Transformer abnormity monitoring method and system based on artificial intelligence

Abstract

The invention provides a transformer abnormity monitoring method and system based on artificial intelligence, relates to the technical field of transformer state monitoring, and solves the technical problems of poor scene adaptability, multi-source data isolation and weak early warning in the prior art. The method comprises the following steps: acquiring real-time data and historical data of a transformer; preprocessing the real-time data through a preprocessing algorithm to obtain standard data; constructing an AI anomaly monitoring model based on historical data; wherein the AI anomaly monitoring model comprises a dynamic baseline module, an anomaly detection module and an anomaly classification module; and inputting the standard data into the AI anomaly monitoring model to generate an anomaly alarm. The method is used in the transformer abnormity monitoring process.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of transformer state monitoring, and in particular to a transformer anomaly monitoring method and system based on artificial intelligence. BACKGROUND

[0002] As the core equipment of the power grid, the operation state of the power transformer directly determines the power supply reliability of the power grid. The current traditional transformer anomaly monitoring method mainly relies on two types of means, namely fixed threshold judgment and artificial experience analysis. The fixed threshold judgment adopts the industry general standard and does not fully consider the individual differences of the equipment and the changes of the operation scene, which is easy to cause misjudgment or omission. The artificial experience analysis adopts a single parameter independent analysis mode, which cannot effectively mine the nonlinear correlation between data and is difficult to accurately identify complex anomalies. At the same time, the sensitivity of the two types of methods to weak anomalies is generally low, and they can only trigger an alarm when the fault develops to an obvious stage, which leads to the loss of the key operation and maintenance window period and easily causes power grid blackout accidents. SUMMARY

[0003] The present application provides a transformer anomaly monitoring method and system based on artificial intelligence, which solves the technical problems of poor scene adaptability, isolated multi-source data and weak early warning of the prior art.

[0004] To achieve the above-mentioned purpose, the present application adopts the following technical solutions: In a first aspect, a transformer anomaly monitoring method based on artificial intelligence is provided, comprising: obtaining real-time data and historical data of the transformer; wherein the real-time data includes scene data and parameter data of the transformer during operation, the scene data includes load rate, environmental temperature and equipment operation life, and the historical data includes normal data and abnormal data; preprocessing the real-time data by a preprocessing algorithm to obtain standard data; constructing an AI anomaly monitoring model based on the historical data; wherein the AI anomaly monitoring model includes a dynamic baseline module, an anomaly detection module and an anomaly classification module; inputting the standard data into the AI anomaly monitoring model to generate an anomaly alarm.

[0005] Based on the technical scheme, in the transformer abnormality monitoring method based on artificial intelligence provided in the application, real-time data and historical data of the transformer are fused, data quality is optimized by combining a preprocessing algorithm, and monitoring is carried out by relying on an AI abnormality monitoring model with a trinity of dynamic baseline, abnormality detection and abnormality classification, which can effectively break through the limitations of traditional monitoring: the dynamic baseline module can adapt to individual differences of equipment and different operating scenarios, greatly reducing false positives caused by scenario mismatch; the abnormality detection module can accurately capture weak abnormalities such as slight temperature rise and slow oil quality exceeding the standard, and greatly advance the early warning time to the fault implicit stage; the abnormality classification module can determine the type of abnormality, cooperate with real-time alarm output, help operation and maintenance personnel quickly lock the fault core, avoid blind troubleshooting, significantly improve fault handling efficiency, and ultimately reduce the risk of sudden power outages caused by transformer faults in the power grid.

[0006] In combination with the first aspect, in a possible implementation manner, the dynamic baseline module is constructed in the following manner: normal data is integrated into a training set; a gradient boosting tree is used to construct a training module, the training set is used to train the training module, and cross-validation is used to optimize hyperparameters of the gradient boosting tree to obtain a trained module; the trained module is corrected according to scene data; and finally, a dynamic baseline module is obtained, which has input including load rate, environmental temperature and equipment running time and output of dynamic normal ranges of each monitoring parameter; wherein the hyperparameters include learning rate and tree depth.

[0007] In combination with the first aspect, in a possible implementation manner, the correction of the training module according to the scene data includes: scene combination and sample extraction are performed on the normal data by using a three-dimensional cross-division method to obtain a scene data set; wherein the scene data set includes load rate, environmental temperature, equipment running time and corresponding monitoring parameter true values of the scene; the scene data set is input into the training module for prediction to obtain a predicted parameter range; deviation analysis is performed on the predicted parameter range and the monitoring parameter true values by using a performance evaluation index to obtain an abnormal scene list; wherein the performance evaluation index includes RMSE, prediction deviation rate and threshold coverage rate; the abnormal scene list is corrected by using a hierarchical correction strategy to obtain a corrected parameter; the training module is updated by using the corrected parameter to obtain a corrected training module.

[0008] In combination with the first aspect, in a possible implementation manner, the scene combination and sample extraction on the normal data by using the three-dimensional cross-division method include: interval division is performed on the load rate, the environmental temperature and the equipment running time to obtain interval division rules; The interval division rules are cross combined by a Cartesian product operation to obtain a subdivision scene ID; Normal data is extracted as a directed sample based on the subdivision scene ID to obtain a scene data set.

[0009] In combination with the first aspect, in a possible implementation manner, the anomaly detection module includes a self-encoder part, an isolation forest part, and a fusion judgment part. The self-encoder part: compresses feature of sample data by an encoder to obtain compressed feature; reconstructs the compressed feature by a decoder to obtain reconstructed feature; calculates reconstruction error of the reconstructed feature and the compressed feature by mean square error; wherein the sample data is any group of data in historical data. The isolation forest part: divides the compressed feature by a random division strategy to obtain a plurality of isolated features; calculates average distance between the sample data and the plurality of isolated features as path depth by distance measurement; normalizes the path depth to obtain an anomaly score. The fusion judgment part: sets a reconstruction error threshold and an anomaly score threshold, compares the reconstruction error and the anomaly score with the reconstruction error threshold and the anomaly score threshold respectively; when any one of the reconstruction error and the anomaly score is greater than the corresponding threshold, marks the sample data as abnormal candidate data.

[0010] In combination with the first aspect, in a possible implementation manner, the anomaly classification module is constructed in the following manner: Integrate the abnormal data and the abnormal candidate data as training data; construct a to-be-trained hybrid module based on a convolutional neural network and a long short-term memory network; train the to-be-trained hybrid module using the training data to obtain a hybrid module; iteratively optimize the hybrid module by an Adam optimizer; finally obtain an anomaly classification module with input of abnormal candidate data and output of abnormal type; wherein the anomaly classification module realizes classification by a Softmax function, the convolutional neural network is used to extract local features of physical signals, and the long short-term memory network is used to extract change features of time series data.

[0011] In combination with the first aspect, in a possible implementation manner, the input of the standard data into the AI anomaly monitoring model includes: Input the scene data into a dynamic baseline module to obtain a dynamic normal range; Compare the parameter data with the dynamic normal range; when the parameter data is out of the dynamic normal range, mark the parameter data as baseline deviation data; Input the baseline deviation data into the anomaly detection model to obtain abnormal candidate data; Input the abnormal candidate data into the anomaly classification module to obtain an abnormal type and generate an abnormal alarm.

