An edge diagnosis method and system for a power distribution equipment mobile inspection terminal

Abstract

The application discloses an edge diagnosis method and system of a power distribution equipment mobile inspection terminal, and relates to the technical fields of intelligent inspection of power distribution equipment and edge computing. The method comprises the following steps: collecting multi-source sensing data of a target power distribution equipment and performing time synchronization; performing state prediction and fusion on each mode data based on an edge diagnosis network, dynamically adjusting a diagnosis threshold according to cross-modal consistency and operating condition information; selecting an edge output, cloud review or alarm mode in combination with diagnosis uncertainty, and using review results to screen samples to update the diagnosis network. The method solves the problems of insufficient multi-source data fusion, low edge diagnosis accuracy and adaptability in the existing power distribution equipment inspection, and improves the intelligent diagnosis capability of the mobile inspection terminal.

Description

Technical Field

[0001] This application relates to the field of intelligent inspection and edge computing technology for power distribution equipment, and in particular to an edge diagnostic method and system for a mobile inspection terminal for power distribution equipment. Background Technology

[0002] As key equipment in the power distribution network, the real-time monitoring and accurate diagnosis of the operating status of power distribution equipment (such as distribution transformers, switchgear, ring main units, etc.) directly affects the reliability of power supply. With the continuous expansion of the scale of the power distribution network and the rapid growth in the number of equipment, the traditional method of relying on regular manual inspections is no longer sufficient to meet the needs of operation and maintenance.

[0003] In recent years, with the rapid development of sensing technology, edge computing, and artificial intelligence, intelligent inspection terminals have begun to be used in the condition monitoring of power distribution equipment. These terminals typically integrate multiple sensors (such as visible light cameras, infrared thermal imagers, acoustic fingerprint sensors, and partial discharge sensors), and complete data acquisition and preliminary analysis locally through edge computing modules. For example, some existing solutions deploy visual recognition and signal processing algorithms in the edge computing module to achieve localized analysis of equipment status, coupling power business logic with computer vision algorithms at the software level to achieve proactive early warning of equipment anomalies. Other solutions integrate multi-source data such as vibration, temperature, and images collected by various types of intelligent devices in the sensing layer, and deploy signal processing and image recognition algorithms in the edge computing layer to achieve localized real-time data analysis.

[0004] However, existing technologies still have the following shortcomings:

[0005] First, the time synchronization and fusion accuracy of multi-source heterogeneous data is insufficient. Visible light images, infrared thermal images, acoustic signature signals, and partial discharge signals originate from different physical sensing mechanisms, and their acquisition time, sampling frequency, and propagation path have inherent differences. Acoustic signature signals experience a propagation delay in the air; without propagation delay compensation, different modal data cannot be truly aligned in the time dimension, affecting the accuracy of subsequent fusion diagnosis. Although existing technologies propose spatiotemporal registration for multi-source heterogeneous data, they still lack effective compensation mechanisms for the time offset caused by the propagation path of acoustic signature signals.

[0006] Secondly, the consistency of multimodal diagnostic results is not effectively utilized. Existing technologies, after obtaining preliminary diagnostic results from each modality, typically perform decision-level fusion or feature-level fusion directly, but fail to fully leverage the consistency information between the diagnostic results of different modalities. When diagnostic conclusions from different modalities contradict each other, existing technologies lack an effective mechanism to identify such cross-modal conflicts and use them as a basis for diagnostic uncertainty or adaptive threshold adjustment.

[0007] Third, diagnostic models cannot be continuously optimized based on actual operating conditions. Existing inspection terminal diagnostic models are typically trained offline before deployment, and their parameters are fixed after deployment. While localized real-time analysis at the edge has become a trend, models with fixed parameters struggle to maintain diagnostic accuracy in the face of dynamic factors such as equipment aging, changing operating conditions, and environmental interference. Although some solutions propose cloud-based collaborative model update strategies, they lack effective sample selection mechanisms and cannot utilize high-value on-site diagnostic data for incremental learning, resulting in low model iteration efficiency.

[0008] Furthermore, edge computing devices have limited computing resources, power, and storage space. Existing technologies still struggle to dynamically adjust computing strategies based on inspection tasks and resource status, while ensuring diagnostic accuracy and rationally allocating computing resources. This also makes it difficult to make optimal decisions between direct output at the edge and uploading to the cloud for review. Summary of the Invention

[0009] The purpose of this application is to provide an edge diagnostic method and system for mobile inspection terminals of power distribution equipment. This addresses the technical problems in existing technologies, such as the difficulty in effectively synchronizing and fusing multi-source heterogeneous sensing data during mobile inspection of power distribution equipment, the difficulty in accurately assessing the reliability of multimodal diagnostic results, the inability to dynamically adjust diagnostic thresholds based on equipment operating status, and the lack of continuous optimization capabilities in edge diagnostic models, leading to insufficient accuracy and adaptability in equipment status diagnosis.

[0010] In view of the above technical problems, this application provides an edge diagnostic method and system for mobile inspection terminals of power distribution equipment.

[0011] A first aspect of this application provides an edge diagnostic method for a mobile inspection terminal for power distribution equipment, the method comprising:

[0012] The multi-source sensing module collects multi-source sensing data of the target power distribution equipment, and the communication module obtains the operating condition information of the target power distribution equipment. The multi-source sensing data includes at least visible light images, infrared thermal images, acoustic fingerprint signals, and partial discharge signals, which correspond to visual mode, thermal imaging mode, acoustic mode, and discharge characteristic mode, respectively. The operating condition information includes at least one of load current, ambient temperature, and ambient humidity.

[0013] Based on the time reference provided by the positioning and timing module, the multi-source sensing data is time-synchronized, and the acoustic propagation delay of the voiceprint signal is compensated to obtain the time-synchronized multi-source sensing data.

[0014] The edge computing module runs an edge diagnostic network to perform inference on each modal data after time synchronization to obtain the diagnostic probability distribution of each modality; it determines the fusion weight based on the data quality index of each modality data, and performs weighted fusion on the diagnostic probability distribution of each modality based on the fusion weight to obtain the fused diagnostic probability distribution.

[0015] Based on the pairwise consistency distance between the various modal diagnostic probability distributions, a cross-modal consistency metric C is determined; multiple random inferences are performed on the fused diagnostic probability distribution using a random inactivation inference method, and an uncertainty metric U is determined based on the degree of dispersion between the state prediction probability distributions obtained from the multiple inferences.

[0016] Based on the operating condition information and the cross-modal consistency metric C, the baseline judgment threshold range is corrected to obtain the corrected judgment threshold range. The cross-modal consistency metric C is used as a damping function control parameter to adjust the adjustment range of the upper and lower limits of the baseline judgment threshold range. Based on the corrected judgment threshold range, the fusion diagnostic probability distribution is classified into state levels to obtain the edge-side state classification result.

[0017] Based on the uncertainty metric U and the cross-modal consistency metric C, select one of the following diagnostic output modes: direct edge output, cloud collaborative review output, or alarm flag;

[0018] When the cloud-based collaborative review mode is selected, the fusion diagnostic probability distribution is sent to the cloud for review, the cloud-side state classification result is obtained, and the cloud-side state classification result is weighted and fused with the edge-side state classification result to output the cloud review fusion result.

[0019] Diagnostic samples whose cross-modal consistency metric C is lower than a preset consistency screening threshold and whose cloud-side state classification results are inconsistent with the edge-side state classification results are included in the local incremental learning sample pool, and the edge diagnostic network parameters are updated based on the local incremental learning sample pool; wherein, the cross-modal consistency metric C also serves as a damping control parameter for the correction of the benchmark judgment threshold interval, a decision control parameter for the selection of the diagnostic output mode, and a control parameter for the screening of incremental learning samples. A multi-source sensing and operating condition acquisition unit is used to collect multi-source sensing data of the target power distribution equipment through the multi-source sensing module, and to acquire the operating condition information of the target power distribution equipment through the communication module. The multi-source sensing data includes at least visible light images, infrared thermal images, acoustic signals, and partial discharge signals, corresponding to visual modality, thermal imaging modality, acoustic modality, and discharge characteristic modality, respectively. The operating condition information includes at least one of load current, ambient temperature, and ambient humidity.

[0020] A time synchronization unit is used to perform time synchronization processing on the multi-source sensing data based on the time reference provided by the positioning and timing module, and to perform acoustic propagation delay compensation on the voiceprint signal to obtain time-synchronized multi-source sensing data.

[0021] A multimodal fusion unit is used by the edge computing module to run the edge diagnostic network, perform inference on each modal data after time synchronization, and obtain the diagnostic probability distribution of each modality; determine the fusion weight according to the data quality index of each modal data, and perform weighted fusion on the diagnostic probability distribution of each modality based on the fusion weight to obtain the fused diagnostic probability distribution.

[0022] A cross-modal consistency and uncertainty calculation unit is used to determine the cross-modal consistency metric C based on the pairwise consistency distance between the probability distributions of each modality diagnosis; to perform multiple random inferences on the fused diagnosis probability distribution using a random inactivation inference method, and to determine the uncertainty metric U based on the degree of dispersion between the state prediction probability distributions obtained from the multiple inferences.

[0023] A threshold correction and classification unit is used to correct the baseline judgment threshold interval based on the operating condition information and the cross-modal consistency metric C to obtain a corrected judgment threshold interval. The cross-modal consistency metric C is used as a damping function control parameter to adjust the adjustment range of the upper and lower limits of the baseline judgment threshold interval. The unit also performs state level classification on the fusion diagnostic probability distribution based on the corrected judgment threshold interval to obtain the edge-side state classification result.

[0024] An output mode selection unit is used to select one of the following diagnostic output modes: edge-side direct output, cloud-based collaborative verification output, or alarm flag, based on the uncertainty metric U and the cross-modal consistency metric C.

[0025] The cloud-based verification unit is used to send the fusion diagnostic probability distribution to the cloud for verification when the cloud collaborative verification mode is selected, obtain the cloud-side state classification result, and perform weighted fusion of the cloud-side state classification result and the edge-side state classification result to output the cloud-based verification fusion result.

