Cable tunnel multi-source data fusion inspection early warning method and system based on big data analysis

Abstract

This invention discloses a method and system for multi-source data fusion inspection and early warning of cable tunnels based on big data analysis. The method includes: quality screening of multi-source monitoring data to construct a big data evidence library characterizing the health status of each area using quality functions; dynamic interference removal of inspection images based on this evidence library, extraction of unknown state confidence values ​​for trend analysis, and generation of preliminary early warning results; simplification of the evidence library and training of a lightweight classifier when computational resources are limited, continuing local early warning analysis; pose correction of inspection equipment to construct a static dense point cloud map; fusion of quality functions of each modality and mapping to the point cloud map, outputting an early warning report containing anomaly type, confidence value, and corresponding section. This invention effectively improves the accuracy, real-time performance, and resource adaptability of cable tunnel inspection and early warning.

Description

Technical Field

[0001] This invention relates to the field of cable tunnel inspection and early warning technology, specifically to a cable tunnel multi-source data fusion inspection and early warning method and system based on big data analysis. Background Technology

[0002] As the core channel for underground power transmission in cities, cable tunnels house high-voltage cables that operate under high loads for extended periods. Affected by factors such as humidity, high temperatures, harmful gases, and mechanical vibrations within the tunnels, safety hazards frequently occur, including overheating of cable joints, partial discharge, insulation aging, and deformation of the tunnel structure, seriously threatening the safe operation of the power grid. Therefore, achieving comprehensive, real-time inspection and early warning of the operating status of cable tunnels is of great significance.

[0003] Currently, cable tunnel inspection and early warning technologies have the following main shortcomings:

[0004] First, existing inspection methods mostly rely on single or loosely coupled monitoring means such as fiber optic temperature measurement, partial discharge detection, gas sensors and video surveillance. Each system operates independently, with different data formats, different sampling frequencies and a lack of a unified spatiotemporal reference. This makes it difficult to effectively correlate and integrate multi-source heterogeneous data, and thus cannot form a comprehensive assessment of the overall health status of the tunnel.

[0005] Second, traditional early warning methods often use fixed threshold comparison or univariate time series analysis, which on the one hand makes it difficult to adapt to the dynamic fluctuation characteristics of tunnel environmental parameters, resulting in high false alarm and false alarm rates; on the other hand, they ignore the complementarity and conflict between different modal data, lack effective assessment of data quality and adaptive screening mechanism, and low-quality or failed sensor data are prone to triggering false early warnings.

[0006] Third, with the introduction of high-dimensional information such as inspection images and point cloud data, the resource consumption of early warning algorithms on edge computing nodes has increased dramatically. Existing methods lack an adaptive simplification mechanism for dynamic changes in computing resources, making it difficult to maintain the continuous operation of early warning functions in resource-constrained scenarios.

[0007] Fourth, existing spatial positioning methods mostly rely on the built-in odometers or inertial navigation units of inspection equipment. During long-term operation, the position and orientation drift error accumulates significantly, resulting in large deviations in the marking of abnormal locations. This makes it impossible to accurately map the early warning results to the physical sections of the tunnel, which restricts the rapid response and precise handling by maintenance personnel.

[0008] To address the aforementioned technical problems, this invention provides a method and system for cable tunnel multi-source data fusion inspection and early warning based on big data analysis. Summary of the Invention

[0009] This invention addresses the technical problems in existing cable tunnel inspection and early warning technologies, such as the difficulty in effectively fusing multi-source heterogeneous data, high false alarm and missed alarm rates for fixed threshold warnings, warning interruptions due to limited computing resources, and abnormal positioning deviations caused by pose drift. It provides a cable tunnel multi-source data fusion inspection and early warning method and system based on big data analysis, aiming to achieve high-precision, real-time, and resource-adaptive comprehensive early warning of cable tunnel health status.

[0010] In a first aspect, embodiments of this application provide a method for multi-source data fusion inspection and early warning of cable tunnels based on big data analysis, the method comprising:

[0011] Quality confidence assessment and adaptive filtering were performed on multi-source monitoring data in cable tunnels to obtain an effective monitoring dataset.

[0012] Extract the point and line features of the tunnel wall, cable joints and support structure in the effective monitoring data, and use distributed evidence clustering to construct a big data evidence library that uses a quality function to characterize the health status of each area;

[0013] Based on the big data evidence database, dynamic interference is removed from the real-time inspection images, the unknown state confidence value is extracted for time-series trend analysis, and a preliminary warning result is generated when the warning conditions are met.

[0014] During the execution of early warning analysis, when computing resources are limited, redundant samples in the big data evidence library are removed and key samples at the boundaries of each segment are retained. A simplified evidence library is constructed and a lightweight classifier is trained to continue the local early warning analysis.

[0015] Bundle adjustment and loop closure detection are performed on the motion trajectory of the inspection equipment to eliminate cumulative pose error and construct a static dense point cloud map.

[0016] The modal quality functions corresponding to the early warning results are fused and converted into probability distributions, which are then mapped to the corresponding spatial locations of the static dense point cloud map. An inspection early warning report containing the anomaly type, confidence value, and corresponding tunnel section is output.

[0017] Secondly, embodiments of this application provide a cable tunnel multi-source data fusion inspection and early warning system based on big data analysis, applied to the method described in the first aspect, the system comprising:

[0018] The data acquisition and quality screening module is used to perform quality confidence assessment and adaptive screening of multi-source monitoring data in cable tunnels to obtain an effective monitoring dataset.

[0019] The evidence library construction module is used to extract point and line features of tunnel walls, cable joints and support structures in the effective monitoring data, and to construct a big data evidence library that uses a quality function to characterize the health status of each area using distributed evidence clustering.

[0020] The dynamic interference removal and trend analysis module is used to dynamically remove interference from real-time inspection images based on the big data evidence library, extract the unknown state confidence value for time series trend analysis, and generate preliminary warning results when the warning conditions are met.

[0021] The adaptive simplification module for computing resources is used to remove redundant samples from the big data evidence library and retain key samples at the boundaries of each segment when computing resources are limited during the execution of early warning analysis. It then constructs a simplified evidence library and trains a lightweight classifier to continue executing local early warning analysis.

[0022] The pose correction and map building module is used to perform bundle adjustment and loop closure detection on the motion trajectory of the inspection equipment, eliminate pose accumulation error and build a static dense point cloud map.

[0023] The multimodal fusion and report output module is used to fuse the modal quality functions corresponding to the early warning results and convert them into probability distributions, map them to the corresponding spatial locations of the static dense point cloud map, and output an inspection early warning report containing anomaly type, confidence value and corresponding tunnel section.

[0024] Thirdly, embodiments of this application provide an electronic device, including:

[0025] processor;

[0026] Memory used to store processor-executable instructions;

[0027] The processor is configured to implement the cable tunnel multi-source data fusion inspection and early warning method based on big data analysis as described in the first aspect when executing the instructions.

[0028] Fourthly, embodiments of this application provide a computer-readable storage medium storing a program that instructs a device to execute the cable tunnel multi-source data fusion inspection and early warning method based on big data analysis as described in the first aspect.

[0029] Compared with existing technologies, this invention has the following advantages: By using multi-source data quality confidence assessment and adaptive filtering, low-quality sensor data is effectively eliminated. Combined with a big data evidence library constructed using distributed evidence clustering, unified representation and correlation of multimodal heterogeneous data are achieved, improving the reliability of comprehensive assessment of the overall health status of the tunnel. By utilizing dynamic interference elimination and time-series trend analysis of the confidence value of unknown states, false triggering of early warnings by dynamic interference from personnel and equipment in inspection images is effectively eliminated, reducing the false alarm rate. Simultaneously, potential fault evolution processes are identified in advance through changes in confidence trends, achieving proactive early warning. Through dynamic monitoring of computing resources and adaptive simplification of the evidence library, high-precision global analysis is maintained when resources are abundant, while automatically downgrading to lightweight local early warning when resources are limited, ensuring the continuous operation of the early warning function under different computing conditions and enhancing the engineering applicability of edge deployment. By using bundle adjustment and loop closure detection to correct pose drift errors, and combining static dense point cloud maps with probability distribution spatial mapping fused with multimodal quality functions, a precise correlation between anomaly types and tunnel physical sections is achieved. This provides maintenance personnel with intuitive and accurate positioning information, effectively improving on-site handling efficiency. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of a cable tunnel multi-source data fusion inspection and early warning method based on big data analysis, provided as an embodiment of this application.

