High-rise facility operation risk monitoring method based on deep learning and point cloud detection

Abstract

The present application relates to the technical field of computer vision, and particularly relates to a high-rise facility operation risk monitoring method based on deep learning and point cloud detection, which collects three-dimensional point clouds in real time, extracts target point clouds through dynamic threshold denoising and template registration; uses a pre-trained space-time graph network combined with an attention mechanism to jointly encode spatial geometry and time sequence motion; identifies high reflectivity beacon points to calculate displacement and angular velocity change rate; constructs a gating fusion model to dynamically weight and couple semantic features and measurement data, generates a risk probability distribution through mutual information consistency testing; uses a membership optimized fuzzy logic classifier to map to a four-level early warning and trigger a graded response; extracts key frames after early warning to fine-tune the model through incremental learning, and reversely optimizes preprocessing parameters to form a closed loop. Through multi-source heterogeneous data fusion, dynamic adaptive weighting and self-evolution mechanism, the present application significantly improves the real-time performance, accuracy and robustness of risk monitoring in complex construction environments.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a method for monitoring operational risks of tall facilities based on deep learning and point cloud detection. Background Technology

[0002] In recent years, with the continuous improvement of national basic road construction, the focus of expressway construction has gradually shifted from building new roads to the reconstruction and expansion of existing roads. Expressway reconstruction and expansion projects face extremely challenging construction environments, requiring expansion work to be carried out while maintaining existing traffic flow. The terrain along the routes is complex, involving multiple tunnels, bridges, and high slopes, making construction organization difficult and posing high safety risks. When large, tall equipment is in operation, an accidental overturning would seriously threaten the safety of adjacent roads, posing a significant safety hazard.

[0003] Under the new transportation development model, there is a lack of technology for automatically monitoring the construction risks of tall structures during highway reconstruction and expansion projects. Most efforts still rely on on-site staff for monitoring and supervision. To address the pain points and difficulties in safety management during highway reconstruction and expansion, we will fully utilize the advantages of next-generation information technology and, through multidisciplinary integration, research technologies for monitoring and early warning of operational risks associated with tall structures during reconstruction and expansion projects, thereby comprehensively improving the safety level of highway reconstruction and expansion projects. Summary of the Invention

[0004] Therefore, the present invention provides a method for monitoring operational risks of tall facilities based on deep learning and point cloud detection, in order to solve the aforementioned problems existing in the prior art.

[0005] To achieve the above objectives, this invention provides a method for monitoring operational risks of tall facilities based on deep learning and point cloud detection, comprising: Step S1: Collect data of the construction scene in real time, and preprocess the acquired 3D point cloud sequence to extract the initial point cloud data of the target facility; Step S2: Input the initial point cloud data into a pre-trained machine learning model to perform joint feature encoding on the spatial structure and temporal changes of the target, and output primary risk feature information; Step S3: Identify the marked feature points arranged at key nodes of the target facility in the point cloud sequence, and calculate the rate of change of their temporal motion parameters; Step S4: Construct a fusion decision model by dynamically weighting and cross-validating the primary risk feature information with the rate of change of motion parameters of the marked feature points to generate a risk probability distribution. Step S5: Based on the risk probability distribution, make a judgment to obtain graded early warning information and trigger an early warning response; Step S6: Use the scene data after the warning is triggered as incremental learning samples to update the machine learning model, and optimize the preprocessing parameters in reverse to form a closed loop for continuous evolution of detection accuracy.

[0006] Furthermore, the process of step S2 includes: A dynamic graph structure is constructed based on the spatial neighborhood relationships of each point in the initial point cloud data, and local geometric features are extracted through graph convolution operations. Recurrent neural network encoding is used to encode the temporal motion patterns of consecutive multi-frame point cloud sequences to generate global motion features; The attention mechanism is used to adaptively allocate and fuse the weights of the local geometric features and global motion features, and outputs primary risk feature information including the semantic category of the risk target, its three-dimensional spatial location, and its motion vector field.

[0007] Furthermore, the process of adaptively assigning and fusing weights to the local geometric features and global motion features using an attention mechanism to output primary risk feature information containing the semantic category of the risk target, its three-dimensional spatial location, and its motion vector field includes: Asymmetric attention structure is constructed by using local geometric features as query vectors and global motion features as key and value vectors. The query vector and key vector are nonlinearly transformed using a learnable weight matrix, and the relevance score between them is calculated. The relevance scores are converted into attention weight coefficients using a normalization function; The value vector is weighted and aggregated using the attention weight coefficients to obtain the fused spatiotemporal feature representation. The spatiotemporal feature representation is decoded through a multilayer perceptron network to obtain the primary risk feature information.

[0008] Furthermore, the process of step S3 includes: Deploy high-reflectivity beacon points at key nodes of the target facility; Candidate beacon points are extracted based on point cloud intensity features, and beacon points with intensity values ​​higher than the background preset multiple are selected by clustering. Cross-frame association matching and unique identification are performed on candidate beacon points in point clouds of multiple consecutive frames to obtain cross-frame association results; The displacement rate of change, angular velocity rate of change, and vibration frequency rate of change of each beacon point are calculated based on the cross-frame correlation results.

[0009] Furthermore, the process of performing cross-frame association matching and unique identification on candidate beacon points in consecutive multi-frame point clouds to obtain cross-frame association results includes: Based on the position and motion state of the beacon point in the previous frame, a prediction-correction mechanism is used to calculate its theoretical position in the current frame. A search neighborhood is established centered on the theoretical location, and candidate beacon points are extracted within the search neighborhood; Calculate the Mahalanobis distance between each candidate beacon point and the theoretical position, and determine the candidate beacon point that meets the preset distance threshold and has the smallest distance as the matching beacon point of the current frame; A unique identifier is assigned to each matching beacon point to maintain its identity consistency across frames; meanwhile, beacon points that fail to match successfully in multiple consecutive frames are marked as lost and a recapture mechanism is triggered.

