on point cloud deep learning according to claim 1, characterized in that, The process of using a volume calculation model based on the Alpha Shape algorithm to calculate and process high-density point cloud data to obtain the volume value of rock mass collapse is as follows: High-density electrical signal point cloud data is input into a volume calculation model based on the Alpha Shape algorithm, and the correspondence between spatial location index and three-dimensional spatial coordinates is established point by point. Based on the spatial location index, the Euclidean distance between each spatial point and other spatial points is calculated with each spatial point as the center, generating a complete distance matrix between spatial points; The spatial scale parameter is gradually increased based on the distance matrix. Once the connection between spatial points is stable, the current spatial scale parameter is used as the final parameter value for the Alpha Shape algorithm. Based on the final determined spatial scale parameters, effective spatial neighborhood relationships are identified point by point in the high-density electrical signal point cloud data, and adjacent spatial points that meet the conditions are connected into a spatial closed geometric grid structure. For each effective triangular facet element of the spatial closed geometric mesh structure, the spatial coordinate origin is connected to the three vertices of the triangular facet element to form a spatial tetrahedral structure, and the element volume value of the spatial tetrahedral structure is calculated one by one. The volume values ​​of all spatial tetrahedral structures are accumulated one by one to obtain the rock mass collapse volume value corresponding to the high-density electrical signal point cloud data.

Images