Search

CN-121981699-A - Transformer respirator predictive maintenance method combined with edge calculation

CN121981699ACN 121981699 ACN121981699 ACN 121981699ACN-121981699-A

Abstract

The invention relates to the field of intelligent operation and maintenance of power equipment, in particular to a predictive maintenance method of a transformer respirator in combination with edge calculation. The method comprises the steps of S1 deploying sensors at edge nodes, collecting parameters such as temperature and humidity of a respirator and related data of a transformer, removing anomalies and standardization through Kalman filtering, building a structured data set, S2 extracting time domain frequency domain characteristics, early warning and marking anomalies through an improved SVM model, S3 abnormal data cloud end training an LSTM model and then optimizing parameters, S4 building a life sub-model to calculate residual life and sequencing, S5 combining load fixed maintenance priority and scheme, and evaluating to form a closed loop after maintenance. The method realizes fault early warning and high-efficiency operation and maintenance, and provides support for equipment operation and maintenance.

Inventors

  • ZHANG QINGJIE
  • REN YIHANG
  • MA SHICHAO
  • ZHANG HENG
  • ZHANG KUNKUN
  • GAO MENGYUAN

Assignees

  • 国网河南省电力公司嵩县供电公司

Dates

Publication Date
20260505
Application Date
20251209

Claims (10)

