Fault identification method and system of track patrol robot in power distribution room and storage medium

Abstract

The application relates to the technical field of data fusion pattern recognition, and particularly discloses a fault identification method and system of a power distribution room track patrol robot and a storage medium. The method comprises the following steps: controlling a track robot to collect operation state observation data of a target power equipment, and triggering the collection of corresponding context environment information according to the data type or a preliminary abnormality determination result; performing time correlation and spatial relationship matching on the operation state observation data and the context environment information, and constructing a data set of the device ontology state, the environment state and the spatial topology relationship; identifying a corresponding environment interference source based on a pre-stored knowledge base, and determining the interference contribution of the environment interference source to the preliminary observation characteristics according to the state attributes of the environment interference source and the spatial topology relationship between the environment interference source and the target power equipment. The application can reduce fault misjudgment caused by factors such as ambient light and adjacent heat sources, and improve the accuracy, reliability and intelligent degree of power distribution room patrol fault identification.

Description

Technical Field

[0001] This invention relates to the field of data fusion pattern recognition technology, specifically to a fault identification method, system, and storage medium for a power distribution room track inspection robot. Background Technology

[0002] With the development of smart grids, unmanned and intelligent inspection of power distribution rooms has become a trend. Track-mounted inspection robots are widely used in power distribution rooms because of their fixed running paths and simple deployment. By carrying sensors such as visible light cameras and infrared thermal imagers, they can realize the automated collection of equipment status. However, existing technical solutions have obvious limitations: the fault judgment logic is simple and isolated, usually relying on threshold comparisons of single-moment and single-sensor data, lacking comprehensive consideration of the equipment's operating environment; for example, the color of the equipment indicator light may be distorted in the image due to ambient light such as reflections from lighting fixtures or interference from natural light, leading to the system misjudging its status; abnormal temperature on the equipment surface may not originate from its own fault, but rather from the heat radiation or heat dissipation of adjacent high-load equipment; existing systems lack an active verification mechanism, and once abnormal data is detected, alarms are often triggered directly, generating a large number of false alarms, interfering with the judgment of maintenance personnel, and reducing the reliability of the system; the inspection strategy is rigid, and the robot usually operates according to preset paths and patterns, unable to intelligently adjust its inspection behavior to conduct in-depth verification based on preliminary detection results. Summary of the Invention

[0003] To overcome the shortcomings of existing technologies, this invention proposes a fault identification method, system, and storage medium for a power distribution room track inspection robot. Based on traditional data acquisition, it simultaneously acquires information on environmental factors that may affect the judgment, and through data fusion and analysis, removes interference and restores the true state of the equipment, thereby significantly improving the accuracy of fault identification. The fault identification method for the power distribution room track inspection robot of the present invention includes the following steps: S1. Control the track robot to collect the operating status observation data of the target power equipment, and trigger the collection of contextual environment information that may affect its interpretation based on the data type of the operating status observation data or the preliminary anomaly judgment result. S2. Perform time correlation and spatial relationship matching between the observed operating status data and the context environment information, establish the spatiotemporal correspondence between the two, and construct a data set containing the device body status, environmental status and spatial topology relationship; S3. Extract preliminary observation features from the operating status observation data, identify corresponding environmental interference sources from the context environment information based on the pre-stored knowledge base, determine the interference contribution of the environmental interference source to the preliminary observation features according to the state attributes of the environmental interference source and its spatial topological relationship with the target power equipment, and verify or correct the preliminary observation features based on the interference contribution to obtain the target features. S4. Construct a dynamic correlation graph model based on the target features, environmental interference sources and their correlation relationships. The dynamic correlation graph model at least represents the state correlation relationship between target features and the interference interpretation relationship of environmental interference sources on the corresponding observed features. S5. Analyze the dynamic correlation graph model to generate fault diagnosis conclusions, and generate subsequent inspection control instructions or knowledge base update instructions for the track robot based on the fault diagnosis conclusions.

[0004] Specifically, in S1: When the operating status observation data is image data involving the color of the device indicator light, image information of potential light sources around the indicator light observation area that may affect its color interpretation is collected; When the temperature data of the target power equipment is detected to reach the preset attention threshold, the status image and temperature data of the neighboring equipment that are spatially adjacent to the target power equipment and may have thermal interaction are collected. When the confidence level of the result obtained after verifying or correcting the preliminary observation features based on the context information is lower than the preset confidence threshold, the control robot changes at least one of the following: acquisition position, gimbal angle, sensor acquisition mode, or acquisition parameters, to supplement the acquisition of the target power equipment or the corresponding environmental interference source.

[0005] Specifically, S3 includes: Based on the color characteristics of the indicator light, the light interference contribution of ambient light to the color observation value of the indicator light is determined according to the color attribute, light intensity attribute and spatial relationship between the potential light source and the indicator light. The color observation value of the indicator light is then compensated according to the light interference contribution to obtain the target color characteristics of the indicator light. Based on the temperature characteristics of the equipment, the equipment type, working status and temperature information of the neighboring equipment are identified, and the thermal interference contribution of the neighboring equipment to the measured temperature of the target power equipment is determined by combining the spatial topological relationship between the neighboring equipment and the target power equipment. Based on the thermal interference contribution, the target temperature characteristics characterizing the heating state of the target power equipment are obtained.

[0006] Specifically, the data set is a structured data frame, whose data field includes at least a subset of device-body multimodal sensing data, a subset of environmental images related to optical interference, a subset of neighboring device data related to thermal interference, and spatial topology metadata. The spatial topology metadata includes at least one of the following: device identifier, monitoring point identifier, acquisition time identifier, and relative position, distance, and relative orientation between the target power equipment, potential light source, and neighboring equipment.

[0007] Specifically, the nodes of the dynamic association graph model include at least: The observation feature nodes corresponding to the preliminary observation features directly extracted from the operational status observation data; Target feature nodes corresponding to the target features obtained after verification or correction of environmental interference; and The interference source nodes corresponding to the environmental interference sources have been identified; The edges of the dynamic association graph model include at least the state association edges between target feature nodes and the interference interpretation edges pointing from the interference source node to the corresponding observed feature node. The weights of the interference interpretation edges are determined or adjusted based on the interference contribution of the corresponding environmental interference sources to the observed features.

[0008] Specifically, compensating for the observed color values ​​of the indicator lights includes: Determine the color components of ambient light based on the potential light source image; The ambient light intensity coefficient is determined based on at least one of the following: the brightness of the potential light source, the imaging area, and the distance and relative angle between the potential light source and the indicator light. The color contribution of ambient light to the color observation value of the indicator light is determined based on the color components and ambient light intensity coefficient. The color observation value of the indicator light is then adjusted or mapped in the RGB color space or HSV color space to obtain the compensated target color features of the indicator light.

[0009] Specifically, determining the contribution of the thermal interference includes: The heat contribution of each neighboring device is determined by using a pre-set empirical coefficient model or a simplified heat transfer model, based on the distance between each neighboring device and the target power equipment, their relative orientation, the type of neighboring device, and the temperature difference between the neighboring device and the target power equipment. The external heat contribution is obtained by combining the heat contribution of each neighboring device, and the measured temperature of the target power equipment is corrected by combining the ambient background temperature and the external heat contribution to obtain the self-temperature rise characteristics that characterize the self-heating state of the target power equipment.

[0010] Specifically, the knowledge base update includes: The false judgment events that are verified to be caused by environmental interference sources are recorded as environmental interference pattern samples, and the environmental interference pattern samples are associated with and stored with the corresponding monitoring point identifiers. The environmental interference pattern sample includes at least the type of interference source, the conditions under which the interference occurs, and the information on the interference impact of the interference source on the corresponding observation features; Based on multiple environmental interference pattern samples associated with the same monitoring point, update at least one of the following: context environment information acquisition triggering conditions, interference contribution calculation parameters, and feature determination thresholds corresponding to the monitoring point.

[0011] Furthermore, the present invention also provides a fault identification system for a power distribution room track inspection robot, comprising: The sensing module is used to collect observation data on the operating status of the target power equipment and contextual information; The context scheduling module is used to trigger the collection task of corresponding context environment information based on the data type of the running status observation data or the preliminary anomaly judgment result. The fusion processing module is used to perform temporal correlation and spatial relationship matching between operational status observation data and contextual environment information, and to construct a data set containing the device's physical status, environmental status, and spatial topology relationships. A knowledge base is used to store device models, fault characteristics, environmental interference patterns, spatial topology information, and association rules. An anti-interference verification module is used to identify environmental interference sources corresponding to preliminary observation features, determine the interference contribution of environmental interference sources to preliminary observation features, and verify or correct preliminary observation features based on the interference contribution; the anti-interference verification module includes an optical interference processing unit and a thermal correlation analysis unit; The feature refinement module is used to output verified or corrected target features; The model building module is used to construct a dynamic correlation graph model based on observation features, target features, and environmental interference sources. The diagnostic module is used to analyze the dynamic correlation graph model and generate fault diagnosis conclusions; The strategy module is used to generate subsequent inspection control instructions or knowledge base update instructions based on the fault diagnosis conclusions. The robot motion control module is used to execute the subsequent inspection control commands to control the movement position of the track robot, the gimbal attitude or the acquisition status of the sensor module, and to feed back the supplementary acquired data to the fusion processing module.

