Method and system for automatically detecting three-dimensional laser scanning datum point at top of elevator shaft based on deep learning and Bayesian fusion framework

Abstract

The invention belongs to the technical field of special equipment safety and intelligent detection, and more specifically relates to an elevator shaft top three-dimensional laser scanning reference point automatic detection method and system based on a deep learning and Bayesian fusion framework, and the method comprises the steps: firstly carrying out the semantic segmentation of a point cloud obtained through three-dimensional laser scanning through a deep learning network; and key areas such as hoistway roof and car roof equipment can be automatically identified. Then, candidate bases-punctuals are extracted in parallel, then a fusion module based on Bayesian inference is used for fusing a plurality of candidate points from different methods and uncertainty of the candidate points, and a posterior optimal reference point with smaller variance and higher precision is calculated in an analytic mode. And finally, the system automatically calculates each safety distance based on the optimal basis-punctuality, verifies the safety space and generates a detection report. According to the method, AI and geometric analysis are deeply fused under a probability framework, so that full-process automation of detection is realized, and objectivity, precision and repeatability of measurement are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of special equipment safety and intelligent detection technology, more specifically, relates to a kind of based on deep learning and bayesian fusion framework's elevator shaft top three-dimensional laser scanning reference point automatic detection method and system. BACKGROUND

[0002] As an indispensable vertical transportation tool in modern cities, the safety of elevator is of vital importance. Various national safety regulations, such as TSG T7001-2009 in China, have put forward strict requirements on the design, manufacture, installation and inspection of elevators. The regulations clearly stipulate that when the car stops at the highest maintenance position, a series of requirements on safety distance and safety space must be met, including the minimum part of the shaft top and the car top area ha, the minimum free vertical distance hb between the minimum part of the shaft top and the highest point of the equipment installed on the car top (such as guardrails, control boxes), the vertical distance hc between the highest part in the shaft (such as guide or suspension system) and the lowest point of the shaft top, and other key dimensions.

[0003] Currently, the traditional inspection method completely relies on the inspector to enter the high-risk shaft top for manual measurement, which has many disadvantages: first, the safety is poor, the inspector works in high altitude, and there is a risk of falling; second, the efficiency is low, the measurement is time-consuming and labor-intensive, and some positions are difficult to reach; third, the subjectivity is strong, the precision and repeatability are poor, and for the judgment of reference points such as "the lowest point of the shaft top" or "the highest point of the part", the selected positions by different inspectors often differ, especially on uneven concrete surfaces or complex-shaped equipment, the difference can reach several centimeters, which directly affects the final measurement results and qualification determination, and lacks scientific and objective basis.

[0004] With the development of three-dimensional laser scanning (3D Laser Scanning) technology, it provides a new possibility to solve the above problems. The inspector can place the scanner on the car top and start the scanning remotely from a safe position, and within a few minutes, a point cloud model containing millions or even hundreds of millions of three-dimensional coordinate points of the entire shaft top space can be obtained. This non-contact measurement method greatly improves the safety and efficiency of the operation.

[0005] However, when directly applying three-dimensional laser scanning technology to this inspection, a new core technical bottleneck emerges: the difficulty of accurately selecting points. The point cloud model is essentially a discrete sampling of the target surface geometry and inevitably contains measurement noise. When measuring the "lowest point of the shaft top" in software, the operator is faced with a seemingly flat but microscopically uneven "point cloud ceiling" composed of countless points. By randomly picking up a point on the screen with a mouse, it is almost impossible to ensure that the point is the true lowest point in the global or specified area. Similarly, for a complex mechanical part, the "highest point" of its point cloud model may just be an abnormally high point caused by noise or scanning angle. This point selection method based on human eye observation and manual clicking is essentially subjective, and its results are not scientifically repeatable and legally indisputable. The accuracy and reliability of the measurement results depend entirely on the operator's experience, responsibility, and "luck".

[0006] Therefore, there is an urgent need in the art for a new technical solution that can overcome the above difficulties. This solution should have the ability to accurately identify and locate key measurement reference points defined by the inspection regulations from a large amount of disordered noisy point cloud data in an intelligent, automatic, objective, and quantifiable manner, and on this basis, complete all size calculations and spatial determinations, thereby elevating the inspection of the elevator shaft top to a new, digital, and intelligent level. SUMMARY

[0007] The purpose of the present application is to overcome the shortcomings of the prior art and provide an elevator shaft top three-dimensional laser scanning reference point automatic detection method and system based on a deep learning and Bayesian fusion framework. This method combines deep learning, robust geometric calculation, and probabilistic Bayesian inference, and improves point cloud data processing from the traditional geometric measurement paradigm to a higher level of probabilistic reasoning based on uncertainty modeling and fusion, with the characteristics of full automation, high precision, and high reliability.

