Transformer fault active early warning method and device based on multi-parameter comprehensive analysis

Abstract

The invention discloses a transformer fault active early warning method and device based on multi-parameter comprehensive analysis, and relates to the technical field of transformer fault monitoring. According to the invention, fault judgment is carried out according to the signal measurement value of each sensor in the fault judgment group, fault judgment early warning and timely removal of the transformer are realized, and then primary characteristic parameters and secondary characteristic parameters of signal sequences of various sensors in the early warning group before the fault are calculated and analyzed. According to the fault detection method, early warning and removal of the transformer with the breakdown risk are achieved, the discharge development stage of the transformer is determined, and the hierarchy, timeliness and accuracy of fault detection are improved through comprehensive judgment and analysis of multiple sensors.

Description

Technical Field

[0001] The present application relates to the technical field of transformer fault monitoring, and in particular to a transformer fault active early warning method and device based on multi-parameter comprehensive analysis. Background Art

[0002] Transformer fault monitoring technology uses a large amount of real-time data such as partial discharge, oil and gas, and temperature monitored online, combined with offline input parameters such as transformer structure and factory test, to achieve a comprehensive assessment of the transformer's operating status and form effective assessment and diagnosis results. It provides a scientific and effective reference for operation and maintenance personnel, and can effectively extend the equipment's operating time, formulate a reasonable maintenance plan, and prevent the occurrence of sudden failures.

[0003] The existing fault assessment method adopts the method of evaluating each parameter separately. Due to the time difference between the occurrence of the fault and the manifestation of the fault, the fault diagnosis result will be delayed. In addition, this method will cause some faults to be ignored and cannot meet the timeliness and accuracy requirements of transformer fault monitoring. Summary of the invention

[0004] The purpose of this application is to provide a transformer fault active early warning method and device based on multi-parameter comprehensive analysis to improve the timeliness and accuracy of transformer fault monitoring.

[0005] To achieve the above objectives, this application provides the following solutions.

[0006] In a first aspect, the present application provides a transformer fault active early warning method based on multi-parameter comprehensive analysis, comprising: Obtaining signal measurement values ​​of various sensors in a fault judgment group; the fault judgment group includes a magnetic leakage sensor, a flow rate sensor, and a pressure sensor installed on the transformer; When the signal measurement values ​​of various sensors in the fault judgment group meet the first warning condition, the transformer is determined to be a breakdown fault transformer, and the breakdown fault transformer is removed; otherwise, the following steps are performed: Acquire signal sequences of various sensors in a pre-fault warning group; the pre-fault warning group includes a high-frequency current sensor, an ultrasonic sensor, and a partial discharge ultra-high frequency sensor installed on the transformer; The signal sequences of the high-frequency current sensor and the partial discharge ultra-high frequency sensor are divided respectively by using a sliding window method, and the growth rates of the primary characteristic parameters of each window of the high-frequency current sensor and the growth rates of the primary characteristic parameters of each window of the partial discharge ultra-high frequency sensor are calculated to form a growth rate set; When the growth rate set meets the early warning conditions of the development stage, the sliding window method is used to divide the signal sequences of various sensors in the pre-fault early warning group, and the secondary characteristic parameters of each window of various sensors in the pre-fault early warning group are calculated; When the secondary characteristic parameters of any window of each type of sensor in the pre-fault warning group meet the second warning condition, the transformer is determined to be a breakdown risk transformer and the breakdown risk transformer is cut off; otherwise, the discharge development stage of the transformer is determined based on the signal sequence of each type of sensor in the pre-fault warning group using a timing weighted decision tree model.

[0007] In a second aspect, the present application provides a transformer fault active early warning device based on multi-parameter comprehensive analysis, wherein the transformer fault active early warning device based on multi-parameter comprehensive analysis applies the above-mentioned transformer fault active early warning method based on multi-parameter comprehensive analysis, and the transformer fault active early warning device based on multi-parameter comprehensive analysis includes: A first signal acquisition module is used to obtain signal measurement values ​​of various sensors in the fault judgment group; the fault judgment group includes a magnetic leakage sensor, a flow rate sensor and a pressure sensor installed on the transformer; The breakdown fault transformer judgment module is used to determine that the transformer is a breakdown fault transformer when the signal measurement values ​​of various sensors in the fault judgment group meet the first warning condition, and remove the breakdown fault transformer; otherwise, call the following modules: The second signal acquisition module is used to obtain the signal sequence of various sensors in the pre-fault warning group; the pre-fault warning group includes a high-frequency current sensor, an ultrasonic sensor and a partial discharge ultra-high frequency sensor installed on the transformer; A growth rate calculation module is used to divide the signal sequences of the high-frequency current sensor and the partial discharge ultra-high frequency sensor respectively by using a sliding window method, and calculate the growth rate of each primary characteristic parameter of each window of the high-frequency current sensor and the growth rate of each primary characteristic parameter of each window of the partial discharge ultra-high frequency sensor to form a growth rate set; The secondary characteristic parameter calculation module is used to divide the signal sequences of various sensors in the pre-fault warning group by using the sliding window method when the growth rate set meets the early warning condition of the development stage, and calculate the secondary characteristic parameters of each window of various sensors in the pre-fault warning group; The development stage determination module is used to determine that the transformer is a breakdown risk transformer and cut off the breakdown risk transformer when the secondary characteristic parameters of any window of various sensors in the pre-fault warning group meet the second warning condition; otherwise, the discharge development stage of the transformer is determined by using a timing weight decision tree model based on the signal sequences of various sensors in the pre-fault warning group.

[0008] According to the specific embodiments provided in this application, this application has the following technical effects.

