A Data Analysis Method for Inspection Robots Based on Digital Twins

Abstract

This invention discloses a data analysis method for inspection robots based on digital twins, relating to the fields of artificial intelligence and robotics. The method includes: S100, constructing a time-axis aligned model of the equipment to be inspected and the inspection robot based on 3D modeling technology, and establishing a standardized interaction interface for multi-source heterogeneous data; S200, constructing a four-layer structured digital twin evaluation system covering geometry, physics, behavior, and rules based on the time-axis aligned model; S300, defining physical perception nodes, virtual mapping nodes, and state evolution nodes, while introducing environmental interference factors, constructing a real-time mapping logic graph, and marking data interaction paths in the graph. By constructing digital twins of physical equipment and inspection robots, real-time mapping between physical and virtual spaces, multi-source heterogeneous data fusion analysis, accurate fault diagnosis and location, predictive maintenance, and dynamic optimization of inspection tasks are achieved, significantly improving the intelligence level of inspections and equipment operation and maintenance efficiency.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and robotics, and in particular to a method for analyzing inspection data of a maintenance robot based on digital twins. Background Technology

[0002] With the deepening of the industrial internet and intelligent manufacturing, maintenance robots, sensor networks and industrial IoT platforms have been widely used in the field of equipment operation and maintenance. These systems record a full range of inspection activities, from visual images and infrared thermal imaging to vibration, temperature and humidity, generating big data on inspections covering the entire life cycle of equipment, laying the foundation for data-driven equipment status perception and intelligent operation and maintenance.

[0003] Among them, the inspection analysis based on multi-source heterogeneous data aims to build a logical relationship between multi-source sensing data and equipment operating status by deeply mining data from various dimensions such as vision, thermal sensing, and sound. Through comprehensive evaluation of equipment feature extraction, anomaly detection, and fault evolution patterns, it provides operation and maintenance personnel with accurate equipment diagnosis conclusions and maintenance suggestions.

[0004] However, existing inspection data analysis technologies suffer from several drawbacks: First, insufficient integration of multi-source data. Different types of inspection data are often stored in heterogeneous systems, lacking a unified data model and interaction standard, making cross-device and cross-system data fusion analysis difficult. Second, weak real-time response and early warning capabilities. Existing methods rely heavily on post-event offline analysis, failing to achieve real-time processing of inspection data, resulting in delayed fault warnings and increasing the risk of equipment downtime. Third, low fault diagnosis and location accuracy. Traditional analysis methods rely heavily on single sensor data and expert experience, making it difficult to accurately identify early, subtle faults and complex coupled faults, and lacking intuitive visualization methods to assist in fault location. Finally, a lack of virtual verification and prediction capabilities. Due to the inability to construct a real-time mapping from physical equipment to virtual space, existing systems struggle to simulate fault evolution in a virtual environment, limiting the accuracy of maintenance plan simulations and remaining service life predictions. Therefore, to address the problems of data fusion difficulties, insufficient real-time performance, low diagnostic accuracy, and lack of prediction and simulation capabilities in traditional inspection data analysis technologies, a digital twin-based inspection robot inspection data analysis method is urgently needed. Summary of the Invention

[0005] To address the problems of data fusion difficulties, insufficient real-time performance, low diagnostic accuracy, and lack of predictive and pre-simulation capabilities in traditional inspection data analysis technologies, this invention provides a digital twin-based inspection robot inspection data analysis method. By constructing a digital twin of physical equipment and the inspection robot, it achieves real-time mapping between physical and virtual spaces, multi-source heterogeneous data fusion analysis, accurate fault diagnosis and location, predictive maintenance, and dynamic optimization of inspection tasks, significantly improving the level of inspection intelligence and equipment operation and maintenance efficiency.

[0006] In a first aspect, the present invention provides a method for analyzing inspection data of a maintenance robot based on digital twins, the method comprising: S100. Based on 3D modeling technology, construct a time axis alignment model of the equipment to be inspected and the inspection robot, and establish a standardized interaction interface for multi-source heterogeneous data. S200, Based on the time axis alignment model, construct a four-layer structured digital twin evaluation system covering geometry, physics, behavior, and rules; S300 defines physical sensing nodes, virtual mapping nodes, and state evolution nodes, while introducing environmental interference factors to construct a real-time mapping logic graph and marking data interaction paths in the graph; S400 uses multi-sensor weighted fusion, temporal residual analysis and feature space mapping methods to identify and quantify the correlation effect between multi-source inspection data and equipment operating status. S500, based on the SHAP additive interpretation method, constructs a state-fault dual-dimensional interpretable diagnostic model, decomposes the fault diagnosis results into various sensor feature indicators, and generates diagnostic output including component contribution ranking, local anomaly explanation and counterfactual maintenance reasoning; S600: Extract the path from the physical sensing node to the state evolution node from the real-time mapping logic graph, screen out significant evolution paths through statistical tests, and calculate the contribution weight of each feature to the fault result by combining the quantified correlation effect value, construct the maintenance evidence chain and generate the fault location view.

