Transformer-oriented data classified storage method, device and equipment and storage medium

Abstract

The invention relates to a transformer-oriented data classified storage method and device, equipment and a storage medium. Data collected for a transformer is input into a diagnosis model, the diagnosis model extracts features in the data and then outputs a data score, and the data is stored according to a target storage strategy corresponding to the data score. Compared with a traditional coarse-grained strategy of timed storage or simple threshold triggering storage, the scheme has the advantages that feature extraction and data scoring are performed on the data of the transformer by combining the diagnosis model, and then different storage strategies can be used for storing the data with different scores; therefore, the technical effect of reducing waste of storage resources is achieved.

Description

Technical Field

[0001] This application relates to the field of power equipment data processing technology, and in particular to a data classification and storage method, apparatus, computer equipment, computer-readable storage medium, and computer program product for transformers. Background Technology

[0002] As a core component of the power grid transmission chain, the reliability of power transformers directly determines the safe and stable operation of the power system. With the deepening of smart grid construction, transformer online monitoring systems are evolving from traditional single-parameter threshold alarm modes to multi-parameter intelligent fusion analysis modes. However, at the data storage level, coarse-grained strategies such as timed storage or simple threshold-triggered storage are commonly used. This approach results in storage space being occupied by a large amount of low-value steady-state data, leading to a waste of storage resources.

[0003] Therefore, the current method of storing transformer data has the drawback of wasting storage resources. Summary of the Invention

[0004] Therefore, it is necessary to provide a data classification and storage method, apparatus, computer equipment, computer-readable storage medium, and computer program product for transformers that can reduce the waste of storage resources, in order to address the above-mentioned technical problems.

[0005] Firstly, this application provides a data classification and storage method for transformers, including:

[0006] Acquire the data to be stored collected from the transformer;

[0007] The data is input into a trained diagnostic model; the diagnostic model is used to extract features corresponding to the data and output a corresponding data score based on the features; the data score characterizes the degree of impact of the data on the stable operation of the transformer.

[0008] Based on the data score, determine the target storage strategy corresponding to the data;

[0009] The data is stored according to the target storage strategy.

[0010] Secondly, this application also provides a data classification and storage device for transformers, comprising:

[0011] The acquisition module is used to acquire the data to be stored collected from the transformer.

[0012] A diagnostic module is used to input the data into a trained diagnostic model; the diagnostic model is used to extract features corresponding to the data and output corresponding data scores based on the features; the data scores characterize the degree of impact of the data on the stable operation of the transformer.

[0013] The determination module is used to determine the target storage strategy corresponding to the data based on the data score;

[0014] A storage module is used to store the data according to the target storage strategy.

[0015] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0016] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0017] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described method.

[0018] The aforementioned data classification and storage method, apparatus, computer equipment, computer-readable storage medium, and computer program product for transformers input data collected from transformers into a diagnostic model. The diagnostic model extracts features from the data and outputs a data score. Based on the target storage strategy corresponding to the data score, the data is stored. Compared to traditional coarse-grained strategies that rely on timed storage or simple threshold-triggered storage, this solution, by combining a diagnostic model to extract features and score transformer data, enables the use of different storage strategies for data with different scores, thereby reducing the waste of storage resources. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating a data classification and storage method for transformers in one embodiment.

[0021] Figure 2 This is a flowchart illustrating the data acquisition steps in one embodiment;

[0022] Figure 3 This is a schematic diagram of the structure of a data classification and storage system for transformers in one embodiment;

[0023] Figure 4 This is a flowchart illustrating a data classification and storage method for transformers in another embodiment;

[0024] Figure 5 This is a flowchart illustrating the monitoring steps of an oil-immersed transformer in one embodiment;

[0025] Figure 6 This is a time-domain plot of the vibration signal in one embodiment;

[0026] Figure 7 This is a frequency domain diagram of the vibration signal in one embodiment;

[0027] Figure 8 This is a time-domain plot of the target data in one embodiment;

[0028] Figure 9 This is a flowchart illustrating the storage steps in one embodiment;

[0029] Figure 10 This is a flowchart illustrating the feature fusion steps in one embodiment;

[0030] Figure 11 This is a schematic diagram of the hierarchical quantization steps in one embodiment;

[0031] Figure 12 This is a flowchart illustrating the data recovery steps in one embodiment;

[0032] Figure 13 This is a structural block diagram of a transformer-oriented data classification and storage device in one embodiment;

[0033] Figure 14 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0035] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0036] In related technologies, power transformers, as core equipment in the power grid transmission chain, directly determine the safe and stable operation of the power system based on their operational reliability. With the deepening of smart grid construction, transformer online monitoring systems are evolving from traditional single-parameter threshold alarm modes to multi-parameter intelligent fusion analysis modes. However, existing technical solutions still face the following bottlenecks:

[0037] At the data acquisition level, the fixed acquisition strategies of monitoring systems cannot be dynamically adjusted according to the actual operating status of the equipment. When the equipment is abnormal, key features may be missed due to insufficient sampling rate, while data redundancy may occur under normal conditions.

[0038] At the data storage level, edge terminals employ coarse-grained strategies such as timed storage or simple threshold-triggered storage. This approach results in storage space being occupied by a large amount of low-value, steady-state data, and when a failure occurs, critical pre-failure data is often lost due to trigger delays or rigid storage strategies. This fails to meet the needs for high-value data preservation and in-depth post-failure analysis.

[0039] Therefore, there is an urgent need in this field for an integrated innovative solution that can connect the entire chain from data collection to storage, and realize intelligent data collection, real-time analysis, and value creation through storage, in order to solve problems such as data silos, response delays, and waste of storage resources.

[0040] Based on this, this application completely solves the problems of data silos, response delays, and storage resource waste by implementing multi-parameter adaptive collaborative acquisition at the data source and dynamic storage scheduling driven by data value, thereby improving the real-time performance, accuracy, and data value density of state awareness.

[0041] In one embodiment, such as Figure 1 As shown, a data classification and storage method for transformers is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and implemented through interaction between the terminal and the server, including the following steps S202 to S208. Wherein:

[0042] Step S202: Obtain the data to be stored collected from the transformer.

[0043] The terminal can be a storage control device. It includes an edge intelligence processing module and a classification storage module. The edge intelligence processing module is used for value assessment and strategy formulation of transformer data, while the classification storage module is used for classifying and storing the data. The terminal can communicate with a data acquisition module, which collects transformer data. These three modules form an organic closed loop through data and control flows, achieving end-to-end optimization from multimodal data perception to intelligent diagnosis and value-added storage. The terminal can acquire data collected from the transformer for storage. For example, the acquisition module can monitor transformer operation data using a series of sensors, collecting corresponding operational monitoring data as data to be stored. The terminal can obtain this data from the acquisition module. This data includes multimodal data.

