Compensation device state monitoring method and system

Abstract

The application discloses a kind of compensation device state monitoring method and system, it is related to electrical equipment state monitoring technical field, wherein method includes: synchronous acquisition device multidimensional image and real-time parameter;AI model is quantized physical defect from image depth analysis;Fusion defect information and parameter constructs multidimensional health vector;Risk is evaluated based on vector, dynamic model and historical data;Risk triggers hierarchical diagnosis alarm above self-adapting threshold value.The application is through comprehensive synchronous data acquisition, AI fine physical defect identification, multimodal information depth fusion, dynamic self-adapting risk evaluation and intelligent hierarchical and diagnostic alarm, effectively solve the state evaluation of compensation device in prior art is not accurate, risk early warning lag and the problem of insufficient operation and maintenance decision support, improve the intelligent, precision level and overall operation and maintenance efficiency of self-check.

Description

Technical Field

[0001] This invention relates to the field of electrical equipment condition monitoring technology, and in particular to a condition monitoring method and system for a compensation device. Background Technology

[0002] In power systems and industrial applications, compensation devices are key equipment for improving power quality and ensuring stable system operation. Real-time monitoring and accurate assessment of their health status are crucial. With the increasing complexity of equipment and the diversification of operating environments, traditional methods such as periodic inspections and monitoring based on single parameter thresholds are gradually showing their limitations in comprehensively understanding the actual operating conditions of compensation devices and early potential risks.

[0003] Existing technologies for monitoring and assessing the condition of compensation devices often suffer from insufficient information dimensions and low levels of intelligence. For example, some methods rely excessively on manual visual inspection or simple sensor data interpretation, making it difficult to capture subtle physical defects or complex correlations between abnormal operating parameters. Furthermore, image data analysis often remains superficial, lacking in-depth feature extraction and quantitative assessment capabilities. The effective fusion and collaborative analysis of multi-source heterogeneous data is also a significant challenge, resulting in a lack of comprehensiveness and accuracy in assessing the overall health of the equipment.

[0004] Therefore, existing technologies are still insufficient in timely detection of early-stage potential faults in compensation devices, accurate assessment of their degradation levels, and provision of intelligent operation and maintenance decision support. These deficiencies may lead to delayed maintenance responses or even unexpected downtime, impacting the system's reliable power supply and production efficiency. Summary of the Invention

[0005] This invention provides a method and system for monitoring the status of a compensation device, which addresses the problems of inaccurate status assessment, delayed risk warning, and insufficient support for operation and maintenance decisions in existing technologies.

[0006] In view of this, the first aspect of the present invention provides a method for monitoring the state of a compensation device, the method comprising:

[0007] The image data of the multi-dimensional physical state of the compensation device and the real-time operating parameters characterizing its dynamic working characteristics are acquired synchronously through the pre-configured image acquisition device and sensor unit and pre-processed.

[0008] By using a first artificial intelligence model optimized for a specific type of physical defect in the compensation device, deep feature extraction, parsing, identification, and quantification are performed on the image data to obtain key feature information of the physical defects.

[0009] The key feature information and the real-time operating parameters are subjected to collaborative multimodal fusion processing to construct a multidimensional fusion state vector that dynamically represents the overall health status of the compensation device;

[0010] Based on the preset dynamic evaluation model, combined with the multi-dimensional fusion state vector, the historical state data of the compensation device, and the real-time operating parameters, the abnormal risk of the compensation device is evaluated and determined, and the determination result is obtained.

[0011] The determination result is matched with a preset adaptive alarm threshold to trigger an alarm signal with hierarchical or diagnostic indication information.

[0012] Optionally, the step of synchronously acquiring image data of the multi-dimensional physical state of the compensation device and real-time operating parameters characterizing its dynamic working characteristics through a pre-configured image acquisition device and sensor unit, and performing preprocessing, includes:

[0013] The image acquisition device acquires multi-dimensional physical state image data, including visible light image data and infrared thermal image data, according to a preset periodic acquisition strategy or in response to a preset event triggering condition, and performs image enhancement and image denoising operations on the image data.

[0014] The sensor unit collects real-time operating parameters, including the temperature parameters of the compensation device body, three-phase voltage parameters, three-phase current parameters, and the temperature and humidity parameters of the environment in which the compensation device is located. It then performs parameter filtering and parameter scaling operations on the real-time operating parameters.

[0015] Optionally, the first artificial intelligence model is a convolutional neural network object detection model;

[0016] The trained convolutional neural network target detection model is used to identify and output the category of physical defects, spatial location coordinates, confidence score representing the reliability of the identification result, and quantitative severity index representing the degree of development or scope of influence of the defect from the image data.

[0017] Optionally, the convolutional neural network object detection model includes the following processing during the training phase:

[0018] A composite loss function is used to guide the adjustment of model parameters. The composite loss function is calculated as follows:

[0019] ;

[0020] In the formula, The total loss value used to evaluate the overall difference between the model predictions and the ground truth labels is the composite loss function. Represents the positioning loss term; For the positioning loss term Weighting coefficients; Represents the confidence level loss term; Represents the category loss term;

[0021] Among them, the positioning loss term The formula for calculating it is expressed as follows:

[0022] ;

[0023] In the formula, For the positioning loss term The calculation results; Represents the defect bounding box predicted by the model. Compared with manually annotated actual defect boundary boxes The crossover ratio between them; Represents the defect bounding box predicted by the model. The center point and the manually annotated actual defect boundary box The Euclidean distance between the center points; This represents the bounding box that can completely enclose the defect predicted by the model. Compared with manually annotated actual defect boundary boxes The length of the diagonal of the smallest bounding rectangle; It is a positive equilibrium parameter; It is used to measure the defect bounding box predicted by the model. Compared with manually annotated actual defect boundary boxes The parameter representing the difference in aspect ratio between the two.

[0024] Optionally, the step of performing collaborative multimodal fusion processing of the key feature information and the real-time operating parameters to construct a multidimensional fusion state vector that dynamically represents the overall health status of the compensation device includes:

[0025] The key feature information is processed through preset encoding rules to form a structured image defect feature vector, which is then denoted as the image defect feature vector.

[0026] The real-time operating parameters are selected and organized to form a structured operating parameter feature vector, which is denoted as the operating parameter feature vector.

[0027] The image defect feature vector and the operating parameter feature vector are fed together as joint inputs to a pre-constructed and trained multimodal fusion network. The multimodal fusion network integrates, correlates and complements the image defect feature vector and the operating parameter feature vector through its internally designed feature learning module and cross-modal information interaction mechanism, generating a multidimensional fusion state vector that reflects the current overall health status of the compensation device.

[0028] Optionally, the multimodal fusion network's processing of the image defect feature vector and the operating parameter feature vector includes:

[0029] A deep feature-level fusion is performed between the defect feature vector and the operating parameter feature vector through a cross-attention mechanism; wherein, the calculation process of the cross-attention mechanism includes: dynamically calculating the correlation weights between different modal features, so that the multimodal fusion network adaptively identifies information combinations that are more indicative of the current health status assessment of the compensation device;

[0030] The expression for the calculation process of the cross-attention mechanism is as follows:

[0031] ;

[0032] In the formula, This represents the fused feature representation calculated through the cross-attention mechanism; The feature vector that serves as the input for the query; The feature vector that serves as the key input; The feature vector represents the input value. , , These are the learnable weight matrices; This represents the matrix transpose operation; Indicates the process The dimension of the projected key feature vector.

