Transformer health prediction system based on multi-dimensional monitoring data fusion

Abstract

The present invention discloses a transformer health prediction system under multi-dimensional monitoring data fusion, which relates to the field of power engineering technology. The system includes: constructing a global prediction model configured on a cloud platform according to a multi-dimensional initial monitoring indicator set. Performing distillation and dimensionality reduction on the global prediction model, creating a simplified prediction model, and distributing it to an edge prediction device, the simplified prediction model includes multiple hierarchical simplified prediction models. Activating a multi-source monitoring device, performing normal multi-source monitoring of a target transformer, and obtaining a real-time multi-source monitoring data set. The real-time data set is transmitted to the edge prediction device for binary risk discrimination, and the global prediction model is called according to the discrimination result to perform full-dimensional positioning prediction, and a health prediction result is output. Thereby achieving the technical effect of improving the prediction responsiveness and accuracy and reducing the prediction cost.

Description

Technical Field

[0001] The present invention relates to the technical field of electric power engineering, and in particular to a transformer health prediction system under multi-dimensional monitoring data fusion. Background Art

[0002] The health of the transformer is crucial to the stability of the entire power system. Existing transformer monitoring technology mainly relies on independent monitoring data from multiple locations, which makes it difficult to fully analyze the health of the transformer. The amount of data to be analyzed is large and the computational burden is heavy. There are technical problems such as high prediction delay, high prediction cost, and low prediction accuracy. Summary of the invention

[0003] The present invention provides a transformer health prediction system under multi-dimensional monitoring data fusion to solve the technical problems of high prediction delay, high prediction cost and low prediction accuracy in the prior art, and achieves the technical effect of improving prediction responsiveness and accuracy and reducing prediction cost.

[0004] The transformer health prediction system under multi-dimensional monitoring data fusion provided by the present invention includes:

[0005] A global model configuration module, wherein the global model configuration module is used to construct a global prediction model based on a multi-dimensional initial monitoring indicator set, wherein the global prediction model is configured on a cloud platform.

[0006] A simplified model construction module, wherein the simplified model construction module is used to perform distillation and dimensionality reduction on the global prediction model, construct a simplified prediction model, and distribute the simplified prediction model to an edge prediction device, wherein the simplified prediction model includes a plurality of hierarchical simplified prediction models.

[0007] The monitoring data acquisition module is used to activate the multi-source monitoring equipment, perform normal multi-source monitoring and acquisition of the target transformer, and obtain a real-time multi-source monitoring data set.

[0008] A step-by-step prediction module is used to transmit the real-time multi-source monitoring data set to the edge prediction device, perform binary risk discrimination on the target transformer, and based on the binary risk discrimination result, call the global prediction model to perform full-dimensional positioning prediction and output a health prediction result.

[0009] In a feasible implementation, before building a global prediction model based on a multi-dimensional initial monitoring indicator set, the execution steps include:

[0010] Determine the multi-source monitoring points, combine them with the target scene monitoring requirements, and define the multi-source monitoring paths.

[0011] Based on the multi-source monitoring path, a multi-dimensional initial monitoring indicator set is extracted.

[0012] The transformer identity feature of the target transformer is obtained, and historical health abnormality record data is obtained based on the transformer identity feature, wherein the historical health abnormality record data includes intrinsic health abnormality record data and homologous health abnormality record data.

[0013] Based on the multi-dimensional initial monitoring indicator set, the historical health abnormality record data is lightened, and the lightened result is output as a standard abnormal sample data set.

[0014] The global prediction model is constructed and trained according to the standard abnormal sample data set, wherein the global prediction model includes a plurality of local health prediction clusters.

[0015] In a feasible implementation, before performing distillation and dimensionality reduction on the global prediction model and constructing a simplified prediction model, the steps include:

[0016] Edge device configuration information of multiple edge prediction devices is obtained, and a performance configuration data set is extracted, wherein the performance configuration data set corresponds to multiple types of edge prediction devices.

[0017] A computing power evaluation is performed based on the performance configuration data set, and a multi-level model scale constraint is defined based on the computing power evaluation result.

[0018] In a feasible implementation, computing power evaluation is performed according to the performance configuration data set, and multi-level model scale constraints are defined in combination with the computing power evaluation results. The execution steps include:

[0019] In an interactive target scene, the plurality of types of edge prediction devices are used as statistical dimensions to obtain the distribution of the number of edge prediction devices in the target scene.

[0020] Based on the preset group distribution constraint, the number distribution of the edge prediction devices is cumulatively divided into groups along the direction of increasing computing power. Whenever the cumulative number of devices of N types of edge prediction devices accounts for more than the group distribution constraint, a group division is performed, and multiple group division results are output.

[0021] According to the group division results multiple times, the lower limit of the computing power within the group is extracted and output as a multi-level computing power target, and the multi-level model scale constraint is defined based on the multi-level computing power target.

[0022] In a feasible implementation, the global prediction model is subjected to distillation and dimensionality reduction to construct a simplified prediction model, and the execution steps include:

[0023] Perform computing power matching and establish a scale mapping table based on the computing power evaluation result and the scale constraint of the multi-level model.

[0024] A plurality of the hierarchical simplified prediction models are constructed, and according to the scale mapping table, the hierarchical simplified prediction models are called to obtain a target hierarchical simplified prediction model.

[0025] Based on the multi-dimensional initial monitoring indicator set, the simplified prediction model is subjected to functional branch pruning, and enhanced training is performed on the branch pruning prediction model to obtain an edge prediction model.

[0026] Based on the scale mapping table, the edge prediction model is transmitted to the edge prediction device.

[0027] In a feasible implementation, a plurality of the hierarchical simplified prediction models are constructed, and the execution step further includes:

[0028] According to the multi-level model scale constraint, a plurality of standard simplified prediction models are defined, wherein the model scales and structures of the plurality of standard simplified prediction models are different.

[0029] The standard abnormal sample data set is softened based on the global prediction model to obtain a softened prediction label set.

[0030] The softened prediction label set is binarized and simplified to obtain a global prediction label set, wherein the global prediction label set includes a binarized prediction result and a prediction probability label.

[0031] A plurality of the standard simplified prediction models are trained with the global prediction label set and the standard abnormal sample data set to obtain a plurality of the hierarchical simplified prediction models.