[0010] Furthermore, the process of step S4 includes: A dual-branch network structure is constructed. The first branch inputs the primary risk feature information and performs high-dimensional semantic encoding, while the second branch inputs the rate of change of motion parameters of the marked feature points and performs temporal signal encoding. The two branches independently process heterogeneous features. A learnable gating mechanism is used to calculate the dynamic weight coefficients of the first and second branches, and the outputs of the two branches are weighted and fused through a weight matrix to generate a fused feature representation. Based on the mutual information consistency test, the correlation analysis between the primary risk feature information and the rate of change of the motion parameters is performed. When the discrepancy between the two prediction results exceeds a preset threshold, a credibility reassessment is triggered to calibrate the dynamic weight coefficient. The fused feature representation is input to a probability regression network for decoding to generate a risk probability distribution function.

[0011] Furthermore, the process of calculating the dynamic weight coefficients of the first and second branches using a learnable gating mechanism, and then weighting and fusing the outputs of the two branches using a weight matrix to generate a fused feature representation includes: The feature importance scores of the first branch and the second branch are calculated respectively using a learnable gating weight matrix; Based on the feature importance score, dynamic weight coefficients are generated through a nonlinear activation function, so that the weights of the two branches are adaptively adjusted according to the characteristics of the input data. The dynamic weight coefficients are multiplied by the corresponding branch outputs and then aggregated into a fused feature representation through a residual connection structure.

[0012] Furthermore, the process of step S5 includes: A fuzzy logic classifier is used to map the risk probability distribution to a preset set of risk levels, and the confidence level of each risk level is calculated using a membership function. Based on the comparison between the confidence level of each risk level and a preset threshold, when the confidence level of a certain level exceeds its threshold, a warning message of the corresponding level is generated. The warning message includes the risk level, warning type and confidence level value. The corresponding warning instruction is triggered based on the level of the warning information.

[0013] Furthermore, the process of mapping the risk probability distribution to a preset set of risk levels using a fuzzy logic classifier, and calculating the confidence level of each risk level through a membership function, includes: Membership functions are constructed based on historical risk event data, and these membership functions adopt triangular, trapezoidal, or Gaussian function forms. A fuzzy rule base is established based on the membership function, and the risk probability value is mapped to each risk level. The number of fuzzy rules and the threshold are dynamically adjusted according to the type of construction scenario. The fuzzy inference results are converted into confidence scores for each risk level through defuzzification operations. The defuzzification method adopts either the weighted average method or the maximum membership principle.

[0014] Furthermore, step S6 includes the following process: Keyframe data containing risk events are extracted from the scene data after the warning is triggered, and samples are screened and labeled based on risk level and classification confidence. The filtered keyframe data is used as incremental samples, and the parameter weights of the machine learning model are fine-tuned and updated using an online learning algorithm. Based on the prediction error distribution of the updated model, the dynamic threshold parameters of the preprocessing stage are optimized in reverse to make the preprocessing results more suitable for the noise characteristics and target distribution of the current scene. By comparing the early warning accuracy of the model before and after optimization on the validation set, we can confirm whether the detection accuracy has been effectively improved, and feed the optimization results back to the preprocessing stage of the next cycle.

[0015] Compared with existing technologies, the advantages of this invention are as follows: By using spatiotemporal synchronous acquisition and dynamic preprocessing of multi-source sensors, the invention ensures the integrity and accuracy of point cloud data, providing high-quality input for subsequent analysis; the pre-trained point cloud spatiotemporal graph network adaptively fuses local geometric structures and global temporal motion features using an attention mechanism, enabling the model to focus on key risk areas and capture early signs of minute displacements; the precise physical measurement data provided by reflection beacon points and the abstract semantic features output by deep learning are dynamically weighted and coupled in a dual-branch gated fusion model. The gating mechanism adjusts the contribution of the two types of features in real time according to data credibility, while the mutual information consistency test effectively identifies and suppresses contradictory predictions. This improves decision-making reliability. The fuzzy logic classifier maps continuous risk probabilities to a tiered early warning set using a membership function. This membership function is adaptively optimized based on historical data and exhibits smooth transitions between levels, avoiding early warning jumps caused by hard thresholds. Combined with a multi-threshold judgment strategy, it ensures the sensitivity and accuracy of triggering different risk levels. After an early warning is triggered, the system automatically extracts keyframes for incremental learning. It consolidates the algorithm with elastic weights that protect important parameters, fine-tuning the model online. Simultaneously, it optimizes preprocessing thresholds and matching parameters in reverse, making denoising and target extraction more closely match the noise distribution and target characteristics of the current scene. After closed-loop verification, the optimization results are fed back to the next acquisition cycle, forming a self-evolving reinforcement loop. The entire method, through a progressive logical chain of hardware synchronization, feature encoding, gating fusion, fuzzy inference, and incremental learning, achieves a fully adaptive closed loop from data acquisition, risk perception, decision generation to system evolution, significantly improving the real-time performance, accuracy, and long-term robustness of risk monitoring in complex construction environments. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the operational risk monitoring method for tall structures based on deep learning and point cloud detection provided by this invention; Figure 2 This is a flowchart illustrating step S2 in the method for monitoring operational risks of tall facilities based on deep learning and point cloud detection provided by the present invention. Figure 3 This is a flowchart illustrating step S3 in the method for monitoring operational risks of tall facilities based on deep learning and point cloud detection provided by the present invention. Figure 4 This is a flowchart illustrating step S4 in the method for monitoring operational risks of tall facilities based on deep learning and point cloud detection provided by the present invention. Detailed Implementation

