Fault diagnosis method, device and equipment for transformer drying equipment and medium

Abstract

The invention discloses a fault diagnosis method and device for transformer drying equipment, equipment and a medium. The method comprises the steps of obtaining operation monitoring data; dividing the operation monitoring data into a plurality of data windows with fixed lengths, wherein each data window is used for representing a section of operation state sequence data of the transformer drying equipment; inputting the plurality of data windows into an LSTM-AE anomaly detection model for window reconstruction and anomaly judgment to obtain abnormal window data; and inputting the abnormal window data into a random forest fault classification model to carry out fault type identification so as to obtain a fault identification result. According to the invention, the fault detection accuracy can be improved, and the fault type can be quickly and accurately identified after the fault occurs.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of transformer drying, in particular to a transformer drying equipment fault diagnosis method, device, equipment and medium. BACKGROUND

[0002] The transformer general drying equipment is a device for removing moisture in the insulation material of a transformer to ensure the safe operation of the transformer. Such equipment is prone to various faults during operation due to environmental factors and complex working conditions.

[0003] In related technologies, the operation monitoring and maintenance of the transformer general drying equipment mostly rely on manual experience and fixed threshold alarms, which has the following problems: the accuracy of fault detection is low and it is difficult to determine the fault type. SUMMARY

[0004] The present application aims to provide a transformer drying equipment fault diagnosis method, device, equipment and medium, which can improve the accuracy of fault detection and quickly and accurately identify the fault type after the fault occurs.

[0005] In a first aspect, an embodiment of the present application provides a transformer drying equipment fault diagnosis method, comprising the following steps:

[0006] Obtain operation monitoring data, which is used to indicate the operation state of the transformer drying equipment;

[0007] Divide the operation monitoring data into a plurality of fixed-length data windows, each data window being used to represent a segment of operation state sequence data of the transformer drying equipment;

[0008] Input the plurality of data windows into an LSTM-AE anomaly detection model for window reconstruction and anomaly judgment to obtain abnormal window data;

[0009] Input the abnormal window data into a random forest fault classification model for fault type identification to obtain a fault identification result, which is used to indicate the fault type of the transformer drying equipment.

[0010] According to some embodiments of the present application, the operation monitoring data is divided into a plurality of data windows, comprising:

[0011] Impute the missing values of the operation monitoring data to obtain first preprocessed data;

[0012] Remove extreme abnormal values in the first preprocessed data to obtain second preprocessed data;

[0013] Denoise the time series of the second preprocessed data to obtain third preprocessed data;

[0014] standardizing the third preprocessed data to obtain fourth preprocessed data;

[0015] dividing the fourth preprocessed data into a plurality of data windows.

[0016] According to some embodiments of the present application, before the plurality of data windows are input into the LSTM-AE anomaly detection model for window reconstruction and anomaly judgment, the method further comprises:

[0017] calculating average instantaneous energy consumption of each drying stage in a drying process of the transformer drying equipment, the each drying stage including a preparation stage, a heating stage, a distillation stage, a pressure reduction stage, a high vacuum stage, and a vacuum breaking stage;

[0018] constructing a reconstruction loss function according to the average instantaneous energy consumption of each stage in the drying process of the transformer drying equipment;

[0019] constructing an LSTM autoencoder model according to the reconstruction loss function;

[0020] obtaining historical operation data of the transformer drying equipment;

[0021] training the LSTM autoencoder model according to the historical operation data to obtain an LSTM-AE anomaly detection model.

[0022] According to some embodiments of the present application, the plurality of data windows are input into the LSTM-AE anomaly detection model for window reconstruction and anomaly judgment, comprising:

[0023] obtaining historical operation data of the transformer drying equipment;

[0024] inputting the historical operation data into the LSTM-AE anomaly detection model to obtain reconstruction errors of the historical operation data;

[0025] calculating an anomaly threshold from the reconstruction errors of the historical operation data according to a Gaussian assumption + 3σ algorithm;

[0026] inputting the plurality of data windows into the LSTM-AE anomaly detection model for window reconstruction to obtain window reconstruction errors;

[0027] comparing the window reconstruction errors with the anomaly threshold for anomaly judgment to obtain anomaly window data.

[0028] According to some embodiments of the present application, the plurality of data windows are input into the LSTM-AE anomaly detection model for window reconstruction and anomaly judgment, comprising:

[0029] The reconstruction error of the LSTM-AE anomaly detection model is divided into a plurality of error sub-distributions according to each drying stage and a vacuum degree interval of a drying process of a transformer drying device, wherein the each drying stage includes a preparation stage, a heating stage, a distillation stage, a pressure reduction stage, a high vacuum stage, and a vacuum breaking stage, and the vacuum degree interval includes a medium-low vacuum interval and a high vacuum degree interval;

[0030] An extreme value fitting is performed on an upper tail of each error sub-distribution by using a generalized Pareto distribution, and an adaptive anomaly threshold is calculated;

[0031] The plurality of data windows are input into the LSTM-AE anomaly detection model for window reconstruction, and window reconstruction error is obtained.

[0032] The window reconstruction error is compared with the anomaly threshold to perform anomaly judgment, and anomaly window data is obtained.

[0033] According to some embodiments of the present application, before the anomaly window data is input into the random forest model to perform fault type identification and obtain a fault identification result, the following further includes:

[0034] Historical fault data of the transformer drying device is obtained.

[0035] The historical fault data is divided into a plurality of fault data windows, and each fault data window is marked with a corresponding fault type.

[0036] Each fault data window is converted into a feature vector of the corresponding fault type.

[0037] The random forest model is trained according to the feature vector, and a random forest fault classification model is obtained.

[0038] According to some embodiments of the present application, the random forest model is trained according to the feature vector, and a random forest fault classification model is obtained, including:

[0039] The random forest model is trained according to the feature vector, and a random forest initial classification model is obtained.

[0040] The average importance of a sensor corresponding to the historical fault data is calculated.

[0041] High-differentiation sensor features are screened according to the average importance of the sensor.

[0042] The random forest initial classification model is trained according to the high-differentiation sensor features, and a random forest fault classification model is obtained.