[0032] In a feasible implementation, the global prediction model is constructed and trained according to the standard abnormal sample data set, and the execution steps include:

[0033] A plurality of classified abnormal sample data sets are obtained, wherein the plurality of classified abnormal sample data sets are clustering results of clustering the standard abnormal sample data set according to healthy abnormal categories.

[0034] Traversing a plurality of the classified abnormal sample data sets, based on an ensemble learning method, constructing and training a plurality of local health prediction clusters, wherein the plurality of local health prediction clusters are specifically adapted to a plurality of types of health abnormalities.

[0035] In a feasible implementation, before training a plurality of the standard simplified prediction models with the global prediction label set and the standard abnormal sample data set, the execution steps include:

[0036] According to the multi-stage model scale constraint, a multi-stage distillation coefficient is calculated, wherein the multi-stage distillation coefficient is a scale ratio of the multi-stage model scale constraint to the global prediction model.

[0037] The multi-level distillation coefficient is used as the fusion weight to define a multi-level loss function, which is:

[0038] ;

[0039] in, Characterizing the multi-level loss function; Characterizing the fusion weight; Characterizing the loss of the standard simplified prediction model training based on the standard abnormal sample data set; Characterize the loss in the training of the standard simplified prediction model based on the global prediction label set.

[0040] The present invention discloses a transformer health prediction system under multi-dimensional monitoring data fusion, including: constructing a global prediction model configured on a cloud platform according to a multi-dimensional initial monitoring indicator set; performing knowledge distillation and dimensionality reduction processing on the global prediction model to generate a simplified prediction model, and distributing it to an edge prediction device, wherein the simplified prediction model includes multiple hierarchical simplified prediction models; activating a multi-source monitoring device to perform normalized multi-source monitoring data collection on a target transformer to obtain a real-time multi-source monitoring data set; transmitting the real-time multi-source monitoring data set to an edge prediction device, performing binary risk discrimination of the target transformer, and calling a global prediction model to perform full-dimensional positioning and health status prediction according to the binary risk discrimination result, and outputting the final health prediction result. The transformer health prediction system under multi-dimensional monitoring data fusion disclosed by the present invention solves the technical problems of high prediction delay, high prediction cost, and low prediction accuracy, and achieves the technical effect of improving prediction responsiveness and accuracy and reducing prediction cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a structural schematic diagram of the transformer health prediction system under multi-dimensional monitoring data fusion of the present invention;

[0042] Figure 2 The present invention is a flowchart of constructing a global prediction model in a transformer health prediction system under multi-dimensional monitoring data fusion.

[0043] Explanation of the accompanying drawings: global model configuration module 11, simplified model construction module 12 monitoring data collection module 13, step-by-step prediction module 14. DETAILED DESCRIPTION

[0044] The technical solution provided in the embodiments of the present invention is to solve the technical problems of high prediction delay, high prediction cost and low prediction accuracy in the prior art. The overall idea adopted is as follows:

[0045] Firstly, a global prediction model is constructed based on the multi-dimensional initial monitoring indicator set and configured on the cloud platform. Secondly, the global prediction model is distilled and dimensionally reduced to construct a simplified prediction model, and multiple hierarchical simplified prediction models are distributed to the edge prediction device. Then, the multi-source monitoring device is activated to perform normal multi-source monitoring and collection of the target transformer to obtain a real-time multi-source monitoring data set. Next, the real-time multi-source monitoring data set is transmitted to the edge prediction device to perform binary risk discrimination of the target transformer. Then, based on the binary risk discrimination result, the global prediction model is called to perform full-dimensional positioning prediction. Finally, the health prediction result of the target transformer is output.

[0046] The above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods of the specification to better understand the above technical solution. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments of the present invention. It should be understood that the present invention is not limited to the example embodiments used only to explain the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. In addition, it should be noted that, for the convenience of description, only the parts related to the present invention are shown in the drawings, rather than all of them.

[0047] Example

[0048] Figure 1 The figure is a flow chart of a transformer health prediction system under multi-dimensional monitoring data fusion of the present invention, wherein the system comprises:

[0049] The global model configuration module 11 is used to construct a global prediction model according to a multi-dimensional initial monitoring indicator set, wherein the global prediction model is configured on a cloud platform.

[0050] Specifically, the health prediction of the transformer is based on the monitoring and collection of a multi-dimensional initial monitoring indicator set. The multi-dimensional initial monitoring indicator set refers to multiple data categories collected through multiple sensors and monitoring devices, covering various physical and chemical parameters of transformer operation, and is used to comprehensively describe the operating status, fault signs and health status of the transformer.

[0051] In other words, the multidimensional initial monitoring indicator set represents different information dimensions of transformer operation. By acquiring information from these dimensions, a complete multidimensional initial monitoring indicator set can be formed, thereby defining the basic data input for the construction of the global prediction model. Exemplarily, the multidimensional initial monitoring indicator set may include electrical indicators (such as partial discharge intensity, high-frequency current signal, ground current, etc.), mechanical indicators (such as vibration characteristics, sound print characteristics, etc.), thermal indicators (such as temperature data, thermal imaging data, etc.), chemical indicators (such as oil chromatograms, spectrograms), etc.

[0052] Specifically, the global prediction model is a complex model built based on a multi-dimensional initial monitoring indicator set to predict the health status and potential faults of the transformer. The model can fuse data from different sensors to provide a global prediction of the transformer status and fault warning.

[0053] Specifically, the global prediction model is deployed on the cloud platform to process the massive data generated by transformers in different locations through the powerful computing resources and remote processing characteristics of cloud computing, and provide real-time prediction and health analysis. It has strong processing capabilities, flexible deployment, convenient model updates and upgrades, and reduces local deployment costs.

[0054] The above steps integrate the multi-dimensional initial monitoring indicator set to build the corresponding global prediction model, and deploy it on the cloud platform. Through the computing and storage capabilities of the cloud platform, it is easy to expand to more transformer equipment and realize centralized management and health prediction of multiple transformers.