[0026] An incremental learning update unit is used to include diagnostic samples whose cross-modal consistency metric C is lower than a preset consistency screening threshold and whose cloud-side state classification results are inconsistent with the edge-side state classification results into a local incremental learning sample pool, and to update the edge diagnostic network parameters based on the local incremental learning sample pool; wherein, the cross-modal consistency metric C also serves as a damping control parameter for the correction of the benchmark judgment threshold range, a decision control parameter for the selection of the diagnostic output mode, and a control parameter for the selection of incremental learning samples.

[0027] A second aspect of this application provides an edge diagnostic system for a mobile inspection terminal of power distribution equipment, the system comprising:

[0028] A multi-source sensing and operating condition acquisition unit is used to collect multi-source sensing data of the target power distribution equipment through the multi-source sensing module, and to acquire the operating condition information of the target power distribution equipment through the communication module. The multi-source sensing data includes at least visible light images, infrared thermal images, acoustic fingerprint signals, and partial discharge signals, which correspond to visual mode, thermal imaging mode, acoustic mode, and discharge characteristic mode, respectively. The operating condition information includes at least one of load current, ambient temperature, and ambient humidity.

[0029] A time synchronization unit is used to perform time synchronization processing on the multi-source sensing data based on the time reference provided by the positioning and timing module, and to perform acoustic propagation delay compensation on the voiceprint signal to obtain time-synchronized multi-source sensing data.

[0030] A multimodal fusion unit is used by the edge computing module to run the edge diagnostic network, perform inference on each modal data after time synchronization, and obtain the diagnostic probability distribution of each modality; determine the fusion weight according to the data quality index of each modal data, and perform weighted fusion on the diagnostic probability distribution of each modality based on the fusion weight to obtain the fused diagnostic probability distribution.

[0031] A cross-modal consistency and uncertainty calculation unit is used to determine the cross-modal consistency metric C based on the pairwise consistency distance between the probability distributions of each modality diagnosis; to perform multiple random inferences on the fused diagnosis probability distribution using a random inactivation inference method, and to determine the uncertainty metric U based on the degree of dispersion between the state prediction probability distributions obtained from the multiple inferences.

[0032] A threshold correction and classification unit is used to correct the baseline judgment threshold interval based on the operating condition information and the cross-modal consistency metric C to obtain a corrected judgment threshold interval. The cross-modal consistency metric C is used as a damping function control parameter to adjust the adjustment range of the upper and lower limits of the baseline judgment threshold interval. The unit also performs state level classification on the fusion diagnostic probability distribution based on the corrected judgment threshold interval to obtain the edge-side state classification result.

[0033] An output mode selection unit is used to select one of the following diagnostic output modes: edge-side direct output, cloud-based collaborative verification output, or alarm flag, based on the uncertainty metric U and the cross-modal consistency metric C.

[0034] The cloud-based verification unit is used to send the fusion diagnostic probability distribution to the cloud for verification when the cloud collaborative verification mode is selected, obtain the cloud-side state classification result, and perform weighted fusion of the cloud-side state classification result and the edge-side state classification result to output the cloud-based verification fusion result.

[0035] An incremental learning update unit is used to include diagnostic samples whose cross-modal consistency metric C is lower than a preset consistency screening threshold and whose cloud-side state classification results are inconsistent with the edge-side state classification results into a local incremental learning sample pool, and to update the edge diagnostic network parameters based on the local incremental learning sample pool; wherein, the cross-modal consistency metric C also serves as a damping control parameter for the correction of the benchmark judgment threshold range, a decision control parameter for the selection of the diagnostic output mode, and a control parameter for the selection of incremental learning samples.

[0036] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0037] The multi-source sensing module collects multi-source sensing data of the target power distribution equipment, and the communication module obtains the operating condition information of the target power distribution equipment. The multi-source sensing data includes at least visible light images, infrared thermal images, acoustic fingerprint signals, and partial discharge signals, corresponding to visual mode, thermal imaging mode, acoustic mode, and discharge characteristic mode, respectively. The operating condition information includes at least one of load current, ambient temperature, and ambient humidity. Based on the time reference provided by the positioning and timing module, the multi-source sensing data undergoes time synchronization processing, and acoustic propagation delay compensation is performed on the acoustic fingerprint signals to obtain… The system obtains time-synchronized multi-source sensing data; the edge computing module runs an edge diagnostic network to infer each modal data after time synchronization, obtaining the diagnostic probability distribution of each modality; fusion weights are determined based on the data quality indicators of each modality data, and the diagnostic probability distributions of each modality are weighted and fused based on the fusion weights to obtain a fused diagnostic probability distribution; a cross-modal consistency metric C is determined based on the pairwise consistency distance between the diagnostic probability distributions of each modality; the fused diagnostic probability distribution is subjected to multiple random inferences using a random inactivation inference method, and the state prediction probability distribution obtained from the multiple inferences is used as the basis for the calculation. The degree of dispersion between the parameters determines the uncertainty metric U; based on the operating condition information and the cross-modal consistency metric C, the baseline judgment threshold interval is corrected to obtain the corrected judgment threshold interval, wherein the cross-modal consistency metric C serves as a damping function control parameter to adjust the adjustment range of the upper and lower limits of the baseline judgment threshold interval; the fusion diagnostic probability distribution is classified into state levels according to the corrected judgment threshold interval to obtain the edge-side state classification result; based on the uncertainty metric U and the cross-modal consistency metric C, a diagnostic output mode is selected from edge-side direct output, cloud collaborative review output, or alarm marking; when the cloud collaborative review mode is selected, the fusion diagnostic probability distribution is sent to the cloud for review to obtain the cloud-side state classification result, and the cloud-side state classification result is weighted and fused with the edge-side state classification result to output the cloud review fusion result; diagnostic samples whose cross-modal consistency metric C is lower than the preset consistency screening threshold and whose cloud-side state classification result is inconsistent with the edge-side state classification result are included in the local incremental learning sample pool, and the edge diagnostic network parameters are updated based on the local incremental learning sample pool;The cross-modal consistency metric C serves as a damping control parameter for the benchmark judgment threshold interval correction, a decision control parameter for the diagnostic output mode selection, and a control parameter for the incremental learning sample screening. A multi-source sensing and operating condition acquisition unit is used to collect multi-source sensing data of the target power distribution equipment through the multi-source sensing module and to acquire the operating condition information of the target power distribution equipment through the communication module. The multi-source sensing data includes at least visible light images, infrared thermal images, acoustic fingerprint signals, and partial discharge signals, corresponding to visual modalities, thermal imaging modalities, acoustic modalities, and discharge characteristic modalities, respectively. The operating condition information includes at least one of load current, ambient temperature, and ambient humidity. A time synchronization unit is used to synchronize the multi-source sensing data based on the time reference provided by the positioning and timing module. The system performs time synchronization processing and acoustic propagation delay compensation on the voiceprint signal to obtain time-synchronized multi-source sensing data; a multimodal fusion unit is used by the edge computing module to run the edge diagnostic network, and to perform inference on each modal data after time synchronization to obtain the diagnostic probability distribution of each modality; the fusion weight is determined according to the data quality index of each modality data, and the diagnostic probability distribution of each modality is weighted and fused based on the fusion weight to obtain the fused diagnostic probability distribution; a cross-modal consistency and uncertainty calculation unit is used to calculate the pairwise consistency and uncertainty between the diagnostic probability distributions of each modality. Consistency distance is used to determine the cross-modal consistency metric C; multiple random inferences are performed on the fusion diagnostic probability distribution using a random inactivation inference method, and the uncertainty metric U is determined based on the dispersion between the state prediction probability distributions obtained from the multiple inferences; a threshold correction and classification unit is used to correct the baseline judgment threshold interval based on the operating condition information and the cross-modal consistency metric C, obtaining a corrected judgment threshold interval, wherein the cross-modal consistency metric C serves as a damping function control parameter to adjust the adjustment range of the upper and lower limits of the baseline judgment threshold interval; based on the corrected judgment threshold... The interval classifies the fusion diagnostic probability distribution into state levels to obtain edge-side state classification results; the output mode selection unit is used to select one of the following diagnostic output modes according to the uncertainty metric U and the cross-modal consistency metric C: edge-side direct output, cloud collaborative review output, or alarm flag; the cloud review unit is used to send the fusion diagnostic probability distribution to the cloud for review when the cloud collaborative review mode is selected, obtain the cloud-side state classification results, and perform weighted fusion of the cloud-side state classification results and the edge-side state classification results to output the cloud review fusion result;An incremental learning update unit is used to include diagnostic samples whose cross-modal consistency metric C is lower than a preset consistency screening threshold and whose cloud-side state classification results are inconsistent with the edge-side state classification results into a local incremental learning sample pool, and to update the edge diagnostic network parameters based on the local incremental learning sample pool. The cross-modal consistency metric C also serves as a damping control parameter for the correction of the baseline judgment threshold range, a decision control parameter for the selection of the diagnostic output mode, and a control parameter for the selection of incremental learning samples. This solves the technical problems of insufficient accuracy and adaptability in equipment status diagnosis caused by the difficulty in effectively synchronizing and fusing multi-source heterogeneous sensing data during mobile inspection of existing power distribution equipment, the difficulty in accurately assessing the reliability of multi-modal diagnostic results, the inability to dynamically adjust diagnostic thresholds according to equipment operating status, and the lack of continuous optimization capability of the edge diagnostic model.

[0038] The above description is merely an overview of the technical solution of this application. In order to more clearly explain the technical means of this application, and to enable its implementation in accordance with the contents of the specification, and to make the above and other objectives, features and advantages of this application more apparent and understandable, specific embodiments of this application are described below. Attached Figure Description

[0039] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings of the embodiments of this disclosure will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0040] Figure 1 A flowchart illustrating the edge diagnostic method for a mobile inspection terminal for power distribution equipment provided in this application embodiment;

[0041] Figure 2 This is a schematic diagram of the edge diagnostic system of the mobile inspection terminal for power distribution equipment provided in an embodiment of this application.