[0009] The present application provides a transformer fault active early warning method and device based on multi-parameter comprehensive analysis. The present application firstly judges the breakdown fault of the transformer according to the signal measurement values ​​of various sensors in the fault judgment group, and timely alarms and cuts off when a breakdown fault exists. When there is no breakdown fault, the signal sequences of the high-frequency current sensor and the partial discharge ultra-high frequency sensor are divided by the sliding window method respectively, and the growth rate of each primary characteristic parameter of each window of the high-frequency current sensor and the growth rate of each primary characteristic parameter of each window of the partial discharge ultra-high frequency sensor are calculated to form a growth rate set; when the growth rate set meets the development stage early warning condition, the signal sequences of various sensors in the pre-fault early warning group are divided by the sliding window method respectively, and the secondary characteristic parameters of each window of each sensor in the pre-fault early warning group are calculated; when the secondary characteristic parameters of any window of each sensor in the pre-fault early warning group meet the second early warning condition, the transformer is determined to be a breakdown risk transformer, and the breakdown risk transformer is cut off, otherwise, according to the signal sequences of various sensors in the pre-fault early warning group, the time series weight decision tree model is used to determine the discharge development stage of the transformer. The present application firstly performs fault judgment based on the signal measurement values ​​of each sensor in the fault judgment group, realizes the fault judgment warning and timely removal of the transformer, and then calculates and analyzes the primary characteristic parameters and secondary characteristic parameters of the signal sequence of each sensor in the pre-fault warning group, realizes the warning and removal of the transformer with breakdown risk, and determines the discharge development stage of the transformer. The present application improves the hierarchy, timeliness and accuracy of fault detection through comprehensive judgment and analysis of multiple sensors. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without creative work.

[0011] Figure 1 A flowchart of a transformer fault active early warning method based on multi-parameter comprehensive analysis is provided in one embodiment of the present application.

[0012] Figure 2 A schematic diagram of a transformer fault active early warning method based on multi-parameter comprehensive analysis provided in one embodiment of the present application.

[0013] Figure 3 A flowchart of the training of a temporal weighted decision tree sub-model provided in one embodiment of the present application.

[0014] Figure 4 A schematic diagram of the calculation results of the peak EMA of the partial discharge ultra-high frequency signal provided in one embodiment of the present application.

[0015] Figure 5 A schematic diagram of the EMA calculation results of the high-frequency current signal skewness provided in one embodiment of the present application.

[0016] Figure 6 A schematic diagram of the calculation results of the acoustic signal peak factor EMA provided in one embodiment of the present application.

[0017] Figure 7 A schematic diagram of feature statistical results within three time windows provided in an embodiment of the present application. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0019] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0020] In an exemplary embodiment, Figure 1 and Figure 2 As shown, a transformer fault active early warning method based on multi-parameter comprehensive analysis is provided, including the following steps 101 to 106.

[0021] Step 101, obtaining signal measurement values ​​of various sensors in a fault judgment group; the fault judgment group includes a magnetic leakage sensor, a flow rate sensor and a pressure sensor installed on the transformer.

[0022] Step 102, when the signal measurement values ​​of various sensors in the fault judgment group meet the first warning condition, determine that the transformer is a breakdown fault transformer, and remove the breakdown fault transformer, otherwise, execute the following steps 103 to 106.

[0023] Step 103, obtaining signal sequences of various sensors in a pre-fault warning group; the pre-fault warning group includes a high-frequency current sensor, an ultrasonic sensor, and a partial discharge ultra-high frequency sensor installed on the transformer.

[0024] Step 104, using a sliding window method to divide the signal sequences of the high-frequency current sensor and the partial discharge ultra-high frequency sensor respectively, and calculate the growth rate of each primary characteristic parameter of each window of the high-frequency current sensor and the growth rate of each primary characteristic parameter of each window of the partial discharge ultra-high frequency sensor to form a growth rate set.

[0025] Step 105, when the growth rate set meets the development stage warning condition, the sliding window method is used to divide the signal sequences of various sensors in the pre-fault warning group, and the secondary characteristic parameters of each window of each sensor in the pre-fault warning group are calculated.

[0026] Step 106: When the secondary characteristic parameters of any window of each type of sensor in the pre-fault warning group meet the second warning condition, the transformer is determined to be a breakdown risk transformer, and the breakdown risk transformer is cut off; otherwise, the discharge development stage of the transformer is determined based on the signal sequence of each type of sensor in the pre-fault warning group using a timing weighted decision tree model.

[0027] By implementing the above steps 101 to 106, it is possible to achieve phased early warning and timely removal of transformer faults. Different judgment methods are used for different discharge development stages to improve the hierarchy, timeliness and accuracy of fault detection.

[0028] The action criterion of effective pressure action protection is based on the pressure information of multiple measuring points under the conditions of normal operation of the transformer, internal short circuit fault, and external short circuit fault. The action value of the measuring point pressure is generally 4-5kPa. For a short circuit with a small number of turns, the transformer protection based on effective pressure information can be identified and acted within hundreds or even tens of milliseconds, while for a short circuit with a large number of turns, the protection device can act within 20ms to 30ms. At the same time, this protection method can distinguish between external short circuit faults and internal short circuit faults, and its protection criterion is not affected by other internal faults such as overheating faults, and has high reliability.

[0029] The oil flow rate is usually applied in the heavy gas action judgment process. Furthermore, the oil flow rate setting value of traditional heavy gas protection is 1-1.5m / s.

[0030] When a breakdown fault occurs inside the transformer and causes a short circuit, the leakage magnetic distribution inside the transformer will be distorted.

[0031] In an embodiment of the present application, the high-frequency current sensor, ultrasonic sensor and partial discharge ultra-high frequency sensor installed on the transformer are set as a pre-fault warning group, and the leakage magnetic sensor, flow rate sensor and pressure sensor installed on the transformer are set as a fault judgment group to perform comprehensive monitoring of transformer faults.

[0032] In another exemplary embodiment, the following settings need to be performed before the above step 101.