[0055] In some embodiments, Figure 2 As shown, based on the multi-dimensional initial monitoring indicator set, a global prediction model is constructed, and the execution steps include:

[0056] Determine the multi-source monitoring points, and define the multi-source monitoring path in combination with the target scenario monitoring requirements. Based on the multi-source monitoring path, extract the multi-dimensional initial monitoring indicator set. Obtain the transformer identity characteristics of the target transformer, and obtain historical health abnormality record data based on the transformer identity characteristics, wherein the historical health abnormality record data includes intrinsic health abnormality record data and homologous health abnormality record data. Based on the multi-dimensional initial monitoring indicator set, perform data lightweighting on the historical health abnormality record data, and output the lightweighting result as a standard abnormal sample data set. Based on the standard abnormal sample data set, construct and train the global prediction model, wherein the global prediction model includes multiple local health prediction clusters.

[0057] Specifically, multi-source monitoring points refer to multiple sensors and monitoring devices arranged on the transformer and its related equipment, covering different dimensions of transformer health monitoring. Exemplary monitoring points include: transformer casing, where infrared sensors can be installed to monitor surface temperature. Near the windings, vibration sensors and UHF sensors can be set to monitor mechanical vibration and partial discharge. At the oil tank location, oil chromatography monitoring equipment can be installed to analyze the composition of dissolved gases in the oil. Grounding system, grounding current sensors can be installed to monitor abnormal changes in grounding current.

[0058] Specifically, the monitoring paths of each monitoring point are determined according to the monitoring requirements of the target scenario, that is, according to different application scenarios (such as monitoring of key transformers in power grids, transformer health management in industrial scenarios, etc.). Among them, the multi-source monitoring path refers to the monitoring method or monitoring method for collecting monitoring data from different monitoring points. By determining the multi-source monitoring path, it is ensured that the monitoring data can fully cover the health status of the transformer, and combined with different monitoring points, collection frequencies, and data types, the efficient transmission and real-time performance of the data are guaranteed to support subsequent health prediction analysis.

[0059] Exemplarily, the multi-source monitoring path may include: a partial discharge monitoring path, including ultrasonic partial discharge sensors, ultra-high frequency partial discharge sensors, high frequency partial discharge sensors, radio frequency partial discharge sensors, etc., which are used for partial discharge monitoring and are arranged in the insulation system or surrounding area of ​​the transformer. An infrared or temperature monitoring path, which is used for overheating monitoring, is usually placed in the casing or high-temperature area of ​​the transformer to monitor overheating points. A vibration monitoring path, which is installed in the mechanical structure of the transformer to monitor mechanical failures or structural looseness. An oil chromatography monitoring path, which is used for gas analysis of transformer oil and is placed outside the oil tank to collect data. A voiceprint monitoring path, which uses a voiceprint detection sensor to monitor the sound waves emitted by the transformer during operation, capture abnormal sound patterns, such as winding displacement, core vibration, etc., to help determine mechanical failures. A grounding current sensor, which is used for current monitoring in the grounding system to detect grounding faults or abnormal current leakage.

[0060] Furthermore, through the defined multi-source monitoring path, the sensor data types of different monitoring points are extracted to form a multi-dimensional initial monitoring indicator set. Among them, the monitoring indicators of the multi-dimensional initial monitoring indicator set include: partial discharge indicators such as discharge intensity, location and frequency. Thermal indicators such as thermal imaging data and temperature values. Vibration indicators such as vibration amplitude and frequency. Oil chromatography indicators of gas components such as hydrogen, methane, carbon monoxide in transformer oil.

[0061] Specifically, the transformer identity feature of the target transformer is obtained, and the historical health abnormality record data is obtained based on the transformer identity feature, wherein the transformer identity feature refers to the key attribute used to uniquely identify a specific transformer, such as the transformer model, manufacturer, year of use, capacity, operating environment, etc. Based on the transformer identity feature, the historical health abnormality data of the transformer can be extracted from the database or monitoring system.

[0062] Specifically, historical health abnormality record data includes: intrinsic health abnormality record data, that is, historical abnormality records of specific transformers in target scenarios, such as past faults, maintenance records, warning records, etc.; homologous health abnormality record data, that is, abnormality records from other similar transformers (same model, similar operating conditions in non-target scenarios). This record composition helps to discover possible common fault modes.

[0063] Specifically, data lightweighting is performed based on the multidimensional initial monitoring indicator set, and data lightweighting is used to extract the most important information for model prediction, thereby reducing the complexity and redundancy of the data. In other words, the historical health abnormality record data obtained above may contain indicator data of indicators that are not in the multidimensional initial monitoring indicator set, so it is necessary to use data lightweighting to filter out the data portion corresponding to the multidimensional initial monitoring indicator set from the historical health abnormality record data and remove redundant or noisy data. The above-mentioned data lightweighting process helps to reduce the amount of sample data, and the obtained standard abnormal sample data set has a high data quality, reducing the risk of overfitting in the subsequent model construction and training process.

[0064] Optionally, the data after light processing forms a standard abnormal sample data set, which includes the historical monitoring indicator data, historical abnormal state and corresponding historical abnormal position of the transformer, and is used for the subsequent training of the global prediction model.

[0065] Furthermore, a global prediction model is constructed using machine learning or deep learning methods based on the standard abnormal sample data set. The global prediction model comprehensively predicts the health status of the transformer by integrating multi-dimensional monitoring data. For example, the model architecture can select classic machine learning models such as decision trees, random forests, and support vector machines (SVM), or select deep learning models based on neural networks (such as LSTM and CNN) to process time series data or image data.

[0066] Specifically, the global prediction model can be refined into multiple local health prediction clusters, each of which is used to make separate predictions for different types of monitoring data or fault types. Exemplarily, multiple local health prediction clusters can be divided into: Mechanical fault prediction cluster, used to predict possible mechanical faults of the transformer, such as loose windings or damaged core components, based on vibration and soundprint data. Electrical fault prediction cluster, used to predict electrical faults or insulation problems based on data such as partial discharge and oil chromatography analysis. Thermal fault prediction cluster, used to predict potential risks due to overheating based on infrared and temperature data.

[0067] Specifically, through the collaborative work of each local health prediction cluster, the prediction results are integrated into the global prediction model to form a comprehensive health prediction result. This comprehensive prediction result can more comprehensively reflect the overall health status of the transformer and improve the accuracy and reliability of the prediction. Through the setting of local health prediction clusters, the system can perform detailed analysis of different fault types of the transformer and provide more accurate decision support for operation and maintenance personnel.