[0007] Preferably, S100 includes: In the modeling of equipment to be repaired, a laser scanner is used to collect point cloud data of the equipment, and a geometric model including bolts, connectors and sealing surfaces is constructed. Material stiffness, damping and thermodynamic parameters are assigned to generate a physical model, and the equipment operation logic and failure criteria are imported to construct a behavior and rule model. In the modeling of the maintenance robot, a kinematic model of the robotic arm is established based on the DH parameter method, and a sensor field of view model covering visible light angle, infrared temperature measurement field of view, and ultrasonic detection radius is established in virtual space. In the standardized data access, Modbus, OPC-UA and MQTT protocol data are obtained through a unified access gateway, and visual image streams, thermal imaging matrices, vibration spectrum and temperature and humidity time series data are converted into standardized JSON messages based on semantic mapping rules. By associating the geometric model, behavior and rule model, robotic arm kinematics model, and sensor field-of-view model based on the inspection time timestamp, a time axis alignment model with the transformer flange coordinates as the core anchor point is constructed.

[0008] Preferably, S200 includes: The four-layer structured digital twin evaluation system includes a geometric layer, a physical layer, a behavioral layer, and a rule layer. The geometric layer comprises basic variables, including component spatial position, deformation displacement, and surface wear. The physical layer consists of intermediate variables, including operating temperature, vibration frequency, insulation resistance, and current harmonic components, reflecting the physical operating state of the equipment. The behavioral layer comprises response variables, including switch action trajectories, valve opening and closing states, and robot inspection coverage. The rule layer comprises constraint variables, including fault discrimination thresholds, safe operating limits, and remaining life decay functions, serving as the judgment criteria for inspection analysis.

[0009] Preferably, S300 includes: Based on the digital twin mapping theory, a real-time mapping logic graph of physical perception, virtual mapping, and state evolution is constructed. Robot sensor data is designated as physical perception nodes, digital twin state pose is designated as virtual mapping nodes, and device health status indicators are designated as state evolution nodes. Environmental noise and light intensity are also included as interference factors in the graph. Directed edges between nodes represent mapping relationships, and significant mapping paths are selected through conditional independence tests, resulting in a real-time mapping logic graph comprising perception nodes, mapping nodes, evolution nodes, and interference factors. A framework for calculating the correlation effect of inspection data mapping on equipment status changes is constructed, and the expression is as follows: ; In the formula, As variables for the evolution of equipment health status, For the robot's inspection perception variables. It is a set of environmental disturbance factors. This is for digital twin mapping operations on physical inspection data.

[0010] Preferably, S400 includes: For the heterogeneous fusion of visual images and thermal imaging, a multi-sensor weighted fusion algorithm is adopted. The average improvement effect of multi-source data on fault identification is calculated by the Euclidean distance difference of the fused feature vectors, as shown in the following expression: ; In the formula, For average recognition gain, and These represent the failure probabilities after fusion and under a single sensor, respectively. The total number of samples; To quantify equipment operating deviations, a time-series residual analysis model is used to compare the output differences between the physical real-time state values ​​and the virtual simulation baseline values ​​on the time axis, thus identifying the net effect of equipment performance degradation. The expression is as follows: ; In the formula, This is an estimate of the residuals due to performance degradation. , These are the physical real-time state value and the virtual simulation baseline value, respectively. , These represent the current and previous inspection times, respectively. For nonlinear coupling faults, the feature space mapping method is adopted, and the positional deviation in the robot's motion trajectory is selected as the instrumental variable to separate the independent impact of the equipment body fault on the fluctuation of the sensor data.

[0011] Preferably, S500 includes: Based on the game-theoretic SHAP additive interpretation method, the average marginal contribution of each sensor feature to the fault diagnosis result is calculated, generating a ranking of feature importance; the SHAP value calculation expression is: ; In the formula, Sensor characteristics The contribution value, For the total number of features, For those without features Feature subset, For the model in the feature set The predicted failure probability value; A counterfactual maintenance reasoning framework is constructed to quantify the expected impact of maintenance parameter adjustments on the remaining life of the equipment. The calculation expression is as follows: ; In the formula, As a counterfactual remaining life expectancy, These are assumed maintenance and adjustment parameters. For the set of control variables, This represents the model residuals.