[0031] Figure 2 This is an architecture diagram of a cable tunnel multi-source data fusion inspection and early warning system based on big data analysis, provided as an embodiment of this application.

[0032] Figure 3 A schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0033] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0034] It should be noted that in the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0035] Based on the embodiments described in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0036] Example 1

[0037] Figure 1 This is a schematic diagram illustrating a method for multi-source data fusion inspection and early warning of cable tunnels based on big data analysis, provided as an embodiment of this application. Figure 1 As shown, a method for multi-source data fusion inspection and early warning of cable tunnels based on big data analysis includes:

[0038] S1. Perform quality confidence assessment and adaptive filtering on multi-source monitoring data within the cable tunnel to obtain a valid monitoring dataset. Assess the quality of various types of sensor data collected, eliminating invalid or low-confidence data to ensure the reliability of subsequent analysis.

[0039] Specifically, in this embodiment, the process of performing quality confidence assessment and adaptive filtering on multi-source monitoring data within the cable tunnel to obtain an effective monitoring dataset includes:

[0040] S1.1 Modal adaptation processing is performed on the fiber optic temperature measurement data, partial discharge data, gas concentration data, and visual image data within the cable tunnel. In this embodiment, the raw data collected by different types of sensors are first standardized and formatted. Since the data structures output by the fiber optic thermometer, partial discharge sensor, gas detector, and camera group are different, including temperature value sequences, discharge pulse waveforms, gas concentration values, and image pixel matrices, these heterogeneous data need to be converted into a unified representation before direct fusion. This step performs timestamp calibration, unit normalization, and data structure reorganization on each modal data, enabling subsequent analysis algorithms to process all data sources without discrimination, laying the foundation for multi-source fusion.

[0041] S1.2. Robust Principal Component Analysis (PPA) is used to decompose the sensor data of each mode into a low-rank effective signal matrix and a sparse noise matrix. Based on the low-rank effective signal matrix, the quality confidence of each data point is calculated according to the reconstruction error. This step uses the robust principal component analysis algorithm to decompose the adapted modal data. The core idea of ​​this algorithm is to split the original data matrix into two parts: one part represents the regular effective signal reflecting the true physical state of the tunnel, and the other part represents the sparse anomaly components caused by sensor noise or occasional disturbances. After decomposition, the effective signal is used to reconstruct the original data in reverse. The difference between the reconstructed value and the original measurement value is calculated as the quantitative basis for the quality confidence. The smaller the reconstruction error, the more consistent the data point is with the overall pattern, and the higher the confidence; conversely, the larger the reconstruction error, the more likely it is caused by sensor failure or communication interference, and the lower the confidence.

[0042] After decomposing the low-rank effective signal matrix L and the sparse noise matrix S, a quality confidence calculation model based on reconstruction error and data structure regularity is introduced to quantify the reliability of each data point. The specific formula is as follows:

[0043] ,

[0044] in, For the first The quality confidence score for each data point, with a value range of [0,1]. The larger the value, the higher the confidence score. For the first The vector of raw monitoring data points for each mode (such as temperature sequence, partial discharge pulse waveform, etc.); This is the low-rank effective signal matrix obtained from robust principal component analysis decomposition; is the Frobenius norm, used to calculate the sum of squares of the reconstruction error; For sparse noise matrix The variance of the corresponding element reflects the noise level of the current mode; The denominator should be a very small positive number to avoid being zero. For inclusion The rank of a local sub-low-rank matrix reflects the strength of structural regularity in the data. It is obtained by decomposing the modal data using robust principal component analysis toolkits (such as RPCA-toolbox). and ; Calculate the Frobenius norm error between the original data and the low-rank reconstruction result point by point; Statistical analysis The variance σs² is used to extract the local submatrix containing the current data point and calculate its rank. Substitute the data into the formula to calculate each data point. This is used for subsequent multi-level threshold filtering.

[0045] S1.3. Set a multi-level confidence threshold including hard thresholds and dynamic soft thresholds, wherein the dynamic soft threshold is adaptively adjusted based on the sliding window statistical characteristics of historical monitoring data. This step simultaneously sets two different types of judgment thresholds: hard thresholds and dynamic soft thresholds. The hard threshold is a fixed baseline, preset by the system based on historical statistical experience, used to filter out obviously distorted data. The dynamic soft threshold is adjusted in real time according to changes in tunnel operating conditions. Its value depends on the statistical characteristics such as the mean and variance of the monitoring data within the recent sliding time window. When the tunnel environment experiences normal fluctuations, this threshold will adapt accordingly, avoiding misjudging normal operating condition changes as data anomalies. The combined use of the two thresholds makes the data filtering strategy both stable and environmentally adaptive.

[0046] S1.4. Data points with a quality confidence level lower than the hard threshold are identified as invalid data and removed. Data points with a quality confidence level between the hard threshold and the dynamic soft threshold are marked as low-confidence data and assigned a lower weight. Data points with a quality confidence level higher than the dynamic soft threshold are marked as valid data. The data points are then aggregated to form the valid monitoring dataset.

[0047] This step categorizes the data into three quality levels based on a comparison of the confidence level of each data point with the aforementioned two-level thresholds. Data with a confidence level below the hard threshold is deemed completely invalid and directly discarded, not participating in any subsequent analysis. Data with a confidence level between the hard threshold and the dynamic soft threshold is marked as low confidence; the system will assign it a lower weight during subsequent fusion processes to reduce its impact on the final judgment but not completely exclude it, thus preserving potentially usable information. Data with a confidence level above the dynamic soft threshold is considered valid data and receives normal weight for comprehensive analysis. The data from these three levels are then aggregated and organized to form a valid monitoring dataset for subsequent steps, ensuring a minimum data quality while preserving usable monitoring information to the greatest extent possible.

[0048] S2. Extract point and line features from the tunnel walls, cable joints, and support structures in the effective monitoring data. Utilize distributed evidence clustering to construct a big data evidence library that characterizes the health status of each region using a quality function. Extract key structural features from the effective data and establish a knowledge base for the health status of each region through clustering algorithms, providing a standard reference for early warning.

[0049] Specifically, in this embodiment, the step of extracting point and line features of tunnel walls, cable joints, and support structures from the effective monitoring data, and constructing a big data evidence library using distributed evidence clustering to characterize the health status of each region with a quality function, includes:

[0050] S2.1. Perform edge detection and corner extraction on the visual image data in the effective monitoring dataset to obtain point and line features that characterize the tunnel wall contour, cable joint gaps and support geometry; perform grid subdivision on the temperature field and partial discharge signal in the effective monitoring dataset, and extract the temperature rise gradient features and discharge energy features of each grid node as point features.

[0051] In this embodiment, for the visual image data in the effective monitoring dataset, edge detection and corner detection algorithms are used to identify pixels with drastic brightness changes as edges or corners, thereby outlining the continuous contour lines of the tunnel wall, the joints between cable connectors, and the intersections of the support structure's edges, forming structural features expressed in the form of points and lines. Meanwhile, for temperature field data and partial discharge signal data, since they do not inherently possess visual point and line structures, this step divides the tunnel space into regular grid cells. At each grid node, the rate of temperature change and the discharge pulse energy value are calculated, serving as point features for the temperature and discharge dimensions, respectively. After different modalities of data undergo their respective adaptive feature extraction methods, they are uniformly converted into a geometric-physical feature set composed of points and lines, facilitating unified clustering analysis in subsequent steps.