[0044] Step S204: Input the above data into the trained diagnostic model; the diagnostic model is used to extract the features corresponding to the above data and output the corresponding data score based on the features; the data score characterizes the degree of influence of the above data on the stable operation of the above transformer.

[0045] The terminal can pre-train the diagnostic model iteratively based on data samples and data scoring samples corresponding to the transformer, resulting in a trained diagnostic model. This diagnostic model can be a lightweight, hierarchical model. The terminal can input the collected data into the trained diagnostic model. The diagnostic model can extract features from the data to obtain the features corresponding to the data to be stored. Specifically, the diagnostic model can extract multiple features through a multi-layer architecture and then fuse these features to obtain the features corresponding to the data. Therefore, the diagnostic model can score the data based on these features.

[0046] For example, based on the aforementioned features, the diagnostic model diagnoses the aforementioned data, obtaining diagnostic results and confidence levels. The diagnostic result represents the value of the data's features (such as the importance of the data to the stable operation of the transformer), and the confidence level represents the probability that the diagnostic result matches the true value of the data. Thus, the terminal can obtain a data score corresponding to the aforementioned features based on the diagnostic results and confidence levels. For example, the terminal might use diagnostic results with a confidence level greater than a preset confidence threshold as the target diagnostic result, and determine the data score of the target diagnostic result based on the correspondence between the diagnostic results and the data scores, using this score as the data score corresponding to the aforementioned feature. In other words, the data score represents the value of the data to be stored, such as characterizing the degree of impact of the data on the stable operation of the transformer. The greater the degree of impact, the higher the value of the data; conversely, the lower the value of the data.

[0047] Step S206: Based on the above data scoring, determine the target storage strategy corresponding to the above data.

[0048] The terminal has multiple preset storage strategies. Different data scores correspond to different storage strategies. The terminal can then determine the target storage strategy based on the data scores. The amount of data stored can differ depending on the storage strategy corresponding to the different data scores. A higher data score indicates greater data value, and the corresponding target storage strategy will store more data; conversely, a lower data value indicates less data storage.

[0049] For example, in one embodiment, the target storage strategy includes one or more of the following: a first storage strategy, a second storage strategy, and a third storage strategy. Each storage strategy has a different data score: the data score of the first storage strategy is less than or equal to a first scoring threshold; the data score of the second storage strategy is greater than the first scoring threshold and less than or equal to a second scoring threshold; and the data score of the third storage strategy is greater than the second scoring threshold. The first and second scoring thresholds can be set according to actual conditions to indicate that the value of the data corresponding to the first storage strategy is less than the value of the data corresponding to the second storage strategy; and the value of the data corresponding to the second storage strategy is less than the value of the data corresponding to the third storage strategy.

[0050] Step S208: Store the data according to the target storage strategy described above.

[0051] Once the terminal determines the target storage strategy corresponding to the aforementioned data, it can store the data to be stored according to the target storage strategy. Under different target storage strategies, the content of the data stored by the terminal and the data retention time can differ. For example, if the target storage strategy is the second storage strategy, it indicates that the value of the aforementioned data is moderate, and the terminal can store the complete data.

[0052] In the aforementioned data classification and storage method for transformers, data collected from transformers is input into a diagnostic model. The model extracts features from the data and outputs a data score. The data is then stored according to the target storage strategy corresponding to the score. Compared to traditional coarse-grained strategies that rely on timed storage or simple threshold-triggered storage, this solution combines a diagnostic model to extract features and score transformer data. This allows for the application of different storage strategies to data with different scores, thereby reducing the waste of storage resources.

[0053] In one embodiment, after determining the target storage strategy corresponding to the data based on the data score, the method further includes: if the target storage strategy is a second storage strategy or a third storage strategy, then sending a target acquisition instruction to the data acquisition device corresponding to the transformer; the target acquisition instruction is used to instruct the data acquisition device to increase the data acquisition frequency for the data.

[0054] In this embodiment, when the target storage strategy belongs to the first or second storage strategy, the terminal can also provide feedback on the acquisition process. For example, when the target storage strategy is the second or third storage strategy, if the terminal determines that the data has high value and needs to be acquired more efficiently, the terminal can send a target acquisition command to the data acquisition device corresponding to the transformer. The data acquisition device can be equipped with an intelligent acquisition module. Through the intelligent acquisition module, the data acquisition device can increase the data acquisition frequency based on the target acquisition command, thereby shortening the detection cycle for the data. In addition, when the storage strategy for the data is the first storage strategy within a preset time period, the terminal can also reduce the acquisition frequency of the data to reduce resource waste.

[0055] Specifically, the above data collection process can be as follows: Figure 2 As shown, Figure 2 This is a flowchart illustrating the data acquisition steps in one embodiment. The aforementioned data acquisition can be a high-precision synchronous acquisition and dynamic adjustment of multi-source heterogeneous data corresponding to the transformer. This includes a precise time synchronization mechanism, a unified interface and standardized data processing, and an adaptive acquisition control mechanism.

[0056] The terminal synchronizes the collected data using a clock. For example, the terminal collects data through multiple sensors in its intelligent acquisition module. The terminal uses a Precision Time Protocol (PTP) time synchronization unit, employing a master-slave clock architecture, to unify the time reference of all sensors. The terminal sets up a global master clock, and each sensor node acts as a slave clock, periodically receiving synchronization messages via the network. A receiver-side verification mechanism is used to calibrate time deviations in real time, ensuring that the time synchronization accuracy of all sensors reaches the millisecond level. The synchronization message carries node hop count and cumulative time error information through a Type-Length-Value (TLV) extended field, further optimizing synchronization accuracy.

[0057] The terminal can also perform standardized data processing. For example, the terminal can collect heterogeneous data such as transformer vibration, ultrasound, electrical parameters, oil and gas composition, and temperature. The terminal can filter, reduce noise, and standardize the format of data from different types of sensors to eliminate data silos.

[0058] The terminal can also adaptively control the data acquisition frequency. The acquisition frequency and parameter combination are not fixed, but dynamically adjusted by the terminal through the adaptive acquisition control unit based on the diagnostic results of the edge intelligent processing module. When the edge module detects suspected discharge characteristics, the data is of high value, and the corresponding storage strategy is the second or third storage strategy. The terminal can automatically send instructions to the acquisition module to increase the sampling rate of ultrasonic and ultra-high frequency (UHF) signals, shorten the detection cycle of the oil gas analysis unit, and start the fast cycle mode. When the equipment is in a long-term stable state, the terminal automatically reduces the sampling frequency and data transmission volume to reduce resource consumption.