  1. 1. A method of predictive maintenance of a transformer respirator in combination with edge calculation, comprising the steps of: The method comprises the steps of S1, deploying multiple types of sensors at edge nodes of a respirator, collecting original data in real time, accessing transformer associated data, removing abnormal values by Kalman filtering, normalizing the data, and generating a structured data set; s2, extracting core features to form a vector library, loading a lightweight model, comparing real-time features with a historical normal threshold, and generating an early warning signal and marking an abnormal position if the threshold is exceeded; S3, uploading abnormal data to a cloud, training a model by combining historical fault data by the cloud, optimizing parameters, and then transmitting the parameters to an edge node to finish model iteration; S4, loading an updated model, establishing a life sub-model for a drying agent, a fan and the like, substituting real-time data to calculate the residual life, and generating a life sorting table; and S5, setting maintenance priority according to service life, generating a task list in combination with transformer substation load, matching equipment parameters, outputting a maintenance scheme, uploading a result to an edge node after maintenance, and synchronizing a cloud to form a closed loop.
  2. 2. The method for predictive maintenance of a transformer respirator in combination with edge calculation according to claim 1, wherein in step S1, the formula for removing outliers by kalman filtering is as follows: ; ; ; ; ; In the formula, For the a priori state estimate at time k, In the form of a state transition matrix, For the a-1 time posterior state estimate, In order to control the input matrix, The input is controlled for the time of k-1, For the a priori error covariance, For the time k-1 posterior error covariance, Is that Is used to determine the transposed matrix of (a), In order to process noise covariance(s), In order for the kalman gain to be achieved, In order to observe the matrix, Is that Is used to determine the transposed matrix of (a), In order to observe the covariance of the noise, As an observation of the time instant k, Is a unit matrix; the data normalization formula is In the following For the data value to be normalized, As the raw sensor data value(s), As the mean value of the original data, Standard deviation of the original data.
  3. 3. The method of edge-computing transformer breather predictive maintenance according to claim 1, wherein in step S1, the plurality of sensor acquisition parameters comprises breather internal temperature Relative humidity of Flow rate of desiccant air flow Pressure difference between inlet and outlet Concentration of dust in the interior The corresponding units are degrees centigrade, percentage, cubic meter per hour, kilopascal, milligram per cubic meter, and the related data of the transformer comprises load current Oil temperature of top layer Oil level The corresponding units are ampere, centigrade and millimeter respectively, the structured data set takes the timestamp as an index, and the format is sensor type-acquisition value-associated data item-data quality identifier, wherein the data quality identifier 0 represents normal, and the data quality identifier 1 represents abnormal.
  4. 4. The method for predictive maintenance of a transformer respirator in combination with edge calculation according to claim 1, wherein in step S2, the core feature extraction formula is as follows, and the sliding window is set to 5 minutes: The time domain characteristic formula is 、 、 、 ; The frequency domain features are obtained through FFT transformation, and the formula is ; The core feature vector is ; In the formula, As the mean value of the window, Take a value of 300 and corresponds to 1 data per second, For the i-th data value within the window, For the window variance to be the value of the window variance, As a peak value of the window, As a slope of the trend, For the data value at the time instant t, Is the amplitude of the signal in the frequency domain, Is the frequency and is in hertz, The value of 300 is taken to be the value of 300, In units of imaginary numbers, Is the dominant frequency and has the unit of hertz, Is the dominant frequency amplitude.
  5. 5. The method for predictive maintenance of a transformer respirator in combination with edge computing according to claim 1, wherein in step S2, the lightweight model is a modified SVM whose linear kernel function formula is In the following For the value of the kernel function, 、 Is a sample feature vector; the model parameters are set as penalty parameters Relaxation variable ; The early warning judgment formula is Wherein Take the value of 6, when Generating early warning signal when in use In order for the euclidean distance to be the same, For the real-time feature vector i-th dimension value, And (5) the i-th dimension value is the average value of the historical normal feature vectors.
  6. 6. The method for predictively maintaining a transformer respirator in combination with edge computing according to claim 1, wherein in step S3, the abnormal data early warning classification rule is that the first-level early warning corresponds to Second-level early warning corresponds to Wherein Is Euclidean distance; The cloud LSTM model parameter is set as an input layer dimension 6, the hidden layer is set as 2 layers, and the neuron numbers are respectively 、 The output layer is 3 kinds of fault probability 、 、 And meet the following ; Adam optimizer parameters set to initial learning rate Number of iterations The learning rate is adjusted 100 times per iteration, and the adjustment rule is that the loss of the verification set is continuously 2 times without drop ; The optimized model parameters are issued in a JSON format and comprise model version numbers, weight matrixes, offset vectors and activation function types.
  7. 7. The method of predictive maintenance of a transformer respirator in combination with edge calculation of claim 1, wherein in step S4, the lifetime submodel formula is as follows: the residual life formula of the desiccant is Wherein The value is taken for 8000 hours, Value taking ; The fan reliability formula is Wherein The value is taken for 12000 hours, The value is 1.8; The remaining life calculation condition of the fan is as follows The formula is ; In the middle of 、 Is the remaining life in hours, Is the run-time and is in hours, In order to be of a relative humidity level, In order to be a flow rate, In order to be able to determine the temperature, The reliability of the fan is obtained.
  8. 8. The method of claim 1, wherein in step S4, the life sequencing formula is Wherein The value of the water-based paint is 0.6, The value of the water-based paint is 0.4, The value is 12000 hours; The ordering rule is In response to the priority of maintenance, In response to the scheduled maintenance, Corresponding to normal monitoring, in The composite life was scored.
  9. 9. The method for predictive maintenance of a transformer respirator in combination with edge computing according to claim 1, wherein in step S5, the maintenance window period screening rule is a load factor The load factor calculation formula is as follows ; Maintaining a priority coefficient formula of ; The maintenance classification rule is as follows In correspondence with the emergency maintenance of the vehicle, Corresponding to the routine maintenance of the device, Corresponding delay maintenance; In the middle of Is the load factor and the unit is the percentage, Is the real-time load current and is in amperes, Is the rated current and is given in amperes, As a coefficient of priority the number of bits, Scoring the composite life; The task list comprises equipment numbers, maintenance grades, recommended maintenance time periods, spare part models and operation step abstracts, wherein the spare part models are drying agents XH-7 and fans FB-200.
  10. 10. The method for predictive maintenance of a transformer respirator in combination with edge calculation according to claim 1, wherein in step S5, the maintenance effect evaluation formula is The judgment standard is The maintenance is qualified; The formula for calculating the ageing degree of spare parts is desiccant Blower fan ; The closed loop iteration rule is that the edge node updates the data set and the feature library, the cloud updates the fault database and retrains the LSTM model in quarters, and the optimized parameters are issued to the edge node; In the middle of As the absolute value of the characteristic deviation, In order to maintain the post-feature vector, As a mean value of the historical normal feature vectors, 、 Is the degree of aging and is given in percent.