[0052] S2.2. Initialize multiple evidence cluster centers, calculate the Jousselme distance between each data point and each cluster center based on the point-line features to use as the relative distance, and calculate the confidence value of each data point to each cluster center based on the relative distance.

[0053] This step first initializes multiple evidence cluster centers randomly or heuristically in the feature space, each center representing a potential state pattern. Then, it calculates the Jousselme distance between each data point in the feature set and each cluster center. This distance is a similarity measure specifically used within the framework of evidence theory; essentially, it measures the degree of difference between the evidence represented by two basic confidence assignment functions. The smaller the distance, the closer the data point is to the state features of its corresponding cluster center. After obtaining the distances of each data point relative to all cluster centers, this step maps these distance values ​​to confidence values ​​using a confidence assignment function. The magnitude of the confidence value directly reflects the credibility of the data point belonging to each cluster center. The sum of the confidence values ​​of a data point with respect to all cluster centers is 1, forming a complete confidence distribution vector.

[0054] To more accurately measure the difference between the two basic confidence assignment functions in evidence theory, and considering the modal imbalance characteristics of multi-source monitoring data for cable tunnels, a sample-weighted improvement is made to the traditional Jousselme distance. The improved distance calculation model is as follows:

[0055] ,

[0056] Where the weighting matrix The diagonal elements are:

[0057] ,

[0058] in, For the two basic confidence assignment functions The improved Jousselme distance between them is used; the smaller the value, the higher the similarity between them. The basic confidence assignment vectors correspond to the two cluster centers / data points respectively, with a dimension of 2K (K=3, corresponding to the three basic propositions of normal, abnormal and uncertain). for A diagonal weighted matrix is ​​used to balance the contribution weights of different propositions; For the first The weight coefficient of each proposition is determined by the number of samples; the smaller the sample size, the higher the weight, to avoid the minority class being ignored. This represents the total number of samples in the evidence database. For the first The number of samples corresponding to each proposition; This is a smoothing factor.

[0059] The number of samples for normal, abnormal, and uncertain propositions in the statistical evidence corpus. Calculate the corresponding weights respectively Construct a diagonal weighted matrix (The weights of the remaining power set propositions are set to 0); After subtracting the two basic confidence assignment vectors to be compared, the improved Jousselme distance is obtained by calculating the quadratic form according to the formula, which is used for subsequent cluster membership calculation.

[0060] S2.3. Based on the confidence value, the data points are divided into subgroups, and the cluster centers are automatically updated until the convergence condition is met, forming multiple clusters.

[0061] This step assigns each data point to the cluster with the highest confidence value based on the confidence distribution vector of each data point, completing one round of soft partitioning. After partitioning, for each cluster, the weighted average position of all data points within that cluster is recalculated as the new cluster center. A positional shift will occur between the old and new cluster centers. The system determines whether the clustering has converged based on the criterion that the confidence entropy tends to stabilize and the average shift of the cluster center positions between two adjacent iterations is less than a preset convergence threshold. If convergence has not occurred, distance calculation, confidence assignment, and membership partitioning are re-executed based on the new cluster centers. This process is repeated iteratively until the positions of all cluster centers are basically stable and no longer change significantly, ultimately forming multiple clusters that can represent different health status patterns in various regions of the tunnel.

[0062] S2.4 For each cluster, based on the probability distribution of data points within the cluster belonging to normal, abnormal, and uncertain categories, construct a basic confidence assignment function to characterize the health status of the region. Summarize the basic confidence assignment functions corresponding to all clusters to construct the big data evidence library.

[0063] This step involves analyzing the type label distribution of each data point within each converged cluster. Specifically, based on the state reflected by the data points within the cluster, they are categorized into three classes: normal, abnormal, or uncertain. The proportion of each of these three types of samples within the cluster is calculated, and a basic confidence assignment function for that region is constructed based on this proportion. This function is a mathematical tool for expressing uncertain information; it assigns total confidence to the three propositions of normal, abnormal, and uncertain, representing the degree of confidence that the current region is in a normal state, the degree of confidence that it is in an abnormal state, and the degree of confidence that a clear judgment cannot be made due to insufficient information, respectively. After constructing the corresponding basic confidence assignment function for each cluster, all functions are aggregated and stored to form a large-scale evidence library covering the entire cable tunnel. In subsequent early warning analysis, any real-time data can find the most matching health status reference in this evidence library.

[0064] Furthermore, the automatic updating of cluster centers until the convergence condition is met includes:

[0065] S2.3.1 Calculate the confidence entropy value of each cluster in the current iteration, and the average offset of the cluster center positions between two adjacent iterations. In this embodiment, after each round of cluster center updates, the system calculates the confidence entropy value of each cluster. Confidence entropy is a generalization of the concept of information entropy within the framework of evidence theory. It is used to quantify the degree of uncertainty in the affiliation of data points within a cluster. A larger entropy value indicates that the data points within the cluster are more dispersed in their affiliation to each cluster center, and the cluster boundaries are more blurred; a smaller entropy value indicates that the data points within the cluster are more concentrated in their affiliation, and the cluster shape is more compact and stable. Simultaneously, the system also calculates the positional difference between the new cluster center obtained in this iteration and the old cluster center in the previous iteration, and takes the average of the displacements of all cluster centers to obtain the average offset. These two indicators characterize the convergence progress of the clustering process from two dimensions: the cohesive stability of the cluster and the global volatility of the center.

[0066] S2.3.2 When the confidence entropy value tends to stabilize and the average offset is less than the preset convergence threshold, the convergence condition is determined to be met. This step determines whether the clustering process should be terminated based on the comprehensive performance of the above two indicators. When the confidence entropy value no longer shows a significant decrease in multiple iterations and enters a stable range, it indicates that the internal structure of each cluster has tended to stabilize; at the same time, when the average offset of the cluster center position continuously decreases to less than the system's preset convergence threshold, it indicates that the position fluctuation of each cluster center between iterations has become negligible. When both of the above conditions are met simultaneously, the system determines that the clustering has reached a stable state, the iteration process terminates, and the current cluster center position and the final membership partitioning results of all data points are output.

[0067] After the clustering process is completed and each cluster is formed, the system assigns a globally unique code to each cluster as its identity identifier within the entire evidence database. This identifier is not only used to distinguish between clusters, but more importantly, it establishes a mapping relationship between the cluster and the actual physical structure of the tunnel. The area corresponding to each cluster can be found in the tunnel ledger with a matching physical entry; for example, one cluster corresponds to cable connector number 3, and another corresponds to the wall section between ventilation openings 5 ​​and 6. The system establishes a bidirectional index relationship between this structural identifier and the corresponding physical ledger information. This allows the system to not only return the cluster number of the anomaly when an anomaly is identified during the early warning phase, but also directly link it to the specific equipment or structural part corresponding to that cluster, providing accurate asset location information for maintenance personnel to handle the situation on-site.

[0068] S3. Based on the aforementioned big data evidence database, dynamically remove interference from real-time inspection images, extract the uncertainty state confidence value for time-series trend analysis, and generate preliminary warning results when the warning conditions are met. Filter dynamic objects such as personnel and equipment in the inspection images, extract indicators reflecting state uncertainty for time-series analysis, and capture potential abnormal trends.

[0069] Specifically, in this embodiment, the step of dynamically removing interference from real-time inspection images based on the big data evidence database, extracting the ignorance state confidence value for time-series trend analysis, and generating a preliminary early warning result when the early warning conditions are met includes:

[0070] S3.1 Input the real-time inspection image into the pre-trained target detection model to identify and segment dynamic interference targets in the image. The dynamic interference targets include the inspection robot body, mobile work vehicles, and floating foreign objects.