[0068] In some implementations, the global prediction model is constructed and trained according to the standard abnormal sample data set, and the execution steps include:

[0069] Acquire multiple classified abnormal sample data sets, wherein the multiple classified abnormal sample data sets are clustering results of clustering the standard abnormal sample data set according to health abnormality categories. Traverse the multiple classified abnormal sample data sets, and construct and train multiple local health prediction clusters based on an ensemble learning method, wherein the multiple local health prediction clusters are specifically adapted to multiple types of health abnormalities.

[0070] Specifically, the classified abnormal sample data set is a subset formed by clustering the standard abnormal sample data set according to different health abnormality categories. Each subset corresponds to a specific health abnormality category, such as partial discharge abnormality, temperature abnormality, vibration abnormality, etc. These classified data sets contain the characteristics of specific categories of health abnormalities, which is helpful for the subsequent construction and training of corresponding prediction models for multiple categories of health abnormalities.

[0071] Exemplarily, the standard abnormal sample data set is clustered and divided. First, a variety of health abnormality labels (such as electrical failure, mechanical failure, thermal failure, etc.) are extracted from the standard abnormal sample data set; then, a clustering algorithm (such as K-means, hierarchical clustering, DBSCAN, etc.) is used to cluster the standard abnormal sample data set according to the extracted multiple health abnormality labels. The clustering result of the clustering division is multiple clusters, and each cluster contains samples of a specific health abnormality category. Multiple clusters correspond to the generation of multiple classified abnormal sample data sets.

[0072] Specifically, the sample data set is traversed, and local health prediction clusters corresponding to specific health abnormality categories are constructed respectively, so as to build the most suitable prediction model for each health abnormality category, thereby improving the accuracy of the overall prediction model.

[0073] Specifically, each local health prediction cluster contains multiple sub-predictors. In other words, the local health prediction cluster is a strong prediction model constructed by combining multiple weak learners (i.e., multiple sub-predictors) based on an ensemble learning method. By integrating multiple sub-predictors, the prediction accuracy and robustness of the local health prediction cluster can be improved. Exemplary ensemble learning methods include Random Forest, Gradient Boosting, XGBoost, Bagging, Boosting, etc.

[0074] Exemplary, construct a local health prediction cluster: First, for each classified abnormal sample data set, select a suitable ensemble learning method. For example, for the local discharge abnormal data set, use random forests and combine multiple decision trees to build a strong prediction model; for the temperature abnormal data set, use a gradient boosting algorithm (such as XGBoost) to capture the characteristics of abnormal temperature changes. Then, on each classified abnormal sample data set, train multiple weak learners (such as decision trees, linear regression models, etc.), and then combine these weak learners into strong learners through ensemble methods to form multiple local health prediction clusters.

[0075] Optionally, each local prediction cluster is used to output the health prediction result for the category according to the input multi-dimensional monitoring data set. By integrating multiple local prediction clusters, a global prediction model can be generated, thereby obtaining the ability to comprehensively evaluate the health status of the entire transformer and providing a comprehensive prediction and diagnosis of the transformer health status. The integration method may include weighted average, voting mechanism, logical operation, etc.

[0076] Specifically, the global prediction model combines the prediction results of multiple local prediction clusters and gives detailed prediction outputs about the overall health status of the transformer and specific monitoring points. The prediction outputs of the global prediction model include: global health prediction results, that is, the Boolean value (True / False) of the overall health prediction of the transformer. A Boolean prediction value of True indicates that the transformer is in a healthy state and has no obvious fault risk; False indicates that a potential health anomaly or fault risk has been detected, requiring further diagnosis and maintenance; Boolean prediction values ​​of the health of each multi-source monitoring point of the transformer, that is, the health status of each multi-source monitoring point. For example, a Boolean prediction value of False for a temperature monitoring point may indicate that the temperature of the point is beyond the normal range; Health anomaly category, that is, a specific anomaly category, indicating in which aspect the transformer has an abnormality. Quantitative output of the degree of abnormality, that is, a quantitative indicator of the degree of abnormality, such as the severity of the abnormality, the risk level of fault development, etc., can be expressed by percentage, score or level (such as mild, moderate, severe).

[0077] Through the above steps, combined with the ensemble learning method, multiple local health prediction clusters are constructed to achieve specific prediction analysis for each health anomaly, ensure the prediction accuracy of different types of anomalies, and improve the efficiency of the training construction process and the robustness of the prediction. At the same time, multiple local health prediction clusters decompose the complex transformer health prediction task into multiple specific small tasks, enhancing the scalability and portability of the global prediction model, making it easier to further expand, modify, and optimize the model according to actual needs in the application to adapt to new health anomaly categories or monitoring scenarios.

[0078] The simplified model construction module 12 is used to perform distillation and dimensionality reduction on the global prediction model, construct a simplified prediction model, and distribute the simplified prediction model to the edge prediction device, wherein the simplified prediction model includes multiple hierarchical simplified prediction models.

[0079] Specifically, global prediction models are usually complex, with a large number of parameters and high computing requirements, and are suitable for running on the cloud or central servers. Compared with cloud platforms, edge devices have relatively low computing power and limited computing resources, and cannot handle such a large model. Therefore, through model distillation and dimensionality reduction technology, complex global prediction models are simplified into lightweight models, which are easy to deploy and run on edge devices.

[0080] Specifically, distillation and dimensionality reduction refers to the process of extracting the knowledge of the global prediction model into a simplified model. The simplified model learns the useful information in the global prediction model by imitating the output of the global prediction model, rather than directly training with the original training data, thereby achieving model compression and simplification. Among them, the simplified prediction model is a lightweight model generated after distillation and dimensionality reduction of the global prediction model. It has fewer parameters and computational complexity, but can still maintain good prediction accuracy, and is suitable for real-time prediction and health monitoring on edge devices.

[0081] Optionally, by analyzing the input data of the global prediction model and combining dimensionality reduction technology, the most important health monitoring indicators (such as partial discharge intensity, temperature, vibration, etc.) can be screened out to reduce the analogy and number of input indicators and further reduce the complexity of the model.

