Power equipment safe operation method based on inspection robot and inspection robot

Abstract

This invention provides a method for safe operation and maintenance of power equipment based on an inspection robot, and the inspection robot itself, relating to the field of equipment inspection technology. The method includes: determining the features of each modality of data and the corresponding dynamic weights of each modality of data based on multimodal data from a target observation area; determining multimodal fusion features based on the dynamic weights of each modality of data and the corresponding features of each modality of data; mapping the multimodal fusion features to a preset equipment semantic map and physical mechanism map and performing joint reasoning to obtain the root cause explanation result of the abnormal event and a comprehensive abnormality score; adjusting the path and posture of the inspection robot according to the re-inspection action, and re-collecting the multimodal data as re-inspection data. This invention improves the anomaly detection rate, diagnostic interpretability, single inspection information density, and operation and maintenance response efficiency.

Description

Technical Field

[0001] This invention relates to the field of equipment inspection technology, and more specifically, to a method for safe operation and maintenance of power equipment based on an inspection robot and the inspection robot itself. Background Technology

[0002] The long-term safe operation of power equipment such as transformers, switchgear, circuit breakers, busbar connection points and cable terminals is directly related to the reliability of power supply and personal safety.

[0003] In related technologies, with the development of sensor technology and artificial intelligence, power equipment can be detected using single or combined methods such as visible light cameras, infrared thermal imagers, acoustic sensors, partial discharge detection modules and vibration sensors, and can be used in conjunction with robots for substation inspection.

[0004] However, the above-mentioned technologies suffer from insufficient reliability of single-modal perception, lack of interpretability in diagnostic results, decoupling of robot motion and detection tasks, and separation of data acquisition and operation and maintenance decisions, resulting in low anomaly detection rate and slow response during equipment inspection. Summary of the Invention

[0005] The problem that this invention aims to solve is that single-modal perception is not reliable enough, diagnostic results lack interpretability, robot motion is decoupled from detection tasks, and data acquisition is separated from operation and maintenance decisions, resulting in at least one of the following problems during equipment inspection: low anomaly detection rate and slow response.

[0006] To address the aforementioned problems, in a first aspect, the present invention provides a method for safe operation and maintenance of power equipment based on an inspection robot, comprising: Based on the multimodal data of the target observation area, determine the features of each modal data and the corresponding dynamic weights of each modal data. Based on the dynamic weights of each modal data and the corresponding features of each modal data, determine the multimodal fusion features. The multimodal fusion features are mapped to a preset device semantic map and physical mechanism map and joint reasoning is performed to obtain the root cause explanation results and comprehensive anomaly score corresponding to the abnormal event; wherein, the device semantic map represents the relationship between the device hierarchical structure and fault maintenance, and the physical mechanism map represents the coupling relationship between the anomaly source and the physical field; Based on the comprehensive anomaly score and the preset fusion risk map, a re-inspection action is determined. The path and posture of the inspection robot are adjusted according to the re-inspection action, and the multimodal data is re-collected as re-inspection data. The comprehensive anomaly score is updated based on the re-inspection data. The updated comprehensive anomaly score and the anomaly root cause explanation results are packaged into an anomaly event package and output. The anomaly event package is used to adjust the operation and maintenance decisions of the power equipment and the inspection strategy of the inspection robot.

[0007] The present invention provides a power equipment safety operation and maintenance method based on an inspection robot. It determines the dynamic weights of each modal data based on its characteristics, and then determines multimodal fusion features based on these dynamic weights and the corresponding modal data characteristics. Since the dynamic weights are determined based on the characteristics of each modality, the contribution of each modality in the fusion result is related to its own characteristics, improving the anomaly detection rate compared to a method where all modalities participate in the fusion with equal weights. The multimodal fusion features are mapped to the equipment semantic graph and physical mechanism graph for joint reasoning to obtain the root cause explanation results and comprehensive anomaly scores corresponding to abnormal events. Because the joint reasoning utilizes both the equipment semantic graph and the physical mechanism graph, the diagnostic results have both semantic and physical basis, thereby enhancing the accuracy of diagnosis. The diagnostic interpretability and false alarm suppression capabilities are enhanced. Based on the comprehensive anomaly score and fused risk map, a re-inspection action is determined. After adjusting the path and posture according to the re-inspection action, data is re-collected as re-inspection data. Since the re-inspection action is determined based on the comprehensive anomaly score and fused risk map, and data is re-collected, the re-inspection process complements the initial detection, thereby increasing the effective information density of a single inspection. The comprehensive anomaly score is updated based on the re-inspection data, and the updated comprehensive anomaly score and anomaly root cause explanation results are packaged into an anomaly event package for output. Since the anomaly event package is used to adjust operational decisions and inspection strategies, anomaly discovery, diagnosis, re-inspection, and output form a closed loop, thereby reducing false alarms and missed detections, and reducing repeated on-site visits and inefficient verification by operational personnel.

[0008] Secondly, the present invention also provides an inspection robot, which applies the power equipment safety operation and maintenance method based on the inspection robot as described in any of the above claims, including: The multimodal quality assessment and fusion module is used to: determine the features of each modal data and the corresponding dynamic weights of each modal data based on the multimodal data of the target observation area; and determine the multimodal fusion features based on the dynamic weights of each modal data and the corresponding features of each modal data. The dual-graph joint reasoning module is used to: map the multimodal fusion features to a preset device semantic graph and physical mechanism graph and perform joint reasoning to obtain the root cause explanation result of the abnormal event and the comprehensive abnormal score; wherein, the device semantic graph represents the relationship between the device hierarchical structure and fault maintenance, and the physical mechanism graph represents the coupling relationship between the abnormal source and the physical field; The proactive re-inspection decision module is used to: determine the re-inspection action based on the comprehensive anomaly score and the preset fusion risk map, adjust the path and posture of the inspection robot according to the re-inspection action, and re-collect the multimodal data as re-inspection data; The operation and maintenance and inspection module is used to: update the comprehensive anomaly score based on the re-inspection data, and package the updated comprehensive anomaly score and the anomaly root cause explanation results into an anomaly event package for output. The anomaly event package is used to adjust the operation and maintenance decisions of the power equipment and the inspection strategy of the inspection robot.

[0009] Thirdly, the present invention provides an electronic device, including a memory and a processor; The memory is used to store computer programs; The processor is configured to, when executing the computer program, implement the power equipment safety operation and maintenance method based on the inspection robot as described in the first aspect.

[0010] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the power equipment safety operation and maintenance method based on an inspection robot as described in the first aspect.