[0012] Preferably, S600 includes: All evolution paths from perception nodes to state evolution nodes are extracted from the real-time mapping logic graph, and statistically significant paths with a significance level of less than 0.05 are selected through hypothesis testing. For fault anomalies, a reverse search is performed along a significant path to locate the initial physical location causing the data fluctuation. Combined with the quantified correlation effect value, the weight of each sensor feature on the anomaly diagnosis result is calculated to achieve fault root cause localization. The inspection images, thermal imaging curves, simulated stress distribution and measured vibration data obtained through tracing are arranged in a time sequence to construct a complete chain of maintenance evidence. Simultaneously, a fault location view is generated, with twin nodes representing equipment components, edge thickness representing fault correlation strength, and yellow and red markings indicating early warning and critical alarm status respectively; a standardized maintenance suggestion report is generated at the same time, clearly indicating the fault location and replacement recommendations.

[0013] Secondly, the present invention also provides a maintenance robot inspection data analysis system based on digital twins, which applies the maintenance robot inspection data analysis method based on digital twins as described above. The system includes a model building module, an evaluation index module, a mapping map module, a correlation effect quantification module, an interpretable diagnosis module, and a maintenance evidence chain module. The model building module constructs a time-axis aligned model of the equipment to be inspected and the inspection robot based on 3D modeling technology, and establishes a standardized interactive interface for multi-source heterogeneous data. The evaluation index module is based on a time axis alignment model to construct a four-layer structured digital twin evaluation system covering geometry, physics, behavior, and rules. The mapping graph module is used to define physical sensing nodes, virtual mapping nodes, and state evolution nodes, while introducing environmental interference factors to construct a real-time mapping logic graph and marking data interaction paths in the graph. The correlation effect quantification module uses multi-sensor weighted fusion, time-series residual analysis, and feature space mapping methods to identify and quantify the correlation effect between multi-source inspection data and equipment operating status. The interpretable diagnostic module, based on the SHAP additive interpretation method, constructs a state-fault dual-dimensional interpretable diagnostic model and generates diagnostic output that includes component contribution ranking, local anomaly explanation, and counterfactual maintenance reasoning. The maintenance evidence chain module extracts paths from the real-time mapping logic graph and filters significant evolution paths for faults, anomalies, or inspection points, calculates contribution weights, constructs maintenance evidence chains, and generates a fault location view.

[0014] Preferably, the hardware support layer of the system includes a lidar sensor, a dual-light infrared gimbal, a 5G edge computing gateway, and a distributed graphics rendering server cluster; the lidar sensor is arranged on top of the maintenance robot for real-time scanning of the spatial contours of physical equipment; the dual-light infrared gimbal is used to simultaneously acquire visible light and thermal imaging signals from the equipment; the 5G edge computing gateway performs protocol parsing and raw data cleaning; the distributed graphics rendering server cluster is configured with high-performance GPU nodes for running digital twin 3D rendering algorithms, deep learning algorithms, and counterfactual reasoning algorithms.

[0015] Compared with the prior art, the present invention has the following advantages and beneficial effects: 1. By integrating LiDAR, dual-light gimbals, and multimodal sensing devices, we completed the construction of digital twin models of physical devices and robots and standardized data access. We innovatively constructed a time axis alignment model with spatial coordinates as the core anchor point, which uniformly maps heterogeneous system data to a millisecond-level time series feature space, effectively eliminating the time deviation and spatial misalignment problems between multi-source data. This provides a standardized and traceable data foundation for cross-dimensional fusion analysis, effectively improves fault location accuracy, and solves the bottleneck of multi-source data collaboration in traditional methods.

[0016] 2. Combining the four-layer evaluation system of digital twins with data-driven mapping maps, a dynamic correlation framework of perception-mapping-evolution is built, and environmental interference factors are incorporated into the analysis model to effectively eliminate the interference of noise on fault identification. The innovative integration of multi-sensor weighted fusion and time-series residual analysis methods allows for the design of differentiated detection strategies for sudden faults and slowly changing damage, significantly improving the accuracy of fault identification. At the same time, it can be flexibly adapted to diverse industrial scenarios such as substation inspection, transmission line maintenance, and indoor computer room monitoring.

[0017] 3. Based on the SHAP additive interpretation method, by calculating the marginal contribution of each sensor feature to the diagnostic results, a global feature importance ranking is generated to achieve a macro-health assessment of the equipment's operating status. At the same time, component-level attribution explanations are generated for specific anomalies, and a counterfactual maintenance reasoning framework is provided to quantify the expected gain of maintenance intervention on equipment lifespan. This completely breaks the black-box nature of deep learning models and solves the industry pain point that maintenance personnel find it difficult to trust AI diagnostic conclusions and lack scientific decision-making basis.