[0059] Through this embodiment, the terminal can adaptively adjust the data collection frequency based on the value of the collected data, thereby shortening the detection cycle of high-value data and improving the safety of transformer operation; and reducing the sampling frequency of long-term stable data to reduce resource waste.

[0060] In one embodiment, storing the data according to the target storage strategy includes: if the target storage strategy is a first storage strategy, obtaining target data greater than a data threshold from the data; extracting a first target feature from the target data; and storing the target data, the first target feature, and the data score.

[0061] In this embodiment, the terminal can store data in different forms depending on the target storage strategy. For example, when the target storage strategy is the first storage strategy, the corresponding data value is low, so the terminal can store only the key data and features. For example, the terminal obtains target data greater than a data threshold from the aforementioned data. Here, target data refers to representative data, such as data that best represents the stable operating state of a transformer. The terminal extracts a first target feature from the target data, thereby the terminal can store the target data, the first target feature, and the data score to complete the storage of the data.

[0062] Through this embodiment, the terminal can store key data and features for low-value data, thereby reducing the waste of storage space resources in the terminal.

[0063] In one embodiment, extracting a first target feature from the target data includes: extracting time-domain features and frequency-domain features from the target data; and obtaining the first target feature based on the time-domain features and the frequency-domain features.

[0064] In this embodiment, the features extracted by the terminal from the target data can include various features. For example, for signal data related to the collected transformer operation information, the terminal can extract time-domain features and frequency-domain features from the target data of the signal data. Specifically, the terminal can extract time-domain features from the time-domain data of the target data and frequency-domain features from the frequency-domain data of the target data. Thus, the terminal can obtain the first target feature based on the time-domain features and the aforementioned frequency-domain features. Both the time-domain features and the frequency-domain features can be features obtained through fusion. For example, after the terminal performs multi-level feature extraction on the time-domain data, it obtains the time-domain features by fusing the multi-level features; similarly, after the terminal performs multi-level feature extraction on the frequency-domain data, it obtains the frequency-domain features by fusing the multi-level features.

[0065] Through this embodiment, the terminal can extract the first target feature of the target data from multiple perspectives, thereby improving the comprehensiveness of the extracted features.

[0066] In one embodiment, storing the data according to the target storage strategy includes: if the target storage strategy is a third storage strategy, acquiring first data for a preset time period before the data acquisition time and second data for a preset time period after the data acquisition time; extracting corresponding second target features based on the first data, the data, and the second data; fusing the second target features corresponding to each type of data to obtain fused target features; sequentially performing first quantization and second quantization on the fused target features to obtain quantized fused target features; the first quantization characterizes the quantization of the data contours of the first data, the data, and the second data; the second quantization characterizes the quantization of the data details of the first data, the data, and the second data; and storing the first data, the data, the second data, and the quantized fused target features.

[0067] In this embodiment, data with a target storage strategy of the third storage strategy is considered to have significant value. Therefore, the terminal can perform comprehensive event storage for this data. That is, the terminal can store the complete data, as well as historical data from the preceding and following periods.

[0068] For example, when the terminal detects that the target storage strategy is the third storage strategy, the terminal can obtain first data (historical data) for a preset time period before the data collection time, and second data (future data) for a preset time period after the data collection time. The terminal can extract the corresponding second target feature based on the first data, the second data, and the third data.

[0069] The terminal can collect various types of data corresponding to transformers, such as vibration, ultra-high frequency, and infrared temperature data. During storage, the terminal fuses the second target features corresponding to each type of data to obtain fused target features. These fused target features are then subjected to a first quantization (e.g., coarse quantization) and a second quantization (e.g., residual quantization) to obtain quantized fused target features. The first quantization represents the quantization of the data outlines of the first, second, and third data sets; for example, the terminal uses coarse quantization to map the data to coarser discrete values. The second quantization represents the quantization of the data details of the first, second, and third data sets; for example, the terminal uses residual quantization to perform hierarchical or dynamic quantization using the residuals between data sets, and iteratively clusters to gradually approximate the original data. Thus, the terminal can store the first, second, and quantized fused target features, achieving comprehensive event storage of the data.

[0070] Through this embodiment, when the target storage strategy is the third storage strategy, the terminal can realize data storage through event panoramic storage, thereby improving the comprehensiveness of high-value data storage.

[0071] In one embodiment, fusing the second target features corresponding to the various types of data to obtain fused target features includes: determining the relevance weights between the various types of data using a relevance learning network; the relevance weights characterizing the degree of relevance between the second target features corresponding to the various types of data; determining the independent weights corresponding to the various types of data using an importance evaluation network; the independent weights characterizing the importance of each of the second target features; determining the target weights of each of the second target features based on the relevance weights and the independent weights; and weighting and fusing the second target features based on their respective target weights to obtain fused target features.

[0072] In this embodiment, the terminal can collect various types of data corresponding to the transformer. When the storage strategy for each type of data belongs to the third storage strategy, the terminal can achieve compressed storage by weighted fusion of the various types of data. The various types of data include correlation weights and independence weights. The correlation weights characterize the degree of correlation between the second target features corresponding to each type of data; the independence weights characterize the importance of each of the aforementioned second target features.

[0073] Regarding relevance weights, the terminal can determine the relevance weights between the aforementioned data types using a relevance learning network. For example, the terminal can learn the relevance between different data types using a relevance learning network, and then determine the relevance weights between each data type. Regarding independent weights, the terminal can determine the independent weights corresponding to the aforementioned data types using an importance evaluation network.

[0074] Therefore, the terminal can determine the target weight of each of the aforementioned second target features based on the aforementioned correlation weight and the aforementioned independent weight. For example, the terminal can fuse the correlation weight and the independent weight to obtain the target weight. The terminal can also perform weighted fusion of the aforementioned second target features based on the aforementioned target weights corresponding to each of the aforementioned second target features to obtain the fused target features.

[0075] Through this embodiment, the terminal can determine the correlation weight and independent weight of various types of data, and then fuse and store the features of various types of data based on the correlation weight and independent weight, thereby reducing the data size when storing the entire event and improving the resource utilization of storage space.