[0043] In a second aspect, the embodiments of the present application provide a fault diagnosis device of a transformer drying device, including:

[0044] a data acquisition module configured to acquire operation monitoring data, the operation monitoring data being used to indicate an operation state of the transformer drying device;

[0045] a data window module configured to divide the operation monitoring data into a plurality of data windows with a fixed length, each data window being used to represent a sequence of operation state data of the transformer drying device;

[0046] an anomaly detection module configured to input the plurality of data windows into an LSTM-AE anomaly detection model to perform window reconstruction and anomaly judgment, and obtain anomaly window data;

[0047] a fault type identification module configured to input the anomaly window data into a random forest fault classification model to perform fault type identification, and obtain a fault identification result, the fault identification result being used to indicate a fault type of the transformer drying device.

[0048] In a third aspect, an electronic device is provided, and the device includes a processor and a memory having stored computer program instructions;

[0049] The processor, when executing the computer program instructions, implements the fault diagnosis method of the transformer drying device according to the first aspect.

[0050] In a fourth aspect, a computer readable storage medium is provided, and the computer readable storage medium has stored computer program instructions, and the computer program instructions, when executed by a processor, implement the fault diagnosis method of the transformer drying device according to the first aspect.

[0051] The fault diagnosis method, device, equipment and medium of the transformer drying device according to the embodiments of the present application have at least the following beneficial effects:

[0052] In the embodiment, first, operation monitoring data is acquired; then, the operation monitoring data is divided into a plurality of data windows with a fixed length; then, the plurality of data windows are input into an LSTM-AE anomaly detection model to perform window reconstruction and anomaly judgment, and anomaly window data is obtained; finally, the anomaly window data is input into a random forest fault classification model to perform fault type identification, and a fault identification result is obtained. The present application can improve the accuracy of fault detection, and can quickly and accurately identify the fault type after the fault occurs.

[0053] Additional aspects and advantages of the present application will be in part apparent and in part pointed out hereinafter. BRIEF DESCRIPTION OF DRAWINGS

[0054] The present application will be further described below in conjunction with the accompanying drawings and embodiments, wherein:

[0055] Figure 1 A flowchart of an embodiment of the transformer drying device fault diagnosis method provided in the present application is shown in the figure;

[0056] Figure 2 A structural diagram of the transformer drying device fault diagnosis device provided in the present application is shown in the figure;

[0057] Figure 3 A structural diagram of the electronic device provided in the present application is shown in the figure. DETAILED DESCRIPTION

[0058] The features and exemplary embodiments of various aspects of the present application will be described in detail below, in order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain the present application, but not to limit the present application. The present application can be implemented without some of these specific details by those skilled in the art. The following description of the embodiments is only to provide a better understanding of the present application by showing examples of the present application.

[0059] In this document, relational terms such as first and second and the like can be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without more limitations, an element defined by the statement "comprising" does not exclude the existence of additional identical elements in the process, method, article, or apparatus that includes the element.

[0060] The transformer universal drying device is a device for removing moisture from the insulation material in the transformer to ensure the safe operation of the transformer. Such devices are prone to various faults during operation due to environmental factors and complex working conditions.

[0061] Specifically, the traditional drying equipment operation monitoring method mostly relies on manual experience and set threshold alarm, and has the following shortcomings: (1) only significant abnormalities or failures can be identified, early micro abnormalities cannot be detected in time and effectively, and the failure is often discovered after the failure expands; (2) the detection means based on fixed threshold or simple statistical method is difficult to handle the complex nonlinear and time-varying characteristics in the equipment operation, and the false positive rate and the false negative rate are relatively high; (3) there is a lack of automatic classification and diagnosis mechanism for abnormalities, so that it is difficult to accurately determine the fault type after the failure occurs, and the maintenance difficulty and cost are increased.

[0062] In summary, in the related art, the operation monitoring and maintenance of the transformer general drying equipment mostly rely on manual experience and fixed threshold alarm, and the following problems exist: the accuracy of fault detection is low and it is difficult to determine the fault type.

[0063] To solve the problems in the prior art, the embodiments of the present application provide a transformer drying equipment fault diagnosis method, device, equipment and medium. First, the transformer drying equipment fault diagnosis provided by the embodiments of the present application is introduced.

[0064] Figure 1 A flowchart of a transformer drying equipment fault diagnosis method provided by the embodiments of the present application is shown. The method is applied to an electronic device.

[0065] A transformer drying equipment fault diagnosis method, comprising the following steps:

[0066] S101, obtaining operation monitoring data, the operation monitoring data being used to indicate the operation state of the transformer drying equipment;

[0067] S102, dividing the operation monitoring data into a plurality of fixed-length data windows, each data window being used to represent a segment of operation state sequence data of the transformer drying equipment;

[0068] S103, inputting the plurality of data windows into an LSTM-AE anomaly detection model to perform window reconstruction and anomaly judgment, and obtaining abnormal window data;

[0069] S104, inputting the abnormal window data into a random forest fault classification model to perform fault type identification, and obtaining a fault identification result, the fault identification result being used to indicate the fault type of the transformer drying equipment.

[0070] In this embodiment, first, the operation monitoring data is acquired; then the operation monitoring data is divided into a plurality of fixed-length data windows; then the plurality of data windows are input into an LSTM-AE anomaly detection model for window reconstruction and anomaly judgment to obtain abnormal window data; finally, the abnormal window data is input into a random forest fault classification model for fault type identification to obtain a fault identification result. The application can improve the accuracy of fault detection, and can quickly and accurately identify the fault type after the fault occurs.

[0071] The acquisition of the operation monitoring data in the above step S101 refers to real-time collection of multi-dimensional time sequence data such as temperature, vacuum degree, flow rate, frequency and the like in the device operation process through a multi-sensor system built in the transformer general drying device. Through these data, the operation state of the device can be reflected.

[0072] It should be noted that the operation monitoring data described above can be raw data of a sensor, or operation monitoring data after preprocessing.

[0073] The division of the operation monitoring data into a plurality of fixed-length data windows in the above step S102 refers to division of the operation monitoring data into data windows with a fixed length of N through a sliding window method. Each window represents a sequence of operation state data of the transformer general drying device, and is used for subsequent anomaly detection and fault diagnosis model construction.

[0074] Specifically, if the operation monitoring data is raw data of a sensor, the operation monitoring data can be preprocessed, such as filtering and normalization, before being divided into a plurality of fixed-length data windows.