[0018] 4. By extracting the entire fault evolution path from the real-time mapping logic graph, screening significant paths through statistical testing, and calculating the contribution weight of each link point by combining the correlation effect value, a complete and traceable evidence chain of perception-mapping-diagnosis is innovatively built. Simultaneously, an interactive 3D visualization view and standardized maintenance report are provided to realize the transformation from passive maintenance to predictive maintenance, predict potential faults 7-30 days in advance, reduce unplanned downtime, and truly realize a data-driven intelligent operation and maintenance closed loop for the entire life cycle of industrial equipment. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the data analysis method for inspection robots based on digital twins.

[0020] Figure 2 This is a structural block diagram of a maintenance robot inspection data analysis system based on digital twins.

[0021] Figure 3 This is a logical flowchart of the process of heterogeneous inspection data interaction and the construction of a four-layer structured evaluation system in this invention. Detailed Implementation

[0022] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention. It should be noted that relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0023] Example 1 Please see Figures 1-3 This embodiment proposes a data analysis method for inspection robots based on digital twins. Taking the inspection of the main transformer and its auxiliary equipment in a large substation as the application background, the scenario involves a high-voltage live environment, complex electromagnetic interference, and the need for high-sensitivity detection of faults such as minor oil leaks and partial discharges. The aim is to achieve deep integration of inspection big data, accurate fault attribution, and predictive maintenance by constructing a full-dimensional digital twin model of the physical transformer and the inspection robot.

[0024] The method is based on a digital twin-based inspection robot data analysis system. At the hardware level, a high-precision lidar sensor is bolted to the top of the inspection robot for real-time construction of equipment spatial point clouds. The robot's end effector integrates a dual-light infrared gimbal, internally connected to a visible light camera and an infrared thermal imager via signal interfaces. A 5G edge computing gateway enables real-time desensitization and preprocessing of inspection images, temperature rise matrices, and vibration frequency signals. At the computing power level, the system is built on a distributed graphics rendering server cluster, with core computing nodes configured with high-performance GPU clusters, providing powerful floating-point computing support for large-scale point cloud rendering, dynamic updates of the digital twin mapping logic graph, and counterfactual reasoning algorithms. Internally, the system is interconnected with the substation's integrated automation system via industrial fiber optic Ethernet, ensuring millisecond-level real-time transmission of multi-source heterogeneous data. At the software level, it includes a model building module, an evaluation index module, a mapping graph module, a correlation effect quantification module, an interpretable diagnostic module, and a maintenance evidence chain module. The method includes: S100. Based on 3D modeling technology, construct a time axis alignment model of the equipment to be inspected and the inspection robot, and establish a standardized interaction interface for multi-source heterogeneous data. Specifically, S100 includes: In the modeling of equipment to be repaired, a high-precision laser scanner is used to acquire point cloud data of the transformer body. A fine geometric model including bolts, heat sinks, flange connections and bushing sealing surfaces is constructed through reverse engineering. Then, transformer oil heat capacity parameters and cooling fin damping coefficients are added to the physical property layer, and failure criteria such as overheating and short circuit impact are imported to form a dynamically updated behavior and rule model. In the modeling of the maintenance robot, a kinematic model of the robotic arm is established based on the DH parameter method, and a sensor field of view model covering visible light angle, infrared temperature measurement field of view, and ultrasonic detection radius is established in virtual space. In the standardized data access, transformer load data of Modbus and OPC-UA protocols are obtained through a unified access gateway. Based on semantic mapping rules, the visual video stream, thermal imaging matrix and audio spectrum data returned by the robot are converted into standardized JSON messages. All messages carry a high-precision global unified timestamp. By associating the refined geometric model, behavior and rule model, robotic arm kinematics model, and sensor field-of-view model based on the inspection time timestamp, a time-axis aligned model is constructed with the transformer flange coordinates as the core anchor point. Taking the detection of oil leakage in the main transformer flange as an example, the system uniformly maps the corresponding visible light leakage image, infrared temperature gradient data, environmental illuminance factor, and robot pose deviation data to a feature vector space indexed by the flange center coordinates, eliminating spatial misalignment between multiple sensor systems and achieving standardized data fusion.

