A Design Method for Adaptive Early Warning Strategy of Blade Tower Sweeping Fault in Wind Turbines

Abstract

The present invention discloses a design method for an adaptive early warning strategy for the fault of a wind turbine blade sweeping the tower, which includes the steps of: sampling the data of an offshore wind turbine based on the SCADA system to obtain a SCADA normal behavior data set; preprocessing the data in the data set to construct training samples; reallocating the weights of the training samples; extracting the vibration frequency domain features of the training samples; modeling normal behavior and calculating the residual sequence; training the residual distribution, counting the maximum number of consecutive ones greater than the upper threshold or less than the lower threshold in the training residual sequence, which is recorded as the maximum consecutive overrun times and used as the final output result in the training process; and entering the online application stage. By designing an adaptive allocation strategy for the weights of training samples, the present invention takes into account that sample imbalance under different wind conditions easily leads to modeling accuracy imbalance, and assigns higher training weights to the severe wind conditions with a smaller proportion in the training samples, ensuring the accuracy and comprehensiveness of the results for faults caused by extreme wind conditions.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind turbine fault monitoring, and particularly relates to a design method for an adaptive early warning strategy for the blade tower sweeping fault of a wind turbine unit. Background Art

[0002] As the most mature new energy power generation technology today, wind power generation has developed rapidly worldwide. The total installed capacity of wind turbines in China has ranked among the top in the world, and the penetration rate of wind power is also increasing year by year. However, the rapid development of the wind power generation market in recent years has also led to insufficient preparation during the R & D period, and the operation and maintenance costs of wind turbines remain high.

[0003] The high failure rate of wind turbines is the main factor leading to high operation and maintenance costs. The units usually operate in remote areas such as far - away suburbs, plains, mountains, and coastal areas. Under harsh wind conditions such as strong winds and turbulence, each key component of the wind turbine is more likely to be subjected to high loads and high disturbances, posing challenges to the safe operation of each key component of the unit. The SCADA system installed in the wind turbine generally has corresponding measuring points in key components, and monitors the status of key components and gives early warnings of faults by monitoring the information of each component, realizes the initial identification of abnormal conditions of the unit, avoids the evolution of initial abnormalities into catastrophic faults, and realizes the timely perception of blade tower sweeping faults, which is of great significance for reducing operation and maintenance costs and realizing intelligent operation and maintenance of wind farms. However, in the historical SCADA operation data, extreme wind conditions occur less frequently, and the imbalance of different wind condition samples may lead to large errors in the modeling process for extreme wind conditions. The existing methods for fault monitoring do not consider the problem of sample imbalance, and mostly only consider time - domain information, and cannot monitor the spectral characteristics of the operation status of detection components, thus it is difficult to ensure the accuracy of the early warning results. Summary of the Invention

[0004] To make up for the deficiencies of the existing technology, the present invention provides a design method for an adaptive early warning strategy for the blade tower sweeping fault of a wind turbine unit, which is based on fault monitoring with adaptive allocation of training sample weights, reduces human uncertain factors, solves the problem of wind condition imbalance in historical training data, and realizes more accurate fault monitoring.

[0005] A design method for an adaptive early warning strategy for the blade tower sweeping fault of a wind turbine unit specifically includes the following steps:

[0006] S1: Sampling the data of an offshore wind turbine unit based on the SCADA system to obtain a SCADA normal behavior data set;

[0007] S2: Pre - process the data in the data set to construct training samples;

[0008] S3: Re - allocate the weights of the training samples;

[0009] S4: Extract the vibration frequency domain features of the training samples;

[0010] S5: Model normal behavior and calculate the residual sequence;

[0011] S6: Train the residual distribution, count the maximum number of consecutive values greater than the upper threshold or less than the lower threshold in the training residual sequence, denoted as the consecutive maximum overrun count train_max_count, and take the upper and lower thresholds in the training set, the maximum training overrun parameter, and the trained model as the final output results in the training process;

[0012] S7: Enter the online application stage.

[0013] Furthermore, in the step S1, key components of the wind turbine to be monitored for blade faults are selected, the frequency domain features of the vibration signals measured at the vibration measurement points of these components in the SCADA system are used as the target variable y, and the parameters X related to the component vibration are used as the relevant variables. N pieces of operation data under the normal operation state of the unit are selected, and a training set is constructed with a ten-minute sliding window.