[0012] In a possible implementation of the first aspect, the pre-processing of the real-time data by the pre-processing algorithm comprises: de-noising the real-time data by a de-noising algorithm to obtain de-noised data; completing missing values of the de-noised data by a linear interpolation method or a K-Nearest Neighbor (KNN) completion method to obtain completed data; normalizing the completed data by a normalization algorithm to obtain normalized data; aligning the normalized data based on a time granularity to obtain standard data.

[0013] In a possible implementation of the first aspect, the method for obtaining the completed data comprises: identifying missing values and determining time lengths of the de-noised data by a missing value detection algorithm to obtain missing sample segments; wherein the missing sample segments comprise parameter data and missing times; comparing the missing times with a preset time length threshold; when the missing time > the time length threshold, marking the missing sample segment as a long-time missing sample segment; when the missing time ≤ the time length threshold, marking the missing sample segment as a short-time missing sample segment; completing numerical values of the short-time missing sample segment by a linear interpolation method to obtain a short-time scene sample; performing a similar day data retrieval operation on the long-time missing sample segment in historical data by a similar day screening rule to obtain a plurality of similar day scene data; performing an optimal sample matching operation on the similar day scene data by a K-Nearest Neighbor (KNN) algorithm to obtain a plurality of similar day samples; completing numerical values of the long-time missing sample segment based on the plurality of similar day samples to obtain a long-time scene sample; performing a splicing operation on the short-time scene sample and the long-time scene sample based on timestamps to obtain the completed data.

[0014] In a second aspect, an electronic device is provided, comprising a communication unit and a processing unit; the communication unit is configured to obtain real-time data and historical data of a transformer; the processing unit is configured to pre-process the real-time data by a pre-processing algorithm to obtain standard data; construct an AI anomaly monitoring model based on the historical data; wherein the AI anomaly monitoring model comprises a dynamic baseline module, an anomaly detection module and an anomaly classification module; input the standard data into the AI anomaly monitoring model to generate an anomaly alarm.

[0015] In a third aspect, the present application provides an electronic device, comprising: a processor and a storage medium; the storage medium comprises instructions, and the processor is configured to execute the instructions to implement the method described in the first aspect and any possible implementation manner of the first aspect. The electronic device can be an electronic device or a chip in the electronic device.

[0016] In a fourth aspect, the present application provides an artificial intelligence-based transformer abnormality monitoring system, comprising: a data acquisition module, a calculation processing module, a man-machine interaction and alarm module, and a data storage module; wherein the data acquisition module is configured to acquire real-time data and historical data of the transformer; the calculation processing module is configured to pre-process the real-time data through a preprocessing algorithm to obtain standard data; and an AI abnormality monitoring model is constructed based on the historical data; wherein the AI abnormality monitoring model comprises a dynamic baseline module, an abnormality detection module, and an abnormality classification module; the standard data is input into the AI abnormality monitoring model to generate an abnormality alarm; the man-machine interaction and alarm module is configured to output abnormality alarm information and provide a local and remote data visualization interface to realize the interaction between an operation and maintenance personnel and the system; and the data storage module is configured to store the real-time data, historical operation data, and AI model files of the transformer to support data tracing and model iteration.

[0017] In a fifth aspect, the present application provides a computer-readable storage medium, which stores instructions, and when the instructions are executed on an electronic device, the electronic device executes the method described in the first aspect and any possible implementation manner of the first aspect.

[0018] In a sixth aspect, the present application provides a computer program product comprising instructions, and when the computer program product is executed on an electronic device, the electronic device executes the method described in the first aspect and any possible implementation manner of the first aspect.

[0019] The present application provides an artificial intelligence-based transformer abnormality monitoring method and system, which can effectively break through the limitations of traditional monitoring by fusing real-time data and historical data of the transformer, optimizing data quality by combining a preprocessing algorithm, and relying on a trinity AI abnormality monitoring model of dynamic baseline, abnormality detection, and abnormality classification to carry out monitoring. The dynamic baseline module can adapt to individual differences of equipment and different operation scenarios, greatly reducing false alarms caused by scene mismatch; the abnormality detection module can accurately capture weak abnormalities such as slight temperature rise and slow oil quality exceeding the standard, and greatly advance the early warning time to the fault implicit stage; the abnormality classification module can clearly identify the type of abnormality, cooperate with real-time alarm output, help operation and maintenance personnel quickly lock the fault core, avoid blind troubleshooting, significantly improve the fault handling efficiency, and ultimately reduce the risk of sudden power outage of the power grid caused by transformer failure.

[0020] It should be understood that the description of technical features, technical solutions, advantages or similar language in the present application does not imply that all features and advantages can be realized in any single embodiment. On the contrary, it can be understood that the description of a feature or advantage means that the specific technical feature, technical solution or advantage is included in at least one embodiment. Therefore, the description of technical features, technical solutions or advantages in the specification does not necessarily refer to the same embodiment. Further, the technical features, technical solutions and advantages described in the embodiments can be combined in any appropriate manner. Those skilled in the art will understand that the embodiments can be implemented without one or more specific technical features, technical solutions or advantages of the specific embodiments. In other embodiments, additional technical features and advantages can be identified in specific embodiments that do not embody all embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 A system architecture diagram of a transformer abnormality monitoring system based on artificial intelligence provided by an embodiment of the present application; Figure 2 A flowchart of a transformer abnormality monitoring method based on artificial intelligence provided by an embodiment of the present application; Figure 3 A flowchart of another transformer abnormality monitoring method based on artificial intelligence provided by an embodiment of the present application; Figure 4 A flowchart of another transformer abnormality monitoring method based on artificial intelligence provided by an embodiment of the present application; Figure 5 A structural schematic diagram of an electronic device provided by an embodiment of the present application; Figure 6 A hardware structural schematic diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0022] In the description of the present application, unless otherwise specified, " / " means "or", for example, A / B can mean A or B. "And / or" in this document is only a description of the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which can mean that A exists alone, A and B exist together, and B exists alone. In addition, "at least one" means one or more, and "multiple" means two or more. "First", "second", etc. do not limit the quantity and execution order, and "first", "second", etc. do not necessarily mean different.

[0023] It should be noted that in this application, the words "exemplary" or "for example" are used to mean serving as an example, instance, or illustration. Any embodiment or design described herein as "exemplary" or "for example" should not be construed as being preferred or advantageous over other embodiments or designs. Rather, the embodied words are used merely to present concepts in a concrete manner.

[0024] The transformer abnormality monitoring method based on artificial intelligence provided in the embodiments of the present application can be applied to the transformer abnormality monitoring system 100 based on artificial intelligence as shown in the figure. Figure 1 As shown in the figure, the communication system comprises a data acquisition module 10, a calculation processing module 20, a man-machine interaction and alarm module 30 and a data storage module 40. Figure 1

[0025] The data acquisition module 10 is configured to acquire real-time data and historical data of the transformer.