[0033] High-frequency current, ultrasonic, partial discharge UHF, magnetic flux leakage, flow velocity and pressure sensors are installed on the transformer. Among them, high-frequency current, ultrasonic and partial discharge UHF sensors are classified as the pre-fault warning group, and magnetic flux leakage, flow velocity and pressure sensors are classified as the post-fault severity judgment group, that is, the fault judgment group, which collects signals during the operation of the transformer in real time.

[0034] In this embodiment, the selected high-frequency current sensor, ultrasonic sensor, partial discharge ultra-high frequency sensor, flow rate sensor and pressure sensor are all optical sensors, and the transformer used for the test is a 110kV real transformer. The installation positions of the sensors are as follows: the high-frequency current sensor is installed at the end screen of the bushing, the ultrasonic sensor is installed at the top of the oil tank, the partial discharge ultra-high frequency sensor is installed at the hand hole position, the flow rate sensor is installed above the winding, and the pressure sensor is installed at the position facing the tank wall between the two-phase windings. After installation, all sensors are started.

[0035] In another exemplary embodiment, Figure 2 As shown, in the above steps 101 to 106, if the signal measurement values ​​of various sensors in the fault judgment group meet the first warning condition, it is determined to be a breakdown fault transformer, and the breakdown fault transformer is removed in time; if not all exceed the threshold, but the high-frequency current and partial discharge ultra-high frequency sensors have a primary characteristic parameter with an absolute value of the growth rate greater than or equal to 20%, then the secondary characteristic parameters of each sensor in the pre-fault warning group are further judged. If the second warning condition is met, it is determined to be a breakdown risk transformer, and the breakdown risk transformer is removed. If the second warning condition is not met, the discharge development stage of the transformer is determined according to the signal sequence of various sensors in the pre-fault warning group using a timing weight decision tree model; if the growth rate of all signals in the pre-fault warning group is less than 20%, it is determined to be an interference signal, and the transformer leakage magnetic flux, pressure and flow rate signals continue to be monitored.

[0036] In another exemplary embodiment, in the above step 104, the sensor signals of the pre-fault warning group are divided by a sliding window method, and the primary characteristic parameters in each window of the high-frequency current and the partial discharge ultra-high frequency signal are calculated. If both signals have primary characteristic parameters with an absolute value of a growth rate greater than 20%, the pre-fault warning group signals are normalized by exponential moving average (EMA) and secondary characteristic calculations are performed (i.e., step 105 is executed), otherwise the signals continue to be monitored.

[0037] The transformer collects sound, high-frequency current, and partial discharge UHF signals through sensors, and uses sliding window division technology to divide the signals according to time. The window length is set to T seconds, and the sliding step is t seconds. The signal information in the window is demodulated and counted to obtain the frequency-current amplitude information corresponding to the sound signal in the window, and the phase-voltage amplitude information corresponding to the high-frequency current and partial discharge UHF signals.

[0038] The signals in the high-frequency current and partial discharge ultra-high frequency sensor window are preliminarily processed to obtain the primary characteristic parameters of the high-frequency current and partial discharge ultra-high frequency sensor.

[0039] Among them, the amplitude average value and window phase width of high-frequency current and partial discharge ultra-high frequency signal are calculated.

[0040] The calculation formula for the amplitude average is shown below.

[0041] .

[0042] in, For the The sensor is in The average signal amplitude in the window, For the The sensor is in The amplitude of the ith data point in the window, When 0 is taken, This type of sensor is a high frequency current sensor. When taking 1, The sensor is an ultra-high sensor; i is the number of the data point sampled in the window, i=1, 2, 3, ..., I; I is the number of data points in the window.

[0043] The calculation formula for the window phase width is shown below.

[0044] .

[0045] in, For the The sensor is in The window phase width within a window is For the The sensor is in The maximum phase point of the signal in the window, For the The sensor is in The minimum phase point of the signal in the window.

[0046] The formula for calculating the growth rate of the average amplitude of two adjacent windows and the window phase width is as follows.

[0047] .

[0048] .

[0049] in, For the Class I sensor The growth rate of the average amplitude of the window, For the The sensor is in The average signal amplitude in the window, For the Class I sensor The growth rate of the window phase width of the window is For the The sensor is in The window phase width within a window.

[0050] For example, for the partial discharge ultra-high frequency sensor, the average amplitudes of two consecutive windows are 0.54V and 0.98V, respectively, with an increase rate of 81.48%, and the phase widths are 102° and 198°, respectively, with an increase rate of 94.12%; for the high-frequency current sensor, the characteristic amplitudes of two consecutive windows are 0.025A and 0.017A, respectively, with an increase rate of -32%, and the phase widths are 217° and 270°, respectively, with an increase rate of 24.42%.

[0051] The exponential moving average (EMA) is used to normalize the signal characteristic parameters and calculate the normalized characteristic values.

[0052] Get the signal amplitude of I consecutive data points in the latest window in real time, and the data quantitative index at the current moment The calculation method of is shown in the following formula.

[0053] .

[0054] in, and are the quantitative indicators of the data in window n and the data in the previous window n-1, respectively. is the secondary characteristic parameter of the data in the current window n, is a smoothing constant, between 0 and 1, generally 0.1 to 0.3. In this embodiment, Take 0.1.

[0055] For partial discharge UHF signals, the secondary characteristic parameter is the maximum amplitude kurtosis, and the calculation formula is shown below.

[0056] .

[0057] in, is the maximum amplitude kurtosis, is the amplitude of the ith data point in the window, is the average value of the amplitude of each data point in the window, is the number of data points in the window.

[0058] For high-frequency current signals, the secondary eigenvalue is the maximum amplitude slope, and the calculation formula is shown below.

[0059] .

[0060] in, is the maximum magnitude inclination.

[0061] For ultrasonic signals, the secondary eigenvalue is the amplitude peak coefficient, and the calculation formula is shown below.