[0075] In the above step S103, the plurality of data windows are input into an LSTM-AE anomaly detection model for window reconstruction and anomaly judgment to obtain abnormal window data. This refers to window reconstruction of the input data window through the trained LSTM-AE anomaly detection model to obtain a window reconstruction error, and then comparing the window reconstruction error with an anomaly threshold. If the window reconstruction error exceeds the threshold, it is determined to be abnormal, and potential faults or working condition deviations can be identified in time.

[0076] It should be noted that the LSTM-AE anomaly detection model refers to a trained LSTM autoencoder model, through which window reconstruction and anomaly judgment can be achieved.

[0077] Specifically, the specific structure of the LSTM autoencoder model includes an LSTM encoding layer and an LSTM decoding layer. The main task of the LSTM encoding layer is to convert the input sequence data into a fixed-length vector representation. This vector captures the key information and patterns in the input sequence. The task of the LSTM decoding layer is to reconstruct the original input sequence or generate a new sequence based on the fixed-length vector generated by the LSTM encoding layer. In the model structure, the LSTM layer at the encoding end gradually extracts the time correlation of multi-dimensional features within the window and obtains a compressed representation. The representation is then copied through the RepeatVector layer to provide the LSTM at the decoding end for layer-by-layer reconstruction of the original data. Finally, the TimeDistributed(Dense(...)) time distribution fully connected layer outputs the restored sequence.

[0078] Specifically, the abnormal threshold can be obtained by the extreme value theory or the Gaussian assumption + 3σ algorithm. The extreme value theory refers to an algorithm that uses a generalized Pareto distribution to fit the upper tail of each error sub-distribution. In an example, the present embodiment uses the Optuna algorithm to automatically search and evaluate the key hyperparameters (such as the number of two-layer LSTM units, the learning rate, the dropout rate, the batch size, etc.) of the LSTM autoencoder. In combination with the extreme value theory, the abnormal threshold is obtained from the reconstruction error distribution of the normal data window. The data window is input into the trained LSTM-AE anomaly detection model. If the window reconstruction error exceeds the abnormal threshold, it is determined to be abnormal. The data window is determined to be an abnormal window data.

[0079] In the above step S104, the abnormal window data is input into the random forest fault classification model for fault type identification to obtain a fault identification result. This means that when step S103 determines that a certain data window is abnormal through the LSTM-AE anomaly detection model, the data window is input into the random forest fault classification model as abnormal window data. The random forest fault classification model refers to a trained random forest model that can automatically determine the fault category of the abnormal window data.

[0080] In some embodiments, dividing the operation monitoring data into multiple data windows can include:

[0081] The missing values of the operation monitoring data are imputed to obtain first preprocessed data;

[0082] The extreme abnormal values in the first preprocessed data are removed to obtain second preprocessed data;

[0083] The time series of the second preprocessed data are denoised to obtain third preprocessed data;

[0084] The third preprocessed data are standardized to obtain fourth preprocessed data;

[0085] The fourth pre-processed data is divided into multiple data windows.

[0086] In this embodiment, first, the missing values of the operation monitoring data are interpolated to obtain first pre-processed data; then, extreme abnormal values in the first pre-processed data are removed to obtain second pre-processed data; then, the time series of the second pre-processed data are denoised to obtain third pre-processed data; then, the third pre-processed data are standardized to obtain fourth pre-processed data; finally, the fourth pre-processed data are divided into multiple data windows. The reliability and accuracy of the abnormality judgment can be further improved.

[0087] The interpolation of the missing values of the operation monitoring data refers to the interpolation of the missing values in the operation monitoring data by a linear interpolation method to ensure the integrity of the time series.

[0088] The removal of the extreme abnormal values in the first pre-processed data refers to the removal of the extreme abnormal values observed by using an expert knowledge method to remove extreme sampling points caused by sensor collection or system interference. The removal by using the expert knowledge method refers to manual removal based on experience.

[0089] The denoising of the time series of the second pre-processed data refers to the denoising of the time series by an Unscented Kalman Filter (UKF) to restore a more true state distribution and avoid phase delay and over-smoothing caused by simple filtering.

[0090] The standardization of the third pre-processed data refers to the Z-score standardization of the processed data to compare and learn different sensor features on the same scale, thereby improving the robustness and accuracy of the subsequent LSTM auto-encoder model and random forest classification model.

[0091] In some embodiments, before the multiple data windows are input into the LSTM-AE abnormality detection model for window reconstruction and abnormality judgment, the method can further include:

[0092] The average instantaneous energy consumption of each drying stage in the drying process of the transformer drying equipment is calculated, and the each drying stage includes a preparation stage, a heating stage, a distillation stage, a pressure reduction stage, a high vacuum stage, and a vacuum breaking stage.

[0093] A reconstruction loss function is constructed according to the average instantaneous energy consumption of each stage in the drying process of the transformer drying equipment.

[0094] An LSTM auto-encoder model is constructed according to the reconstruction loss function.

[0095] The historical operation data of the transformer drying equipment are obtained.

[0096] The LSTM-AE anomaly detection model is obtained by training the LSTM auto-encoder model according to historical operation data.

[0097] In this embodiment, first, the average instantaneous energy consumption of each drying stage in the drying process of the transformer drying device is calculated, then a reconstruction loss function is constructed according to the average instantaneous energy consumption of each stage in the drying process of the transformer drying device, then an LSTM auto-encoder model is constructed according to the reconstruction loss function, then historical operation data of the transformer drying device is obtained, and finally the LSTM auto-encoder model is trained according to the historical operation data to obtain the LSTM-AE anomaly detection model. The accuracy of anomaly judgment can be further improved.

[0098] It should be noted that the process of the transformer general drying device can be divided into six stages: "preparation P, heating H, distillation D, pressure reduction PR, high vacuum FV, and vacuum breaking E". Therefore, the above calculation of the average instantaneous energy consumption of each drying stage in the drying process of the transformer drying device refers to the calculation of the average instantaneous energy consumption of each stage, which is as follows:

[0099] Assuming that the drying process is divided into K stages (K=6 in this embodiment: 1 preparation stage-2 heating stage-3 intermediate distillation stage-4 pressure reduction stage-5 high vacuum stage-6 vacuum breaking stage), the average instantaneous energy consumption of each stage is as follows:

[0100]

[0101] where k∈{1,2,3,4,5,6} is the number of each stage, P(t) is the real-time power of the stage (calculated from the current and voltage of the device electric box), stagek is the set of all time points contained in the kth stage; N k is the number of sampling points of the kth stage.