[0076] In one exemplary embodiment, such as Figure 3 As shown, Figure 3 This is a schematic diagram of a data classification and storage system for transformers in one embodiment. The system adopts a layered distributed architecture design: the bottom layer consists of various sensor arrays, including vibration sensors, ultrasonic sensors, grounding current sensors, ultra-high frequency sensors, high frequency pulse current sensors, oil and gas monitoring units, and temperature and humidity sensors; the middle layer consists of three major functional modules, namely, an intelligent acquisition module, an edge intelligent processing module, and a classification and storage module; the top layer consists of a historical database and a remote monitoring center.

[0077] The system's data flow is as follows: sensors, intelligent acquisition module, edge intelligent processing module, classification and storage module, historical database / monitoring center. The control flow is a reverse closed loop: edge intelligent processing module, intelligent acquisition module, and simultaneously edge intelligent processing module and classification and storage module, forming a complete closed loop of perception, analysis, decision-making, and execution.

[0078] The terminal synchronously collects data from all sensors via an intelligent acquisition module. Through parallel acquisition, it preprocesses the data from each sensor to form a standardized data stream before transmitting it to the edge intelligent processing module. This solves the problem of data silos and provides high-quality, standardized raw materials for subsequent fusion analysis.

[0079] The terminal achieves multimodal fusion through a clock synchronization unit. For example, the terminal uses Time-Sensitive Networking (TSN), satellite timing technology, or the IEEE 1588 protocol to provide a unified global clock for all sensors.

[0080] The terminal uses a data preprocessing unit to package output data from sensors of different manufacturers and using different protocols into a unified data packet with precise timestamps. It can also receive control commands from the edge intelligence module through an adaptive acquisition and control unit to dynamically adjust its acquisition behavior.

[0081] The edge intelligence processing module is responsible for understanding data locally in real time, solving the response latency problem, performing real-time and high-precision diagnosis at the data source, and outputting instructive conclusions.

[0082] The terminal also achieves real-time edge computing through a layered lightweight diagnostic model architecture. The first layer of the layered lightweight diagnostic model is single-modal feature extraction, which uses a lightweight model to process different types of data in parallel and extract local feature patterns. The second layer is cross-modal feature fusion and diagnosis, which concatenates and fuses the high-level feature vectors extracted from each modality for analysis. Finally, the diagnostic results and confidence scores are output.

[0083] The categorized storage module is the system's memory system, responsible for saving valuable information, solving the problem of wasted storage resources, realizing intelligent management of the data lifecycle, and ensuring the maximization of storage space value.

[0084] The terminal receives diagnostic results and confidence levels from the edge intelligence module through the data value assessment unit, and calculates a dynamic value score (data score) accordingly.

[0085] The terminal, through a dynamic storage strategy execution unit, executes different storage strategies based on the data scoring: For low-value data, only extracted features, diagnostic results, and a short data snapshot are stored, using high compression rates and automatically overwritten after short-term storage; this is summary storage (first storage strategy), which saves significant space. For medium-value data, the original data is stored completely for medium- to long-term in-depth analysis and traceability; this is high-fidelity storage (second storage strategy). High-value data triggers a "red alert" level operation. Not only is the current data stored with the highest quality, but data from the past period (i.e., the process before the failure) and data from the future period (i.e., the failure development process) in the cache are also saved together, forming a complete event black box, which is permanently stored and prioritized for uploading; this is event panorama storage (third storage strategy), ensuring that no critical information is missed.

[0086] The feedback path can be divided into two parts: diagnostic-driven acquisition and classification-based storage. Diagnostic-driven acquisition involves sending the diagnostic conclusions from the edge intelligence module as control commands back to the adaptive control unit of the intelligent acquisition module, instructing it to increase the sampling rate of relevant sensors. Classification-based storage uses high-scoring data obtained from the value assessment unit, which can also be sent back as control commands to the intelligent acquisition module. For example, when triggering event-based panoramic storage, the terminal commands the acquisition module through the storage module to immediately activate the highest-rate acquisition mode for all sensors and clear the buffer, preparing for the storage of historical data.

[0087] The process by which the terminal classifies and stores the data of the aforementioned transformers can be as follows: Figure 4 As shown, Figure 4 This is a flowchart illustrating a data classification and storage method for transformers in another embodiment. The core of the storage mechanism is a value assessment unit, which receives diagnostic labels and confidence levels output by the intelligent diagnostic module and, combined with data characteristics, calculates a real-time data score representing the data's value.

[0088] The terminal obtains diagnostic results and confidence levels through a diagnostic model, and calculates a data score through a data value assessment unit. The data value assessment unit integrates the input diagnostic results and confidence levels into a logically sound and reliable value scoring calculation model, such as a multi-factor weighted assessment model, to obtain a data score.

[0089] The terminal can also determine the grading thresholds for data scoring. Using historical data backtracking and sample database analysis, the terminal identifies which data segments are critical fault data, useful analytical data, and discardable normal data. Then, it uses a data value assessment unit to calculate the data score for all main clauses and analyzes the distribution of these scores. Ideally, three relatively concentrated areas will be observed. The terminal categorizes these three areas into low, medium, and high value zones. Note that the aforementioned grading thresholds are not fixed, nor are they determined for every value assessment; rather, they are updated periodically.

[0090] When determining data scores, the terminal employs multiple tiers of assessment. For low scores (e.g., a diagnosis of normal status with high confidence): the terminal enables summary storage, storing only key feature data (the first target feature), the diagnosis result, and a snapshot (target data), using high compression and a short storage period. For medium scores (e.g., a diagnosis of gradual status change and a low-confidence warning): high-fidelity storage is enabled, storing the original data completely with a medium-to-long-term storage period. For high scores (e.g., a diagnosis of extremely high uncertainty, a fault alarm, and high confidence): the terminal enables event-wide storage, immediately storing the original data of all parameters at the highest sampling rate, backtracking cached data, and extending storage time to form a complete event record, marking it for permanent storage and uploading it with the highest priority. This transforms storage decisions from being based on fixed rules to being based on dynamic value assessment, achieving a fundamental shift from storing data to managing value, ensuring that storage resources are always prioritized for the most valuable data.

[0091] like Figure 5 As shown, Figure 5 This is a flowchart illustrating the monitoring steps for an oil-immersed transformer in one embodiment. Taking the monitoring of an oil-immersed transformer as an example, a core grounding current sensor, temperature and humidity sensor, noise sensor, oil temperature sensor, oil level sensor, vibration sensor, load voltage sensor, ultrasonic sensor, ultra-high frequency sensor, and high-frequency pulse current sensor are pre-installed on the oil-immersed transformer. All sensors are connected to the aforementioned terminal.