[0082] Optionally, multiple hierarchical simplified prediction models are constructed based on the computing power limitations of edge devices. Each simplified model is optimized for different levels of complexity and computing power to adapt to different types of edge devices and prediction requirements. Exemplarily, for edge devices with limited computing power and memory, the simplest prediction model is constructed to reduce input features and model parameters, retaining only key health monitoring indicators. For devices with medium computing power, more health monitoring indicators and prediction layers are retained, and prediction functions for local monitoring points are added, such as health status predictions for vibration, temperature, and partial discharge. For high-computing power edge devices, most of the monitoring indicators and model layers are retained, and refined predictions can be made, including the overall health of the transformer, the abnormality category and degree of abnormality at each monitoring point, etc. Multiple hierarchical simplified prediction models ensure that the model can be flexibly deployed according to the computing power of different edge devices, which can not only ensure the lightweight of the model, but also improve the prediction accuracy when resources permit.

[0083] Furthermore, the appropriate hierarchical simplified model is dynamically selected for deployment to ensure that the computing requirements of the model match the computing power of the device, so that the simplified prediction model can run on the edge device.

[0084] Through the above steps, a hierarchical simplified prediction model is built and deployed, which reduces the computing burden of edge devices, optimizes the operating efficiency of equipment, extends the service life of equipment, and enhances the overall performance of the transformer health monitoring system.

[0085] In some embodiments, before performing distillation and dimensionality reduction on the global prediction model and constructing a simplified prediction model, the steps include:

[0086] Obtain edge device configuration information of multiple edge prediction devices and extract performance configuration data sets, wherein the performance configuration data sets correspond to multiple types of edge prediction devices. Perform computing power evaluation based on the performance configuration data sets, and define multi-level model scale constraints based on the computing power evaluation results.

[0087] Specifically, edge device configuration information is data related to the hardware and software capabilities of edge prediction devices, which characterizes the computing power and operating environment of each device and is used to evaluate the computing power, storage resources, network bandwidth and other performance indicators that affect model operation. Among them, the methods for obtaining edge device configuration information include device self-test reports, manual configuration input and database queries.

[0088] Specifically, the performance configuration data set corresponds to multiple types of edge prediction devices, that is, the performance configuration data set contains performance configuration data of multiple types of edge devices.

[0089] Exemplarily, edge device configuration information includes hardware information (such as processor type, main frequency, number of cores, memory size, storage capacity, GPU availability, etc.), software environment information, network bandwidth and latency.

[0090] Specifically, computing power evaluation is a process of evaluating computing resources based on the hardware and software capabilities of edge devices to determine the model complexity and reasoning tasks that each device can support. Computing power evaluation is used to guide the distillation and deployment of simplified prediction models. The computing power evaluation process involves processing power evaluation, storage capacity evaluation, network bandwidth evaluation, etc.

[0091] Exemplarily, the processing power evaluation is performed based on the device's CPU, GPU, and memory configuration to determine the number of floating point operations per second (FLOPS) that the edge prediction device can perform. The processing power determines the complexity of the model that the device can run. The storage capacity evaluation is performed based on the available storage capacity of the device to determine the size of the model that the edge prediction device can accommodate. The network bandwidth evaluation includes the evaluation of network bandwidth and latency, which is used to determine whether real-time model updates can be performed or whether it is suitable to run prediction tasks locally on the edge device.

[0092] Furthermore, based on the computing power evaluation results, the model scale limit suitable for each type of edge device is defined, namely the multi-level model scale constraint. The multi-level model scale constraint determines the complexity of the simplified prediction model that can be run on edge prediction devices of different performance levels to ensure that each device can effectively perform the prediction reasoning task of transformer health.

[0093] In some implementations, computing power evaluation is performed according to the performance configuration data set, and multi-level model scale constraints are defined in combination with the computing power evaluation results. The execution steps include:

[0094] Interactive target scenario, taking the multiple types of edge prediction devices as statistical dimensions, obtain the distribution of the number of edge prediction devices in the target scenario. Based on the preset group distribution constraints, the distribution of the number of edge prediction devices is cumulatively divided into groups along the direction of increasing computing power. Whenever the cumulative number of devices of N types of edge prediction devices accounts for more than the group distribution constraints, a group division is performed, and multiple group division results are output. Based on the multiple group division results, the lower limit of the computing power within the group is extracted and output as a multi-level computing power target, and the multi-level model scale constraint is defined based on the multi-level computing power target.

[0095] Specifically, in different application scenarios or working environments, the distribution of types and quantities of deployed edge prediction devices is not the same, and may include edge prediction devices of different types, quantities, and compositions. The number distribution of edge prediction devices refers to the types and proportions of edge prediction devices deployed in the target scenario. Exemplarily, to obtain the number distribution of edge prediction devices for the target scenario, first classify the edge devices based on the target scenario. For example, classify the devices according to the device model or device computing power, and the proportion of each device in the scenario is different; then, use "device classification" and "device quantity" as statistical dimensions to count the number distribution of each type of edge device in the target scenario.

[0096] Furthermore, based on the group distribution constraint, the devices are arranged in order from low to high in terms of computing power, and the number of devices at different computing power levels is accumulated. When the cumulative number of devices reaches or exceeds the preset group distribution constraint, a group division is performed. Multiple group division results correspond to multiple device groups. Among them, the group distribution constraint refers to a pre-set cumulative device number ratio threshold, which is used to determine the division ratio of each group of devices in the edge device number distribution. Usually, the group distribution constraint is set to a percentage, such as 20%, 30%, indicating the proportion of the number of devices in each group to the total number of devices.

[0097] Exemplarily, the group distribution constraint is set to 30%. When the cumulative number of edge devices from Category 1 to Category N reaches 30% of the total number of devices (for example, the cumulative number of edge devices from Category 1 to Category N accounts for 32% of the total number of edge devices), the first N groups of edge devices are divided into one group to complete the first group division. Then, multiple divisions are performed in sequence according to the cumulative proportion of the number of devices until all edge devices are grouped.