[0102] The above construction of the LSTM auto-encoder model according to the reconstruction loss function refers to the normalization of the average instantaneous energy consumption of each stage to a weight, and then the weight is added to the reconstruction loss function of the LSTM auto-encoder model for training.

[0103] The above acquisition of the historical operation data of the transformer drying device refers to the acquisition of the multi-dimensional time series data of the historical operation of the transformer drying device through a database or other means.

[0104] The LSTM-AE anomaly detection model is obtained by training the LSTM auto-encoder model according to historical operation data, which means that the LSTM-AE anomaly detection model capable of accurately reconstructing a window and judging an anomaly is obtained by training the LSTM auto-encoder model constructed above according to the historical operation multi-dimensional time series data of the transformer drying device.

[0105] Specifically, in the training stage, the preprocessed multi-dimensional time series data is divided into windows according to a fixed time step (for example, 30 data for a window and 10 data for a sliding step), and is input into the auto-encoder (LSTM-AE) containing two layers of LSTM encoding layers and two layers of LSTM decoding layers.

[0106] In the model structure, the LSTM layer at the encoding end gradually extracts the time correlation of the multi-dimensional features in the window and obtains a compressed representation, which is copied through the RepeatVector layer for the LSTM at the decoding end to gradually reconstruct the original data, and finally the TimeDistributed(Dense(...)) outputs the restored sequence.

[0107] In some embodiments, inputting the plurality of data windows into the LSTM-AE anomaly detection model to reconstruct a window and judge an anomaly can include:

[0108] Obtaining historical operation data of the transformer drying device;

[0109] Inputting the historical operation data into the LSTM-AE anomaly detection model to obtain reconstruction error of the historical operation data;

[0110] Calculating an anomaly threshold from the reconstruction error of the historical operation data according to a Gaussian assumption + 3σ algorithm;

[0111] Inputting the plurality of data windows into the LSTM-AE anomaly detection model to reconstruct a window to obtain window reconstruction error;

[0112] Comparing the window reconstruction error with the anomaly threshold to judge an anomaly to obtain abnormal window data.

[0113] In this embodiment, first, the historical operation data of the transformer drying device is obtained, and then the historical operation data is input into the LSTM-AE anomaly detection model to obtain reconstruction error of the historical operation data; then, an anomaly threshold is calculated from the reconstruction error of the historical operation data according to a Gaussian assumption + 3σ algorithm, the plurality of data windows are input into the LSTM-AE anomaly detection model to reconstruct a window to obtain window reconstruction error, and the window reconstruction error is compared with the anomaly threshold to judge an anomaly to obtain abnormal window data. The accuracy of the anomaly judgment can be further improved.

[0114] The historical operation data of the transformer drying device refers to multi-dimensional time series data of historical operation of the transformer drying device obtained through a database or other means.

[0115] The abnormal threshold is calculated from the reconstruction error of the historical operation data according to the Gaussian assumption + 3σ algorithm, and the specific principle is: based on the central limit theorem, if the reconstruction error sequence is approximately unimodal and symmetrically distributed under normal working conditions, it can be considered that its probability distribution tends to be Gaussian distribution N(mu, sigma 2 ). At this time, the probability of samples exceeding the mean value by 3σ is only about 0.27%, which can be considered as a rare anomaly. Therefore, in this embodiment, the abnormal threshold T is set as:

[0116] T = mu + 3sigma;

[0117] That is, the normal and abnormal operation of the transformer drying device can be distinguished at a confidence level of 99.7%.

[0118] Specifically, after calculating the abnormal threshold, the LSTM-AE anomaly detection model needs to be trained according to the abnormal threshold so that it can accurately judge the anomaly. The specific implementation steps are as follows:

[0119] 1) Training phase: for the training set containing only normal samples, the reconstruction error epsilon i of each sliding window is calculated The error mean mu and standard deviation sigma can be obtained from the following formula, from which a single global threshold T can be obtained. The specific formula is as follows:

[0120]

[0121] 2) Staged expansion: the drying process is divided into multiple stages, and mu k , sigma k can be calculated for each stage respectively, so as to obtain stage-specific thresholds.

[0122] T k = mu k + 3sigma k , k = 1, 2,..., K;

[0123] 3) Detection phase: for any to-be-detected sliding window j, the reconstruction error epsilon j is calculated. If epsilon j <= T, it is normal; otherwise, it is marked as abnormal.

[0124] In some embodiments, inputting multiple data windows into the LSTM-AE anomaly detection model for window reconstruction and anomaly judgment can include:

[0125] The reconstruction error of the LSTM-AE anomaly detection model is divided into multiple error sub-distributions according to the drying stages and the vacuum degree intervals of the transformer drying device drying process, wherein each drying stage includes a preparation stage, a heating stage, a distillation stage, a pressure reduction stage, a high vacuum stage, and a vacuum breaking stage, and the vacuum degree intervals include a medium-low vacuum interval and a high vacuum degree interval.

[0126] The upper tail of each error sub-distribution is fitted with a generalized Pareto distribution to calculate an adaptive anomaly threshold.

[0127] Multiple data windows are input into the LSTM-AE anomaly detection model for window reconstruction to obtain window reconstruction errors.

[0128] The window reconstruction errors are compared with the anomaly threshold to make an anomaly judgment to obtain abnormal window data.

[0129] In this embodiment, first, the reconstruction error of the LSTM-AE anomaly detection model is divided into multiple error sub-distributions according to the drying stages and the vacuum degree intervals of the transformer drying device drying process; then the upper tail of each error sub-distribution is fitted with a generalized Pareto distribution to calculate an adaptive anomaly threshold, and then multiple data windows are input into the LSTM-AE anomaly detection model for window reconstruction to obtain window reconstruction errors; finally, the window reconstruction errors are compared with the anomaly threshold to make an anomaly judgment to obtain abnormal window data. The adaptive anomaly threshold can be realized, and the anomaly threshold can be dynamically adjusted according to the drying stages and the vacuum state, realizing sensitive and robust detection of different working conditions, and further improving the accuracy of anomaly judgment.