Description

Transformer respirator predictive maintenance method combined with edge calculation Technical Field The invention relates to the field of intelligent operation and maintenance of power equipment, in particular to a predictive maintenance method of a transformer respirator in combination with edge calculation. Background In the field of operation and maintenance of power equipment, an oil immersed transformer respirator is used as a core component for guaranteeing the performance of insulating oil, and the operation and maintenance technology of the oil immersed transformer respirator has obvious limitations for a long time and is disjointed with the development requirement of the current intelligent power grid. The operation and maintenance of the existing breather mainly adopt manual regular inspection, and a system scheme is lacking in a data acquisition link, namely, the operation and maintenance of the existing breather does not have cooperative deployment of multiple types of sensors, only the color change of a drying agent and the state of an oil cup are observed manually, core parameters such as the internal temperature, humidity, airflow flow, pressure difference and dust concentration of the breather cannot be acquired, associated data such as transformer load current and top oil temperature are not accessed, and further, the processes of eliminating abnormal Kalman filtering values and standardizing the data do not exist, so that raw data are disordered and have large errors, a structured data set is difficult to form, and a reliable basis cannot be provided for subsequent analysis. In the aspect of data processing and early warning, the prior art has no core feature extraction mechanism, does not calculate the features such as time domain mean value, variance, peak value and the like through a sliding window, does not acquire frequency domain information by means of FFT conversion, does not introduce a lightweight model, namely lacks the support of models such as an improved SVM and the like, only relies on artificial experience to judge data trend, does not have early warning logic of Euclidean distance comparison real-time features and historical threshold values, cannot generate a grading early warning signal, and often has excessive reaction or response lag, thereby wasting resources and being difficult to suppress fault expansion. In terms of model iteration and life management, the prior art has no 'edge-cloud' collaborative architecture, namely, an edge side can only simply collect data, abnormal data cannot be uploaded to the cloud, the cloud cannot train an LSTM model and optimize parameters to issue, so that the model cannot be upgraded iteratively, meanwhile, a life sub-model is not established for a drying agent and a fan, the influence of temperature, humidity and flow on the residual life of the drying agent is not considered, the fan replacement time is not calculated through a reliability formula, and only a periodic replacement mode is adopted, so that the problems of non-failure part waste and potential safety hazards existing near the life limit part are caused. In terms of maintenance decision and closed-loop management, the prior art has no scientific maintenance priority division, namely a low-load maintenance window period is not screened through comprehensive service life grading and transformer substation load rate, power grid load fluctuation and user loss are easy to cause due to blind power outage overhaul, a maintenance effect evaluation mechanism is not adopted, the characteristic deviation and the spare part aging degree after maintenance are not calculated, a maintenance result cannot be synchronized to an edge and cloud update data set and an optimization model, a closed loop of acquisition, analysis, maintenance and update is difficult to form, the operation and maintenance technology cannot self-evolve, and the problem of resource mismatch is outstanding. Disclosure of Invention In order to thoroughly solve the problems of no system for data acquisition, missing processing early warning, difficult iteration of a model, rough service life management and no closed loop maintenance in the prior art, the invention provides a predictive maintenance method of a transformer breather combining edge calculation, and the operation and maintenance are optimized through multi-sensor acquisition, intelligent analysis, cloud iteration, service life modeling and closed loop management, and the specific technical scheme is as follows: a method of predictive maintenance of a transformer breather in combination with edge calculation, comprising the steps of: The method comprises the steps of S1, deploying multiple types of sensors at edge nodes of a respirator, collecting original data in real time, accessing transformer associated data, removing abnormal values by Kalman filtering, normalizing the data, and generating a structured data set; s2, extracting core features to form a