[0025] S200, Based on the time axis alignment model, construct a four-layer structured digital twin evaluation system covering geometry, physics, behavior, and rules; Specifically, S200 includes: The four-layer structured digital twin evaluation system includes a geometric layer, a physical layer, a behavioral layer, and a rule layer. The geometric layer includes basic variables such as the wear degree of the main transformer casing surface and the deformation and displacement of the heat sink. The physical layer includes 15 core dimensions, such as real-time oil temperature, dissolved gas composition in the oil, partial discharge ultrasonic intensity, and converter current harmonic components, to reflect the intrinsic physical state of the equipment. The behavioral layer includes the on / off status of the cooling fan, robot inspection coverage, and robotic arm operation trajectory, reflecting the interactive response between the equipment and the inspection system. The rule layer consists of constraint variables, including the transformer tripping threshold, insulation remaining life decay function, and IEC fault diagnosis criteria, serving as the basis for judgment in inspection analysis.

[0026] S300 defines physical sensing nodes, virtual mapping nodes, and state evolution nodes, while introducing environmental interference factors to construct a real-time mapping logic graph and marking data interaction paths in the graph; Specifically, S300 includes: Based on digital twin mapping theory, a real-time mapping logic graph of physical perception, virtual mapping, and state evolution is constructed. Infrared and visible light signals collected by robot sensors are designated as physical perception nodes, the simulated temperature field and pose state inside the digital twin are designated as virtual mapping nodes, and the overall health rating of the equipment is designated as state evolution nodes. Strong magnetic field environments and ambient light fluctuations are also included as interference factors in the graph. Directed edges between nodes represent mapping relationships, and significant paths are selected through conditional independence testing, resulting in a real-time mapping logic graph containing 20 perception nodes, 12 mapping nodes, and 5 evolution nodes. Based on the full probability criterion, a framework for calculating the correlation effect of inspection data mapping on equipment state changes is constructed, expressed as follows: ; In the formula, As variables for the evolution of equipment health status, For the robot's inspection perception variables. It is a set of environmental disturbance factors. This involves a digital twin mapping operation for physical inspection data. This process is completed in a high-performance graphics rendering server through large-scale topology search, utilizing GPU to accelerate logical graph inference, and can complete a full-link state mapping update within 100ms.

[0027] S400 uses multi-sensor weighted fusion, temporal residual analysis and feature space mapping methods to identify and quantify the correlation effect between multi-source inspection data and equipment operating status. Specifically, S400 includes: To address the heterogeneous fusion of visual images and thermal imaging, a multi-sensor weighted fusion algorithm is employed. First, a convolutional neural network is used to extract oil stain feature points and hotspot distributions on the flange surface, and the representation weights of each sensor for fault features are calculated. Then, a pixel-level registration method is used to spatially align the visible light and infrared layers at a 1:1 ratio. Finally, the average improvement effect of multi-source data on fault identification is calculated based on the Euclidean distance difference between the fused feature vectors, as expressed below: ; In the formula, For average recognition gain, and These represent the failure probabilities after fusion and under a single sensor, respectively. The total number of samples; To quantify equipment operating deviations, a time-series residual analysis model is employed. Using the simulated temperature rise output of the digital twin under standard current load as the baseline group and the real-time infrared data collected by the robot as the experimental group, spurious interference from load fluctuations is eliminated, and the net effect of equipment overheating is identified. The expression is as follows: ; In the formula, This is an estimate of the residuals due to performance degradation. , These are the physical real-time state value and the virtual simulation baseline value, respectively. , These represent the current and previous inspection times, respectively. For nonlinear coupling faults, pose deviation in the robot's motion trajectory is selected as a tool variable to isolate the independent impact of abnormal equipment structure on sensor data fluctuations.

[0028] S500, based on the SHAP additive interpretation method, constructs a state-fault dual-dimensional interpretable diagnostic model, decomposes the fault diagnosis results into various sensor feature indicators, and generates diagnostic output including component contribution ranking, local anomaly explanation and counterfactual maintenance reasoning; Specifically, the S500 includes: Based on the game-theoretic SHAP additive interpretation method, the average marginal contribution of each sensor feature to the transformer fault outcome is calculated. The calculated SHAP value for the local temperature rise index of the transformer bushing is +4.5, indicating that for every unit increase in this feature's abnormality, the fault probability increases by 4.5%. The SHAP value calculation expression is as follows: ; In the formula, Sensor characteristics The contribution value, For the total number of features, For those without features Feature subset, For the model in the feature set The predicted failure probability value is calculated below; for a single transformer oil leakage alarm, a component-level failure attribution explanation is generated, for example: loose flange bolts contribute +12.2%, and excessive environmental vibration contributes +3.1%; A counterfactual maintenance reasoning framework is constructed to quantify the expected impact of maintenance parameter adjustments on the remaining life of the equipment. The calculation expression is as follows: ; In the formula, As a counterfactual remaining life expectancy, These are assumed maintenance and adjustment parameters. For the set of control variables, This represents the model residuals. For example, the model outputs a counterfactual conclusion: if bolt tightening and sealing surface replacement are performed at the current node, the probability of unplanned downtime for this component is expected to decrease from 15% to 0.2%.

