An intelligent testing method for monorail crane inspection robot based on data analysis

Abstract

The present invention discloses an intelligent testing method for a monorail inspection robot based on data analysis, which relates to the field of intelligent detection technology. The method includes synchronously collecting track images, point clouds, and posture data, performing time alignment and preprocessing, and outputting standardized data packets; using a YOLO detection network to detect inspection targets, combining a point cloud feature hierarchical extraction network with a temporal convolutional network to output multi-scale feature vectors; combining an improved A‑star algorithm with a long-short-term memory network to predict fault development trends, optimizing inspection paths through reinforcement learning, and outputting maintenance decision plans; converting maintenance decision plans into control instructions, driving the robot to perform inspection tasks and providing real-time feedback on operating status, combining incremental learning with point cloud reconstruction, and outputting a visual diagnostic report. The present invention uses a dynamic threshold algorithm for adaptive analysis and combines a cross-modal attention mechanism to calculate key geometric indicators, thereby improving the ability to identify structural anomalies.

Description

Technical Field

[0001] The present invention relates to the field of intelligent detection technology, and in particular to an intelligent testing method for a monorail crane inspection robot based on data analysis. Background Art

[0002] As the operating environment for rail transit equipment becomes increasingly complex, real-time perception of track structure status and fault identification have become critical to ensuring safe system operation. Currently, existing technologies mostly combine visual sensors, lidar, and inertial measurement units for data acquisition, achieving multimodal data fusion through time synchronization and coordinate transformation. Some advanced systems also incorporate reinforcement learning to optimize inspection paths. These systems are gradually evolving from traditional manual inspection to automated, intelligent inspection models.

[0003] While current track inspection systems can simultaneously collect image data, point cloud data, and posture data, the spatiotemporal alignment and fusion mechanisms for cross-modal features are incomplete, and they are unable to achieve the temporal coordination and spatial alignment of multiple sensor information. This directly limits the inspection system's ability to accurately identify structural anomalies and comprehensively judge complex faults, becoming a major factor hindering the development of intelligent automated track inspection. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides an intelligent testing method for a monorail inspection robot based on data analysis to solve the problem that track inspection lacks a multimodal data fusion mechanism, resulting in insufficient accuracy in structural fault identification.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides an intelligent testing method for a monorail crane inspection robot based on data analysis, which includes synchronously collecting track images, point clouds, and posture data, performing time alignment and preprocessing, and outputting standardized data packets;

[0008] Use the YOLO detection network to detect inspection targets, combine the point cloud feature hierarchical extraction network with the temporal convolutional network, and output multi-scale feature vectors;

[0009] The dynamic threshold algorithm is used to detect the number of rail bolts, joint width and horizontal misalignment, and output the initial fault judgment result;

[0010] Use evidence theory to fuse multi-source information, calculate comprehensive confidence and output the final fault diagnosis report;

[0011] Combine the improved A-star algorithm with the long short-term memory network to predict fault development trends, optimize inspection paths through reinforcement learning, and output maintenance decision plans;

[0012] The maintenance decision plan is converted into control instructions, driving the robot to perform inspection tasks and provide real-time feedback on the operating status. Combined with incremental learning and point cloud reconstruction, a visual diagnostic report is output.

[0013] As a preferred solution of the data analysis-based intelligent testing method for the monorail inspection robot of the present invention, the following specific steps are used to synchronously collect track images, point clouds and posture data, perform time alignment and preprocessing, and output standardized data packets:

[0014] The sensor data is unified into the world coordinate system by combining the timestamp and the coordinate transformation matrix, and a motion compensation algorithm is used to construct a multimodal dataset.

[0015] Preprocess and extract features from multimodal datasets, outputting images, cloud point data, inertial measurement unit preprocessing data, and track geometry features;

[0016] The protocol buffer format is used to encapsulate data, combined with optimized compression algorithm and cyclic redundancy check, to output standardized data packets.

[0017] As a preferred solution of the data analysis-based intelligent testing method of the monorail inspection robot described in the present invention, the YOLO detection network is used to detect the inspection target, and the point cloud feature hierarchical extraction network and the temporal convolutional network are combined to output the multi-scale feature vector. The specific steps are as follows:

[0018] Use the YOLOv11 target detection network to process images, perform three-level voxelized point cloud processing to extract point cloud features, and output image feature maps and point cloud level features;

[0019] Combining pose matrix projection and bilinear interpolation algorithm, the point cloud coordinates are transformed and then features are aligned to construct a bimodal feature group;

[0020] Use the temporal sliding mechanism to establish cross-frame feature associations, 3D convolution kernels to extract spatiotemporal features, and use multi-level feature fusion and adaptive weighting to fuse temporal features to output a multimodal fusion feature pyramid.

[0021] The multimodal fusion feature pyramid is normalized after removing redundant information through feature distillation to output a multi-scale feature vector.

[0022] As a preferred solution of the data analysis-based intelligent testing method for the monorail inspection robot of the present invention, the method detects the number of rail bolts, joint width and horizontal misalignment by a dynamic threshold algorithm and outputs the preliminary fault judgment result. The specific steps are as follows:

[0023] Establish cross-modal associations, calculate cross-attention matrices and weighted fusion features, and output fusion feature vectors;

[0024] The sliding window is used to calculate the mean and standard deviation, and the detection threshold is adaptively adjusted to generate a dynamic threshold.

[0025] The YOLOv11 detection algorithm is used to calculate the abnormal number of bolts. The YOLOv11-seg segmentation algorithm outputs the excessive joint width and abnormal horizontal misalignment, and outputs the initial fault judgment result.