[0008] To achieve the above purpose, the technical solutions specifically adopted by the present application are as follows: An elevator shaft top three-dimensional laser scanning reference point automatic detection method based on a deep learning and Bayesian fusion framework, comprising the following steps: Step 1, data acquisition: place a three-dimensional laser scanning device on the elevator car top, and perform one or more scans of the elevator shaft top space when the elevator car runs to the highest maintenance position defined by the inspection regulations, to obtain original point cloud data containing three-dimensional coordinate information, reflectivity information, and color information of each point; Step 2, Data Preprocessing: The original point cloud data is stitched and registered, outlier points are removed, downsampled, and the coordinate system is aligned based on principal component analysis (PCA) to ensure that the Z-axis of the coordinate system is consistent with the extension direction of the main structure of the shaft (such as the guide rail). Step 3, Point Cloud Semantic Segmentation: Input the preprocessed point cloud data into a pre-trained point cloud semantic segmentation network, perform category prediction on each point in the point cloud, and segment it into multiple predefined semantic regions including at least "shaft top", "car top maximum standing area plane", "car top guardrail", "car top fixed equipment", and "highest component of the guidance or suspension system". Step 4: Extraction of multi-source candidate key points: For the measurement benchmarks required by the inspection regulations, candidate key points with uncertainty quantification are extracted by executing at least two of the following methods in parallel: (1) Geometric analysis method: For the semantic region point cloud that is divided into “shaft top” and “car top highest standing area plane”, the corresponding spatial plane equation is fitted by the Iterative Reweighted Least Squares (IRLS) algorithm or the Random Sampling Consensus (RANSAC) algorithm. The lowest or highest point in the convex hull of the semantic region point cloud is calculated analytically according to the equation as the first type of candidate key point. At the same time, its uncertainty is evaluated according to the fitting residual. (2) AI prediction method: For the point cloud of semantic regions with irregular shapes that are segmented into “car roof fixed equipment” and “highest component of guide or suspension system”, the point cloud is input into a pre-trained probabilistic key point prediction network (P-CPPN). The probabilistic key point prediction network (P-CPPN) directly regresses the expected value (mean vector) of the three-dimensional coordinates of the highest point of the component and its covariance matrix of prediction uncertainty, which are used as the second type of candidate key points. (3) Physical anchor point method: Before data acquisition, a high-precision target sticker with a unique code is pre-attached to the physical location to be measured (such as the estimated lowest point of the shaft top); during data processing, the three-dimensional coordinates of the center point of the target sticker are accurately located as the third type of candidate key point through image recognition and point cloud mapping algorithms, and uncertainty is assigned to it according to the positioning accuracy model of target recognition. Step 5: Key point fusion based on Bayesian inference: Establish a multi-source data fusion decision module, which will uniformly model each type of candidate key point and its uncertainty information extracted in Step 4 as a multidimensional Gaussian probability distribution; The multi-source data fusion decision module applies the Gaussian product rule, and by calculating the product of these multidimensional Gaussian probability distributions, obtains a posterior Gaussian distribution with smaller covariance, and uses the mean vector of this posterior distribution as the unique and high-precision final measurement benchmark point; Step 6, Dimension Calculation and Spatial Judgment: Based on the final measurement reference point, automatically calculate the key dimensions such as the free vertical distance ha, hb, and hc required by the inspection regulations, and compare them with the regulatory standard values ​​to determine whether they are qualified; and, within the free space point cloud defined by the semantic regions related to "shaft top" and "car top", execute the octree-based maximum inscribed space search (Octree-based MISS) algorithm to determine whether there exists a non-interference cuboid space of not less than 0.50m × 0.60m × 0.80m.

[0009] Furthermore, in step 3, the point cloud semantic segmentation network is trained using a weighted cross-entropy loss function. Its mathematical definition is: ; in, The total number of points in the training batch. The total number of semantic categories. It is a point Category The true label (one-hot encoded form). It is a network prediction point Category The probability, This is a category The assigned category weights, which are related to the category The frequency of occurrence in the training dataset is inversely proportional to the frequency of occurrence, which is used to alleviate the data class imbalance problem caused by the large difference in the number of points between large area categories such as shaft top and walls and small component categories such as guide shoes and rope ends.

[0010] Furthermore, in step S4, the Iterative Reweighted Least Squares (IRLS) algorithm solves the plane equations through the following iterative process. parameters : In the In this iteration, a weighted least squares problem is solved. ,in The number of points in a planar point cloud. It is a point To the parameter Algebraic distance of the defined plane, weight Based on the previous round of residuals Calculated using a robust weighting function (such as Tukey's biweight function), this function can significantly reduce the impact of outliers and noise points on the final plane fitting result.

[0011] Furthermore, in step S4, the loss function of the Probabilistic Keypoint Prediction Network (P-CPPN) It is designed to minimize the negative log-likelihood, and its specific form is: ; in, It is the input component point cloud. These are the three-dimensional coordinates of the actual key points. and These are the network's responses to input. The mean vector and covariance matrix of the predicted keypoint coordinates. The first term of the loss function is the square of the Mahalanobis distance between the predicted mean and the true value, used to penalize prediction errors; the second term is the logarithm of the determinant of the covariance matrix, which serves as a regularization term, encouraging the network to output a larger covariance when prediction is uncertain, while preventing the covariance from expanding indefinitely.