[0014] Furthermore, in the step S2, the preprocessing includes:

[0015] S21: Combine the characteristics of the prior abnormal operation state determined manually to eliminate the data of abnormal operation in the training set that has not been detected by the SCADA system;

[0016] S22: Select the standard wind speed section. Set the time sliding window length to be the same as that of the training samples. When the overall wind speed within the sliding window does not exceed 10 m / s, the wind speed data within this sliding window is used as the standard wind speed section. Perform Fourier transform on the standard wind speed section and remove the fundamental frequency component to obtain the standard wind speed spectrum;

[0017] S23: Extract the wind speed spectrum. Traverse all samples in the training set, perform Fourier transform on the wind speed data within each 10-minute sample sliding window, and remove the fundamental frequency component to obtain the spectrum information as the wind speed spectrum of the training samples;

[0018] S24: Traverse all samples in the training set, calculate the inner product of the wind speed spectrum within each sample sliding window and the selected standard wind speed spectrum to obtain a sequence with the same number as the training samples, which is used as the inner product sequence of the training samples.

[0019] Furthermore, in the step S3, the reallocation of sample weights is to perform sample weight allocation based on the obtained inner product sequence of the training samples. Divide the range between the maximum and minimum values of the training sample inner product sequence, set N h interval numbers, and perform histogram statistics to obtain the number of training samples in each interval. The number of samples in the i-th interval is hi ; Calculate the weight W of the samples included in the i-th interval i :

[0020] Further, in step S4, the vibration frequency domain feature extraction is performed by calculating the spectra of multi-channel vibration signals, performing Fourier transform on four vibration data within the sliding window of the training samples, and combining the spectral data of the four channels as the original features of the vibration signal data, and the feature dimension is where T is the length of the sliding window of the training samples, in minutes; then use PCA to reduce the dimension of the spectral features, traverse the training data samples, and use the PCA principal component analysis method to reduce the original frequency domain features from 120T to M dimensions, where M is the output feature dimension of the normal behavior model.

[0021] Further, in step S5, the normal behavior modeling is performed during the offline training process. The constructed training sample set is input, the input data is the training data set with timestamps removed, and the output is the obtained vibration frequency domain features, so as to perform real-time estimation on the frequency domain features. The variable estimation models include machine learning and deep learning methods such as SVR, GBRT, ANN, LSTM, GRU, or WTS-GRU, and calculate the residual sequence of the actual running value and the model estimated value in the training text set.

[0022] Further, in step S6, the training residual distribution includes:

[0023] S61: Obtain the training set residual sequence, and obtain the probability distribution model of the residual sequence through the kernel density estimation method. Among them, the probability density at the residual d x is calculated as follows: where h is the bandwidth, N is the number of training samples, and d i is the training residual of the i-th sample;

[0024] S62: Based on the probability distribution of the obtained training set residual sequence, by setting the confidence level α, obtain the upper threshold upperLimit and lower threshold lowerLimit of the training set residual, where the mathematical calculation formula is: Statistical the maximum number of consecutive times that the training residual sequence is greater than the threshold upper limit or less than the threshold lower limit as the maximum consecutive overrun times train_max_count, and use the upper and lower thresholds of the training set, the maximum training overrun parameter, and the trained model as the final output results in the training process for subsequent use.

[0025] Further, in step S7, in the online application stage, obtain the online real-time running data set as the test set [X test ,y test, the parameter selection is the same as the training set selection, and it is input into the trained variable estimation model to obtain the residual sequence of the actual value of the current operating data set minus the model estimation value.

[0026] Further, in step S7, during the application phase, the input is the real-time operating second-level data collected by the wind turbine SCADA system. When the discrimination criterion is continuous over-limit, it is the 0 / 1 sequence of whether the "blade sweeping the tower alarm" is reported at the current moment, where 0 means no early warning and 1 means early warning, as well as the auxiliary decision-making information corresponding to the real-time operating data; based on the training set, the continuous maximum over-limit times train_max_count is the normal situation. Therefore, the lower limit of the continuous over-limit threshold parameter in the test set is set to train_max_count, and the discrimination criterion is whether the real-time over-limit degree is greater than the training maximum over-limit degree. Finally, the application model outputs the 0 / 1 sequence of real-time alarm, where 0 means no alarm and 1 means alarm.