[0026] The calculation processing module 20 is configured to pre-process the real-time data through a pre-processing algorithm to obtain standard data, and construct an AI abnormality monitoring model based on the historical data, wherein the AI abnormality monitoring model comprises a dynamic baseline module, an abnormality detection module and an abnormality classification module; and input the standard data into the AI abnormality monitoring model to generate an abnormality alarm.

[0027] The man-machine interaction and alarm module 30 is configured to output the abnormality alarm information and provide a local and remote data visualization interface to realize the interaction between an operation and maintenance personnel and the system.

[0028] The data storage module 40 is configured to store the real-time data, historical operation data and AI model file of the transformer to support data tracing and model iteration.

[0029] To solve the technical problems of poor scene adaptability, isolated multi-source data and weak early warning in the prior art, the embodiments of the present application provide a transformer abnormality monitoring method based on artificial intelligence, which comprises the following steps: Acquiring real-time data and historical data of the transformer, wherein the real-time data comprises scene data and parameter data of the transformer during operation, the scene data comprises a load rate, an environmental temperature and a device operation life, and the historical data comprises normal data and abnormal data; Pre-processing the real-time data through a pre-processing algorithm to obtain standard data; Constructing an AI abnormality monitoring model based on the historical data, wherein the AI abnormality monitoring model comprises a dynamic baseline module, an abnormality detection module and an abnormality classification module; Inputting the standard data into the AI abnormality monitoring model to generate an abnormality alarm.

[0030] ​Therefore, the technical problems of poor scene adaptability, isolated multi-source data and weak early warning in the prior art are solved.

[0031] As Figure 2 shown, the transformer abnormality monitoring method based on artificial intelligence provided by the embodiments of the present application comprises: S201, acquiring real-time data and historical data of the transformer.

[0032] The real-time data comprises scene data and parameter data of the transformer during operation, the scene data comprises load rate, environmental temperature and equipment operation time, and the historical data comprises normal data and abnormal data.

[0033] It should be noted that the parameter data of the transformer during operation comprises temperature data, electrical quantity data, cooling system data and oil gas data, etc.; the scene data comprises load rate, environmental temperature and equipment operation time; and the historical data comprises parameter data and scene data during normal operation and operation abnormality.

[0034] For example, taking the monitoring of a 220kV oil-immersed power transformer in a substation as an example, when performing the S201 step, two types of data are synchronously acquired through the sensor network deployed on the transformer body and the periphery, the substation SCADA system and the equipment operation account: in the real-time data, the scene data is the environmental temperature of the current working environment of the transformer, the real-time load rate and the equipment operation time, and the parameter data is the real-time collected temperature data (winding hot spot temperature, top oil temperature, bottom oil temperature), electrical quantity data (three-phase load current, high and low voltage side voltage, active power and reactive power), cooling system data (fan start-stop state, oil pump running current, cooler inlet and outlet temperature difference) and oil gas data (oil dissolved gas component concentration); in the historical data, the parameter data is the historical record stored since the operation of the transformer, including the temperature, electrical quantity, cooling system and oil gas data during normal operation in the past years, and the abnormal data (such as three-phase current imbalance data when turn-to-turn short circuit occurred, oil ethylene concentration sudden rise data) occurred before, and the scene data is specifically the current real-time load rate, the current environmental temperature and the operation time of the transformer obtained from the equipment account.

[0035] S202, preprocessing the real-time data through a preprocessing algorithm to obtain standard data.

[0036] In some implementations, the preprocessing of the real-time data through the preprocessing algorithm comprises: de-noising the real-time data through a de-noising algorithm to obtain de-noised data; completing missing values of the de-noised data through a linear interpolation method or a K-nearest neighbor completion method to obtain completed data; The normalization algorithm is used to normalize the completed data to obtain normalized data. The normalized data is aligned based on the time granularity to obtain standard data.

[0037] For example, first, the real-time data of the transformer is processed to remove noise, among which the time series of the top oil temperature, three-phase load current and other data are filtered by moving average method to remove the instantaneous fluctuations caused by sensor device vibration, and the vibration amplitude data of the transformer body is removed by wavelet transform to remove the environmental noise generated by the operation of other devices in the substation, to obtain the denoised oil temperature, current, vibration and other data; then the integrity of the denoised data is checked, if it is found that the top oil temperature data in a certain period of time is missing for 8 minutes, the linear interpolation method is used to complete the missing value based on the effective oil temperature before and after the missing, if the dissolved gas analysis (DGA) data exists for 40 minutes due to temporary failure of the monitoring device, the historical DGA data of the transformer under the same load rate and environment temperature is referred to, and the K-nearest neighbor completion method is used to complete the missing value, to obtain the completed data without data gaps; then the completed data is normalized, the continuous data such as oil temperature (30-110℃) and three-phase current (200-1000A) are mapped to the interval [0,1] by Min-Max normalization, the dimensional difference of the gas concentration ratio data such as methane and ethylene in oil is eliminated by Z-score standardization, and the discrete state such as the cooling system "4 fans all on / 2 running / stop" is converted into a numerical vector by one-hot encoding, to obtain normalized data; finally, the normalized data is aligned according to the time granularity of "1 time / minute", the three-phase current data sampled at 10 seconds is taken as the minute average value, and the DGA data sampled at 1 hour is repeated and filled according to the minute, to finally obtain the standard data with unified time stamp and standardized format.

[0038] S203, constructing an AI anomaly monitoring model based on historical data.

[0039] The AI anomaly monitoring model includes a dynamic baseline module, an anomaly detection module and an anomaly classification module.

[0040] In some implementations, the dynamic baseline module is constructed in the following manner: The normal data is integrated into a training set; a to-be-trained module is constructed based on gradient boosting tree, the training set is used to train the to-be-trained module, the hyperparameters of the gradient boosting tree are optimized by cross-validation, a trained module is obtained; the trained module is corrected according to the scene data; finally, a dynamic baseline module is obtained, which inputs the load rate, environment temperature and device running time, and outputs the dynamic normal range of each monitoring parameter; wherein the hyperparameters include learning rate and tree depth.

[0041] For example, first, normal data (including winding hotspot temperature, top oil temperature, three-phase current, ethylene concentration in oil, and corresponding load rate, ambient temperature) without protection action and operation and maintenance abnormal record is screened from the transformer 5-year historical data, and the equipment running time of each data acquisition period is supplemented by combining with the equipment account to form a training set. Then, based on the gradient boosting tree algorithm, the training module is constructed, the load rate, ambient temperature, and equipment running time in the training set are used as input features, and the true value of each monitoring parameter is used as a label. The training set is used to iteratively train the training module, and 5-fold cross-validation is used to optimize the hyperparameters. Within the value range of the learning rate (0.01-0.1) and the tree depth (3-10), the optimal hyperparameters are determined by minimizing the mean square error, which are the learning rate 0.05 and the tree depth 6, and the preliminary training module is obtained. Then, typical scene data of the transformer (such as high load rate 90%+extreme high temperature 38℃+running time 12 years, low load rate 30%+low temperature 5℃+running time 5 years, etc.) is selected, the scene data is input into the preliminary training module, the deviation of the monitoring parameter range output by the module from the actual normal data in the scene is compared, and the training module is corrected by adjusting the feature weight (such as improving the importance of the “load rate” feature in the high load scene). Finally, the dynamic baseline module is obtained, which inputs the load rate, ambient temperature, and equipment running time, and outputs the dynamic normal range of each monitoring parameter such as the winding hotspot temperature (such as the dynamic range 105-118℃ in the high load high temperature scene) and the three-phase current unbalance degree (such as the dynamic range 0.5%-1.2% in the low load low temperature scene).