[0042] Figure labeling: Multi-source sensing and operating condition acquisition unit 10, time synchronization unit 20, multi-modal fusion unit 30, cross-modal consistency and uncertainty calculation unit 40, threshold correction and classification unit 50, output mode selection unit 60, cloud verification unit 70, incremental learning and update unit 80. Detailed Implementation

[0043] This application provides an edge diagnostic method and system for mobile inspection terminals of power distribution equipment, which solves the technical problems of insufficient accuracy and adaptability of equipment status diagnosis caused by the difficulty in effectively synchronizing and integrating multi-source heterogeneous sensing data, the difficulty in accurately assessing the reliability of multimodal diagnostic results, the inability to dynamically adjust diagnostic thresholds according to the equipment operating status, and the lack of continuous optimization capability of edge diagnostic models during the existing mobile inspection of power distribution equipment.

[0044] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0045] Example 1, Figure 1 As shown, this application provides an edge diagnostic method for a mobile inspection terminal for power distribution equipment, wherein the method includes:

[0046] The multi-source sensing module collects multi-source sensing data of the target power distribution equipment, and the communication module obtains the operating condition information of the target power distribution equipment. The multi-source sensing data includes at least visible light images, infrared thermal images, acoustic fingerprint signals, and partial discharge signals, which correspond to visual mode, thermal imaging mode, acoustic mode, and discharge characteristic mode, respectively. The operating condition information includes at least one of load current, ambient temperature, and ambient humidity.

[0047] Specifically, the mobile inspection terminal for power distribution equipment uses the multi-source sensing module to collect on-site data from the target power distribution equipment, obtaining multi-source sensing data reflecting the target power distribution equipment's appearance, temperature, operating sound characteristics, and electrical discharge characteristics. The multi-source sensing module includes a visible light camera unit, an infrared thermal imaging unit, an acoustic signature acquisition unit, and a partial discharge detection unit. Different types of sensing units acquire equipment status information under different physical characteristics, and the collected data is categorized according to their corresponding data types: visual mode, thermal imaging mode, acoustic mode, and discharge characteristic mode. For example, during the inspection of a distribution transformer, the visible light camera unit acquires visible light images of the transformer's casing, bushings, terminals, and surrounding environment to form visual modal data, which reflects the appearance of the equipment, such as surface damage, dirt, rust, and foreign matter adhesion. The infrared thermal imaging acquisition unit acquires infrared thermal image data during the operation of the distribution transformer to form thermal imaging modal data, which reflects temperature distribution states such as overheating of terminals, abnormal local temperature rise, and abnormal heat dissipation. The acoustic signature acquisition unit acquires sound signals generated during the operation of the distribution transformer to form acoustic modal data, which reflects the operating sound characteristics of the equipment, such as internal mechanical vibration, abnormal discharge, and electromagnetic noise. The partial discharge detection unit acquires partial discharge signals during the operation of the distribution equipment to form discharge characteristic modal data, which reflects electrical characteristics such as insulation defects and abnormal partial discharge. The multi-source sensing data collected by the multi-source sensing module is transmitted to the edge computing module for processing via wired or wireless communication. For example, it can be connected to the edge computing module through a serial communication interface, Ethernet interface, wireless communication interface, or industrial fieldbus interface, enabling the edge computing module to acquire the multi-source sensing data in real time. Simultaneously, the mobile inspection terminal for power distribution equipment obtains the operating condition information of the target power distribution equipment through the communication module. This operating condition information describes the current operating environment and load status of the target power distribution equipment, including at least one of load current, ambient temperature, and ambient humidity. Specifically, the load current is obtained through a current detection device or power distribution system operation data interface installed on the power distribution equipment side, reflecting the current load level of the target power distribution equipment; the ambient temperature is obtained through a temperature sensor or environmental monitoring equipment, reflecting changes in the operating environment temperature; and the ambient humidity is obtained through a humidity sensor or environmental monitoring equipment, reflecting changes in the ambient humidity of the environment where the equipment is located. When some operating condition information cannot be obtained, such as when the target power distribution equipment is not equipped with an ambient humidity sensor, the operating condition information may only include load current and ambient temperature; when load current cannot be obtained, ambient temperature and ambient humidity may also be used in the subsequent diagnostic process.

[0048] Based on the time reference provided by the positioning and timing module, the multi-source sensing data is time-synchronized, and the acoustic propagation delay of the voiceprint signal is compensated to obtain the time-synchronized multi-source sensing data.

[0049] Furthermore, the acoustic propagation delay compensation includes:

[0050] The edge computing module acquires the distance d between the acoustic signature acquisition unit in the multi-source sensing module and the target power distribution equipment, as well as the current ambient sound speed v, and calculates the acoustic propagation delay. ;

[0051] Using the time reference provided by the positioning and timing module as a reference time, the acquisition time of the voiceprint signal is subtracted from the time referenced by the time referenced by the positioning and timing module. This aligns the acquisition time of the compensated acoustic signature signal with the acquisition times of the visible light image, infrared thermal image, and partial discharge signal to the same time reference.

[0052] Specifically, the positioning and timing module provides a unified time reference for the multi-source sensing data to eliminate time offsets caused by differences in acquisition clocks between different sensing units, enabling the visible light image, infrared thermal image, acoustic signature signal, and partial discharge signal to be analyzed correspondingly under the same time reference. The positioning and timing module may include at least one of a satellite timing module, a network timing module, and a high-precision clock module, used to provide unified timestamp information to the edge computing module. In one embodiment, after receiving the multi-source sensing data acquired by the multi-source sensing module, the edge computing module first corrects the acquisition timestamps corresponding to each modality of data based on the time reference provided by the positioning and timing module, and establishes a time correspondence based on the acquisition time information of each modality of data. For example, if the acquisition time of the visible light image is t1, the acquisition time of the infrared thermal image is t2, the acquisition time of the acoustic signature signal is t3, and the acquisition time of the partial discharge signal is t4, the edge computing module converts the above acquisition times to the same time reference system through a unified time reference, enabling different modal data to correspond to the same device operating state. Since the acoustic signature signal is acquired by the acoustic signature acquisition unit after propagating through the air medium, there is a propagation delay between the actual acquisition time and the time when the target power distribution equipment generates the sound signal. Therefore, this embodiment further corrects the acoustic signature signal in time through acoustic propagation delay compensation. Specifically, the edge computing module obtains the distance d between the acoustic signature acquisition unit in the multi-source sensing module and the target power distribution equipment, as well as the current ambient sound speed v, and calculates the acoustic propagation delay based on the relationship between sound wave propagation distance and propagation speed. The acoustic propagation delay satisfy: Where d represents the spatial distance between the voiceprint acquisition unit and the target power distribution equipment, the unit can be meters; v represents the speed of sound in the current environment, the unit can be meters per second. The current environmental speed of sound v can be calculated based on the ambient temperature, or it can be directly obtained through a sound speed detection device. For example, in a normal temperature environment, the speed of sound in air is approximately 343 m / s. When the voiceprint acquisition unit is 3.43 m away from the target power distribution equipment, the corresponding acoustic propagation delay is approximately 10 ms. Further, the edge computing module uses the time reference provided by the positioning and timing module as a reference time to compensate for the acquisition time of the voiceprint signal. Specifically, the acquisition time of the voiceprint signal is subtracted from the acoustic propagation delay. This advances the acquisition time of the compensated acoustic signature signal to the corresponding sound source generation time, thereby aligning the acquisition time of the compensated acoustic signature signal with the acquisition time of the visible light image, infrared thermal image, and partial discharge signal to the same time reference. In another embodiment, when the distance between the acoustic signature acquisition unit and the target power distribution equipment changes, for example, when the mobile inspection terminal of the power distribution equipment gradually approaches or moves away from the target power distribution equipment during the inspection process, the edge computing module can dynamically calculate the acoustic propagation delay based on the real-time acquired distance d between the acoustic signature acquisition unit and the target power distribution equipment. The system updates the time compensation of the voiceprint signal in real time to ensure that the voiceprint signals collected at different locations during the inspection process are consistent with the visible light image, infrared thermal image and partial discharge signal in time.

[0053] The edge computing module runs an edge diagnostic network to perform inference on each modal data after time synchronization to obtain the diagnostic probability distribution of each modality; it determines the fusion weight based on the data quality index of each modality data, and performs weighted fusion on the diagnostic probability distribution of each modality based on the fusion weight to obtain the fused diagnostic probability distribution.

[0054] Furthermore, the data quality indicators include:

[0055] The edge gradient degradation index of a visible light image is determined by extracting the gradient magnitude distribution of the visible light image and then statistically analyzing the proportion of pixels that are below a preset gradient threshold.

[0056] The temperature gradient entropy of the infrared thermal image is determined by statistically analyzing the gradient entropy after extracting the temperature field gradient of the infrared thermal image.

[0057] The signal-to-noise ratio of the voiceprint signal is determined by the ratio of the effective power of the voiceprint signal to the background noise power.

[0058] The concentration of the pulse phase distribution of the partial discharge signal is determined by the reciprocal of the variance of the phase distribution of the partial discharge pulse within the power frequency cycle.

[0059] When any modality data is missing, the data quality index score for that modality is recorded as zero, and the data quality scores of the remaining modalities are renormalized to determine the fusion weight of each modality data.

[0060] Furthermore, the edge computing module determines the remaining inference count based on the remaining inspection path length between the current inspection position and the preset inspection endpoint, combined with the number of diagnostic samples corresponding to the unit path length in the historical inspection process.

[0061] Obtain the resource consumption parameters corresponding to a single diagnostic inference when the high-precision diagnostic inference path is executed, and determine the amount of high-precision diagnostic resources required to complete the remaining inspection stage based on the remaining inference count and the single diagnostic inference resource consumption parameters.