[0130] The transformer drying device drying process refers to six stages of "preparation-heating-intermediate distillation-pressure reduction-high vacuum-vacuum breaking" according to the process of the transformer drying device; the vacuum degree interval refers to dividing the drying process into different intervals according to the vacuum degree, for example, dividing it into two levels of medium-low vacuum interval and high vacuum interval with 40 mbar as the boundary.

[0131] The reconstruction error of the LSTM-AE anomaly detection model is divided into multiple error sub-distributions according to the drying stages and the vacuum degree intervals of the transformer drying device drying process, which refers to dividing it into 6x2=12 error sub-distributions according to the six drying stages and the two vacuum degree intervals.

[0132] The extreme value fitting of the upper tail of each error sub-distribution with the generalized Pareto distribution refers to the extreme value fitting of the upper tail of each error sub-distribution with the generalized Pareto distribution (Tail-GPD), and the adaptive abnormal threshold is calculated according to VaR (confidence 95%) + a ES (expected loss, a ≈ 0.1-0.2). In this way, both the drying energy consumption and the vacuum physical constraints are embedded, and the threshold is dynamically adjusted according to the stage-vacuum state, so that sensitive and robust detection of abnormalities in different working conditions is realized.

[0133] In the step of inputting multiple data windows into the LSTM-AE anomaly detection model for window reconstruction to obtain window reconstruction errors, the LSTM-AE anomaly detection model can use the aforementioned part to construct a reconstruction loss function according to the average instantaneous energy consumption of each stage in the drying process of the transformer drying equipment, and then construct and train the generated LSTM-AE anomaly detection model. The reconstruction loss function of the LSTM-AE anomaly detection model is weighted according to the average instantaneous energy consumption of the six drying stages.

[0134] For example, an example of anomaly judgment of the embodiment is described as follows:

[0135] In the foregoing embodiment, the reconstruction error sequence ε t weighted according to the average instantaneous energy consumption of each drying stage can be obtained, and the formula is as follows:

[0136]

[0137] where ω t is the aforementioned energy consumption weight sequence, and x t , is the original and reconstructed feature vector. If s t ∈{1,2,...,k},k=6 is the label of each drying stage (preparation P, heating H, distillation D, pressure reduction PR, high vacuum FV, and vacuum breaking E), then the vacuum category is divided by , where p t is the vacuum degree at a certain time step. If v t =0, it is medium / low vacuum; if v t =1, it is high vacuum. Therefore, there are 12 error sub-distributions, and the formula is as follows:

[0138] D k，h ={ε t |s t =k,v t =h},(k,h)∈(1,...,6)×{0,1};

[0139] where the q-th quantile of each error sub-distribution D k，h is taken As the threshold point (take q = 0.95), the exceedance value z i With the generalized Pareto distribution, the exceedance value z i The formula is as follows:

[0140] z i = ε i -u k,h , z i > 0;

[0141] Then the cumulative distribution function of the exceedance value sample is:

[0142]

[0143] Where ξ is the GPD shape parameter, and σ is the GPD scale parameter. Then the generalized Pareto distribution G(z; ξ, σ) is used to model the extreme value of the exceedance value z, and the parameters (ξ k，h , σ k,h ) are obtained by MLE, that is, for the threshold exceedance residual sample z i , the corresponding GPD probability density function is:

[0144]

[0145] Where the log-likelihood (omitting the constant term) is:

[0146]

[0147] Take the partial derivative of l with respect to ξ and σ respectively, and set it to zero to get the MLE equation group, as follows:

[0148]

[0149] Solve the equation group by Newton-Raphson, and let:

[0150]

[0151] Iterative update, as follows:

[0152]

[0153] Until convergence, that is, the MLE estimate of the parameters (ξ k,h , σ k,h ) is obtained. Next, the adaptive abnormal threshold based on VaR+α·ES is calculated, and given the confidence level p to get:

[0154]

[0155] From which the phase-vacuum adaptive abnormal threshold is obtained:

[0156] Tk，h = VaR k，h (p) + a [ES k，h (p) - VaR k，h (p)], 0 < a < 1;

[0157] where a is determined by grid search on validation set to maximize Precision ≥ 90% and Recall under each sub-distribution. Specifically, when judging an instance (s t = k, v t = h) and e t > T k,h , the instance is labeled as abnormal, otherwise, it is labeled as normal. The model detects in a sliding window manner, if the number of abnormal instances in a window exceeds 30% of the window, the window is labeled as abnormal.

[0158] In some embodiments, before obtaining the fault identification result, the abnormal window data can be input into a random forest model for fault type identification, and the method can further include:

[0159] obtaining historical fault data of the transformer drying equipment;

[0160] dividing the historical fault data into a plurality of fault data windows, each fault data window being labeled with a corresponding fault type;

[0161] converting each fault data window into a feature vector of the corresponding fault type;

[0162] training a random forest model according to the feature vector to obtain a random forest fault classification model.

[0163] In the embodiment, first, historical fault data of the transformer drying equipment is obtained, and then the historical fault data is divided into a plurality of fault data windows; then each fault data window is converted into a feature vector of the corresponding fault type, and finally a random forest model is trained according to the feature vector to obtain a random forest fault classification model. The accuracy of fault type identification can be further improved.

[0164] The above-mentioned obtaining historical fault data of the transformer drying equipment refers to collecting original data of historical known fault periods of the transformer drying equipment by a database or other statistical methods, such as filter clogging, coil temperature sensor failure, oil pump 311 failure, etc.

[0165] The above-mentioned dividing the historical fault data into a plurality of fault data windows, each fault data window being labeled with a corresponding fault type, refers to cutting the historical fault data into data windows with a length of N on a time axis, which can have partial overlap, and labeling the fault type in each window range.

[0166] The above converting each fault data window into a feature vector corresponding to a fault type refers to converting the fault data window into a high-dimensional feature vector. Specifically, time domain features and frequency domain features are extracted from the data (temperature, flow, vacuum, etc.) of each sensor in the fault data window, the time domain features are mean, standard deviation, maximum and minimum value, peak-to-peak value, skewness, kurtosis, maximum difference, root mean square, etc., the frequency domain features are main frequency, main amplitude, spectral centroid, energy bandwidth, total energy, etc. calculated by FFT, then the time domain features and the frequency domain features are spliced to form a high-dimensional feature vector, and finally the corresponding fault label is attached. Thus, a feature vector containing an explicit fault type label is obtained.