[0042] In some implementations, the anomaly detection module includes a self-encoder part, an isolation forest part, and a fusion judgment part. The self-encoder part: compresses the feature of the sample data through the encoder to obtain compressed features; reconstructs the compressed features through the decoder to obtain reconstructed features; and calculates the reconstruction error of the reconstructed features and the compressed features through the mean square error. The sample data is any group of data in the historical data. The isolation forest part: divides the compressed features through a random division strategy to obtain a plurality of isolated features; calculates the average distance between the sample data and the plurality of isolated features as the path depth through distance measurement; and normalizes the path depth to obtain an anomaly score. The fusion judgment part: sets a reconstruction error threshold and an anomaly score threshold, compares the reconstruction error and the anomaly score with the reconstruction error threshold and the anomaly score threshold respectively; and marks the sample data as abnormal candidate data when either of the reconstruction error and the anomaly score is greater than the corresponding threshold.

[0043] It should be noted that the fusion model of the autoencoder and the isolation forest is adopted, the historical data is taken as the input, and the multi-dimensional feature correlation mode of the transformer during normal operation is learned; the reconstruction error of the abnormal data is large, the isolation forest part identifies the samples deviating from the normal feature distribution, and the early weak abnormality is preliminarily identified.

[0044] For example, first, 5-year historical data of the transformer is selected as sample data, which includes normal operation data (such as load rate 60%-80%, winding temperature under environmental temperature 20-30℃, three-phase current, methane / ethylene concentration in oil, and multi-dimensional data) and early weak abnormality data (such as temperature increase of 2℃ in the early stage of turn-to-turn short circuit, and slow increase of ethylene concentration in oil from 15μL / L to 22μL / L); the autoencoder part takes these sample data as the input, the encoder compresses 18-dimensional original features (including 3 temperature data, 4 electrical quantity data, 5 DGA data, 2 vibration data, and 4 cooling system data) into 6-dimensional core compression features (such as “temperature-current cooperative feature” and “DGA-insulation state correlation feature”), and the decoder reconstructs the 6-dimensional compression features into 18-dimensional reconstruction features, and calculates the difference between the reconstruction features and the original features through mean square error (MSE); the reconstruction error of the normal data is mostly between 0.02 and 0.05, while the reconstruction error of the early weak abnormality data increases to 0.12-0.18 due to the deviation of the feature correlation from the normal mode; the isolation forest part constructs 100 isolated trees based on the above compression features, adopts a random division strategy (such as randomly selecting “winding temperature-current cooperative feature”, taking the mean value 0.6 of the feature in the normal data as the split value, and dividing the data into “≤0.6” and “>0.6” two groups), calculates the path depth of each sample data in all isolated trees and normalizes it to obtain the abnormal score; the abnormal score of the normal data is mostly between 0.2 and 0.4 due to the compliance with the feature distribution, and the abnormal score of the weak abnormality data is increased to 0.7-0.9 due to the deviation from the distribution; the fusion judgment part first sets the reconstruction error threshold to 0.08 based on the 3σ principle of the historical normal data, and sets the abnormal score threshold to 0.5 based on the 2σ principle, and then judges the real-time input sample data: if the reconstruction error of a certain data is 0.15 (>0.08) and the abnormal score is 0.82 (>0.5), or only the reconstruction error is greater than 0.08 and only the abnormal score is greater than 0.5, it is marked as abnormal candidate data.

[0045] In some implementations, the abnormal classification module is constructed in the following manner: Integrate the abnormal data and the abnormal candidate data into training data; construct a to-be-trained mixed module based on a convolutional neural network and a long short-term memory network; train the to-be-trained mixed module using the training data to obtain a mixed module; iteratively optimize the mixed module through an Adam optimizer; and finally obtain an abnormal classification module with the input being the abnormal candidate data and the output being an abnormal type; wherein the abnormal classification module realizes classification through a Softmax function, the convolutional neural network is used to extract local features of physical signals, and the long short-term memory network is used to extract change features of time-series data.

[0046] It should be noted that the convolutional neural network is used to extract local features of vibration and sound signals; the long short-term memory network is used to capture time-series change features of temperature, current and DGA data; the module adopts a Softmax activation function to realize multi-classification, minimizes a cross-entropy loss function through an Adam optimizer, and improves the classification accuracy.

[0047] For example, first, integrate the training data, combine the transformer historical abnormal data (such as the three-phase current imbalance + vibration sudden change data when the inter-turn short circuit occurs, the slow increase of water in oil + dielectric loss increase data when the insulation is damp, the oil temperature sudden rise + fan current to zero data when the cooling system fails) and the abnormal candidate data (such as early micro-abnormal data of temperature micro-increase + DGA gas ratio anomaly) output by the abnormal detection module, divide the training set and the validation set according to 8:2, and each data contains vibration waveform (10 seconds of local signal sampling), sound spectrum, 1 hour time sequence of temperature / current / DGA; Then construct the mixed module to be trained, the convolutional neural network (CNN) adopts 3-layer convolutional layer + pooling layer structure to extract local features of physical signals such as vibration waveform and sound spectrum (such as capturing 200Hz feature peak value of vibration signal when inter-turn short circuit occurs, abnormal noise frequency band when cooling fails), and the long short-term memory network (LSTM) adopts 2-layer bidirectional LSTM structure to capture the time sequence change characteristics of temperature (every 5 minutes, a total of 12 time steps), current and DGA data (such as the slow upward trend of 0.5ppm of water per week when the insulation is damp, the 3℃ sudden rise mode of temperature within 10 minutes when the inter-turn short circuit occurs); Then input the training data into the mixed module to be trained, the local features extracted by CNN and the time sequence features extracted by LSTM are fused through the full connection layer, and then the abnormal type probability distribution (such as inter-turn short circuit, insulation damp, cooling failure, normal fluctuation) is output through the Softmax function, the classification accuracy is evaluated by the validation set, and the parameters are adjusted in the reverse direction; Then, through the Adam optimizer (learning rate 0.001, attenuation coefficient 0.9), the cross-entropy loss function is minimized, and the iteration optimization is performed for 100 rounds until the loss converges (the accuracy of the training set is 92%, and the accuracy of the validation set is 88%); Finally, the abnormal classification module is obtained - when the abnormal candidate data (such as 200Hz peak value of vibration signal + 3℃ temperature rise within 10 minutes) is input, the module outputs “inter-turn short circuit” (probability 91%).

[0048] S204, input the standard data into the AI abnormal monitoring model to generate an abnormal alarm.