[0017] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0018] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0019] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0020] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0021] Please see Figure 1 As shown, this invention provides a method for monitoring operational risks of tall facilities based on deep learning and point cloud detection, comprising: Step S1: Collect data of the construction scene in real time, and preprocess the acquired 3D point cloud sequence to extract the initial point cloud data of the target facility; Specifically, a radar-camera measurement system is deployed, consisting of a LiDAR and a global shutter camera. The LiDAR is a solid-state LiDAR with a wavelength of 905nm, a frame rate ≥10Hz, and a ranging accuracy of ±2cm. The camera is an industrial global shutter camera with a resolution ≥1920×1080, and its frame rate is synchronized with the LiDAR. An FPGA chip generates a microsecond-level synchronous square wave signal to simultaneously trigger both the LiDAR scan and the camera exposure, ensuring a trigger time deviation of ≤0.2ms between the two sensors, achieving pixel-level spatiotemporal alignment. The camera lens is coated with an anti-reflection film in the 850-1100nm band, which has a transmittance >95%, effectively suppressing visible light interference and enhancing the LiDAR's reflected signal reception efficiency. The acquired 3D point cloud sequence and image data are synchronously transmitted to the embedded processor via Gigabit Ethernet, with a data latency <50ms, ensuring real-time processing. For any point p in the point cloud sequence, the KD-Tree algorithm is used to search for all points within its spherical neighborhood of radius r (r ranges from 0.1 to 0.3m), and the number of neighborhood points is counted as the local point density ρ(p). Calculate the mean μ and standard deviation σ of the local point density of all points in the current frame. The dynamic threshold function is: The adjustment coefficient k is dynamically adjusted according to the scene complexity: k=1.5 for static scenes, and k∈[1.8,2.0] for high dust or vibration scenes. If the local point density ρ(p) < T(p) of point p, it is identified as an outlier and removed. The noise rate of the processed point cloud is reduced to below 5%. A standard template point cloud Pt for tall facilities (such as drilling rigs, tower cranes, etc.) is pre-established, and the template includes the geometric structure and dimensional parameters of key components. The FPFH (Fast Point Feature Histogram) feature descriptor is used to perform coarse matching between the scene point cloud Ps and the template point cloud Pt to obtain the initial value of the initial transformation matrix. The optimal transformation matrix is ​​solved by the SVD algorithm. Iterative calculations minimize the registration error. The registration convergence condition is that the difference in error between two adjacent iterations is <0.001m or the maximum number of iterations (30) is reached. The transformed template point cloud T·Pt and the scene point cloud Ps are then subjected to nearest neighbor search, and the matching score formula is... The distance threshold σ is set to 0.05m, and N is the total number of template point clouds. When Score ≥ 0.75, the matching is considered successful, and all points in the scene with a distance < σ from the template point cloud are extracted as the initial point cloud data of the target facility; if Score < 0.75, the template library is updated or the registration step is re-executed.

[0022] Step S2: Input the initial point cloud data into a pre-trained machine learning model to perform joint feature encoding on the spatial structure and temporal changes of the target, and output primary risk feature information; Specifically, step S2 includes the following process: A dynamic graph structure is constructed based on the spatial neighborhood relationships of each point in the initial point cloud data, and local geometric features are extracted through graph convolution operations. Specifically, for the initial point cloud data For each point pi, a dynamic adjacency matrix A is constructed using the K-nearest neighbor algorithm (K=20-30). The neighborhood relationships are adaptively adjusted according to the point cloud density: K=20 when the point density ρ>500 points / m³, and K=30 when ρ<200 points / m³. A geometric feature vector is calculated for each edge (pi, pj). ,in , The corresponding point normal vector is calculated in the local neighborhood using the PCA method. Feature aggregation is performed using the EdgeConv structure, and the graph convolution operation in the l-th layer is as follows: Where: MLP is a three-layer perceptron (channel count 64→128→256), max is element-wise maximum pooling, and N(i) represents the set of neighborhood points of point pi. Let represent the feature vector of point pi in the l-th layer. Stack three layers of EdgeConv to extract multi-scale local geometric features, and output a feature map H with dimension N×256.

[0023] Recurrent neural network encoding is used to encode the temporal motion patterns of consecutive multi-frame point cloud sequences to generate global motion features; Specifically, a continuous T=8 frame point cloud sequence {P_t, P_{t-1}, ..., P_{t-7}} is taken, and frame-by-frame local feature extraction {H_t, H_{t-1}, ..., H_{t-7}} is performed on each frame. A dual-constraint matching algorithm based on spatial distance and feature similarity is used to perform point-level association between adjacent frames, establish cross-frame point correspondences, and generate a trajectory matrix. The 8-frame trajectory sequence of each point is input into a bidirectional LSTM network (128 hidden layer dimensions). The forward LSTM models the historical motion trend, and the backward LSTM models the future motion dependency. The bidirectional outputs are concatenated to obtain a global motion feature vector M with a dimension of N×256.

[0024] The attention mechanism is used to adaptively allocate and fuse the weights of the local geometric features and global motion features, and outputs primary risk feature information including the semantic category of the risk target, its three-dimensional spatial location, and its motion vector field.

[0025] Specifically, the process of adaptively assigning and fusing weights to the local geometric features and global motion features using an attention mechanism to output primary risk feature information containing the semantic category of the risk target, its three-dimensional spatial location, and its motion vector field includes: Asymmetric attention structure is constructed by using local geometric features as query vectors and global motion features as key and value vectors. Specifically, local geometric features As the query vector, it is mapped by the first layer MLP as ; global motion features As key vectors and value vectors, they are mapped to the following values ​​via the second and third layers of the MLP: and .in This forms an asymmetric attention structure. Three-dimensional spatial position encoding is added to Q, K, and V. ,in pos represents the coordinates of a point, enhancing the model's geometric perception of spatial structures.

[0026] The query vector and key vector are nonlinearly transformed using a learnable weight matrix, and the relevance score between them is calculated. Specifically, the query vector and key vector are linked by a learnable weight matrix W_Q. Perform a linear transformation and calculate the unnormalized correlation score: ,in, As the bias matrix, the tanh activation function introduces nonlinearity to capture complex feature interactions. This represents the result of transforming the query vector using the weight matrix; This represents the result of transforming the key vector using the weight matrix; This represents the transpose matrix of the key vectors. The computational complexity is O(N·d_k·d_k).

[0027] The relevance scores are converted into attention weight coefficients using a normalization function; Specifically, a sparse attention mechanism is used to perform Top-K filtering (K=50) on the relevance scores S, and then Softmax normalization is applied to generate the attention weight matrix. : , in This represents the attention weight of the i-th geometric feature point to the j-th motion feature point.

[0028] The value vector is weighted and aggregated using the attention weight coefficients to obtain the fused spatiotemporal feature representation. Specifically, the value vector V is weighted and aggregated using the attention weight matrix A to obtain an enhanced spatiotemporal feature representation: Fenhanced = A⋅V+H. The residual connection H preserves the original geometric features and prevents gradient vanishing. The aggregated feature dimension is N×256. An 8-head attention mechanism is used for parallel computation, with each head outputting an N×32 dimension. After concatenation, the output is projected back to N×256 through a linear layer to capture risk patterns in different subspaces.

[0029] The spatiotemporal feature representation is decoded through a multilayer perceptron network to obtain the primary risk feature information.