[0098] Through group division, edge devices are stratified according to computing power and quantity, so as to assign a suitable simplified prediction model to the group division result (i.e. each edge device group). Each group division result (i.e. each edge device group) includes an appropriate number of edge devices. This division method based on the proportion of device quantity helps to ensure that the simplified prediction models constructed subsequently can be fully deployed and applied. At the same time, the computing power resources are evenly distributed to avoid overloading of certain device groups or waste of model resources.

[0099] Specifically, the computing power lower limit of the group division result (i.e., each edge device group) refers to the computing power value of the edge prediction device with the weakest computing power in the group. Extracting the computing power lower limit helps to set the minimum model complexity constraint for each group to ensure that all devices can run the corresponding simplified prediction model. Then, based on the computing power lower limit of each device group, multi-level computing power targets are defined, that is, corresponding computing power requirements are set for different device groups, achieving the technical effect of providing a basis for subsequent model scale constraints.

[0100] Specifically, the multi-level model scale constraint is a model complexity limit defined according to the computing power target of each group of devices, which is used to determine the parameter scale, input dimension and computational complexity of the simplified prediction model, and ensure that each device can smoothly run the corresponding simplified prediction model. Exemplarily, for a group of devices with lower computing power, the model scale needs to be greatly simplified to reduce the number of parameters and input features. For example, the number of model parameters is limited to no more than 100K, and the input data dimension is limited to no more than 10 features. For medium computing power devices, the model scale can be moderately increased, for example, allowing the model to contain 500K parameters and the input dimension to reach 20 features. For high computing power devices, the model scale can be relatively large, allowing the use of complex deep learning models, such as the number of parameters can exceed 1M and the input dimension exceeds 30 features.

[0101] According to the multi-level computing power targets, the system can assign corresponding simplified prediction models to each device group to ensure that the computing requirements of the model match the computing power of the device. For example, low-computing power devices may only run lightweight linear regression models, while high-computing power devices run complex neural network models to achieve efficient and accurate predictions on devices with limited resources, thereby improving the performance of the overall prediction system.

[0102] In some embodiments, the global prediction model is subjected to distillation and dimensionality reduction to construct a simplified prediction model, and the execution steps include:

[0103] Perform computing power matching and establish a scale mapping table based on the computing power evaluation results and the scale constraints of the multi-level model. Construct multiple hierarchical simplified prediction models, and call the hierarchical simplified prediction model according to the scale mapping table to obtain the target hierarchical simplified prediction model. Perform functional branch pruning on the simplified prediction model based on the multi-dimensional initial monitoring indicator set, and perform enhanced training on the branch pruning prediction model to obtain an edge prediction model. Based on the scale mapping table, transmit the edge prediction model to the edge prediction device.

[0104] Specifically, first, the computing power of each device is determined based on the computing power evaluation results of different edge devices. Then, according to the previously defined multi-level model scale constraints, a suitable model scale is assigned to each type of edge device, and the matching allocation results are output as the above-mentioned scale mapping table to ensure that the model can run efficiently on the device. Among them, the scale mapping table is a relationship table that maps devices of different computing power levels to the corresponding model scales. This table clarifies what scale of simplified prediction models should be deployed on devices of different computing power levels in the target scenario.

[0105] Specifically, hierarchical simplified models of different complexity are constructed to adapt to different computing power devices. Then, according to the guidance of the scale mapping table, a simplified prediction model suitable for the computing power of the edge device is selected and output as the target hierarchical simplified prediction model.

[0106] Furthermore, when constructing a global prediction model, multi-dimensional health monitoring indicators may be used, while in the simplified model, only a small number of key monitoring indicators may need to be processed, so the simplified prediction model is functionally pruned to reduce unnecessary input dimensions. Exemplarily, functional pruning is achieved by pruning specific branches, feature extraction layers, or neural network layers of the simplified model to ensure that the model further reduces computational complexity while meeting prediction accuracy.

[0107] Specifically, after functional branch pruning, the model performance of the branch pruning prediction model is further improved by increasing the amount of data, adjusting training parameters, or using enhanced training methods such as transfer learning, so that it can still achieve efficient prediction results under the new simplified structure.

[0108] Specifically, through the scale mapping table, the type of edge prediction model that should be deployed on each edge device can be determined, and the edge prediction model can be transmitted to the corresponding edge prediction device for edge deployment.

[0109] Optionally, when the computing power and load of the edge prediction device change, the system dynamically adjusts the scale mapping table and updates the transmission edge prediction model to ensure that the deployed model can adapt to the current state of the device.

[0110] Optionally, when the global prediction model or the simplified model is updated, the new edge prediction model is sent to the device through the same transmission mechanism to ensure that the edge device uses the latest and most optimized model.

[0111] The above steps establish a scale mapping table so that the simplified prediction model can be flexibly allocated according to the computing power of the edge device and efficiently transmitted to the device, ensuring that the model runs smoothly on the edge device with limited resources. Combined with functional pruning and enhanced training, the prediction accuracy is maintained while further reducing the model size.

[0112] In some implementations, constructing a plurality of the hierarchical simplified prediction models, the execution step further includes:

[0113] According to the multi-level model scale constraint, multiple standard simplified prediction models are defined, wherein the model scales and structures of the multiple standard simplified prediction models are different. The standard abnormal sample data set is softened based on the global prediction model to obtain a softened prediction label set. The softened prediction label set is binarized and simplified to obtain a global prediction label set, wherein the global prediction label set includes a binarized prediction result and a prediction probability label. Multiple standard simplified prediction models are trained with the global prediction label set and the standard abnormal sample data set to obtain multiple hierarchical simplified prediction models.

[0114] Specifically, according to the multi-level model scale constraints, multiple standard simplified prediction models are defined. These models differ in scale (number of parameters, number of layers) and structure (model hierarchy, feature extraction capability). They include linear models, shallow neural networks, deep convolutional neural networks, etc., to meet the needs of different computing resource devices.

[0115] Furthermore, the global prediction model is used as the teacher model to soften the standard abnormal sample data set to generate a softened prediction label set. The softened prediction label set is a probability distribution label determined based on the output results of the global model, which indicates the probability that the global prediction model believes that each sample belongs to a different health prediction result category, and is used to help simplify the prediction model for training.