[0092] The terminal continuously and synchronously collects all parameters, with the data temporarily stored in a buffer. The terminal continuously runs a lightweight model for diagnosis via an edge intelligence module. At a certain moment, the model identifies an abnormal frequency component in the signal, but the confidence level is only 60% (data score). The terminal activates a high-fidelity storage mode via the classification storage module to fully record subsequent data. Simultaneously, the intelligent module sends instructions to the acquisition module to increase the sampling rate. A dozen minutes later, the oil and gas unit detects a sudden increase in acetylene (C2H2) content. The model's overall assessment is a high-energy discharge with a confidence level of 99% (data score is high). The terminal immediately triggers an event-wide storage mode via the storage module, completely storing the data from the past half hour (i.e., the initial stage of the fault) and data for the future period from the buffer into a secure storage area, and generating the highest-level alarm.

[0093] Taking the storage of vibration sensing signals from a transformer under normal operating conditions as an example, the time-domain graph of the vibration signal is as follows: Figure 6 As shown, Figure 6 This is a time-domain plot of the vibration signal in one embodiment. The frequency-domain plot of the vibration signal is shown below. Figure 7 As shown, Figure 7 This is a frequency domain diagram of a vibration signal in one embodiment. The terminal analyzes the data and finds the score to be low. Therefore, the terminal enables summary storage, storing only feature data (the first target feature), diagnostic results, and snapshots (target data), using a high compression ratio and a short storage period. Thus, for this vibration signal, only the key time domain (or time series) data features need to be saved. The key data features (the first target feature) of this vibration signal can be calculated during data analysis and fault diagnosis in the edge intelligent processing module. They are stored numerically and in a structured manner. Integers use TS-2DIFF differential encoding or Gorilla XOR encoding; floating-point numbers use precision-based power transform encoding, are converted to integers, compressed, and stored in memory. Compared to the original storage method, this storage space utilization is improved by 70%.

[0094] Among them, such as Figure 8 As shown, Figure 8 This is a time-domain plot of the target data in one embodiment. For feature extraction in the key time domain (target data), the terminal can use pre-set threshold analysis to obtain the signal in the key time domain for the portion of the signal exceeding the threshold. By only storing the signal in the key time domain, the terminal can effectively reduce the data volume of this segment from thousands of data points to hundreds of data points, effectively saving 40% of storage space. The terminal can extract signal features (time-domain features) from the key time domain.

[0095] For example, the terminal performs data preprocessing and basic waveform feature extraction. First, it eliminates linearly or slowly changing interference components introduced by temperature variations, sensor zero-point drift, etc. The process is as follows: The terminal processes the original signal x... raw(t) Perform a k-th order (usually 1st order) polynomial fitting to obtain the trend term x. trend The processed signal is: x(t) = x raw (t)-x trend (t). Then, the terminal uses the sliding window method to calculate the mean μ and standard deviation σ of the data within the window. Data points exceeding μ±3σ are considered outliers and corrected using the mean of neighboring points or linear interpolation.

[0096] The terminal can also extract basic time-domain statistical features. For example, the terminal describes the signal from the overall perspective of amplitude distribution. The mean is the DC component characterizing the signal, and ideally should be close to zero. The mean is calculated as follows: The root mean square (RMS) value reflects the average vibrational energy or intensity of a signal. The RMS value X rms的 The calculation is as follows: X rms =√ The peak value is the maximum absolute value of the signal and is highly sensitive to transient shocks. The peak value X... peak The calculation is as follows: X peak =max(|x i |). Peak-to-valley value X p-p It is the difference between the maximum and minimum values ​​of the signal, representing the overall fluctuation range of the signal. Its calculation method is as follows: X p-p =max(x i )-min(x i Skewness measures the asymmetry of the probability density distribution of signal amplitude. Positive skewness indicates the presence of a positive pulse with a large amplitude, and it is calculated as follows:

[0097] Skewness= .

[0098] Kurtosis measures the sharpness of a distribution curve and is extremely sensitive to impact characteristics. It is calculated as follows:

[0099] Kurtosis= .

[0100] The Impulse Factor can more effectively reveal the impulse component in a waveform.

[0101] Wherein, the peak factor CrestFactor=X peak / X rms ImpulseFactor= .

[0102] The terminal can also perform deep time-domain feature mining. Deep feature mining aims to reveal more complex structural patterns and dynamic characteristics within the signal. Based on these decomposed features, the terminal employs adaptive signal decomposition methods (such as ensemble empirical mode decomposition) to decompose the original signal into a series of intrinsic mode functions (IMFs) from high to low frequencies. The process is as follows: x(t) = IMF j (t)+r K (t), where IMF j (t) is the t-th intrinsic mode function, r K (t) represents the residual term. The first few high-frequency IMF components usually contain sensitive information about the winding mechanical state (such as impact and resonant modes). The key to feature reconstruction calculation lies in the statistical characteristics (such as kurtosis and root mean square value) of the IMF components (such as IMF1, IMF2), which constitute the feature set of the decomposition domain. For example, the kurtosis of IMF2 may be more sensitive than the kurtosis of the original signal. Permutation entropy is a method to measure the randomness and dynamic mutation of a time series. The calculation process is as follows: The terminal performs phase space reconstruction on the time series {x(i), i=1,2,...,N} to obtain the delay vector X(i): X(i)=[x(i),x(i+τ),…,x(i+(m-1)τ)]. The terminal arranges the elements in each vector X(i) in ascending order to obtain a set of symbol sequences (permutation patterns) π. The terminal calculates the frequency p(π) of each permutation pattern π. Calculate the permutation entropy H. p H p (m)=-∑p(π)lnp(π), the more regular the signal, the better H p The smaller the value, the more random and chaotic H becomes. p The larger the value, the greater the potential for winding faults to introduce nonlinearity, causing changes in its permutation entropy. Sample entropy measures the complexity of a time series and is calculated as follows: The probability that the distance between template vectors X(i) and X(j) is less than the tolerance r is defined as B. i m (r). Sample entropy is defined as: SampEn(m,r,N)=-ln[(B m+1 (r)) / (B m (r)). The higher the sequence self-similarity, the fewer the patterns, and the smaller the sample entropy. The terminal can construct high-dimensional feature vectors. For example, the terminal combines all the above features into a comprehensive feature vector F (temporal domain features) to fully characterize the vibration state. F=[X rms ,Skewness,Kurtosis,CrestFactor,IMF2 rms IMF2 Kurtosis H p SampEn].