[0071] In this embodiment, when real-time inspection images enter the system, they are first input into a pre-trained target detection model. This model is trained using a large number of image samples labeled with common objects found in various tunnels, and can automatically identify and box-frame regions in the image that belong to dynamic interference targets. The dynamic interference targets of interest in this step mainly include the robotic arm and body of the inspection robot itself, vehicles or personnel temporarily entering the tunnel for maintenance work, and foreign objects such as suspended cables or falling dust clumps that are blown by airflow or vibration. The target detection model outputs precise segmentation masks or bounding boxes for these targets, distinguishing these regions from the static structural regions of the tunnel itself, providing a location basis for subsequent steps to accurately locate the interference regions that need to be removed.

[0072] To quantify the degree of uncertainty within clusters and avoid the problem of traditional entropy values ​​being overly sensitive to small fluctuations, a confidence entropy calculation method with a smoothing factor is introduced. This method, combined with the cluster center offset, serves as a convergence criterion. The specific formula is as follows:

[0073] ,

[0074] The convergence condition is:

[0075] ,

[0076] in, For the first The confidence entropy of a cluster is such that the larger the value, the higher the uncertainty within the cluster. To identify the frame, ={Normal, Abnormal, Uncertain}; For the first In the basic confidence assignment function of a cluster, the proposition The corresponding reliability value; The smoothing coefficient (set to 0.1) is used to avoid... When = 0, the logarithm is meaningless; To identify the number of basic propositions in the framework, K=3 in this invention; For the first During the nth iteration The confidence entropy of each cluster; The confidence entropy convergence threshold (taken as 0.01) can be adjusted according to the complexity of the tunnel environment; This represents the total number of clusters. For the first During the nth iteration The center vector of each cluster; The L2 norm is used to calculate the positional offset of the cluster centers; The cluster center offset convergence threshold is set to 0.05, which can be adjusted according to the feature space scale.

[0077] After each iteration, the confidence entropy with a smoothing factor is calculated for each cluster. Calculate the absolute value of the difference between the confidence entropy of the current iteration and the previous iteration, and determine whether it is less than 1. Simultaneously calculate the average L2 offset of all cluster centers and determine if it is less than 1 / 2. When both conditions are met, the clustering is considered converged, and the final clustering result is output.

[0078] S3.2 Based on the quality function of the corresponding static scene in the big data evidence library, construct the background model of the inspection real-time image, use epipolar geometric constraints to calculate the feature point matching relationship between the inspection real-time image and adjacent frames, remove the dynamic interference features corresponding to the dynamic interference target and the feature point matching relationship, and retain the static effective features characterizing the tunnel wall, cable joint and support structure.

[0079] This step utilizes the quality functions of corresponding static scenes already stored in the big data evidence library as prior knowledge to construct a background model for the current real-time inspection image. This background model reflects the distribution of static structural features that the current observation position should present in the absence of dynamic interference. Based on this, the system uses epipolar geometry constraints to calculate the feature point matching relationship between the current image and adjacent frames. Epipolar geometry is a theory in computer vision that describes the geometric constraint relationship between the projection points of the same 3D point on two different viewpoint imaging planes. Through this constraint, it can determine which feature points conform to the motion law of the static scene between frames and which feature points exhibit abnormal motion that does not conform to this law. The system collectively refers to feature points that fall within the dynamic interference target area and those that exhibit abnormal inter-frame motion as dynamic interference features and removes them. What is retained are those feature points with fixed positions and inter-frame motion laws that conform to the static scene. These static effective features correspond to the fixed texture of the tunnel wall, the stable outline of the cable joint, and the stationary geometry of the support.

[0080] To accurately distinguish between static effective features and dynamic interference features in inspection images, a dual judgment criterion for dynamic interference features is established by combining target detection results and epipolar geometry constraints. The specific formula is as follows:

[0081] ,

[0082] The epipolar constraint residual is:

[0083] ,

[0084] in, This is a set of dynamic interference feature points; For the first frame of the image There are 1 feature points, with coordinates ( ). , ,1)T (homogeneous coordinates); For adjacent frames of images and Homogeneous coordinates of matched feature points; for The corresponding mask area; This is the set of dynamic interference target bounding boxes output by the target detection model in S3.1; The fundamental matrix between the two frames is calculated using epipolar geometric constraints; The epipolar constraint residual threshold (set to 1.5 pixels) can be adjusted according to the camera resolution; For feature points and The larger the value of the epipolar constraint residual, the less it conforms to the motion law of the static scene; For vectors of Components. The ORB feature extraction algorithm is used to obtain matching feature point pairs between the current frame and adjacent frames. Estimate the fundamental matrix F between two frames using the RANSAC algorithm; for each feature point First, determine whether it falls within the masked area of ​​a dynamically interfering target; if so, classify it directly. If it is not within the mask region, calculate its epipolar constraint residual. ,like > If it is, it is determined to be a dynamic interference feature and classified as... The remaining feature points are retained as static effective features for subsequent unknown state reliability extraction.

[0085] S3.3. Based on the retained static valid features, retrieve the most similar cluster in the big data evidence database, and extract the uncertain term representing the unknown state in the basic confidence assignment function corresponding to the cluster as the unknown state confidence value.

[0086] This step involves performing a nearest neighbor search on the static, valid features retained from the previous step within the big data evidence database to find the cluster most similar to this set of static features. After finding a matching cluster, the uncertainty term in the basic confidence assignment function corresponding to that cluster is extracted. In this theoretical framework, this term is called the "ignorance state confidence value," which represents the portion of confidence that the system cannot definitively classify the current region as normal or abnormal based on the current observation information. The magnitude of the ignorance state confidence value directly reflects the degree of matching between the current observation and existing state patterns in the evidence database: a low value indicates a good match between the current state and typical normal or abnormal patterns in the evidence database, meaning the system has sufficient information to make a clear judgment; a high value indicates that the current state does not match any pattern in the evidence database, and the system is in a state of cognitive ambiguity. This ambiguity often foreshadows the emergence of new anomalies or changes in operating conditions.

[0087] S3.4. Smooth the unknown state reliability value with a preset time window and calculate its slope over time. When the unknown state reliability value exceeds the preset reliability threshold and the slope is greater than zero and continues to exceed the set time, it is determined that the warning condition is met, and a preliminary warning result containing the initial estimation information of the abnormal location and the current unknown state reliability value is generated.

[0088] This step performs a cumulative analysis of the extracted unknown state reliability values ​​over time. A high unknown state reliability value at a single moment may be an occasional fluctuation caused by transient noise or brief occlusion, and has no warning value. Therefore, the system smooths and filters the unknown state reliability value sequence with a fixed-length time window to eliminate random jitter, and then calculates the slope of the sequence over time, i.e., the rate of increase of the reliability value. When the smoothed unknown state reliability value exceeds the system's preset reliability threshold, and the rate of increase remains greater than zero for more than the set duration, it indicates that the unknown state reliability value is not an occasional jump at a certain moment, but shows a continuous upward trend over a valid period of time, reflecting that the state of the tunnel area is gradually deviating from all known patterns and entering an unknowable state range. At this time, the system determines that the warning triggering condition is met and generates a preliminary warning result, which includes a rough estimate of the current location of the anomaly and the unknown state reliability value at the current moment, for subsequent steps to perform multimodal verification and precise positioning.

[0089] S4. During the early warning analysis process, when computing resources are limited, redundant samples in the big data evidence base are removed, and key samples at the boundaries of each segment are retained. A simplified evidence base is constructed, and a lightweight classifier is trained to continue local early warning analysis. When device computing power is insufficient, the knowledge base size is automatically compressed, key boundary information is retained, and a simplified model is trained to maintain basic early warning capabilities.

[0090] Specifically, in this embodiment, when computing resources are limited, the process of removing redundant samples from the big data evidence library and retaining key samples at the boundaries of each segment, constructing a simplified evidence library and training a lightweight classifier, and continuing to perform local early warning analysis includes:

[0091] S4.1 The system monitors the CPU utilization, available memory, and remaining battery power of the current device in real time. When any of these indicators falls below a preset resource threshold, the system determines that computing resources are limited. In this embodiment, the system continuously monitors three core resource indicators of the inspected device in the background: CPU utilization reflects the current load on the processor, available memory determines whether large-scale data processing can be accommodated, and remaining battery power limits the device's continuous operating time. The system presets a resource threshold for each indicator. When the real-time monitored CPU utilization remains high, available memory is below the safety margin, or remaining battery power is about to be exhausted, and any of the three indicators reaches its respective warning line, the system determines that the device is currently in a state of limited computing resources. This determination triggers subsequent streamlining processes to prevent the entire warning system from crashing or the task from being interrupted due to resource exhaustion.