[0012] Furthermore, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the processor performs the fault identification method for the power distribution room track inspection robot as described above.

[0013] The beneficial effects of this invention are as follows: 1. Based on the data type of the target power equipment's operating status observation data or the preliminary anomaly judgment result, this invention actively triggers the collection of contextual environment information, and establishes the correspondence between the equipment's own data and environmental information through time correlation and spatial relationship matching, so that subsequent fault judgment can be combined with specific environmental conditions for analysis, avoiding threshold judgment based solely on a single moment or a single sensor data.

[0014] 2. This invention identifies environmental interference sources corresponding to preliminary observation features based on a pre-stored knowledge base, and determines the interference contribution based on the state attributes of the environmental interference sources and their spatial topological relationship with the target power equipment. Based on this, the preliminary observation features are verified or corrected, thereby distinguishing between equipment anomalies and false anomalies caused by external factors such as ambient light and nearby heat sources, reducing the probability of false alarms, and improving the accuracy and reliability of fault identification results.

[0015] 3. This invention constructs a dynamic correlation graph model that simultaneously represents the state correlation between target features and the interpretation relationship of environmental interference sources on observed features. It performs joint analysis on multiple target features and their environmental impacts, and generates subsequent inspection control instructions or knowledge base update instructions based on fault diagnosis conclusions. This enables the track robot to perform supplementary data collection, fine retesting, or adjustment of inspection strategies for suspected anomalies. At the same time, it uses verified environmental interference patterns to optimize the subsequent judgment process, thereby forming a closed-loop inspection process with a feedback mechanism, improving the intelligence level and operation and maintenance efficiency of power distribution room inspection. Attached Figure Description

[0016] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0017] Figure 1 This is a flowchart of the method in this invention; Figure 2 This is a flowchart of the indicator light interference verification process for the present invention; Figure 3 This is a flowchart of the temperature thermal interference verification process for the device of the present invention; Figure 4 This is a diagram showing the ambient light interference compensation for indicator light colors in an embodiment of the present invention. Figure 5 This is a thermal interference stripping diagram of a neighboring device according to an embodiment of the present invention; Figure 6 This is a grouped bar chart comparing the diagnostic effects of two methods in an embodiment of the present invention; Figure 7 This is a curve of the convergence error of the confidence level of the active retest in an embodiment of the present invention. Detailed Implementation

[0018] To make the technical means, data processing relationships, and their resulting technical effects clearer, the present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be noted that the following embodiments are used to illustrate the technical solution of the present invention, and not to limit the scope of protection of the present invention. Without departing from the basic technical concept of the present invention, those skilled in the art can adaptively adjust the corresponding parameters, data structures, and calculation methods according to the specific type of power distribution room equipment, track layout, sensor configuration, and environmental conditions.

[0019] This invention primarily addresses the problem of misjudgments in track-based patrol robots when relying on single, isolated device observation data for fault diagnosis, which is easily influenced by external environmental factors such as ambient light and nearby heat sources. To address this, instead of directly treating the initial observed anomaly as a device malfunction, this invention further acquires contextual information related to the observed data type or preliminary anomaly result. It establishes a temporal and spatial correspondence between device observation data and environmental interference sources, determines the interference contribution of environmental sources to the preliminary observation features, and, after stripping or interpreting this interference contribution, forms target features reflecting the device's own state. Furthermore, it utilizes the state correlation between target features and the interpretation relationship of environmental interference sources on the observed features to complete fault diagnosis.

[0020] The color compensation algorithm, thermal contribution calculation model, and dynamic correlation graph model in this invention can all be implemented using deterministic calculation methods, without requiring trained artificial intelligence models such as neural networks as a necessary condition. Information such as equipment type and operating status can be determined through equipment ledger and spatial map queries, nameplate character recognition, template matching, color or shape feature judgment, and motion feature analysis. In another optional embodiment, a trained image recognition model can also be used for recognition; its specific model structure and training process will be further explained later.

[0021] Example 1:

[0022] like Figures 1-7 As shown, the fault identification method of the power distribution room track inspection robot includes S1 to S5.

[0023] S1. Acquisition of operational status observation data and triggering acquisition of context environment information; The track-mounted robot runs along a pre-set track within the power distribution room and moves to the monitoring position corresponding to the target power equipment according to a pre-stored equipment map of the power distribution room. The equipment map records at least the equipment identifier, equipment type, monitoring point identifier, spatial location of the monitoring point, corresponding stopping position on the track, and spatial adjacency relationships between equipment or environmental objects around the monitoring point.

[0024] Once the robot reaches the monitoring location, it collects operational status observation data of the target power equipment through its sensing modules. These sensing modules can include visible light cameras and infrared thermal imagers, and can also be configured with one or more of the following, depending on the inspection task: sound sensors, ultrasonic detection modules, and gas sensors.

[0025] For visible light monitoring, operational status observation data may include overall equipment images, indicator light area images, instrument area images, or switch status images; for infrared monitoring, operational status observation data may include infrared thermal images of the target equipment and temperature values ​​of the monitoring area determined based on the thermal images.

[0026] The robot control unit determines whether to trigger context environment information collection based on the data type of the operational status observation data or the preliminary anomaly judgment results.

[0027] When operational status observation data involves indicator light color determination, the indicator light color is defined as an observation type easily affected by ambient light. Therefore, a light environment acquisition task is triggered after obtaining the indicator light image. The robot controls the pan-tilt unit to change the observation direction of the visible light camera, or calls upon a wide-angle auxiliary camera to capture images of areas above, to the side, and other areas that may generate incident or reflected light from the indicator light, thereby obtaining images of potential light sources. Potential light sources include power distribution room lighting fixtures, natural light areas formed by windows, adjacent equipment displays, other high-brightness indicator lights, and surfaces that can generate significant reflected light.

[0028] When the infrared thermal imager detects that the temperature at the target device's monitoring point reaches a preset attention threshold, a thermal environment acquisition task is triggered. The attention threshold is lower than the direct fault alarm threshold; its purpose is to allow the robot to determine whether the external thermal environment can explain the current temperature increase before identifying a fault. The robot uses the device map to locate devices spatially adjacent to the target monitoring point that may exchange heat, and moves to the corresponding acquisition position to obtain visible light images and infrared thermal image data from one or more nearby devices.

[0029] For example, if the target cable joint fault alarm temperature threshold is 80℃, 60℃ can be set as the attention threshold. When the measured temperature of the target joint is 65℃, a fault conclusion is not directly drawn based on this temperature. Instead, the operating status and surface temperature of transformer cabinets, capacitor cabinets, or other high-heat equipment near the target joint are collected to determine how much of the current temperature of the target joint is due to the heat generated by external equipment.

[0030] S2. Establishment of spatiotemporal correspondence and construction of structured data frames; The data generated by the sensing module and the context acquisition task are first correlated over time.

[0031] Each target equipment inspection task generates a unique data frame identifier, and simultaneously records the target equipment identifier, monitoring point identifier, and target observation time. For contextual environment data triggered by this target observation, it is bound to the corresponding data frame within a preset association time range.

[0032] For data from nearby devices that need to be acquired after the robot moves, it is not required that the data be acquired at the exact same physical moment as the target device data. As long as the time difference between the two acquisitions is within a preset time range that can represent the same stage of device operation, a time correlation is established. The time range can be preset based on the load change rate of the monitored device, the robot's moving speed, and the inspection cycle of the power distribution room.

[0033] Further spatial relationship matching is performed on the various data. Based on the location of the tracked robot, the gimbal attitude, and the pre-stored equipment map, the spatial relationships between the target device, indicator lights, potential light sources, and nearby devices are determined. The spatial relationships can include at least one of relative position, distance, and relative orientation.

[0034] After temporal correlation and spatial relationship matching, a structured data frame is generated. A data frame includes at least a subset of device-specific multimodal sensor data, a subset of optical environment images, a subset of thermal environment data, and spatial topology metadata.

[0035] The device's multimodal sensing data subset is used to store the target device's visible light image, infrared thermal image, temperature measurement results, and other device status data obtained according to the configuration.

[0036] A subset of the ambient light image is used to store images of potential light sources corresponding to the current indicator light observation.

[0037] A subset of thermal environment data is used to store images, temperatures, and operating status of nearby devices corresponding to the current temperature observation.

[0038] Spatial topology metadata is used to store device identifiers, monitoring point identifiers, data frame identifiers, acquisition time, data source object identifiers, and the relative position, distance, and orientation between the target device and environmental objects.

[0039] In this way, different environmental data are not added to the analysis as ordinary supplementary data unrelated to the target equipment, but rather a one-to-one or one-to-many correspondence is established with the specific target observation data that it may affect.

[0040] S3. Preliminary observation feature extraction, interference contribution calculation and target feature purification; First, preliminary observation features are extracted from the target equipment's operational status observation data. For indicator light images, the preliminary observation features include at least the observed color of the indicator light; for infrared thermal imaging data, the preliminary observation features include at least the measured temperature of the target monitoring point. Then, based on the knowledge base and contextual information, the environmental interference sources corresponding to these preliminary observation features are determined.

[0041] In this embodiment, "interference contribution" refers to the calculable impact of external environmental factors causing the initial observation characteristics obtained by the sensor to deviate from the true state of the target device. For different observation characteristics, the interference contribution can be represented by numerical values ​​with different physical meanings. For example, ambient light interference contribution can be represented as the degree of shift of each color component in the RGB or HSV color space; external heat source interference contribution can be represented as the temperature increment caused by a nearby heat source.