[0030] Specifically, the fused spatiotemporal features will be represented Input three-layer perceptron network decoding:

[0031] The MLP_out structure consists of a linear layer (256→512) → ReLU → Dropout (0.3) → linear layer (512→256) → ReLU → task-specific output layer.

[0032] Output risk category probability k=6 categories (normal, drill pipe tilt, hull displacement, crossbeam vibration, overturning precursor, collision risk); output bounding box parameters (x,y,z,w,h,l)∈ The VoteNet architecture is used to generate candidate bounding boxes through keypoint voting; the output is 6-DOF motion parameters (v_x,v_y,v_z,ω_x,ω_y,ω_z)∈ Acceleration and angular acceleration are calculated using time-series difference.

[0033] Step S3: Identify the marked feature points arranged at key nodes of the target facility in the point cloud sequence, and calculate the rate of change of their temporal motion parameters; Specifically, step S3 includes the following process: Deploy high-reflectivity beacon points at key nodes of the target facility; Specifically, two beacon points (P1, P2) are arranged at the top of the drill pipe to monitor verticality; two beacon points (P3, P4) are arranged at the hinge point of the machine body to monitor base displacement; and one beacon point (P5) is arranged at the stress concentration point of the crossbeam to monitor structural vibration. The spacing between the beacon points is ≥1.5m to avoid point cloud adhesion.

[0034] Candidate beacon points are extracted based on point cloud intensity features, and beacon points with intensity values ​​higher than the background preset multiple are selected by clustering. Specifically, histogram statistics are performed on the intensity attribute I(p) of each point in the point cloud sequence to calculate the mean background intensity. with standard deviation Dynamic intensity threshold Extract candidate point set DBSCAN clustering is performed on the candidate point set C, with parameters set to ε=0.15m (neighborhood radius) and MinPts=10 (minimum number of points). After removing noisy points, the cluster centers are retained as the initial beacon point extraction results. The geometric center coordinates and radius of each cluster are calculated, and outlier clusters with radii >0.2m or <0.03m are filtered out, retaining the candidate beacon point set C' that meets the beacon point size constraints.

[0035] Cross-frame association matching and unique identification are performed on candidate beacon points in point clouds of multiple consecutive frames to obtain cross-frame association results; Specifically, the process of performing cross-frame association matching and unique identification on candidate beacon points in consecutive multi-frame point clouds to obtain cross-frame association results includes: Based on the position and motion state of the beacon point in the previous frame, a prediction-correction mechanism is used to calculate its theoretical position in the current frame. Specifically, a state vector is established for each beacon point. The theoretical position of the current frame is predicted using a constant velocity model: , Wherein, the state transition matrix Δt is the frame interval (0.1s), and I is the identity matrix. Prediction covariance matrix. The process noise covariance Q = diag(0.01, 0.01, 0.01, 0.1, 0.1, 0.1).

[0036] A search neighborhood is established centered on the theoretical location, and candidate beacon points are extracted within the search neighborhood; Specifically, in terms of theoretical position Establish a spherical search neighborhood centered on a given point, with a radius of... ,in This is used to predict the standard deviation of the location.

[0037] Calculate the Mahalanobis distance between each candidate beacon point and the theoretical position, and determine the candidate beacon point that meets the preset distance threshold and has the smallest distance as the matching beacon point of the current frame; Specifically, for each candidate beacon point p_c within the search neighborhood, calculate its Mahalanobis distance to the theoretical location:

[0038] in, Let the candidate point coordinate vector be... To predict the position vector, To innovate the covariance matrix, (H=[I_{3×3}, 0_{3×3}] is the observation matrix, and R=diag(0.02,0.02,0.02) is the observation noise).

[0039] Select the option that satisfies D_M(p_c) < =7.81 (a chi-square threshold with 95% confidence) and the candidate point with the smallest D_M is selected as the matching beacon point for the current frame.

[0040] A unique identifier is assigned to each matching beacon point to maintain its identity consistency across frames; meanwhile, beacon points that fail to match successfully in multiple consecutive frames are marked as lost and a recapture mechanism is triggered.

[0041] Specifically, a unique ID is assigned to each matching beacon point, and this ID remains unchanged across frames. If an ID fails to match for three consecutive frames, it is marked as lost, and a recapture mechanism is initiated. The search radius is increased by two times, and the intensity threshold is reduced to [a lower threshold]. The system continuously monitors for 5 frames. If a point that meets the Mahalanobis distance constraint is detected again, its original ID is restored.

[0042] The displacement rate of change, angular velocity rate of change, and vibration frequency rate of change of each beacon point are calculated based on the cross-frame correlation results.

[0043] Specifically, for the successfully matched beacon point trajectory {(x_t, y_t, z_t)}, calculate the instantaneous velocity vector: Rate of change of acceleration: The vector formed by two points P1 and P2 at the top of the drill pipe Calculate the angular velocity: Angular acceleration rate of change: An FFT analysis was performed on the z-coordinate sequence of the beam beacon point P5, with a window length of 32 frames (3.2 seconds), to calculate the dominant frequency. With amplitude Monitor the frequency drift rate Δf / Δt and the amplitude growth rate ΔA / Δt.

[0044] Step S4: Construct a fusion decision model by dynamically weighting and cross-validating the primary risk feature information with the rate of change of motion parameters of the marked feature points to generate a risk probability distribution. Specifically, step S4 includes the following process: A dual-branch network structure is constructed. The first branch inputs the primary risk feature information and performs high-dimensional semantic encoding, while the second branch inputs the rate of change of motion parameters of the marked feature points and performs temporal signal encoding. The two branches independently process heterogeneous features. Specifically, inputting primary risk characteristic information Where C=256 is the feature dimension, including semantic category probability, 3D bounding box parameters, and motion vector field. This branch uses a three-layer MLP network (channel count 256→512→256) for high-dimensional semantic encoding, with each layer followed by LayerNorm and ReLU activation, outputting a semantic feature encoding vector. It is used to capture abstract risk patterns in deep learning models.

[0045] Input the sequence of motion parameter change rates of the marked feature points Where T=5 is the temporal window length (corresponding to 0.5 seconds), and K=9 is the dimension of motion parameters (3 rates of change of displacement, 3 rates of change of angular velocity, and 3 rates of change of vibration frequency). This branch uses a one-dimensional convolutional network (kernel size 3, number of channels 9→32→64→128) superimposed with a bidirectional GRU network (128 hidden layers) to encode the temporal signal and output a motion feature encoding vector. It is used to capture the temporal evolution of the physical parameters of beacon points.