[0062] When the demand for high-precision diagnostic resources exceeds the current available computing resources of the edge computing module and cannot support the high-precision diagnostic inference path to complete the remaining inference times, the current inference path will be switched from the high-precision diagnostic inference path to the lightweight diagnostic inference path.

[0063] During the K diagnostic cycles of the inference path switching transition phase, the diagnostic probability distributions before and after the switch are weighted and fused according to linearly gradual weights. The high-precision diagnostic inference path fusion weights corresponding to the i-th diagnostic cycle are: , where i=1,2,...,K, and K is the preset number of transitional diagnostic cycles.

[0064] Specifically, the edge computing module predicts the number of diagnostic inferences required for the remaining inspection phase based on the remaining inspection path length between the current inspection position and the preset inspection endpoint, combined with the number of diagnostic samples per unit path length during historical inspections. Further, the edge computing module obtains resource consumption parameters corresponding to a single diagnostic inference execution on the high-precision diagnostic inference path. These resource consumption parameters include at least one of the following: the number of computation cycles required for a single diagnostic inference, storage usage, and energy consumption. Based on the product of the predicted number of diagnostic inferences and the single diagnostic inference resource consumption parameters, the high-precision diagnostic resource requirement for completing the remaining inspection phase is determined. In one embodiment, the edge diagnostic network performs feature extraction and state prediction on the data corresponding to the visual modality, thermal imaging modality, acoustic modality, and discharge characteristic modality to obtain the diagnostic probability distribution for each modality. Specifically, the visual modality is used to identify surface defects or structural anomalies in equipment; the thermal imaging modality is used to identify local overheating or abnormal temperature rise; the acoustic modality is used to identify abnormal vibrations or abnormal discharge sound patterns; and the discharge characteristic modality is used to identify insulation anomalies or enhanced partial discharge states. Specifically, the data quality indicators are used to evaluate the effectiveness of the corresponding modal data, and can be determined based on the physical characteristics of different modal data. For the visible light image corresponding to the visual modality, the data quality indicators include the edge gradient degradation index. The edge gradient degradation index is determined by extracting the gradient amplitude distribution of the visible light image and statistically analyzing the proportion of pixels below a preset gradient threshold. For example, when the inspection environment has insufficient lighting, lens contamination, or image blurring, the edge gradient in the visible light image decreases, the proportion of low-gradient pixels increases, and the corresponding edge gradient degradation index decreases, indicating a decline in the quality of the visual modality data. For the infrared thermal image corresponding to the thermal imaging modality, the data quality indicators include the temperature gradient entropy. The temperature gradient entropy is obtained by extracting the temperature field gradient information in the infrared thermal image and calculating it based on the gradient distribution, and is used to reflect the effectiveness of temperature change areas in the infrared thermal image. For example, when there are local abnormal heating areas in the power distribution equipment, there will be obvious temperature gradient changes in the infrared thermal image, corresponding to a high temperature gradient entropy; when the infrared thermal image is affected by the environment, causing the temperature distribution to tend to be uniform, the corresponding temperature gradient entropy decreases. For the acoustic signature signal corresponding to the acoustic mode, the data quality metric includes the signal-to-noise ratio (SNR). The SNR is determined by the ratio of the effective power in the acoustic signature signal to the background noise power, and is used to evaluate the proportional relationship between the effective equipment operating sound and the environmental noise in the acoustic signature signal. For example, in a substation environment, when there is noise interference from large equipment operation, the background noise power increases, leading to a decrease in the SNR of the acoustic signature signal. In this case, the edge computing module reduces the fusion weight corresponding to the acoustic mode. For the partial discharge signal corresponding to the discharge characteristic mode, the data quality metric includes the pulse phase distribution concentration.The pulse phase distribution concentration is determined by calculating the reciprocal of the variance of the phase distribution of the partial discharge pulses within the power frequency cycle, and is used to evaluate the regularity of the partial discharge pulse distribution. For example, when the partial discharge pulses are concentrated in a specific phase interval, it indicates that the collected partial discharge signal has high effectiveness, and the corresponding pulse phase distribution concentration is high. After obtaining the data quality indicators for each modality, the edge computing module performs normalization processing based on each data quality indicator, and determines the fusion weight corresponding to each modality based on the normalized data quality score. Specifically, the data quality scores for each modality can be denoted as follows. The corresponding fusion weight can then be expressed as: .in, Let represent the fusion weight corresponding to the i-th modal data, and qi represent the data quality score of the i-th modal data. Further, when any modal data is missing, the edge computing module sets the data quality score of that modal data to zero and re-normalizes the data quality scores of the remaining modal data to redetermine the fusion weights for each modal data. For example, when a partial discharge signal cannot be collected at the inspection site due to a sensor malfunction, the data quality score corresponding to the discharge characteristic modality is set to zero, and the fusion weights are recalculated based on the data quality scores of the visual modality, thermal imaging modality, and acoustic modality, ensuring that the missing modality does not affect the participation of other valid modal data in diagnosis. Through the above method, the edge computing module dynamically adjusts the fusion weights based on the data quality of each modality and performs weighted fusion of the diagnostic probability distributions of each modality to obtain a fused diagnostic probability distribution. Furthermore, to adapt to the continuous inspection needs under the limited computing resources of the mobile inspection terminal, the edge diagnostic network includes a high-precision diagnostic inference path and a lightweight diagnostic inference path. Both the high-precision and lightweight diagnostic inference paths are used to extract features and predict the state of each modality in the multi-source sensing data, and output the corresponding device state diagnostic probability distribution. The lightweight diagnostic inference path reduces the number of network parameters or the amount of floating-point operations, so that the computing resources required for diagnostic inference are lower than those required by the high-precision diagnostic inference path. Specifically, the edge computing module predicts the number of diagnostic inferences required for the remaining inspection stage based on the remaining inspection path length between the current inspection position and the preset inspection endpoint, combined with the number of diagnostic samples per unit path length in historical inspection processes. Furthermore, the edge computing module acquires the resource consumption parameters corresponding to a single diagnostic inference execution of the high-precision diagnostic inference path. These resource consumption parameters include at least one of the following: the number of computation cycles required for a single diagnostic inference, storage usage, and energy consumption. Based on the product of the predicted diagnostic inference count and the resource consumption parameters for a single diagnostic inference, the module determines the high-precision diagnostic resource requirement needed to complete the remaining inspection phase. The edge computing module further acquires currently available computing resources, including at least one of remaining power, available storage space, and remaining computing power. When the high-precision diagnostic resource requirement exceeds the currently available computing resources and cannot support the high-precision diagnostic inference path in completing the diagnostic tasks of the remaining inspection phase, the currently executing inference path is switched from the high-precision diagnostic inference path to the lightweight diagnostic inference path to reduce computational resource consumption during the inspection process. Furthermore, to reduce diagnostic output fluctuations caused by changes in the inference path during the switching process, K diagnostic cycles are set as path switching transition cycles during the transition phase from the high-precision diagnostic inference path to the lightweight diagnostic inference path.During the i-th diagnostic cycle, the diagnostic probability distributions of the two inference paths before and after the switch are weighted and fused, where the fusion weight is the one corresponding to the high-precision diagnostic inference path. Determined in the following manner: Where i = 1, 2, ..., K; correspondingly, the fusion weight of the output results of the lightweight diagnostic inference path is... By progressively reducing the weight of the high-precision diagnostic inference path output and gradually increasing the weight of the lightweight diagnostic inference path output, the diagnostic results remain continuous and stable during the inference path switching process.

[0065] Based on the pairwise consistency distance between the various modal diagnostic probability distributions, a cross-modal consistency metric C is determined; multiple random inferences are performed on the fused diagnostic probability distribution using a random inactivation inference method, and an uncertainty metric U is determined based on the degree of dispersion between the state prediction probability distributions obtained from the multiple inferences.

[0066] Furthermore, the cross-modal consistency metric C is obtained in the following manner:

[0067] The modal diagnostic probability distributions corresponding to the visual modality, thermal imaging modality, acoustic modality, and discharge characteristic modality are compared pairwise to calculate the consistency distance between each modal diagnostic probability distribution. The cross-modal consistency metric C is obtained based on the weighted average of the pairwise modal consistency distances.