[0042] First, analysis of ambient light interference with indicator light colors; First, the potential light source image is preprocessed by converting the image to RGB, HSV, or grayscale representation, and then the bright areas are determined by using brightness threshold segmentation.

[0043] In one implementation, the brightness distribution of all pixels in the environmental image can be statistically analyzed, and the set of pixels with higher brightness that can form a continuous area can be identified as candidate light source areas. Then, by combining the area and shape of the candidate areas with the known locations of lamps in the device map, high-brightness reflection points that are obviously not light sources can be eliminated.

[0044] For each potential light source, its color attributes, brightness, imaging area, distance from the target indicator light, and relative angle are obtained.

[0045] The degree of influence of ambient light can be determined according to the following process: First, determine the degree of brightness influence based on the brightness of the potential light source. The brighter the potential light source, the greater the likelihood that it will affect the image color of the target indicator light.

[0046] Secondly, the degree of area influence is determined based on the effective area of ​​the potential light source in the image. Under otherwise similar conditions, a light source that can form a larger effective luminous area has a higher degree of influence than a light source that only forms a small bright spot.

[0047] Secondly, the degree of distance influence is determined based on the distance between the potential light source and the target indicator light. The closer the distance, the higher the interference weight is usually assigned; as the distance increases, the interference weight gradually decreases according to a preset distance attenuation relationship.

[0048] Then, the degree of directional influence is determined based on the relative angle between the light source's illumination direction and the indicator light's surface. The directional influence is higher when the primary illumination direction of the potential light source faces the indicator light's surface; it decreases when the light source is located behind the indicator light or in a position where it cannot be directly illuminated. Finally, the aforementioned influence degrees of brightness, area, distance, and directional influence are combined with the corresponding light source type or the historical ambient light influence coefficients stored in the knowledge base to obtain the comprehensive interference intensity of the potential light source on the current indicator light color observation result. For cases with multiple potential light sources, the comprehensive interference intensity of each potential light source is calculated separately, and then the influence on the RGB color components or HSV color attributes is determined based on the color attributes of each potential light source.

[0049] Taking the RGB color space as an example, we can first obtain the observed red, green, and blue components of the current indicator light area. Then, based on the red, green, and blue components of each potential light source and the corresponding comprehensive interference intensity, we can estimate the additional amount caused by each potential light source to the three color components of the current indicator light. After combining the additional amounts of all effective potential light sources on the same color component, we can subtract or cancel them from the corresponding component of the currently observed color to obtain the compensated color component.

[0050] If the compensation result exceeds the allowed range of RGB color components, it will be limited to the valid range.

[0051] In cases using the HSV color space, one or more of the hue, saturation, and brightness can be adjusted according to the color attributes of the ambient light and the overall intensity of interference. When the ambient light primarily causes an overall brightness bias, brightness and saturation can be adjusted first; when the ambient light has a significant color bias, hue can also be corrected accordingly.

[0052] After compensation, the obtained target color is compared with the color ranges corresponding to normal green light, warning yellow light, fault red light or other states stored in the knowledge base, and the color state that best matches is taken as the target color feature of the indicator light.

[0053] For example, if the original indicator light observation results are generally whitish, while environmental data acquisition indicates the presence of a bright, cool white light above the indicator light, the system determines, based on the light's color, brightness, effective area, distance, and orientation, that it can produce a strong white additive component in the indicator light image. After canceling out this additive component, the compensation result falls within the color range corresponding to a red fault indicator light, and the target color of that indicator light is then determined to be red.

[0054] The color compensation method described above represents only one specific feasible implementation. Alternatively, a lookup table or segmented mapping relationship between different ambient light conditions and color shifts can be established based on the on-site calibration results. However, all of these processes follow the data processing procedure of "obtaining ambient light attributes, determining the degree of ambient light influence, calculating the shift caused by ambient light to the currently observed color, correcting the observed color, and obtaining the target color."

[0055] Second, identification of equipment type and working status; The type of nearby equipment can be determined first based on the equipment map and equipment identification. If the equipment around the target monitoring point has been recorded in the equipment map, the equipment identification and type corresponding to the image can be directly determined based on the robot's current track position, camera orientation, and the spatial area corresponding to the image.

[0056] When the equipment type cannot be directly determined from the equipment map, the equipment nameplate area can be collected and the equipment model can be obtained using character recognition. Alternatively, visual features such as the equipment's outer contour, cabinet door structure, fan location, and instrument distribution can be matched with equipment templates in a knowledge base. The operating status of adjacent equipment can be obtained using deterministic image processing methods. For example, for a cooling fan, multiple frames of images can be continuously collected, and the area where the fan blades are located can be compared between frames. When a clear and continuous change in the blade area appears in adjacent images, the fan is determined to be in operation; when the corresponding area remains basically stable in multiple consecutive frames, the fan is determined to be in a stopped state.

[0057] For equipment operation indicator lights, color and brightness can be extracted based on pre-stored indicator light positions and compared with operation status templates in the knowledge base. For equipment with digital displays or meters, load, current, power, or other display values ​​that can characterize the operation status can be obtained through character recognition. Therefore, this implementation does not require the use of an artificial intelligence model to complete equipment type and operating status identification.

[0058] Third, calculation of thermal interference contribution from nearby equipment; For each neighboring device identified as a potential external heat source, the following information is obtained: neighboring device type, neighboring device operating status, neighboring device surface temperature, distance between the neighboring device and the target temperature measurement point, orientation of the neighboring device relative to the target temperature measurement point, and ambient background temperature.

[0059] The following steps are then used to determine the thermal contribution of the nearby device to the target temperature measurement result.

[0060] First, the corresponding thermal impact coefficient for the adjacent equipment type is retrieved from the knowledge base. Different types of equipment have different heat dissipation methods, heat dissipation structures, and surface radiation characteristics, so different base thermal impact coefficients can be configured. Second, the base thermal impact coefficient is corrected based on the operating status of the adjacent equipment. For example, when the same transformer cabinet is operating under high load, a higher correction coefficient can be used than when it is operating under low load; when it is stopped, a lower correction coefficient can be used. Third, the distance attenuation is determined based on the distance between the adjacent equipment and the target temperature measurement point. The closer the distance, the higher the thermal impact; as the distance increases, the heat contribution decreases according to a preset attenuation law. The distance attenuation relationship can be implemented by looking up a table, segmented decrease, or continuous attenuation. Then, the azimuth correction is determined based on the relative orientation between the main heat dissipation direction of the adjacent equipment and the target temperature measurement point. When the target temperature measurement point is located on the main heat dissipation direction of the adjacent equipment, a higher azimuth correction value is assigned; when the target temperature measurement point deviates from the main heat dissipation direction, the thermal impact is reduced.

[0061] Further calculate the temperature difference between the nearby equipment temperature and the ambient background temperature. The greater the difference between the nearby equipment temperature and the ambient background temperature, the more significant its impact as an external heat source; if the nearby equipment temperature is close to the ambient background temperature, its heat contribution should be correspondingly reduced.

[0062] The thermal influence coefficient of the equipment type, the degree of correction for operating conditions, the degree of distance attenuation, the degree of orientation correction, and the temperature rise of nearby equipment relative to the ambient background are combined to obtain the thermal contribution of the nearby equipment to the target temperature measurement point. This combination can be achieved through a product relationship, or by creating a lookup table or a piecewise model through on-site calibration. The key is to determine a corresponding external thermal contribution based on the above inputs. When multiple effective external heat sources exist around the target equipment, the thermal contribution of each nearby equipment is calculated separately. Then, the thermal contributions are added together or synthesized according to preset combination rules in the knowledge base to obtain the total external thermal contribution of all nearby heat sources to the target temperature measurement point.

[0063] To prevent the total external heat contribution from exceeding the actual temperature rise of the target temperature measurement point relative to the ambient background temperature, an upper limit can be set on the total external heat contribution so that it does not exceed the difference between the measured temperature of the target device and the ambient background temperature.

[0064] The temperature rise of the target device itself is obtained in the following way: First, the total temperature rise of the target temperature measurement point relative to the environment is obtained by subtracting the ambient background temperature from the measured temperature of the target device. Then, the estimated external heat contribution is subtracted from the total temperature rise. The remaining portion represents the temperature rise characteristic caused by the target device's own heat generation.

[0065] For example, if the measured temperature of the target connector is 65℃ and the ambient background temperature is 30℃, then the total temperature rise of the target connector relative to the environment is 35℃.

[0066] If the adjacent transformer cabinet contributes approximately 10°C of external heat and the adjacent capacitor cabinet contributes approximately 2°C of external heat, then the total external heat contribution is 12°C.

[0067] After subtracting the external heat contribution of 12℃ from the total temperature rise of 35℃, the temperature rise of the target joint itself is approximately 23℃.

[0068] The system then uses the self-temperature rise, instead of the measured temperature of 65°C which has not been stripped of environmental conditions, to compare with the normal self-temperature rise range of the same equipment type, the same monitoring point, or similar operating conditions in the knowledge base, thereby determining whether the target equipment itself has abnormal heat generation.