[0062] .

[0063] in, is the amplitude peak factor, is the maximum value in the window, is the RMS value of the signal amplitude within the window.

[0064] When setting the threshold interval near the breakdown stage, the partial discharge ultra-high frequency signal peak threshold interval is 0.638-1, the high-frequency current signal slope threshold interval is 0-0.238, and the ultrasonic signal peak coefficient threshold interval is 0.803-1.

[0065] In this embodiment, based on a large number of experimental results, the thresholds set for the three sensors are: the upper limit of the partial discharge UHF signal peak threshold is 0.6, the lower limit of the high-frequency current signal slope threshold is 0.25, and the upper limit of the ultrasonic signal peak coefficient threshold is 0.8. In order to more clearly illustrate the change trend of the EMA calculation results of the signal characteristic value at each fault development stage, a test set data (calculation results of the whole fault development stage) is selected for illustration, as shown in the figure. Figure 4-Figure 6 As shown, Figure 4-Figure 6 They are respectively the EMA calculation results of the peak value of the partial discharge ultra-high frequency signal, the EMA calculation results of the high-frequency current signal skewness, and the EMA calculation results of the acoustic signal peak coefficient.

[0066] Exemplarily, the latest time window is selected to calculate the kurtosis of the partial discharge ultra-high frequency signal, the maximum amplitude slope of the high-frequency current signal, and the peak coefficient of the ultrasonic signal. The calculated results are 0.12, 0.63, and 0.03, respectively, all of which do not exceed the set threshold. Therefore, the high-frequency current, the amplitude and phase information of the partial discharge ultra-high frequency signal, and the amplitude and frequency information of the ultrasonic signal in the latest time window are combined into a feature vector.

[0067] In another exemplary embodiment, in the above step 106, if all secondary characteristic parameters of the signals of various sensors in the pre-fault warning group are within the set threshold range, it is determined to be close to the breakdown stage (that is, the transformer is determined to be a breakdown risk transformer) and an alarm is issued, and the breakdown risk transformer is cut off at the same time; otherwise, the probability of the discharge development stage of each signal in the pre-fault warning group is calculated using the time series weight gradient decision tree model, the signal addition weight is assigned using the entropy weight method, the maximum probability stage is selected as the judgment result, an alarm is issued, and the signal continues to be monitored.

[0068] This application proposes a recursive multi-classification fault diagnosis overall framework, and constructs a transformer fault diagnosis method based on a gradient boosting decision tree in the overall framework.

[0069] In another exemplary embodiment, in the above step 106, during the fault development stage, the difference in the characteristic signal values ​​is not very obvious and there are fluctuations, and there is also a period of silence, which makes it difficult to accurately identify the detection based on the conventional threshold method. However, through the measured values, it can be found that the amplitude, corresponding phase and frequency information of different signal development stages are different, so the algorithm can be used to identify this part of the signal to more accurately determine the fault state.

[0070] If no secondary characteristic parameter of any sensor signal in the pre-fault warning group exceeds the set threshold, that is, the second warning condition is not met, the amplitude, phase or amplitude, and frequency information of each sensor signal are converted into a feature vector. Within a time window, the feature vectors of the partial discharge ultra-high frequency signal, high-frequency current signal, and acoustic signal are as follows.

[0071] .

[0072] .

[0073] .

[0074] Among them, U represents the local discharge ultra-high frequency signal; H represents the high-frequency current signal; A represents the acoustic signal; y is the development stage of the signal, y∈{1, 2, 3}, when y=1, the signal is in the discharge start stage, when y=2, the signal is in the intense period, when y=3, the signal is in the silent period, and N is the number of windows. In the embodiment of the present application, the discharge development stage is divided into the discharge start stage, the intense period and the silent period.

[0075] For partial discharge ultra-high frequency signals and high-frequency current signals, the maximum signal amplitude and the maximum phase shift in the window are selected as the signal characteristics of the window. The specific calculation formula is shown in the following formula.

[0076] .

[0077] in, is the maximum phase shift of the partial discharge UHF signal in window n, is the phase of the i-th data point of the partial discharge UHF signal in window n, and I is the number of data points in the window.

[0078] .

[0079] in, is the maximum phase shift of the high-frequency current signal in window n, is the phase of the i-th data point of the high-frequency current signal in window n.

[0080] .

[0081] in, is the maximum amplitude of the partial discharge UHF signal in window n, is the amplitude of the i-th data point of the partial discharge UHF signal in window n.

[0082] .

[0083] in, is the maximum amplitude of the high-frequency current signal in window n, is the amplitude of the i-th data point of the high-frequency current signal in window n.

[0084] For the acoustic signal, the maximum signal amplitude in the window and its corresponding phase are selected as the signal features of the window.

[0085] .

[0086] .

[0087] in, is the amplitude of the ith data point of the acoustic signal in window n, is the maximum amplitude of the acoustic signal in window n, is the phase of the i-th data point of the acoustic signal in window n, is the maximum phase value of the acoustic signal in window n.

[0088] In another exemplary embodiment, time dimension information and a time decay function are introduced to balance feature timeliness.

[0089] The time dimension is measured using the sampling interval It captures the dynamic change of features and provides a basis for subsequent weight calculation. For samples far away from the current time, its influence on the final result is weakened. After adding time weight, the calculated probability distribution is smoother than that without adding time weight.

[0090] The formula for calculating the basic time weight is shown below.

[0091] .

[0092] in, is the basic time weight at time t, is the time decay factor; is the difference between the sample time and the current time.

[0093] For the input signal, calculate the basic time weight corresponding to the window it contains, and calculate the average time weight of the same type of signal .

[0094] .

[0095] Among them, m is the type of sensor, that is, the signal type, y is the stage of the signal, is the number of samples in the discharge development stage y in the signal of the m-th sensor, is the basic time weight of the sample in the discharge development stage y in the signal of the m-th sensor at time t.