[0011] The inspection robot, electronic device, and computer-readable storage medium provided by this invention have the same beneficial effects as the power equipment safety operation and maintenance method based on the inspection robot compared to the prior art, and will not be repeated here. Attached Figure Description

[0012] Figure 1 This invention illustrates a flowchart of a power equipment safety operation and maintenance method based on an inspection robot, as shown in an embodiment of the present invention. Figure 1 ; Figure 2 This diagram illustrates the process of selecting a re-inspection action to maximize the benefit function in an embodiment of the present invention. Figure 3 A schematic diagram of the process for generating an exception event package in an embodiment of the present invention is shown; Figure 4 This invention illustrates a flowchart of a power equipment safety operation and maintenance method based on an inspection robot, as shown in an embodiment of the present invention. Figure 2 ; Figure 5 A schematic diagram of the inspection robot in an embodiment of the present invention is shown; Figure 6 A schematic diagram of the structure of an electronic device according to an embodiment of the present invention is shown. Detailed Implementation

[0013] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0014] It should be noted that relational terms such as "first" and "second" in this invention are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0015] In the description of this specification, references to terms such as "embodiment," "one embodiment," and "one implementation" indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or implementation is included in at least one embodiment or illustrative implementation of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or implementation. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or implementations.

[0016] Reference Figure 1 As shown in the figure, this invention proposes a method for safe operation and maintenance of power equipment based on an inspection robot, including: Based on the multimodal data of the target observation area, the features of each modal data and the corresponding dynamic weights of each modal data are determined. Based on the dynamic weights of each modal data and the corresponding features of each modal data, the multimodal fusion features are determined.

[0017] Specifically, based on the multimodal data of the target observation area, the features of each modal data and the corresponding dynamic weights of each modal data are determined. Then, based on the dynamic weights of each modal data and the corresponding features, the multimodal fusion features are determined. For example, the multimodal data includes at least three of the following: visible light, infrared, acoustic, partial discharge, and vibration. The dynamic weights of each modal data are calculated using a normalized exponential function based on the quality score of each modal data. The quality score is determined using an S-shaped growth curve function based on the quality feature vectors of each modal data, such as sharpness, occlusion rate, temperature drift, and signal-to-noise ratio. The dynamic weights are multiplied by the corresponding encoder output features and then summed to obtain the multimodal fusion features. The core of this step is to dynamically adjust the contribution of each modality according to the real-time operating conditions, so that the fusion features are biased towards the modality with the highest current quality.

[0018] The multimodal fusion features are mapped to a preset device semantic map and physical mechanism map and joint reasoning is performed to obtain the root cause explanation results and comprehensive anomaly score corresponding to the abnormal event; wherein, the device semantic map represents the relationship between the device hierarchical structure and fault maintenance, and the physical mechanism map represents the coupling relationship between the anomaly source and the physical field.

[0019] Specifically, the multimodal fusion features are mapped to preset equipment semantic maps and physical mechanism maps, and joint reasoning is performed to obtain the root cause explanation results and comprehensive anomaly scores corresponding to the abnormal events. The equipment semantic map represents the correlation between equipment hierarchical structure and fault maintenance, while the physical mechanism map represents the coupling relationship between the anomaly source and the physical field. For example, the equipment semantic map describes the subordinate and influence relationships between station areas, equipment, components, measuring points, fault types, and maintenance actions; the physical mechanism map describes physical laws such as heat conduction, mechanical coupling, sound propagation, and electric field coupling. After joint reasoning, a comprehensive anomaly score (e.g., 0.85 points) and anomaly root cause explanation results (e.g., "increased contact resistance at connection points causes overheating") are obtained.

[0020] Based on the comprehensive anomaly score and the preset fusion risk map, a re-inspection action is determined. The path and posture of the inspection robot are adjusted according to the re-inspection action, and the multimodal data is re-collected as re-inspection data.

[0021] Specifically, a re-inspection action is determined based on the comprehensive anomaly score and a preset fusion risk map. The path and posture of the inspection robot are adjusted according to this re-inspection action, and the multimodal data is re-collected as re-inspection data. For example, the fusion risk map comprehensively considers terrain traversability costs, electrical hazard costs, thermal hazard exposure costs, body stability penalties, and observable benefits from different viewpoints. When the comprehensive anomaly score exceeds a threshold and the confidence level is insufficient, the system automatically selects the optimal re-inspection action (such as close-range re-photographing or detour re-inspection), adjusts the robot to move closer to the target equipment or changes the observation angle, and re-collects data to obtain higher-quality evidence.

[0022] The comprehensive anomaly score is updated based on the re-inspection data. The updated comprehensive anomaly score and the anomaly root cause explanation results are packaged into an anomaly event package and output. The anomaly event package is used to adjust the operation and maintenance decisions of the power equipment and the inspection strategy of the inspection robot.

[0023] Specifically, the comprehensive anomaly score is updated based on the re-inspection data. The updated comprehensive anomaly score and the root cause explanation results are packaged into an anomaly event package and output. The anomaly event package is used to adjust the operation and maintenance decisions of the power equipment and the inspection strategy of the inspection robot. For example, after the re-inspection, the comprehensive anomaly score increases from 0.65 to 0.92, and the root cause explanation results are packaged into an anomaly event package. Based on the severity of the anomaly and the importance of the asset in the event package, the cloud platform increases the operation and maintenance priority of the equipment and shortens the re-inspection cycle.

[0024] In practical application, this embodiment determines the dynamic weights of each modality's data based on its characteristics, and then determines the multimodal fusion features based on these dynamic weights and the corresponding modal data features. Since the dynamic weights are determined based on the characteristics of each modality's data, the contribution of each modality in the fusion result is related to its own characteristics, improving the anomaly detection rate compared to a method where all modalities participate in the fusion with equal weights. The multimodal fusion features are then mapped to the device semantic graph and physical mechanism graph for joint reasoning to obtain the root cause explanation results and comprehensive anomaly scores corresponding to the abnormal events. Because the joint reasoning utilizes both the device semantic graph and the physical mechanism graph, the diagnostic results have both semantic and physical basis, thereby enhancing the interpretability of the diagnosis. The system enhances false alarm suppression capabilities. Based on the comprehensive anomaly score and the fused risk map, a re-inspection action is determined. The path and attitude are adjusted according to the re-inspection action, and data is re-collected as re-inspection data. Since the re-inspection action is determined based on the comprehensive anomaly score and the fused risk map, and data is re-collected, the re-inspection process complements the initial detection, thereby increasing the effective information density of a single inspection. The comprehensive anomaly score is updated based on the re-inspection data, and the updated comprehensive anomaly score and the anomaly root cause explanation results are packaged into an anomaly event package for output. Since the anomaly event package is used to adjust operational decisions and inspection strategies, anomaly discovery, diagnosis, re-inspection, and output form a closed loop, thereby reducing false alarms and missed detections, and reducing repeated on-site visits and inefficient verification by operational personnel.