[0092] S4.2. Traverse the big data evidence database, calculate the average Jousselme distance between samples within each cluster, and remove samples with an average distance less than a preset density threshold as redundant samples within the cluster. This step involves traversing and filtering all samples in the big data evidence database. Within each cluster, the system calculates the Jousselme distance between any two samples. This distance, within the framework of evidence theory, measures the degree of difference between two pieces of evidence; the smaller the distance, the more similar the state information expressed by the two samples. Subsequently, the system calculates the average Jousselme distance between each sample and all other samples within the same cluster. The smaller the average distance, the more numerous the neighboring samples with highly similar states are clustered around the sample, belonging to a highly dense region within the cluster. The system identifies samples with an average distance less than the preset density threshold as redundant samples within the cluster and deletes them. The state information carried by these samples has been fully covered by other samples within the cluster; removing them does not result in the loss of key state features but effectively reduces the storage size of the evidence database and the overhead of subsequent calculations.

[0093] S4.3 Identify the decision boundaries of each cluster, and determine the samples whose distance from the cluster center is within the boundary interval as key samples of each segment boundary and retain them. The boundary interval is dynamically defined by the minimum Jousselme distance between the cluster and other adjacent clusters.

[0094] While eliminating redundant samples, this step identifies and retains the class of samples most critical to the overall classification decision. Each cluster occupies a certain region in the feature space, and the boundary between different clusters is called the decision boundary. This step defines a boundary interval for each cluster, the range of which is dynamically determined by the minimum Jousselme distance between the cluster and its nearest neighboring clusters. Specifically, the boundary interval is defined as the region within a certain proportional range of distance from the cluster center—a middle zone that is neither too close nor too far from the cluster center. Samples located within this interval are the supporting samples on the decision boundary; they carry the key discriminative information that distinguishes the state differences between this cluster and its neighboring clusters. The system retains the key samples at each boundary segment, ensuring that the streamlined evidence base maintains accurate classification boundaries while compressing its size.

[0095] S4.4. Summarize all retained key boundary samples to construct the simplified evidence library. This step summarizes and organizes all remaining key boundary samples after the above redundancy removal and boundary retention to form a simplified evidence library. The number of samples in this evidence library is much smaller than the original big data evidence library, but it retains the key boundary information between each cluster in the state space distribution, which is equivalent to maintaining the maximum distinguishing ability of the health status of each section of the tunnel with minimal sample cost.

[0096] S4.5 Based on the simplified evidence base, the reconstruction coefficients of each sample in the feature space are solved by sparse reconstruction optimization to construct a sparse reconstruction evidence K-nearest neighbor classifier. The sparse reconstruction evidence K-nearest neighbor classifier is used to perform localized inference on real-time monitoring data and continue to perform early warning analysis.

[0097] This step trains a lightweight sparse reconstruction evidence K-nearest neighbor classifier based on a simplified evidence base. The classifier works by using a sparse reconstruction optimization algorithm to solve for a set of reconstruction coefficients for each sample in the simplified evidence base, allowing newly input real-time monitoring data to be approximated by a linear combination of the few most relevant samples in the simplified base. This mechanism eliminates the need to load the complete original evidence base during the classification phase; sparse retrieval within the simplified base is sufficient to determine the category of new samples. The system utilizes this classifier for localized inference on real-time monitoring data, maintaining early warning analysis with low computational overhead even under resource constraints, until resources are restored before switching back to a high-precision analysis mode based on the complete evidence base.

[0098] To achieve lightweight early warning inference in computationally limited scenarios, a sparse reconstruction evidence K-nearest neighbor classifier is constructed based on a simplified evidence base. Its core principle is to minimize the reconstruction error to solve for the sparse representation coefficients of the samples, thereby obtaining the classification decision result. The specific formula is as follows:

[0099] ,

[0100] in, To input real-time monitoring samples The prediction categories (1=normal, 2=abnormal, 3=uncertain); To streamline the evidence database In the basic confidence assignment function of the boundary key samples, the first... The reliability values ​​of the basic propositions; The real-time monitoring feature vector to be classified; To streamline the evidence database Feature vectors of key boundary samples; To reduce the total number of samples in the evidence database; For the first The sparse reconstruction coefficients corresponding to each sample reflect the similarity between the sample to be classified and the sample itself. is the sparsity regularization coefficient (taken as 0.01), which controls the sparsity of the coefficients; is the L1 norm of the coefficient vector, used to constrain the sparsity of the reconstructed coefficients.

[0101] The feature vectors of all samples in the simplified evidence base are used to form a dictionary matrix. For the input samples to be classified The sparse reconstruction coefficients are solved by LASSO regression. Minimize the sum of the reconstruction error and the L1 regularization term for each category; Calculate the weighted sum of the reconstruction coefficients and corresponding reliability values ​​for all samples. The category with the largest weighted sum is selected as the prediction result to achieve lightweight local early warning inference.

[0102] S5. Perform bundle adjustment and loop closure detection on the motion trajectory of the inspection equipment to eliminate cumulative pose errors and construct a static dense point cloud map. Correct positioning deviations through multi-frame constraints and loop closure detection, establish a high-precision three-dimensional spatial model, and eliminate cumulative errors generated during long-term operation.

[0103] In this embodiment, during the long-term movement of the inspection equipment within the tunnel, its inertial navigation unit and odometer gradually accumulate pose errors due to factors such as integral drift, wheel slippage, or uneven ground. This causes the equipment's position estimate to deviate increasingly from the true value over time. To eliminate this accumulated error, this step employs a combination of bundle adjustment and loop closure detection to optimize the overall motion trajectory.

[0104] Bundle adjustment is a global optimization algorithm that simultaneously optimizes device pose and 3D spatial point coordinates. Its basic idea is to use image feature points captured by the device at different locations as observation constraints. By minimizing the sum of reprojection errors of all observation points, the device pose parameters and the positions of all 3D spatial points corresponding to each frame of the image are jointly adjusted. During the optimization process, bundle adjustment incorporates all observation relationships into a large system of equations for solving, ensuring that each pose and each point is constrained by multiple adjacent observations. This evenly distributes local measurement errors across the entire trajectory range, significantly suppressing the cumulative effect of drift.

[0105] Building upon this, the system further utilizes a loop closure detection mechanism to eliminate long-term drift. Loop closure detection refers to the process where, after traveling a certain distance within the tunnel, the inspection equipment returns to a previously visited position. The system identifies the similarity between the current image and a historical frame through visual feature matching, thus establishing a closed-loop constraint. This constraint forcibly associates the current pose with historical poses, forming a closed loop. Based on this constraint, the system performs a comprehensive readjustment of the poses at all intermediate moments within the loop, ensuring seamless spatial connection between the start and end points of the closed loop. This fundamentally eliminates the positional misalignment between the start and end points caused by drift accumulation, i.e., the so-called drift error.

[0106] After dual optimization using bundle adjustment and loop closure detection, each image frame stored in the system corresponds to a precisely corrected device pose. Based on these precisely located image sequences, the system uses triangulation to reconstruct the 3D spatial coordinates of the pixel coordinates corresponding to the same 3D spatial point in each frame, thereby generating dense 3D position information for each point. All spatial points are aggregated to form a static dense point cloud map covering the entire tunnel wall, cable joints, and pipe gallery structure. Each point in this map has precise 3D spatial coordinates, providing a high-precision spatial reference for the accurate location of subsequent anomaly areas.