[0096] Building a predictive model: .

[0097] in, is the prediction model obtained from the k-th training, is the prediction model obtained from the k+1th training, is the gradient coefficient, is the negative gradient of the previous prediction for the k+1th training, is the input feature vector.

[0098] Enter the initial forecast model: ;in, is the initial prediction model.

[0099] Sample initialization probability distribution: .

[0100] in, is the distribution probability of each sample in the k-th training process, , and are the probabilities that each sample belongs to the first, second, and third prediction results during the k-th training process.

[0101] Compute the negative gradient of the previous prediction , as the learning objective of the new decision tree, ;in, is the actual probability of the sample. When the sample belongs to category 2, is [0,1,0].

[0102] Arrange the signals of the same type in a group of signals in ascending order, calculate the middle value of the two data as the splitting point, and take the maximum amplitude value of the partial discharge ultra-high frequency signal as an example. The calculation formula of the splitting point is shown in the following formula.

[0103] .

[0104] in, is the splitting point vector of the maximum amplitude of the partial discharge UHF signal, N is the number of windows, , , , and They are the maximum amplitude values ​​of the partial discharge UHF signals in windows 1, 2, 3, N-1, and N respectively.

[0105] Calculate the gain and set the gain threshold. Select the split point with the maximum gain each time you split. If the gain is greater than the threshold, continue splitting. Otherwise, stop splitting. If the gains are the same, give priority to the smaller value.

[0106] .

[0107] .

[0108] in, For gain, represents the information entropy of the parent node sample set S; |S| is the total number of samples; is the left subtree subset at the split point Information entropy of; | | is the number of samples in the left subtree subset; is the right subtree subset at the split point Information entropy of; | | is the number of samples in the right subtree subset.

[0109] The gain calculation of this application is applied in the growth process of the decision tree to find a balance between model complexity and prediction accuracy. The gain and the loss function are a function of each other, but sometimes the loss function is still decreasing, and if the gain brought by the split is too small, the split will be stopped.

[0110] After selecting the split point, calculate the average negative gradient of the left and right subtrees of the sample split point as the leaf node of the decision tree.

[0111] .

[0112] .

[0113] in, is the forward predicted negative gradient of the left subtree of the split point in the k+1th training, It is the forward predicted negative gradient of the right subtree of the split point in the k+1th training.

[0114] Input the classification sample, compare it with the decision tree split point, and use the corresponding leaf node to update the decision tree model: .

[0115] .

[0116] The sample probability is calculated using the softmax function with time weight, so that more recent samples have a higher influence and improve the adaptability to the current environment.

[0117] .

[0118] in, is the distribution probability of each sample in the k+1th training process, For the The prediction model obtained by training.

[0119] According to actual requirements, input model parameters, including the number of trees, learning rate, maximum depth, subsampling rate, and minimum number of sample leaves.

[0120] In order to illustrate the specific implementation of the technical solution of the present application, the present application sets up the following examples.

[0121] Data processing: partial discharge UHF amplitude and phase information of 5 adjacent time windows, samples 1-5 respectively: ,in, is the input vector of the partial discharge UHF sensor.

[0122] The corresponding stage labels are: ;in, The label corresponding to the input vector of the partial discharge UHF sensor.

[0123] Extract data features: , ;in, is the UHF amplitude characteristic vector of partial discharge, is the UHF phase eigenvector of partial discharge.

[0124] Calculate the median split point of the data after sorting: , , is the median splitting point of the UHF amplitude eigenvector of partial discharge, is the median splitting point of the partial discharge UHF phase eigenvector.

[0125] The two sets of data are arranged in order and correspond to the true labels: , , is the label corresponding to the UHF amplitude feature vector of partial discharge, is the label corresponding to the partial discharge UHF phase feature vector.

[0126] Establish an initialization model for a certain stage of signal : .

[0127] Set the initial value of the probability that a sample is in the four stages : .

[0128] Time decay factor Set to 0.1; the sampling interval is 0.01s, and the time weight of each sample is calculated : Sample 5: .

[0129] Sample 4: .

[0130] Sample 3: .

[0131] Sample 2: .

[0132] Sample 1: .

[0133] For each signal, a decision tree needs to be established in each stage, taking stage 2 as an example.

[0134] First iteration.

[0135] The calculation formula of the negative gradient (residual) of category 2 for all samples is shown below.

[0136] .

[0137] in, is the negative gradient of category 2, is the actual probability of the sample of category 2, is the predicted probability of the sample.

[0138] Calculate the split point gain and set the gain threshold to 0.1. For the amplitude signal, the split point is an information gain of 0.35.

[0139] The left subtree sample includes 1 (0.2), and the corresponding gradient is .

[0140] The right subtree samples include 4 (0.5, 1.3, 0.75, 2.1), and the corresponding gradients are , , , For the parent node, the positive residual probability for: , negative residual probability for: .

[0141] Parent node entropy: .

[0142] Left subtree entropy: .

[0143] Right subtree entropy: .

[0144] .

[0145] Similarly, the gain of the split point 0.625 is 0.0202; the gain of the split point 1.025 is 0.4202; the gain of the split point 1.7 is 0.171. By comparison, the information gain of the split point 1.025 is the largest, which is 0.4202. Therefore, the optimal gain split point for this set of data is 1.025.

[0146] The split samples are split again according to the amplitude characteristics: after the split, the right subtree contains only one sample, and the average negative gradient is , cannot be split, split the left subtree. The amplitude sample corresponding to the left subtree for: , and its corresponding split point for: , calculate the point information gain.

[0147] The information gain is 0.3115 when the split point is 23.5, 0 when the split point is 36.5, and 0.3115 when the split point is 48.5. 23.5 is selected as the split point. The average negative gradient of the left subtree after the split is , the average negative gradient of the right subtree is .