[0167] The above training the random forest model according to the feature vector to obtain a random forest fault classification model refers to training the random forest model using the feature vector containing the explicit fault type label, so that the random forest model automatically learns the response mode of the sensor in the time domain and the frequency domain under different fault scenarios. Thus, after detecting a new abnormal data window, the same feature extraction method can be used to convert it into a feature vector to input the random forest model, and automatic determination of the fault category is realized. The random forest fault classification model obtained by the above training method can improve the accuracy of fault type recognition.

[0168] In some embodiments, training the random forest model according to the feature vector to obtain a random forest fault classification model can include:

[0169] training the random forest model according to the feature vector to obtain a random forest initial classification model;

[0170] calculating the average importance of the sensors corresponding to the historical fault data;

[0171] selecting high-discriminative sensor features according to the average importance of the sensors;

[0172] training the random forest initial classification model according to the high-discriminative sensor features to obtain a random forest fault classification model.

[0173] In this embodiment, the random forest model is then trained according to the feature vector to obtain a random forest initial classification model, the average importance of the sensors corresponding to the historical fault data is then calculated, high-discriminative sensor features are then selected according to the average importance of the sensors, and the random forest initial classification model is then trained according to the high-discriminative sensor features to obtain a random forest fault classification model. In this embodiment, the feature importance is used to evaluate each sensor, the most discriminative sensor features are selected according to the average importance of the sensors, and the random forest model is retrained, which can further improve the accuracy and computational efficiency of fault classification.

[0174] The training of the random forest model according to the feature vector is the same as the training process in the foregoing, and will not be described in detail here.

[0175] The average importance of the sensor corresponding to the historical failure data refers to obtaining the sensor features corresponding to the historical failure data, then calculating the importance of the sensor features, and taking the average value of the importance as a threshold for feature selection.

[0176] The high-discrimination sensor features selected according to the average importance of the sensor refer to evaluating the feature importance by measuring the average contribution of each sensor feature to the purity improvement at all decision tree split nodes, and selecting the features according to the evaluation.

[0177] The training of the random forest initial classification model according to the high-discrimination sensor features refers to taking the average value of the importance as a threshold for feature selection, and only selecting the most discriminative sensor features with an average value to train the random forest model. Through the above feature selection, redundancy and noise can be removed, the random forest model can learn more reliable discrimination boundaries in a cleaner feature space, and a large amount of unnecessary calculation is saved, so that the training and reasoning are faster and more resource-saving while improving the fault classification accuracy.

[0178] For example, a specific example of selecting the most discriminative sensor features according to the average importance of the sensor and training the random forest model is given as follows:

[0179] First, data preparation is performed, historical failure data is read and missing value filling, time alignment and normalization processing are performed, then the samples are stratified according to the fault label to divide the training set and the validation set.

[0180] Then, cross-validation training is performed, and K-fold (commonly used K=5) stratified cross-validation is adopted: a random forest model is trained in each fold, and after the training is completed, the importance vector of each sensor feature given by the fold is recorded.

[0181] Then, the average importance is summarized, the n×K importance matrix obtained from K folds is averaged by column to obtain the "average importance" ranking, and the screening rules are set according to business requirements, for example: the cumulative contribution rate of importance reaches 95%, or the top N (such as the top 15) is taken, or the importance is higher than one standard deviation of the average value of all features.

[0182] Then, the feature subset is determined and the final model is trained, only the selected high-discrimination sensor columns are retained, and the random forest is retrained on the feature subset using the complete training set (without folding) to generate the final model, and the "selected feature index / name" is saved separately for column screening during online reasoning.

[0183] Finally, verification and deployment can be performed, and the Precision / Recall / F1 of the new model is evaluated on the independent verification set. If the performance meets the threshold and the inference delay is reduced, the model file and the feature list can be published to the running environment together. During online inference, the sensor column is extracted according to the feature list first, and then sent to the random forest to complete fault classification.

[0184] In summary, the present application realizes joint detection through the LSTM-AE anomaly detection model + random forest fault classification model. First, the real-time collected device data is preprocessed, such as missing value interpolation, outlier removal, and noise reduction by unscented Kalman filter. Then, these time series data are sent to the LSTM-AE anomaly detection model in the form of fixed window length and sliding step for reconstruction. Whether the reconstruction error exceeds the abnormal threshold is used to determine whether the data window belongs to abnormal data. Once a data window is determined to be abnormal, the multi-sensor sequence of the data window is further extracted for multi-dimensional time domain and frequency domain features, combined into a feature vector, and input to the pre-trained random forest model for specific fault type classification. Finally, the entire joint detection process can be evaluated by independent test data, and the performance of the system in high-precision detection and fault type identification is comprehensively measured by using Precision, Recall, F1-score and other indicators, so as to ensure the stability and generalization performance of the series model in actual operation.

[0185] It should be noted that the independent test data refers to data that has not been used in the model training and verification stage. The performance of the system in high-precision detection and fault type identification is measured by using Precision, Recall, F1-score and other indicators, which means that the confusion matrix of the anomaly detection and fault classification results is calculated by combining the real labels of the data, and the precision, recall and F1-score indicators are calculated to evaluate the detection performance and model performance. If both of them are more than 80%, it means that the model has good performance in detecting abnormal data and identifying faults.

[0186] The following describes an example of joint detection by LSTM-AE anomaly detection model + random forest fault classification model, and the specific steps are as follows:

[0187] Table 1 sensor raw data

[0188]

[0189] Table 1 is part of the original data collected from multiple sensors, wherein the sampling period is 5s, the window size is 30s, and the step is 10s.

[0190] The corresponding data window in Table 1 is input into the LSTM-AE anomaly detection model for anomaly detection, and the anomaly detection result is obtained as shown in Table 2:

[0191] Table 2 Anomaly detection result

[0192]

[0193] According to the reconstruction error threshold It can be seen that the window is an abnormal window.

[0194] Then feature extraction is performed, and statistical features and frequency domain features of the window are calculated. Assuming that there are 15 features to be calculated, the total features input into the random forest are 15*8 = 120, but some of them may not be useful for the model to identify the fault type. Therefore, the random forest can perform feature selection on these features, extract features with high contribution to the model identification of the fault, and then use the extracted features to form a feature vector to retrain the random forest.