[0026] As a preferred solution of the data analysis-based intelligent testing method for the monorail inspection robot of the present invention, the method utilizes evidence theory to fuse multi-source information, calculates the comprehensive confidence and outputs the final fault diagnosis report. The specific steps are as follows:

[0027] Combining the multi-scale feature vectors and multimodal datasets, the Z-score normalization method is used to normalize the multi-scale feature vectors and multimodal datasets, and then the statistical threshold method is used to remove outliers and output the feature matrix;

[0028] The Dirichlet distribution parameters are fitted to the feature matrix through a three-layer fully connected neural network, the KL divergence is used to calculate the conflict degree, and the initial evidence strength vector and evidence conflict matrix are output;

[0029] Combining the feature matrix, the initial evidence strength vector, and the evidence conflict matrix, the gated attention mechanism is used to reweight the features, and the exponential moving average algorithm is applied for time series fusion to output the optimized evidence distribution and uncertainty measure.

[0030] Diagnosis rules are constructed based on Bayesian decision theory, network parameters are updated using stochastic gradient descent optimizer, and fault diagnosis reports are output.

[0031] As a preferred solution of the data analysis-based intelligent testing method for the monorail inspection robot described in the present invention, the improved A-star algorithm is combined with the long short-term memory network to predict the fault development trend, and the inspection path is optimized through reinforcement learning to output the maintenance decision plan. The specific steps are as follows:

[0032] Filter effective fault points from the fault diagnosis report and establish spatial mapping, map the fault location to the map coordinate system, and construct a fault dataset;

[0033] Generate risk-aware inspection paths, calculate dynamic fault weights using the A-star algorithm, and output optimized inspection paths;

[0034] Combining fault datasets with real-time sensor data, a long-short-term memory network is used to predict fault development trends and generate high-risk prediction reports.

[0035] Combine fault data sets, optimized inspection routes, and high-risk prediction reports, classify them by urgency, associate maintenance resources, and output maintenance decision plans.

[0036] As a preferred solution of the data analysis-based intelligent testing method for the monorail inspection robot described in the present invention, the maintenance decision plan is converted into a control instruction, the robot is driven to perform the inspection task and the operating status is fed back in real time, and a visual diagnostic report is output by combining incremental learning and point cloud reconstruction. The specific steps are as follows:

[0037] Analyze maintenance decision plans, extract fault locations, task execution priorities, and maintenance actions, and output robot-executable instruction sets;

[0038] Drive the robot to perform tasks and monitor status, collect images, point clouds, and inertial measurement unit data in real time, and build real-time status data packages;

[0039] Based on real-time status data packets, incremental learning is used to dynamically adjust information weights, combined with point cloud reconstruction for three-dimensional rendering, and a visual diagnostic report is output.

[0040] In a second aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the intelligent testing method of the monorail crane inspection robot based on data analysis as described in the first aspect of the present invention.

[0041] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the intelligent testing method of the monorail crane inspection robot based on data analysis as described in the first aspect of the present invention.

[0042] The beneficial effects of the present invention are as follows: the present invention adopts a dynamic threshold algorithm for adaptive analysis, and combines it with a cross-modal attention mechanism to calculate geometric indicators, thereby improving the ability to identify structural anomalies, solving the defects of traditional single-modal methods that are limited by environmental interference and detection redundancy, and improving the accuracy and robustness of track structural fault identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0044] Figure 1 This is a flow chart of the intelligent testing method for the monorail crane inspection robot based on data analysis.

[0045] Figure 2 The flowchart of multimodal feature fusion and diagnosis.

[0046] Figure 3 Implement a flow chart for path planning and maintenance decision making.

[0047] Figure 4 This is the data collection and preprocessing flow chart. DETAILED DESCRIPTION

[0048] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0049] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0050] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0051] Reference Figures 1 to 4 , is an embodiment of the present invention, which provides an intelligent testing method for a monorail crane inspection robot based on data analysis, comprising the following steps:

[0052] S1. Synchronously collect track images, point clouds, and attitude data, perform time alignment and preprocessing, and output standardized data packets.

[0053] Furthermore, the sensor data is unified into the world coordinate system by combining the timestamp and the coordinate transformation matrix, and a motion compensation algorithm is used to construct a multimodal dataset;

[0054] Specifically, a precise time protocol is used for synchronization to make the sensor data time consistent, and the Kalman filter is used to dynamically compensate for the hardware clock drift, reduce the sampling delay of different sensors, and output the time-aligned multi-sensor data stream; a sensor external parameter calibration matrix is ​​established, and the relative position and posture of each sensor are determined through hand-eye calibration. A three-dimensional rigid body motion group is used to convert the sensor data into the world coordinate system to construct multimodal data in a unified coordinate system; the inertial measurement unit data is used in combination with the B-spline interpolation algorithm to calculate the instantaneous position and posture of the lidar during the scanning process in real time, and the point cloud data is corrected by inverse motion according to the estimated position and posture changes to eliminate the point cloud stretching and compression distortion caused by the movement of the monorail crane, and output a distortion-free point cloud. The angular velocity data provided by the inertial measurement unit data is combined to improve the image clarity, output a deblurred image, and construct a multimodal data set;

[0055] It should be noted that a three-dimensional rigid body motion group is used to transform the sensor data into the world coordinate system. First, the relative position and posture of each sensor are measured to form a fixed rigid body transformation matrix. Then, the inertial measurement unit data is used to dynamically compensate for the coordinate offset caused by the robot motion. Finally, the sensor data is transformed into a unified world coordinate system through matrix multiplication.

[0056] The better approach is to reduce motion blur or point cloud distortion caused by time asynchrony through time alignment, improve the accuracy of subsequent feature extraction, and reduce dynamic errors. Unifying data to the world coordinate system reduces sensor installation errors and ensures consistency in multimodal data space. The motion compensation algorithm solves the point cloud stretching and compression problems caused by high-speed motion, improving the accuracy of track geometric feature extraction. At the same time, multi-sensor data fusion improves detection robustness.