[0049] In some implementations, the inputting the standard data into the AI abnormal monitoring model, such as Figure 3 As shown, comprising: Inputting the scene data into the dynamic baseline module to obtain a dynamic normal range; Comparing the parameter data with the dynamic normal range; when the parameter data is out of the dynamic normal range, marking the parameter data as baseline deviation data; Inputting the baseline deviation data into the abnormal detection model to obtain abnormal candidate data; Inputting the abnormal candidate data into the abnormal classification module to obtain an abnormal type, and generating an abnormal alarm.

[0050] Exemplarily, the scene data (load rate 85%, ambient temperature 38℃, operation age 12 years) in the standard data is input into the dynamic baseline module to obtain the dynamic normal range of each monitoring parameter (for example, winding hot spot temperature 105-118℃, three-phase current unbalance degree 0.8%-1.5%, ethylene concentration in oil <20 μL / L); the parameter data (winding hot spot temperature 122℃, three-phase current unbalance degree 2.1%, ethylene concentration in oil 23 μL / L) in the standard data is compared with the dynamic normal range, and because the winding temperature and the current unbalance degree both exceed the corresponding range, the two parameter data are marked as baseline deviation data; the baseline deviation data is input into the anomaly detection module, and the reconstruction error (0.16) calculated by the autoencoder and the anomaly score (0.83) output by the isolation forest both exceed the set threshold, so the abnormal candidate data are obtained; the abnormal candidate data are input into the anomaly classification module, the convolutional neural network extracts the 200Hz feature peak in the vibration signal, the long short-term memory network captures the 5℃ sudden rising trend within 10 minutes of temperature, and the classification result (inter-turn short circuit probability 93%) is output by the Softmax function, and finally the abnormal alarm containing the abnormal type, the associated parameter and the confidence is generated.

[0051] Based on the above technical solution, the transformer abnormality monitoring method based on artificial intelligence provided in the application can effectively break through the limitations of traditional monitoring by fusing real-time data and historical data of the transformer, optimizing data quality by combining a preprocessing algorithm, and relying on an AI abnormality monitoring model with dynamic baseline, anomaly detection and anomaly classification. The dynamic baseline module can adapt to individual differences of equipment and different operating scenarios, and greatly reduces false positives caused by scene mismatch. The anomaly detection module can accurately capture weak abnormalities such as temperature rise and slow oil quality exceeding the standard, and greatly advance the early warning time to the fault implicit stage. The anomaly classification module can determine the abnormal type and cooperate with real-time alarm output to help operation and maintenance personnel quickly lock the fault core, avoid blind troubleshooting, significantly improve fault handling efficiency, and ultimately reduce the risk of sudden power outages caused by transformer faults in the power grid.

[0052] In a possible implementation manner of the embodiment of the application, S202 can be implemented through the following S301, which is specifically described as follows. S301, missing value completion is performed on the denoised data by a linear interpolation method or a K-nearest neighbor completion method to obtain completed data.

[0053] In some implementation manners, the completed data is obtained in the following manner. A missing value detection algorithm is used to identify and determine the time length of missing values in the denoised data to obtain a missing sample segment; wherein the missing sample segment includes parameter data and missing time. compare the missing time with a preset time length threshold; when the missing time > the time length threshold, mark the missing sample segment as a long-time missing sample segment; when the missing time ≤ the time length threshold, mark the missing sample segment as a short-time missing sample segment; complete the short-time missing sample segment by a linear interpolation method to obtain a short-time scene sample; perform a similar day data retrieval operation on the long-time missing sample segment in historical data by a similar day screening rule to obtain a plurality of similar day scene data; perform an optimal sample matching operation on the similar day scene data by a K-nearest neighbor algorithm to obtain a plurality of similar day samples; complete the long-time missing sample segment based on the plurality of similar day samples to obtain a long-time scene sample; perform a splicing operation on the short-time scene sample and the long-time scene sample based on a timestamp to obtain the completed data.

[0054] For example, by scanning the denoised data through the missing value detection algorithm (based on time series continuity verification), two missing sample segments are identified: one is the winding temperature data (14:05-14:10), missing for 5 minutes, containing parameter data (winding temperature) and missing time (5 minutes); the other is the ethylene concentration in oil data (16:20-17:00), missing for 40 minutes, containing parameter data (ethylene concentration in oil) and missing time (40 minutes). Compare the missing time with the preset duration threshold (30 minutes), because 5 minutes ≤ 30 minutes, the winding temperature missing segment is marked as a short-time missing sample segment; because 40 minutes > 30 minutes, the ethylene concentration in oil missing segment is marked as a long-time missing sample segment. For the short-time missing sample segment, linear interpolation is used to calculate the temperature values (108.2°C, 108.4°C, 108.6°C, 108.8°C, 109.0°C, respectively) for 14:05-14:10 based on the winding temperature of 108°C at 14:04 before the missing and the winding temperature of 110°C at 14:11 after the missing, to obtain a short-time scenario sample. For the long-time missing sample segment, 10 pieces of oil ethylene concentration data of similar days (normal operation days within 3 months that meet the conditions) are retrieved according to the similar day screening rule (matching the same load rate interval of 80%-90%, the same environment temperature of 35-40°C, and the same running age of 12 years); through the K-nearest neighbor algorithm (K=5, based on the Euclidean distance of load rate and environment temperature to calculate similarity), 5 most similar samples are matched from the 10 pieces of data, and the average ethylene concentration of 16:20-17:00 is 18.2 μL / L (calculated after removing the maximum value of 20.5 μL / L and the minimum value of 16.0 μL / L); the average value is used to complete the long-time missing segment to obtain a long-time scenario sample. Based on the timestamp, the short-time scenario sample (winding temperature for 14:05-14:10) and the long-time scenario sample (ethylene concentration in oil for 16:20-17:00) are spliced into the original denoised data to finally obtain the complete data without missing.

[0055] Based on the above technical solution, the missing segment is accurately identified and the duration is determined through the missing value detection, laying a foundation for differentiated completion; linear interpolation is used for short-time missing, which can effectively retain the time series continuity and change trend of the data (such as the smooth feature of slow temperature rise); for long-time missing, similar day screening is used to match historical data of the same scenario, combined with the K-nearest neighbor algorithm to select the optimal sample, and then the average value is calculated to complete the missing, which can greatly improve the adaptability of the completed value to the real scenario (such as parameter completion in high load and high temperature scenarios that are more consistent with actual working conditions); the completed data obtained through timestamp splicing not only reduces the abrupt error of short-time missing, but also reduces the scene mismatch risk of long-time missing, significantly improving data integrity and accuracy, providing high-quality data support for subsequent normalization processing and AI model input, and reducing model monitoring deviation caused by data missing.

[0056] In a possible implementation of the embodiment of the present application, S203 can be implemented by S401, which is specifically described as follows. S401, correcting the training module according to the scene data.