[0068] Specifically, after the edge computing module runs the edge diagnostic network to obtain the modal diagnostic probability distributions, in order to measure the consistency of the judgment results of the target power distribution equipment status among different modes, the edge computing module further determines the cross-modal consistency metric C based on the pairwise consistency distance between the modal diagnostic probability distributions. Here, the visual modality, thermal imaging modality, acoustic modality, and discharge characteristic modality correspond to the modal diagnostic probability distributions obtained from visible light images, infrared thermal images, acoustic signature signals, and partial discharge signals, respectively. For example, for a certain target power distribution equipment, the modal diagnostic probability distribution output by the visual modality can be expressed as... The modal diagnostic probability distribution of the thermal imaging modal output can be expressed as: Each element represents the predicted probability corresponding to a state level such as normal, attentive, abnormal, and severe abnormal; similarly, the acoustic modal output modal diagnostic probability distribution... The probability distribution of output mode diagnosis of the discharge characteristic mode. Each element represents a predicted probability corresponding to a state level such as normal, attention, abnormal, and severe abnormality. The edge computing module calculates... as well as The pairwise consistency distance is used to obtain the degree of consistency between different modal diagnostic probability distributions. Specifically, the consistency distance can be obtained using probability distribution distance metrics, including but not limited to KL divergence, JS divergence, cosine distance, Euclidean distance, or distance functions constructed based on differences in probability distributions. The consistency distance is used to characterize the degree of difference between the diagnostic probability distribution of the i-th modality and the diagnostic probability distribution of the j-th modality. The smaller the value, the closer the diagnostic probability distributions of the corresponding two modal outputs are. The edge computing module determines the cross-modal consistency metric C based on the weighted average of the pairwise modal consistency distances mentioned above. For example, in one embodiment, the cross-modal consistency metric C can be calculated as follows: The cross-modal consistency metric C is calculated based on the consistency distance, and the specific calculation formula is as follows: Where M represents the number of modes, and in this embodiment M=4, corresponding to the visual mode, thermal imaging mode, acoustic mode and discharge characteristic mode, respectively; This represents the consistency distance between the diagnostic probability distribution of the i-th modality and the diagnostic probability distribution of the j-th modality; This represents the consistency distance weight between the i-th mode and the j-th mode. In one implementation, the consistency distance... Perform normalization so that 0 ≤ The cross-modal consistency metric C satisfies 0 ≤ C ≤ 1. Further, in one embodiment, the consistency distance weight wij is determined based on the data quality index of the corresponding modality, so that the consistency distance between modalities with higher data quality has a greater impact on the cross-modal consistency metric C. Specifically, let the data quality score of the i-th modality be Qi, and the data quality score of the j-th modality be Qj, then the consistency distance weight wij between the corresponding two modalities can be determined as follows: Where Qi and Qj represent the data quality scores for the corresponding modalities, and the data quality scores are obtained by normalization based on the data quality indicators, and satisfy the following: The above methods enable modalities with higher data quality to contribute more to the cross-modal consistency metric C. Among them, each... The sum is 1. When the diagnostic probability distributions of each modality tend to be consistent, Dij decreases and the cross-modal consistency measure C increases; when there are significant differences, As the cross-modal consistency metric C increases, the overall consistency decreases. Furthermore, to avoid fluctuations in diagnostic results due to random factors during a single inference process, the edge computing module employs a random deactivation inference method, performing multiple random inferences on the same diagnostic input and determining the uncertainty metric U based on the dispersion between the state prediction probability distributions obtained from these multiple inferences. The edge computing module performs multiple random inferences on the diagnostic input corresponding to the multi-source sensing data synchronized at the same time, obtaining multiple state prediction probability distributions. For example, for a set of fused diagnostic input data for the same target power distribution equipment, the edge computing module performs N random deactivation inferences, obtaining state prediction probability distributions respectively: The edge computing module further calculates the degree of dispersion among the predicted probability distributions of the multiple states, such as calculating the variance, mean absolute deviation, or distribution entropy of the predicted probability for each state level, and determines the uncertainty metric U based on the degree of dispersion. The specific calculation formula is as follows: Where N represents the number of random deactivation inferences; Pn represents the probability distribution of the state prediction obtained from the nth random deactivation inference; Let represent the mean of the probability distribution of state predictions obtained from N random inferences, and its calculation formula is as follows: ;in, The squared distance between probability distributions is represented by the random inactivation inference count N. The number of random inactivation inference counts N can be set according to the computing resource configuration of the edge computing module, for example, 20, 30, or 50 times. The random inactivation inference method can be implemented by setting random inactivation probabilities in one or more hidden layers of the edge diagnostic network. When the state prediction probability distribution obtained from multiple random inferences tends to be consistent, the uncertainty metric U decreases; when the state prediction probability distribution obtained from multiple random inferences fluctuates significantly, the uncertainty metric U increases. In this way, the cross-modal consistency metric C is used to describe the degree of consistency in the state judgment of the target power distribution equipment among different sensing modalities, and the uncertainty metric U is used to describe the stability of the inference results of the edge diagnostic network itself. Both serve as the control basis for subsequent judgment threshold interval correction, diagnostic output mode selection, and incremental learning sample screening, thereby achieving adaptive edge diagnostics for multi-source heterogeneous sensing data.

[0069] Based on the operating condition information and the cross-modal consistency metric C, the baseline judgment threshold range is corrected to obtain the corrected judgment threshold range. The cross-modal consistency metric C is used as a damping function control parameter to adjust the adjustment range of the upper and lower limits of the baseline judgment threshold range. Based on the corrected judgment threshold range, the fusion diagnostic probability distribution is classified into state levels to obtain the edge-side state classification result.

[0070] Furthermore, the benchmark determination threshold range is: The correction includes:

[0071] Based on the reference load current Reference ambient temperature Reference ambient humidity For the baseline operating condition, calculate:

[0072] Load factor Temperature coefficient Humidity coefficient ,

[0073] in This is the current load current. The current ambient temperature. Define the current ambient humidity; define the time coefficient. This is the ratio of the time interval between the current data collection time and the most recent maintenance time to the preset standard maintenance cycle.

[0074] Define the damping function:

[0075] ;

[0076] in and The parameters are based on historical labeled samples, and the damping function is... The cross-modal consistency metric C approaches zero as it decreases;

[0077] The revised lower threshold is:

[0078] ;

[0079] The revised upper limit of the threshold is:

[0080] ;

[0081] in, , , , The coefficients are obtained based on regression calibration of historical labeled sample sets, and the corrected judgment threshold range is... ;

[0082] The edge computing module is configured with a verification window. Within this window, the proportion of abnormal state samples determined based on the revised judgment threshold range is statistically analyzed. When the proportion of abnormal state samples exceeds the statistical range determined by the average abnormality rate obtained from historical operational data and a preset multiple standard deviation, the revised judgment threshold range is reverted to the baseline judgment threshold range. .

[0083] Specifically, after obtaining the cross-modal consistency metric C, to adapt to the state determination requirements of the target power distribution equipment under different operating conditions, equipment aging levels, and changes in the reliability of multimodal diagnostic results, the edge computing module adaptively corrects the pre-set benchmark judgment threshold interval based on the operating condition information and the cross-modal consistency metric C, obtaining a corrected judgment threshold interval. Then, based on the corrected judgment threshold interval, the module classifies the fusion diagnostic probability distribution into state levels, obtaining the edge-side state classification result. The benchmark judgment threshold interval represents the initial judgment range for classifying the fusion diagnostic probability distribution into state levels under standard operating conditions, and can be expressed as (…). ),in This represents the lower limit of the baseline judgment threshold range. This indicates the upper limit of the baseline judgment threshold range. In one embodiment, the edge computing module determines the baseline judgment threshold range based on the historical normal operation data of the target power distribution equipment. For example, under the conditions of rated load, standard ambient temperature, and standard ambient humidity, the module statistically analyzes the range of changes in the fusion diagnostic probability distribution within the historical operating cycle and determines the baseline judgment threshold range based on the statistical results. Furthermore, the edge computing module calculates operating condition change parameters based on the operating condition information. Specifically, using the reference load current... Reference ambient temperature and baseline ambient humidity As a baseline operating condition, the load factor β, temperature factor γ, and humidity factor δ are calculated based on the load current I, ambient temperature T, and ambient humidity H from the current operating condition information. The specific calculation methods are as follows: ; Where I represents the current load current, This represents the reference load current, and T represents the current ambient temperature. H represents the reference ambient temperature, and H represents the current ambient humidity. This represents the baseline ambient humidity. The load factor... The load factor β is used to reflect the degree of change of the current load current relative to the reference load current. The temperature coefficient γ reflects the degree of change of the current ambient temperature relative to the reference ambient temperature, and the humidity coefficient δ reflects the degree of change of the current ambient humidity relative to the reference ambient humidity. For example, when the target power distribution equipment changes from a low-load state to a high-load operating state, the load current I increases, causing the load factor β to rise. The edge computing module can adjust the judgment threshold range according to a preset correction coefficient to adapt to the changes in equipment state characteristics under high-load operating conditions. Further, the edge computing module defines a time coefficient τ to characterize the change in operating time of the target power distribution equipment since the most recent maintenance. Specifically, the time coefficient τ is the ratio of the time interval between the current acquisition time and the most recent maintenance time to a preset standard maintenance cycle, and its calculation method is as follows: Where t represents the current data collection time, The time of the most recent maintenance is indicated by τ, and Ts represents the preset standard maintenance cycle. By setting the time coefficient τ, the edge computing module can adjust the state judgment threshold according to changes in equipment operating time. Furthermore, to avoid excessive threshold adjustment due to low-reliability diagnostic information when there are conflicts in diagnostic results from different modalities, the edge computing module uses the cross-modal consistency metric C as a damping function control parameter to limit the upper and lower limits of the baseline judgment threshold range. Specifically, a damping function is defined... : in and These are parameters obtained based on historical labeled samples. In one embodiment, the... and The calibration is performed as follows: Historical diagnostic samples of the target power distribution equipment under at least two different operating conditions are collected. For each historical diagnostic sample, its cross-modal consistency metric C and the corresponding optimal damping coefficient value are calculated. The optimization objective is to minimize the sum of squares of the threshold correction errors on each sample. The gradient descent method is used to solve for the optimization objective to obtain... and The value of is determined by performing a grid search on the calibration dataset for each historical diagnostic sample. C represents the cross-modal consistency metric. A higher cross-modal consistency metric C indicates a higher degree of consistency among the modal diagnostic probability distributions corresponding to the visual modality, thermal imaging modality, acoustic modality, and discharge characteristic modality. Increasing this value enhances the impact of operating condition factors on the correction of the judgment threshold range; when the cross-modal consistency metric C decreases, it indicates a significant diagnostic difference between different modes, and the damping function... The threshold is reduced and approaches zero, thereby reducing the threshold adjustment range. Further, the edge computing module adjusts the threshold based on the operating parameters and the damping function. The benchmark threshold range is corrected. Specifically, the corrected lower threshold L is obtained as follows: The corrected upper threshold U is obtained as follows: ;in, , , , These are coefficients obtained through regression calibration based on a historical labeled sample set. In one implementation, the... , , , The calibration is performed as follows: using the mean square error between the fusion diagnostic probability distribution of each sample in the historical labeled sample set and the manually labeled state level as the loss function, and using the baseline judgment threshold interval (L0, U0) as the initial value, the loss function is minimized using a multiple linear regression method. The normal equations are then solved using the least squares method to obtain the values ​​of kL, kU, a, and b. The historical labeled sample set includes diagnostic samples of the target power distribution equipment under various operating conditions. Each sample is labeled with a manually verified label corresponding to four state levels: normal, attention, abnormal, and severe abnormal. The corrected judgment threshold interval is (L, U). Through this method, the edge computing module can dynamically adjust the state judgment range based on the current load status, environmental conditions, equipment operating time, and the consistency of multimodal diagnostics. Subsequently, the edge computing module compares the state probabilities in the fusion diagnostic probability distribution with the corrected judgment threshold interval, and determines the edge-side state classification result based on the comparison result, such as normal state, attention state, abnormal state, or severe abnormal state. Furthermore, to prevent abnormal operating conditions from causing the corrected judgment threshold range to deviate from the actual equipment operating state for an extended period, the edge computing module sets a verification window and statistically analyzes the proportion of abnormal state samples determined based on the corrected judgment threshold range within the verification window. The edge computing module statistically obtains the average abnormality rate μ and a preset multiple standard deviation range σ based on historical operating data. When the proportion of abnormal state samples exceeds the statistical range determined by the average abnormality rate and the preset multiple standard deviation, the current threshold correction result is determined to be abnormal, and the corrected judgment threshold range is reverted to the baseline judgment threshold range (L0, U0). For example, if, within multiple consecutive verification windows, the proportion of abnormal state samples obtained based on the corrected judgment threshold range is significantly higher than the normal fluctuation range corresponding to historical operating data, it indicates that the current threshold correction may be affected by sudden environmental factors, multimodal data anomalies, or model errors. In this case, the edge computing module automatically cancels the current threshold correction result and reverts to using the baseline judgment threshold range for state level classification.