[0033] Optionally, the assessment and determination of the abnormal risk of the compensation device based on the preset dynamic evaluation model, combined with the multi-dimensional fused state vector, the historical state data of the compensation device, and the real-time operating parameters, includes:

[0034] Based on a preset dynamic evaluation model, specific parameters are selected from the real-time operating parameters or specific indicators are selected from the multi-dimensional fused state vector. The baseline range for normal operation of the specific parameters or specific indicators is determined based on historical state data and current operating conditions.

[0035] Based on a preset risk assessment logic, the current value of the specific parameter or the specific indicator is compared with the baseline range, and the current abnormal risk level of the compensation device is determined by combining the key feature information.

[0036] Optionally, the upper limit of the baseline range The calculation method is as follows:

[0037] ;

[0038] In the formula, This represents the specific parameter or indicator at the current time of observation. The upper limit of the dynamic baseline; This represents the specific parameter or indicator at the current time of observation. Exponentially weighted moving average; It is a preset standard deviation factor; This represents a series of historical observations of the specific parameter or the specific indicator extracted from the historical state data.

[0039] Optionally, matching the determination result with a preset adaptive alarm threshold to trigger an alarm signal with hierarchical or diagnostic indication information includes:

[0040] When the determination result matches a preset complex anomaly risk or a preset key anomaly risk, the multi-dimensional physical state image data, the real-time operating parameters, the key feature information of the physical defect, and the multi-dimensional fused state vector are input into a preset second artificial intelligence model; the second artificial intelligence model is then fine-tuned and optimized by the compensation device therein before outputting a diagnostic report.

[0041] The diagnostic report includes operational guidance suggestions for the identified abnormal risks and serves as a component of the diagnostic indication information; the second artificial intelligence model is a multimodal large language model.

[0042] A second aspect of the present invention provides a status monitoring system for a compensation device, the system comprising:

[0043] The data acquisition unit is used to synchronously acquire image data of the multi-dimensional physical state of the compensation device and real-time operating parameters characterizing its dynamic working characteristics through a pre-configured image acquisition device and sensor unit, and to perform preprocessing.

[0044] The defect identification unit is used to extract, parse, identify and quantify the image data using a first artificial intelligence model optimized for a specific type of physical defect in the compensation device, so as to obtain key feature information of the physical defect.

[0045] The data fusion unit is used to perform collaborative multimodal fusion processing on the key feature information and the real-time operating parameters to construct a multidimensional fusion state vector that dynamically represents the overall health status of the compensation device.

[0046] The risk assessment unit is used to assess and determine the abnormal risks of the compensation device based on a preset dynamic assessment model, combined with the multi-dimensional fusion state vector, the historical state data of the compensation device, and the real-time operating parameters, and to obtain the determination result.

[0047] An alarm triggering unit is used to match the determination result with a preset adaptive alarm threshold and trigger an alarm signal with hierarchical or diagnostic indication information.

[0048] As can be seen from the above technical solutions, the present invention has the following advantages:

[0049] 1. This invention achieves comprehensive perception of equipment status information by synchronously acquiring multi-dimensional physical state images and real-time operating parameters of the compensation device through a preset image acquisition device and sensor unit. Compared with existing technologies that rely on a single data source or asynchronous data acquisition, this invention effectively solves the problem of inaccurate equipment status assessment caused by incomplete information or misaligned timing, laying a solid foundation for subsequent accurate diagnosis.

[0050] 2. This invention employs a first artificial intelligence model optimized for specific physical defects to perform deep analysis of image data to identify and quantify key defect features, thereby improving the automation and precision of physical defect identification. Compared with traditional manual inspection or simple image processing techniques, it overcomes the limitations of strong subjectivity, low efficiency, and difficulty in quantifying defect details, making defect assessment more objective and in-depth.

[0051] 3. This invention innovatively integrates identified physical defect features with real-time operating parameters through collaborative multimodal fusion processing, constructing a multidimensional fused state vector that dynamically characterizes the overall health status of the compensation device. This differs from existing technologies that typically analyze visual information and sensor data separately. By deeply fusing multi-source information, it uncovers the intrinsic correlations between different modal data, significantly improving the comprehensiveness and accuracy of the health status assessment of the compensation device.

[0052] 4. This invention employs an intelligent risk assessment method based on multi-dimensional fusion state vectors, combined with a dynamic evaluation model and historical data. This method dynamically determines the normal operating baseline of equipment and accurately identifies abnormal risks accordingly. Compared to existing technologies that commonly use fixed thresholds for alarms, this invention solves the problem of frequent false alarms and missed alarms caused by their inability to adapt to changes in equipment operating conditions and aging trends, thus improving the reliability and adaptability of risk warnings.

[0053] 5. When an anomaly risk reaches a preset threshold, this invention can not only automatically trigger tiered alarms, but also, for complex or critical anomalies, invoke a second artificial intelligence model optimized with domain knowledge to perform in-depth diagnostic analysis, outputting detailed cause analysis and operation and maintenance guidance. Compared to traditional alarm systems that only provide simple alarm signals and lack in-depth fault diagnosis capabilities and specific solution guidance, this significantly improves the intelligence level of fault handling and the efficiency of operation and maintenance decisions. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 A flowchart illustrating a state monitoring method for a compensation device provided in an embodiment of the present invention;

[0056] Figure 2 This is a schematic diagram of the structure of a state monitoring system for a compensation device provided in an embodiment of the present invention. Detailed Implementation

[0057] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0058] This invention provides a method for monitoring the state of a compensation device, the method comprising the following steps:

[0059] S1. Through the pre-configured image acquisition device and sensor unit, synchronously acquire image data of the multi-dimensional physical state of the compensation device and real-time operating parameters characterizing its dynamic working characteristics, and perform preprocessing.

[0060] In this embodiment, in order to achieve a comprehensive perception of the state of the compensation device, the core of step S1 is to synchronously acquire multi-dimensional physical state image data of the compensation device and real-time operating parameters that accurately characterize its dynamic working characteristics through a pre-configured image acquisition device and sensor unit, and to perform preliminary processing on these raw data in order to improve the quality and efficiency of subsequent analysis.

[0061] Specifically, the acquisition and preliminary processing of multi-dimensional physical state image data of the compensation device are achieved as follows:

[0062] Image acquisition devices, such as industrial cameras or infrared thermal imagers, are deployed in locations capable of effectively monitoring critical parts or the overall appearance of the compensation device. These image acquisition devices perform image capture tasks according to preset acquisition logic.