[0107] S6. The modal quality functions corresponding to the early warning results are fused and converted into probability distributions, mapped to the corresponding spatial locations of the static dense point cloud map, and an inspection early warning report containing the anomaly type, confidence value, and corresponding tunnel section is output. Various monitoring evidences are cross-verified and converted into risk probabilities, then superimposed onto the three-dimensional spatial model to generate early warning conclusions with specific section coordinates.

[0108] Specifically, in this embodiment, fusing the modal quality functions corresponding to the early warning results and converting them into probability distributions, and mapping them to the corresponding spatial locations of the static dense point cloud map, includes:

[0109] S6.1 Extract the basic confidence assignment functions corresponding to the temperature monitoring data, partial discharge data, gas concentration data and visual image data within the same tunnel section at the same time stamp, and use them as the quality functions of each mode to be fused.

[0110] In this embodiment, after the system generates a preliminary warning result for a certain tunnel section at a certain time, it enters the multimodal fusion stage. The system first uses that time as the time reference and the tunnel section corresponding to the warning result as the spatial reference. It then extracts the basic confidence assignment functions corresponding to temperature monitoring data, partial discharge data, gas concentration data, and visual image data at the same time stamp and spatial location from the big data evidence database. These four basic confidence assignment functions represent the judgment opinions of their respective modes on the current region's state: the temperature mode reflects the thermal state, the partial discharge mode reflects the insulation state, the gas mode reflects the leakage state, and the image mode reflects the appearance and structural state. Each of the four modes independently observes different aspects of the same region, each carrying evidence regarding whether the region is normal or abnormal, as input for fusion.

[0111] S6.2. The Dempster combination rule is used to perform orthogonal summation on the quality functions of each modality to obtain the fused global quality function. This step employs the Dempster combination rule from evidence theory to perform orthogonal summation on the basic confidence assignment functions of the four modalities. The core idea of ​​this combination rule is to orthogonally synthesize the confidence assignments from multiple independent information sources, and obtain a new global quality function through cross-multiplication and normalization. The essence of the orthogonal summation operation is: if two pieces of evidence support the same proposition, their joint support effect will reinforce each other; if two pieces of evidence provide contradictory judgments on the same proposition, the conflicting parts will be proportionally allocated, and the normalization process will readjust the total confidence to 1. After multiple pairwise combination iterations, the evidence from the four modalities is merged into a unified global quality function. This function integrates the judgment information from all available sensors, and its reliability and confidence level are generally higher than the individual judgments of any single modality.

[0112] To address the conflict issue in multimodal data fusion, an adaptive conflict weight is introduced based on the traditional Dempster combination rule to reduce the interference of anomalous modalities on the fusion result. The improved fusion formula is as follows:

[0113] ,

[0114] The conflict factor and adaptive weight are respectively:

[0115] ,

[0116] ,

[0117] in, For the global quality function after fusion, the proposition The corresponding reliability value; These correspond to propositions in the basic confidence assignment functions for the four modalities of temperature, partial emission, gas, and vision, respectively. For the first Adaptive weights for each modality ( =1,2,3,4), the greater the conflict, the smaller the weight; For the first The proposition in the basic confidence assignment function of each modality The reliability value; The improved conflict factor reflects the overall degree of conflict among multimodal evidence; This is the weight adjustment coefficient (set to 2), which controls the strength of the influence of distance on weight. For the first Improved Jousselme distance between observational evidence for each modality and evidence for the corresponding cluster; To identify the frame {normal, abnormal, uncertain}.

[0118] For each modality's observational evidence, calculate its improved Jousselme distance to the evidence of the corresponding cluster. Substituting the values ​​into the formula yields the adaptive weights for each mode. ; Calculate the improved conflict factor ,like If the value is greater than 0.8, the weight of the modality with the greatest conflict will be automatically reduced to 50% of its original value, and the modality will be recalculated. The evidence from the four modalities is orthogonally summed according to the improved Dempster combination rule to obtain the fused global quality function, which is used for subsequent probability transformation and report generation.

[0119] S6.3. Using the Pignistic probability transformation method, the global quality function is converted into a probability distribution that includes normal, abnormal, and unknown states. This step transforms the fused global quality function into an intuitive probability distribution form. The basic confidence allocation function in evidence theory allows confidence to be assigned to the power set of propositions; for example, it allows a portion of confidence to be assigned to the composite proposition "normal or abnormal." This expression is advantageous in the evidence reasoning stage, but it is inconvenient to connect with conventional probabilistic understanding in the final decision-making process. This step uses the Pignistic probability transformation method to evenly distribute the confidence of all composite propositions contained in the global quality function according to the number of basic propositions they contain. This decomposes the confidence that might have been distributed on composite propositions and distributes it across the three basic propositions: normal, abnormal, and unknown. The final result is a probability distribution whose sum is 1, representing the subjective probability estimates of the current region being in a normal state, an abnormal state, and an uncertain state, respectively.

[0120] S6.4. Based on the pose information of the inspection equipment in the static dense point cloud map, determine the physical space coordinates corresponding to the early warning result, and associate and bind the probability distribution with the physical space coordinates. This step is based on the completed static dense point cloud map to find the precise pose information of the inspection equipment at that moment. The static dense point cloud map is constructed by combining the pose information corrected by the bundle adjustment and loop closure detection in the early stage with the environmental point cloud collected by the lidar or depth camera. Each point in the point cloud corresponds to a three-dimensional coordinate in the physical space of the tunnel. According to the pose information of the inspection equipment and the installation parameters of the sensors, the system calculates the physical space coordinates of the abnormal position corresponding to the preliminary early warning result in the point cloud map through coordinate transformation, and then associates and binds the probability distribution obtained in S5.3 with the coordinates, so that the coordinate position carries the probability estimate of three states at the same time, realizing the mapping from abstract data to specific spatial location.

[0121] S6.5. According to the physical segmentation rules of cable tunnels, determine the tunnel segment to which the physical spatial coordinates belong, and generate an inspection and early warning report containing the anomaly type, confidence value, and the tunnel segment. This step determines the segment affiliation of the physical spatial coordinates obtained in S5.4 according to the predefined physical segmentation rules of cable tunnels. In operation and maintenance management, cable tunnels are usually segmented and numbered according to joint numbers, ventilation opening intervals, fire compartments, or mileage markers, etc., and each segment has a clear start and end coordinate range. The system compares the anomaly coordinates with the coordinate range of each segment one by one to determine the specific tunnel segment to which the coordinates belong. Subsequently, the final inspection and early warning report is generated, which clearly marks the anomaly type, confidence value, and corresponding tunnel segment number, enabling operation and maintenance personnel to directly reach the problem location based on the report and improve on-site handling efficiency.

[0122] In this embodiment, before formally performing Dempster fusion, the system first performs preprocessing operations for evidence alignment and conflict removal. Specifically, the system uses the same timestamp as the time reference and the spatial grid in the static dense point cloud map as the spatial reference to uniformly interpolate and resample temperature monitoring data, partial discharge data, gas concentration data, and visual image data to the same spatial resolution, generating multimodal observation vectors corresponding to the spatial grid one-to-one, ensuring that the data of each modality are comparable in the spatiotemporal dimensions during subsequent fusion.

[0123] Subsequently, the system calculates the Jousselme distance between each modality's observation vector and its corresponding cluster in S2 to quantify the degree of agreement between each modality's evidence and existing state patterns. If the Jousselme distance between a modality's observation vector and its corresponding cluster exceeds a system-preset conflict threshold, it indicates a significant deviation between the observation results of that modality and existing knowledge in the evidence base, and the modality's data is determined to be conflicting evidence.

[0124] For modal data identified as conflicting evidence, the system further performs a consistency check, examining the degree of deviation between this modal evidence and the majority of other modal evidence. If the deviation exceeds a set proportion, it indicates a systematic contradiction between the observations of this modality and those of other modalities, likely caused by factors such as sensor malfunction, communication interference, or sudden environmental changes. In this case, the system reduces the weight of this conflicting evidence in subsequent Dempster fusion, or removes it directly in extreme deviation cases, to avoid a single anomalous modality misleading the overall fusion result, thereby ensuring the consistency and reliability of the quality functions of all modalities participating in the fusion at the evidence level.