[0084] Based on the uncertainty metric U and the cross-modal consistency metric C, select one of the following diagnostic output modes: direct edge output, cloud collaborative review output, or alarm flag;

[0085] When the cloud-based collaborative review mode is selected, the fusion diagnostic probability distribution is sent to the cloud for review, the cloud-side state classification result is obtained, and the cloud-side state classification result is weighted and fused with the edge-side state classification result to output the cloud review fusion result.

[0086] Furthermore, the cloud-based collaborative review includes:

[0087] When the uncertainty metric U exceeds the first uncertainty threshold, or the cross-modal consistency metric C is lower than the first consistency threshold, the edge diagnostic network performs alignment projection mapping on the feature information of the diagnostic sample to generate alignment activation values.

[0088] The alignment activation value is uploaded to the cloud teacher network through the communication module. The cloud teacher network performs verification and inference based on the alignment activation value and outputs the cloud-side state classification result.

[0089] The cloud-side state classification result and the edge-side state classification result are weighted and fused to obtain the fused state classification result.

[0090] The alignment projection mapping layer is obtained through joint training during the training phase, with the edge diagnosis network as the student network and the cloud teacher network as the teacher network, so as to minimize the mean square error between the alignment activation value output by the alignment projection mapping layer and the preset intermediate layer output features of the cloud teacher network.

[0091] Furthermore, the selection of the diagnostic output mode includes:

[0092] When the uncertainty measure Below the first uncertainty threshold and the cross-modal consistency metric When the value exceeds the first consistency threshold, the edge-side direct output mode is adopted.

[0093] When the uncertainty measure The uncertainty is between the first uncertainty threshold and the second uncertainty threshold, or the cross-modal consistency measure. When the value is between the first consistency threshold and the second consistency threshold, a cloud-based collaborative review output mode is adopted.

[0094] When the uncertainty measure The cross-modal consistency metric is higher than the second uncertainty threshold. When the threshold is below the second consistency threshold, an alarm flagging mode is used;

[0095] Specifically, in the cloud-based collaborative review output mode, the edge diagnostic network outputs the edge-side state probability distribution. The cloud-side state probability distribution output by the cloud-based teacher network. Weighted fusion is performed to obtain the final state probability distribution. for:

[0096] ;

[0097] Weight Determine according to the following formula:

[0098] ;

[0099] in, and As a preset constant, The weights are parameters calibrated based on historical labeled samples. It decreases as the uncertainty metric U increases, and decreases as the cross-modal consistency metric increases. As it increases, the final diagnostic conclusion is based on the probability distribution of the final state. The state level with the highest probability.

[0100] Specifically, after obtaining the edge-side state classification result, to further improve the diagnostic reliability under complex operating conditions and in cases of conflicting multi-source sensing data, the edge computing module comprehensively evaluates the reliability of the current diagnostic task based on the uncertainty metric U and the cross-modal consistency metric C, and selects a diagnostic output mode from the edge-side direct output mode, cloud-based collaborative review output mode, and alarm marking mode based on the evaluation result. When the uncertainty metric U is lower than a first uncertainty threshold and the cross-modal consistency metric C is higher than a first consistency threshold, the edge computing module selects the edge-side direct output mode and directly outputs the edge-side state classification result. For example, when the modal diagnostic probability distributions output by the visible light image, infrared thermal image, acoustic signature signal, and partial discharge signal of the target power distribution equipment are relatively consistent, and the changes in the multiple state prediction probability distributions obtained by the edge diagnostic network through random inactivation inference are small, the edge computing module determines that the current diagnostic result is credible and directly outputs the edge-side state classification result. When the uncertainty metric U is higher than the second uncertainty threshold and the cross-modal consistency metric C is lower than the second consistency threshold, the edge computing module selects an alarm marking mode, marks the current target power distribution equipment status as abnormal, and generates alarm information. The alarm information includes the target power distribution equipment identifier, abnormal status level, corresponding abnormal confidence probability, and alarm trigger time. The first uncertainty threshold is lower than the second uncertainty threshold, and the first consistency threshold is higher than the second consistency threshold, forming a normal diagnostic region, a collaborative review region, and an abnormal alarm region. In one embodiment, the first uncertainty threshold and the second uncertainty threshold are determined as follows: The uncertainty metric values ​​obtained by the edge diagnostic network through random inactivation inference on multiple diagnostic samples on the validation set are acquired; the distribution of the uncertainty metric values ​​is statistically analyzed; the 75th percentile of the uncertainty metric values ​​is used as the first uncertainty threshold, and the 95th percentile is used as the second uncertainty threshold. The first consistency threshold and the second consistency threshold are determined as follows: Cross-modal consistency metrics corresponding to multiple diagnostic samples on the validation set are obtained; the distribution of the cross-modal consistency metrics is statistically analyzed; the 25th percentile of the cross-modal consistency metrics is used as the second consistency threshold; and the 5th percentile is used as the first consistency threshold. Further, when selecting the cloud-based collaborative review output mode, the edge computing module generates a review request based on the diagnostic task information corresponding to the fused diagnostic probability distribution, extracts key feature information affecting the state judgment of the target power distribution equipment based on the time-synchronized multi-source sensing data, sends the key feature information to the cloud for review to obtain a cloud-side state classification result, and weights and fuses the cloud-side state classification result with the edge-side state classification result to obtain a cloud-based review fusion result.Specifically, to reduce the data transmission volume between the edge computing module and the cloud-based teacher network, the edge computing module does not directly upload all time-synchronized multi-source sensing data. Instead, it extracts features from the time-synchronized multi-source sensing data through the edge diagnostic network to obtain edge-side feature vectors. The alignment projection mapping layer is connected after the feature extraction part of the edge diagnostic network. Using the edge-side feature vectors as input, it maps the edge-side feature vectors to a feature space consistent with the feature space of the cloud-based teacher network, generating alignment activation values. These alignment activation values ​​are then uploaded to the cloud-based teacher network through the communication module. Further, after receiving the alignment activation values, the cloud-based teacher network performs verification inference based on the alignment activation values ​​and outputs cloud-side state classification results. The cloud-side state classification results include the probability distribution of each state level corresponding to the target power distribution equipment. For example, for the four state levels of normal, attention, abnormal, and severe abnormality, the cloud-based teacher network outputs the cloud-side state probability distribution Pc=(pc1,pc2,pc3,pc4). Further, the alignment projection mapping layer is obtained through joint training during the training phase. Specifically, the edge diagnostic network deployed in the mobile inspection terminal serves as the student network, and the cloud-based teacher network deployed in the cloud serves as the teacher network. During the training phase, the same diagnostic samples are input into both the student and teacher networks to align the aligned activation values ​​output by the alignment projection mapping layer with the preset intermediate layer output features of the cloud-based teacher network. The parameters of the alignment projection mapping layer are optimized by minimizing the mean squared error loss function. The loss function can be expressed as: in, Indicates the first Alignment activation values ​​corresponding to each sample This indicates the output characteristics of the preset intermediate layer corresponding to the cloud-based teacher network. This indicates the number of training samples. Through the aforementioned joint training method, the aligned activation values ​​generated by the edge diagnostic network can retain key feature information related to the inference of the cloud-based teacher network. Furthermore, in the cloud-based collaborative review output mode, the edge computing module will distribute the edge-side state probability. Cloud-side state probability distribution output by the cloud-based teacher network Perform weighted fusion to obtain the final state probability distribution. The specific calculation formula is as follows: ;in, This represents the edge-side state probability distribution obtained based on the fusion diagnostic probability distribution and the corrected decision threshold interval. This represents the cloud-side state probability distribution of the teacher's network output. The fusion weight λ represents the probability distribution of the edge-side states. Further, the fusion weight λ is dynamically determined based on the uncertainty metric U and the cross-modal consistency metric C, and the specific calculation formula is as follows: ;in, and As a preset constant, These are parameters obtained based on historical labeled samples. When the uncertainty measure... When the weight increases, it indicates that the inference stability of the edge diagnostic network itself decreases, and the weight... This reduces the proportion of cloud-based teacher network output in the fusion results; when the cross-modal consistency measure Increasing the weight indicates improved consistency of diagnostic results across different modalities, improved reliability of the edge state probability distribution, and improved weighting. The increase in the proportion of edge diagnostic results enhances the diagnostic accuracy. For example, in a ring main unit, when infrared thermal imaging detects an abnormal local temperature rise, but acoustic signature and partial discharge signals do not show significant abnormalities, the differences between the diagnostic probability distributions of different modes are large, reducing the cross-modal consistency metric C. Simultaneously, the probability distributions of multiple state predictions fluctuate significantly during random inactivation inference, increasing the uncertainty metric U. In this case, the edge computing module selects the cloud-based collaborative verification output mode, uploads the aligned activation value to the cloud-based teacher network, and fuses the edge-side state probability distribution Pe and the cloud-side state probability distribution Pc according to the dynamically calculated weight λ. Finally, the edge computing module selects the state level with the highest probability in the final state probability distribution Pf as the final diagnostic conclusion for the target power distribution equipment, and outputs the edge-side state classification result, the cloud-based verification fusion result, or the alarm labeling result according to the diagnostic output mode. Through this method, adaptive diagnostic output mode selection based on the uncertainty metric U and the cross-modal consistency metric C is achieved, improving the accuracy and reliability of target power distribution equipment state diagnosis in complex operating environments while ensuring real-time edge-side performance.