[0069] Fourth, on-site calibration of heat-affected parameters; The equipment type thermal influence coefficient, operating condition correction coefficient, and distance attenuation parameter used in the thermal influence model can be obtained through on-site calibration.

[0070] In one implementation, a relatively stable time period for equipment operation is selected, and the target equipment is used as a fixed temperature measurement object. The temperature of the adjacent equipment, the temperature of the target temperature measurement point, the ambient background temperature, the distance, and the relative orientation are recorded when the adjacent equipment is in a stopped, low-load, normal operation, and high-load state, respectively.

[0071] Under the condition that the load of the target equipment remains basically unchanged, compare the temperature changes of the target temperature measurement point under different operating conditions of the adjacent equipment, and take the temperature increment that can establish a stable correspondence with the state changes of the adjacent equipment as the reference value of the thermal influence of the adjacent equipment.

[0072] Then, the thermal influence coefficient of the equipment type, the degree of correction of the working state, and the degree of distance attenuation are adjusted respectively to gradually reduce the overall difference between the thermal contribution output by the model and the reference temperature increment measured on site.

[0073] When the calculation results under multiple consecutive sets of calibration data are within the preset allowable error range of the actual temperature increment on site, the current parameters are saved to the knowledge base as the initial thermal influence parameters for the corresponding equipment type or the corresponding monitoring point.

[0074] Therefore, the aforementioned thermal impact parameters can be repeatedly obtained from field-collected data, rather than relying on subjective settings whose source cannot be determined.

[0075] Target feature confidence determination: For target features corrected for environmental interference, the corresponding target feature confidence is determined.

[0076] The confidence level of a target feature can be determined by combining the following three types of information: The first category is raw data quality, used to evaluate whether the input data itself is reliable. For example, whether the image is clear, whether it is overexposed, or whether there are any obstructions; whether the infrared thermal image is complete; and whether the temperature measurement area accurately covers the target location.

[0077] The second category is the reliability of interference contribution calculation, which is used to evaluate whether the current environmental interference model has sufficiently reliable input and historical data. For example, whether the spatial location of potential light sources is clear, whether the types of nearby devices have been successfully identified, and whether the thermal contribution parameters have been calibrated with sufficient samples.

[0078] The third category is repeatability consistency, used to evaluate whether the target features obtained under different acquisition conditions remain consistent after environmental compensation. The evaluation results of the above three categories can be converted into a unified confidence level or a unified numerical range, and then weighted and combined according to a preset ratio.

[0079] For example, the weight of the original data quality can be set to 30%, the weight of the confidence of the interference contribution calculation can be set to 40%, and the weight of the consistency of repeated observations can be set to 30%.

[0080] The final confidence level is obtained by summing the three evaluation results after multiplying them by their respective weights.

[0081] The weights mentioned are only for illustrating one specific implementation method and can be adjusted according to device type, sensor characteristics, and historical verification results.

[0082] Active supplementary data collection under low confidence conditions: When the confidence of the target feature reaches or exceeds the preset confidence threshold, the current target feature is input into the subsequent dynamic association graph model.

[0083] When the confidence level of the target feature is lower than the preset confidence threshold, it is considered that the existing data is insufficient to effectively distinguish the target device's own state from environmental interference, and the strategy module generates a supplementary collection instruction.

[0084] Regarding the indicator light color, the robot can be controlled to change its acquisition position along the track or adjust the gimbal attitude so that the camera forms different observation angles relative to the indicator light and potential light source, and then acquire images of the target indicator light and related potential light source again.

[0085] You can also change the exposure time, gain, or white balance parameters according to the actual functions of the camera and then take another picture.

[0086] If the original indicator light color changes significantly after the acquisition conditions are changed, but the target color remains within the same color range after compensation is performed again based on the new ambient light data, it indicates that the original change mainly comes from the change in ambient light, thereby improving the consistency evaluation of repeated observations of the target color.

[0087] In case of temperature anomalies, the robot can be controlled to change the infrared temperature measurement position or angle, and expand the scanning range of nearby devices to collect temperature and spatial relationship information of the target device and nearby heat sources again.

[0088] The heat contribution of each nearby heat source, the total external heat contribution, and the temperature rise of the target equipment itself were then recalculated.

[0089] If the self-temperature rise obtained from different acquisition locations all fall within the same abnormal state range, the consistency evaluation of repeated observations of the target temperature characteristic will be improved.

[0090] The supplementary collected data is correlated with the original data frame, or a new data frame with a retest relationship with the original data frame is created, and spatiotemporal correlation, interference contribution calculation and target feature purification are re-executed, thereby forming an active verification closed loop.

[0091] S4. Construction of dynamic association graph model; After completing the environmental interference verification, a dynamic correlation graph model is constructed based on the preliminary observation characteristics, environmental interference sources, and target characteristics.

[0092] The dynamic association graph model consists of nodes and the connections between them.

[0093] Nodes must include at least three types.

[0094] The first category is observation feature nodes, used to store preliminary observation features directly obtained from the original operating status observation data, such as "the A-phase indicator light of #1 incoming cabinet is whitish" or "the measured temperature of the X connector of #1 incoming cabinet is 65℃". The second category is interference source nodes, used to indicate identified environmental interference sources, such as "the L1 ceiling LED light is strong white light" or "the #2 transformer cabinet is operating at high temperature". The third category is target feature nodes, used to indicate features reflecting the state of the target equipment after interference contribution verification or correction, such as "the A-phase indicator light of #1 incoming cabinet is red after compensation" or "the temperature rise of the X connector of #1 incoming cabinet is 23℃". Each node can record the node type, corresponding equipment, monitoring point, feature type, feature value, acquisition time, and target feature confidence level. At least interference interpretation relationships and state correlation relationships must be established between nodes.

[0095] The interference interpretation relationship points from the interference source node to the corresponding observation feature node, and is used to indicate the extent to which a certain environmental interference source can interpret a certain original observation anomaly.

[0096] For example, an optical interference interpretation relationship is established between the L1 node of the ceiling LED light and the node of "indicator light color is too white"; an electrical interference interpretation relationship is established between the #2 transformer cabinet high temperature operation node and the node of "connection measured temperature 65℃".

[0097] Each interference interpretation relationship is assigned a corresponding interpretation weight. For thermal interference interpretation weights, the proportion of the thermal contribution from a nearby device to the total temperature rise of the target temperature measurement point relative to the ambient background can be calculated first. The higher the proportion, the better the nearby device can explain the current temperature rise of the target temperature measurement point, and therefore the higher the corresponding interference interpretation weight.

[0098] When the thermal contribution calculated by a nearby device can explain most of the temperature rise at the target temperature measurement point, a higher thermal interference interpretation weight can be assigned; when its thermal contribution accounts for only a small part of the total temperature rise, a lower thermal interference interpretation weight is assigned.

[0099] To interpret the color interference weights, we can compare the color shift before and after ambient light compensation, as well as the total shift of the original color relative to the standard state color.

[0100] If ambient light compensation can eliminate most of the original color shift, it indicates that the ambient light can strongly explain the original color anomaly, and the corresponding interference explanation weight is high. If the color remains basically unchanged after compensation, it indicates that the ambient light's ability to explain the current color anomaly is weak, and the corresponding weight is low. The above explanation weights can be uniformly converted to a numerical range of 0 to 1, or they can be divided into discrete levels of low, medium, and high.

[0101] State association is used to connect two target feature nodes, indicating that there is an electrical, spatial, or temporal relationship between the actual state features of the two devices after environmental interference purification.

[0102] State association weights can be determined by comprehensively considering electrical association, spatial association, and temporal association.

[0103] The degree of electrical correlation is determined based on whether the two target characteristics belong to the same equipment, the same phase, the same circuit, or have an upstream or downstream electrical connection. The degree of spatial correlation is determined based on the distance and installation location between the two monitoring points; the closer the spatial distance and the more likely they are to influence each other structurally, the higher the degree of spatial correlation.

[0104] The degree of temporal correlation is determined by the proximity of the occurrence times of two abnormal features. When two anomalies occur in the same inspection or within a short period of time, the degree of temporal correlation is high; when the interval is long, the degree of correlation decreases accordingly.

[0105] In one specific implementation, electrical correlation can be used as the primary factor, temporal correlation as a secondary factor, and spatial correlation as a further auxiliary factor. For example, electrical correlation accounts for 50% of the total state correlation evaluation, temporal correlation accounts for approximately 30%, and spatial correlation accounts for approximately 20%. The three correlation levels are weighted and combined according to their corresponding proportions to obtain the state correlation weight between the two target feature nodes.

[0106] In a further embodiment, a feature correction relationship can be established between the observed feature node and its corresponding target feature node to record the process of obtaining the target feature after the original observed features have been corrected for environmental interference. The dynamic association graph model is not fixed. After the robot obtains new observation data, it can update the node attributes of the same monitoring point within a preset time range, or it can create new time-series nodes and recalculate the interference interpretation weight and state association weight related to the node. Historical nodes that exceed the current effective diagnostic time range can be retained in the knowledge base or removed from the active subgraph used in the current diagnosis.

[0107] The dynamic association graph model in this embodiment is a data structure that is analyzed based on node attributes, node relationships, weights, and knowledge base rules, and does not require training using a graph neural network.