[0025] This invention is particularly suitable for legged inspection robots and can be promoted in power grids, power generation, energy storage and integrated energy stations. It can also be extended to equipment inspection in petrochemical, rail transit, mining, port and other high-risk industrial inspection scenarios.

[0026] As an optional embodiment of the present invention, the step of determining the features of each modal data and the corresponding dynamic weights of each modal data based on the multimodal data of the target observation area, and determining the multimodal fusion features based on the dynamic weights of each modal data and the corresponding features of each modal data, includes: Based on the multimodal data, extract the quality feature vector corresponding to each modality data, and determine the corresponding quality score of each modality data through a preset activation function; Specifically, based on the multimodal data, quality feature vectors corresponding to each modality are extracted, and a quality score for each modality is determined using a preset activation function. The quality feature vector is a set of indicators extracted from each modality to evaluate data quality. Examples include sharpness and occlusion rate from visible light images, temperature drift from infrared thermal images, signal-to-noise ratio from acoustic signals, attitude jitter from vibration signals, and sampling stability from partial discharge signals. The preset activation function is the sigmoid function, used to map the linear transformation result of the quality feature vector to the interval between 0 and 1, obtaining the quality score. A quality score closer to 1 indicates greater reliability of the modality data under the current operating conditions. The quality score is determined according to the following formula: (1) in, This represents the quality score of the i-th modal data. This represents the quality feature vector of the i-th modality data. This represents the learnable weight parameters. Indicates vector transpose; Indicates the bias term. (·) represents the S-shaped growth curve function (activation function).

[0027] Based on the determined quality scores of each modal data, the corresponding dynamic weights of each modal data are determined through a preset normalization function; Specifically, based on the determined quality scores of each modality data, the corresponding dynamic weights of each modality data are determined using a preset normalization function. The preset normalization function is a normalized exponential function (Softmax function), used to convert the quality scores of each modality into normalized dynamic weights, such that the sum of all weights is 1. The dynamic weights are determined according to the following formula: (2) in, Indicates the first Dynamic weights of each modality data Indicates the first Quality scores for each modality of data. This represents the temperature coefficient, used to adjust the smoothness of the weight distribution. Temperature coefficient The larger the value, the more uniform the distribution of modal weights; The smaller the value, the more prominent the modal weight with a high quality score. For example, when a modality's quality score is significantly lower due to occlusion or noise, its dynamic weight automatically decreases; when all modal quality scores are similar, the weight distribution tends to be uniform.

[0028] The feature vectors corresponding to each modality data are extracted, and the dynamic weights of each modality data are multiplied by the corresponding feature vectors and then summed to obtain the multimodal fusion features.

[0029] Specifically, feature vectors corresponding to each modality of data are extracted. The dynamic weights of each modality are multiplied by their corresponding feature vectors, and then summed to obtain the multimodal fusion feature. The feature vectors corresponding to each modality of data are extracted using an encoder, which can be a convolutional neural network, a visual Transformer, or a time-frequency convolutional network. The dynamic weights are multiplied modally by modality of the feature vectors and then summed to achieve weighted fusion, ensuring that features from high-quality modalities dominate the fusion result. The multimodal fusion feature is determined according to the following formula: (3) in, Indicates multimodal fusion features, Indicates the first Dynamic weights of each modality data Indicates the first The feature vector corresponding to each modality data.

[0030] In practical applications, this embodiment utilizes a quality self-assessment and dynamic weighted fusion mechanism to automatically adjust the contribution of each mode based on real-time operating conditions. For example, in low-light environments at night, the clarity of visible light images decreases, automatically reducing their quality score and correspondingly decreasing their dynamic weight, thus automatically favoring the infrared mode. In environments with strong electromagnetic interference, the signal-to-noise ratio of partial discharge signals decreases, reducing their quality score, and the system automatically reduces the weight of the partial discharge mode to prevent interference signals from dominating the fusion result. This mechanism effectively solves the problem of insufficient reliability of single modes under complex operating conditions, significantly improving the robustness and adaptability of multimodal sensing. Simultaneously, the temperature coefficient... The smoothness of the adjustable weight distribution avoids drastic changes in fusion characteristics due to fluctuations in the quality scores of individual modes, thus enhancing the stability of the system.

[0031] For example, unmanned station night inspection: In low-light conditions at night, the robot automatically increases the weight of infrared and acoustic modes and reduces the weight of visible light modes, which are more affected by noise; when the transition between dawn and dusk causes changes in the thermal imaging background, the system corrects the thermal imaging threshold through environmental parameters to maintain stable anomaly detection performance.

[0032] As an optional embodiment of the present invention, the method further includes: during the model training phase, identifying cross-modal consistency loss and updating model parameters based on the cross-modal consistency loss to constrain the consistency of representation of the same device part by different modalities.

[0033] Cross-modal consistency loss is determined by the following formula: (4) in, Indicates cross-modal consistency loss. Indicates the first The feature vector output by the modal encoder Indicates the first The feature vector output by the modal encoder Indicates from modality To mode The mapping operator, ||·|2², represents the square of the L2 norm (i.e., the sum of squares of the Euclidean distance).

[0034] In practical applications, this embodiment introduces cross-modal consistency loss during the training phase to force different modalities to maintain consistent feature representations of the same device location. When a certain modality produces abnormal representations due to occlusion, noise, or other reasons, constraints from other modalities can effectively suppress these contradictory representations, thereby improving the stability and reliability of anomaly detection. This embodiment is particularly suitable for scenarios where there is a natural correlation between multimodal data (e.g., infrared thermal images and visible light images should present consistent temperature distributions and appearance characteristics at the same device location).

[0035] As an optional embodiment of the present invention, the multimodal data includes at least three of visible light, infrared, acoustic, partial discharge, and vibration; the quality feature vector includes at least one of sharpness, occlusion rate, temperature drift, signal-to-noise ratio, attitude jitter, and sampling stability.

[0036] Specifically, the multimodal data includes at least three of the following: visible light, infrared, acoustic, partial discharge, and vibration. The quality feature vector includes at least one of the following: sharpness, occlusion rate, temperature drift, signal-to-noise ratio, attitude jitter, and sampling stability. Sharpness is used to evaluate the image quality of visible light images and can be calculated using Laplacian variance or the Brenner gradient function; occlusion rate is used to evaluate the proportion of key parts of the target device that are occluded and can be calculated using semantic segmentation or target detection algorithms; temperature drift is used to evaluate the degree to which infrared thermal images are affected by changes in ambient temperature and can be calculated using a blackbody reference or ambient temperature compensation; signal-to-noise ratio is used to evaluate the ratio of useful signal to noise in acoustic or partial discharge signals; attitude jitter is used to evaluate the stability of the robot body during sampling and can be calculated using the variance of the angular velocity and acceleration of the inertial measurement unit; sampling stability is used to evaluate the continuity of sampling vibration or partial discharge signals and can be calculated using the variance of the sampling interval.