[0116] Specifically, the simplified model is limited in scale and computing power and cannot handle overly complex softened labels. Therefore, the softened prediction label set is binary-simplified and converted into simpler classification labels (such as binary classification of healthy / abnormal) and prediction probability labels. Furthermore, the binary labels and probability labels (i.e., global prediction label set) generated by the global prediction model are used for training to ensure that the simplified model can inherit the prediction capability from the global model while adapting to the computing power limitations of edge devices.

[0117] Exemplarily, a hierarchical simplified prediction model is obtained, including taking a standard abnormal sample data set and a global prediction label set (including binarization results and prediction probabilities) as input data, and training multiple standard simplified prediction models. During the training process, the standard simplified prediction model is optimized by minimizing the gap between the prediction output and the input data through objective functions such as the cross entropy loss function. This allows a more efficient transformer health prediction capability to be obtained, that is, a prediction effect close to that of the global model can be achieved with fewer parameters and computing resources.

[0118] The multiple hierarchical simplified prediction models trained through the above steps are adaptively optimized according to different device computing power levels. The complexity and scale of the models are gradually increased to adapt to edge devices with different computing power levels.

[0119] Furthermore, before training a plurality of the standard simplified prediction models with the global prediction label set and the standard abnormal sample data set, the execution step further includes:

[0120] According to the multi-level model scale constraint, a multi-level distillation coefficient is calculated, wherein the multi-level distillation coefficient is the scale ratio of the multi-level model scale constraint to the global prediction model. The multi-level distillation coefficient is used as the fusion weight to define a multi-level loss function, which is:

[0121] ;

[0122] in, Characterizing the multi-level loss function; Characterizing the fusion weight; Characterizing the loss of the standard simplified prediction model training based on the standard abnormal sample data set; Characterize the loss in the training of the standard simplified prediction model based on the global prediction label set.

[0123] Specifically, the multi-level distillation coefficient is defined by comparing the scale difference between the global prediction model and each hierarchical simplified prediction model, and is used to balance the knowledge transfer between the simplified model and the global model during the model distillation process. In other words, the multi-level distillation coefficient determines how the student model weighs the true labels from the standard abnormal sample data set and the soft labels from the global prediction model during training. Generally, the smaller the simplified model, the smaller the corresponding distillation coefficient, indicating that the simplified model relies more on the soft labels of the global model during training. Exemplarily, if the model scale ratio is 1 / 10, the distillation coefficient is 0.1. Through reasonable weights, it helps to ensure that the simplified prediction model can learn both the true label information and inherit the knowledge of the global prediction model.

[0124] Specifically, Represents the total loss of the multi-level loss function, which is the optimization target of the simplified model. represents the loss based on the global set of predicted labels (i.e., soft labels), such as the difference between the output of the simplified model and the global model soft labels measured using the cross entropy loss function. Represents the loss based on the standard abnormal sample dataset, that is, the prediction error of the simplified model on the true label.

[0125] The monitoring data collection module 13 is used to activate the multi-source monitoring equipment, perform normal multi-source monitoring and collection of the target transformer, and obtain a real-time multi-source monitoring data set.

[0126] Optionally, the multi-source monitoring equipment includes at least a UHF sensor, an ultrasonic sensor, a high-frequency current transformer, a voiceprint detection sensor, a ground current sensor, a vibration sensor, an infrared sensor, and an oil chromatography monitoring device.

[0127] Optionally, before performing normal multi-source monitoring and acquisition of the target transformer, first, ensure that the multi-source monitoring equipment is safe and sound, and calibrate the equipment to ensure that all sensors can work accurately. Then, configure the data acquisition frequency and data transmission method. Preferably, a relatively low data acquisition frequency is used for normal multi-source monitoring and acquisition.

[0128] For example, the real-time multi-source monitoring data set includes partial discharge signals, discharge frequency and intensity; mechanical vibration signals, sound intensity and spectrum characteristics; soundprint characteristics; ground current waveform, fault current characteristics; temperature distribution diagram, hot spot temperature value; transformer oil dissolved gas concentration, gas type and change trend, etc.

[0129] The step-by-step prediction module 14 is used to transmit the real-time multi-source monitoring data set to the edge prediction device, perform binary risk discrimination on the target transformer, and based on the binary risk discrimination result, call the global prediction model to perform full-dimensional positioning prediction and output a health prediction result.

[0130] Specifically, first, the edge prediction device receives multi-dimensional monitoring data from the multi-source monitoring path, including sensor data such as temperature, vibration, partial discharge, and oil chromatography analysis, and performs a preliminary prediction of the transformer health status through the edge prediction model deployed on the edge prediction device to obtain the binary risk discrimination result of the target transformer. Then, if the binary risk discrimination result of the target transformer is that there is a health risk, the real-time multi-source monitoring data set is updated and collected, and the real-time multi-source monitoring data set before and after the update is transmitted to the cloud platform and input into the configured global prediction model; then, the global prediction model passes the received input data (i.e., the real-time multi-source monitoring data set before and after the update) to each local health prediction cluster according to the response relationship between the multi-dimensional initial monitoring indicator set and multiple types of health anomalies; then, each local health prediction cluster runs in parallel, processes the data related to the corresponding health anomaly category, and outputs the prediction results for multiple types of health anomalies. Finally, the prediction results of each local prediction cluster are summarized to generate a global prediction result.

[0131] In the above steps, the global prediction model comprehensively evaluates the health status of the transformer through multiple local prediction clusters, judges the overall health status of the transformer, and outputs specific abnormality categories, abnormality locations, and abnormality levels, converting complex monitoring data into easy-to-understand health status indicators, which helps maintenance personnel to more intuitively understand the operating status of the transformer.

[0132] In summary, the transformer health prediction system under multi-dimensional monitoring data fusion provided by the present invention has the following technical effects:

[0133] Based on the multi-dimensional initial monitoring indicator set, a global prediction model configured on the cloud platform is constructed. The global prediction model is subjected to knowledge distillation and dimensionality reduction processing to generate a simplified prediction model, which is then distributed to the edge prediction device. The simplified prediction model includes multiple hierarchical simplified prediction models. The multi-source monitoring device is activated to perform normalized multi-source monitoring data collection on the target transformer to obtain a real-time multi-source monitoring data set. The real-time multi-source monitoring data set is transmitted to the edge prediction device, and the binary risk discrimination of the target transformer is performed. According to the binary risk discrimination results, the global prediction model is called to perform full-dimensional positioning and health status prediction, and the final health prediction result is output. This achieves the technical effect of improving the prediction responsiveness and accuracy and reducing the prediction cost.