[0108] S5. Dynamic correlation diagram analysis and fault diagnosis; After the dynamic association graph is established, the degree of abnormality of each target feature is first determined based on the normal state range of the equipment in the knowledge base.

[0109] For temperature target characteristics, the degree of anomaly can be determined by the extent to which the target device's own temperature rise exceeds the upper limit of normal temperature rise. When the temperature rise is within the normal range, the degree of anomaly is low; the greater the exceedance of the upper limit, the higher the degree of anomaly. For indicator light target color characteristics, the degree of anomaly can be determined by the device status corresponding to the compensated color. For example, a normal green after compensation indicates a low degree of anomaly, while a faulty red after compensation indicates a high degree of anomaly.

[0110] The effective anomaly level of a single target feature is further determined by combining the target feature confidence level. When the target feature has a high anomaly level and a high confidence level, its contribution to subsequent fault diagnosis is greater; when the target feature shows anomalies but the confidence level is low, its contribution to fault diagnosis is correspondingly reduced.

[0111] Then, the state correlations between multiple target features are analyzed. When two or more high-confidence target anomaly features have high state correlation weights, these nodes are combined into a candidate fault subgraph.

[0112] For example, if the target characteristics of the fault indicator light and the target characteristics of the abnormal temperature rise of the cable joint appear simultaneously in the same phase circuit, and the two are strongly correlated in terms of electrical relationship and occurrence time, they are analyzed as the same candidate fault event.

[0113] Candidate fault scores can be determined using the following steps: First, calculate the combined result of the anomaly degree of each target feature and its corresponding confidence level; Secondly, the combined results of all target features in the candidate fault subgraphs are summarized; Secondly, correlation weighting is added to target features that have strong state correlations with each other, so that multiple mutually supportive anomalous features have higher fault confidence than a single isolated anomalous feature; finally, the obtained comprehensive results are converted into a unified fault confidence level or a preset numerical range.

[0114] Therefore, when multiple high-confidence real anomaly features mutually support each other electrically, temporally, or spatially, the system can improve the diagnostic confidence of the corresponding fault candidates, rather than directly triggering an alarm based solely on a single sensor threshold. Simultaneously, the system analyzes the interpretability of environmental interference nodes.

[0115] If a certain original observation feature shows an anomaly, but the corresponding interference interpretation weight is high, and the target feature obtained after environmental interference correction is within the normal range, then the original anomaly is judged as a pseudo-anomaly mainly caused by environmental interference.

[0116] When the original observed anomaly still corresponds to a high degree of target anomaly after environmental interference correction, and this target feature forms a strong state correlation with other real anomaly target features, the diagnostic confidence of the real fault candidate is increased. The candidate fault type is further determined based on the fault association rules stored in the knowledge base.

[0117] For example, when the target characteristics of a fault indicator light and the target characteristics of abnormal temperature rise of a connector coexist in the same phase circuit, and the correlation between their states exceeds a preset correlation condition, a suspected fault conclusion can be generated for that phase circuit, indicating a contact abnormality, overload, or other faults corresponding to the knowledge base rules. If the current target characteristics cannot uniquely determine the specific fault type, only the suspected fault type and its corresponding confidence level are output; a unique fault cause is not forcibly generated. The final diagnostic conclusion may include the target equipment identifier, target monitoring point, abnormal target characteristics, identified or eliminated major environmental interference factors, suspected fault type, and diagnostic confidence level.

[0118] Adjust robot behavior based on diagnostic results: When the diagnostic confidence reaches the high confidence fault threshold, the strategy module generates fine retest and alarm instructions, controls the robot to return to or remain in the target monitoring area, performs multi-angle, high-resolution retests on abnormal areas, and further collects sound, ultrasound, or other data that can provide corroboration based on the robot's existing sensor configuration.

[0119] When the initial anomaly can be primarily explained by environmental interference, no alarm of the same level as a real equipment failure is generated. Instead, relevant interference information is sent to the knowledge base update module. When the diagnostic result is still in a suspected state or the confidence level of the target feature is insufficient, the strategy module generates a supplementary data acquisition task, causing the robot to change its position, gimbal angle, acquisition range, or sensor acquisition parameters. The new acquired data is then re-input into the fusion processing module, anti-interference verification module, and dynamic correlation graph model. This forms a feedback loop of "equipment observation, environmental verification, fault diagnosis, active retesting, and re-diagnosis."

[0120] Knowledge base structure and updates: The knowledge base of this implementation method may include at least an equipment model table, a monitoring point spatial relationship table, an environmental interference mode table, and a fault association rule table.

[0121] The equipment model table records the equipment type, normal parameter range, typical fault characteristics, and corresponding judgment conditions.

[0122] The monitoring point spatial relationship table records the monitoring point identifier, corresponding equipment, monitoring location, standard camera orientation, list of nearby equipment, and potential environmental interference objects. The environmental interference pattern table records at least the monitoring point identifier, interference source identifier, interference source type, interference occurrence conditions, historical interference contribution parameters, and the number of verified samples. The fault association rule table records the electrical or state association relationships between different target anomaly characteristics and their corresponding candidate fault types.

[0123] When an anomaly is confirmed through supplementary data collection or other verification methods to be primarily caused by a specific environmental factor, this data is compiled into an environmental interference pattern sample. For indicator light interference samples, the monitoring point, ambient light source type, occurrence time, light source color, light source brightness, spatial orientation, original color, compensated color, and ambient light influence parameters used in this instance can be recorded. For thermal interference samples, the monitoring point, type of nearby equipment, operating status, distance, orientation, temperature of nearby equipment, temperature of the target equipment, ambient background temperature, and the calculated thermal contribution parameters can be recorded. After accumulating multiple verified environmental interference samples for the same monitoring point, the interference contribution parameters corresponding to that monitoring point can be updated.

[0124] In one specific update method, the parameters of the saved historical samples of the same type are averaged, and then the parameters corresponding to the new validation samples are added to the calculation. As the number of validation samples increases, the influence of each new sample on the final parameters gradually decreases, causing the updated parameters to gradually stabilize.

[0125] For example, a monitoring point has obtained a stable thermal impact parameter based on 10 validation samples. When the 11th validation sample is obtained, the historical parameters formed by the previous 10 samples can be considered with a weight of 10, and the new sample parameters can be considered with a weight of 1. Then, the average value can be calculated again to obtain the updated thermal impact parameter.

[0126] The same incremental update method can be used for ambient light impact parameters. Knowledge base parameter updates only use verified samples that can be confirmed as environmental interference to avoid mistakenly writing real equipment malfunctions as environmental interference into the knowledge base. When the same type of environmental interference repeatedly occurs at a monitoring point historically, the triggering conditions for collecting contextual environmental information can be changed. For example, if an indicator light monitoring point is repeatedly affected by a fixed overhead light, the original "collect the overhead light information only after detecting color anomalies" can be adjusted to "collect the overhead light information directly when the robot arrives at the monitoring point to perform indicator light inspection." For a temperature monitoring point, if historical data indicates that an adjacent transformer will stably generate significant external thermal impact under high load conditions, its temperature information can be obtained in advance and included in the current thermal contribution calculation when the adjacent transformer is subsequently detected to be operating under high load.

[0127] In the above embodiments, device type and device operating status are not necessarily determined by artificial intelligence models. When computing resources permit and there is a need to improve visual recognition capabilities in complex environments, a trained convolutional neural network can be used to identify the types and operating status of neighboring devices.

[0128] In one specific implementation, a residual convolutional neural network with a common image feature extraction part and two classification output parts is used.

[0129] The model input is an RGB image of the device region cropped from a neighboring device image. The input image is uniformly scaled or cropped to 224 pixels × 224 pixels, and the pixel values ​​of the three RGB channels are normalized.

[0130] The common image feature extraction part first sets up a convolutional layer with a kernel size of 7×7, a stride of 2, and 64 output channels, and then sets up a max pooling layer with a kernel size of 3×3 and a stride of 2.

[0131] Four residual stages are set sequentially after the max pooling layer.

[0132] The first residual stage includes two basic residual blocks, with a total of 64 output channels.

[0133] The second residual stage includes two basic residual blocks, with a total of 128 output channels.

[0134] The third residual stage includes two basic residual blocks, with a total of 256 output channels.

[0135] The fourth residual stage includes two basic residual blocks, with a total of 512 output channels.

[0136] Each basic residual block consists of two convolutional layers with a kernel size of 3×3. The input features of this residual block are residually concatenated with the output features of the second convolutional layer, thus preserving previous feature information and improving gradient propagation during deep network training. Between adjacent residual stages, a convolution with a stride of 2 is used to reduce the feature map size and adjust the number of channels. After the output of the fourth residual stage, global average pooling is used to convert the two-dimensional feature map into 512-dimensional image features. The resulting 512-dimensional image features are then input into the device type classification branch and the operating status classification branch, respectively.

[0137] The device type classification branch includes a fully connected layer and a Softmax output layer, used to output the probability that the current image belongs to each preset device type. Preset device types include, for example, transformer cabinets, capacitor cabinets, and switch cabinets. The operating state classification branch includes a fully connected layer and a Softmax output layer, used to output the probability that the current device belongs to the stop, low load, normal operation, and high load operating states, respectively. For devices that only need to determine whether a specific actuator is working, the operating state output can also be set to two categories: "Running" and "Stopped".