[0012] Furthermore, in step 5, the specific method for modeling candidate keypoints from different sources using Gaussian distribution is as follows: For candidate points extracted by geometric analysis Its covariance matrix is ​​constructed as a diagonal matrix. The values ​​of its diagonal elements are proportional to the square of the root mean square error (RMSE) of the plane fitted by the IRLS algorithm, i.e. ,in Scaling factor It is the identity matrix; For candidate points extracted by AI prediction, their mean Covariance Matrix Directly output from the P-CPPN network; For candidate points extracted by the physical anchor point method Its covariance matrix The positioning accuracy evaluation model is determined based on the target recognition algorithm. It is usually a diagonal matrix, and the diagonal elements reflect the positioning error variance on each coordinate axis.

[0013] Furthermore, the Gaussian product rule is used for fusion. Independent observations that follow a Gaussian distribution The posterior probability distribution is obtained by calculating their product. The posterior distribution after fusion remains a Gaussian distribution. Its mean vector Covariance Matrix Calculated using the following analytical formula: ; in, It is the first The accuracy matrix (information matrix) of each observation, and the final adopted coordinates of the measurement reference points are the posterior mean. .

[0014] Furthermore, the specific implementation of the octree-based maximum inline space search (Octree-based MISS) algorithm includes: a) Construct an octree data structure from the free space point cloud defined by the semantic region below “shaft top” and the semantic region above “car top”, where each leaf node represents a spatial voxel and is marked as “free” or “occupied”. b) Starting from the root node of the octree, recursively search the largest cuboid subspace in the tree that consists entirely of leaf nodes marked "free" in a depth-first or breadth-first manner. c) During the search process, an efficient pruning strategy is implemented using the hierarchical structure of an octree: when the size of the spatial cube represented by a node is smaller than any dimension of the target safe space of 0.50m×0.60m×0.80m, it is not necessary to explore all its descendant nodes, thereby significantly improving search efficiency.

[0015] The present invention also provides an automatic detection system for three-dimensional laser scanning reference points on the top of an elevator shaft based on a deep learning and Bayesian fusion framework. The system includes: at least one processor; a memory communicatively connected to the processor; and program instructions stored in the memory and executable on the processor, the program instructions being configured to perform the above-described method.

[0016] The present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps of the above-described method.

[0017] This invention has the following characteristics and beneficial effects: 1) High precision and objectivity: By employing a Bayesian fusion framework, candidate key points and their uncertainties extracted through geometric analysis, AI prediction, and physical anchoring methods are uniformly modeled as a multidimensional Gaussian probability distribution. The posterior distribution is then calculated using the Gaussian product rule, yielding a mathematically optimal and repeatable measurement benchmark. This method not only eliminates the subjectivity of manual point selection but also obtains the optimal solution with the minimum covariance through mathematical optimization, thus providing a quantified uncertainty assessment (posterior covariance matrix) for the final measurement benchmark. This characteristic provides a traceable confidence index for the reliability of the verification results, which is of significant value in forensic identification and authentication scenarios.

[0018] 2) Full-process automation and high efficiency: From loading the original point cloud to generating the inspection report, the entire process is automated. Semantic segmentation, key point extraction, fusion decision-making, size calculation, and spatial determination are all completed automatically by algorithms, reducing the original manual measurement and data processing time of several hours to just a few minutes, greatly improving inspection efficiency.

[0019] 3) Enhanced security: Inspectors only need to set up and retrieve the scanner on the car roof, eliminating the need to enter the top of the shaft for high-altitude operations, thus fundamentally eliminating safety risks such as falls and collisions. This design aligns with the "zero-contact" safety concept for special equipment inspection.

[0020] 4) Robustness and strong adaptive fault tolerance: The hybrid engine architecture design enables dynamic weight allocation through uncertainty quantification. When an algorithm module outputs abnormally (e.g., when the AI ​​model has a large prediction deviation for rare components), the increase in its covariance matrix will automatically reduce the decision weight of that module in Bayesian fusion, avoiding a catastrophic impact on the final result. This mechanism allows the system to maintain stable output even in complex wellbore environments (such as those with obstructions or non-standard components). Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall process of an automatic detection method for three-dimensional laser scanning reference points on the top of an elevator shaft based on a deep learning and Bayesian fusion framework, according to an embodiment of the present invention.

[0022] Figure 2 This is a hardware block diagram of an automatic detection system for three-dimensional laser scanning reference points on the top of an elevator shaft based on a deep learning and Bayesian fusion framework, according to an embodiment of the present invention.