[0057] Preprocess and extract features from multimodal datasets, outputting images, cloud point data, inertial measurement unit preprocessing data, and track geometry features;

[0058] Specifically, a bilateral filtering algorithm is used to eliminate image noise and retain edge information, adaptive histogram equalization is used to solve the problem of uneven illumination, and lens distortion correction is performed based on camera calibration parameters to output a clear image with denoised, balanced illumination and no distortion; a statistical outlier removal algorithm is used to remove outliers, and random sampling consistency plane detection is used for ground segmentation. A three-dimensional voxelization method is used to reduce the amount of data and perform voxel grid downsampling to output a denoised and streamlined track structure point cloud; a static calibration method is used to eliminate sensor zero bias, and a Kalman filter algorithm is used for noise filtering to output accurate attitude estimation data; a directional gradient histogram is used for feature description, and a track surface is constructed based on a Poisson reconstruction algorithm, and images, cloud point data, inertial measurement unit preprocessing data, and track geometric features are output;

[0059] Optimally, through denoising, illumination balancing, and geometric correction, environmental interference can be effectively eliminated, enabling the visual inspection algorithm to accurately identify subtle defects on the track surface, such as cracks, rust, and other minor damage. Outlier removal and ground segmentation can accurately extract track structural features, reducing false detections and missed detections. At the same time, zero-bias correction and noise filtering can significantly improve the stability of attitude estimation. Multimodal feature fusion can cross-validate detection results, reduce misjudgments caused by single sensor failure, and improve the credibility of fault diagnosis. The entire preprocessing process can adapt to different lighting conditions and different vibration environments, improving detection accuracy, enhancing adaptability, and improving the reliability of fault diagnosis.

[0060] Use protocol buffer format to encapsulate data, combine optimized compression algorithm and cyclic redundancy check, and output standardized data packets;

[0061] Specifically, it receives structured data from the feature extraction module, classifies and organizes the data, distinguishes different types of data features, uses protocol buffer encoding to compress numerical data using variable-length integer encoding, establishes a nested message structure for geometric feature data, and sets a fixed field type for text metadata, and performs binary serialization conversion; analyzes the serialized data, implements a hierarchical compression strategy, calculates a cyclic redundancy check code for the compressed data, adds a data packet sequence number and timestamp, verifies and validates key fields, and generates a standardized data packet;

[0062] The best approach is to minimize data size through binary encoding and intelligent compression. The hierarchical data structure supports fast retrieval and parsing. The multiple verification mechanisms reduce data transmission errors while supporting breakpoint resumption and error recovery. The standardized format supports cross-platform exchange. The compact storage format saves space.

[0063] It should be noted that the data processing process is fully automated and standardized. The detection accuracy and reliability are improved through multi-sensor collaborative enhancement, dynamic error compensation and intelligent anti-interference. High-efficiency data compression and edge computing optimization make resource utilization more optimized. At the same time, in terms of engineering practicality and maintainability, it has also been enhanced through standardized data packets, cyclic redundancy check and block processing mechanisms.

[0064] S2. Use the YOLO detection network to detect inspection targets, combine the point cloud feature hierarchical extraction network with the temporal convolutional network, and output multi-scale feature vectors.

[0065] Furthermore, the image is processed using the YOLOv11 object detection network, and the three-level voxelized point cloud processing is used to extract point cloud features, outputting image feature maps and point cloud level features;

[0066] Specifically, the collected sensor detection images are enhanced and distorted through an open source computer vision library, and the detection results and preliminary sensor detection feature maps are output. The feature pyramid network is then used to enhance the small target detection capability, and the image feature map and detection results are output. Point cloud rasterization and feature calculation are used to output low-resolution voxel features. The first-level voxel features are scanned using three-dimensional sparse convolution to extract local geometric features. The grid resolution is reduced by combining maximum pooling, and medium-resolution voxel features are output. The channel attention module is used to automatically identify important voxels. The third-level voxel features are spliced ​​with the second-level features to generate a high-dimensional feature vector, and a multi-scale feature vector is output.

[0067] The preferred method is to use first-level voxelization to reduce data volume, improve computational efficiency, and preserve the overall geometric structure of the track. Second-level voxelization captures details at track joints and suppresses noise points, such as outliers caused by dust and flying insects. Third-level voxelization enhances the feature expression of fault-sensitive areas and provides a geometric basis for subsequent multimodal fusion. Through this three-level voxelization, the overall track structure can be captured while local defects can be accurately located. The point cloud data provides precise 3D geometric features, improving detection accuracy while also having anti-interference capabilities and optimizing computational efficiency.

[0068] Combining pose matrix projection and bilinear interpolation algorithm, the point cloud coordinates are transformed and then features are aligned to construct a bimodal feature group;

[0069] Specifically, a calibration plate is installed on the robot, and camera images and lidar point clouds are collected at the same time. The corner points of the calibration plate are detected by computer vision algorithms, and the corresponding points scanned by the lidar are matched. The spatial relationship between the camera and the lidar is optimized and calculated, and the camera lidar pose matrix is ​​output; the lidar point cloud is mapped to the camera image coordinate system using the pose matrix, and the three-dimensional point cloud data scanned by the lidar is converted to the camera coordinate system according to the pose matrix. According to the camera parameters, the three-dimensional points are projected onto the two-dimensional image plane to generate a sparse depth map, and invalid points that exceed the actual range of the track are eliminated to construct a sparse depth map aligned with the image pixels; on the sparse depth map, for each missing pixel point, the four nearest valid depth values ​​around it are found, the depth value of the pixel point is calculated according to the distance weight, the interpolation result is optimized in combination with the image features, and a complete dense depth map is output; a convolutional neural network is used to extract features from the image RGB data and depth data respectively, and the most relevant image and depth features are automatically selected for fusion through the attention mechanism to construct a dual-modal feature group;

[0070] Use the temporal sliding mechanism to establish cross-frame feature associations, 3D convolution kernels to extract spatiotemporal features, and use multi-level feature fusion and adaptive weighting to fuse temporal features to output a multimodal fusion feature pyramid.