[0063] In one possible implementation, the acquisition logic is a preset periodic acquisition strategy. For example, the system is set to automatically trigger an image acquisition action every fixed time interval (e.g., several minutes or several hours, the specific interval is determined according to the type and importance of the compensation device) to obtain image data reflecting the physical state of the compensation device at that moment.

[0064] As an alternative or supplement, the acquisition logic also responds to preset event triggering conditions. For example, when one or more real-time operating parameters (such as temperature and current) detected by the sensor unit show abnormal fluctuations or exceed the warning threshold, the system immediately triggers the image acquisition device to capture the physical state of the equipment at the moment the abnormality occurs.

[0065] The acquired multi-dimensional physical state image data includes at least visible light image data that reflects information such as the external structure of the compensation device, the surface condition of connecting components, and whether there is deformation or contamination, as well as infrared thermal image data that reflects information such as the temperature distribution of each component of the compensation device and the presence of abnormal heat points in a non-contact manner. These two modalities of image data can characterize the physical state of the compensation device from different dimensions.

[0066] After acquiring the raw image data, preliminary image processing operations need to be performed on the acquired image data in order to improve image quality and reduce the interference of noise on subsequent recognition and analysis.

[0067] These operations include image enhancement, which aims to improve the visual quality of an image, highlight details, and make potential physical defects easier for subsequent artificial intelligence models to identify. For example, histogram equalization can be used to expand the dynamic range of an image's grayscale, or contrast stretching can be used to enhance the contrast between the target and the background in an image.

[0068] These operations also include image denoising, which aims to remove various noises that may be introduced during image acquisition and transmission, such as random or fixed-pattern noise caused by electromagnetic interference, uneven lighting, or sensor characteristics. For example, Gaussian filtering is used to smooth high-frequency noise, or median filtering is used to remove impulse noise, while better preserving image edge details.

[0069] Meanwhile, the acquisition and preliminary parameter processing of real-time operating parameters that accurately characterize the dynamic working characteristics of the compensation device are achieved as follows:

[0070] Sensor units, such as temperature sensors, voltage transformers, current transformers, and humidity sensors, are installed at key measuring points on the compensation device body (such as the surface of capacitors, reactor coils, and near switch contacts) and in the operating environment of the compensation device. These sensor units are used to collect a series of real-time operating parameters that can accurately characterize the dynamic operating characteristics of the compensation device and its operating environment in real time and continuously.

[0071] The collected real-time operating parameters include at least the body temperature parameter, which directly reflects the health status of the core components of the compensation device; the three-phase voltage and three-phase current parameters, which reflect the power supply quality and load conditions of the compensation device; and the ambient temperature and humidity parameters, which affect the operational stability and lifespan of the compensation device. These multi-dimensional operating parameters together constitute a quantitative description of the working status of the compensation device.

[0072] The raw operating parameter data collected may contain noise due to interference during the measurement process or the accuracy limitations of the sensor itself, or the numerical range may vary greatly due to differences in the physical meaning and units of different parameters. Therefore, preliminary parameter processing operations need to be performed on the collected operating parameters.

[0073] These operations include parameter filtering, which aims to remove high-frequency noise and random disturbances from the sequence of operating parameters, extract the true trend of parameter changes, and make the data smoother and more reliable. For example, the Kalman filter algorithm can be used to make optimal estimates of parameters, or an exponentially weighted moving average filter can be used to smooth the data.

[0074] Simultaneously, these operations also include parameter scaling, which aims to transform operating parameters of different physical units and orders of magnitude to similar numerical ranges or a unified scale. This prevents parameters with larger numerical ranges from dominating the analysis results during subsequent multimodal data fusion, and also helps improve the training efficiency and performance of subsequent artificial intelligence models. For example, the min-max normalization method can be used to linearly map the parameters to the interval [0, 1] or [-1, 1], or the Z-score standardization method can be used to transform the parameters into a distribution with a mean of 0 and a standard deviation of 1.

[0075] The synchronous acquisition and preliminary processing of the aforementioned image data and operational parameters provide a high-quality, multi-dimensional data foundation for subsequent AI-based physical defect identification, multimodal data fusion, and health status assessment. Synchronous acquisition ensures the temporal consistency between the physical state reflected in the image data and the operational characteristics characterized by the operational parameters.

[0076] S2. Using a first artificial intelligence model optimized for specific physical defect types of the compensation device, deep feature extraction, analysis, identification and quantification of image data are performed to obtain key feature information of physical defects.

[0077] In this embodiment, step S2 uses a specially optimized first artificial intelligence model to perform deep feature extraction and analysis on the multi-dimensional physical state image data obtained and preprocessed in step S1. Its core objective is to accurately and efficiently identify and quantify various physical defects that may exist on the compensation device, and output their key feature information.

[0078] Specifically, the first artificial intelligence model is a convolutional neural network object detection model. Such models, such as deep learning networks based on the YOLO (You-Only-Look-Once) architecture, Faster-R-CNN architecture, or SSD (Single-Shot-MultiBox-Detector) architecture, integrate components such as convolutional layers, pooling layers, and fully connected layers into their network structure. This allows them to automatically learn and extract complex features from input image data, ranging from low-order to high-order. This model was specifically optimized for training on common, specific types of physical defects found in compensation devices. The training dataset typically contains a large number of precisely labeled images of compensation devices, indicating the type and location of the defects.

[0079] When the multi-dimensional physical state image data (e.g., containing visible light and infrared thermal images) preprocessed in step S1 is fed into the trained first AI model, the model parses the image content using its internal deep feature extraction network. This parsing process identifies visual patterns in the image that are associated with predefined defect categories. The model then outputs a series of key feature information about the identified physical defects, providing direct evidence for subsequent multimodal fusion and risk assessment. These key feature information outputs typically include:

[0080] The classification of physical defects, for example, clearly indicating that the identified defects are predefined specific defect types such as "capacitor casing bulge", "terminal overheating", "insulator surface contamination" or "localized corrosion of the enclosure";

[0081] The precise spatial coordinates of a physical defect in an image are usually given in the form of a bounding box. For example, the region of the defect in the image is precisely defined by the coordinates (x, y) of the top left corner of the bounding box and its width (w) and height (h).

[0082] A confidence score is a numerical value between 0 and 1 that characterizes the reliability of the identification result. It reflects the model's confidence in its judgment (i.e., the presence of this type of defect in the region). A higher confidence score means that the identification result is more reliable.

[0083] And at least one quantitative severity indicator characterizing the extent of the defect’s development or potential impact range, such as, for overheating defects, the magnitude of the temperature exceeding the normal range or the area of ​​the high-temperature region; for deformation defects, the degree or size of the deformation; for dirt or rust, the area or density of the area it covers.