[0029] S600: For specific faults, anomalies, or inspection points, extract the path from the physical sensing node to the state evolution node from the real-time mapping logic graph, screen out significant evolution paths through statistical tests, and calculate the contribution weight of each feature to the fault result by combining the quantified correlation effect value, construct the maintenance evidence chain, and generate a visualized fault location view.

[0030] Specifically, the S600 includes: Extract all evolution paths from the sensing node to the state evolution node from the real-time mapping logic graph, and filter them through hypothesis testing. p Statistically significant pathways with a value <0.05; Taking abnormal noise from the B-phase bushing of a transformer as an example, the system traces back to obtain a complete chain of evidence: abnormal spectrum of high-frequency sound sensor (contribution 40%) → distortion of local electric field intensity distribution inside the digital twin (contribution 30%) → discharge evolution calculation results exceeding the standard (significance level 0.01) → finally determined to be moisture in the bushing insulation; a fault location view is generated, with twin nodes representing equipment components and edge thickness representing fault correlation strength, and red marking of severe alarm status; a standardized maintenance suggestion report is generated simultaneously, clearly indicating that the damaged location is the middle of the B-phase bushing.

[0031] Example 2 Please see Figures 1-3 This embodiment takes the maintenance and inspection of indoor precision power distribution cabinets in data centers as the application scenario. This scenario requires extremely high-precision real-time monitoring and diagnosis of circuit breaker contact temperature, cable insulation aging, and switch action logic inside the cabinet, aiming to verify the adaptability of the present invention in indoor closed, high-density equipment environments.

[0032] In this embodiment, a rail-mounted maintenance robot is deployed on the top track of each row of distribution cabinets, equipped with a high-resolution macro infrared camera and a partial discharge detection probe; data is transmitted back via a 5G edge computing gateway. The system constructs a time-axis aligned model with the modular coordinates of the switches inside the distribution cabinet as anchor points. During the inspection process, the robot visually identifies the deviation between the switch handle status and the action commands issued by the system, and uses thermal imaging to collect real-time data on contact temperature rise and current thermal effects at cable connections, while simultaneously recording ambient humidity and background electromagnetic noise.

[0033] The three-layer evaluation system constructed in this embodiment is customized for the characteristics of indoor computer rooms: the physical layer focuses on busbar contact resistance, neutral-to-ground voltage deviation, and switch contact temperature difference; the rule layer includes fire protection specifications for power distribution rooms and three-phase imbalance judgment criteria. Based on the real-time mapping logic graph, significant paths such as: load current fluctuation → contact thermal resistance → infrared anomaly features are identified. For samples with excessively high contact temperatures, SHAP additive interpretation is used to generate failure reports, locating the root cause as loosening of screws due to vibration (feature contribution 55%), and providing counterfactual reasoning: if torque tightening is performed, the fire risk level of the switch node will decrease by three orders of magnitude. The fault evidence chain generated by the system fully presents the evolution process of "physical loosening - thermal accumulation - insulation aging".

[0034] Example 3 Please see Figures 1-3 This embodiment is aimed at a scenario of joint inspection of UAVs and maintenance robots on cross-regional ultra-high voltage transmission lines. This scenario is characterized by extremely complex geographical environment, large fluctuations in communication bandwidth, and huge scale of inspection tasks. It aims to verify the value of the method of the present invention in distributed, large-scale collaborative inspection.

[0035] This embodiment employs a cloud-edge collaborative architecture, utilizing 5G and satellite communication links to achieve globally distributed data processing. The data acquisition module generates a digital twin corridor for long-distance lines using a lidar mounted on a drone, while maintenance robots are deployed on tower supports to perform detailed inspections. The system standardizes inspection data across the entire process, using NLP to parse inspection commands and extract the focus weights for different components such as insulator strings and vibration dampers.

[0036] In the construction of the real-time mapping logic graph, extreme weather (strong winds, freezing) is incorporated as a confounding factor into the model to eliminate the interference of the natural environment on equipment health assessment. For corrosion assessment of different sections, a difference-in-differences (DID) model is used to compare the deterioration rate between sections with and without anti-corrosion measures, accurately quantifying the actual protective effect of anti-corrosion materials. Based on the SHAP method, actionable maintenance suggestions are provided to inspection personnel. For example, for abnormal conductor galloping, the root cause is traced back to the displacement of the vibration damper (contributing 70%), and highlighted in the 3D view.