[0138] Artificial intelligence model training data: The training data can consist of equipment images collected from the current power distribution room site and similar power distribution equipment under different viewing angles, different lighting conditions and different operating states.

[0139] Each training image must have at least a device type label and a device operating status label.

[0140] Equipment type labels are determined based on equipment ledgers, nameplates, or manual verification results.

[0141] The equipment operating status label is determined based on the equipment operation records, load data, or on-site verification results.

[0142] Images whose working status cannot be reliably confirmed from operation records or on-site information will not be used as valid training samples for the corresponding working status classification task, in order to avoid incorrect labels affecting the model.

[0143] The training data can be divided into training set, validation set and test set according to the proportions of 70%, 15% and 15% respectively.

[0144] Before training, random cropping, slight translation, brightness adjustment, and contrast adjustment can be performed on the training set images to enhance the model's ability to adapt to different shooting positions and lighting changes in the power distribution room.

[0145] Image transformation methods that may alter the meaning of the actual state of the device are not used.

[0146] Artificial intelligence model training process: During the training process, the classification error of equipment type and the classification error of equipment working status are calculated simultaneously.

[0147] Equipment type classification error is used to evaluate the difference between the equipment type probability output by the model and the actual equipment type label.

[0148] The job status classification error is used to evaluate the difference between the job status probability output by the model and the actual job status label.

[0149] In one specific approach, the two classification errors are combined with the same weights to form the overall training error.

[0150] The ratio of the two can also be adjusted according to the importance of the actual task. For example, when the device type can be reliably obtained through a map, and the identification of the working status is more important, the proportion of the working status classification error in the overall training error can be increased.

[0151] The model parameters are updated using the Adam optimization method.

[0152] One specific training method is as follows: the initial learning rate is set to 0.001, 32 training images are input in each batch, the weight decay coefficient is set to 0.0001, and the maximum number of training rounds is set to 50 rounds.

[0153] After each round or several rounds of training, the classification accuracy and validation error are calculated on the validation set. Training continues as long as the validation error continues to decrease; if the validation error stops decreasing after several consecutive rounds of training, the learning rate can be reduced or training can be stopped early to avoid the model overfitting to the training data.

[0154] After training, the accuracy of device type identification and working status identification is checked using a test set that was not involved in parameter updates.

[0155] Once the test results meet the preset application requirements, the trained model parameters are fixed and deployed on the local industrial control computer of the track robot or an edge computing server connected to the track robot. The input-output relationship of the artificial intelligence model in this invention is as follows: During actual operation, the RGB image of the adjacent device area is input into the aforementioned artificial intelligence model. The device type classification branch outputs the probability corresponding to each device type and selects the device type with the highest probability as the device type identification result. The working state classification branch outputs the probability corresponding to each working state and selects the working state with the highest probability as the device working state identification result. Simultaneously, the highest probability is used as the reliability of the corresponding identification result.

[0156] When the highest probability is lower than the preset recognition confidence threshold, such as below 80%, the recognition result is not directly used for subsequent thermal contribution calculation. Instead, the device map, nameplate recognition, or the robot is controlled to re-collect a clearer image for confirmation.

[0157] The output of the AI ​​model is not used independently to draw fault conclusions. Equipment type identification results are used to select the corresponding thermal impact coefficient from the knowledge base. Operating status identification results are used to select the appropriate operating status correction level. Subsequently, the thermal contribution of the neighboring equipment to the target equipment is calculated by combining the temperature, distance, relative orientation, and ambient background temperature of the neighboring equipment.

[0158] Therefore, the input to the artificial intelligence model is an image of nearby devices, and the output is the device type and operating status. This output is further used as input for calculating the thermal interference contribution, thus participating in the separation of the target device's own temperature rise. Through this data association, the artificial intelligence model in this invention plays a technical role in providing device attributes and operating status parameters for external thermal impact assessment, rather than directly equating image classification results with device malfunctions. In other embodiments, other convolutional neural networks, object detection networks, or visual classification networks can be used instead of the aforementioned residual convolutional neural network, as long as they can output the device type and operating status information required for subsequent thermal contribution calculation based on the image of nearby devices.

[0159] Taking the inspection of the #1 incoming line cabinet by a track robot as an example.

[0160] After the robot reaches the monitoring position of the #1 incoming line cabinet, it acquires visible light images and infrared thermal images of the panel.

[0161] The A-phase indicator light in the panel image appears whitish; the X-connector temperature is measured at 65℃, which is above the 60℃ threshold.

[0162] The system establishes a data frame corresponding to this inspection. Since the observation content includes indicator light colors, the robot collects environmental images of the ceiling LED light L1 and the window light area. Because the X-connector temperature reaches the attention threshold, the robot further collects images and infrared temperatures of the adjacent #2 transformer cabinet and #3 capacitor cabinet. All data is written to the corresponding data frame, and the acquisition time, equipment identification, monitoring point identification, distance, and relative orientation are recorded.

[0163] Regarding the indicator light color, the system determines the color attributes, brightness, effective area, distance, and relative orientation of LED L1 and the window light, and calculates the overall interference level they cause to the observed indicator light color. Then, based on the color components and corresponding interference levels of the two ambient light sources, the system determines the additional components they cause to the observed indicator light color in the RGB color space, and cancels these additional components out of the original observed color. The compensated color is then added to the red fault indicator light color range in the knowledge base, thus generating the target feature "#1 incoming line cabinet A-phase indicator light target color is red". Regarding temperature anomalies, it is determined that #2 equipment is a transformer cabinet operating under high load with a surface temperature of 85℃; #3 equipment is a capacitor cabinet with a surface temperature of 40℃. The system calculates the external thermal contribution of each of the two devices to the target connector based on their type, operating status, temperature, distance, and relative orientation.

[0164] For example, calculations show that transformer cabinet #2 contributes approximately 10°C to the heat contribution of the target joint, and capacitor cabinet #3 contributes approximately 2°C. The total external heat contribution from the two adjacent devices is approximately 12°C. With an ambient background temperature of 30°C and a measured temperature of 65°C at the target joint, the total temperature rise of the target joint relative to the environment is 35°C.

[0165] After subtracting approximately 12°C of external heat contribution from the total temperature rise of 35°C, the temperature rise of the target connector itself is approximately 23°C. This results in the target characteristic that "the temperature rise of the #1 incoming cabinet X connector itself is approximately 23°C".

[0166] The dynamic correlation graph model should establish at least the following observation feature nodes: "Indicator light color is white", "LED light L1 strong white light" interference source node, "A phase indicator light target color is red", "X joint measured temperature 65℃" observation feature node, "#2 transformer cabinet high heat operation" and "#3 capacitor cabinet heat up" interference source nodes, and "X joint temperature rise of about 23℃".

[0167] An optical interference interpretation relationship is established between the LED light L1 node and the indicator light observation color node. Thermal interference interpretation relationships are established between the #2 transformer cabinet and #3 capacitor cabinet nodes and the measured temperature node of the X-connector. Since the heat contribution from the #2 transformer cabinet is higher than that from the #3 capacitor cabinet, the thermal interference interpretation weight corresponding to the #2 transformer cabinet is higher than that of the #3 capacitor cabinet. Simultaneously, since the A-phase indicator light and the X-connector belong to the same electrical circuit, a state correlation relationship with a higher weight is established between the "A-phase indicator light target color red" node and the "X-connector's own temperature rise abnormal" node.

[0168] During the initial analysis, if the system's overall fault diagnosis confidence level is lower than the preset verification threshold, a final high-confidence fault conclusion is not immediately formed. Instead, active supplementary data acquisition is triggered. The robot changes its acquisition position along the track and adjusts the camera's observation angle to acquire data on the indicator light, LED L1, target connector, and nearby devices again. If the color of the indicator light in the second, uncompensated instance differs significantly from the first due to the change in observation angle, but is still determined to be red after corresponding ambient light compensation, it indicates that the compensation result has high repeatability consistency. Simultaneously, if the temperature rise of the target connector itself, recalculated based on the second thermal environment data, is still within the abnormal range, it indicates that the abnormal temperature rise is not primarily caused by nearby heat sources or single observation errors.

[0169] Based on this, the system improves the consistency evaluation of repeated observations of the two target features and the confidence level of the comprehensive fault diagnosis. The dynamic association graph model updates node attributes and association weights based on the second data collection. After re-analysis, if the comprehensive fault diagnosis confidence level exceeds the preset high confidence threshold, the system outputs a diagnostic conclusion that there is a suspected equipment abnormality in the A-phase circuit of the #1 incoming cabinet, indicating that the indicator light abnormality and the abnormal temperature rise of the connector are mutually supported, and that the main ambient light interference and the thermal impact of adjacent equipment have been identified and corrected. Based on this, the strategy module generates detailed fault retesting and alarm control commands. For the confirmed LED light L1 light interference and #2 transformer cabinet thermal interference, knowledge base update data is generated simultaneously, and the corresponding environmental interference mode is associated with and saved with the corresponding monitoring point of the #1 incoming cabinet for subsequent inspection to optimize interference contribution parameters or context acquisition trigger conditions.