[0037] In practical applications, the various indicators of the quality feature vector can be flexibly selected and combined according to the sensor type and actual operating conditions. For example, in visible light image quality assessment, sharpness and occlusion rate can be selected as the main indicators; in infrared thermal image quality assessment, temperature drift can be selected as the main indicator; and in acoustic signal quality assessment, signal-to-noise ratio can be selected as the main indicator. By extracting these lightweight quality features, the system can calculate the quality scores of each modality in real time on the edge computing unit without consuming a large amount of computing resources. The specific calculation method of the quality feature vector can be adapted according to the sensor characteristics and deployment environment, exhibiting good scalability.

[0038] like Figure 3 As shown, in an optional embodiment of the present invention, the step of mapping the multimodal fusion features to a preset device semantic map and physical mechanism map and performing joint reasoning to obtain the abnormal root cause explanation result and comprehensive abnormal score corresponding to the abnormal event includes: The multimodal fusion features are attached to the corresponding nodes in the device semantic graph to obtain the first attachment result; wherein, the nodes of the device semantic graph include at least one of station area, equipment, component, measurement point, fault type and maintenance action, and the edges of the device semantic graph include at least one of subordinate relationship, adjacency relationship, influence relationship and maintenance relationship; Specifically, the multimodal fusion features are attached to the corresponding nodes in the equipment semantic graph to obtain a first attachment result. The nodes in the equipment semantic graph include at least one of station area, equipment, component, measurement point, fault type, and maintenance action. The edges in the equipment semantic graph include at least one of dependency relationship, adjacency relationship, influence relationship, and maintenance relationship. For example, when the multimodal fusion features indicate that a transformer bushing has an abnormal temperature, it is attached to the fault type node corresponding to "transformer-bushing-overheating" in the equipment semantic graph, resulting in a first attachment result indicating that "bushing overheating" is a candidate fault type.

[0039] The multimodal fusion features are attached to the corresponding nodes in the physical mechanism graph to obtain a second attachment result; wherein, the nodes of the physical mechanism graph include at least one of heat source, vibration source, discharge source, noise source, conductor and insulator, and the edges of the physical mechanism graph include at least one of heat conduction, mechanical coupling, sound propagation and electric field coupling; Specifically, the multimodal fusion feature is attached to the corresponding node in the physical mechanism map to obtain a second attachment result. The nodes in the physical mechanism map include at least one of heat sources, vibration sources, discharge sources, noise sources, conductors, and insulators, and the edges of the physical mechanism map include at least one of heat conduction, mechanical coupling, sound propagation, and electric field coupling. For example, attaching a temperature anomaly feature to the "heat source" node in the physical mechanism map and associating it along the "heat conduction" edge to the adjacent conductor node yields a second attachment result indicating the path of heat propagation along the conductor.

[0040] Based on the propagation constraints in the physical mechanism map and the second connection result, it is determined whether the multimodal fusion feature conforms to the physical mechanism of the device, and a judgment result is obtained. Specifically, based on the propagation constraints in the physical mechanism map and the second connection result, it is determined whether the multimodal fusion characteristics conform to the physical mechanism of the device, and a judgment result is obtained. Propagation constraints include the directionality of heat conduction (heat flows from the high-temperature region to the low-temperature region), the vibration transmission relationship of mechanical coupling (vibration source is transmitted to adjacent components through mechanical connections), the attenuation law of sound propagation (sound pressure attenuates with propagation distance), and the intensity distribution of electric field coupling (electromagnetic waves generated by partial discharge propagate along the conductor and attenuate). For example, when a significant temperature rise is detected at a connection point but the heat diffusion gradient along the conductor propagation direction does not conform to the heat conduction law, the judgment result is that it does not conform to the physical mechanism, indicating a possible measurement error or local anomaly.

[0041] Based on the first connection result, the second connection result, and the judgment result, the comprehensive anomaly score and the anomaly root cause explanation result are determined.

[0042] Specifically, based on the first connection result, the second connection result, and the judgment result, the comprehensive anomaly score and the anomaly root cause explanation result are determined. The first connection result provides a candidate set of fault types, the second connection result provides the node association paths in the physical map, and the judgment result verifies whether it conforms to the propagation law. All three are used together for subsequent score calculation and root cause confirmation.

[0043] In practical applications, this embodiment leverages the dual constraints of a device semantic graph and a physical mechanism graph. This allows the system to not only identify the presence of anomalies but also trace their root causes. The device semantic graph provides prior knowledge of the device's hierarchical structure and fault maintenance relationships, while the physical mechanism graph provides the propagation laws of heat, mechanics, sound, and electricity. The joint reasoning of these two graphs effectively suppresses false alarms inherent in purely data-driven methods. For example, when a connection point experiences a significant thermal rise but corresponding acoustic and vibration anomalies are low, the system can determine that it is more likely due to increased contact resistance than a heat dissipation system malfunction. Similarly, when a temperature rise is accompanied by periodic vibration and abnormal noise in the fan area, the system can determine that it is more likely related to a heat dissipation system anomaly. This interpretable diagnostic result significantly increases maintenance personnel's confidence in the robot's conclusions.

[0044] The following is an application example: overheating inspection of contacts in the main transformer area. When the robot is performing an inspection around the main transformer, it detects through infrared thermography that the temperature of a certain contact is higher than that of adjacent contacts, and at the same time, visible light images show oxidation marks near the bolts. Combining the relationship of "joint-fastener-increased contact resistance-heating" in the semantic graph and the heat diffusion law in the physical graph, the system judges that the increased contact resistance is more likely caused by loose connections or oxidation, and therefore suggests arranging a power outage for tightening or replacement.

[0045] Here's another application example: Anomaly monitoring of air-cooled systems in energy storage devices: During inspections of energy storage containers or cabinets, robots collect visible light, infrared, and vibration data from the fan area. If abnormal vibration of the fan casing is detected and the infrared image shows an expanding temperature rise boundary, the system further invokes the acoustic module for continuous monitoring and outputs the result "Increased risk of fan bearing wear or blade imbalance," recommending a shorter re-inspection cycle.