[0148] The established decision tree is shown in the series of program segments.

[0149] if , ; else if , ; else ; in, is the model output probability corresponding to the UHF amplitude feature vector of partial discharge.

[0150] Input sample 2 belonging to category 2, and the extracted data features are [0.5, 32].

[0151] The learning rate η in the score is set to 0.1, and it is calculated that ;in, It is the timing weight decision tree sub-model of the discharge development stage 2 corresponding to the partial discharge UHF sensor after one iteration.

[0152] Similarly, the score of this sample in other types of decision trees can be obtained.

[0153] .

[0154] .

[0155] in, , They are the timing weighted decision tree sub-models of the discharge development stages 1 and 3 corresponding to the partial discharge UHF sensor after one round of iteration.

[0156] The probability of sample 2 being in each discharge development stage is calculated using the softmax function considering time weight: , , ;in, , , They are the probabilities that sample 2 is in each discharge development stage 1, 2, and 3 output by the timing weighted decision tree sub-model corresponding to the discharge development stages 1, 2, and 3 of the partial discharge UHF sensor after one round of iteration.

[0157] For example, according to the actual situation, the model parameters are input, and the initial model is , the initial probability of the three stages of the sample , the gain threshold Gain is (greater than) 0.1, the learning rate η is 0.05, and the maximum depth of the leaf node is 5. The historical experimental data of each signal of the 110kV real transformer used in the experiment will be used to train the gradient decision tree. The high-frequency current, ultrasound, and partial discharge UHF signal data of the historical faults of transformers of the same voltage level at each stage are divided into three parts, 80% for training set, 10% for validation set, and 10% for test set. The training set data is used to train the time-series weighted gradient decision tree sub-models of high-frequency current, ultrasound, and partial discharge UHF signals in the three discharge development stages of discharge initial stage, intense stage, and silent stage, and the validation set and test set are used to test the accuracy of the time-series weighted gradient decision tree sub-model.

[0158] The entropy method is used to calculate the weight of the sound, high-frequency current, and partial discharge UHF sensor signal judgment results in the total result, and the three types of signal information are comprehensively considered. The specific calculation formula is shown in the following formula.

[0159] .

[0160] .

[0161] .

[0162] .

[0163] .

[0164] .

[0165] .

[0166] In the above formula, is the final probability result, is the probability that the transformer belongs to the discharge development stage y, is the weight of the discharge development stage y, is the probability that the signal sequence of the mth sensor in the pre-fault warning group belongs to the discharge development stage y, is the information entropy redundancy of the discharge development stage y, Discharge development stage The information entropy redundancy is is the information entropy of discharge development stage y, is the proportion of samples used to train the time series weighted decision tree sub-model of the discharge development stage y corresponding to the m-th sensor in the pre-fault warning group, is the number of samples used to train the time series weighted decision tree sub-model of the discharge development stage y corresponding to the m-th sensor in the pre-fault warning group, is the discharge development stage corresponding to the mth sensor in the pre-fault warning group for training The number of samples of the temporal weighted decision tree sub-model, is the normalization factor, is the total number of categories.

[0167] For example, the specific training process is described in the following embodiments. Figure 3 The specific steps shown are as follows.

[0168] The partial discharge UHF signal is input into three time-series weighted gradient decision tree sub-models of the partial discharge UHF signal discharge initial stage, the partial discharge UHF signal intense period, and the partial discharge UHF signal silent period. The probabilities of the signal being in the three stages are calculated to be 37%, 46%, and 27%, respectively.

[0169] The high-frequency current signal is input into three time-series weight gradient decision tree sub-models: the initial stage of high-frequency current signal discharge, the intense stage of high-frequency current signal, and the silent stage of high-frequency current signal. The probabilities of the signal being in the three stages are calculated to be 11%, 49%, and 44%, respectively.

[0170] The acoustic signal is input into three time-series weight gradient decision tree sub-models at the initial stage of acoustic signal discharge, the acute stage of acoustic signal, and the silent stage of acoustic signal. The probabilities of the signal being in the three stages are calculated to be 51%, 63%, and 21%, respectively.

[0171] The entropy method is used to calculate the weight of the judgment results of partial discharge ultra-high frequency sensor signal, high-frequency current sensor signal, and sound sensor signal in the total result, and the three types of signal information are comprehensively considered.

[0172] Calculated partial discharge UHF sensor signal weight for: .

[0173] Calculate the signal weight of the high-frequency current sensor for: .

[0174] Calculated sound sensor signal weight for: .

[0175] The three signals are added together according to the calculated weights, and the calculation results are shown below.

[0176] P1=37%×31.54%+11%×48.79%+51%×19.67%=27.07%.

[0177] P2=46%×31.54%+49%×48.79%+63%×19.67%=50.81%.

[0178] P3=27%×31.54%+44%×48.79%+21%×19.67%=34.13%.

[0179] Among them, P1, P2, and P3 are the probabilities that the partial discharge UHF sensor is in the discharge initial stage, intense period, and silent period, respectively.

[0180] Using the formula P=max (P1, P2, P3)=50.81%, where P is the final judgment result, the severe period is selected as the final judgment result and an alarm is issued.

[0181] To illustrate the differences in fault amplitude characteristics at different fault development stages, the characteristic statistics within three time windows after visualization of part of the test set data are shown in the following figure. Figure 7 shown.

[0182] In another exemplary embodiment, the above step 202 specifically includes the following steps.

[0183] During the whole process of fault monitoring, after starting the sensor, if the monitoring value of any signal of the leakage magnetic field, pressure and flow rate sensors in the fault severity judgment group is greater than 0, further judgment of the signal is required, otherwise continuous monitoring is performed.

[0184] When the three signal sensors all exceed the threshold value, the fault is judged as a serious fault stage, the transformer is a breakdown fault transformer, and the breakdown fault transformer is removed in time.