[0046] A learnable gating mechanism is used to calculate the dynamic weight coefficients of the first and second branches, and the outputs of the two branches are weighted and fused through a weight matrix to generate a fused feature representation. Specifically, the process of using a learnable gating mechanism to calculate the dynamic weight coefficients of the first and second branches, and then weighting and fusing the outputs of the two branches using a weight matrix to generate a fused feature representation includes: The feature importance scores of the first branch and the second branch are calculated respectively using a learnable gating weight matrix; Specifically, two sets of learnable gating weight matrices W_g1, W_g2∈R^(256×64) and biases b_g1, b_g2∈R^(64) are designed and applied to the outputs of the two branches respectively:

[0047] in, This is the compressed feature importance score vector.

[0048] Based on the feature importance score, dynamic weight coefficients are generated through a nonlinear activation function, so that the weights of the two branches are adaptively adjusted according to the characteristics of the input data. Specifically, the dynamic weight coefficients α∈[0,1] and β∈[0,1] are calculated using a nonlinear activation function, satisfying α+β=1: ,in This is the gated control vector. When the confidence level of deep learning is low (e.g., due to occlusion or blurring), α automatically decreases to enhance the weights of the physical parameters.

[0049] The dynamic weight coefficients are multiplied by the corresponding branch outputs and then aggregated into a fused feature representation through a residual connection structure.

[0050] Specifically, the outputs of the two branches are dynamically weighted: To enhance gradient flow, a residual structure is introduced to generate the final fused feature representation:

[0051] The FC layer is a fully connected layer that projects the spliced ​​features to 256 dimensions, and LayerNorm stabilizes the training process.

[0052] Based on the mutual information consistency test, the correlation analysis between the primary risk feature information and the rate of change of the motion parameters is performed. When the discrepancy between the two prediction results exceeds a preset threshold, a credibility reassessment is triggered to calibrate the dynamic weight coefficient. Specifically, it is based on information theory to compute primary risk characteristic information. With the rate of change of motion parameters mutual information Estimate the statistical dependence between the two:

[0053] The joint probability distribution p(f_i,m_j) is calculated using kernel density estimation (KDE), with a bandwidth parameter h=0.1. A prediction discrepancy is identified when the mutual information value MI < the threshold θ_MI (θ_MI=0.5 bits) or the risk levels of the two branches are inconsistent. In this case, a credibility reassessment mechanism is triggered, resetting the dynamic weight coefficients α and β to a baseline value of 0.5 and initiating multi-frame smoothing filtering to reduce the impact of single-frame misjudgments.

[0054] The fused feature representation is input to a probability regression network for decoding to generate a risk probability distribution function.

[0055] Specifically, the feature representation will be fused. The input is a probabilistic regression network, which consists of two MLP layers (256→128→k), and the output is the probability distribution of k risk levels:

[0056] Where z_c is the logit value of category c, and k=4 corresponds to risk levels {safe, caution, warning, danger}. Simultaneously, the prediction entropy H(P) = -∑P(c)logP(c) is output as a measure of uncertainty in risk assessment; samples with H(P) > 0.8 are marked as low-confidence samples. The classification loss, location regression loss, and consistency loss are jointly optimized.

[0057] Among them, consistency loss To penalize cases with low mutual information, λ_1=1.0 and λ_2=0.5. Adversarial examples are introduced to enhance robustness by adding a small perturbation δ (‖δ‖2≤0.1) to the input features to maximize the loss function and improve the model's noise resistance.

[0058] Step S5: Based on the risk probability distribution, make a judgment to obtain graded early warning information and trigger an early warning response; Specifically, step S5 includes the following process: A fuzzy logic classifier is used to map the risk probability distribution to a preset set of risk levels, and the confidence level of each risk level is calculated using a membership function. Specifically, a four-level risk level set is preset: {Safe, Caution, Warning, Danger}, corresponding to risk probability intervals [0, 0.25], (0.25, 0.5], (0.5, 0.75], respectively. (0.75, 1.0). A triangular membership function is constructed for each risk level, with the center points located at 0.125, 0.375, 0.625, and 0.875, and the base width is 0.25. For example, the membership function for the attention level peaks at 1.0 when the risk probability is 0.375, and drops to 0 at 0.25 and 0.5; the warning level peaks at 0.625, and drops to 0 at 0.5 and 0.75. There is a 25% overlap between adjacent levels to avoid classification jumps caused by hard boundaries. The risk probability distribution is input into the membership function to calculate the confidence score for each level. For example, if the risk probability of a frame is 0.55, then its membership score for the attention level is 0.2, for the warning level is 0.8, and for the danger and safety levels it is 0. The final confidence score is normalized to form a quadruple [0.0, 0.2, 0.8, ...]. [0.0]. Based on the comparison between the confidence level of each risk level and a preset threshold, when the confidence level of a certain level exceeds its threshold, a warning message for the corresponding level is generated. The warning message includes the risk level, warning type, and confidence level value. Specifically, the process of mapping the risk probability distribution to a preset set of risk levels using a fuzzy logic classifier, and calculating the confidence level of each risk level through a membership function, includes: Membership functions are constructed based on historical risk event data, and these membership functions adopt triangular, trapezoidal, or Gaussian function forms. Specifically, 2,000 risk event records were selected from the construction monitoring data of the past 12 months. Each record includes the probability distribution value at the time of the risk occurrence, the final risk level, the event type (overturning, displacement, vibration, etc.), and environmental parameters (wind speed, temperature, dust concentration). Data were grouped into four construction scenarios: tunnel excavation, bridge hoisting, high slope operations, and road surface breaking, with no fewer than 400 records in each group. The membership function type was selected based on the transition characteristics of the risk level. For scenarios with clear risk level boundaries and strict transition requirements (such as bridge hoisting), a trigonometric function was used, with its parameters determined by the historical data mean. For scenarios with ambiguous transition zones (such as high slope operations), a trapezoidal function was used, with the flattened term covering the probability fluctuation range. For scenarios with a continuous and smooth risk probability distribution (such as road surface breaking), a Gaussian function was used, with its standard deviation determined by the data variance. Taking the warning level as an example, the center point a4 of the trigonometric function was set as the weighted average of the historical risk probability values ​​for that level, and the bottom endpoints a3 and a5 were taken as the 25th and 75th quartile values, respectively. The parameters are automatically updated every 30 days, with new data accounting for 20% of the weight and old data accounting for 80%, ensuring that the function adapts to seasonal construction changes.