[0057] In some implementations, the correction of the training module according to the scene data includes: Figure 4 as shown, comprising: combining the normal data by a three-dimensional cross-division method and extracting samples to obtain a scene data set; wherein the scene data set includes a load rate, an environmental temperature, a device running time and corresponding monitoring parameter true values of a scene; inputting the scene data set into the training module for prediction to obtain a predicted parameter range; performing deviation analysis on the predicted parameter range and the monitoring parameter true values by a performance evaluation index to obtain an abnormal scene list; wherein the performance evaluation index includes RMSE, a prediction deviation rate and a threshold coverage rate; correcting the abnormal scene list by a hierarchical correction strategy to obtain a correction parameter; updating the training module by the correction parameter to obtain a corrected training module of the scene.

[0058] It should be noted that the hierarchical correction strategy corrects the training module by feature weight adjustment (for slight deviation), weighted incremental training (for moderate deviation) and scene sub-model fusion (for severe deviation) according to different deviation degrees; and the abnormal scene list is a scene and corresponding scene data set with deviation exceeding a standard.

[0059] Exemplary, the transformer 5-year historical normal data is processed by three-dimensional cross-division method: first, the load rate is divided into [0, 30%), [30%, 60%), [60%, 90%), [90%, 100%] four intervals, the ambient temperature is divided into [≤10℃], (10℃, 30℃], (30℃, 50℃] three intervals, the equipment running time is divided into [0, 5 years), [5 years, 10 years), [10 years, 20 years] three intervals, 36 sub-scenarios are obtained by Cartesian product operation; then based on the scene ID matching condition, the scene samples are extracted from the normal data to obtain the scene data set (for example, the "load rate 90%-100% + ambient temperature 30-50℃ + running time 10-20 years" scene contains 1200 samples, each containing corresponding load rate, ambient temperature, running time and winding temperature, oil ethylene concentration and other monitoring parameter true value). The scene data set is input into the training module for prediction to obtain the prediction parameter range of each scene (for example, the above high load high temperature old equipment scene, the predicted winding temperature range is 102-115℃, and the actual true value is concentrated in 105-118℃). The deviation analysis is carried out on the prediction range and the true value by the performance evaluation index: the RMSE of this scene is 4.2 (threshold 3.0), the prediction deviation rate is 12% (threshold 10%), and the threshold coverage rate is 88% (threshold 95%), because the multiple indicators are over standard, it is listed in the abnormal scene list (a total of 6 similar deviation over-standard scenes and corresponding data sets are marked). The hierarchical correction strategy is adopted for the abnormal scene list: for 1 mild deviation scene (prediction deviation rate 11%), the "ambient temperature" feature weight is adjusted (from 0.2 to 0.3); for 3 moderate deviation scenes (prediction deviation rate 15%-20%), weighted incremental training is adopted (300 samples of this scene are added, and the weight is set to 1.5 times of the normal sample); for 2 severe deviation scenes (prediction deviation rate > 20%), scene sub-model is constructed (training the data of this scene alone) and fused with the original model (fusion weight 0.7:0.3) to obtain correction parameters such as feature weight adjustment coefficient, incremental training weight and sub-model fusion weight. The training module is updated by the correction parameters, and finally the corrected training module of the scene is obtained.

[0060] Based on the above technical scheme, the three-dimensional cross division method is used to extract the accurate scene data set, which provides scene data support for subsequent correction. After the prediction parameter range is obtained by inputting the scene data set into the training module, the RMSE and prediction deviation rate are used to accurately locate the abnormal scene with excessive deviation between the predicted value and the true value. Then, the hierarchical correction strategy is used to solve different deviation problems (lightweight adjustment, moderate increment training, and severe sub-model fusion), and finally the corrected training module obtained through parameter updating can better adapt to different working condition scenes, greatly improve the prediction accuracy of monitoring parameters, effectively reduce the risk of false negative and false positive in subsequent abnormal monitoring caused by insufficient scene adaptation, and lay a foundation for reliable operation of the AI abnormal monitoring model.

[0061] In a possible implementation manner of the embodiment of the application, S401 can be implemented by S701, which is specifically described as follows. S701, scene combination and sample extraction are performed on the normal data by using the three-dimensional cross division method to obtain a scene data set.

[0062] In some implementation manners, the scene combination and sample extraction on the normal data by using the three-dimensional cross division method include: interval division is performed on the load rate, the environmental temperature, and the equipment running time to obtain interval division rules; cross combination is performed on the interval division rules by using Cartesian product operation to obtain a subdivided scene ID; directed sample extraction is performed on the normal data based on the subdivided scene ID to obtain a scene data set.

[0063] It should be noted that interval division is performed on the load rate, the environmental temperature, and the equipment running time in combination with industry standards and operation and maintenance experience of those skilled in the art; and the subdivided scene ID includes the coded representation of the load rate, the environmental temperature, and the equipment running time.

[0064] Exemplarily, first, in combination with DL / T572-2010 "Power Transformer Operation Regulations" and the experience of operation and maintenance personnel, three dimensions are divided into intervals: the load rate is divided into [0, 30%), [30%, 60%), [60%, 90%), [90%, 100%] according to "light load-medium load-heavy load-full load", corresponding to the codes 0-3; the ambient temperature is divided into [≤10℃], (10℃, 30℃], (30℃, 50℃] according to "low temperature-normal temperature-high temperature", corresponding to the codes 0-2; the equipment running time is divided into [0, 5 years), [5 years, 10 years), [10 years, 20 years] according to "newly put into operation-medium term-aging", corresponding to the codes 0-2, forming a clear interval division rule. Then, through Cartesian product operation, the interval division rules of the three dimensions are cross combined to generate 36 unique subdivided scene IDs, in the format of "load rate code-ambient temperature code-equipment running time code", for example, "3-2-2" represents the "full load (90%-100%)+high temperature (30-50℃)+aging equipment (10-20 years)" scene. Finally, based on the matching conditions of the subdivided scene ID (the data needs to meet the load rate, ambient temperature and running time intervals corresponding to the ID at the same time), the historical normal data is extracted for directional samples, such as extracting the winding temperature, three-phase current, gas in oil and other monitoring data that meet the conditions from the "3-2-2" scene, a total of 1500 valid samples are obtained, and finally a scene dataset containing 36 scenes, each scene with 800-2000 samples is formed, and each sample is associated with the corresponding load rate, ambient temperature, equipment running time and monitoring parameter true value.

[0065] Based on the above technical solution, the intervals are divided in combination with industry standards and operation and maintenance experience to ensure that they are in line with the actual working conditions; the subdivided scene ID containing codes is generated through Cartesian product to realize full working condition coverage without omission; the samples are extracted according to the ID to ensure data homogeneity; and finally the obtained scene dataset can accurately reflect the normal state of the transformer under different working conditions, providing high-quality data support for subsequent dynamic baseline construction and AI abnormal monitoring model training, and effectively reducing the misjudgment and missed judgment caused by insufficient scene adaptation.