[0125] In this embodiment, the entire inspection and early warning system is deployed in a multi-node inspection scenario using an architecture that combines centralized training and distributed execution. The cloud server serves as the centralized training center, utilizing historical multi-source monitoring data accumulated from various edge nodes to offline train a distributed evidence clustering model and a sparse reconstructed evidence K-nearest neighbor classifier, generating unified model parameters. These model parameters cover core configurations such as the initial cluster center positions, the classifier's reconstruction coefficient matrix, and the fusion weights of each modality's quality function. These parameters are then distributed to the edge inspection nodes deployed in various tunnel sections via the communication network.

[0126] After receiving unified model parameters, each edge inspection node independently runs the early warning analysis process locally, including dynamic interference removal, time-series trend analysis of unknown state reliability values, and local early warning analysis. When the edge node has sufficient computing resources, it uploads the updated cluster center locations and sample reliability distributions from the local process to the cloud server as the data source for global model updates.

[0127] After receiving incremental information uploaded by each edge node, the cloud server updates the global big data evidence base using an experience replay mechanism. The core idea of ​​this mechanism is to not only use the newly uploaded samples during the update but also randomly select a certain proportion of old samples from the historical experience pool for retraining. This balances the impact of new and old data on the model, preventing the model from becoming overly biased towards recently received data and forgetting previously learned knowledge. After the cloud completes the global evidence base update, the latest global model parameters are synchronized to each edge node, enabling all nodes to share the global learning results and achieving collaborative optimization and resource sharing among multiple nodes.

[0128] Example 2

[0129] like Figure 2 As shown in the diagram, this application provides an architecture diagram of a cable tunnel multi-source data fusion inspection and early warning system based on big data analysis, which is applied to the cable tunnel multi-source data fusion inspection and early warning method based on big data analysis as described in Embodiment 1, including:

[0130] The data acquisition and quality screening module 210 is used to perform quality confidence assessment and adaptive screening of multi-source monitoring data in cable tunnels to obtain an effective monitoring dataset.

[0131] The evidence library construction module 220 is used to extract the point and line features of the tunnel wall, cable joints and support structures in the effective monitoring data, and to construct a big data evidence library that uses a quality function to characterize the health status of each area using distributed evidence clustering.

[0132] The dynamic interference removal and trend analysis module 230 is used to dynamically remove interference from real-time inspection images based on the big data evidence library, extract the unknown state confidence value for time series trend analysis, and generate preliminary warning results when the warning conditions are met.

[0133] The computational resource adaptive simplification module 240 is used to remove redundant samples from the big data evidence library and retain key samples at the boundaries of each segment when computational resources are limited during the execution of early warning analysis, to build a simplified evidence library and train a lightweight classifier, and to continue to execute local early warning analysis.

[0134] The pose correction and map building module 250 is used to perform bundle adjustment and loop closure detection on the motion trajectory of the inspection equipment, eliminate pose accumulation errors, and build a static dense point cloud map.

[0135] The multimodal fusion and report output module 260 is used to fuse the modal quality functions corresponding to the early warning results and convert them into probability distributions, map them to the corresponding spatial locations of the static dense point cloud map, and output an inspection early warning report containing anomaly type, confidence value and corresponding tunnel section.

[0136] Figure 3 This is an electronic device provided in one embodiment of this application. For example... Figure 3 As shown, the electronic device includes at least the following components: processor 301 and memory 300, communication interface 303, and bus 302.

[0137] In this embodiment of the application, memory 300 is used to store executable instructions of processor 301, which, when configured to execute instructions, implements the method as described in the first aspect.

[0138] In embodiments of this application, a computer-readable storage medium includes instructions that instruct a device to perform the method as described in the first aspect. For example, the instructions instruct the device to perform... Figure 1 The method is shown in the process steps.

[0139] In one embodiment of this application, the program operating in the electronic device may be a program that controls a central processing unit (CPU) or similar device to achieve the functions of the above-described embodiments of the present invention (a program that enables the computer to function). Information processed by these systems is then temporarily stored in random access memory (RAM) during processing, and subsequently stored in various ROMs such as read-only memory (FlashROM) and hard disk drives (HDDs), and read, corrected, and written by the CPU as needed.

[0140] It should be noted that a portion of the electronic device described above can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.

[0141] It should be noted that the computer mentioned here refers to a computer built into an electronic device, employing hardware including an operating system and peripheral devices. Furthermore, computer-readable recording media refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage systems such as hard drives built into the computer.

[0142] Furthermore, computer-readable recording media can include: media that dynamically stores programs for short periods of time, such as communication lines used when transmitting programs via networks like the Internet or communication lines like telephone lines; and media that store programs for fixed periods of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining them with programs already recorded in the computer.

[0143] Furthermore, the electronic device in the above embodiments can also be implemented as an assembly (system group) composed of multiple systems. Each system constituting the system group can possess some or all of the functions or functional blocks of the electronic device in the above embodiments. As a system group, it is sufficient to have all the functions or functional blocks of the electronic device.

[0144] Those skilled in the art should recognize that the above embodiments are only used to illustrate this application and are not intended to limit this application. Any appropriate changes and variations made to the above embodiments within the essential spirit and scope of this application fall within the scope of protection claimed in this application.

Claims

1. A method for multi-source data fusion inspection and early warning of cable tunnels based on big data analysis, characterized in that, Includes the following steps: Quality confidence assessment and adaptive filtering were performed on multi-source monitoring data in cable tunnels to obtain an effective monitoring dataset. Extract the point and line features of the tunnel wall, cable joints and support structure in the effective monitoring data, and use distributed evidence clustering to construct a big data evidence library that uses a quality function to characterize the health status of each area; Based on the big data evidence database, dynamic interference is removed from the real-time inspection images, the unknown state confidence value is extracted for time-series trend analysis, and a preliminary warning result is generated when the warning conditions are met. During the execution of early warning analysis, when computing resources are limited, redundant samples in the big data evidence library are removed and key samples at the boundaries of each segment are retained. A simplified evidence library is constructed and a lightweight classifier is trained to continue the local early warning analysis. Bundle adjustment and loop closure detection are performed on the motion trajectory of the inspection equipment to eliminate cumulative pose error and construct a static dense point cloud map. The modal quality functions corresponding to the early warning results are fused and converted into probability distributions, which are then mapped to the corresponding spatial locations of the static dense point cloud map. An inspection early warning report containing the anomaly type, confidence value, and corresponding tunnel section is output.

2. The method according to claim 1, characterized in that, The process involves quality confidence assessment and adaptive filtering of multi-source monitoring data within the cable tunnel to obtain an effective monitoring dataset, including: Modal adaptation processing was performed on fiber optic temperature measurement data, partial discharge data, gas concentration data, and visual image data within the cable tunnel. Robust principal component analysis is used to decompose the sensor data of each mode into a low-rank effective signal matrix and a sparse noise matrix. The quality confidence of each data point based on the reconstruction error is calculated according to the low-rank effective signal matrix. A multi-level confidence threshold including a hard threshold and a dynamic soft threshold is set, wherein the dynamic soft threshold is adaptively adjusted according to the sliding window statistical characteristics of historical monitoring data; Data points with a quality confidence level below the hard threshold are identified as invalid data and removed. Data points with a quality confidence level between the hard threshold and the dynamic soft threshold are marked as low-confidence data and assigned a lower weight. Data points with a quality confidence level above the dynamic soft threshold are marked as valid data. The data points are then aggregated to form the valid monitoring dataset.