[0027] Compared with the prior art, the present invention has the following advantages:

[0028] 1) By designing an adaptive allocation strategy for training sample weights, considering the problem that sample imbalance under different wind conditions is likely to cause modeling accuracy imbalance under different wind conditions, higher training weights are given to the poor wind conditions with a smaller proportion in the training samples, ensuring the accuracy and comprehensiveness of the results for faults caused by extreme wind conditions;

[0029] 2) Monitoring the frequency domain characteristics of multi-dimensional vibration variables, the operating conditions of high loads usually first cause vibration anomalies of key components. Effective feature extraction in the frequency domain can ensure the robustness and accuracy of early warning;

[0030] 3) The present invention is a warning strategy design method based on normal behavior modeling for blade sweeping the tower faults. This process is applicable to all variable estimation models and is applicable to all key components of wind turbines with vibration characteristics, having scalability. Description of the Drawings

[0031] Figure 1 is the implementation flowchart of the training model of the present invention;

[0032] Figure 2 is the implementation flowchart of the application model of the present invention;

[0033] Figure 3 is the estimation result diagram of the selected variable estimation model in the embodiment of the present invention;

[0034] Figure 4 is the estimation result residual diagram of the selected variable estimation model in the embodiment of the present invention;

[0035] Figure 5 is the final early warning result diagram in the embodiment of the present invention. Detailed implementation mode

[0036] The following further explains a method for designing an adaptive early warning strategy for the blade tower sweeping fault of a wind turbine generator set of the present invention in conjunction with the accompanying drawings.

[0037] In this embodiment, fault monitoring is carried out on a wind turbine generator set in a wind farm where a blade tower sweeping fault has occurred. At 11:11:00 on December 21, 2022, during routine maintenance, the staff found that a blade tower sweeping fault occurred. It is speculated that the time when the tower sweeping fault occurred was around 21:00 on December 16, 2022. The data collected by the SCADA system of this wind turbine generator set in 2022 is selected for fault monitoring. Among them, the data sampling interval of the SCADA system is 1 s, the data information is for 12 months, and the time range is from 00:00:00 on January 1, 2022 to 23:59:59 on December 31, 2022. The frequency domain characteristics of the relevant parameters used to estimate the vibration frequency domain characteristics of the wind turbine generator set are selected as the target variables, and all parameters that affect the target variable value, such as other operating parameters of the generator and system parameters, are used as relevant variables. The input and output variables of the model are shown in Tables 1 and 2:

[0038] Table 1: Input variable table of the training model of a certain wind turbine in a certain wind farm

[0039]

[0040]

[0041]

[0042] Table 2: Output variable table of the training model of a certain wind turbine in a certain wind farm

[0043] Variable name Variable name (Chinese) model WTS-SP-GRU normal behavior model upperLimit Upper limit of training set residuals lowerLimit Lower limit of training set residuals max_count Maximum number of consecutive over-limit data in the training set

[0044] As Figure 1 and Figure 2 described, the implementation data set of the method for designing the adaptive early warning strategy for the blade tower sweeping fault of the wind turbine generator set in this embodiment is the 12 - month operation data of the above - mentioned wind turbine generator set. The implementation steps of this method are specifically as follows:

[0045] 1) Obtain the SCADA normal behavior data set. Obtain the operation data recorded in the SCADA system of this wind turbine generator set, including four vibration channels and all relevant variables. Select the data in the first 10 months in the normal operation state, that is, the data from 00:00:00 on January 1, 2022 to 23:59:59 on October 31, 2022 as the training set, and the data in the last two months, that is, the data from 00:00:00 on November 1, 2022 to 23:55:00 on December 30, 2022 is constructed as the test set and used as the real - time data set during online application;

[0046] 2) Preprocessing of data. Based on the fault operation and maintenance records of the unit during the data recording period of the SCADA training set, screen the start time and reset time of the faults reported in the fault operation and maintenance records corresponding to the blade-related fault records in the training set during the corresponding period. Take the moment 30 minutes before the fault start time as the starting moment and the moment 30 minutes after the fault reset time as the ending moment to construct the time interval for extracting fault data. Screen out the data corresponding to the relevant fault data time interval in the SCADA training set and eliminate it. Concatenate the finally remaining SCADA data sets and sort them by time as the preprocessed data.

[0047] Specifically, combined with the characteristics of the prior abnormal operating state determined manually, eliminate the data in the training set that is not detected by the SCADA system but does not conform to normal operation; first, eliminate the obvious abnormal data values of "all temperature variables exceed 150 °C or are less than -50 °C", then eliminate the corresponding abnormal data of "the measured wind speed of the wind turbine is greater than the cut-in wind speed, but the power value is too low", then eliminate the corresponding abnormal data of "the measured wind speed of the wind turbine is less than the cut-in wind speed, but the power value is too high", and finally eliminate the corresponding abnormal data of "the measured wind speed of the wind turbine is below the rated wind speed and is operating at power limit". Concatenate the output result data sets and arrange them in chronological order.