[0071] Specifically, the temporal modeling capability is implicitly introduced into the spatial convolution to establish feature associations between adjacent frames. Some channels of the current frame features are replaced with the corresponding channels. Conventional 2D convolution is performed on the shifted features to mix spatiotemporal information and output a feature tensor containing temporal associations. The temporal feature tensor is scanned with a 3D convolution kernel to extract short-term motion patterns. 3D pooling compression is used to downsample along the time dimension to reduce the impact of redundant frames, suppress transient noise, focus on persistent faults, and output a spatiotemporal feature map. The 3D convolution output is deconvolved, and the temporal sliding mechanism output and the upsampled features are spliced ​​along the channel dimension. The weights of the two-level features are automatically assigned through adaptive weighting to output a multimodal fusion feature map.

[0072] The superior spatiotemporal feature extraction capability of 3D convolution can distinguish instantaneous interference from real faults, improving detection stability. The multi-level feature fusion pyramid structure achieves an optimized combination of features at different time granularities, enabling rapid response to sudden faults while continuously tracking slowly developing defect types. The spatiotemporal fusion detection mechanism improves environmental adaptability and fault coverage.

[0073] For the multimodal fusion feature pyramid, redundant information is removed through feature distillation and then normalized to output a multi-scale feature vector.

[0074] Specifically, the knowledge distillation framework and channel attention mechanism are used to compress feature dimensions, retain information that is effective for fault detection, and output a compact feature map after distillation. Mean-variance normalization is performed on each feature channel, and feature expression capabilities are restored through learnable parameters. Nonlinearity is enhanced by combining nonlinear activation, and a standardized multi-scale feature vector is output.

[0075] The better one uses knowledge distillation technology to effectively extract key information from multimodal fusion features, remove redundant data interference, and focus more accurately on the core characteristics of track faults; through the adaptive screening of the channel attention mechanism, it can intelligently distinguish important features from noise features, and improve the ability to identify subtle faults; standardization processing greatly reduces the feature distribution differences caused by different sensors and data sources, and enhances stability and generalization capabilities under various environmental conditions.

[0076] S3. Detect the number of rail bolts, joint width, and horizontal misalignment through a dynamic threshold algorithm, and output the initial fault judgment result.

[0077] Furthermore, we establish cross-modal associations, calculate the cross-attention matrix and weighted fusion features, and output the fusion feature vector;

[0078] Specifically, the YOLOv11 backbone network is used to extract multi-scale feature maps, a point cloud processing network is used to extract local geometric features, and a one-dimensional convolutional neural network is used to process inertial measurement unit data to extract temporal features. A multi-head cross-attention mechanism is used to establish the correlation between visual, geometric, and temporal data, and output weighted cross-modal features. The dimensions of the modal features are checked to see if they are the same. If not, the dimensions are unified through linear transformation, and the arithmetic mean of the three modal features is directly taken to output a fused feature vector.

[0079] The best approach is to integrate visual, geometric, and time series data to more comprehensively analyze track conditions and improve detection accuracy and stability. Dynamically adjust the weights of different modal data to adapt to complex and changing inspection environments and enhance environmental adaptability. Visual inspection of bolt count, geometric data verification of installation angles, and time series data analysis of vibration trends enhance the comprehensiveness of fault diagnosis.

[0080] The sliding window is used to calculate the mean and standard deviation, and the detection threshold is adaptively adjusted to generate a dynamic threshold.

[0081] Specifically, a sliding window is used to calculate local statistics. The arithmetic mean of the data within the window reflects the current local trend. The standard deviation measures the degree of data dispersion and is used to determine abnormal fluctuations. The local mean and standard deviation of each window are output to form a time series statistics sequence. An adaptive threshold is used to adjust the sensitivity coefficient and threshold constraint. Anomaly detection logic is added and combined with incremental calculation to generate a dynamic threshold.

[0082] The best approach is to improve the intelligence and refinement of fault detection through sliding window statistics and adaptive threshold adjustment, taking into account real-time performance, robustness, and explainability.

[0083] The YOLOv11 detection algorithm is used to calculate the number of bolt anomalies. The YOLOv11-seg segmentation algorithm outputs the joint width exceeding the limit and the horizontal misalignment anomaly, and outputs the initial fault judgment result.

[0084] Specifically, the YOLO11 algorithm is used to detect bolt joints and bolts. If a bolt joint is detected, the bolts are counted. The segmentation algorithm is used to calculate the gap range, extract the gap range, and perform parameter calculation through maximum inscribed quadrilateral fitting. The length of the upper base of the quadrilateral is converted into the actual physical size as the joint width. The angle between the two sides of the quadrilateral is calculated to determine whether the installation angle exceeds the limit. The bottom and top sides of the upper and lower quadrilaterals are extracted respectively, and the horizontal projection difference of the two sides is calculated to determine whether the horizontal misalignment exceeds the limit. The angle between the upper left side and the lower left side and the right side are calculated, and the average is taken to determine whether the angle exceeds the limit. The initial fault judgment result is output.

[0085] S4. Use evidence theory to fuse multi-source information, calculate the comprehensive confidence and output the final fault diagnosis report.

[0086] Furthermore, the multi-scale feature vectors and multi-modal datasets are combined and normalized using the Z-score normalization method. Statistical thresholding is then used to remove outliers and output a feature matrix.