[0101] Diagnostic samples whose cross-modal consistency metric C is lower than a preset consistency screening threshold and whose cloud-side state classification results are inconsistent with the edge-side state classification results are included in the local incremental learning sample pool, and the edge diagnostic network parameters are updated based on the local incremental learning sample pool; wherein, the cross-modal consistency metric C also serves as the damping control parameter for the correction of the benchmark judgment threshold range, the decision control parameter for the selection of the diagnostic output mode, and the control parameter for the selection of incremental learning samples.

[0102] Specifically, after obtaining the cloud-side state classification result and the edge-side state classification result, the edge computing module further performs incremental learning sample screening on the current diagnostic sample based on the cross-modal consistency metric C and the consistency relationship between the two state classification results. Specifically, the edge computing module determines whether the cross-modal consistency metric C corresponding to the current diagnostic sample is lower than a preset consistency screening threshold, and determines whether the cloud-side state classification result and the edge-side state classification result are consistent. In one embodiment, the preset consistency screening threshold is determined as follows: samples from historical diagnostic samples that have been reviewed and corrected by the cloud and are inconsistent with the edge-side output state classification result are obtained, and the distribution of the cross-modal consistency metric C value corresponding to the sample is statistically analyzed. The 90th percentile of the distribution is used as the preset consistency screening threshold. When the cross-modal consistency metric C is lower than the preset consistency screening threshold, and the cloud-side state classification result is inconsistent with the edge-side state classification result, it indicates that the current diagnostic sample has a large modal conflict or that the edge diagnostic network has not fully learned the state features. The edge computing module then includes the diagnostic sample in the local incremental learning sample pool. For example, for a specific power distribution device, the edge diagnostic network outputs an edge-side state classification result of "Attention" based on the multi-source sensing data, while the cloud-based teacher network outputs a cloud-side state classification result of "Abnormal" after verification based on the aligned activation value. Simultaneously, there are significant differences in the modal diagnostic probability distributions corresponding to the visual modality, thermal imaging modality, acoustic modality, and discharge characteristic modality, causing the cross-modal consistency metric C to fall below the preset consistency screening threshold. In this case, the edge computing module stores the diagnostic sample as an incremental learning sample in the local incremental learning sample pool. Further, when the number of diagnostic samples in the local incremental learning sample pool reaches a preset update condition or meets a preset update cycle, the edge computing module updates the edge diagnostic network parameters based on the diagnostic samples in the local incremental learning sample pool, enabling the edge diagnostic network to learn new device state features. The diagnostic samples in the local incremental learning sample pool may include at least one of the feature information corresponding to the time-synchronized multi-source sensing data, the edge-side state classification result, and the cloud-based state classification result, serving as the data basis for updating the edge diagnostic network parameters. Furthermore, the cross-modal consistency metric C also serves as the damping control parameter for the correction of the benchmark judgment threshold interval, the decision control parameter for the selection of the diagnostic output mode, and the control parameter for the selection of incremental learning samples.Specifically, during the baseline judgment threshold range correction process, the cross-modal consistency metric C is used to adjust the threshold correction magnitude; during the diagnostic output mode selection process, the cross-modal consistency metric C is used to determine the degree of conflict between different modal diagnostic results; during the incremental learning sample screening process, the cross-modal consistency metric C is used to screen diagnostic samples with high learning value. Through the correlation control of the cross-modal consistency metric C in the above different processing stages, the edge diagnostic network can achieve coordinated adaptation of threshold adjustment, diagnostic decision-making, and model optimization based on changes in multimodal diagnostic consistency.

[0103] Example 2, based on the same inventive concept as the edge diagnosis method of the mobile inspection terminal for power distribution equipment in the foregoing examples, such as... Figure 2 As shown, this application provides an edge diagnostic system for a mobile inspection terminal for power distribution equipment. The system and method embodiments in this application are based on the same inventive concept. The system includes:

[0104] The multi-source sensing and operating condition acquisition unit 10 is used to collect multi-source sensing data of the target power distribution equipment through the multi-source sensing module and to acquire the operating condition information of the target power distribution equipment through the communication module. The multi-source sensing data includes at least visible light images, infrared thermal images, acoustic fingerprint signals and partial discharge signals, which correspond to visual mode, thermal imaging mode, acoustic mode and discharge characteristic mode, respectively. The operating condition information includes at least one of load current, ambient temperature and ambient humidity.

[0105] The time synchronization unit 20 is used to perform time synchronization processing on the multi-source sensing data based on the time reference provided by the positioning and timing module, and to perform acoustic propagation delay compensation on the voiceprint signal to obtain the time-synchronized multi-source sensing data.

[0106] The multimodal fusion unit 30 is used by the edge computing module to run the edge diagnostic network, perform inference on each modal data after time synchronization, obtain the diagnostic probability distribution of each modality, determine the fusion weight according to the data quality index of each modal data, and perform weighted fusion on the diagnostic probability distribution of each modality based on the fusion weight to obtain the fused diagnostic probability distribution.

[0107] A cross-modal consistency and uncertainty calculation unit 40 is used to determine the cross-modal consistency metric C based on the pairwise consistency distance between the various modal diagnostic probability distributions; to perform multiple random inferences on the fused diagnostic probability distribution using a random inactivation inference method; and to determine the uncertainty metric U based on the degree of dispersion between the state prediction probability distributions obtained from the multiple inferences.

[0108] The threshold correction and classification unit 50 is used to correct the baseline judgment threshold interval based on the operating condition information and the cross-modal consistency metric C to obtain the corrected judgment threshold interval. The cross-modal consistency metric C is used as a damping function control parameter to adjust the adjustment range of the upper and lower limits of the baseline judgment threshold interval. The unit 50 also performs state level classification on the fusion diagnostic probability distribution based on the corrected judgment threshold interval to obtain the edge-side state classification result.

[0109] The output mode selection unit 60 is used to select one of the following diagnostic output modes: edge-side direct output, cloud-based collaborative verification output, or alarm flag, based on the uncertainty metric U and the cross-modal consistency metric C.

[0110] The cloud-based verification unit 70 is used to send the fusion diagnostic probability distribution to the cloud for verification when the cloud collaborative verification mode is selected, obtain the cloud-side state classification result, and perform weighted fusion of the cloud-side state classification result and the edge-side state classification result to output the cloud-based verification fusion result.

[0111] The incremental learning update unit 80 is used to include diagnostic samples whose cross-modal consistency metric C is lower than a preset consistency screening threshold and whose cloud-side state classification results are inconsistent with the edge-side state classification results into a local incremental learning sample pool, and to update the edge diagnostic network parameters based on the local incremental learning sample pool; wherein, the cross-modal consistency metric C also serves as a damping control parameter for the correction of the benchmark judgment threshold range, a decision control parameter for the selection of the diagnostic output mode, and a control parameter for the selection of incremental learning samples.

Claims

1. An edge diagnostic method for a mobile inspection terminal for power distribution equipment, wherein the mobile inspection terminal for power distribution equipment includes a multi-source sensing module, an edge computing module, a positioning and timing module, and a communication module, characterized in that, include: The multi-source sensing module collects multi-source sensing data of the target power distribution equipment, and the communication module obtains the operating condition information of the target power distribution equipment. The multi-source sensing data includes at least visible light images, infrared thermal images, acoustic fingerprint signals, and partial discharge signals, which correspond to visual mode, thermal imaging mode, acoustic mode, and discharge characteristic mode, respectively. The operating condition information includes at least one of load current, ambient temperature, and ambient humidity. Based on the time reference provided by the positioning and timing module, the multi-source sensing data is time-synchronized, and the acoustic propagation delay of the voiceprint signal is compensated to obtain the time-synchronized multi-source sensing data. The edge computing module runs an edge diagnostic network to perform inference on each modal data after time synchronization to obtain the diagnostic probability distribution of each modality; it determines the fusion weight based on the data quality index of each modality data, and performs weighted fusion on the diagnostic probability distribution of each modality based on the fusion weight to obtain the fused diagnostic probability distribution. Based on the pairwise consistency distance between the various modal diagnostic probability distributions, a cross-modal consistency metric C is determined; multiple random inferences are performed on the fused diagnostic probability distribution using a random inactivation inference method, and an uncertainty metric U is determined based on the degree of dispersion between the state prediction probability distributions obtained from the multiple inferences. Based on the operating condition information and the cross-modal consistency metric C, the baseline judgment threshold range is corrected to obtain the corrected judgment threshold range. The cross-modal consistency metric C is used as a damping function control parameter to adjust the adjustment range of the upper and lower limits of the baseline judgment threshold range. Based on the corrected judgment threshold range, the fusion diagnostic probability distribution is classified into state levels to obtain the edge-side state classification result. Based on the uncertainty metric U and the cross-modal consistency metric C, select one of the following diagnostic output modes: direct edge output, cloud collaborative review output, or alarm flag; When the cloud-based collaborative review mode is selected, the fusion diagnostic probability distribution is sent to the cloud for review, the cloud-side state classification result is obtained, and the cloud-side state classification result is weighted and fused with the edge-side state classification result to output the cloud review fusion result. Diagnostic samples whose cross-modal consistency metric C is lower than a preset consistency screening threshold and whose cloud-side state classification results are inconsistent with the edge-side state classification results are included in the local incremental learning sample pool, and the edge diagnostic network parameters are updated based on the local incremental learning sample pool; wherein, the cross-modal consistency metric C also serves as the damping control parameter for the correction of the benchmark judgment threshold range, the decision control parameter for the selection of the diagnostic output mode, and the control parameter for the selection of incremental learning samples.