[0084] To improve the accuracy and robustness of the first artificial intelligence model in detecting specific physical defect types in the compensation device, an important aspect of its optimization process is the use of a composite loss function during the model training phase. This guides the iterative adjustment of model internal parameters (such as convolution kernel weights, bias terms, etc.). This composite loss function comprehensively evaluates the model's performance in multiple aspects, including defect localization, target-background differentiation, and defect classification. The composite loss function is calculated as follows:

[0085] ;

[0086] In the formula, The total loss value, or composite loss function, represents the overall difference between the model's predicted output and the human-annotated ground truth labels. The goal of model training is to minimize this total loss value by adjusting the parameters. The localization loss term is represented by its core function of accurately measuring the positional deviation between the physical defect bounding box predicted by the model and the real bounding box annotated by the human. A smaller localization loss means that the model predicts the position more accurately. To locate the loss item The weighting coefficient is a hyperparameter used to balance the contribution of localization loss to the total loss. Its value can be determined based on experience or through experimental tuning to ensure that the loss of different tasks receives reasonable attention. The confidence loss term is used to train the model to distinguish between target regions (foreground) containing physical defects to be detected and background regions that do not contain any defects, and to penalize the model for misclassifying the foreground as background (missed detection) or misclassifying the background as foreground (false detection). The classification loss term measures how well the model's predicted defect category matches the actual defect category for identified physical defect areas, aiming to improve the accuracy of defect type identification.

[0087] To more accurately and comprehensively evaluate and optimize the model's localization performance, the localization loss term... This is specifically implemented through a calculation method called Complete Intersection-over-Union-Loss (CIoU-Loss). Compared to the traditional intersection-over-union ratio (CIoU-Loss), this method offers a more comprehensive approach. )loss, The loss function not only considers the overlap between the predicted and ground truth bounding boxes, but also incorporates considerations of the distance between their center points and the consistency of their aspect ratios. This allows the model to more effectively learn accurate defect localization. The formula is expressed as:

[0088] ;

[0089] In the formula, This is the location loss term. The specific calculation results; Represents the defect bounding box predicted by the model. Compared with manually annotated actual defect boundary boxes The intersection-union ratio between two bounding boxes is equal to the area of ​​their intersection divided by the area of ​​their union. It is the core indicator for measuring the overlap of bounding boxes. Defect bounding boxes predicted by the model The center point and the manually annotated actual defect boundary box The Euclidean distance between the center points of two boxes is used to measure how close the center points of the two boxes are. This represents the bounding box that can completely enclose the defect predicted by the model. Compared with manually annotated actual defect boundary boxes The diagonal length of the minimum bounding rectangle, which is used to normalize the center point distance so that its penalty effect is independent of the scale of the bounding box; It is a positive balancing parameter used to adjust the aspect ratio consistency penalty term. The relative importance of losses;

[0090] It is a defect bounding box used to measure the model's predictions. Compared with manually annotated actual defect boundary boxes The parameter representing the difference in aspect ratio between the two is calculated as follows:

[0091] ;

[0092] In the formula, and These represent the actual defect bounding boxes annotated manually. The width and height, and and These represent the defect bounding boxes predicted by the model, respectively. Width and height. This item Penalizing cases where the aspect ratio of the predicted bounding box is inconsistent with that of the ground truth bounding box helps the model learn to generate bounding boxes with shapes that are closer to the actual defects.

[0093] Balance parameters The calculation method depends on the aspect ratio difference parameter. and intersection Specifically:

[0094] ;

[0095] This calculation method ensures that when the intersection-union ratio between the predicted and ground truth boxes is high (i.e., good overlap), the model pays more attention to aspect ratio consistency. This is achieved by employing a composite loss function and its components. The positioning loss calculation method can effectively improve the accuracy of the first artificial intelligence model in locating physical defects of various compensation devices in complex background environments, its adaptability to defects of different shapes and sizes, and the accuracy of defect classification, thereby providing more reliable physical defect information for subsequent system status assessment.

[0096] S3. Perform collaborative multimodal fusion processing of key feature information and real-time operating parameters to construct a multidimensional fusion state vector that dynamically represents the overall health status of the compensation device;

[0097] In this embodiment, the core task of step S3 is to perform deep collaborative multimodal fusion processing on the key feature information about physical defects output in step S2 and the real-time operating parameters obtained in step S1. The aim is to construct a multidimensional fusion state vector that can comprehensively and dynamically characterize the overall health status of the compensation device. This fusion process lays a solid data foundation for subsequent intelligent risk assessment.

[0098] Specifically, this fusion process includes the following operations:

[0099] First, the key feature information of physical defects identified and quantified in step S2, such as defect category, spatial location, confidence score, and quantified severity index, is structured using a pre-defined encoding rule. This encoding rule aims to transform raw defect information of different types and scales into a unified numerical vector suitable for processing by a neural network model. For example, for categorical information such as defect category, one-hot encoding is used to convert it into a binary vector; for numerical information such as location coordinates, confidence score, and severity index, normalization is performed to ensure they fall within a specific numerical range. After this encoding process, a structured image defect feature vector is formed, denoted as [vector name missing]. This vector This vividly illustrates the physical anomaly of the compensation device as perceived from a visual perspective.

[0100] Simultaneously, the real-time operating parameters that accurately characterize the dynamic working properties of the device, obtained and preliminarily processed in step S1, such as the temperature parameters of the compensation device itself, three-phase voltage parameters, three-phase current parameters, and ambient temperature and humidity parameters, are subjected to feature selection and organization. Feature selection aims to select a subset of parameters most relevant to the health status of the compensation device from among numerous operating parameters, thereby reducing redundant information and improving fusion efficiency. Subsequently, these selected operating parameters, after scale normalization, are organized into a structured operating parameter feature vector, denoted as [vector name missing]. This vector This represents the operating conditions and environmental impact of the compensation device as perceived from the sensor data dimension.

[0101] After obtaining the structured image defect feature vector With running parameter feature vector Subsequently, these two feature vectors, each carrying information from a different modality, are used as joint inputs and fed into a pre-constructed and fully trained multimodal fusion network. The core function of this multimodal fusion network is to learn and reveal the complex dependencies and complementarities between feature information from two different sources.

[0102] In one possible implementation, the network, through its internal feature learning module (e.g., containing several fully connected layers or other types of neural network layers) and an efficient cross-modal information interaction mechanism, processes information derived from visual perception (by... (embodied) and sensor perception (by) This approach involves deep integration, correlation analysis, and complementary enhancement of feature information from two different modalities. In this way, the network can surpass the effects of simple feature splicing, achieving effective information fusion. For example, if a minor physical defect in an image occurs simultaneously with a significant anomaly in the operating parameters, its weight in the fused representation may be adaptively enhanced.