[0048] 3) Training sample weight assignment. In order to balance the sample weights of various different wind conditions in historical data, ensure the model performance under special wind conditions such as strong wind and gust, reduce model overfitting, and enhance the reliability of normal behavior modeling, it is necessary to assign weights to training samples. Select the standard wind speed section in the training samples, set the time sliding window length to be the same as the training samples (both are 10 minutes in this example), ensure that the overall wind speed within the sliding window does not exceed 10 m / s, perform Fourier transform on the standard wind speed section and remove the fundamental frequency component to obtain the standard wind speed spectrum.

[0049] Traverse all samples in the training set, perform Fourier transform on the wind speed data within each 10-minute sample sliding window, and remove the fundamental frequency component to obtain the spectral information as the wind speed spectrum of the training sample; then calculate the inner product of the wind speed spectrum within each 10-minute sliding window of the training sample and the standard wind speed spectrum selected in step 4 to obtain a sequence with the same number as the training samples as the inner product sequence of the training sample. The inner product sequence reflects the distance between the wind speed spectrum of each training sample and the standard wind speed spectrum.

[0050] Based on the obtained inner product sequence of the training samples, perform training sample weight assignment, divide the interval within the maximum and minimum values of the training sample inner product sequence, and set N h= 20 number of intervals, and perform histogram statistics to obtain the number of training samples in each interval. The number of samples in the i-th interval is h i ; Calculate the weight W of the samples included in the i-th interval i : The more samples in the interval, it indicates that the wind condition in this interval appears more frequently, and the weight of the samples included in the interval is less. On the contrary, due to the low frequency of extreme wind conditions in the training set, the number of samples in the corresponding interval is relatively small, and the assigned weight Wi is relatively high. This solves the problem that the model cannot accurately model for such wind conditions due to the low frequency of extreme wind conditions in the training set, and ensures the accuracy and comprehensiveness of the early warning.

[0051] 4) Frequency domain feature extraction. Calculate the spectrum of the multi-channel vibration signal, perform Fourier transform on the four vibration data within the sliding window of the training samples, and merge the spectrum data of the four channels as the original features of the vibration signal data. The feature dimension is where T = 10 is the length of the training sample sliding window in minutes; then use PCA to reduce the dimension of the spectrum features, traverse the training data samples, and use the PCA principal component analysis method to reduce the original frequency domain features from 1200 to 63 dimensions.

[0052] 5) Modeling based on normal behavior. During the offline training process, input the constructed training set. The input data is the training data set with timestamps removed, and the output is the vibration frequency domain features obtained in step 7, so as to perform real-time estimation on the frequency domain features. The variable estimation model used in this embodiment is WTS-SP-GRU. This model can comprehensively consider the influence of relevant variables and historical information on the target variable, and assign different influence weights through the attention mechanism, so as to accurately perform real-time estimation on the target variable of the constructed test set during online application. The variable estimation results are as Figure 3 shown. The three subgraphs respectively show the fitting situation of the model for the frequency domain features of 3 dimensions; the solid line is the true value of the target variable of the test set, and the dashed line is the model estimated value of the target variable of the test set. The residual results are as Figure 4 shown.

[0053] 6) Obtain the residual sequence of the training set. Since it is impossible to assume that the residual sequence satisfies any known assumed distribution, the kernel density estimation method is used to obtain the probability distribution model of the residual sequence. Among them, the probability density at the residual d x is calculated as follows: where h is the bandwidth, N is the number of training samples, d iis the training residual of the i-th sample. Based on the probability distribution of the obtained training set residual sequence, by setting the confidence level α, the upper threshold upperLimit = 0.082 and the lower threshold lowerLimit = 0.003 of the training set residual are obtained. The maximum number of consecutive values greater than the upper threshold or less than the lower threshold in the training residual sequence is counted, train_max_count = 3, that is, the maximum consecutive overrun times in the training set is 3. When the maximum consecutive overrun times in the online application stage is greater than 3, an alarm will be issued.

[0054] As Figure 2 shown, in the online application stage, an online real-time operation data set is obtained as the test set. After a series of steps such as preprocessing, [X test , y test is obtained. The parameter selection is the same as that of the training set. It is input into the trained variable estimation model to obtain the residual sequence of the actual value of the current operation data set minus the model estimation value. According to the upper threshold upperLimit and the lower threshold lowerLimit of the training set residual obtained in step 6), it is judged whether the residual of each 10-minute window data is within this interval, and the maximum consecutive overrun value max_count at the current moment is obtained. The discrimination criterion is whether the real-time overrun degree is greater than the training maximum overrun degree. If it is greater, an alarm will be issued at this point. Figure 5 is the final warning result. The maximum consecutive overrun times in the training set are given in the upper left corner. The vertical coordinate gives the time corresponding to the alarm data point. The earliest alarm time is 2022.12.12 21:30:00. An alarm is given at this moment earliest, and a warning is achieved 4 days before the fault occurs. In addition, on the first day after the fault occurs, 2022.12.17 14:30:00, an alarm is also given when the fan is in the fault operation state. This result shows that the invention can not only give accurate warnings before the fault occurs, but also monitor the operation state of the unit, and the result is effective and reliable.