[0023] Figure 3 This is a schematic diagram illustrating the pruning principle of the maximum inline space search algorithm based on an octree in an embodiment of the present invention. Detailed Implementation

[0024] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0025] like Figure 1 As shown, in a first aspect, to address the problem of subjective and low-precision selection of key measurement reference points on noisy point clouds in the prior art, this invention provides an automatic detection method for three-dimensional laser scanning reference points on the top of elevator shafts based on a deep learning and Bayesian fusion framework, comprising the following steps: S1: Data Acquisition A high-precision 3D laser scanner (such as a Faro Focus or Leica RTC360) is fixed in a stable position on the top of the elevator car. The car is moved to the highest maintenance position required by national standards, and the scanner is started remotely to perform a 360° panoramic scan of the space above the shaft. The scanning parameters are set to high density and high resolution to ensure sufficient detail is obtained. The scan simultaneously acquires the 3D coordinates (X, Y, Z), reflectance, and RGB color information of each point. If a single scan cannot cover the entire area, the car or scanner position is moved to perform multiple scans, and high-precision stitching and registration are performed using the equipment's built-in software or third-party software (such as Cyclone) by utilizing overlapping points in common areas or deploying spherical targets.

[0026] S2, Data Preprocessing The merged original point cloud is preprocessed, including: using statistical filtering or radius filtering algorithms to remove obvious outlier noise points; to speed up subsequent processing, voxel downsampling can be used to reduce the point cloud density while preserving features; the most critical step is coordinate system alignment. Principal component analysis (PCA) is performed on the point cloud subset of vertical, structurally significant components (such as guide rails) within the shaft to find the principal direction with the largest variance, and this direction is aligned with the Z-axis of the global coordinate system, thereby ensuring that all subsequent vertical distance calculations are performed in a unified and meaningful coordinate system.

[0027] S3, AI-based point cloud semantic segmentation The purpose of this step is to enable the computer to "understand" what each part of the point cloud represents. The preprocessed point cloud is then fed into a pre-trained deep learning network for point cloud semantic segmentation. This network can employ current state-of-the-art architectures such as PointNet++, RandLA-Net, or DGCNN, which are capable of effectively handling unordered, unstructured point cloud data. To address the issue of significant differences in the number of points across different categories in shaft scenarios (e.g., the number of points on the shaft top and walls may account for over 90% of the total, while the number of points on key small components such as guide shoes and rope heads is extremely low), this invention employs a weighted cross-entropy loss function during the network training phase. For a given point cloud containing… The training batch of points has a loss function Defined as: ; in, It is the total number of predefined semantic categories (such as "shaft top", "car top guardrail", etc.). It is a one-hot vector, if point The true category is If the value is 1, then its value is 1; otherwise, it is 0. These are the points output by the network through the final Softmax layer. Category The predicted probability. The key lies in the weights. Its calculation method is usually related to the reciprocal of the category frequency, for example... Alternatively, a more complex median-frequency balancing strategy can be employed. By assigning higher weights to rarer classes, the network pays more attention to the errors generated by these rarer class points during backpropagation parameter updates, thus preventing the model from being dominated by large classes and significantly improving the segmentation accuracy of key small components such as guide shoes and speed limiter cables. After segmentation, the original cluttered point cloud is assigned semantic labels, forming several logically independent subsets of the point cloud.

[0028] S4. Parallelized Multi-Source Candidate Keypoint Extraction For different measurement reference points, the system calls different algorithm modules in parallel to extract candidate points and their uncertainties based on their geometric characteristics.

[0029] I. For large-scale planes (such as the top of the shaft or the top of the car where people can stand), a robust geometric analysis method is adopted: For areas requiring planar geometry, such as the shaft top and car top standing areas, this invention employs the Iterative Reweighted Least Squares (IRLS) algorithm for highly robust planar fitting. Unlike RANSAC, which searches for interior points through random sampling, IRLS is a more refined deterministic optimization method. It aims to minimize a weighted sum of squared residuals: , in, These are points in a planar point cloud, and the equation of the plane is: .

[0030] The iterative process of IRLS is as follows: (a) Initialization: Set all weights The initial plane parameters are obtained using the standard least squares method; (b) Residual calculation: in the first... In each iteration, the residual from each point to the current fitted plane is calculated. .

[0031] (c) Weight update: Based on the residuals, use a robust weight function (such as Tukey's biweight function) to update the weights. To update the weights This function assigns very small or even zero weights to points with large residuals (outliers).

[0032] (d) Parameter update: Solve the weighted least squares problem with the new weights to obtain the new plane parameters.

[0033] (e) Iterative convergence: Repeat (b)-(d) until the plane parameters converge.

[0034] This method utilizes interior point information more fully than RANSAC, resulting in a more accurate planar model. After fitting, the system uses linear programming to find the theoretical minimum point of the plane within the two-dimensional convex hull (XY plane projection) of the semantic region point cloud; this point is the first type of candidate keypoint. Its uncertainty is quantified by the root mean square error (RMSE) of the fit.