[0134] It should be understood that the embodiments disclosed in the present invention and the above description can enable those skilled in the art to use the present invention to implement the present invention. At the same time, the present invention is not limited to the above-mentioned embodiments. It should be understood that those skilled in the art can still modify the technical solutions recorded in the above-mentioned embodiments, or replace some of the technical features therein by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.

Claims

1. The transformer health prediction system based on multi-dimensional monitoring data fusion is characterized by: The system comprises: A global model configuration module, the global model configuration module is used to build a global prediction model according to a multi-dimensional initial monitoring indicator set, wherein the global prediction model is configured on a cloud platform; A simplified model construction module, the simplified model construction module is used to perform distillation and dimensionality reduction on the global prediction model, construct a simplified prediction model, and distribute the simplified prediction model to an edge prediction device, wherein the simplified prediction model includes a plurality of hierarchical simplified prediction models; A monitoring data acquisition module, which is used to activate the multi-source monitoring equipment, perform normal multi-source monitoring and acquisition of the target transformer, and obtain a real-time multi-source monitoring data set; A step-by-step prediction module, which is used to transmit the real-time multi-source monitoring data set to the edge prediction device, perform binary risk discrimination of the target transformer, and call the global prediction model to perform full-dimensional positioning prediction based on the binary risk discrimination result, and output a health prediction result; Based on the multi-dimensional initial monitoring indicator set, a global prediction model is constructed. The execution steps include: Determine the multi-source monitoring points, combine the target scene monitoring requirements, and define the multi-source monitoring path; Based on the multi-source monitoring path, extracting a multi-dimensional initial monitoring indicator set; Obtaining a transformer identity feature of a target transformer, and obtaining historical health abnormality record data based on the transformer identity feature, wherein the historical health abnormality record data includes intrinsic health abnormality record data and homologous health abnormality record data; Based on the multi-dimensional initial monitoring indicator set, the historical health abnormality record data is lightened, and the lightened result is output as a standard abnormal sample data set; According to the standard abnormal sample data set, construct and train the global prediction model, wherein the global prediction model includes a plurality of local health prediction clusters; Before performing distillation and dimensionality reduction on the global prediction model and constructing a simplified prediction model, the steps include: Obtain edge device configuration information of multiple edge prediction devices, and extract a performance configuration data set, wherein the performance configuration data set corresponds to multiple types of edge prediction devices; Performing computing power evaluation based on the performance configuration data set, and defining multi-level model scale constraints based on the computing power evaluation results; The computing power is evaluated according to the performance configuration data set, and the multi-level model scale constraints are defined in combination with the computing power evaluation results. The execution steps include: Interact with the target scene, taking the multiple types of edge prediction devices as statistical dimensions to obtain the distribution of the number of edge prediction devices in the target scene; Based on the preset group distribution constraint, the number distribution of the edge prediction devices is cumulatively divided into groups along the direction of increasing computing power, and each time the cumulative number of devices of the N types of edge prediction devices accounts for more than the group distribution constraint, a group division is performed, and multiple group division results are output; According to the group division results multiple times, the lower limit of the computing power within the group is extracted and output as a multi-level computing power target, and the multi-level model scale constraint is defined based on the multi-level computing power target.

2. The transformer health prediction system under multi-dimensional monitoring data fusion according to claim 1 is characterized in that: The global prediction model is subjected to distillation and dimensionality reduction to construct a simplified prediction model, and the execution steps include: Perform computing power matching and establish a scale mapping table based on the computing power evaluation result and the scale constraint of the multi-level model; Constructing a plurality of the hierarchical simplified prediction models, and calling the hierarchical simplified prediction models according to the scale mapping table to obtain a target hierarchical simplified prediction model; Based on the multi-dimensional initial monitoring indicator set, the simplified prediction model is functionally pruned, and the pruned prediction model is enhanced and trained to obtain an edge prediction model; Based on the scale mapping table, the edge prediction model is transmitted to the edge prediction device.

3. The transformer health prediction system under multi-dimensional monitoring data fusion as claimed in claim 2 is characterized in that: Constructing a plurality of said hierarchical simplified prediction models, the execution step further includes: According to the multi-level model scale constraint, a plurality of standard simplified prediction models are defined, wherein the model scales and structures of the plurality of standard simplified prediction models are different; Softening the standard abnormal sample data set based on the global prediction model to obtain a softened prediction label set; Binarizing and simplifying the softened prediction label set to obtain a global prediction label set, wherein the global prediction label set includes a binary prediction result and a prediction probability label; A plurality of the standard simplified prediction models are trained with the global prediction label set and the standard abnormal sample data set to obtain a plurality of the hierarchical simplified prediction models.

4. The transformer health prediction system under multi-dimensional monitoring data fusion according to claim 1 is characterized in that: According to the standard abnormal sample data set, the global prediction model is constructed and trained, and the execution steps include: Acquire multiple classified abnormal sample data sets, wherein the multiple classified abnormal sample data sets are clustering division results of clustering the standard abnormal sample data set according to healthy abnormal categories; Traversing a plurality of the classified abnormal sample data sets, based on an ensemble learning method, constructing and training a plurality of local health prediction clusters, wherein the plurality of local health prediction clusters are specifically adapted to a plurality of types of health abnormalities.

5. The transformer health prediction system under multi-dimensional monitoring data fusion as claimed in claim 3, characterized in that: Prior to training a plurality of the standard simplified prediction models with the global prediction label set and the standard abnormal sample data set, the execution steps include: Calculating a multi-stage distillation coefficient according to the multi-stage model scale constraint, wherein the multi-stage distillation coefficient is a scale ratio of the multi-stage model scale constraint to the global prediction model; The multi-level distillation coefficient is used as the fusion weight to define a multi-level loss function, which is: ； in, represents the multi-level loss function; α represents the fusion weight; Characterizing the loss of the standard simplified prediction model training based on the standard abnormal sample data set; Characterize the loss in the training of the standard simplified prediction model based on the global prediction label set.

Images