[0037] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for analyzing inspection data of a maintenance robot based on digital twins, characterized in that, The method includes: S100. Based on 3D modeling technology, construct a time axis alignment model of the equipment to be inspected and the inspection robot, and establish a standardized interaction interface for multi-source heterogeneous data. S200, Based on the time axis alignment model, construct a four-layer structured digital twin evaluation system covering geometry, physics, behavior, and rules; S300 defines physical sensing nodes, virtual mapping nodes, and state evolution nodes, while introducing environmental interference factors to construct a real-time mapping logic graph and marking data interaction paths in the graph; S400 uses multi-sensor weighted fusion, temporal residual analysis and feature space mapping methods to identify and quantify the correlation effect between multi-source inspection data and equipment operating status. S500, based on the SHAP additive interpretation method, constructs a state-fault dual-dimensional interpretable diagnostic model, decomposes the fault diagnosis results into various sensor feature indicators, and generates diagnostic output including component contribution ranking, local anomaly explanation and counterfactual maintenance reasoning; S600: Extract the path from the physical sensing node to the state evolution node from the real-time mapping logic graph, screen out significant evolution paths through statistical tests, and calculate the contribution weight of each feature to the fault result by combining the quantified correlation effect value, construct the maintenance evidence chain and generate the fault location view.

2. The method for analyzing inspection robot data based on digital twins according to claim 1, characterized in that, The S100 includes: In the modeling of equipment to be repaired, a laser scanner is used to collect point cloud data of the equipment, and a geometric model including bolts, connectors and sealing surfaces is constructed. Material stiffness, damping and thermodynamic parameters are assigned to generate a physical model, and the equipment operation logic and failure criteria are imported to construct a behavior and rule model. In the modeling of the maintenance robot, a kinematic model of the robotic arm is established based on the DH parameter method, and a sensor field of view model covering visible light angle, infrared temperature measurement field of view, and ultrasonic detection radius is established in virtual space. In the standardized data access, Modbus, OPC-UA and MQTT protocol data are obtained through a unified access gateway, and visual image streams, thermal imaging matrices, vibration spectrum and temperature and humidity time series data are converted into standardized JSON messages based on semantic mapping rules. By associating the geometric model, behavior and rule model, robotic arm kinematics model, and sensor field-of-view model based on the inspection time timestamp, a time axis alignment model with the transformer flange coordinates as the core anchor point is constructed.

3. The method for analyzing inspection robot data based on digital twins according to claim 2, characterized in that, The S200 includes: The four-layer structured digital twin evaluation system includes a geometric layer, a physical layer, a behavioral layer, and a rule layer. The geometric layer comprises basic variables, including component spatial position, deformation displacement, and surface wear. The physical layer consists of intermediate variables, including operating temperature, vibration frequency, insulation resistance, and current harmonic components, reflecting the physical operating state of the equipment. The behavioral layer comprises response variables, including switch action trajectories, valve opening and closing states, and robot inspection coverage. The rule layer comprises constraint variables, including fault discrimination thresholds, safe operating limits, and remaining life decay functions, serving as the judgment criteria for inspection analysis.

4. The method for analyzing inspection robot data based on digital twins according to claim 3, characterized in that, The S300 includes: Based on the digital twin mapping theory, a real-time mapping logic graph of physical perception, virtual mapping, and state evolution is constructed. Robot sensor data is designated as physical perception nodes, digital twin state pose is designated as virtual mapping nodes, and device health status indicators are designated as state evolution nodes. Environmental noise and light intensity are also included as interference factors in the graph. Directed edges between nodes represent mapping relationships, and significant mapping paths are selected through conditional independence tests, resulting in a real-time mapping logic graph comprising perception nodes, mapping nodes, evolution nodes, and interference factors. A framework for calculating the correlation effect of inspection data mapping on equipment status changes is constructed, and the expression is as follows: ； In the formula, As variables for the evolution of equipment health status, For the robot's inspection perception variables. It is a set of environmental disturbance factors. This is for digital twin mapping operations on physical inspection data.