[0035] 2. For components with complex shapes (such as car top equipment and guide shoes), a probabilistic AI prediction method is adopted. For components with irregular shapes that are difficult to describe using simple geometry, this invention trains an innovative probabilistic keypoint prediction network (P-CPPN). The core idea of ​​this network is not only to predict the location of keypoints but also to quantify its "confidence" in that prediction. Its network architecture is typically an encoder-decoder structure (such as a variant based on PointNet), with the input being a local point cloud of the component, and the output designed to predict parameters of a three-dimensional Gaussian distribution: the mean vector. (i.e., the predicted keypoint coordinates) and covariance matrix (Representing the uncertainty of the prediction). To enable the network to learn this "honest" prediction, its loss function is cleverly designed to minimize the negative log-likelihood: ; in, These are the coordinates of the actual keypoints in the training data. This loss function consists of two terms: the first is the predicted mean. truth value The square of the Mahalanobis distance between them. If the network is unsure about its prediction for a particular sample (e.g., the input point cloud is occluded or heavily noisy), it can "actively" increase the predicted covariance matrix. This will cause the inverse of the covariance matrix to The value decreases, thus reducing the penalty from the first error. However, the second term... As a regularization term, it penalizes excessive uncertainty (i.e., the determinant of the covariance matrix). (Too large). The interplay between these two factors allows the network, after training, to provide accurate predictions with low uncertainty for clear inputs, and relatively conservative predictions with high uncertainty for ambiguous inputs. This uncertainty caused by the inherent characteristics of the data is called Aleatoric uncertainty. Network output. of This represents the second type of candidate key points and their uncertainties.

[0036] III. Optional auxiliary methods: Physical anchor point method To further improve the accuracy of specific key points, or to provide strong priors when the confidence level of the AI ​​model is low, the physical anchor point method can be used. The specific operation is as follows: 1) Target deployment: Before scanning, attach a high-precision target sticker with a special pattern and code to the estimated critical area (such as the lowest point of the shaft top); 2) Coordinate Calculation: The target is accurately detected from the panoramic image using image recognition algorithms (such as template matching or deep learning-based detectors). Combined with camera parameters and point cloud data, the precise coordinates of the target's center point in three-dimensional space are calculated. ; 3) Uncertainty Assessment: Based on factors such as camera calibration error, sub-pixel positioning accuracy, and depth estimation error, the uncertainty is pre-assessed. This forms a very definite Gaussian distribution.

[0037] S5. Intelligent Fusion Decision Based on Bayesian Inference In step S4, we obtained multiple candidate points for keypoint locations using various methods, including geometric analysis, AI prediction, and physical anchors. To integrate information from these heterogeneous sources, we introduced a Bayesian inference framework to uniformly model all candidate points as probability distributions.

[0038] Specifically, each candidate point is treated as the mean of a Gaussian distribution, and its uncertainty is quantified by the corresponding covariance matrix. The fusion process is formalized as a multi-observation inference problem: that is, inferring the true location of keypoints based on multiple independent observations (outputs of each algorithm).

[0039] According to Bayesian statistical theory, if both the prior distribution and the likelihood function follow a Gaussian distribution, then the posterior distribution must also be a Gaussian distribution. Here, we treat the output of each algorithm as an independent "observation," with a probability distribution of... To find the best estimate that combines all observation information, we need to calculate the posterior probability. It is proportional to the product of the probabilities of all independent observations: .

[0040] Because the product of multiple Gaussian distributions has an analytical solution, this property is known as the Gaussian product rule. The posterior distribution after fusion remains a Gaussian distribution. The formula for calculating its parameters is: ; Here Known as the precision matrix (or information matrix), it represents the first... The amount of information contained in each observation. As can be seen from the formula, the final mean... It is the mean of all inputs. The weighted average is calculated, and the weight of each mean is exactly its corresponding precision matrix. This means that a small uncertainty (covariance) is achieved. Small, precision Larger (or less) observations will play a dominant role in determining the final result. For example, if the positioning accuracy of the physical anchor point is much higher than the prediction accuracy of the AI ​​model, then the final fusion point will be very close to the coordinates of the physical anchor point. Conversely, if the AI ​​model gives a prediction with very high confidence, while the geometric analysis results in high uncertainty due to surface unevenness, then the AI's prediction will dominate the outcome. The final calculated posterior mean This is the unique, objective, and highly accurate final measurement reference point adopted in this invention. Posterior covariance. This quantifies the remaining uncertainty of the final result.

[0041] S6. Automated Dimension Calculation and Space Determination After accurately determining all key benchmarks (such as) using Bayesian fusion methods After that (etc.), the system enters the automated dimension calculation and compliance judgment stage. The system directly uses the coordinates of these points to calculate the vertical distances such as ha, hb, and hc, and compares the results with the regulatory standard values, automatically giving a "qualified" or "unqualified" judgment.