[0087] Specifically, it receives multi-scale feature vectors from the YOLOv11 detection network, synchronously receives heterogeneous data such as point cloud data and inertial measurement unit posture data, combines timestamp alignment and coordinate transformation matrix, and outputs a time-aligned multimodal data matrix; calculates the global mean and standard deviation for each feature dimension, transforms each data point, uses a sliding window to correct, and outputs a normalized matrix; uses a statistical threshold method to remove outliers, combines static anomaly detection with dynamic anomaly filtering, directly eliminates sensor failure data and uses linear interpolation to output the cleaned feature matrix;

[0088] The better one provides a high-credibility input foundation for subsequent evidence fusion through strict time alignment, standardization and anomaly cleaning;

[0089] The Dirichlet distribution parameters are fitted to the feature matrix through a three-layer fully connected neural network, the KL divergence is used to calculate the conflict degree, and the initial evidence strength vector and evidence conflict matrix are output;

[0090] Specifically, numerical scaling and anomaly removal are combined to calculate the mean and standard deviation of each column of data, scale each data point, and discard or replace any data points that are out of range with adjacent valid values ​​to construct a standardized feature matrix. Frequency statistics are used to count the number of faults such as missing bolts and track misalignment, calculate the proportion of each type of fault, adjust the weights, and output an initial evidence strength vector. A consistency rule library is used to define physical constraints between sensors and mark conflicts. All conflicting sensor combinations and their degree of deviation are recorded, and an evidence conflict matrix is ​​constructed. The initial evidence strength vector and the evidence conflict matrix are output.

[0091] The better one is that it uses a three-layer fully connected neural network to fit the parameters of the Dirichlet distribution and combines it with the KL divergence to calculate the degree of evidence conflict, thus achieving highly robust fusion of multi-source data, solving the problems of small samples and class imbalance, and improving practicality.

[0092] Combining the feature matrix, the initial evidence strength vector, and the evidence conflict matrix, the gated attention mechanism is used to reweight the features, and the exponential moving average algorithm is applied for time series fusion to output the optimized evidence distribution and uncertainty measure.

[0093] Specifically, the feature matrix is ​​fused with the initial evidence strength through a fully connected layer, and the evidence conflict matrix is ​​used to generate a gating mask. Gating is applied to the fused features, and a gating weight matrix is ​​output. The weights are adjusted using the evidence conflict matrix, and weighted evidence is calculated for each type of fault. The reweighted evidence vector and attention weight matrix are output. The exponential moving average algorithm is used to smooth the evidence and calculate the cumulative uncertainty, and the optimized evidence distribution and uncertainty measurement value are output.

[0094] It should be noted that the gated attention mechanism consists of three parts: feature enhancement layer, conflict gate generator and gated attention calculation;

[0095] The synergy between gated attention and exponential moving average improves dynamic adaptability and optimizes computational efficiency.

[0096] Construct diagnostic rules based on Bayesian decision theory, use stochastic gradient descent optimizer to update network parameters, and output fault diagnosis reports;

[0097] Specifically, based on the optimized evidence distribution and uncertainty measure, the Dirichlet parameters are converted into class conditional probabilities, the posterior probability is calculated, and combined with uncertainty adjustment, a fault probability vector and a decision basis report are output. The fault diagnosis report is output by minimizing the Bayesian risk optimization objective, using stochastic gradient descent with momentum and a parameter update strategy.

[0098] The better one solves the consistency problem of heterogeneous sensor data through spatiotemporal alignment, standardization and outlier elimination, ensures the reliability of subsequent analysis, and can automatically adapt to complex environmental interference such as dust and vibration; the dynamic fusion mechanism improves the robustness of the diagnostic results; through gated attention and temporal fusion technology, it achieves dual optimization of spatial and temporal dimensions, which can not only quickly respond to sudden failures, but also capture slowly developing potential defects, forming a comprehensive health status assessment.

[0099] S5. Combine the improved A-star algorithm with the long short-term memory network to predict the development trend of faults, optimize the inspection path through reinforcement learning, and output the maintenance decision plan.

[0100] Furthermore, the fault diagnosis report is used to screen valid fault points and establish a spatial mapping, mapping the fault location to the map coordinate system to construct a fault dataset;

[0101] Specifically, fault diagnosis reports are sorted by confidence, high-confidence faults are retained, repeated fault points in the same area are spatially clustered to remove redundancy, and a de-redundant fault list is output. Synchronous positioning and mapping are used to transform the fault points from the robot's local coordinate system to the global map coordinate system. The global coordinates are calculated based on the robot's real-time pose and the local coordinates of the fault points, and the set of fault locations in the global coordinate system is output. A fault association graph is constructed with the fault points as nodes and spatial temporal continuity as edges. The fault information is written into a time series database, and a fault association network is established based on the graph database to construct a fault dataset.

[0102] The best solution is to improve the reliability of fault data through a multi-level screening mechanism, improve positioning accuracy through precise coordinate conversion, and improve operation and maintenance level through intelligent processing;

[0103] Generate risk-aware inspection paths, calculate dynamic fault weights using the A-star algorithm, and output optimized inspection paths;

[0104] It should be noted that the risk-aware inspection path is generated, the dynamic weight of the fault is calculated by combining the A-star algorithm, and the optimized inspection path is output;

[0105] Furthermore, we use the fault data set to screen high-confidence fault points, calculate dynamic risk values ​​to quantify fault risks, and construct a risk heat map;

[0106] Specifically, a confidence threshold is set. Statistical analysis is used to calculate the mean and standard deviation based on the confidence distribution of historical detection data. A confidence threshold is then set. For each target detection result, a minimum confidence threshold is set. Detection boxes below this threshold are discarded, and low-reliability detection results are filtered. Adjacent fault points are merged through spatial clustering to output a set of high-confidence fault points. A dynamic risk value is calculated for each fault point in the set. The time attenuation factor is combined to quantify the comprehensive impact of the risk. A list of fault points with dynamic risk scores is constructed. Spatial discretization and kernel density estimation are used to construct a risk heat map.

[0107] The best solution is to effectively filter out false detections caused by environmental noise through confidence threshold screening and spatial clustering, ensure the authenticity of fault data, improve anti-interference capabilities, and enhance detection reliability. It also integrates multiple indicators such as real-time confidence, historical statistics, and fault severity to perform multi-dimensional quantification and enhance the comprehensiveness of risk assessment.