[0185] The optional upper threshold value range of the magnetic flux leakage sensor is 1.5-2.0B / mT, the upper threshold value range of the flow rate sensor is 0-1m / s, and the upper threshold value range of the pressure sensor is 0-2kPa.

[0186] In this embodiment, the upper threshold value range of the magnetic flux leakage sensor is set to 1.5 B / mT, the upper threshold value range of the flow rate sensor is set to 0.1 m / s, and the upper threshold value range of the pressure sensor is set to 0.5 kPa. The measured values ​​of the pressure and flow rate sensors do not exceed the threshold value.

[0187] In the method in the above embodiment of the present application, high-frequency current, ultrasound, partial discharge ultra-high frequency, leakage magnetic, flow rate and pressure sensors are installed on the transformer, and the sensors are divided into a fault judgment group and a pre-fault warning group. If the signal measurement values ​​of various sensors in the fault judgment group do not meet the first warning condition, the pre-fault warning group sensor signals are divided by the sliding window method, and the preliminary characteristic values ​​in each window of the high-frequency current and the partial discharge ultra-high frequency signal are calculated. If both signals have a primary characteristic parameter with an absolute value of the growth rate greater than 20%, the pre-fault warning group signal is EMA normalized and the secondary characteristic is calculated, otherwise the signal continues to be monitored. If all the secondary characteristic parameters of the pre-fault warning group signal are within the set threshold range, it is determined to be a breakdown risk transformer and an alarm is issued, and the breakdown risk transformer is cut off at the same time, otherwise the probability of the discharge development stage of each signal in the pre-fault warning group is calculated using the time series weight gradient decision tree model, and the signal addition weight is allocated using the entropy weight method, and the maximum probability stage is selected as the judgment result, and an alarm prompt is issued, and the signal is continued to be monitored.

[0188] This method realizes phased early warning and timely removal of transformer faults through multi-sensor comprehensive judgment and algorithm analysis. Different judgment methods are used for different discharge development stages to improve the hierarchy, timeliness and accuracy of fault detection.

[0189] Based on the same inventive concept, the embodiment of the present application also provides a transformer fault active early warning device based on multi-parameter comprehensive analysis for implementing the transformer fault active early warning method based on multi-parameter comprehensive analysis mentioned above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more embodiments of the transformer fault active early warning device based on multi-parameter comprehensive analysis provided below can refer to the limitations of the transformer fault active early warning method based on multi-parameter comprehensive analysis above, and will not be repeated here.

[0190] In an exemplary embodiment, a transformer fault active early warning device based on multi-parameter comprehensive analysis is provided, including the following modules.

[0191] The first signal acquisition module is used to obtain signal measurement values ​​of various sensors in the fault judgment group; the fault judgment group includes a magnetic leakage sensor, a flow rate sensor and a pressure sensor installed on the transformer.

[0192] The breakdown fault transformer judgment module is used to determine that the transformer is a breakdown fault transformer and remove the breakdown fault transformer when the signal measurement values ​​of various sensors in the fault judgment group meet the first warning condition. Otherwise, the following module is called.

[0193] The second signal acquisition module is used to obtain signal sequences of various sensors in the pre-fault warning group; the pre-fault warning group includes a high-frequency current sensor, an ultrasonic sensor and a partial discharge ultra-high frequency sensor installed on the transformer.

[0194] The growth rate calculation module is used to divide the signal sequences of the high-frequency current sensor and the partial discharge ultra-high frequency sensor respectively by using the sliding window method, and calculate the growth rate of each primary characteristic parameter of each window of the high-frequency current sensor and the growth rate of each primary characteristic parameter of each window of the partial discharge ultra-high frequency sensor to form a growth rate set.

[0195] The secondary characteristic parameter calculation module is used to divide the signal sequences of various sensors in the pre-fault warning group by the sliding window method when the growth rate set meets the early warning conditions of the development stage, and calculate the secondary characteristic parameters of each window of various sensors in the pre-fault warning group.

[0196] The development stage determination module is used to determine that the transformer is a breakdown risk transformer and cut off the breakdown risk transformer when the secondary characteristic parameters of any window of various sensors in the pre-fault warning group meet the second warning condition; otherwise, the discharge development stage of the transformer is determined by using a timing weight decision tree model based on the signal sequences of various sensors in the pre-fault warning group.

[0197] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0198] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0199] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A transformer fault active early warning method based on multi-parameter comprehensive analysis, characterized in that: include: Obtain signal measurement values ​​of various sensors in the fault judgment group; The fault judgment group includes a magnetic leakage sensor, a flow rate sensor and a pressure sensor installed on the transformer; When the signal measurement values ​​of various sensors in the fault judgment group meet the first warning condition, the transformer is determined to be a breakdown fault transformer, and the breakdown fault transformer is removed; otherwise, the following steps are performed: Acquire signal sequences of various sensors in a pre-fault warning group; the pre-fault warning group includes a high-frequency current sensor, an ultrasonic sensor, and a partial discharge ultra-high frequency sensor installed on the transformer; The signal sequences of the high-frequency current sensor and the partial discharge ultra-high frequency sensor are divided respectively by using a sliding window method, and the growth rates of the primary characteristic parameters of each window of the high-frequency current sensor and the growth rates of the primary characteristic parameters of each window of the partial discharge ultra-high frequency sensor are calculated to form a growth rate set; When the growth rate set meets the early warning conditions of the development stage, the sliding window method is used to divide the signal sequences of various sensors in the pre-fault early warning group, and the secondary characteristic parameters of each window of various sensors in the pre-fault early warning group are calculated; When the secondary characteristic parameters of any window of each type of sensor in the pre-fault warning group meet the second warning condition, the transformer is determined to be a breakdown risk transformer and the breakdown risk transformer is cut off; otherwise, the discharge development stage of the transformer is determined based on the signal sequence of each type of sensor in the pre-fault warning group using a timing weighted decision tree model.