[0066] The above describes the scheme of the embodiments of the present application mainly from the perspective of device implementation. It can be understood that each device, for example, an electronic device, includes at least one of a corresponding hardware structure and a software module for implementing each function. Those skilled in the art should easily realize that, in combination with the units and algorithm steps of the examples described in the embodiments disclosed herein, the present application can be implemented in the form of hardware or a combination of hardware and computer software. Whether a certain function is implemented in hardware or computer software driven hardware depends on the specific application and design constraints of the technical scheme. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0067] The embodiments of the present application can divide the functional units of the electronic device according to the above method examples. For example, each functional unit can be divided according to each function, or two or more functions can be integrated in one processing unit. The integrated unit can be implemented in the form of hardware or software functional unit. It should be noted that the division of units in the embodiments of the present application is illustrative, and is only a logical functional division. Actual implementation can have another division method.

[0068] In the case of using an integrated unit, Figure 5 A possible structure diagram of the electronic device (denoted as electronic device 50) involved in the above embodiments is shown, which includes a processing unit 501 and a communication unit 502, and can also include a storage unit 503. Figure 5 The structure diagram shown can be used to illustrate the structure of the electronic device involved in the above embodiments.

[0069] When Figure 5 When the structure diagram shown is used to illustrate the structure of the electronic device involved in the above embodiments, the processing unit 501 is used to control and manage the actions of the electronic device, the communication unit 502 is used for communication between the electronic device and other devices, and the storage unit 503 is used to store the program code and data of the electronic device.

[0070] For example, the communication unit 502 is configured to acquire real-time data and historical data of the transformer; The processing unit 501 is configured to pre-process the real-time data by a pre-processing algorithm to obtain standard data, and construct an AI anomaly monitoring model based on the historical data; the AI anomaly monitoring model includes a dynamic baseline module, an anomaly detection module and an anomaly classification module; the standard data is input into the AI anomaly monitoring model to generate an anomaly alarm.

[0071] The processing unit 501 can be a processor or a controller, and the communication unit 502 can be a communication interface, transceiver, transceiver circuit, transceiver device, etc. The term "communication interface" is a general term and may include one or more interfaces. The storage unit 503 can be a memory. When the electronic device 50 is a chip, the processing unit 501 can be a processor or a controller, and the communication unit 502 can be an input interface and / or an output interface, pins, or circuits, etc. The storage unit 503 can be a storage unit within the chip (e.g., a register, cache, etc.) or a storage unit located outside the chip (e.g., read-only memory (ROM), random access memory (RAM, etc.).

[0072] The communication unit can also be called a transceiver unit. The antenna and control circuit with transceiver functions in the electronic device 50 can be considered as the communication unit 502 of the electronic device 50, and the processor with processing functions can be considered as the processing unit 501 of the electronic device 50. Optionally, the device in the communication unit 502 used to implement the receiving function can be considered as the communication unit. The communication unit is used to execute the receiving steps in the embodiments of this application, and the communication unit can be a receiver, a receiver circuit, etc. The device in the communication unit 502 used to implement the transmitting function can be considered as the transmitting unit. The transmitting unit is used to execute the transmitting steps in the embodiments of this application, and the transmitting unit can be a transmitter, a transmitter, a transmitting circuit, etc.

[0073] Figure 5 If the integrated units in the process are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, in essence, or the parts that contribute to the prior art, or all or part of the technical solutions, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. Storage media for storing computer software products include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0074] Figure 5 The units in the process can also be called modules; for example, a processing unit can be called a processing module.

[0075] This application also provides a hardware structure diagram of an electronic device (denoted as electronic device 60), see [link to diagram]. Figure 6The electronic device 60 comprises a processor 601, and optionally further comprises a memory 602 connected with the processor 601.

[0076] In the first possible implementation, referring to Figure 6 The electronic device 60 further comprises a transceiver 603. The processor 601, the memory 602 and the transceiver 603 are connected through a bus. The transceiver 603 is used for communicating with other devices or communication networks. Optionally, the transceiver 603 can comprise a transmitter and a receiver. The device for realizing the receiving function in the transceiver 603 can be regarded as a receiver, and the receiver is used for executing the receiving steps in the embodiments of the present application. The device for realizing the sending function in the transceiver 603 can be regarded as a transmitter, and the transmitter is used for executing the sending steps in the embodiments of the present application.

[0077] Based on the first possible implementation, Figure 6 The structure diagram shown can be used for illustrating the structure of the electronic device involved in the above embodiments.

[0078] Among them, Figure 6 The system chip in the electronic device can also be illustrated. In this case, the actions performed by the above electronic device can be realized by the system chip, and the specific actions performed can be referred to in the above, and will not be described here again.

[0079] In the implementation process, each step in the method provided by the embodiment can be completed by the integrated logic circuit of the hardware in the processor or the instructions in the software form. The steps of the method disclosed in the embodiments of the present application can be directly embodied as the execution completed by the hardware processor, or executed by the combination of the hardware and the software modules in the processor.

[0080] The processor in the present application can include, but is not limited to, at least one of the following: a central processing unit (CPU), a microprocessor, a digital signal processor (DSP), a microcontroller unit (MCU), or an artificial intelligence processor, and various computing devices running software, each of which can include one or more cores for executing software instructions to perform operations or processing. The processor can be a separate semiconductor chip, or can be integrated with other circuits as a semiconductor chip, for example, it can be integrated with other circuits (such as coding and decoding circuits, hardware acceleration circuits, or various bus and interface circuits) to form a SoC (system on chip), or it can be integrated as a built-in processor in an ASIC. The ASIC integrated with the processor can be packaged separately or packaged together with other circuits. In addition to including cores for executing software instructions to perform operations or processing, the processor can further include necessary hardware accelerators, such as field programmable gate arrays (FPGAs), PLDs (programmable logic devices), or logic circuits that implement special logic operations.

[0081] The memory in the embodiments of the present application can include at least one of the following types: read-only memory (ROM) or other types of static storage devices that can store static information and instructions, random access memory (RAM) or other types of dynamic storage devices that can store information and instructions, and electrically erasable programmable read-only memory (EEPROM). In some scenarios, the memory can also be a compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compact discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, etc.), magnetic disk storage medium or other magnetic storage device, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and capable of being accessed by a computer, but not limited to.

[0082] The embodiments of the present application also provide a computer readable storage medium including instructions, which, when executed on a computer, cause the computer to perform any of the above methods.

[0083] The embodiments of the present application also provide a computer program product including instructions, which, when executed on a computer, cause the computer to perform any of the above methods.

[0084] The embodiments of the present application further provide a chip, comprising a processor and an interface circuit, the interface circuit being coupled with the processor, the processor being configured to run computer programs or instructions to implement the method described above, and the interface circuit being configured to communicate with other modules outside the chip.

[0085] In the above embodiments, the implementation can be achieved by software, hardware, firmware or any combination thereof, entirely or partially. When implemented by software, the implementation can be in the form of a computer program product, entirely or partially. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the entire or partial process or function described in the embodiments of the present application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another, for example, the computer instructions can be transferred from one website, computer, server or data center to another website, computer, server or data center through wired (such as coaxial cable, optical fiber, digital subscriber line (digital subscriber line, DSL)) or wireless (such as infrared, wireless, microwave, etc.) mode. The computer-readable storage medium can be any available medium that can be accessed by a computer or data storage device including one or more servers, data centers, etc. integrated with the medium. The available medium can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid state disk (solid state disk, SSD)), etc.