[0108] Combine the A-star algorithm with the long short-term memory network prediction correction to output the predicted risk area;

[0109] Specifically, a sliding window method is used to generate time series samples based on historical fault data. Continuous sensor data is Z-value standardized, discrete fault categories are one-hot encoded, synthetic minority class oversampling is used to handle sample imbalance, a bidirectional long short-term memory network layer is used to capture forward and reverse time series dependencies, cross entropy loss is used for classification tasks, and mean square error is used for regression tasks. The early stopping method is combined to prevent overfitting, and a fault probability prediction map is output. A cost function is used to calculate the sequence nodes that minimize the global risk. For the area around the execution task path, areas with predicted risk values ​​higher than the threshold are extracted, and clustering is used to form continuous risk areas. Scattered high-risk points are merged into continuous risk areas with engineering significance. Obvious outliers are removed from all high-risk points extracted from the execution task path, and dense areas are downsampled. Spatial domain and risk similarity are calculated, and core points are identified. Starting from any core point, all its adjacent high-risk points are incorporated into the current area. The area is recursively expanded until no new points can be absorbed. The coordinate extreme values ​​of all points are calculated, and the minimum enclosing rectangle or polygon is generated to generate the boundary. Validity verification is performed, and a continuous risk area is output. A predicted risk area report is constructed based on the continuous risk area, and the predicted risk area is output.

[0110] When it is explained that, for the areas around the execution path, the areas with predicted risk values ​​higher than the threshold are marked as high-risk areas, and the rest are normal areas. This can be done through dynamic environmental adaptive adjustment, or by performing kernel density estimation on the predicted probability output by the long-short-term memory network in historical fault data to determine the value range of the high-risk area. The execution path is the movement trajectory planned by the monorail inspection robot during the execution of the task, including spatial dimension, time dimension and attribute extension.

[0111] The best approach is to predict future fault development trends through long-short-term memory networks, enabling robots to avoid high-risk areas that may worsen in advance, improving the foresight and long-term effectiveness of inspections and enhancing the predictability of path planning; improving adaptability to complex environments; optimizing resource allocation efficiency, and prioritizing high-risk areas through risk probability weighting, while reasonably skipping safe sections to reduce ineffective inspection time and energy consumption;

[0112] It should be noted that the cost function expression is as follows:

[0113] ;

[0114] in, is the total cost, , evaluates the quality of path planning, the smaller the value, the better the path. Indicates the actual moving cost, including distance, energy consumption, turning penalty, etc. To estimate the cost to reach the destination, Represents the predicted risk probability, 0 P 1, is the risk-efficiency balance coefficient, 0.5 0.8.

[0115] Use the proximal strategy optimization algorithm to dynamically adjust the path planning strategy and output the optimized inspection path;

[0116] Specifically, the risk-efficiency balance factor is combined to adjust the weight ratio of risk avoidance and inspection efficiency in path planning. The dynamic weight calculation basic weight is combined with environmental adaptive adjustment to output the dynamic adjustment coefficient. The emergency risk avoidance rules, energy efficiency optimization rules and regional coverage rules are combined with the optimized inspection path and the predicted risk area to generate the initial path. The dynamic adjustment mechanism and exception handling protocol are used to output the real-time optimized path.

[0117] It should be noted that the emergency avoidance rule is to enable the robot to take immediate obstacle avoidance action to prevent collision or derailment when a high-risk fault is detected; the energy efficiency optimization rule is to maximize the endurance while ensuring the basic inspection quality; the area coverage rule is to ensure full coverage of key components to avoid missed inspections;

[0118] It should be noted that the dynamic adjustment mechanism and exception handling protocol, the dynamic adjustment mechanism is to automatically optimize the robot's behavior parameters according to real-time environmental changes and task requirements; the core principles of the exception handling protocol are hierarchical response, minimal intervention and traceability;

[0119] The better ones improve real-time response capabilities through rule engines, improve adaptability through dynamic weight mechanisms, improve reliability through hierarchical rule bases, and improve real-time performance, reliability, and maintainability through rule-based design.

[0120] Combining fault datasets with real-time sensor data, a long-short-term memory network is used to predict fault development trends and generate high-risk prediction reports.

[0121] Specifically, a sliding window method is used to generate time series samples, fault categories are one-hot encoded, environmental parameters are Z-value normalized, and the problem of sample imbalance is addressed through synthetic minority class oversampling technology. A standardized time series feature matrix is ​​output. A bidirectional long short-term memory network layer is used to capture forward and reverse time series dependencies, combined with the attention mechanism layer to weight time step features. The fully connected layer outputs the probability of fault deterioration, and a high-risk prediction report is generated based on risk grading.

[0122] It should be noted that fault categories refer to the specific fault types that need to be detected and predicted in the monorail track system. They are generally divided into two categories: structural faults and functional faults. They include missing and loose bolts, cracked track joints, track surface corrosion, horizontal track misalignment, insulation damage, and track geometric deformation. Environmental parameters are external environmental factors that affect the occurrence and development of faults, including meteorological conditions, mechanical loads, electrical parameters, chemical factors, and time characteristics (accumulated operating time, last maintenance interval).

[0123] The best approach is to capture the dynamic trends of fault development through long-short-term memory networks, achieve early warning, and enable maintenance personnel to take intervention measures before the fault worsens, reducing the probability of sudden equipment failure and improving the timeliness and foresight of predictions. Combined with the attention mechanism, it automatically focuses on key influencing factors, making prediction results more targeted and reducing misjudgments. Using incremental learning technology, it dynamically absorbs new data and adjusts the prediction logic to adapt to changes in the track environment, enhancing the initiative and accuracy of risk management.

[0124] Combine fault data sets, optimized inspection routes, and high-risk prediction reports to output maintenance decision plans based on urgency levels and associated maintenance resources.