2. The transformer fault active early warning method based on multi-parameter comprehensive analysis according to claim 1 is characterized in that: Each primary characteristic parameter includes the amplitude average value and the window phase width; The secondary characteristic parameter of the high-frequency current sensor is the maximum amplitude slope; The secondary characteristic parameter of the ultrasonic sensor is the amplitude peak factor; The secondary characteristic parameter of the partial discharge UHF sensor is the maximum amplitude kurtosis.

3. The transformer fault active early warning method based on multi-parameter comprehensive analysis according to claim 2 is characterized in that: The first warning condition is: the signal measurement value of the magnetic leakage sensor is greater than the magnetic leakage threshold, the signal measurement value of the flow rate sensor is greater than the flow rate threshold, and the signal measurement value of the pressure sensor is greater than the pressure threshold; The development stage warning condition is: both the high-frequency current sensor and the partial discharge ultra-high frequency sensor in the adjacent windows have a primary characteristic parameter whose absolute value of growth rate is greater than the growth rate threshold; The second warning condition is that the maximum amplitude inclination is within the inclination interval of the stage approaching breakdown, the amplitude peak coefficient is within the coefficient interval of the stage approaching breakdown, and the maximum amplitude kurtosis is within the kurtosis interval of the stage approaching breakdown.

4. The transformer fault active early warning method based on multi-parameter comprehensive analysis according to claim 1 is characterized in that: According to the signal sequence of various sensors in the pre-fault warning group, the time series weight decision tree model is used to determine the discharge development stage of the transformer, including: According to the signal sequences of various sensors in the pre-fault warning group, the characteristic vectors of the signal sequences of various sensors in the pre-fault warning group are constructed; The characteristic vectors of the signal sequences of various sensors in the pre-fault warning group are respectively input into the time series weight decision tree sub-models of various discharge development stages corresponding to various sensors, and the probability that the signal sequences of various sensors in the pre-fault warning group belong to each discharge development stage is determined; The following formula is used to perform weighted summation of the probability that the signal sequence of each type of sensor in the pre-fault warning group belongs to each discharge development stage, and the probability that the transformer belongs to each discharge development stage is determined; ； in, is the probability that the transformer belongs to the discharge development stage y, is the weight of the discharge development stage y, is the probability that the signal sequence of the mth sensor in the pre-fault warning group belongs to the discharge development stage y; According to the probability that the transformer belongs to each discharge development stage, the discharge development stage of the transformer is determined.

5. The transformer fault active early warning method based on multi-parameter comprehensive analysis according to claim 4 is characterized in that: The calculation formula of the weight of the discharge development stage y is: ； ； ； ； in, is the information entropy redundancy of the discharge development stage y, Discharge development stage The information entropy redundancy is is the information entropy of discharge development stage y, is the proportion of samples used to train the time series weighted decision tree sub-model of the discharge development stage y corresponding to the m-th sensor in the pre-fault warning group, is the number of samples used to train the time series weighted decision tree sub-model of the discharge development stage y corresponding to the m-th sensor in the pre-fault warning group, is the discharge development stage corresponding to the mth sensor in the pre-fault warning group for training The number of samples of the temporal weighted decision tree sub-model, is the normalization factor.

6. The transformer fault active early warning method based on multi-parameter comprehensive analysis according to claim 4 is characterized in that: The time series weight decision tree sub-models for each discharge development stage corresponding to each type of sensor in the pre-fault warning group are trained in the following way: Constructing a target sample set; the target sample set is a set of time-series weighted decision tree sub-models trained to obtain a target discharge development stage, the target discharge development stage is any discharge development stage among various discharge development stages, and the target sensor is any sensor among various types of sensors in the pre-fault warning group; A decision tree model is trained based on the target sample set to obtain a timing weighted decision tree sub-model of the target discharge development stage.

7. A transformer fault active early warning device based on multi-parameter comprehensive analysis, characterized in that: The transformer fault active early warning device based on multi-parameter comprehensive analysis applies the transformer fault active early warning method based on multi-parameter comprehensive analysis described in any one of claims 1 to 6, and the transformer fault active early warning device based on multi-parameter comprehensive analysis includes: A first signal acquisition module is used to obtain signal measurement values ​​of various sensors in the fault judgment group; the fault judgment group includes a magnetic leakage sensor, a flow rate sensor and a pressure sensor installed on the transformer; The breakdown fault transformer judgment module is used to determine that the transformer is a breakdown fault transformer when the signal measurement values ​​of various sensors in the fault judgment group meet the first warning condition, and remove the breakdown fault transformer; otherwise, call the following modules: The second signal acquisition module is used to obtain the signal sequence of various sensors in the pre-fault warning group; the pre-fault warning group includes a high-frequency current sensor, an ultrasonic sensor and a partial discharge ultra-high frequency sensor installed on the transformer; A growth rate calculation module is used to divide the signal sequences of the high-frequency current sensor and the partial discharge ultra-high frequency sensor respectively by using a sliding window method, and calculate the growth rate of each primary characteristic parameter of each window of the high-frequency current sensor and the growth rate of each primary characteristic parameter of each window of the partial discharge ultra-high frequency sensor to form a growth rate set; The secondary characteristic parameter calculation module is used to divide the signal sequences of various sensors in the pre-fault warning group by using the sliding window method when the growth rate set meets the early warning condition of the development stage, and calculate the secondary characteristic parameters of each window of various sensors in the pre-fault warning group; The development stage determination module is used to determine that the transformer is a breakdown risk transformer and cut off the breakdown risk transformer when the secondary characteristic parameters of any window of various sensors in the pre-fault warning group meet the second warning condition; otherwise, the discharge development stage of the transformer is determined by using a timing weight decision tree model based on the signal sequences of various sensors in the pre-fault warning group.

Images