[0046] As an optional embodiment of the present invention, determining the comprehensive anomaly score and the anomaly root cause explanation result based on the first connection result, the second connection result, and the judgment result includes: A first score is determined based on the support of each modality data extracted from the multimodal fusion features and the corresponding first weight. A second score is determined based on the joint reasoning results of the device semantic graph and the physical mechanism graph and the corresponding second weight. The comprehensive anomaly score is obtained by combining the first score and the second score. Specifically, a first score is determined based on the support scores of each modality extracted from the multimodal fusion features and their corresponding first weights. The support scores for each modality include visible light support, infrared support, acoustic support, partial discharge support, and vibration support, representing the confidence level of each modality in the current anomalous event. For example, when the infrared image shows a significant temperature rise while the visible light image shows no obvious appearance abnormality, the infrared support is set to 0.9, and the visible light support is set to 0.2.

[0047] The comprehensive anomaly score is determined according to the following formula: (5); in, Indicates the fault type The overall abnormality score, , , , , These represent the support of visible light, infrared, acoustic, partial discharge, and vibration modes for fault type k (i.e., the support of each mode data). This represents the graph inference gain term (i.e., the joint inference result). to The weight coefficients (i.e., the first weights) are the weights corresponding to the support of each modality. This is the weighting coefficient (i.e., the second weight) corresponding to the graph inference gain term.

[0048] Based on the first connection result, the fault type matched by the device semantic graph is determined as a candidate fault type; based on the second connection result and the judgment result, it is determined whether the candidate fault type conforms to the propagation constraints in the physical mechanism graph; if it does, the candidate fault type is confirmed as the abnormal root cause explanation result.

[0049] Specifically, based on the first connection result, the fault type matched by the equipment semantic map is determined as a candidate fault type. Each fault type node in the equipment semantic map corresponds to a specific equipment anomaly, such as "bushing overheating," "circuit breaker contact wear," or "transformer oil overheating." The first connection result associates multimodal fusion features with one or more nodes, and the system uses the fault type corresponding to the node with the highest matching degree as the candidate fault type. For example, when infrared thermal imaging shows that the temperature of the A-phase bushing is significantly higher than that of the B and C phases, and visible light images show discharge traces on the bushing surface, the first connection result will match the "bushing overheating" node, and the candidate fault type will be determined as "bushing overheating."

[0050] Based on the second connection result and the judgment result, it is determined whether the candidate fault type conforms to the propagation constraints in the physical mechanism diagram. The propagation constraints in the physical mechanism diagram are the basis for judging whether an anomaly conforms to physical laws. For example, for the candidate fault type "sleeve overheating," the heat conduction constraint in the physical mechanism diagram requires that heat should propagate outward from the heat source (such as a poor contact point of the internal conductor) along the sleeve axis, and the temperature should decrease from the inside to the outside. If the second connection result and the judgment result show that the temperature distribution conforms to this law, it is judged to conform; if the hot spot is located on the sleeve surface and has no inward conduction trend (possibly caused by external environmental radiation), it is judged to not conform.

[0051] If the conditions are met, the candidate fault type is confirmed as the root cause explanation result. That is, the system ultimately outputs this fault type as the diagnostic conclusion. For example, if the above-mentioned "sleeve overheating" candidate fault type is verified by physical mechanism to conform to the laws of heat conduction, then the root cause explanation result is confirmed as "sleeve overheating, possibly caused by poor contact of internal conductors." If the conditions are not met, the current candidate fault type is discarded, and the system returns to the first connection result to continue evaluating other candidate fault types, or marks them as requiring manual review.

[0052] In practical application, this embodiment uses a weighted sum for comprehensive anomaly scoring, considering both direct observational evidence from multiple modalities and incorporating prior knowledge and propagation patterns from dual graphs, making the scoring results more comprehensive and accurate. The weight coefficients corresponding to the support of each modality can be optimized according to the sensitivity of different fault types. The confirmation of the anomaly root cause explanation results requires simultaneous satisfaction of both semantic graph matching and physical mechanism verification, avoiding misjudgments caused by relying solely on semantic matching. For example, when the semantic graph indicates a certain fault type but the physical propagation pattern does not match, the system will not output the root cause but will instead prompt for further re-examination or manual verification. This dual confirmation mechanism significantly improves the reliability of the diagnostic results.

[0053] As an optional embodiment of the present invention, the step of determining the re-inspection action based on the comprehensive anomaly score and the preset fusion risk map, adjusting the path and posture of the inspection robot according to the re-inspection action, and re-collecting the multimodal data as re-inspection data includes: Obtain the comprehensive anomaly score and its corresponding confidence level; Specifically, the comprehensive anomaly score and its corresponding confidence level are obtained. For example, when the comprehensive anomaly score is 0.65 and the confidence level is 0.7, if the preset confidence level threshold is 0.8, it is determined to be below the threshold, triggering an active re-examination. The confidence level can be calculated using uncertainty estimation methods known in the art, such as the output variance of a Bayesian neural network, the statistical variance of multiple forward propagations in Monte Carlo Dropout, etc. In practical applications, the confidence level threshold can be dynamically adjusted according to the consistency of support among different modalities: when there is a significant conflict between multiple modalities, the threshold is lowered to trigger a re-examination; when the conclusions of each modality are highly consistent, the threshold is increased to avoid unnecessary re-examination.

[0054] When the confidence level is lower than a preset threshold, an active re-inspection is triggered: the re-inspection action that maximizes the re-inspection benefit function is selected from the re-inspection strategy library according to the preset re-inspection benefit function and executed to re-collect the multimodal data as re-inspection data. The re-inspection strategy library includes at least one of close-range re-shooting, detour re-inspection, high-position re-inspection, fixed-point dwell, low-speed rescanning and contact local detection.

[0055] Specifically, when the confidence level is lower than a preset threshold, an active re-examination is triggered: The fusion risk map is determined according to the following formula: (6) in, Indicates the robot's location From a perspective The risks and costs associated with performing detection tasks. Indicates the cost of traversing the terrain. This indicates the cost of the electrical safety hazard. This indicates the cost of exposure to thermal hazards. This indicates a penalty for maintaining the stability of the organism. This represents the observable benefits from a specific perspective. to These are the weighting coefficients for each item. During planning, the objective is to minimize the cumulative risk integral, while satisfying constraints such as foot friction, stride length, body tilt angle, joint torque, and safety distance.

[0056] During path planning, the system aims to minimize the cumulative risk score. Within the boundaries that satisfy foot friction constraints, step length constraints, body tilt angle constraints, joint torque constraints, and safety distance constraints, the system collaboratively calculates the robot's travel path, foot landing point, dwell pose, and gimbal orientation.