[0103] The terminal can also use deep learning models to extract frequency domain features of vibration signals. Specifically, the terminal can perform signal segmentation and standardization. For example, the terminal can segment continuously acquired transformer vibration signals into sample segments of fixed length L: x=[x1,x2,…,x…]. L The terminal performs standardization on each sample segment as follows: x* i =(x i -μ) / σ, μ=1 / L x i , σ=√(1 / L (x i -μ) 2 The terminal utilizes frequency domain transformation and Fast Fourier Transform (FFT) to apply FFT to the standardized signal, transforming it into frequency domain data X(k): X(k) = x*(n)e -j2πkn / L Let k = 0, 1, ..., L-1. Extract the amplitude spectrum A(k) as a frequency domain feature: A(k) = |X(k)| = √(Re(X(k))) 2 +Im(X(k)) 2 Due to the symmetry of the spectrum, the first N is usually taken. Add 1 more frequency points to obtain the frequency domain feature vector A∈R N Next, logarithmic compression and normalization are performed at the terminal. To enhance the model's sensitivity to different amplitude features, logarithmic compression is applied: A log (k)=log(1+A(k)), and finally normalization is performed to obtain the final frequency domain input feature F. in .

[0104] The terminal can also extract frequency domain features based on deep learning using a one-dimensional convolutional neural network (1D-CNN). The terminal first inputs the normalized frequency domain amplitude spectrum F. in ∈R N×1 The first feature extraction layer is performed: local frequency domain pattern learning. The convolution operation uses C1 convolution kernels of width K1. The output of the j-th convolution kernel at frequency point m is: z j 1 (m)= w j 1 (k)F in (m+k)+b j 1 Next, nonlinear activation is performed, using the ReLU function to enhance the nonlinear expressive power, calculated as follows: a j 1 (m)=max(0,z j1 (m)) uses max pooling of width P1 to compress the feature dimensions while preserving the main frequency domain modes: p j 1 (n)= {a j 1 (nS1+r)}, where S1 is the pooling stride. Then, the terminal performs depthwise convolution: using C2 convolution kernels of width K2 to process multi-channel features: z j 2 (m)= w j 2 (c,k)p c 1 (m+k)+b j 2 Combining ReLU activation with Dropout regularization: a j 2 (m)=Dropout(max(0,z j 2 (m)),θ), where θ is the dropout rate. Finally, global average pooling is used to obtain the global statistical features in the frequency domain: g j =1 / T a j 2 (t), where T is the length of the time dimension.

[0105] The terminal performs frequency domain feature fusion on the aforementioned frequency domain features. Specifically, the terminal concatenates frequency domain features from different levels to form a rich feature representation (frequency domain features): f concat =[g 1 ⊕g 2 ⊕…⊕g M Where ⊕ represents the vector concatenation operation, and M is the number of feature layers. A frequency domain attention mechanism (optional enhancement) is introduced, adaptively weighting the important frequency components: α j =exp(v T tanh(Wg j +b)) / ( exp(v T tanh(Wg i +b))), f att = α j g j Finally, the frequency domain features f are obtained by adjusting and refining the feature dimensions through a fully connected layer. final :f final =Φ(W fconcat f att +b fconcat), where Φ is the final activation function (such as tanh or linear activation).

[0106] Among them, such as Figure 9 As shown, Figure 9 This is a flowchart illustrating the storage steps in one embodiment. Assume the storage module immediately triggers the event panorama storage mode, immediately storing the raw data of all parameters at the highest sampling rate, and backtracking the cache data while extending the storage time backward to form a complete event record, which is marked for permanent storage and uploaded with the highest priority. At this point, it is necessary to completely save the multimodal data. Here, the terminal can perform multimodal data storage. Specifically, the terminal treats the transformer's multimodal monitoring data (taking vibration Xv, ultra-high frequency Xu, and infrared temperature Xt as examples) as an organic whole, guiding the compression process through cross-modal correlation features, and utilizing a quantized embedded deep recovery network to reconstruct high-fidelity data from low bit-rate measurements.

[0107] For example, the terminal performs adaptive block partitioning of multimodal data: First, the terminal aligns all modal data on the time axis and performs normalization. Then, the data stream of each modality is divided into non-overlapping blocks. For example, the vibration signal Xv: sampling rate fs=51.2kHz, performs a short-time Fourier transform with a window length of 1024 and an overlap of 512, generating a time-frequency graph Sv(t,f), which is then divided into 32×32 non-overlapping image blocks. The ultra-high frequency signal Xu: extracts the pulse phase distribution map and also divides it into 32×32 blocks. The infrared temperature Xt: the temperature distribution map is directly divided into 32×32 blocks. Finally, all blocks are flattened into a vector I of length 1024. m (i, t).

[0108] The terminal can perform feature extraction. For example, the terminal passes each modality's data block through a lightweight convolutional encoder E. m Preliminary feature extraction is performed, mapping the data from 1024 dimensions to 256-dimensional latent features F. m (i, t). For example: Vibration encoder E v Using one-dimensional convolution with a kernel size of 7 and a stride of 2, the impact characteristics in vibration waveforms are captured. (Ultra-high frequency encoder E) u Using 2D convolution with a 3x3 kernel, the texture pattern of the discharge pulse is learned. Temperature encoder E t Using two-dimensional convolution with a kernel size of 5x5, we focus on the gradient distribution of the temperature field.

[0109] The terminal can also perform cross-modal feature fusion. Specifically, the terminal will use F... v F u F tMultimodal data features are concatenated along the feature dimension, and the weights of each modal feature are calculated through a cross-modal channel attention layer. The weight calculation formula is: α m =σ(MLP(GAP(F m In this context, GAP stands for Global Average Pooling, MLP stands for Multilayer Perceptron, and σ is the Sigmoid function. The joint features after weighted fusion (the fusion target features) are: F joint =α v F v ⊕α u F u ⊕α t F t , where ⊕ represents feature splicing.

[0110] The feature fusion process can be as follows: Figure 10 As shown, Figure 10 This is a flowchart illustrating the feature fusion steps in one embodiment. The terminal, based on the concept of deep canonical correlation analysis, achieves intelligent fusion of multimodal features through an end-to-end trainable network. It includes six core stages.