[0042] For the three-dimensional safety space (size not less than 0.50m × 0.60m × 0.80m) required by regulations, the system uses an octree-based maximum inline space search algorithm for verification. The specific process is as follows: Spatial modeling: The system first voxels the free space defined below the point cloud of "shaft top" and above the point cloud of "car top", and constructs it as an octree data structure.

[0043] Efficient Search: The octree recursively divides the cubic space into eight sub-cubes, achieving hierarchical organization of the three-dimensional space. The system performs a depth-first search starting from the root node to find the largest inscribed space that can accommodate the target cuboid.

[0044] Intelligent pruning: During the search process, if the size of the cube represented by a node is smaller than the safe space size required by regulations, then the node and all its child nodes will be pruned directly without further exploration.

[0045] Thanks to the hierarchical structure and pruning strategy of the octree, the search efficiency of this invention is much higher than that of brute-force search on a dense three-dimensional voxel grid, and it can accurately determine whether there is a safe space that meets regulatory requirements in a short time.

[0046] like Figure 2 As shown, in a first aspect, this invention provides an automatic detection system for three-dimensional laser scanning reference points on the top of an elevator shaft based on a deep learning and Bayesian fusion framework. The system includes: at least one processor; a memory communicatively connected to the processor; and program instructions stored in the memory and executable on the processor, configured to perform the method described in the first aspect of this invention. The detection system acquires data using a three-dimensional laser scanner and enables user interaction and output of detection reports through input / output devices.

[0047] A third aspect of the present invention provides a complete automated detection process for the top space of an elevator shaft to further illustrate the present invention, the process comprising the following steps: Step 1: Data Acquisition and Preprocessing Inspectors placed a Faro Focus S70 3D laser scanner at the center of the elevator car top and remotely controlled it via tablet. The elevator car was driven to the position of the forced deceleration switch at the highest maintenance terminal (i.e., the measurement benchmark acquisition position defined by the inspection regulations), and the inspector started the scan. The scan parameters were set to 1 / 4 resolution and 3x quality, taking approximately 3 minutes to acquire raw point cloud data (.fls format) of approximately 80 million points and a panoramic HDR image.

[0048] After initial noise reduction using Faro SCENE software, the data is exported in the universal .e57 format. The detection system software of this invention is then launched, and the .e57 file is loaded. The software automatically executes the preprocessing procedure: Outlier Removal: Statistical outlier removal algorithms (such as distance-based anomaly detection) are used to filter out flight points that significantly deviate from the main point cloud distribution; Coordinate system alignment: Automatically identify the point clouds of the two main guide rails, calculate their first principal component vector through principal component analysis (PCA), and rotate the Z-axis of the global coordinate system to be parallel to this vector to ensure that the coordinate system is consistent with the extension direction of the main shaft structure (guide rail).

[0049] Step 2: Point Cloud Semantic Segmentation The preprocessed point cloud data is fed into a pre-trained semantic segmentation network based on the RandLA-Net architecture. During training, this network employs a weighted cross-entropy loss function optimized for elevator shaft scenarios. Supervised learning is conducted using a large amount of elevator shaft point cloud data (covering different shaft structures, equipment types, and environmental conditions) to ensure robust recognition of complex geometric features. After approximately 2 minutes of processing, the network outputs segmentation results. The point cloud data is automatically labeled with 12 categories, including: "shaft top," "wall," "guide rail," "car top," "car top guardrail," "car top inspection box," "counterweight," "speed governor," "wire rope," and "guide shoe." On the visualization interface, the shaft top is displayed in gray, the car top guardrail in blue, and the guide shoe in red, etc.

[0050] Step 3: Run the parallelized multi-source candidate key point extraction system according to the inspection regulations to extract the key benchmark points required for calculation, such as ha, hb, and hc. Specifically: 1) Determining the endpoints of ha: ; Lowest point of the shaft top ( The extraction process using geometric analysis is as follows: Data preparation: The system extracts all point cloud data marked "well top" after semantic segmentation; Plane Fitting: The Iterative Reweighted Least Squares (IRLS) algorithm is applied to fit the spatial plane. After 5 iterations, it converges to a stable solution. The plane equation is: .

[0051] Accuracy assessment: The root mean square error (RMSE) of the fitted plane and interior point data is 2.1 mm; Key point calculation: Based on the aforementioned plane equation, the system calculates the coordinates of the lowest point within the XY projection convex hull of the "shaft top" point cloud to obtain candidate points. rice.

[0052] Uncertainty modeling: The uncertainty of the geometric analysis method is quantified into a Gaussian distribution, and its covariance matrix is ​​Σ_geom = diag(2.1², 2.1², 2.1²) mm².

[0053] The highest point of the car roof where people can stand ( ): Similarly, the IRLS algorithm is used to perform plane fitting on the region marked "car roof" to obtain the plane equation and candidate points of the highest point. and its covariance .