[0125] Specifically, the system combines fault data sets, optimized inspection routes, and high-risk prediction reports to perform a weighted calculation of the urgency index. Adjacent fault points are then merged and processed, classified by urgency, and a graded fault list is output. The system also calculates the minimum total response time, dynamically adjusts resource weights, and outputs a resource allocation plan based on matching logic. The resource allocation plan is manually reviewed, and natural language generation is used to convert structured data into spoken instructions, ultimately outputting a maintenance decision plan.

[0126] The best approach is to automate the entire process, reducing manual reliance and improving the speed and consistency of operational response. The spatiotemporal alignment and fusion of multimodal data mitigates the limitations of single sensors and enhances comprehensive detection. Combined with the A-star algorithm for path optimization, it balances fault risk and inspection efficiency, achieving dynamic adaptive planning. Incremental learning allows for continuous adaptation to environmental changes.

[0127] It should be noted that the urgency level can be divided into three levels: emergency, high-risk and routine; the matching logic can be set by yourself. For example, emergency tasks are solved by dispatching the nearest three senior technicians and a full-function maintenance vehicle, high-risk tasks are solved by two ordinary technicians and a light tool kit, and routine tasks are included in the next planned inspection.

[0128] S6. Convert maintenance decision plans into control instructions, drive the robot to perform inspection tasks and provide real-time feedback on operating status. Combine incremental learning with point cloud reconstruction to output a visual diagnostic report.

[0129] Furthermore, it analyzes the maintenance decision plan, extracts the fault location, task execution priority and maintenance action, and outputs the robot executable instruction set;

[0130] Specifically, based on global positioning and local point cloud registration using simultaneous localization and mapping, the precise location of the fault in the global map is determined and the 3D coordinates of the fault are output. Dynamic weights are used to output a list of fault priorities. Combining the knowledge graph with the robotic arm motion template library, the preset motion templates are called to output an executable instruction set.

[0131] It should be noted that the preset action template content is called, such as the torque for tightening a bolt and the motion trajectory of the robot arm;

[0132] Optimally, fault location analysis can provide high-precision target points for path planning; dynamic weighting can be used to improve the adaptability of fault response strategies; knowledge graphs and preset action templates can be used to improve the efficiency of standardized maintenance processes and reduce the risk of operational errors;

[0133] Drive the robot to perform tasks and monitor status, collect image data, point cloud data, and inertial measurement unit data in real time, and build a real-time status data package;

[0134] Specifically, the instruction set is parsed and decomposed into subtasks such as chassis movement, robotic arm movement, and sensor adjustment. The dynamic window method is used for real-time path planning to avoid sudden obstacles and output the status feedback of each execution unit of the robot. Hardware time synchronization and sensor calibration and fusion are used in combination with real-time sensor data to align the timestamps and coordinate system of real-time data and output a time-aligned multimodal data stream. The sensor data stream and robot state feedback are combined with threshold detection, trend analysis and logical judgment to set the safety threshold of the adjustable parameters of the control task operation, including algorithm parameters, hardware control parameters, etc. The sliding window statistics are used to detect data mutations, and compound anomalies are judged by logical rules. The real-time health status report and anomaly type classification are output to construct a real-time status data packet.

[0135] Based on real-time status data packets, it uses incremental learning to dynamically adjust information weights, combines point cloud reconstruction with 3D rendering, and outputs a visual diagnostic report.

[0136] Specifically, it combines incremental learning with dynamic adjustment of information weights, dynamically updates feature weights based on new data, adopts a sliding window mechanism to retain recent key data, automatically reduces the weight of abnormal data through the credibility of information entropy sensor data, and outputs the dynamically adjusted feature weight matrix and; performs time series registration on continuous frame point clouds to eliminate motion distortion, adopts adaptive voxel filtering to balance accuracy and computational complexity, fuses multi-view scanning data, completes occluded areas, outputs point cloud quality assessment reports and high-integrity three-dimensional point cloud data; converts point clouds into colored three-dimensional grid patterns, adds visual markers according to fault types, and outputs interactive three-dimensional visualization reports in combination with integrated multi-view observation functions; automatically extracts indicators such as misalignment and number of missing bolts, sorts fault lists by urgency, generates fault distribution heat maps, repair priority recommendations and predictive maintenance plans, and outputs visualization diagnostic reports;

[0137] It should be noted that the point cloud is converted into a colored three-dimensional grid pattern;

[0138] Specifically, floating noise points are removed, track structure points are retained, and point cloud smear caused by robot movement is corrected using inertial measurement unit data based on motion compensation. Missing areas such as track joints are filled with holes and interpolated using surrounding point cloud data. Surface reconstruction is performed by combining triangular mesh generation and mesh optimization. 2D image pixels captured by the camera are projected onto a 3D mesh surface based on camera parameters. Multi-view image fusion is performed, and the weighted average of the colors of each view is taken for color matching, faults are highlighted, and the image is converted into a colored 3D mesh pattern.

[0139] The best solution is to improve the automation level of monorail crane inspection through multi-sensor fusion and intelligent decision-making algorithms; improve the accuracy of detection and positioning through spatiotemporal synchronization and fusion analysis of multi-source data; shorten the average fault handling time and improve the overall operation and maintenance efficiency through dynamic priority evaluation and task optimization scheduling.

[0140] This embodiment also provides a computer device, which is suitable for the case of a monorail crane inspection robot intelligent testing method based on data analysis, including: a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute computer executable instructions to implement the monorail crane inspection robot intelligent testing method based on data analysis proposed in the above embodiment.

[0141] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0142] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the intelligent testing method for a monorail inspection robot based on data analysis proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, disk or optical disk.

[0143] In summary, the present invention improves the ability to identify structural anomalies by using a dynamic threshold algorithm for adaptive analysis and combining it with a cross-modal attention mechanism to calculate geometric indicators. This addresses the limitations of traditional single-modality methods, which are limited by environmental interference and detection redundancy, and improves the accuracy and robustness of track structural fault identification.