3. The method according to claim 1, characterized in that, The step involves extracting point and line features from the tunnel walls, cable joints, and support structures in the effective monitoring dataset, and constructing a big data evidence library using distributed evidence clustering to characterize the health status of each region with a quality function. This includes: Edge detection and corner extraction are performed on the visual image data in the effective monitoring dataset to obtain point and line features that characterize the tunnel wall contour, cable joint gaps and support geometry; the temperature field and partial discharge signal in the effective monitoring dataset are meshed, and the temperature rise gradient features and discharge energy features of each mesh node are extracted as point features. Multiple evidence cluster centers are initialized. The Jousselme distance between each data point and each cluster center is calculated based on the point-line features to serve as the relative distance. The confidence value of each data point to each cluster center is then calculated based on the relative distance. Based on the confidence value, each data point is assigned a membership, and the cluster centers are automatically updated until the convergence condition is met, forming multiple clusters. For each cluster, a basic confidence assignment function is constructed to characterize the health status of the region based on the probability distribution of data points within the cluster belonging to normal, abnormal, and uncertain categories. The basic confidence assignment functions corresponding to all clusters are then aggregated to construct the big data evidence library.

4. The method according to claim 3, characterized in that, The automatic updating of cluster centers until the convergence condition is met includes: Calculate the confidence entropy value of each cluster at the current iteration number, and the average offset of the cluster center position between two adjacent iterations; When the confidence entropy value tends to stabilize and the average offset is less than the preset convergence threshold, it is determined that the convergence condition is met. It also includes: configuring a unique structural identifier for each cluster, and establishing an index relationship between the structural identifier and the physical ledger information of the tunnel wall, cable joint or support structure.

5. The method according to claim 1, characterized in that, The process of dynamically removing interference from real-time inspection images based on the big data evidence database, extracting the ignorance state reliability value for time-series trend analysis, and generating preliminary early warning results when the early warning conditions are met specifically includes: The real-time inspection images are input into a pre-trained target detection model to identify and segment dynamic interference targets in the images. The dynamic interference targets include the inspection robot body, mobile work vehicles, and floating foreign objects. Based on the quality function of the corresponding static scene in the big data evidence library, a background model of the real-time inspection image is constructed. The feature point matching relationship between the real-time inspection image and adjacent frames is calculated using epipolar geometric constraints. Dynamic interference features corresponding to the dynamic interference target and the feature point matching relationship are removed, while static effective features representing the tunnel wall, cable joint and support structure are retained. Based on the retained static valid features, the most similar cluster is retrieved in the big data evidence database, and the uncertain term representing the unknown state in the basic confidence assignment function corresponding to the cluster is extracted as the unknown state confidence value. The unknown state reliability value is smoothed using a preset time window, and its slope over time is calculated. When the unknown state reliability value exceeds a preset reliability threshold and the slope is greater than zero and continues for a set duration, the warning condition is determined to be met, and a preliminary warning result containing the initial estimate of the abnormal location and the current unknown state reliability value is generated.

6. The method according to claim 1, characterized in that, When computing resources are limited, redundant samples in the big data evidence library are removed, while key samples at the boundaries of each segment are retained. A simplified evidence library is constructed, and a lightweight classifier is trained. Local early warning analysis continues, including: Real-time monitoring of the device's CPU utilization, available memory, and remaining power; when any of these indicators falls below the preset resource threshold, the device is deemed to have limited computing resources. Traverse the big data evidence database, calculate the average Jousselme distance between samples within each cluster, and identify samples whose average distance is less than a preset density threshold as redundant samples within the cluster and remove them. The decision boundaries of each cluster are identified, and samples whose distance from the cluster center is within the boundary interval are identified as key samples of each segment boundary and retained. The boundary interval is dynamically defined by the minimum Jousselme distance between the cluster and other adjacent clusters. Summarize all retained boundary key samples to construct the simplified evidence library; Based on the simplified evidence base, the reconstruction coefficients of each sample in the feature space are solved by sparse reconstruction optimization, and a sparse reconstruction evidence K-nearest neighbor classifier is constructed. The sparse reconstruction evidence K-nearest neighbor classifier is used to perform localized inference on real-time monitoring data and continue to perform early warning analysis.

7. The method according to claim 2, characterized in that, The step of fusing the modal quality functions corresponding to the early warning results and converting them into probability distributions, and mapping them to the corresponding spatial locations of the static dense point cloud map, includes: Extract the basic confidence assignment functions corresponding to temperature monitoring data, partial discharge data, gas concentration data and visual image data within the same tunnel section at the same time stamp, and use them as the quality functions of each mode to be fused. The Dempster combination rule is used to perform orthogonal summation on the modal mass functions to obtain the fused global mass function. The global quality function is transformed into a probability distribution that includes normal, abnormal, and unknown states using the Pignistic probability transformation method. Based on the pose information of the inspection equipment in the static dense point cloud map, the physical space coordinates corresponding to the early warning result are determined, and the probability distribution is associated and bound with the physical space coordinates. According to the physical segmentation rules of cable tunnels, the tunnel segment to which the physical spatial coordinates belong is determined, and an inspection and early warning report containing the anomaly type, confidence value, and the tunnel segment is generated.

8. The method according to claim 7, characterized in that, Before performing orthogonal summation on the modal quality functions using the Dempster combination rule, a multimodal evidence alignment screening step is also included: Temperature monitoring data, partial discharge data, gas concentration data, and visual image data at the same time stamp are spatiotemporally aligned, and various modal data are interpolated and resampled based on the spatial grid in the static dense point cloud map to generate multimodal observation vectors with consistent spatial resolution. Calculate the Jousselme distance between each modal observation vector and its corresponding cluster, and determine the modal data with a Jousselme distance greater than a preset conflict threshold as conflict evidence; The conflicting evidence is subjected to a consistency test. If the degree of deviation between the conflicting evidence and the other majority modal evidence exceeds a set proportion, the weight of the conflicting evidence in the subsequent Dempster fusion is reduced or it is removed, thereby achieving the alignment and screening of the multimodal quality function.

9. The method according to claim 1, characterized in that, The method is implemented in a multi-node inspection system using a centralized training and distributed execution architecture. The cloud server uses historical multi-source monitoring data to train the distributed evidence clustering model and the sparse reconstructed evidence K-nearest neighbor classifier offline, generating unified model parameters and distributing them to each edge inspection node. Each edge inspection node independently performs dynamic interference removal, time-series trend analysis, and local early warning analysis based on locally collected real-time monitoring data, and uploads the locally updated cluster centers and sample reliability distribution to the cloud server when computing resources are sufficient. The cloud server updates the global big data evidence database based on the information uploaded by each edge inspection node using an experience replay mechanism, and synchronizes the updated global model parameters to each edge inspection node, thereby achieving multi-node collaborative optimization and resource sharing.

10. A cable tunnel multi-source data fusion inspection and early warning system based on big data analysis, applied to the method described in any one of claims 1 to 9, characterized in that, The system includes: The data acquisition and quality screening module is used to perform quality confidence assessment and adaptive screening of multi-source monitoring data in cable tunnels to obtain an effective monitoring dataset. The evidence library construction module is used to extract point and line features of tunnel walls, cable joints and support structures in the effective monitoring data, and to construct a big data evidence library that uses a quality function to characterize the health status of each area using distributed evidence clustering. The dynamic interference removal and trend analysis module is used to dynamically remove interference from real-time inspection images based on the big data evidence library, extract the unknown state confidence value for time series trend analysis, and generate preliminary warning results when the warning conditions are met. The adaptive simplification module for computing resources is used to remove redundant samples from the big data evidence library and retain key samples at the boundaries of each segment when computing resources are limited during the execution of early warning analysis. It then constructs a simplified evidence library and trains a lightweight classifier to continue executing local early warning analysis. The pose correction and map building module is used to perform bundle adjustment and loop closure detection on the motion trajectory of the inspection equipment, eliminate pose accumulation error and build a static dense point cloud map. The multimodal fusion and report output module is used to fuse the modal quality functions corresponding to the early warning results and convert them into probability distributions, map them to the corresponding spatial locations of the static dense point cloud map, and output an inspection early warning report containing anomaly type, confidence value and corresponding tunnel section.

Images