[0195] The corresponding abnormal data window in Table 2 is input into the random forest fault classification model, and the fault classification judgment result is obtained as shown in Table 3:

[0196] Table 3 Fault classification judgment result

[0197] Category Confidence Temperature sensor failure 0.89 Fine filter clogging 0.06 Fuel oil pump 311 failure 0.03 Normal 0.02

[0198] According to the confidence output by the random forest, the fault type with the highest confidence is output. According to Table 3, the model detects an abnormality and finally determines that it is a "coil temperature sensor fault".

[0199] In summary, the present application realizes real-time monitoring of the transformer drying equipment through the LSTM-AE anomaly detection model and the random forest fault classification model, and automatically performs anomaly detection and fault classification on the real-time device operation data. When an abnormality or a specific fault is detected, a warning mechanism can be automatically triggered to prompt maintenance personnel to intervene in a timely manner. In addition, device operation data can be continuously collected, and the model can be continuously optimized and updated to adapt to changes in the device operating environment and improve the safety of the device operation.

[0200] Based on the fault diagnosis method of the transformer drying equipment provided in the above embodiments, the present application also provides a specific implementation mode of a fault diagnosis device of the transformer drying equipment.

[0201] As shown in Figure 2 The fault diagnosis device 200 of the transformer drying equipment provided in the embodiments of the present application can include:

[0202] A data acquisition module 201 is configured to acquire operation monitoring data, and the operation monitoring data is used to indicate the operation state of the transformer drying equipment.

[0203] The data window module 202 is configured to divide the operation monitoring data into a plurality of fixed-length data windows, and each data window is used to represent a sequence of operation state data of the transformer drying device;

[0204] The anomaly detection module 203 is configured to input the plurality of data windows into an LSTM-AE anomaly detection model to perform window reconstruction and anomaly judgment, and obtain anomaly window data.

[0205] The fault type identification module 204 is configured to input the anomaly window data into a random forest fault classification model to perform fault type identification, and obtain a fault identification result, which is used to indicate the fault type of the transformer drying device.

[0206] The fault diagnosis apparatus 200 of the transformer drying device provided in the embodiments of the present application is used to execute the fault diagnosis method of the transformer drying device in the above embodiments, and the specific processing process is the same as that of the fault diagnosis method of the transformer drying device in the above embodiments, which will not be repeated here.

[0207] Figure 3 A hardware structure schematic diagram of an electronic device provided in the embodiments of the present application is shown.

[0208] The electronic device can include a processor 301 and a memory 302 having computer program instructions stored therein.

[0209] Specifically, the processor 301 can include a central processing unit (CPU), or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application.

[0210] The memory 302 can include a mass storage for data or instructions. By way of example and not limitation, the memory 302 can include a hard disk drive (HDD), a floppy disk drive, a flash memory, an optical disk, a magneto-optical disk, a magnetic tape, or a universal serial bus (USB) drive or a combination of two or more of these. Where appropriate, the memory 302 can include removable or non-removable (or fixed) media. Where appropriate, the memory 302 can be internal or external to the integrated gateway disaster recovery device. In a particular embodiment, the memory 302 is a non-volatile solid-state memory.

[0211] In some embodiments, the memory 302 can include read-only memory (ROM), random access memory (RAM), disk storage media devices, optical storage media devices, flash memory devices, electrical, optical, or other physical / tangible memory storage devices. Thus, generally, the memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software that, when executed (e.g., by one or more processors), is operable to perform the operations described with reference to the methods according to an aspect of the present disclosure.

[0212] The processor 301 implements the transformer drying device fault diagnosis method of any of the above embodiments by reading and executing the computer program instructions stored in the memory 302.

[0213] In one example, the electronic device can further include a communication interface 303 and a bus 310. As shown, the processor 301, the memory 302, and the communication interface 303 are connected through the bus 310 and complete communication with each other. Figure 3

[0214] The communication interface 303 is mainly used to realize the communication between the modules, devices, units and / or equipment in the embodiments of the present application.

[0215] The bus 310 includes hardware, software, or both, that couples components of the online data traffic billing device to each other. By way of example, and not limitation, the bus can include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an InfiniBand (IB) interconnect, a Low Pin Count (LPC) bus, a memory bus, a Micro Channel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association local (VLB) bus, or another suitable bus or a combination of two or more of these. Where suitable, the bus 310 can include one or more buses. Although particular buses are described and shown in the embodiments of the present application, the present application contemplates any suitable bus or interconnect.

[0216] In addition, in combination with the transformer drying device fault diagnosis method in the above embodiments, the embodiments of the present application can provide a computer storage medium to implement. The computer storage medium has computer program instructions stored thereon; the computer program instructions are executed by the processor to implement the transformer drying device fault diagnosis method of any of the above embodiments.

[0217] ​It is to be understood that the application is not limited to particular configurations and processes described herein and shown in the drawings. The detailed description is not to be taken in a limiting sense, and the scope of the present application is defined by the appended claims. In the above embodiments, several specific steps are described and illustrated in order to provide a thorough disclosure of the application. However, it can be appreciated that the method process of the present application is not limited to the specifically enumerated steps, and that various changes, modifications and additions can be made thereto by those skilled in the art without departing from the scope of the present application. Further, the scope of the present application is not intended to be limited to particular configurations and process steps described herein and shown in the drawings.

[0218] The functions shown in the block diagrams of the above described structures can be implemented as hardware, software, firmware or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application specific integrated circuits (ASICs), appropriate firmware, plug-ins, functional cards, etc. When implemented in software, the elements of the present application are program or code segments that are used to perform the required tasks. The program or code segments can be stored in a machine readable medium or transmitted through a data signal carried in a carrier wave over a transmission medium or communication link. A "machine readable medium" includes any medium that can store or transfer information. Examples of machine readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. The code segments can be downloaded via computer networks such as the Internet, intranet, etc.

[0219] It is also to be understood that the example embodiments described herein are based on a series of steps or apparatuses to describe some methods or systems. However, the present application is not limited to the order of the steps described above, that is, the steps can be performed in the order mentioned in the embodiments, or in an order different from the embodiments, or several steps can be performed simultaneously.