2. The edge diagnosis method for a mobile inspection terminal of power distribution equipment according to claim 1, characterized in that, The acoustic propagation delay compensation includes: The edge computing module acquires the distance d between the acoustic signature acquisition unit in the multi-source sensing module and the target power distribution equipment, as well as the current ambient sound speed v, and calculates the acoustic propagation delay. ; Using the time reference provided by the positioning and timing module as a reference time, the acquisition time of the voiceprint signal is subtracted from the time referenced by the time referenced by the positioning and timing module. This aligns the acquisition time of the compensated voiceprint signal with the acquisition times of the visible light image, infrared thermal image, and partial discharge signal to the same time reference.

3. The edge diagnosis method for a mobile inspection terminal of power distribution equipment according to claim 1, characterized in that, The data quality indicators include: The edge gradient degradation index of a visible light image is determined by extracting the gradient magnitude distribution of the visible light image and then statistically analyzing the proportion of pixels that are below a preset gradient threshold. The temperature gradient entropy of the infrared thermal image is determined by statistically analyzing the gradient entropy after extracting the temperature field gradient of the infrared thermal image. The signal-to-noise ratio of the voiceprint signal is determined by the ratio of the effective power of the voiceprint signal to the background noise power. The concentration of the pulse phase distribution of the partial discharge signal is determined by the reciprocal of the variance of the phase distribution of the partial discharge pulse within the power frequency cycle. When any modality data is missing, the data quality index score for that modality is recorded as zero, and the data quality scores of the remaining modalities are renormalized to determine the fusion weight of each modality data.

4. The edge diagnosis method for a mobile inspection terminal of power distribution equipment according to claim 1, characterized in that, The edge computing module determines the remaining inference times based on the remaining inspection path length between the current inspection location and the preset inspection endpoint, combined with the number of diagnostic samples corresponding to the unit path length in the historical inspection process. Obtain the resource consumption parameters corresponding to a single diagnostic inference when executing a high-precision diagnostic inference path, and determine the high-precision diagnostic resource requirements required to complete the remaining inspection stage based on the remaining inference count and the resource consumption parameters corresponding to the single diagnostic inference. When the demand for high-precision diagnostic resources exceeds the current available computing resources of the edge computing module and cannot support the high-precision diagnostic inference path to complete the remaining inference times, the current inference path will be switched from the high-precision diagnostic inference path to the lightweight diagnostic inference path. During the K diagnostic cycles of the inference path switching transition phase, the diagnostic probability distributions before and after the switch are weighted and fused according to linearly gradual weights. The high-precision diagnostic inference path fusion weights corresponding to the i-th diagnostic cycle are: , where i=1,2,...,K, and K is the preset number of transitional diagnostic cycles.

5. The edge diagnosis method for a mobile inspection terminal of power distribution equipment according to claim 1, characterized in that, The cross-modal consistency metric C is obtained in the following manner: The modal diagnostic probability distributions corresponding to the visual modality, thermal imaging modality, acoustic modality, and discharge characteristic modality are compared pairwise to calculate the consistency distance between each modal diagnostic probability distribution. The cross-modal consistency metric C is obtained based on the weighted average of the pairwise modal consistency distances.

6. The edge diagnosis method for a mobile inspection terminal of power distribution equipment according to claim 1, characterized in that, The benchmark judgment threshold range is: The correction includes: Based on the reference load current Reference ambient temperature Reference ambient humidity For the baseline operating condition, calculate: Load factor Temperature coefficient Humidity coefficient , in This is the current load current. The current ambient temperature. Define the current ambient humidity; define the time coefficient. This is the ratio of the time interval between the current data collection time and the most recent maintenance time to the preset standard maintenance cycle. Define the damping function: ； in and The parameters are based on historical labeled samples, and the damping function is... The cross-modal consistency metric C approaches zero as it decreases; The revised lower threshold is: ； The revised upper limit of the threshold is: ； in, , , , The coefficients are obtained based on regression calibration of historical labeled sample sets, and the corrected judgment threshold range is... ; The edge computing module is configured with a verification window. Within this window, the proportion of abnormal state samples determined based on the revised judgment threshold range is statistically analyzed. When the proportion of abnormal state samples exceeds the statistical range determined by the average abnormality rate obtained from historical operational data and a preset multiple standard deviation, the revised judgment threshold range is reverted to the baseline judgment threshold range. .

7. The edge diagnosis method for a mobile inspection terminal of power distribution equipment according to claim 1, characterized in that, The cloud-based collaborative review includes: When the uncertainty metric U exceeds the first uncertainty threshold, or the cross-modal consistency metric C is lower than the first consistency threshold, the edge diagnostic network performs alignment projection mapping on the feature information of the diagnostic sample to generate alignment activation values. The alignment activation value is uploaded to the cloud teacher network through the communication module. The cloud teacher network performs verification and inference based on the alignment activation value and outputs the cloud-side state classification result. The cloud-side state classification result and the edge-side state classification result are weighted and fused to obtain the fused state classification result. The alignment projection mapping layer is obtained through joint training during the training phase, with the edge diagnosis network as the student network and the cloud teacher network as the teacher network, so as to minimize the mean square error between the alignment activation value output by the alignment projection mapping layer and the preset intermediate layer output features of the cloud teacher network.

8. The edge diagnosis method for a mobile inspection terminal of power distribution equipment according to claim 7, characterized in that, The selection of the diagnostic output mode includes: When the uncertainty measure Below the first uncertainty threshold and the cross-modal consistency metric When the value exceeds the first consistency threshold, the edge-side direct output mode is adopted. When the uncertainty measure The uncertainty is between the first uncertainty threshold and the second uncertainty threshold, or the cross-modal consistency measure. When the value is between the first consistency threshold and the second consistency threshold, a cloud-based collaborative review output mode is adopted. When the uncertainty measure The cross-modal consistency metric is higher than the second uncertainty threshold. When the threshold is below the second consistency threshold, an alarm flagging mode is used; Specifically, in the cloud-based collaborative review output mode, the edge diagnostic network outputs the edge-side state probability distribution. The cloud-side state probability distribution output by the cloud-based teacher network. Weighted fusion is performed to obtain the final state probability distribution. for: ； Weight Determine according to the following formula: ； in, and As a preset constant, The weights are parameters calibrated based on historical labeled samples. It decreases as the uncertainty metric U increases, and decreases as the cross-modal consistency metric increases. As it increases, the final diagnostic conclusion is based on the probability distribution of the final state. The state level with the highest probability.

9. An edge diagnostic system for a mobile inspection terminal of power distribution equipment, characterized in that, The system is used to implement the edge diagnosis method for a mobile inspection terminal of power distribution equipment as described in any one of claims 1 to 8, and the system includes: A multi-source sensing and operating condition acquisition unit is used to collect multi-source sensing data of the target power distribution equipment through the multi-source sensing module, and to acquire the operating condition information of the target power distribution equipment through the communication module. The multi-source sensing data includes at least visible light images, infrared thermal images, acoustic fingerprint signals, and partial discharge signals, which correspond to visual mode, thermal imaging mode, acoustic mode, and discharge characteristic mode, respectively. The operating condition information includes at least one of load current, ambient temperature, and ambient humidity. A time synchronization unit is used to perform time synchronization processing on the multi-source sensing data based on the time reference provided by the positioning and timing module, and to perform acoustic propagation delay compensation on the voiceprint signal to obtain time-synchronized multi-source sensing data. A multimodal fusion unit is used by the edge computing module to run the edge diagnostic network, perform inference on each modal data after time synchronization, and obtain the diagnostic probability distribution of each modality; determine the fusion weight according to the data quality index of each modal data, and perform weighted fusion on the diagnostic probability distribution of each modality based on the fusion weight to obtain the fused diagnostic probability distribution. A cross-modal consistency and uncertainty calculation unit is used to determine the cross-modal consistency metric C based on the pairwise consistency distance between the probability distributions of each modality diagnosis; to perform multiple random inferences on the fused diagnosis probability distribution using a random inactivation inference method, and to determine the uncertainty metric U based on the degree of dispersion between the state prediction probability distributions obtained from the multiple inferences. A threshold correction and classification unit is used to correct the baseline judgment threshold interval based on the operating condition information and the cross-modal consistency metric C to obtain a corrected judgment threshold interval. The cross-modal consistency metric C is used as a damping function control parameter to adjust the adjustment range of the upper and lower limits of the baseline judgment threshold interval. The unit also performs state level classification on the fusion diagnostic probability distribution based on the corrected judgment threshold interval to obtain the edge-side state classification result. An output mode selection unit is used to select one of the following diagnostic output modes: edge-side direct output, cloud-based collaborative verification output, or alarm flag, based on the uncertainty metric U and the cross-modal consistency metric C. The cloud-based verification unit is used to send the fusion diagnostic probability distribution to the cloud for verification when the cloud collaborative verification mode is selected, obtain the cloud-side state classification result, and perform weighted fusion of the cloud-side state classification result and the edge-side state classification result to output the cloud-based verification fusion result. An incremental learning update unit is used to include diagnostic samples whose cross-modal consistency metric C is lower than a preset consistency screening threshold and whose cloud-side state classification results are inconsistent with the edge-side state classification results into a local incremental learning sample pool, and to update the edge diagnostic network parameters based on the local incremental learning sample pool; wherein, the cross-modal consistency metric C also serves as a damping control parameter for the correction of the benchmark judgment threshold range, a decision control parameter for the selection of the diagnostic output mode, and a control parameter for the selection of incremental learning samples.

Images