[0059] A fuzzy rule base is established based on the membership function, and the risk probability value is mapped to each risk level. The number of fuzzy rules and the threshold are dynamically adjusted according to the type of construction scenario. Specifically, fuzzy rules adopt the "IF-THEN" format, with the premise being a risk probability value and the conclusion being a risk level. Each rule contains three elements: input variable (risk probability), fuzzy set (membership function), and output level. The basic rule base contains 16 core rules, covering four scenarios × four risk levels. Rule subsets are dynamically activated based on the current construction scenario type. For example, in the tunnel excavation scenario, which focuses on displacement and vibration risks, the weight coefficient for "medium risk probability and high displacement change rate" mapped to "warning level" is set to 0.8, while the weight coefficient for "medium risk probability and high vibration frequency" mapped to "warning level" is set to 0.6. Scenario switching is achieved by reading construction logs or manual input, with a switching latency of <100 milliseconds. The consistency between the past 100 frames of warning results and actual manual confirmation results is statistically analyzed. When the trigger accuracy of a rule is below 70%, the activation threshold of that rule is automatically increased by 0.05; when the accuracy is above 95% and the trigger frequency is too low, the threshold is decreased by 0.03, continuously optimizing the rule base.

[0060] The fuzzy inference results are converted into confidence scores for each risk level through defuzzification operations. The defuzzification method adopts either the weighted average method or the maximum membership principle.

[0061] Specifically, a weighted average is calculated for the membership values ​​of each risk level output by fuzzy inference. Let the membership values ​​of the four risk levels be... , , , Given the corresponding level center values ​​c1=0.125, c2=0.375, c3=0.625, and c4=0.875, the final output risk probability value P_risk is:

[0062] Substitute P_risk back into each membership function to calculate the normalized confidence scores for each level, ensuring the sum of confidence scores is 1.0. When the membership value of a certain risk level is significantly higher than other levels (difference > 0.4), directly set the confidence score of that level to 1.0, and set the remaining levels to 0, achieving a hard-decision output, suitable for emergency situations requiring a clear warning level. This mode is manually switched, and the weighted average method is used by default to maintain smoothness. The generated confidence score sequence is filtered using a 5-frame moving average to eliminate instantaneous jitter. If the confidence score of a certain level jumps from 0.2 to 0.8 and then back within 5 frames, the filtered output is 0.45, avoiding false alarms caused by single-frame anomalies.

[0063] The corresponding warning instruction is triggered based on the level of the warning information.

[0064] Specifically, establishing a mapping relationship between risk levels and early warning actions includes attention level, warning level, and danger level.

[0065] Attention level: Triggers a continuous yellow LED warning light, sends a status log to the monitoring center, and records point cloud data to a temporary buffer (retaining the most recent 10 seconds). Warning level: Triggers yellow LED flashing (frequency 2Hz), buzzer sounding at low frequency (interval 2 seconds), pushes warning information to the on-site safety officer's mobile phone, and starts recording image data (retaining the most recent 30 seconds); Danger level: Triggers high-frequency flashing of red LED (frequency 5Hz), continuous buzzer sound, broadcasts evacuation instructions to all construction personnel, automatically sends an emergency shutdown request to the supervision platform, and saves all point cloud, image, and log data to permanent storage.

[0066] Set a command trigger cooldown period so that the same type of warning at the same beacon point will not be triggered repeatedly within 5 seconds, thus avoiding frequent false alarms caused by mechanical vibration.

[0067] Step S6: Use the scene data after the warning is triggered as incremental learning samples to update the machine learning model, and optimize the preprocessing parameters in reverse to form a closed loop for continuous evolution of detection accuracy.

[0068] Specifically, step S6 includes the following process: Keyframe data containing risk events are extracted from the scene data after the warning is triggered, and samples are screened and labeled based on risk level and classification confidence. Specifically, after an alert is triggered, the system automatically extracts scene data from 5 seconds before to 10 seconds after the trigger time, forming a 15-second keyframe segment. This window covers the complete process of risk germination, development, and triggering, ensuring that the sample contains sufficient contextual information. The extracted keyframe data undergoes a three-layer screening process. The first layer filters by risk level, retaining only "Warning" and "Danger" level events and removing low-confidence "Attention" level samples. The second layer filters by classification confidence, retaining only samples with model prediction confidence between 0.6 and 0.9, avoiding overly easy (confidence > 0.9) or overly difficult (confidence < 0.6) samples that could interfere with the model. The third layer filters by consistency, removing anomalous samples with severe discrepancies between the two source data (MI value < 0.3) in the mutual information consistency test. The system automatically generates preliminary annotations, including 12 fields such as risk type, level, location, and motion parameters. High-risk events (hazard level) are automatically submitted to the cloud and reviewed and confirmed by safety officers within 24 hours. Upon confirmation, the event status changes to "verified." Warning-level events are reviewed in batches weekly. Unreviewed samples are marked "pending confirmation" and used only for temporary training, with their weight reduced by 50%. Trajectory similarity comparison is used. If the Euclidean distance between a new sample and a beacon point trajectory in the historical database is less than 0.1 meters and the time interval is within 1 hour, it is considered a duplicate event, and only the higher-quality one is retained. If the automatically labeled and manually labeled events are inconsistent, an arbitration mechanism is triggered: a senior engineer reviews the case and makes a final judgment based on on-site video recordings and point cloud playback.