[0057] According to a preset re-inspection benefit function, the re-inspection action that maximizes the re-inspection benefit function is selected from the re-inspection strategy library and executed to re-collect the multimodal data as re-inspection data. The re-inspection strategy library includes at least one of the following: close-range re-shooting (i.e., controlling the robot to move closer to the target device, shortening the observation distance to improve resolution and signal-to-noise ratio), bypass re-inspection (i.e., changing the robot's horizontal azimuth angle to observe from different angles to eliminate the effects of occlusion and reflection), high-position re-inspection (i.e., raising the gimbal or robotic arm to observe from above to obtain a top-down view), fixed-point dwelling (i.e., the robot stays stably in the current pose to extend the sampling time to capture intermittent abnormal signals), low-speed rescanning (i.e., reducing the robot's moving speed to continuously collect multiple frames of data during movement to improve the signal-to-noise ratio), and contact-type local detection (i.e., using the robotic arm to bring the sensor probe close to the device surface for precise measurement). For example, when the abnormal part is obscured by other devices, bypass re-inspection is preferred; when it is necessary to confirm intermittent partial discharge signals, fixed-point dwelling re-inspection is preferred.

[0058] In practical applications, this embodiment integrates risk maps to optimize navigation and detection tasks: when navigability is high, a safe path is prioritized; when electrical or thermal risks are high, a safe distance is automatically maintained and the gimbal orientation is adjusted; and when anomaly confirmation is required, the dwell pose that maximizes observation benefits is prioritized. Compared to existing solutions where path planning only considers navigability, this embodiment enables the robot to proactively optimize the detection perspective while ensuring safety, significantly improving the quality of evidence collection for suspected anomalous equipment, and ensuring operational safety in hazardous scenarios such as high pressure, high temperature, and complex terrain.

[0059] Below is an application example: substation switchgear inspection. The robot enters the switchgear passage according to the substation inspection plan, using a visible light camera to identify the status of indicator lights, meters, and cabinet doors. An infrared thermal imager screens the temperature rise of the cabinet door's outer surface and key connection areas. An acoustic and partial discharge module performs combined sensing of abnormal discharge. When an abnormal localized temperature rise accompanied by partial discharge pulses is detected in a cabinet, the system automatically approaches and switches to a high-precision re-inspection mode, outputting a diagnostic conclusion of "suspected insulation degradation with discharge," and elevating the work order priority.

[0060] Below is an application example: comprehensive inspection of the environment and equipment in a power tunnel. The robot uses its four-legged mobility to traverse ramps, waterlogged areas, and narrow bridge sections within the tunnel, continuously collecting thermal images of cable joints, ambient humidity, and visible light appearance information. When the system detects abnormal temperature and increased humidity near a junction box, it can automatically generate an environmental anomaly alarm, alerting maintenance personnel to potential sealing and leakage issues.

[0061] like Figure 2 As shown, in an optional embodiment of the present invention, the re-inspection benefit function is a weighted sum of the expected confidence improvement value, the observability improvement value, the time and energy consumption cost, the environmental risk, and the motion risk; wherein, the expected confidence improvement value is the expected increase in the confidence of the comprehensive anomaly score after performing the re-inspection action, and the observability improvement value is the expected increase in the visibility of key parts of the target equipment after performing the re-inspection action.

[0062] Specifically, the re-inspection benefit function is determined according to the following formula: (7) in, This represents the benefit score for performing the re-inspection action a; This indicates the expected increase in confidence level, that is, the expected increase in the confidence level of the comprehensive anomaly score after performing the re-examination action a; This indicates the expected increase in observability, specifically the expected increase in the visibility of key parts of the target equipment after performing re-inspection action a. This indicates the time and energy costs, including the additional time and energy consumed for re-inspection. This indicates environmental and motion risks, including electrical conductivity risks when near high-risk equipment, thermal radiation exposure risks, and motion risks of the robot itself in complex terrain. to These are the weighting coefficients for each item, which can be configured by operations and maintenance personnel according to the actual scenario.

[0063] Increased confidence level The improvement in feature quality can be estimated by simulating the re-inspection process. For example, close-range re-imaging can improve image resolution, thereby increasing the confidence in target detection; bypass re-inspection can eliminate occlusion, thereby improving the confidence in partial discharge localization. The expected improvement in observability is also significant. The cost can be estimated based on the changes in coverage and resolution of key parts of the target equipment in images or point clouds before and after the re-inspection process. (Time and energy consumption costs are also considered.) Environmental risks and motion risks can be predicted based on robot motion models and gimbal control models. The assessment can be based on the robot's current position, ground friction coefficient, safe distance from high-voltage equipment, and the heat radiation intensity of the equipment.

[0064] The system calculates the re-examination action for each candidate from the re-examination strategy library. Value, selection The largest action is executed.

[0065] In practical application, this embodiment quantifies the re-examination decision into a calculable benefit score using the re-examination benefit function. Weighting coefficients. to It can be dynamically adjusted according to the scenario: it can improve efficiency during nighttime inspections. The weight of (observability improvement) can be increased when battery power is low. The weighting of (time and energy costs) can be increased when near high-voltage equipment. The weighting of (environmental risks) is determined by this function. This function ensures that the selection of re-inspection actions seeks the optimal solution between benefits and costs, improving the reliability of anomaly confirmation while controlling operational risks and resource consumption. By adjusting the weighting coefficients, this embodiment can be adapted to the special safety requirements and task constraints of different power scenarios (such as substations, switchgear channels, and power tunnels).

[0066] like Figure 3 and Figure 4As shown, in an optional embodiment of the present invention, the steps of updating the comprehensive anomaly score based on the re-inspection data, packaging the updated comprehensive anomaly score and the anomaly root cause explanation result into an anomaly event package for output, and the preceding steps are all performed in the edge computing unit of the inspection robot; after updating the comprehensive anomaly score based on the re-inspection data and packaging the updated comprehensive anomaly score and the anomaly root cause explanation result into an anomaly event package for output, the method further includes: The edge computing unit uploads the abnormal event packet to the cloud, so that the cloud can generate operation and maintenance decisions for the power equipment based on the abnormal event packet, and adjust the inspection strategy of the inspection robot according to the operation and maintenance decisions.

[0067] Specifically, the inspection robot has a built-in edge computing unit. The aforementioned processes of determining the characteristics and dynamic weights of each modal data, obtaining multimodal fusion features, joint inference, determining the comprehensive anomaly score, triggering proactive re-inspection, updating the comprehensive anomaly score, and packaging and outputting anomaly event packages are all performed within this edge computing unit. After outputting the anomaly event package, the edge computing unit uploads it to the cloud, enabling the cloud to generate maintenance decisions for the power equipment based on the anomaly event package and adjust the inspection strategy of the inspection robot accordingly. For example, the cloud can increase the maintenance priority of a certain piece of equipment and shorten its re-inspection cycle based on the severity of the anomaly and the asset importance in the anomaly event package.