5. The method for analyzing inspection robot data based on digital twins according to claim 4, characterized in that, The S400 includes: For the heterogeneous fusion of visual images and thermal imaging, a multi-sensor weighted fusion algorithm is adopted. The average improvement effect of multi-source data on fault identification is calculated by the Euclidean distance difference of the fused feature vectors, as shown in the following expression: ； In the formula, For average recognition gain, and These represent the failure probabilities after fusion and under a single sensor, respectively. The total number of samples; To quantify equipment operating deviations, a time-series residual analysis model is used to compare the output differences between the physical real-time state values ​​and the virtual simulation baseline values ​​on the time axis, thus identifying the net effect of equipment performance degradation. The expression is as follows: ； In the formula, This is an estimate of the residuals due to performance degradation. , These are the physical real-time state value and the virtual simulation baseline value, respectively. , These represent the current and previous inspection times, respectively. For nonlinear coupling faults, the feature space mapping method is adopted, and the positional deviation in the robot's motion trajectory is selected as the instrumental variable to separate the independent impact of the equipment body fault on the fluctuation of the sensor data.

6. The method for analyzing inspection robot data based on digital twins according to claim 5, characterized in that, The S500 includes: Based on the game-theoretic SHAP additive interpretation method, the average marginal contribution of each sensor feature to the fault diagnosis result is calculated, generating a ranking of feature importance; the SHAP value calculation expression is: ； In the formula, Sensor characteristics The contribution value, For the total number of features, For those without features Feature subset, For the model in the feature set The predicted failure probability value; A counterfactual maintenance reasoning framework is constructed to quantify the expected impact of maintenance parameter adjustments on the remaining life of the equipment. The calculation expression is as follows: ； In the formula, As a counterfactual remaining life expectancy, These are assumed maintenance and adjustment parameters. For the set of control variables, This represents the model residuals.

7. The method for analyzing inspection data of a maintenance robot based on digital twins according to claim 6, characterized in that, The S600 includes: All evolution paths from perception nodes to state evolution nodes are extracted from the real-time mapping logic graph, and statistically significant paths with a significance level of less than 0.05 are selected through hypothesis testing. For fault anomalies, a reverse search is performed along a significant path to locate the initial physical location causing the data fluctuation. Combined with the quantified correlation effect value, the weight of each sensor feature on the anomaly diagnosis result is calculated to achieve fault root cause localization. The inspection images, thermal imaging curves, simulated stress distribution and measured vibration data obtained through tracing are arranged in a time sequence to construct a complete chain of maintenance evidence. Simultaneously, a fault location view is generated, with twin nodes representing equipment components, edge thickness representing fault correlation strength, and yellow and red markings indicating early warning and critical alarm status respectively; a standardized maintenance suggestion report is generated at the same time, clearly indicating the fault location and replacement recommendations.

8. A maintenance robot inspection data analysis system based on digital twins, characterized in that, The system of the inspection robot inspection data analysis method based on digital twin as described in any one of claims 1-7 includes a model building module, an evaluation index module, a mapping spectrum module, a correlation effect quantification module, an interpretable diagnosis module, and an inspection evidence chain module. The model building module constructs a time-axis aligned model of the equipment to be inspected and the inspection robot based on 3D modeling technology, and establishes a standardized interactive interface for multi-source heterogeneous data. The evaluation index module is based on a time axis alignment model to construct a four-layer structured digital twin evaluation system covering geometry, physics, behavior, and rules. The mapping graph module is used to define physical sensing nodes, virtual mapping nodes, and state evolution nodes, while introducing environmental interference factors to construct a real-time mapping logic graph and marking data interaction paths in the graph. The correlation effect quantification module uses multi-sensor weighted fusion, time-series residual analysis, and feature space mapping methods to identify and quantify the correlation effect between multi-source inspection data and equipment operating status. The interpretable diagnostic module, based on the SHAP additive interpretation method, constructs a state-fault dual-dimensional interpretable diagnostic model and generates diagnostic output that includes component contribution ranking, local anomaly explanation, and counterfactual maintenance reasoning. The maintenance evidence chain module extracts paths from the real-time mapping logic graph and filters significant evolution paths for faults, anomalies, or inspection points, calculates contribution weights, constructs maintenance evidence chains, and generates a fault location view.

9. The intelligent prediction system for temperature compensation of a probe station according to claim 8, characterized in that, The hardware support layer of the system includes a lidar sensor, a dual-light infrared gimbal, a 5G edge computing gateway, and a distributed graphics rendering server cluster. The lidar sensor is positioned on top of the maintenance robot to scan the spatial outline of the physical equipment in real time. The dual-light infrared gimbal is used to simultaneously acquire visible light and thermal imaging signals from the equipment. The 5G edge computing gateway performs protocol parsing and raw data cleaning. The distributed graphics rendering server cluster is configured with high-performance GPU nodes to run digital twin 3D rendering algorithms, deep learning algorithms, and counterfactual reasoning algorithms.

Images