[0054] 2) Determining the endpoints of hb: ; The highest point of the car roof fixing equipment ( ): AI Prediction Method: The system extracts all point cloud clusters labeled "car roof guardrail" and "car roof inspection box". The "car roof guardrail" point cloud is fed into a P-CPPN network. The network outputs the predicted mean of the highest point. Meters, and a non-diagonal covariance matrix. This indicates that the prediction errors are correlated in different directions. Similarly, predicting the point cloud of the "car roof inspection box" yields... and The system compares the Z values ​​of the two predicted points and selects the one with the larger Z value (along with its corresponding mean and covariance) as the candidate point for that location.

[0055] Step 4: Keypoint Fusion Based on Bayesian Inference Suppose that before scanning, the inspector affixed a high-precision target to the estimated lowest area on the top of the shaft. The system also obtained a candidate point during processing using the physical anchoring method. The covariance given by the positioning accuracy model is meters. .

[0056] Now, regarding the lowest point of the shaft top We have two candidate observations: A. Derived from geometric analysis ; B. From physical anchor points ; The multi-source data fusion decision module performs fusion calculations: 1) Calculate their respective precision matrices: and .

[0057] 2) Calculate the fused precision matrix: .

[0058] 3) Calculate the fused covariance matrix: .

[0059] 4) Calculate the mean after fusion: .

[0060] Because the target has higher positioning accuracy (smaller covariance, larger accuracy matrix), it will have a larger weight in the final weighted average. The calculated final lowest point at the top of the wellbore. Meters, and then the covariance of the test The diagonal element values ​​are all smaller than any of the previous sources, indicating that the fused result is more certain. Adopted as final Coordinates. Other key points were also ultimately determined in a similar (or single-source) manner.

[0061] Step 5: Size Calculation and Space Determination Step 5.1, Dimension Calculation: The system uses the finalized high-precision reference point coordinates for calculations: .

[0062] .

[0063] … (Calculations for other dimensions) Step 5.2, Spatial Determination: The system extracts all free-space point clouds between the lower surface of the "shaft top" and the upper surface of the "car top" that are not segmented into objects. An octree with a resolution of 2 cm is constructed based on these point clouds. Then, an octree-based MISS algorithm is launched to search for the existence of an inscribed cuboid with dimensions no smaller than 0.50m × 0.60m × 0.80m. The algorithm completes the search within 0.5 seconds and reports finding a maximum interference-free space with dimensions of 0.72m × 0.88m × 1.20m.

[0064] Step Six: Report Generation Finally, the system automatically generates a PDF inspection report. The first page of the report clearly lists the measured values ​​of all key dimensions such as ha, hb, and hc, the regulatory requirements, and the "pass / fail" judgment result. The appendix provides more detailed information, including: bird's-eye view and side view of the 3D point cloud, visualization of the semantic segmentation results, enlarged views of key reference points, and verification results of the safety space.

[0065] As can be seen from the above embodiments, the method of the present invention transforms a complex, experience-dependent testing process into a highly automated, data-driven, objective, and quantifiable scientific process.

[0066] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An elevator shaft top three-dimensional laser scanning reference point automatic detection method based on a deep learning and Bayesian fusion framework, characterized in that, The method comprises the following steps: Step 1, data acquisition: a three-dimensional laser scanning device is arranged on the top of the elevator car, and the top space of the elevator shaft is scanned one or more times when the elevator car runs to the highest maintenance position defined by the inspection regulation, to obtain original point cloud data containing three-dimensional coordinate information, reflectivity information and color information of each point; Step 2, data preprocessing: the original point cloud data is subjected to splicing registration, outlier elimination, downsampling and coordinate system alignment based on principal component analysis, so that the Z-axis of the coordinate system is consistent with the extension direction of the main structure of the shaft; Step 3, point cloud semantic segmentation: the preprocessed point cloud data is input into a pre-trained point cloud semantic segmentation network, each point in the point cloud is classified and predicted, and the point cloud is segmented into at least a plurality of predefined semantic regions including "shaft top", "top of the car highest standing area plane", "car top guardrail", "car top fixed equipment", "highest part of the guide or suspension system"; Step 4, multi-source candidate key point extraction: candidate key points with uncertainty quantification are extracted by parallel execution of at least two methods: Geometric analysis method: for the semantic region point cloud segmented as "shaft top" and "top of the car highest standing area plane", the corresponding spatial plane equation is fitted by applying the iterative reweighted least squares algorithm or the random sample consensus algorithm, and the lowest point or the highest point in the convex hull of the semantic region point cloud is calculated as the first type of candidate key point according to the equation, and the uncertainty is evaluated according to the fitting residual; AI prediction method: for the semantic region point cloud segmented as "car top fixed equipment" and "highest part of the guide or suspension system", the point cloud is input into a pre-trained probabilistic key point prediction network to obtain the three-dimensional coordinate expectation value of the part geometric highest point and the covariance matrix of the prediction uncertainty, which are the second type of candidate key points; Physical anchor point method: before data acquisition, a target sticker with a unique code is pasted at the physical position to be measured, the three-dimensional coordinates of the center point of the target sticker are located by image recognition and point cloud mapping algorithm, which are the third type of candidate key points, and the uncertainty is given according to the positioning accuracy model of the target recognition; Step 5, key point fusion based on Bayesian inference: a multi-source data fusion decision module is constructed, and the candidate key points and their uncertainty information extracted by the geometric analysis method, the AI prediction method and the physical anchor point method in step 4 are modeled as multi-dimensional Gaussian probability distribution; the multi-source data fusion decision module applies the Gaussian product rule to calculate the product of these multi-dimensional Gaussian probability distributions to obtain a posterior Gaussian distribution with smaller covariance, and the mean vector of the posterior distribution is taken as the only and high-precision final measurement reference point. Step 6, dimension calculation and space determination: based on the final measurement fiducials, automatically calculate the critical dimensions required by the inspection regulation and compare them with the regulatory standard values to determine whether they are qualified; and, within the free space point cloud defined by the "well top" and "car top" related semantic regions, perform an octree-based maximum inscribed space search algorithm to determine whether there is a non-interfering cuboid space no smaller than 0.50m x 0.60m x 0.80m.