[0103] To achieve this effective cross-modal information interaction and deep feature-level fusion, the multimodal fusion network employs a cross-attention mechanism. This mechanism allows the model to dynamically calculate the relevance or importance weights of one modality's feature information to another during the fusion process, enabling the model to adaptively focus on information combinations that are more indicative of the overall health status assessment of the current compensation device. Its core computational process is described as follows:

[0104] ;

[0105] In the formula, This represents the fused feature representation calculated through this cross-attention mechanism, which integrates the interaction information between features of different modalities; The feature vector representing the input to the query originates from the image defect feature vector. or running parameter feature vector The result is obtained through a corresponding linear transformation and is used to actively explore relevant information in another mode. The feature vector representing the key input originates from and Another set of feature vectors for different modalities (i.e., if) Depend on If generated, then Depend on (Generated, and vice versa) is obtained through a corresponding linear transformation and used for similarity or relevance matching with the query vector; The feature vector representing the value input, and its relationship to the key input. Originating from the same modality and obtained through corresponding linear transformations, it represents the actual information content to be applied by the weights calculated based on query-key relevance; , , These are learnable weight matrices used to linearly project the original query, key, and value feature vectors onto new feature subspaces that are more suitable for attention calculation. The parameters of these weight matrices are learned through backpropagation during the training of the multimodal fusion network to enhance the model's feature extraction and expressive capabilities. This represents the matrix transpose operation; Indicates the weight matrix The dimension of the projected key feature vector is calculated by dividing by the dot product similarity between the query and the key. It is a scaling operation designed to stabilize gradients and prevent the softmax function from entering the saturation region due to excessively large dot product results, thereby contributing to the stability of model training.

[0106] The softmax function converts the calculated raw attention score (i.e., the similarity between the query and each key) into a set of probability weights, representing the importance of each part of the value vector to the current query.

[0107] Through the effective operation of this cross-attention mechanism, the multimodal fusion network can deeply capture and utilize the potential conditional dependencies or synergistic indicative effects between one modality of information (e.g., specific physical defect features observed in an image) and another modality of information (e.g., anomalous changes in specific operating parameters associated with it). This dynamic and selective information integration method enables the network's final output—a multidimensional fused state vector dynamically representing the overall health of the compensation device—to more comprehensively, accurately, and discriminatively reflect the true state of the compensation device, providing high-quality input for subsequent precise risk assessment and early warning decisions.

[0108] S4. Based on the preset dynamic evaluation model, combined with the multi-dimensional fusion state vector, the historical state data of the compensation device and the real-time operating parameters, the abnormal risk of the compensation device is evaluated and judged, and the judgment result is obtained.

[0109] In this embodiment, the core objective of step S4 is to intelligently, multidimensionally, and with high precision assess and determine the current operating status and potential anomaly risks of the compensation device based on the multi-dimensional fusion state vector constructed in step S3, combined with a preset dynamic evaluation model and historical state data. The output of this step is a clear conclusion regarding the current risk level of the compensation device.

[0110] Specifically, this assessment and judgment process unfolds as follows:

[0111] The system first employs a pre-defined dynamic evaluation model. The key idea behind this model is the recognition that the "normal" operating state of the compensation device is not constant but is influenced by various factors such as service life, cumulative operating time, seasonal changes in environmental conditions, and load fluctuations, exhibiting slow, normal time-varying characteristics. Therefore, simply using fixed thresholds to judge anomalies is often inaccurate and inflexible. The dynamic evaluation model targets key parameters in real-time operating parameters (e.g., capacitor temperature, current harmonic content in critical circuits, device power factor, etc.) or specific indicators in the multi-dimensional fused state vector generated in step S3 that comprehensively reflect the device's health status (e.g., one or more scalar values ​​output by the fusion network, designed as health metrics or anomaly scores). Based on accumulated historical state data of the compensation device (e.g., a sequence of corresponding parameters or indicators collected and stored over a long period when the device is confirmed to be in good operating condition) and current operating conditions (e.g., current load level, ambient temperature, and other factors that may affect the parameter baseline), the model dynamically determines a reasonable baseline range for normal operation at the current moment. This baseline range typically includes a dynamically adjusted upper limit and a dynamically adjusted lower limit, or more complexly, a probability distribution range that evolves over time.

[0112] After dynamically determining the baseline range for normal operation, the system compares the current real-time observed values ​​of specific parameters (real-time monitored physical quantities that directly reflect the operating status of the compensation device) or specific indicators (quantitative values ​​that comprehensively reflect the overall health status of the equipment after multimodal fusion processing) with the dynamically determined baseline range. If the current observed value exceeds this dynamic baseline range (e.g., higher than the dynamic upper limit or lower than the dynamic lower limit), it initially indicates that the parameter or indicator may have deviated abnormally.

[0113] However, the final risk assessment does not solely rely on whether a single parameter or indicator exceeds the dynamic baseline. More importantly, the system simultaneously incorporates key feature information about physical defects output from step S2, such as the existence of identified physical defects, their type, severity, and whether their location is critical. This direct physical evidence from the image modality and the indirect operational status representation from the parameter / indicator modality will work together to create a pre-designed and configured risk assessment logic.

[0114] This risk assessment logic aims to comprehensively consider multiple aspects of information to make a final risk level judgment. In one possible implementation, the risk assessment logic is based on a set of expert rules, such as, "Rule 1: If the current temperature of a capacitor exceeds its dynamic baseline upper limit by 5 degrees Celsius, and (AND) a visible light image identifies a 'shell bulge' defect at the location of the capacitor with a severity index greater than threshold A, then the risk level of the compensation device is determined to be 'critical alarm'."

[0115] As a more advanced implementation, the risk assessment logic is also a trained machine learning model, such as a decision tree, support vector machine, or a small neural network classifier. The input to the model is a multi-dimensional fused state vector (which itself contains a fusion of parameter and defect information), or a quantitative representation of independent parameter exceedances and a structured representation of key physical defect features. Its output is the current abnormal risk level of the compensation device, such as predefined levels like "normal," "minor warning," "general alarm," and "serious alarm."

[0116] This includes the process of dynamically determining the baseline range for normal operation of a specific parameter or indicator, with the upper limit of this baseline range being... The calculation is specifically implemented through a statistical method that combines Exponentially-Weighted Moving Average (EWMA) with historical data volatility. The purpose of this method is to ensure that the established baseline smoothly adapts to the inherent, slow, normal trend of the parameter or indicator, while maintaining sufficient sensitivity to sudden, abnormal fluctuations that may indicate a failure. The specific calculation method is as follows:

[0117] ;

[0118] In the formula, This represents a specific parameter or indicator at the current moment of observation. The calculated upper limit of the dynamic baseline; This represents a specific parameter or indicator at the current moment of observation. The exponentially weighted moving average. This average reflects the recent smoothed trend of a parameter or indicator. Its calculation and update method is typically as follows:

[0119] ;

[0120] In the formula, For a specific parameter or indicator at the current time of observation The actual observed value; This is a preset smoothing factor, with values ​​ranging from 0 to 1 (for example, a common range is 0.1 to 0.3). This factor is used to adjust the current observation relative to the historical average when updating the current exponentially weighted moving average. The proportion of the smaller A higher value means that historical data has a greater weight and the average value changes more smoothly; conversely, a lower value means that the current value changes more quickly. For a specific parameter or indicator at the immediately preceding time of observation The calculated exponentially weighted moving average.

[0121] This is a preset standard deviation factor, a positive real number, which can be selected as 2, 2.5, or 3, depending on the specific application scenario and the tolerance for false positives and false negatives. This factor is used to set the upper limit of the dynamic baseline deviation from the smoothed mean, and its value directly affects the balance between the sensitivity and specificity of anomaly detection. The historical standard deviation is the statistically calculated value of a series of historical observations (usually data from a period when the equipment is in a confirmed healthy and stable operating condition) of a specific parameter or indicator extracted from historical state data. This standard deviation reflects the inherent fluctuation or dispersion of the parameter or indicator under normal operating conditions.