[0111] Taking vibration, ultra-high frequency, and temperature signals as examples, the terminal input in the input feature preprocessing stage includes feature maps of three modes: vibration feature F... v ∈R B×256×H×W UHF characteristics F u ∈R B×256×H×W and temperature characteristics F t ∈R B ×256×H×W Where B represents the batch size, 256 is the number of feature channels, and H×W is the spatial dimension. These features first enter a modality-specific encoder, which extracts unique feature representations for each modality through convolutional layers, batch normalization, and the PReLU activation function, while simultaneously enhancing important features and suppressing noise through a modality attention mechanism. Mathematically, this is expressed as: F ~ m =ModalEncoder m (F m )=LayerNorm(ResBlock(ModalAttn(Conv(F m form∈{u,v,t}. The modal attention mechanism combines channel attention and spatial attention to ensure that the features of each modality are optimally represented before fusion.

[0112] In the cross-modal correlation learning phase, the terminal will encode the features F. ~ v F ~ u F ~t The system then enters the cross-modal correlation learning network. This stage calculates the correlation weights between all mode pairs, including three sets of mode pairs: vibration-UHF, vibration-temperature, and UHF-temperature. For each mode pair (m,n), the correlation calculation process is as follows: C m,n =σ(Conv 3×3 (Conv 3×3 (Conv 3×3 (Concat(F ~ m ,F ~ n ))))). Here, σ represents the Sigmoid activation function, ensuring the output is in the range [0,1]. These local correlation maps are then processed by a global correlation aggregation network to generate global correlation weights w. m,n corr This captures the overall statistical dependencies between modalities.

[0113] During the adaptive weight generation phase, the terminal obtains cross-modal correlation information and also evaluates the independent importance of each modality. The independent weights for each modality are calculated using an importance evaluation network. The independent weights are represented as: I m =σ(Linear(ReLU(Linear(GAP(F ~ m ))))). Here, GAP represents global average pooling. Subsequently, the terminal adaptively fuses the independent importance and relevance weights to generate the final modality fusion weights (target weights): w m =αI m +(1-α)∑ n≠m w m,n corr Here, α is a learnable balancing parameter that allows the network to dynamically adjust the relative contributions of independent importance and correlation effects.

[0114] In the weighted feature fusion stage, after the terminal obtains the fusion weights for each modality, it performs weighted feature fusion: F fused =w v F ~ v ⊕w u F ~ u ⊕w t F ~ t Here, ⊕ represents the concatenation operation along the channel dimension. This weighted concatenation method ensures that modalities with higher information content occupy a more important position in the fused features, while preserving the complete information of all modalities.

[0115] During the feature fusion refinement stage, the terminal will initially fuse the feature Ffused ∈R B×384×H×W Entering the fusion refinement network, this network facilitates information exchange between features of different modalities through cross-modal interaction blocks: F interacted =Conv 1×1 (ReLU(Conv 3×3 (ChannelShuffle(F fused The channel rearrangement operation breaks down the isolation between modes, enabling full interaction of vibration, ultra-high frequency, and temperature features across the channel dimension. Finally, the features are further refined through a series of residual blocks: F joint =ResBlock(ResBlock(Conv 3×3 (F interacted The output is the joint feature F. joint ∈R B×256×H×W It contains complementary information from each modality and eliminates potential feature conflicts and redundancy through a refining process.

[0116] The terminal also includes a supervision and optimization mechanism. The entire fusion process is supervised by a correlation consistency loss to ensure that the learned correlation weights conform to the physical laws of transformer condition monitoring. The loss function is defined as: L corr =∑ m,n∈P ||w m,n corr -w m,n prior ||2 2 Where P represents the set of all modal pairs, w m,n prior It uses prior relevance weights set based on domain knowledge. This supervision mechanism ensures that the fusion process is not only data-driven but also has clear physical meaning.

[0117] The terminal can also perform layered quantization encoding. For example... Figure 11 As shown, Figure 11 This is a schematic diagram of the hierarchical quantization steps in one embodiment. The terminal includes a hierarchical quantization controller, which is divided into a coarse quantization layer and residual quantization. The coarse quantization layer... joint Perform 4-bit scalar quantization (coarse quantization) to map the feature values ​​to a 16-level codebook C. coarse Above. This preserves the main outline of the data. The second residual quantization layer calculates the residual after coarse quantization (residual quantization) R=F. joint -Q coarse (F joint Then, perform another 4-bit quantization on the residual R, resulting in codebook C. fine This captures detailed information about the data loss. Finally, the final data is output and stored, consisting of a coarse quantization index, a fine quantization index, and two codebooks (C).coarse and C fine Composition. Compared to the original data, the compression ratio can reach 32:1.

[0118] The terminal can also recover stored data. Specifically, the terminal uses quantization technology to split the recovery network, solving the inverse problem of achieving high-fidelity recovery from heavily quantized and compressed data. This can be achieved as follows: Figure 12 As shown, Figure 12 This is a flowchart illustrating the data recovery steps in one embodiment.

[0119] In this process, the terminal performs network initialization, decodes the stored quantization index using the codebook, and recovers the joint feature F. ~ joint This serves as input for network recovery.

[0120] The terminal can perform high-fidelity solutions: I k+1 = ||F ~ joint -ΦI||2 2 +μ / 2||Iw k ||2 2 This formula simulates the solution to the quadratic optimization problem using a lightweight convolutional subnetwork. This high-fidelity subnetwork consists of two convolutional layers: a first layer with a 1x1 kernel and a second layer with a 3x3 kernel, using the PReLU activation function.

[0121] The terminal solves for w through cross-modal attention priors. k+1 = ||wI k+1 ||2 2 +λR(w). The regularization term R(w) here is implemented by the CMA prior network. The CMA prior network structure is as follows: Input: High-fidelity solution output I k+1 (Current estimate), and other high-quality reconstruction results of the previous modalities are used as references. Then, a multi-head self-attention mechanism is used to compute feature associations within the current modality. Next, cross-attention is introduced, where the features of the current modality are used as the query, and the features of the reference modality are used as the key and value. This allows the network to query the temperature and UHF signal features at the same moment when reconstructing the vibration signal, and to find patterns of coordinated change. The final output is a feature w enhanced with cross-modal information. k+1 .

[0122] The terminal can also perform iterative network expansion. For example, the terminal sequentially connects the high-fidelity solution and the CMA prior module, expanding it into a 6-stage deep network. Each iteration is equivalent to an iteration of the optimization algorithm, progressively refining the reconstruction results. The final output is the reconstructed data block I*. minit That is, the recovered data.

[0123] Through the above embodiments, the terminal combines a diagnostic model to extract features and score data from transformer data, thereby enabling the use of different storage strategies for data with different scores, thus achieving the technical effect of reducing the waste of storage resources.