2. The method of claim 1, wherein the method is based on a deep learning and Bayesian fusion framework for automatic detection of reference points of an elevator shaft top three-dimensional laser scanning. In the step 3, the training of the point cloud semantic segmentation network adopts a weighted cross-entropy loss function The mathematical definition is: ； where, is the total number of training batch midpoints, is the total number of semantic classes, is a point with a true label belonging to a class is the network's prediction of the probability that the point belongs to a class is the probability assigned to a class is the class weight assigned to a class 3.The method of claim 1, wherein, In step S4, the iterative reweighted least squares algorithm solves the plane equation through the following iterative process. parameters : In the first iteration, a weighted least squares problem is solved where is the sum of squared residuals for the points in the point cloud is the algebraic distance of a point to the plane defined by the parameters and the weight is computed based on the previous round of residuals by a robust weight function.

4. The method of claim 1, wherein the method is based on a deep learning and Bayesian fusion framework for automatic detection of reference points of an elevator shaft top 3D laser scan. In the step S4, the loss function of the probabilistic key point prediction network is designed to minimize the negative log-likelihood, which is specifically in the form of: ； wherein, is an inputted component point cloud, is a three-dimensional coordinate of a real key point, and are a mean vector and a covariance matrix of key point coordinates predicted by the network for the input, respectively.

5. The method of claim 1, wherein the method is based on a deep learning and Bayesian fusion framework for automatic detection of reference points of an elevator shaft top 3D laser scan. The specific way of Gaussian distribution modeling of candidate key points of different sources in step 5 is as follows: For the candidate points extracted by the geometric analysis method whose covariance matrix is constructed as a diagonal matrix whose diagonal elements are proportional to the square of the root mean square error after the IRLS algorithm fits the plane, i.e. where is a scaling factor, is the identity matrix; For the candidate points extracted by the AI prediction method, the mean value and the covariance matrix are directly output by the P-CPPN network; For the candidate points extracted by the physical anchor method whose covariance matrix is determined in accordance with a positioning accuracy evaluation model of the target recognition algorithm.

6. The method of claim 1, wherein the method is based on a deep learning and Bayesian fusion framework for automatic detection of reference points of an elevator shaft top 3D laser scan. The Gaussian multiplication rule is used to fuse individual, Gaussian-distributed observations The posterior distribution is obtained by computing their product The fused posterior distribution is still Gaussian with mean vector and covariance matrix computed analytically ； wherein, is the first observed precision matrix (information matrix), and the finally adopted coordinate of the measured reference point is the posterior mean .

7. The method of claim 1, wherein the method is based on a deep learning and Bayesian fusion framework for automatic detection of reference points of an elevator shaft top 3D laser scan. The specific implementation of the octree-based maximum inscribed space search algorithm includes: a) Construct the free space point cloud defined by the "well top" semantic region below and the "car top" related semantic region above into an octree data structure, where each leaf node represents a space voxel and is labeled as "free" or "occupied"; b) Starting from the root node of the octree, recursively search for the largest cuboid subspace composed entirely of leaf nodes labeled "free" in the tree in a depth-first or breadth-first manner; c) During the search process, implement an efficient pruning strategy using the hierarchical structure of the octree: when the size of the space cube represented by a node has already been smaller than any dimension of the target safety space 0.50m x 0.60m x 0.80m, there is no need to further explore all its descendant nodes.

8. An elevator shaft top three-dimensional laser scanning reference point automatic detection system based on a deep learning and Bayesian fusion framework, the system comprising: at least one processor; a memory communicatively connected to the processor; the memory stores program instructions in the memory and executable on the processor, characterized in that: the program instructions are configured to perform the method of any one of claims 1 to 7.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the method of any one of claims 1 to 7. The computer program is executed by the processor to realize the steps of the method of any one of claims 1 to 7.

Images