[0170] The experiment of this invention was carried out in a pilot area of ​​a 10kV substation, using a combination of laboratory test bench and field verification. All experimental conditions can be reproduced in reality, and the configuration is shown in Table 1. Table 1. Laboratory-related configuration parameters

[0171] Experiment 1: Verification of Ambient Light Interference Compensation for Indicator Light Colors Experimental Procedure: A simulated power distribution room indicator panel was constructed on a laboratory bench, with three typical lighting scenarios set up: cool white LED lighting (color temperature approximately 6000K), warm LED lighting (color temperature approximately 3000K), and mixed lighting (cool white LED above + warm LED on the side lit simultaneously). The target indicator was a standard red fault indicator (RGB reference value (200, 30, 30)). Thirty images of the indicator were captured for each scenario, and the observed color features before compensation (original observed features) and the target color features after compensation were extracted. The compensation process was performed according to the steps described in S3 of this invention: identifying potential light sources → calculating the color components of the light source and the overall interference intensity → subtracting additional components in the RGB color space → obtaining the target color. The color distance was defined as the RGB Euclidean distance (normalized 0-100) between the observed color and the standard red reference value. The smaller the distance, the closer the interpretation is to the true red, as shown in Table 2.

[0172] Table 2 Experimental data for three typical lighting scenarios

[0173] Figure 4 The changes in color distance before and after compensation are visually illustrated using a dumbbell diagram. All three dumbbell lines cross the interpretation tolerance threshold (25): before compensation, the color distances for the three scenarios are 46-64, all exceeding the threshold and would be misjudged as white / orange or other non-red states using traditional methods; after compensation, the color distances decrease to 12-18, all falling within the red interpretation range. The standard deviation of the distance after compensation (2.1-2.6) is significantly smaller than before compensation (3.8-5.1), indicating that after removing ambient light interference, color interpretation is not only accurate but also exhibits good consistency under different lighting conditions. This invention compensates for indicator light colors by actively collecting potential light source information and calculating the contribution of light interference, increasing the color interpretation accuracy in three lighting scenarios from 0% to 100% (30 / 30 correctly judged as red).

[0174] Experiment 2: Verification of Thermal Interference Stripping from Adjacent Equipment Experimental Procedure: Using the X cable connector of the #1 incoming line cabinet as the target temperature measurement point, 30 sets of infrared thermal imaging data were collected under four scenarios: ① neighboring equipment stopped; ② neighboring equipment under low load; ③ neighboring equipment under high load; ④ target equipment with minor fault + neighboring equipment under high load. The ambient background temperature was kept constant at 30℃. For each set of data, the following were calculated: the measured temperature of the target temperature measurement point (including external thermal interference), and the equivalent self-state temperature after removing the external thermal influence according to the thermal contribution calculation process described in S3 of this invention (= ambient background temperature 30℃ + self-temperature rise). The equivalent self-state temperature reflects the true heating state of the target equipment, as shown in Table 3.

[0175] Table 3 Results of 30 sets of infrared thermal imaging data collected in each of the four scenarios

[0176] Figure 5 The dumbbell diagram illustrates the comparison between the measured temperature and the equivalent self-state temperature after stripping. The key scenario is the third row, where nearby equipment is under high load: the measured temperature of 65℃ exceeds the attention threshold (60℃), and the traditional single-feature threshold method would trigger a suspected anomaly alarm. However, after calculating the thermal contribution from nearby equipment (approximately 10℃ from the high-load operation of transformer #2, approximately 2℃ from capacitor #3, and approximately 12℃ from the total external thermal contribution), the equivalent self-state temperature is only 53℃, within the normal range (≤60℃), indicating that the anomaly is actually caused by thermal radiation from nearby equipment and is a false anomaly. In the fourth row scenario, the equivalent self-state temperature of 63℃ after stripping external thermal contributions still exceeds the normal upper limit, indicating a genuine anomaly in the target equipment itself. This invention, by identifying nearby heat sources and calculating thermal interference contributions, corrects the equivalent self-state temperature in scenario ③ from the measured 65℃ to 53℃ (within the normal range), effectively eliminating false alarms caused by high load from nearby equipment; simultaneously, the genuine anomaly in scenario ④ is not masked (the equivalent temperature of 63℃ still exceeds the upper limit), ensuring that genuine anomalies are not missed and false anomalies are not falsely reported. The results of 30 repeated experiments were consistent (Scenario ③ 30 / 30 was judged as normal, Scenario ④ 30 / 30 was judged as abnormal).

[0177] Experiment 3: Dynamic Association Graph Joint Diagnosis and Active Retesting Closed-Loop Validation Experimental process: 60 simulated inspection scenario samples were constructed (samples were generated by combining on-site data according to working conditions, including three types of observations: indicator light color, connector temperature, and status of adjacent equipment). Among them, 30 were real fault scenarios (A phase indicator light red + X connector itself temperature rise abnormal, the two belong to the same electrical circuit), and 30 were environmental interference pseudo-abnormal scenarios (only ambient light interference or adjacent heat source interference exists, the target equipment itself is normal).

[0178] Two methods were used to process each sample: Method A (traditional single-feature threshold method): Thresholds are set independently for indicator light color and connector temperature. An alarm is triggered if any feature exceeds the threshold. There is no environmental verification or retesting mechanism. Method B (this invention): Execution context environment information collection → Interference contribution calculation and feature correction → Dynamic association graph construction → Joint analysis of state association and interference interpretation → Active supplementary collection and retesting when confidence is insufficient.

[0179] The false alarm rate (the proportion of false alarms that are mistaken for faults in pseudo-anomaly scenarios), the false negative rate (the proportion of false negatives that are missed in real fault scenarios), and the diagnostic accuracy of the two methods were statistically analyzed. Meanwhile, the average diagnostic confidence of the method of this invention as the number of active supplementary data collections increased (0, 1, 2, and 3 times, the average of 30 real fault scenarios) is shown in Tables 4 and 5.

[0180] Table 4. Comparison of diagnostic efficacy between the two methods (60 samples)

[0181] Table 5. Number of times the invention actively supplements data collection and the average diagnostic confidence (mean of 30 real fault scenarios).

[0182] Figure 6 The false alarm rate, false negative rate, and diagnostic accuracy of the two methods were compared. The method of this invention reduces the false alarm rate from 33.3% to 3.3% (a reduction of approximately 90%), the false negative rate from 16.7% to 3.3%, and the diagnostic accuracy from 75.0% to 96.7%. The significant reduction in the false alarm rate is directly due to the environmental interference verification mechanism—of the 10 false anomaly samples initially misjudged as faults, 9 were correctly identified as being caused by environmental interference after calculation and correction of the contribution of light / thermal interference, and no longer triggered fault alarms. The reduction in the false negative rate is due to the compensation for the masking of indicator light colors by ambient light, allowing the true fault characteristics to be restored (in the traditional method, 4 true faults were missed because the indicator light colors were diluted by strong light; this invention correctly identified them after color compensation). Figure 7 This paper demonstrates the change in average diagnostic confidence of the method of this invention with the increase of the number of active supplementary data collections. During the initial inspection, the average confidence was only 0.61, lower than the high-confidence fault judgment threshold (0.85), insufficient to form a final conclusion. After one supplementary data collection, it rose to 0.75, reached 0.86 after two collections and crossed the threshold, and stabilized at 0.92 after three collections. The standard deviation of the confidence level narrowed from 0.045 to 0.028, indicating that the retesting mechanism not only improved the confidence level but also enhanced the stability of the judgment. Traditional methods lack a retesting mechanism, and the confidence level remains at the initial judgment level (mean 0.70 ± 0.05), unable to eliminate uncertainty through supplementary observations.

[0183] Example 2:

[0184] This embodiment also provides a fault identification system for a power distribution room track inspection robot, including a sensing module, a context scheduling module, a fusion processing module, an anti-interference verification module, a knowledge base, a feature purification module, a model building module, a diagnostic module, a strategy module, and a robot motion control module.

[0185] The sensing module is located on the main body of the tracked robot and is used to collect observation data on the operating status of the target power equipment and to obtain environmental information based on the context of the task.

[0186] The context scheduling module receives standard inspection tasks and preliminary observation results, and generates corresponding light environment acquisition tasks, thermal environment acquisition tasks, or low-confidence supplementary acquisition tasks based on the data type of the observation or preliminary anomaly results.

[0187] The fusion processing module receives target device data and context environment data respectively, performs time correlation according to the corresponding data frames and spatial relationship matching according to the device map to generate structured data frames.

[0188] The anti-interference verification module includes an optical interference processing unit and a thermal correlation analysis unit.

[0189] The light interference processing unit receives data on the observed color of the indicator lights, the color, brightness, effective area, and spatial relationships of potential light sources. It calculates the ambient light color contribution based on the interference level of each potential light source and outputs the compensated target color and its corresponding confidence level.

[0190] The thermal correlation analysis unit receives the measured temperature of the target device, the ambient background temperature, the type, operating status, temperature, distance and orientation of the neighboring devices, calculates the thermal contribution of each neighboring device, and then obtains the total external thermal contribution and the temperature rise characteristics of the target device itself.

[0191] The knowledge base stores normal equipment parameters, equipment spatial topology relationships, color status templates, thermal contribution model parameters, historical environmental interference patterns, and fault association rules.

[0192] The feature purification module organizes the anti-interference verification results into a target feature set and adds corresponding confidence scores to the target features.

[0193] The model building module establishes a dynamic correlation graph model based on observation features, environmental interference sources, and target features, and updates the nodes and the correlation weights between nodes based on new inspection data.