[0122] The dynamic baseline upper limit is calculated using this method, which combines exponentially weighted moving averages and historical volatility. It can adaptively track normal changes in parameters or indicators and set reasonable warning lines based on their historical behavior, thereby helping to improve the accuracy and reliability of anomaly risk assessment and reduce misjudgments or omissions caused by fixed thresholds. The calculation of the lower limit also uses a similar method. Finally, by comprehensively comparing the current state with the dynamic baseline and integrating physical defect information, the system can more accurately and intelligently assess and determine the current and potential anomaly risks of the compensation device.

[0123] S5. Match the judgment result with the preset adaptive alarm threshold and trigger an alarm signal with hierarchical or diagnostic indication information.

[0124] In this embodiment, the core of step S5 is that, based on the results of the intelligent assessment and accurate judgment completed in step S4, when the current or potential abnormal risk of the compensation device reaches the preset adaptive alarm threshold, the corresponding alarm signal can be automatically and timely triggered, and under specific conditions, in-depth diagnostic analysis can be further initiated to provide more detailed fault analysis and operation and maintenance guidance.

[0125] Specifically, when the abnormal risk assessment result of the compensation device output in step S4—such as a quantified risk score or a specific risk level—reaches or exceeds one or more preset adaptive alarm thresholds, the system will automatically trigger an alarm mechanism. The adaptive alarm thresholds are preset based on factors such as the compensation device's model, importance, operating environment, and historical fault data, and are designed as multi-level thresholds to correspond to different levels of risk severity. For example, "attention," "warning," and "critical" level thresholds are set. The system will trigger an alarm of the corresponding level based on which level of threshold the currently assessed risk exceeds.

[0126] The alarm signals that are triggered are not just simple prompts, but also contain hierarchical or diagnostic indications.

[0127] "Classification" is reflected in the fact that the intensity, presentation method, or notification scope of alarms may vary depending on the risk level. For example, lower-level risks may only display a prompt message on the local monitoring interface, while higher-level risks may be accompanied by audible and visual alarms, send alarm information to the superior dispatch system, or notify relevant maintenance personnel through mobile communication networks.

[0128] "Diagnostic indication information" means that the alarm signal carries preliminary indicative information about the abnormal condition. For example, it may indicate which part of the parameter is out of limit, or what type of physical defect has been identified, or a comprehensive health status summary.

[0129] In a more in-depth approach to fault diagnosis and decision support, if the results of the intelligent assessment and accurate judgment in step S4 indicate the presence of a pre-set complex anomaly risk (e.g., a risk involving multiple component-related faults, or a risk scenario where the cause is difficult to determine directly through simple rules) or a pre-set key anomaly risk (e.g., a high-risk risk that may lead to serious equipment damage or system shutdown), the system will further execute a deep diagnostic analysis process.

[0130] In this case, the system will submit a more comprehensive set of information to a pre-defined second artificial intelligence model. This information set specifically includes:

[0131] The multi-dimensional physical state image data acquired and preliminarily processed in step S1, such as visible light images and infrared thermal images, provides direct evidence of the device's physical appearance and thermal condition.

[0132] The real-time operating parameters acquired and preliminarily processed in step S1, such as temperature, voltage, current, and humidity, reflect the dynamic operating characteristics of the equipment and environmental conditions.

[0133] The key feature information of the physical defects identified and quantified in step S2, such as the type, location, and severity of the defects, indicates the specific physical damage point.

[0134] The multidimensional fusion state vector constructed in step S3, as the product of deep fusion of the aforementioned image information and parameter information, comprehensively characterizes the overall health status of the compensation device.

[0135] The second artificial intelligence model is preferably a multimodal large language model. This model has the ability to process and understand data from different modalities (e.g., images, text, numerical sequences). Crucially, this multimodal large language model has been pre-tuned and optimized using specialized knowledge related to the operation, maintenance, and fault diagnosis of the compensation device. This fine-tuning process allows the model to learn from a large amount of professional literature, historical fault cases, equipment manuals, and operation and maintenance procedures in this field, enabling it to perform domain-specific deep understanding and reasoning on the input compensation device status data.

[0136] After receiving the comprehensive information including images, parameters, defect features, and fusion vectors, the specially fine-tuned multimodal large language model will perform in-depth diagnostic analysis. The analysis results will be output in the form of a structured diagnostic report.

[0137] The diagnostic report is very comprehensive, containing at least one or more of the following:

[0138] A detailed analysis of the causes of the identified complex or critical anomalies; for example, the model may infer the most likely root cause event, fault propagation path, or interaction between multiple factors based on the input multimodal data.

[0139] Specific operation and maintenance guidance suggestions, such as the model may provide recommended inspection steps, maintenance plans, spare parts replacement suggestions based on the analyzed causes of failures and equipment status, and even predict the possible consequences if the failure is not handled in time, as well as corresponding emergency response plans.

[0140] This diagnostic report will then be integrated into the final alarm information as part of the diagnostic indication information, or presented to relevant personnel through a dedicated maintenance support interface. In this way, maintenance personnel can not only be promptly informed of equipment anomalies, but also receive in-depth analysis results from artificial intelligence models and actionable decision support, thereby significantly improving the efficiency and accuracy of fault handling and ensuring the safe and stable operation of the compensation device.

[0141] Please see the appendix Figure 2 The present invention also provides a status monitoring system for a compensation device, the system comprising:

[0142] The data acquisition unit 201 is used to synchronously acquire image data of the multi-dimensional physical state of the compensation device and real-time operating parameters characterizing its dynamic working characteristics through a pre-configured image acquisition device and sensor unit, and to perform preprocessing.

[0143] The data acquisition unit comprehensively and synchronously captures raw information reflecting the status of the compensation device. This unit is internally equipped with selected and strategically placed image acquisition devices and a series of sensor units.

[0144] On the one hand, image acquisition devices, such as high-resolution visible light industrial cameras and infrared thermal imagers capable of sensing temperature distribution, capture images of key parts and the overall appearance of the compensation device according to preset strategies (e.g., periodic scanning or event-triggered snapshots). This yields multi-dimensional physical state image data, including the device's surface structure, connection status, presence of obvious deformation, color anomalies, and temperature field distribution of various components. This image data provides direct visual evidence for subsequent physical defect identification.

[0145] On the other hand, sensor units, such as temperature sensors, voltage transformers, current transformers, and humidity sensors installed on the compensation device itself or in its operating environment, continuously and in real time acquire a series of real-time operating parameters that accurately characterize the dynamic operating characteristics of the compensation device, such as the temperature of key components, the three-phase voltage and current values ​​of input and output, power factor, harmonic content, and the temperature and humidity of the environment in which the device is located. These parameter data dynamically reflect the operating conditions of the compensation device and the influence of the external environment. "Synchronous acquisition" is an important feature of this unit, ensuring that the acquired image data can accurately correspond to the operating parameters at the same moment, laying the foundation for the effective fusion and correlation analysis of subsequent multimodal information.