[0069] The filtered keyframe data is used as incremental samples, and the parameter weights of the machine learning model are fine-tuned and updated using an online learning algorithm. Specifically, the Elastic Weight Consolidation (EWC) algorithm is used for online incremental learning. This algorithm protects important parameters during model fine-tuning, preventing catastrophic forgetting caused by new samples. Important parameters are calculated using the Fisher information matrix, preserving key knowledge from past tasks. An extremely low learning rate is used during the incremental learning phase, initially set to 1e-5, decreasing by 5% every 100 new samples to ensure smooth model updates. Compared to the offline training learning rate (1e-3), the incremental learning rate is reduced by a factor of 100 to prevent drastic parameter fluctuations. Every 16 selected new samples form a mini-batch, and a parameter update is performed. The update frequency is controlled at once every 10 minutes to avoid excessive computational load affecting real-time monitoring performance. To preserve the model's original generalization ability, a knowledge distillation loss is incorporated into incremental training. The pre-update model is used as the teacher model, and its output serves as a soft label, which, along with the hard labels of the new samples, guides training. The distillation temperature parameter is set to 2.0 to balance the retention ratio of old and new knowledge.

[0070] Based on the prediction error distribution of the updated model, the dynamic threshold parameters of the preprocessing stage are optimized in reverse to make the preprocessing results more suitable for the noise characteristics and target distribution of the current scene. Specifically, after processing every 500 frames of data, the statistical model predicts the error distribution. The average deviation μ_err and standard deviation σ_err between the predicted risk location and the actual beacon point location are calculated. If μ_err > 0.05 meters or σ_err > 0.03 meters, a systematic bias in the preprocessing stage is identified. If the error distribution indicates insufficient target extraction (false positive rate > 10%), the template matching score threshold in step S1 is lowered from 0.75 to 0.70 to expand the target extraction range. If severe noise interference is indicated (false positive rate > 15%), the k-coefficient for outlier removal is increased from 1.5 to 1.8 to enhance noise suppression. The step size is adjusted to 0.05, and evaluation is performed every 500 frames to avoid frequent jitter. The system has a built-in library of four scene parameter configurations (tunnel, bridge, high slope, road surface), automatically identifying the current scene by analyzing the point cloud height distribution and beacon point movement patterns. If the scenario is identified as a bridge hoisting operation, the system automatically switches to a bridge-specific parameter library (k=1.3, σ=0.04 meters) to achieve scene-specific adaptation of preprocessing parameters. New threshold parameters do not take effect immediately but are smoothly transitioned using a moving average method. New parameters account for only 30% of the weight, while historical parameters account for 70%, with the transition completed gradually over 50 frames (5 seconds) to avoid sudden parameter changes that could cause system output oscillations.

[0071] By comparing the early warning accuracy of the model before and after optimization on the validation set, we can confirm whether the detection accuracy has been effectively improved, and feed the optimization results back to the preprocessing stage of the next cycle.

[0072] In this embodiment, 200 frames are randomly selected monthly from the confirmed and labeled sample library to form the monthly validation set. The validation set covers various risk levels and scenario types, with 30% being hazardous, 40% being warning, and 30% being normal, ensuring comprehensive evaluation. Three core indicators are calculated: Accuracy rate: The percentage of warning levels that are completely consistent with the actual warning levels; Underreporting rate: The proportion of real dangerous events that do not trigger an early warning; False alarm rate: The percentage of warnings or alerts triggered under normal conditions.

[0073] If the accuracy of the graded model improves by more than 2% and the false negative rate decreases by more than 1% after incremental learning, the optimization is considered effective, and the new model replaces the old model. If the metric decreases or fluctuates by less than 1%, the model is rolled back to the previous version, and the learning rate is reduced by 50% for a retry. The optimization results are written to the system in the form of a parameter configuration file, and the new parameters are automatically loaded when the next preprocessing stage starts. An optimization log is also generated, recording information such as the adjustment time, metrics before and after the adjustment, and scenario type, for auditing and analysis by operations and maintenance personnel. The log is uploaded to the cloud, supporting parameter synchronization and experience sharing across multiple devices.

[0074] Specifically, this invention ensures the integrity and accuracy of point cloud data through spatiotemporal synchronous acquisition and dynamic preprocessing of multi-source sensors, providing high-quality input for subsequent analysis. The pre-trained point cloud spatiotemporal graph network adaptively fuses local geometric structures and global temporal motion features using an attention mechanism, enabling the model to focus on key risk areas and capture early signs of minute displacements. Precise physical measurement data provided by reflection beacon points and abstract semantic features output by deep learning are dynamically weighted and coupled in a dual-branch gating fusion model. The gating mechanism adjusts the contribution of the two types of features in real time based on data credibility, while the mutual information consistency test effectively identifies and suppresses contradictory predictions, thereby improving decision-making. Reliability: The fuzzy logic classifier maps continuous risk probabilities to a tiered early warning set through a membership function. This membership function is adaptively optimized based on historical data and exhibits smooth transitions between levels, avoiding early warning jumps caused by hard thresholds. Combined with a multi-threshold judgment strategy, it ensures the sensitivity and accuracy of triggering different risk levels. After an early warning is triggered, the system automatically extracts keyframes for incremental learning. Elastic weights that protect important parameters are used to consolidate the algorithm and fine-tune the model online. Simultaneously, the preprocessing threshold and matching parameters are optimized in reverse, making denoising and target extraction more consistent with the noise distribution and target characteristics of the current scene. After closed-loop verification, the optimization results are fed back to the next acquisition cycle, forming a self-evolving reinforcement loop. The entire method, through a progressive logical chain of hardware synchronization, feature encoding, gating fusion, fuzzy inference, and incremental learning, achieves a fully adaptive closed loop from data acquisition, risk perception, decision generation to system evolution, significantly improving the real-time performance, accuracy, and long-term robustness of risk monitoring in complex construction environments.