[0194] The diagnostic module analyzes the dynamic correlation graph model based on the target's anomaly level, target feature confidence, state correlation, and interference interpretation to generate fault diagnosis conclusions.

[0195] The strategy module generates subsequent inspection control instructions based on the diagnostic conclusions or the confidence level of the target features.

[0196] When additional data collection is needed, the robot motion control module controls the robot to move to a new track position, adjust the gimbal attitude, or change the sensor data collection status.

[0197] The supplementary data is sent again from the sensing module to the fusion processing module, and then sequentially enters the anti-interference verification module, feature purification module, model building module and diagnosis module, thus forming a feedback fault verification closed loop.

[0198] Example 3:

[0199] This embodiment also provides a computer-readable storage medium having a computer program stored thereon.

[0200] The computer-readable storage medium may be a solid-state drive, an embedded multimedia card, flash memory, or other non-transitory storage medium capable of storing computer programs.

[0201] When the computer program is executed by the processor in the local industrial control computer of the track robot or the edge computing server connected to the track robot, the processor performs the above-mentioned steps of operating status observation data acquisition control, context acquisition task generation, spatiotemporal data association, structured data frame construction, environmental interference source identification, interference contribution calculation, target feature purification, dynamic association graph model construction and updating, fault diagnosis, active supplementary acquisition control, and knowledge base update.

[0202] The specific values, network layers, weight ratios, thresholds, and calculation processes described in the above embodiments are used to illustrate at least one specific way to achieve the technical solution of the present invention. Those skilled in the art can calibrate or adjust the corresponding parameters based on the type of power distribution room equipment, sensor characteristics, robot track position, and historical inspection data. As long as the corresponding environmental information is obtained according to the observation results of the target equipment, the interference contribution is determined based on the environmental interference state and spatial relationship, and the target characteristics of the equipment itself are obtained after stripping or interpreting the environmental interference, and fault diagnosis is completed based on the state correlation between target characteristics and the interference interpretation relationship of environmental interference sources on the observed characteristics, the technical solution described in the present invention can be achieved.

[0203] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A fault identification method for a power distribution room track inspection robot, characterized in that, Includes the following steps: S1. Control the track robot to collect the operating status observation data of the target power equipment, and trigger the collection of contextual environment information that may affect its interpretation based on the data type of the operating status observation data or the preliminary anomaly judgment result. S2. Perform time correlation and spatial relationship matching between the observed operating status data and the context environment information, establish the spatiotemporal correspondence between the two, and construct a data set containing the device body status, environmental status and spatial topology relationship; S3. Extract preliminary observation features from the operating status observation data, identify corresponding environmental interference sources from the context environment information based on the pre-stored knowledge base, determine the interference contribution of the environmental interference source to the preliminary observation features according to the state attributes of the environmental interference source and its spatial topological relationship with the target power equipment, and verify or correct the preliminary observation features based on the interference contribution to obtain the target features. S4. Construct a dynamic correlation graph model based on the target features, environmental interference sources and their correlation relationships. The dynamic correlation graph model at least represents the state correlation relationship between target features and the interference interpretation relationship of environmental interference sources on the corresponding observed features. S5. Analyze the dynamic correlation graph model to generate fault diagnosis conclusions, and generate subsequent inspection control instructions or knowledge base update instructions for the track robot based on the fault diagnosis conclusions.

2. The fault identification method for the power distribution room track inspection robot according to claim 1, characterized in that, In S1: When the operating status observation data is image data involving the color of the device indicator light, image information of potential light sources around the indicator light observation area that may affect its color interpretation is collected; When the temperature data of the target power equipment is detected to reach the preset attention threshold, the status image and temperature data of the neighboring equipment that are spatially adjacent to the target power equipment and may have thermal interaction are collected. When the confidence level of the result obtained after verifying or correcting the preliminary observation features based on the context information is lower than the preset confidence threshold, the control robot changes at least one of the following: acquisition position, gimbal angle, sensor acquisition mode, or acquisition parameters, to supplement the acquisition of the target power equipment or the corresponding environmental interference source.

3. The fault identification method for the power distribution room track inspection robot according to claim 2, characterized in that, S3 includes: Based on the color characteristics of the indicator light, the light interference contribution of ambient light to the color observation value of the indicator light is determined according to the color attribute, light intensity attribute and spatial relationship between the potential light source and the indicator light. The color observation value of the indicator light is then compensated according to the light interference contribution to obtain the target color characteristics of the indicator light. Based on the temperature characteristics of the equipment, the equipment type, working status and temperature information of the neighboring equipment are identified, and the thermal interference contribution of the neighboring equipment to the measured temperature of the target power equipment is determined by combining the spatial topological relationship between the neighboring equipment and the target power equipment. Based on the thermal interference contribution, the target temperature characteristics characterizing the heating state of the target power equipment are obtained.

4. The fault identification method for the power distribution room track inspection robot according to claim 1, characterized in that, The data set is a structured data frame, whose data field includes at least a subset of device-body multimodal sensing data, a subset of environmental images related to optical interference, a subset of neighboring device data related to thermal interference, and spatial topology metadata. The spatial topology metadata includes at least one of the following: device identifier, monitoring point identifier, acquisition time identifier, and relative position, distance, and relative orientation between the target power equipment, potential light source, and neighboring equipment.

5. The fault identification method for the power distribution room track inspection robot according to claim 4, characterized in that, The nodes of the dynamic association graph model include at least: The observation feature nodes corresponding to the preliminary observation features directly extracted from the operational status observation data; Target feature nodes corresponding to the target features obtained after verification or correction of environmental interference; and The interference source nodes corresponding to the environmental interference sources have been identified; The edges of the dynamic association graph model include at least the state association edges between target feature nodes and the interference interpretation edges pointing from the interference source node to the corresponding observed feature node. The weights of the interference interpretation edges are determined or adjusted based on the interference contribution of the corresponding environmental interference sources to the observed features.

6. The fault identification method for the power distribution room track inspection robot according to claim 3, characterized in that, Compensation for indicator light color observations includes: Determine the color components of ambient light based on the potential light source image; The ambient light intensity coefficient is determined based on at least one of the following: the brightness of the potential light source, the imaging area, and the distance and relative angle between the potential light source and the indicator light. The color contribution of ambient light to the color observation value of the indicator light is determined based on the color components and ambient light intensity coefficient. The color observation value of the indicator light is then adjusted or mapped in the RGB color space or HSV color space to obtain the compensated target color features of the indicator light.

7. The fault identification method for the power distribution room track inspection robot according to claim 3, characterized in that, Determining the contribution of the thermal interference includes: The heat contribution of each neighboring device is determined by using a pre-set empirical coefficient model or a simplified heat transfer model, based on the distance between each neighboring device and the target power equipment, their relative orientation, the type of neighboring device, and the temperature difference between the neighboring device and the target power equipment. The external heat contribution is obtained by combining the heat contribution of each neighboring device, and the measured temperature of the target power equipment is corrected by combining the ambient background temperature and the external heat contribution to obtain the self-temperature rise characteristics that characterize the self-heating state of the target power equipment.

8. The fault identification method for the power distribution room track inspection robot according to claim 1, characterized in that, The knowledge base update includes: The false judgment events that are verified to be caused by environmental interference sources are recorded as environmental interference pattern samples, and the environmental interference pattern samples are associated with and stored with the corresponding monitoring point identifiers. The environmental interference pattern sample includes at least the type of interference source, the conditions under which the interference occurs, and the information on the interference impact of the interference source on the corresponding observation features; Based on multiple environmental interference pattern samples associated with the same monitoring point, update at least one of the following: context environment information acquisition triggering conditions, interference contribution calculation parameters, and feature determination thresholds corresponding to the monitoring point.

9. A system applicable to the fault identification method of the power distribution room track inspection robot according to any one of claims 1-8, characterized in that, include: The sensing module is used to collect observation data on the operating status of the target power equipment and contextual information; The context scheduling module is used to trigger the collection task of corresponding context environment information based on the data type of the running status observation data or the preliminary anomaly judgment result. The fusion processing module is used to perform temporal correlation and spatial relationship matching between operational status observation data and contextual environment information, and to construct a data set containing the device's physical status, environmental status, and spatial topology relationships. A knowledge base is used to store device models, fault characteristics, environmental interference patterns, spatial topology information, and association rules. An anti-interference verification module is used to identify environmental interference sources corresponding to preliminary observation features, determine the interference contribution of environmental interference sources to preliminary observation features, and verify or correct preliminary observation features based on the interference contribution; the anti-interference verification module includes an optical interference processing unit and a thermal correlation analysis unit; The feature refinement module is used to output verified or corrected target features; The model building module is used to construct a dynamic correlation graph model based on observation features, target features, and environmental interference sources. The diagnostic module is used to analyze the dynamic correlation graph model and generate fault diagnosis conclusions; The strategy module is used to generate subsequent inspection control instructions or knowledge base update instructions based on the fault diagnosis conclusions. The robot motion control module is used to execute the subsequent inspection control commands to control the movement position of the track robot, the gimbal attitude or the acquisition status of the sensor module, and to feed back the supplementary acquired data to the fusion processing module.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, the processor performs the fault identification method of the power distribution room track inspection robot according to any one of claims 1-8.

Images