[0220] The computer program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other processing devices to cause a series of operational steps to be performed on the computer, other programmable apparatus or other processing devices to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions / acts specified in the flowchart and / or block diagram block or blocks. Alternatively, computer program instructions can be downloaded to or transferred into a general use portable computer, other programmable data processing apparatus, or other processing devices to cause a series of operational steps to be performed on the computer, other programmable apparatus or other processing devices to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions / acts specified in the flowchart and / or block diagram block or blocks.

[0221] The above merely describes a specific implementation of the present application. Those skilled in the art can clearly understand the specific working processes of the system, modules and units described above for the convenience and brevity of description, and the corresponding processes in the foregoing method embodiments can be referred to, which will not be described herein again. It should be understood that the protection scope of the present application is not limited in this way, and any person skilled in the art can easily think of various equivalent modifications or replacements within the technical range disclosed in the present application, and these modifications or replacements should be covered in the protection scope of the present application.

Claims

1. A fault diagnosis method for transformer drying equipment, characterized in that, Includes the following steps: Acquire operational monitoring data, which is used to indicate the operating status of the transformer drying equipment; The operational monitoring data is divided into multiple fixed-length data windows, each of which represents a sequence of operational status data for the transformer drying equipment. Multiple data windows are input into the LSTM-AE anomaly detection model for window reconstruction and anomaly detection to obtain anomaly window data. The abnormal window data is input into a random forest fault classification model to identify the fault type and obtain the fault identification result, which is used to indicate the fault type of the transformer drying equipment.

2. The fault diagnosis method for transformer drying equipment according to claim 1, characterized in that, The process of dividing the operational monitoring data into multiple data windows includes: The missing values ​​in the operation monitoring data are imputed to obtain the first preprocessed data; Remove extreme outliers from the first preprocessed data to obtain the second preprocessed data; The time series of the second preprocessed data is denoised to obtain the third preprocessed data; The third preprocessed data is standardized to obtain the fourth preprocessed data; The fourth preprocessed data is divided into multiple data windows.

3. The fault diagnosis method for transformer drying equipment according to claim 1, characterized in that, Before inputting the multiple data windows into the LSTM-AE anomaly detection model for window reconstruction and anomaly detection, the method further includes: Calculate the average instantaneous energy consumption of each drying stage in the drying process of the transformer drying equipment. The drying stages include the preparation stage, heating stage, distillation stage, pressure reduction stage, high vacuum stage, and vacuum breaking stage. The reconstruction loss function is constructed based on the average instantaneous energy consumption at each stage of the transformer drying process. Construct an LSTM autoencoder model based on the reconstruction loss function; Obtain historical operating data of the transformer drying equipment; The LSTM autoencoder model is trained based on the historical operating data to obtain the LSTM-AE anomaly detection model.

4. The fault diagnosis method for transformer drying equipment according to claim 1, characterized in that, The step of inputting multiple data windows into the LSTM-AE anomaly detection model for window reconstruction and anomaly detection includes: Obtain historical operating data of the transformer drying equipment; Input historical running data into the LSTM-AE anomaly detection model to obtain the reconstruction error of the historical running data; The anomaly threshold is calculated from the reconstruction error of historical running data based on the Gaussian hypothesis and the 3σ algorithm. Multiple data windows are input into the LSTM-AE anomaly detection model for window reconstruction to obtain the window reconstruction error. The window reconstruction error is compared with the anomaly threshold to determine anomalies and obtain abnormal window data.

5. The fault diagnosis method for transformer drying equipment according to claim 1, characterized in that, The step of inputting multiple data windows into the LSTM-AE anomaly detection model for window reconstruction and anomaly detection includes: The reconstruction error of the LSTM-AE anomaly detection model is divided into multiple error sub-distributions according to the various drying stages and vacuum ranges of the transformer drying equipment drying process. The various drying stages include the preparation stage, heating stage, distillation stage, pressure reduction stage, high vacuum stage and vacuum breaking stage. The vacuum ranges include the medium-low vacuum range and the high vacuum range. For the upper tail of each error sub-distribution, extreme values ​​are fitted using a generalized Pareto distribution, and an adaptive outlier threshold is calculated. Multiple data windows are input into the LSTM-AE anomaly detection model for window reconstruction to obtain the window reconstruction error. The window reconstruction error is compared with the anomaly threshold to determine anomalies and obtain abnormal window data.

6. The fault diagnosis method for transformer drying equipment according to claim 1, characterized in that, Before inputting the abnormal window data into the random forest model for fault type identification and obtaining the fault identification result, the method further includes: Obtain historical fault data of transformer drying equipment; The historical fault data is divided into multiple fault data windows, and each fault data window is marked with a corresponding fault type. Each of the aforementioned fault data windows is converted into a feature vector corresponding to the fault type; The random forest model is trained based on the feature vectors to obtain a random forest fault classification model.

7. The fault diagnosis method for transformer drying equipment according to claim 6, characterized in that, The step of training the random forest model based on the feature vector to obtain the random forest fault classification model includes: The random forest model is trained based on the feature vectors to obtain the initial classification model of the random forest. Calculate the average importance of the sensors corresponding to historical fault data; High-discrimination sensor features are selected based on the average importance of the sensors; The initial random forest classification model is trained based on the highly discriminative sensor features to obtain a random forest fault classification model.

8. A fault diagnosis device for transformer drying equipment, characterized in that, include: The data acquisition module is used to acquire operation monitoring data, which is used to indicate the operating status of the transformer drying equipment; The data window module is used to divide the operation monitoring data into multiple fixed-length data windows, each data window representing a segment of the operation status sequence data of the transformer drying equipment; Anomaly detection module is used to input multiple data windows into the LSTM-AE anomaly detection model for window reconstruction and anomaly judgment to obtain anomaly window data; The fault type identification module is used to input the abnormal window data into the random forest fault classification model to identify the fault type and obtain the fault identification result, which is used to indicate the fault type of the transformer drying equipment.

9. An electronic device, characterized in that, The device includes: a processor and a memory storing computer program instructions; When the processor executes the computer program instructions, it implements the fault diagnosis method for the transformer drying equipment as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions, which, when executed by a processor, implement the fault diagnosis method for the transformer drying equipment as described in any one of claims 1-7.

Images