[0146] The defect identification unit 202 is used to extract, parse, identify and quantify deep features of image data through a first artificial intelligence model optimized for specific physical defect types of the compensation device, so as to obtain key feature information of physical defects.

[0147] The defect identification unit's core function is to automatically and intelligently identify and quantify various physical defects from multi-dimensional physical state image data collected by the data acquisition unit using advanced artificial intelligence technology. This unit is equipped with a primary artificial intelligence model, typically a deep learning model specifically trained and optimized for specific types of physical defects that may occur in the compensation device (such as casing bulging, oil leakage, connection point corrosion, insulator damage, abnormal heating points, etc.), with convolutional neural network object detection models being a prime example.

[0148] Upon receiving image data, the defect identification unit utilizes this first artificial intelligence model for deep feature extraction and analysis. This process involves the model performing multi-level abstraction and understanding of image information, automatically learning and recognizing visual patterns and texture features associated with various predefined defects. After analysis, the unit can output key feature information about the identified physical defects. This information typically includes: the defect's category (clearly identifying the type of defect), the defect's precise spatial coordinates in the image (e.g., indicated by a bounding box), the model's confidence level in the identification result (i.e., confidence score), and one or more quantitative severity indicators characterizing the degree of defect development or potential impact range (e.g., the temperature value of an overheated spot, the length of a crack, the area of ​​a corroded region, etc.). This structured defect information is a crucial input for subsequent risk assessment.

[0149] The data fusion unit 203 is used to perform collaborative multimodal fusion processing of key feature information and real-time operating parameters to construct a multidimensional fusion state vector that dynamically represents the overall health status of the compensation device.

[0150] The data fusion unit's main task is to effectively perform collaborative multimodal fusion processing on information from different sensing sources—namely, the key feature information about physical defects output by the defect identification unit and the real-time operating parameters acquired by the data acquisition unit. Its ultimate goal is to construct a multidimensional fused state vector that can comprehensively and dynamically characterize the overall health status of the compensation device.

[0151] In this unit, necessary preprocessing and feature engineering are first performed on the data from two different modalities. For example, structured image defect features (such as defect category, location, and severity) and pre-processed operational parameter features (such as normalized values ​​of temperature, voltage, and current) are organized into a unified vector representation. These feature vectors are then fed as joint input into a pre-built and trained multimodal fusion network. This network, through its specifically designed internal structure, including modules incorporating attention mechanisms (such as cross-attention), learns and captures the intrinsic correlations and complementarities between features from different modalities. It is not merely a simple data concatenation, but rather achieves a "1+1>2" effect through deep information interaction and integration. For instance, a minor physical defect accompanied by significant operational parameter anomalies may have its weight amplified in the fusion state. Finally, this unit outputs a highly condensed multidimensional fusion state vector, which aims to comprehensively reflect the current overall health level of the compensation device in a unified form.

[0152] Risk assessment unit 204 is used to assess and determine the abnormal risks of the compensation device based on a preset dynamic assessment model, combined with multi-dimensional fusion state vectors, historical state data of the compensation device and real-time operating parameters, and obtain the judgment result.

[0153] The risk assessment unit is positioned to intelligently and accurately assess and determine the current abnormal risks and potential future abnormal risks of the compensation device based on the multi-dimensional fusion state vector constructed by the data fusion unit, combined with a preset dynamic assessment model and the historical state data accumulated by the system.

[0154] The dynamic evaluation model employed by this unit takes into account the non-stationary characteristics of the compensation device's operating state. For example, it dynamically adjusts the baseline or threshold used to determine anomalies based on historical data (such as the range and trend of parameter fluctuations during normal operation) and current operating conditions (such as load and environmental factors). Upon receiving the latest multidimensional fused state vector, the risk assessment unit compares it with the dynamic baseline and performs a comprehensive analysis using pre-defined risk assessment logic or models (such as rule-based systems, machine learning classifiers, or regression models). This analysis process considers not only deviations from single indicators but also the combined impact of multidimensional information, thereby enabling more accurate identification of early signs of failure or complex failure modes. Ultimately, the unit outputs a clear assessment result of the current and potential anomaly risks of the compensation device, such as a specific risk score, a predefined risk level (e.g., "Normal," "Attention," "Warning," "Critical"), or a probability estimate of a specific failure mode.

[0155] The alarm triggering unit 205 is used to match the judgment result with the preset adaptive alarm threshold and trigger an alarm signal with hierarchical or diagnostic indication information.

[0156] As the final action output link of the monitoring system, the alarm triggering unit's core responsibility is to automatically trigger the corresponding alarm signal in a timely and accurate manner when the abnormal risk of the compensation device reaches the preset adaptive alarm threshold, based on the intelligent assessment and accurate judgment results output by the risk assessment unit.

[0157] This unit maintains one or more sets of adaptive alarm thresholds. These thresholds are not fixed but are dynamically adjusted based on the results of the risk assessment model, historical experience, and operational strategies. When the risk level or risk score determined by the risk assessment unit exceeds the corresponding threshold, the alarm triggering unit will be activated.

[0158] The triggered alarm signals have several key characteristics: First, they are "tiered," meaning that the form and intensity of the alarm vary depending on the severity of the risk, ranging from simple interface prompts to audible and visual alarms, and even automatic notifications to relevant maintenance personnel or higher-level management systems. Second, they include "diagnostic indication information," meaning that the alarm signal is not merely a general warning, but also provides preliminary diagnostic clues about the abnormal risk, such as indicating which key parameters have exceeded limits, identifying the type of physical defect, or revealing specific abnormal patterns through multi-dimensional fusion state vectors.

[0159] In some embodiments, if the evaluation results indicate the presence of a pre-defined complex or critical anomaly risk, the alarm triggering unit may also link with a more advanced diagnostic module. For example, it may submit more comprehensive data (including original images, parameters, defect information, and fusion vectors) to a multimodal large language model fine-tuned with domain-specific knowledge of the compensation device for in-depth diagnostic analysis. The detailed causal analysis or specific operation and maintenance guidance suggestions generated by the analysis may be used as part of the diagnostic indication information, thereby further enhancing the practical value and guiding significance of the alarm.

[0160] In the embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0161] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0162] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0163] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0164] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for monitoring the state of a compensation device, characterized in that, include: The image data of the multi-dimensional physical state of the compensation device and the real-time operating parameters characterizing its dynamic working characteristics are acquired synchronously through the pre-configured image acquisition device and sensor unit and pre-processed. By using a first artificial intelligence model optimized for a specific type of physical defect in the compensation device, deep feature extraction, parsing, identification, and quantification are performed on the image data to obtain key feature information of the physical defects. The key feature information and the real-time operating parameters are subjected to collaborative multimodal fusion processing to construct a multidimensional fusion state vector that dynamically represents the overall health status of the compensation device; Based on the preset dynamic evaluation model, combined with the multi-dimensional fusion state vector, the historical state data of the compensation device, and the real-time operating parameters, the abnormal risk of the compensation device is evaluated and determined, and the determination result is obtained. The determination result is matched with a preset adaptive alarm threshold to trigger an alarm signal with hierarchical or diagnostic indication information.