[0068] In practical applications, the edge computing unit can independently complete all detection and decision-making tasks in offline mode, ensuring the continuity of inspection in unattended scenarios. After the network is restored, the event packets are automatically uploaded, centrally processed in the cloud, and feedback is provided to adjust the strategy, achieving synergy between edge real-time performance and cloud-based global optimization.

[0069] like Figure 4 As shown, in an optional embodiment of the present invention, the cloud is used for: Based on the severity of the anomaly in the anomaly package, the frequency of historical recurrence, the importance of the asset, and the load sensitivity, determine at least one of the following: maintenance priority, generate maintenance work orders, spare parts recommendations, and equipment lifespan trend analysis results; Specifically, the cloud platform combines multiple factors to automatically assign operational priority levels to the generated abnormal event packages: (8) Severity of abnormality The frequency of historical recurrence is determined based on the calculated comprehensive anomaly score. Asset importance is determined by the number of historical occurrences of the same device or similar anomalies within a preset time period, based on cloud-based statistics. Based on the criticality of the equipment in the power system, load sensitivity is preset. The abnormal risk weighting coefficient reflects the equipment under the current load conditions. to It can be configured by operations and maintenance personnel according to actual needs; based on the calculated operations and maintenance priorities. The system automatically distinguishes the handling level, which includes at least one of the following: immediate shutdown, arranging maintenance as soon as possible, inclusion in the observation list, and continued monitoring.

[0070] When generating a maintenance work order, the work order shall at least include the abnormal equipment number, the abnormal location, the root cause explanation result of the abnormality, the recommended handling measures, and the suggested re-inspection cycle as described in claim 1; the spare parts recommendation shall be generated based on the root cause explanation result of the abnormality and the equipment life trend analysis result; the equipment life trend analysis result shall be obtained by comprehensively evaluating the ambient temperature, equipment load curve, maintenance interval, duration of abnormality, and historical failure modes.

[0071] Adjust the target observation area and re-inspection cycle in the subsequent inspection plan of the inspection robot.

[0072] Specifically, when the maintenance priority indicates that a certain device needs to be repaired or shut down immediately, the cloud will adjust the target observation area to the location of the device and shorten the re-inspection cycle of the device; when the device lifespan trend analysis results show that the remaining lifespan of a certain device is lower than a preset threshold, the cloud will increase the observation frequency of the device in subsequent inspection plans and adjust the re-inspection cycle to half or less of the original cycle; when the severity of the anomaly is low and the frequency of historical recurrence is stable, the cloud will maintain or extend the re-inspection cycle of the device to optimize the allocation of inspection resources.

[0073] In practical application, this embodiment achieves collaboration between real-time processing at the edge and in-depth analysis in the cloud. After the edge completes anomaly screening and re-inspection decision-making, it uploads the event package to the cloud. The cloud automatically calculates the maintenance priority based on the maintenance priority formula, distinguishes the handling level, generates work orders and spare parts suggestions, and dynamically adjusts the inspection strategy. This closed-loop mechanism shortens the time from anomaly discovery to handling decision-making, concentrates limited inspection resources on high-risk equipment, and continuously optimizes the statistical parameters in the calculation formula through historical data accumulation, improving the accuracy of maintenance decisions. In addition, the cloud interface pushes anomaly evidence for on-duty personnel to remotely review, forming a highly efficient human-machine collaborative review closed loop.

[0074] like Figure 5 As shown, the present invention also provides an inspection robot 200, which applies the power equipment safety operation and maintenance method based on the inspection robot as described in the above embodiments, including: The multimodal quality assessment and fusion module 210 is used to: determine the features of each modal data and the corresponding dynamic weights of each modal data based on the multimodal data of the target observation area; and determine the multimodal fusion features based on the dynamic weights of each modal data and the corresponding features of each modal data. The dual-graph joint reasoning module 220 is used to: map the multimodal fusion features to a preset device semantic graph and physical mechanism graph and perform joint reasoning to obtain the root cause explanation result of the abnormal event and the comprehensive abnormal score; wherein, the device semantic graph represents the relationship between the device hierarchical structure and fault maintenance, and the physical mechanism graph represents the coupling relationship between the abnormal source and the physical field; The proactive re-inspection decision module 230 is used to: determine the re-inspection action based on the comprehensive anomaly score and the preset fusion risk map, adjust the path and posture of the inspection robot according to the re-inspection action, and re-collect the multimodal data as re-inspection data; The operation and maintenance and inspection module 240 is used to: update the comprehensive anomaly score according to the re-inspection data, and package the updated comprehensive anomaly score and the anomaly root cause explanation results into an anomaly event package for output. The anomaly event package is used to adjust the operation and maintenance decisions of the power equipment and the inspection strategy of the inspection robot.

[0075] The specific implementation method of this embodiment can also refer to the corresponding implementation method described above, and will not be described again here.

[0076] like Figure 6 As shown in the figure, an electronic device 300 provided in this embodiment of the invention includes a memory 310 and a processor 320; the memory 310 is used to store a computer program; the processor 320 is used to implement the power equipment safety operation and maintenance method based on the inspection robot as described above when the computer program is executed.

[0077] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0078] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

[0079] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A method for safe operation and maintenance of power equipment based on an inspection robot, characterized in that, include: Based on the multimodal data of the target observation area, determine the features of each modal data and the corresponding dynamic weights of each modal data. Based on the dynamic weights of each modal data and the corresponding features of each modal data, determine the multimodal fusion features. The multimodal fusion features are mapped to a preset device semantic map and physical mechanism map and joint reasoning is performed to obtain the root cause explanation results and comprehensive anomaly score corresponding to the abnormal event; wherein, the device semantic map represents the relationship between the device hierarchical structure and fault maintenance, and the physical mechanism map represents the coupling relationship between the anomaly source and the physical field; Based on the comprehensive anomaly score and the preset fusion risk map, a re-inspection action is determined. The path and posture of the inspection robot are adjusted according to the re-inspection action, and the multimodal data is re-collected as re-inspection data. The comprehensive anomaly score is updated based on the re-inspection data. The updated comprehensive anomaly score and the anomaly root cause explanation results are packaged into an anomaly event package and output. The anomaly event package is used to adjust the operation and maintenance decisions of the power equipment and the inspection strategy of the inspection robot.