[0144] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An intelligent testing method for a monorail crane inspection robot based on data analysis, characterized by: include, Synchronously collect track images, point clouds, and attitude data, perform time alignment and preprocessing, and output standardized data packages; Use the YOLO detection network to detect the inspection target, combine the point cloud feature level extraction network and the temporal convolutional network, and output multi-scale feature vectors. The specific steps are as follows: Use the YOLOv11 target detection network to process images, perform three-level voxelized point cloud processing to extract point cloud features, and output image feature maps and point cloud level features; Combining pose matrix projection and bilinear interpolation algorithm, the point cloud coordinates are transformed and then features are aligned to construct a bimodal feature group; Use the temporal sliding mechanism to establish cross-frame feature associations, 3D convolution kernels to extract spatiotemporal features, and use multi-level feature fusion and adaptive weighting to fuse temporal features to output a multimodal fusion feature pyramid. For the multimodal fusion feature pyramid, redundant information is removed through feature distillation and then normalized to output a multi-scale feature vector. The dynamic threshold algorithm is used to detect the number of rail bolts, joint width and horizontal misalignment, and output the initial fault judgment result; Use evidence theory to fuse multi-source information, calculate the comprehensive confidence and output the final fault diagnosis report. The specific steps are as follows: Combining the multi-scale feature vectors and multimodal datasets, the Z-score normalization method is used to normalize the multi-scale feature vectors and multimodal datasets, and then the statistical threshold method is used to remove outliers and output the feature matrix; The Dirichlet distribution parameters are fitted to the feature matrix through a three-layer fully connected neural network, the KL divergence is used to calculate the conflict degree, and the initial evidence strength vector and evidence conflict matrix are output; Combining the feature matrix, the initial evidence strength vector, and the evidence conflict matrix, the gated attention mechanism is used to reweight the features, and the exponential moving average algorithm is applied for time series fusion to output the optimized evidence distribution and uncertainty measure. Construct diagnostic rules based on Bayesian decision theory, use stochastic gradient descent optimizer to update network parameters, and output fault diagnosis reports; Combine the improved A-star algorithm with the long short-term memory network to predict the fault development trend, optimize the inspection path through reinforcement learning, and output the maintenance decision plan. The specific steps are as follows: Filter fault points in the fault diagnosis report and establish spatial mapping, map the fault location to the map coordinate system, and build a fault dataset; Generate risk-aware inspection paths, calculate dynamic fault weights using the A-star algorithm, and output optimized inspection paths; Combining fault datasets with real-time sensor data, a long-short-term memory network is used to predict fault development trends and generate high-risk prediction reports. Combine fault data sets, optimized inspection routes, and high-risk prediction reports to output maintenance decision plans based on urgency levels and associated maintenance resources. The maintenance decision plan is converted into control instructions, driving the robot to perform inspection tasks and provide real-time feedback on the operating status. Combined with incremental learning and point cloud reconstruction, a visual diagnostic report is output.

2. The method for intelligent testing of a monorail inspection robot based on data analysis according to claim 1, characterized in that: The synchronous acquisition of track images, point clouds and attitude data, and the time alignment and preprocessing to output standardized data packets are as follows: The sensor data is unified into the world coordinate system by combining the timestamp and the coordinate transformation matrix, and a motion compensation algorithm is used to construct a multimodal dataset. Preprocess and extract features from multimodal datasets, outputting images, cloud point data, inertial measurement unit preprocessing data, and track geometry features; The protocol buffer format is used to encapsulate data, combined with optimized compression algorithm and cyclic redundancy check, to output standardized data packets.

3. The method for intelligent testing of a monorail inspection robot based on data analysis according to claim 2, characterized in that: The dynamic threshold algorithm is used to detect the number of rail bolts, joint width and horizontal misalignment, and output the initial fault judgment result. The specific steps are as follows: Establish cross-modal associations, calculate cross-attention matrices and weighted fusion features, and output fusion feature vectors; Use sliding windows to calculate mean and standard deviation, and combine with adaptive threshold adjustment to generate dynamic thresholds; The YOLOv11 detection algorithm is used to calculate the abnormal number of bolts. The YOLOv11-seg segmentation algorithm outputs the excessive joint width and abnormal horizontal misalignment, and outputs the initial fault judgment result.

4. The method for intelligent testing of a monorail inspection robot based on data analysis according to claim 3, characterized in that: The maintenance decision plan is converted into control instructions, driving the robot to perform inspection tasks and provide real-time feedback on the operating status. Incremental learning and point cloud reconstruction are combined to output a visual diagnostic report. The specific steps are as follows: Analyze maintenance decision plans, extract fault locations, task execution priorities, and maintenance actions, and output robot-executable instruction sets; Drive the robot to perform tasks and monitor status, collect images, point clouds, and inertial measurement unit data in real time, and build real-time status data packages; Based on real-time status data packets, incremental learning is used to dynamically adjust information weights, combined with point cloud reconstruction for three-dimensional rendering, and a visual diagnostic report is output.

5. The method for intelligent testing of a monorail inspection robot based on data analysis according to claim 4, characterized in that: The risk-aware inspection path is generated, the dynamic weight of the fault is calculated by combining the A-star algorithm, and the optimized inspection path is output. The specific steps are as follows: Utilize the fault data set to screen high-confidence fault points, calculate dynamic risk values ​​to quantify fault risks, and construct a risk heat map; Combine the A-star algorithm with the long short-term memory network prediction correction to output the predicted risk area; Use the proximal strategy optimization algorithm to dynamically adjust the path planning strategy and output the optimized inspection path.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the data analysis-based intelligent testing method for a monorail inspection robot according to any one of claims 1 to 5 are implemented.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the data analysis-based intelligent testing method for a monorail inspection robot according to any one of claims 1 to 5 are implemented.

Images