[0124] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0125] Based on the same inventive concept, this application also provides a transformer-oriented data classification and storage device for implementing the above-described transformer-oriented data classification and storage method. The solution provided by this device is similar to the implementation described in the above method; therefore, the specific limitations of one or more transformer-oriented data classification and storage device embodiments provided below can be found in the limitations of the transformer-oriented data classification and storage method described above, and will not be repeated here.

[0126] In one exemplary embodiment, such as Figure 13 As shown, a data classification and storage device for transformers is provided, including: an acquisition module 500, a diagnostic module 502, a determination module 504, and a storage module 506, wherein:

[0127] The acquisition module 500 is used to acquire the data to be stored collected from the transformer.

[0128] The diagnostic module 502 is used to input the above data into the trained diagnostic model; the diagnostic model is used to extract the features corresponding to the above data and output the corresponding data score based on the above features; the data score characterizes the degree of influence of the above data on the stable operation of the above transformer.

[0129] The determination module 504 is used to determine the target storage strategy corresponding to the above data based on the above data scoring.

[0130] Storage module 506 is used to store the above data according to the target storage strategy.

[0131] In one embodiment, the storage module 506 is configured to, if the target storage strategy is a first storage strategy, obtain target data greater than a data threshold from the data; extract a first target feature from the target data; and store the target data, the first target feature, and the data score.

[0132] In one embodiment, the storage module 506 is used to extract time-domain features and frequency-domain features from the target data; and to obtain the first target feature based on the time-domain features and the frequency-domain features.

[0133] In one embodiment, the storage module 506 is configured to, if the target storage strategy is a third storage strategy, acquire first data for a preset time period before the data acquisition time and second data for a preset time period after the data acquisition time; extract corresponding second target features based on the first data, the data, and the second data; fuse the second target features corresponding to each type of data to obtain fused target features; sequentially perform first quantization and second quantization on the fused target features to obtain quantized fused target features; the first quantization characterizes the quantization of the data contours of the first data, the data, and the second data; the second quantization characterizes the quantization of the data details of the first data, the data, and the second data; and store the first data, the data, the second data, and the quantized fused target features.

[0134] In one embodiment, the storage module 506 is configured to: determine the correlation weights between different types of data using a correlation learning network; the correlation weights characterize the degree of correlation between the second target features corresponding to different types of data; determine the independent weights corresponding to different types of data using an importance evaluation network; the independent weights characterize the importance of each of the second target features; determine the target weights of each of the second target features based on the correlation weights and the independent weights; and perform weighted fusion of each of the second target features based on the target weights corresponding to each of the second target features to obtain fused target features.

[0135] In one embodiment, the above-mentioned apparatus further includes: a feedback module, configured to send a target acquisition instruction to the data acquisition device corresponding to the transformer if the target storage strategy is a second storage strategy or a third storage strategy; the target acquisition instruction is used to instruct the data acquisition device to increase the data acquisition frequency for the data.

[0136] Each module in the aforementioned data classification and storage device for transformers can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0137] In one exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 14 As shown, the computer device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When executed by the processor, the computer program implements a data classification and storage method for transformers. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0138] Those skilled in the art will understand that Figure 14 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0139] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described data classification and storage method for transformers.

[0140] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described transformer-oriented data classification storage method.

[0141] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the above-described transformer-oriented data classification and storage method.

[0142] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0143] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0144] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0145] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A data classification and storage method for transformers, characterized in that, The method includes: Acquire the data to be stored collected from the transformer; The data is input into a trained diagnostic model; the diagnostic model is used to extract features corresponding to the data and output a corresponding data score based on the features; the data score characterizes the degree of impact of the data on the stable operation of the transformer. Based on the data score, determine the target storage strategy corresponding to the data; The data is stored according to the target storage strategy.

2. The method according to claim 1, characterized in that, The target storage strategy includes one or more of the following: a first storage strategy, a second storage strategy, and a third storage strategy; wherein: The data score of the first storage strategy is less than or equal to the first score threshold; The data score of the second storage strategy is greater than the first score threshold and less than or equal to the second score threshold; The data score of the third storage strategy is greater than the second score threshold.

3. The method according to claim 1, characterized in that, The step of storing the data according to the target storage strategy includes: If the target storage strategy is the first storage strategy, then obtain target data greater than the data threshold from the data; Extract the first target feature from the target data; The target data, the first target feature, and the data score are stored.

4. The method according to claim 3, characterized in that, Extracting the first target feature from the target data includes: Extract time-domain and frequency-domain features from the target data; The first target feature is obtained based on the time-domain features and the frequency-domain features.

5. The method according to claim 1, characterized in that, The step of storing the data according to the target storage strategy includes: If the target storage strategy is the third storage strategy, then the first data of the preset time period before the data collection time and the second data of the preset time period after the data collection time are obtained. Based on the first data, the data, and the second data, extract the corresponding second target features; The second target features corresponding to the data of each type are fused to obtain the fused target features; The fusion target features are sequentially subjected to a first quantization and a second quantization to obtain quantized fusion target features; the first quantization represents the quantization of the data contours of the first data, the data, and the second data; the second quantization represents the quantization of the data details of the first data, the data, and the second data. The first data, the data, the second data, and the quantized fused target features are stored.

6. The method according to claim 5, characterized in that, The step of fusing the second target features corresponding to the various types of data to obtain fused target features includes: A correlation learning network is used to determine the correlation weights between the data of each type; the correlation weights characterize the degree of correlation between the second target features corresponding to the data of each type. An importance assessment network is used to determine independent weights for each type of data; these independent weights characterize the importance of each of the second target features. The target weights for each of the second target features are determined based on the correlation weights and the independence weights. Based on the target weights corresponding to each second target feature, the second target features are weighted and fused to obtain the fused target features.

7. The method according to any one of claims 1 to 6, characterized in that, After determining the target storage strategy corresponding to the data based on the data score, the method further includes: If the target storage strategy is the second storage strategy or the third storage strategy, then a target acquisition instruction is sent to the data acquisition device corresponding to the transformer; the target acquisition instruction is used to instruct the data acquisition device to increase the data acquisition frequency for the data.

8. A data classification and storage device for transformers, characterized in that, The device includes: The acquisition module is used to acquire the data to be stored collected from the transformer. A diagnostic module is used to input the data into a trained diagnostic model; the diagnostic model is used to extract features corresponding to the data and output corresponding data scores based on the features; the data scores characterize the degree of impact of the data on the stable operation of the transformer. The determination module is used to determine the target storage strategy corresponding to the data based on the data score; A storage module is used to store the data according to the target storage strategy.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images