2. The method for monitoring the state of the compensation device according to claim 1, characterized in that, The process involves synchronously acquiring image data of the multi-dimensional physical state of the compensation device and real-time operating parameters characterizing its dynamic working characteristics through a pre-configured image acquisition device and sensor unit, and then preprocessing the data, including: The image acquisition device acquires multi-dimensional physical state image data, including visible light image data and infrared thermal image data, according to a preset periodic acquisition strategy or in response to a preset event triggering condition, and performs image enhancement and image denoising operations on the image data. The sensor unit collects real-time operating parameters, including the temperature parameters of the compensation device body, three-phase voltage parameters, three-phase current parameters, and the temperature and humidity parameters of the environment in which the compensation device is located. It then performs parameter filtering and parameter scaling operations on the real-time operating parameters.

3. The method for monitoring the state of the compensation device according to claim 1, characterized in that, The first artificial intelligence model is: a convolutional neural network object detection model; The trained convolutional neural network target detection model is used to identify and output the category of physical defects, spatial location coordinates, confidence score representing the reliability of the identification result, and quantitative severity index representing the degree of development or scope of influence of the defect from the image data.

4. The method for monitoring the state of the compensation device according to claim 3, characterized in that, The convolutional neural network object detection model, during the training phase, includes the following processing: A composite loss function is used to guide the adjustment of model parameters. The composite loss function is calculated as follows: ； In the formula, The total loss value used to evaluate the overall difference between the model predictions and the ground truth labels is the composite loss function. Represents the positioning loss term; For the positioning loss term Weighting coefficients; Represents the confidence level loss term; Represents the category loss term; Among them, the positioning loss term The formula for calculating it is expressed as follows: ； In the formula, For the positioning loss term The calculation results; Represents the defect bounding box predicted by the model. Compared with manually annotated actual defect boundary boxes The crossover ratio between them; Represents the defect bounding box predicted by the model. The center point and the manually annotated actual defect boundary box The Euclidean distance between the center points; This represents the bounding box that can completely enclose the defect predicted by the model. Compared with manually annotated actual defect boundary boxes The length of the diagonal of the smallest bounding rectangle; It is a positive equilibrium parameter; It is used to measure the defect bounding box predicted by the model. Compared with manually annotated actual defect boundary boxes The parameter representing the difference in aspect ratio between the two.

5. The method for monitoring the state of the compensation device according to claim 1, characterized in that, The step of performing collaborative multimodal fusion processing of the key feature information and the real-time operating parameters to construct a multidimensional fusion state vector that dynamically represents the overall health status of the compensation device includes: The key feature information is processed through preset encoding rules to form a structured image defect feature vector, which is then denoted as the image defect feature vector. The real-time operating parameters are selected and organized to form a structured operating parameter feature vector, which is denoted as the operating parameter feature vector. The image defect feature vector and the operating parameter feature vector are fed together as joint inputs to a pre-constructed and trained multimodal fusion network. The multimodal fusion network integrates, correlates and complements the image defect feature vector and the operating parameter feature vector through its internally designed feature learning module and cross-modal information interaction mechanism, generating a multidimensional fusion state vector that reflects the current overall health status of the compensation device.

6. The method for monitoring the state of the compensation device according to claim 5, characterized in that, The multimodal fusion network's processing of the image defect feature vector and the operating parameter feature vector includes: A deep feature-level fusion is performed between the defect feature vector and the operating parameter feature vector through a cross-attention mechanism; wherein, the calculation process of the cross-attention mechanism includes: dynamically calculating the correlation weights between different modal features, so that the multimodal fusion network adaptively identifies information combinations that are more indicative of the current health status assessment of the compensation device; The expression for the calculation process of the cross-attention mechanism is as follows: ； In the formula, This represents the fused feature representation calculated through the cross-attention mechanism; The feature vector that serves as the input for the query; The feature vector that serves as the key input; The feature vector represents the input value. , , These are the learnable weight matrices; This represents the matrix transpose operation; Indicates the process The dimension of the projected key feature vector.

7. The method for monitoring the state of the compensation device according to claim 1, characterized in that, The method, based on a preset dynamic evaluation model and combining the multi-dimensional fused state vector, the historical state data of the compensation device, and the real-time operating parameters, assesses and determines the abnormal risk of the compensation device, including: Based on a preset dynamic evaluation model, specific parameters are selected from the real-time operating parameters or specific indicators are selected from the multi-dimensional fused state vector. The baseline range for normal operation of the specific parameters or specific indicators is determined based on historical state data and current operating conditions. Based on a preset risk assessment logic, the current value of the specific parameter or the specific indicator is compared with the baseline range, and the current abnormal risk level of the compensation device is determined by combining the key feature information.

8. The method for monitoring the state of the compensation device according to claim 7, characterized in that, The upper limit of the baseline range The calculation method is as follows: ； In the formula, This represents the specific parameter or indicator at the current time of observation. The upper limit of the dynamic baseline; This represents the specific parameter or indicator at the current time of observation. Exponentially weighted moving average; It is a preset standard deviation factor; This represents a series of historical observations of the specific parameter or the specific indicator extracted from the historical state data.

9. The method for monitoring the state of the compensation device according to claim 2, characterized in that, The step of matching the determination result with a preset adaptive alarm threshold to trigger an alarm signal with hierarchical or diagnostic indication information includes: When the determination result matches a preset complex anomaly risk or a preset key anomaly risk, the multi-dimensional physical state image data, the real-time operating parameters, the key feature information of the physical defect, and the multi-dimensional fused state vector are input into a preset second artificial intelligence model; the second artificial intelligence model is then fine-tuned and optimized by the compensation device therein before outputting a diagnostic report. The diagnostic report includes operational guidance suggestions for the identified abnormal risks and serves as a component of the diagnostic indication information; the second artificial intelligence model is a multimodal large language model.

10. A status monitoring system for a compensation device, characterized in that, include: The data acquisition unit is used to synchronously acquire image data of the multi-dimensional physical state of the compensation device and real-time operating parameters characterizing its dynamic working characteristics through a pre-configured image acquisition device and sensor unit, and to perform preprocessing. The defect identification unit is used to extract, parse, identify and quantify the image data using a first artificial intelligence model optimized for a specific type of physical defect in the compensation device, so as to obtain key feature information of the physical defect. The data fusion unit is used to perform collaborative multimodal fusion processing on the key feature information and the real-time operating parameters to construct a multidimensional fusion state vector that dynamically represents the overall health status of the compensation device. The risk assessment unit is used to assess and determine the abnormal risks of the compensation device based on a preset dynamic assessment model, combined with the multi-dimensional fusion state vector, the historical state data of the compensation device, and the real-time operating parameters, and to obtain the determination result. An alarm triggering unit is used to match the determination result with a preset adaptive alarm threshold and trigger an alarm signal with hierarchical or diagnostic indication information.

Images