[0086] Although the present application is described herein in conjunction with various embodiments, other variations of the disclosed embodiments can be understood and implemented by those skilled in the art through viewing the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "one" does not exclude a plurality. A single processor or other unit can implement several functions listed in the claims. Some measures are described in mutually different dependent claims, but this does not mean that these measures cannot be combined to produce good results.

[0087] Although the present application has been described in connection with the preferred embodiments thereof with reference to the specific content thereof, it will be apparent to those skilled in the art that various modifications and changes can be made thereto without departing from the spirit and scope of the application. Accordingly, it is intended that the present application cover all such modifications and changes as fall within the scope of the application. It should be understood that various holidays and alterations can be made to the application disclosed in this specification without departing from the spirit or ambit of the present application. It is intended that the present application embrace all such alternates, modifications and fall within the scope of the claims accompanying this specification.

Claims

1. An artificial intelligence-based transformer abnormality monitoring method, characterized by, include: Acquire real-time and historical data of the transformer; wherein, the real-time data includes scene data and parameter data when the transformer is working, the scene data includes load rate, ambient temperature and equipment operating years, and the historical data includes normal data and abnormal data; Standard data is obtained by preprocessing real-time data using a preprocessing algorithm. An AI anomaly monitoring model is constructed based on historical data; wherein, the AI ​​anomaly monitoring model includes a dynamic baseline module, an anomaly detection module, and an anomaly classification module; Input standard data into the AI ​​anomaly monitoring model to generate anomaly alerts.

2. The method of claim 1, wherein, The construction method of the dynamic baseline module includes: Normal data is integrated into a training set; a module to be trained is constructed based on a gradient boosting tree, and the module to be trained is trained using the training set. Cross-validation is used to optimize the hyperparameters of the gradient boosting tree to obtain the training module; the training module is corrected according to the scenario data; finally, a dynamic baseline module is obtained with inputs including load rate, ambient temperature and equipment operating years, and outputs the dynamic normal range of each monitoring parameter; wherein, the hyperparameters include learning rate and tree depth.

3. The method of claim 2, wherein, The step of correcting the training module based on scene data includes: Normal data is combined into scenarios and samples are extracted using a three-dimensional cross-segmentation method to obtain a scenario dataset; wherein, the scenario dataset includes the scenario's load rate, ambient temperature, equipment operating years, and the corresponding actual values ​​of monitoring parameters; The scene dataset is input into the training module for prediction, and the range of prediction parameters is obtained. A list of abnormal scenarios is obtained by analyzing the deviation between the predicted parameter range and the actual value of the monitored parameters through performance evaluation indicators; wherein, the performance evaluation indicators include RMSE, prediction deviation rate and threshold coverage. The abnormal scenario list is targeted for correction using a hierarchical correction strategy to obtain correction parameters; The training module is updated by correcting the parameters to obtain the scene-corrected training module.

4. The method of claim 3, wherein, The method of scene combination and sample extraction of normal data through three-dimensional cross-partitioning includes: The load rate, ambient temperature, and equipment service life are divided into intervals to obtain the interval division rules; The interval division rules are cross-combined by Cartesian product operation to obtain the subdivided scene ID; Based on the segmented scene ID, targeted sample extraction is performed on normal data to obtain the scene dataset.

5. The method of claim 1, wherein, The anomaly detection module includes an autoencoder section, an isolated forest section, and a fusion determination section. The autoencoder part: compresses the sample data using an encoder to obtain compressed features; The compressed features are reconstructed using a decoder to obtain reconstructed features; the reconstruction error between the reconstructed features and the compressed features is calculated using mean squared error; wherein, the sample data is any set of data from historical data; The isolated forest part: compressed features are divided into several isolated features using a random partitioning strategy; the average distance between sample data and several isolated features is calculated using a distance metric and marked as path depth; the path depth is normalized to obtain anomaly scores; The fusion determination part sets a reconstruction error threshold and an anomaly score threshold, compares the reconstruction error and the anomaly score with the reconstruction error threshold and the anomaly score threshold respectively, and marks the sample data as abnormal candidate data when any one of the reconstruction error and the anomaly score is greater than the corresponding threshold.

6. The method of claim 5, wherein, The abnormal classification module is constructed in the following manner: The abnormal data and the abnormal candidate data are integrated into training data, a hybrid module to be trained is constructed based on a convolutional neural network and a long short-term memory network, the hybrid module to be trained is trained using the training data to obtain a hybrid module, and the hybrid module is iteratively optimized by an Adam optimizer to finally obtain an abnormal classification module with input of abnormal candidate data and output of abnormal types; the abnormal classification module is classified by a Softmax function, the convolutional neural network is used to extract local features of physical signals, and the long short-term memory network is used to extract change features of time series data.

7. The method of claim 1, wherein, The standard data is input into the AI anomaly monitoring model in the following manner: The scene data is input into the dynamic baseline module to obtain a dynamic normal range; The parameter data is compared with the dynamic normal range, and the parameter data is marked as baseline deviation data when the parameter data is out of the dynamic normal range; The baseline deviation data is input into the anomaly detection model to obtain abnormal candidate data; The abnormal candidate data is input into the abnormal classification module to obtain an abnormal type and generate an abnormal alarm.

8. The method of claim 1, wherein, The real-time data is preprocessed by a preprocessing algorithm in the following manner: The real-time data is denoised by a denoising algorithm to obtain denoised data; The denoised data is completed for missing values by a linear interpolation method or a K-nearest neighbor completion method to obtain completed data; The completed data is normalized by a normalization algorithm to obtain normalized data; The normalized data is aligned based on a time granularity to obtain standard data.

9. The method of claim 8, wherein, The completed data is obtained in the following manner: The missing sample segment is obtained by missing value detection and time length determination on the denoised data; the missing sample segment includes parameter data and a missing time; The missing time is compared with a preset time length threshold; when the missing time is greater than the time length threshold, the missing sample segment is marked as a long-time missing sample segment; when the missing time is less than or equal to the time length threshold, the missing sample segment is marked as a short-time missing sample segment; The short-time missing sample segment is completed for numerical values by a linear interpolation method to obtain a short-time scene sample; The long-time missing sample segment is searched for similar day data in historical data by a similar day filtering rule to obtain a plurality of similar day scene data; The similar day scene data is matched for optimal samples by a K-nearest neighbor algorithm to obtain a plurality of similar day samples; The long-time missing sample segment is completed for numerical values based on the plurality of similar day samples to obtain a long-time scene sample; The short-time scene sample and the long-time scene sample are spliced based on timestamps to obtain the completed data.

10. An electronic device, comprising: It comprises: A communication unit and a processing unit; The communication unit is configured to acquire real-time data and historical data of a transformer; The processing unit is configured to: pre-process the real-time data by a pre-processing algorithm to obtain standard data; construct an AI anomaly monitoring model based on historical data; wherein the AI anomaly monitoring model comprises a dynamic baseline module, an anomaly detection module, and an anomaly classification module; and input the standard data into the AI anomaly monitoring model to generate an anomaly alarm.

Images