[0075] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0076] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for monitoring operational risks of tall facilities based on deep learning and point cloud detection, characterized in that, include: Step S1: Collect data of the construction scene in real time, and preprocess the acquired 3D point cloud sequence to extract the initial point cloud data of the target facility; Step S2: Input the initial point cloud data into a pre-trained machine learning model to perform joint feature encoding on the spatial structure and temporal changes of the target, and output primary risk feature information; Step S3: Identify the marked feature points arranged at key nodes of the target facility in the point cloud sequence and calculate the rate of change of their motion parameters; The process of step S3 includes: Deploy high-reflectivity beacon points at key nodes of the target facility; Candidate beacon points are extracted based on point cloud intensity features, and beacon points with intensity values ​​higher than the background preset multiple are selected by clustering. Cross-frame association matching and unique identification are performed on candidate beacon points in point clouds of multiple consecutive frames to obtain cross-frame association results; Calculate the displacement rate of change, angular velocity rate of change, and vibration frequency rate of change for each beacon point based on the cross-frame correlation results; Step S4: Construct a fusion decision model by dynamically weighting and cross-validating the primary risk feature information with the rate of change of motion parameters of the marked feature points to generate a risk probability distribution. The process of step S4 includes: A dual-branch network structure is constructed. The first branch inputs the primary risk feature information and performs high-dimensional semantic encoding, while the second branch inputs the rate of change of motion parameters of the marked feature points and performs temporal signal encoding. The two branches independently process heterogeneous features. A learnable gating mechanism is used to calculate the dynamic weight coefficients of the first and second branches, and the outputs of the two branches are weighted and fused through a weight matrix to generate a fused feature representation. Based on the mutual information consistency test, the correlation analysis between the primary risk feature information and the rate of change of the motion parameters is performed. When the discrepancy between the two prediction results exceeds a preset threshold, a credibility reassessment is triggered to calibrate the dynamic weight coefficient. The fused feature representation is input to a probability regression network for decoding to generate a risk probability distribution function; Step S5: Based on the risk probability distribution, make a judgment to obtain graded early warning information and trigger an early warning response; Step S6: Use the scene data after the warning is triggered as incremental learning samples to update the machine learning model, and optimize the preprocessing parameters in reverse to form a closed loop for continuous evolution of detection accuracy.

2. The method for monitoring operational risks of tall facilities based on deep learning and point cloud detection according to claim 1, characterized in that, The process of step S2 includes: A dynamic graph structure is constructed based on the spatial neighborhood relationships of each point in the initial point cloud data, and local geometric features are extracted through graph convolution operations. Recurrent neural network encoding is used to encode the temporal motion patterns of consecutive multi-frame point cloud sequences to generate global motion features; The attention mechanism is used to adaptively allocate and fuse the weights of the local geometric features and global motion features, and outputs primary risk feature information including the semantic category of the risk target, its three-dimensional spatial location, and its motion vector field.

3. The method for monitoring operational risks of tall facilities based on deep learning and point cloud detection according to claim 2, characterized in that, The process of adaptively assigning and fusing weights to the local geometric features and global motion features using an attention mechanism, and outputting primary risk feature information including the semantic category of the risk target, its three-dimensional spatial location, and its motion vector field, includes: Asymmetric attention structure is constructed by using local geometric features as query vectors and global motion features as key and value vectors. The query vector and key vector are nonlinearly transformed using a learnable weight matrix, and the relevance score between them is calculated. The relevance scores are converted into attention weight coefficients using a normalization function; The value vector is weighted and aggregated using the attention weight coefficients to obtain the fused spatiotemporal feature representation. The spatiotemporal feature representation is decoded through a multilayer perceptron network to obtain the primary risk feature information.

4. The method for monitoring operational risks of tall facilities based on deep learning and point cloud detection according to claim 3, characterized in that, The process of performing cross-frame association matching and unique identification on candidate beacon points in consecutive multi-frame point clouds to obtain cross-frame association results includes: Based on the position and motion state of the beacon point in the previous frame, a prediction-correction mechanism is used to calculate its theoretical position in the current frame. A search neighborhood is established centered on the theoretical location, and candidate beacon points are extracted within the search neighborhood; Calculate the Mahalanobis distance between each candidate beacon point and the theoretical position, and determine the candidate beacon point that meets the preset distance threshold and has the smallest distance as the matching beacon point of the current frame; A unique identifier is assigned to each matching beacon point to maintain its identity consistency across frames; meanwhile, beacon points that fail to match successfully in multiple consecutive frames are marked as lost and a recapture mechanism is triggered.

5. The method for monitoring operational risks of tall facilities based on deep learning and point cloud detection according to claim 4, characterized in that, The process of calculating the dynamic weight coefficients of the first and second branches using a learnable gating mechanism, and then weighting and fusing the outputs of the two branches using a weight matrix to generate a fused feature representation includes: The feature importance scores of the first branch and the second branch are calculated respectively using a learnable gating weight matrix; Based on the feature importance score, dynamic weight coefficients are generated through a nonlinear activation function, so that the weights of the two branches are adaptively adjusted according to the characteristics of the input data. The dynamic weight coefficients are multiplied by the corresponding branch outputs and then aggregated into a fused feature representation through a residual connection structure.

6. The method for monitoring operational risks of tall facilities based on deep learning and point cloud detection according to claim 5, characterized in that, The process of step S5 includes: A fuzzy logic classifier is used to map the risk probability distribution to a preset set of risk levels, and the confidence level of each risk level is calculated using a membership function. Based on the comparison between the confidence level of each risk level and a preset threshold, when the confidence level of a certain level exceeds its threshold, a warning message of the corresponding level is generated. The warning message includes the risk level, warning type and confidence level value. The corresponding warning instruction is triggered based on the level of the warning information.

7. The method for monitoring operational risks of tall facilities based on deep learning and point cloud detection according to claim 6, characterized in that, The process of mapping the risk probability distribution to a preset set of risk levels using a fuzzy logic classifier, and calculating the confidence level of each risk level through a membership function, includes: Membership functions are constructed based on historical risk event data, and these membership functions adopt triangular, trapezoidal, or Gaussian function forms. A fuzzy rule base is established based on the membership function, and the risk probability value is mapped to each risk level. The number of fuzzy rules and the threshold are dynamically adjusted according to the type of construction scenario. The fuzzy inference results are converted into confidence scores for each risk level through defuzzification operations. The defuzzification method adopts either the weighted average method or the maximum membership principle.

8. The method for monitoring operational risks of tall facilities based on deep learning and point cloud detection according to claim 7, characterized in that, The process of step S6 includes: Keyframe data containing risk events are extracted from the scene data after the warning is triggered, and samples are screened and labeled based on risk level and classification confidence. The filtered keyframe data is used as incremental samples, and the parameter weights of the machine learning model are fine-tuned and updated using an online learning algorithm. Based on the prediction error distribution of the updated model, the dynamic threshold parameters of the preprocessing stage are optimized in reverse to make the preprocessing results more suitable for the noise characteristics and target distribution of the current scene. By comparing the early warning accuracy of the model before and after optimization on the validation set, we can confirm whether the detection accuracy has been effectively improved, and feed the optimization results back to the preprocessing stage of the next cycle.

Images