2. The method for safe operation and maintenance of power equipment based on inspection robots according to claim 1, characterized in that, The process of determining the features of each modal data and the corresponding dynamic weights of each modal data based on the multimodal data of the target observation area, and determining the multimodal fusion features based on the dynamic weights of each modal data and the corresponding features of each modal data, includes: Based on the multimodal data, extract the quality feature vector corresponding to each modality data, and determine the corresponding quality score of each modality data through a preset activation function; Based on the determined quality scores of each modal data, the corresponding dynamic weights of each modal data are determined through a preset normalization function; The feature vectors corresponding to each modality data are extracted, and the dynamic weights of each modality data are multiplied by the corresponding feature vectors and then summed to obtain the multimodal fusion features.

3. The method for safe operation and maintenance of power equipment based on inspection robots according to claim 2, characterized in that, The multimodal data includes at least three of the following: visible light, infrared, acoustic, partial discharge, and vibration; the quality feature vector includes at least one of the following: sharpness, occlusion rate, temperature drift, signal-to-noise ratio, attitude jitter, and sampling stability.

4. The method for safe operation and maintenance of power equipment based on inspection robots according to claim 1, characterized in that, The process of mapping the multimodal fused features to a preset device semantic map and physical mechanism map and performing joint reasoning to obtain the root cause explanation result of the abnormal event and the comprehensive abnormal score includes: The multimodal fusion features are attached to the corresponding nodes in the device semantic graph to obtain the first attachment result; wherein, the nodes of the device semantic graph include at least one of station area, equipment, component, measurement point, fault type and maintenance action, and the edges of the device semantic graph include at least one of subordinate relationship, adjacency relationship, influence relationship and maintenance relationship; The multimodal fusion features are attached to the corresponding nodes in the physical mechanism graph to obtain a second attachment result; wherein, the nodes of the physical mechanism graph include at least one of heat source, vibration source, discharge source, noise source, conductor and insulator, and the edges of the physical mechanism graph include at least one of heat conduction, mechanical coupling, sound propagation and electric field coupling; Based on the propagation constraints in the physical mechanism map and the second connection result, it is determined whether the multimodal fusion feature conforms to the physical mechanism of the device, and a judgment result is obtained. Based on the first connection result, the second connection result, and the judgment result, the comprehensive anomaly score and the anomaly root cause explanation result are determined.

5. The method for safe operation and maintenance of power equipment based on inspection robots according to claim 4, characterized in that, The step of determining the comprehensive anomaly score and the anomaly root cause explanation result based on the first connection result, the second connection result, and the judgment result includes: A first score is determined based on the support of each modality data extracted from the multimodal fusion features and the corresponding first weight. A second score is determined based on the joint reasoning results of the device semantic graph and the physical mechanism graph and the corresponding second weight. The comprehensive anomaly score is obtained by combining the first score and the second score. Based on the first connection result, the fault type matched by the device semantic graph is determined as a candidate fault type; based on the second connection result and the judgment result, it is determined whether the candidate fault type conforms to the propagation constraints in the physical mechanism graph; if it does, the candidate fault type is confirmed as the abnormal root cause explanation result.

6. The method for safe operation and maintenance of power equipment based on an inspection robot according to any one of claims 1-5, characterized in that, The process of determining a re-inspection action based on the comprehensive anomaly score and a preset fusion risk map, adjusting the path and posture of the inspection robot according to the re-inspection action, and re-collecting the multimodal data as re-inspection data includes: Obtain the comprehensive anomaly score and its corresponding confidence level; When the confidence level is lower than a preset threshold, an active re-inspection is triggered: the re-inspection action that maximizes the re-inspection benefit function is selected from the re-inspection strategy library according to the preset re-inspection benefit function and executed to re-collect the multimodal data as re-inspection data. The re-inspection strategy library includes at least one of close-range re-shooting, detour re-inspection, high-position re-inspection, fixed-point dwell, low-speed rescanning and contact local detection.

7. The method for safe operation and maintenance of power equipment based on inspection robots according to claim 6, characterized in that, The re-inspection benefit function is a weighted sum of the expected confidence improvement, the observability improvement, and the time and energy costs, environmental risks, and motion risks; wherein, the expected confidence improvement is the expected increase in the confidence of the comprehensive anomaly score after performing the re-inspection action, and the observability improvement is the expected increase in the visibility of key parts of the target equipment after performing the re-inspection action.

8. The method for safe operation and maintenance of power equipment based on an inspection robot according to any one of claims 1-5, characterized in that, The steps of updating the comprehensive anomaly score based on the re-inspection data, packaging the updated comprehensive anomaly score and the anomaly root cause explanation result into an anomaly event package for output, and the preceding steps are all performed in the edge computing unit of the inspection robot; after updating the comprehensive anomaly score based on the re-inspection data, packaging the updated comprehensive anomaly score and the anomaly root cause explanation result into an anomaly event package for output, the steps further include: The edge computing unit uploads the abnormal event packet to the cloud, so that the cloud can generate operation and maintenance decisions for the power equipment based on the abnormal event packet, and adjust the inspection strategy of the inspection robot according to the operation and maintenance decisions.

9. The method for safe operation and maintenance of power equipment based on an inspection robot according to claim 8, characterized in that, The cloud is used for: Based on the severity of the anomaly in the anomaly package, the frequency of historical recurrence, the importance of the asset, and the load sensitivity, determine at least one of the following: maintenance priority, generate maintenance work orders, spare parts recommendations, and equipment lifespan trend analysis results; Adjust the target observation area and re-inspection cycle in the subsequent inspection plan of the inspection robot.

10. An inspection robot, characterized in that, The method for safe operation and maintenance of power equipment based on inspection robots as described in any one of claims 1-9 includes: The multimodal quality assessment and fusion module is used to: determine the features of each modal data and the corresponding dynamic weights of each modal data based on the multimodal data of the target observation area; and determine the multimodal fusion features based on the dynamic weights of each modal data and the corresponding features of each modal data. The dual-graph joint reasoning module is used to: map the multimodal fusion features to a preset device semantic graph and physical mechanism graph and perform joint reasoning to obtain the root cause explanation result of the abnormal event and the comprehensive abnormal score; wherein, the device semantic graph represents the relationship between the device hierarchical structure and fault maintenance, and the physical mechanism graph represents the coupling relationship between the abnormal source and the physical field; The proactive re-inspection decision module is used to: determine the re-inspection action based on the comprehensive anomaly score and the preset fusion risk map, adjust the path and posture of the inspection robot according to the re-inspection action, and re-collect the multimodal data as re-inspection data; The operation and maintenance and inspection module is used to: update the comprehensive anomaly score based on the re-inspection data, and package the updated comprehensive anomaly score and the anomaly root cause explanation results into an anomaly event package for output. The anomaly event package is used to adjust the operation and maintenance decisions of the power equipment and the inspection strategy of the inspection robot.

Images