[0055] The design method of the adaptive early warning strategy for the blade tower sweeping fault of the wind turbine unit of the present invention mainly includes data preprocessing, training sample weight assignment, frequency domain feature extraction, normal behavior modeling, determination of the parameter range of the discrimination criterion for the test set, and other links. Figure 1 is the implementation flowchart of the training model of the adaptive early warning strategy for the blade tower sweeping fault of the present invention, Figure 2 is the implementation flowchart of the application model of the adaptive early warning strategy for the blade tower sweeping fault of the present invention, Figure 3 is the estimation result diagram of the variable estimation model selected in the embodiment of the present invention, Figure 4 is the residual diagram of the estimation result of the variable estimation model selected in the embodiment of the present invention, Figure 5It is the final warning result diagram in the embodiments of the present invention. The result shows that the invention can not only give accurate warnings before a fault occurs, but also monitor the operating state of the unit, and the result is effective and reliable.

[0056] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A design method for an adaptive early warning strategy for the fault of a wind turbine blade sweeping the tower, characterized in that, The method includes the following steps: S1: Based on the data sampling of the offshore wind turbine generator set in the SCADA system, obtain the SCADA normal behavior data set; select the key components of the wind turbine generator set to be monitored for blade faults. The frequency domain characteristics of the vibration signals measured by the vibration measurement points of this component in the SCADA system are used as the target variable y, and the parameters X related to the component vibration are used as the relevant variables. Select N pieces of operation data under the normal operation state of the unit, and construct a training set with a ten-minute sliding window i = 1, 2,..., N; S2: Preprocess the data in the data set to construct training samples; S3: Reassign the weights of training samples; the weight reassignment of samples is to perform training sample weight assignment based on the obtained inner product sequence of training samples, divide the interval within the maximum and minimum values of the training sample inner product sequence, set N h the number of intervals, and conduct histogram statistics to obtain the number of training samples in each interval. The number of samples in the i-th interval is h i ; calculate the weight W of the samples included in the i-th interval i : S4: Extract the vibration frequency domain features of the training samples; the vibration frequency domain features are extracted by calculating the spectra of multi-channel vibration signals, performing Fourier transforms on four vibration data within the sliding window of the training samples, and merging the spectral data of the four channels as the original features of the vibration signal data, with the feature dimension being where T is the length of the sliding window of the training samples, in minutes; then use PCA to reduce the dimensionality of the spectral features, traverse the training data samples, and use the PCA principal component analysis method to reduce the original frequency domain features from 120T to M dimensions, where M is the output feature dimension of the normal behavior model; S5: Model normal behavior and calculate the residual sequence; S6: Train the residual distribution, count the maximum number of consecutive values greater than the upper threshold or less than the lower threshold in the training residual sequence, denoted as the consecutive maximum overrun count train_max_count, and use the upper and lower thresholds in the training set, the maximum training overrun parameter, and the trained model as the final output results in the training process; S7: Enter the online application stage.

2. The design method of an adaptive early warning strategy for the tower-sweeping fault of the blade of a wind turbine generator set according to claim 1, characterized in that In step S2, the preprocessing includes: S21: Combine the characteristics of the prior abnormal operating state determined manually to eliminate the abnormal operating data in the training set that has not been detected by the SCADA system; S22: Select the standard wind speed section. Set the time sliding window length to be the same as the training sample. When the overall wind speed within the sliding window does not exceed 10 m / s, use the wind speed data within this sliding window as the standard wind speed section, perform Fourier transform on the standard wind speed section and remove the fundamental frequency component to obtain the standard wind speed spectrum; S23: Extract the wind speed spectrum. Traverse all samples in the training set, perform Fourier transform on the wind speed data within each 10-minute sample sliding window, and remove the fundamental frequency component to obtain the spectral information as the wind speed spectrum of the training sample; S24: Traverse all samples in the training set, calculate the inner product of the wind speed spectrum within each sample sliding window and the selected standard wind speed spectrum to obtain a sequence with the same number as the training samples as the inner product sequence of the training samples.

3. The design method of an adaptive early warning strategy for the blade tower sweeping fault of a wind turbine unit according to claim 1, characterized in that, In step S5, normal behavior modeling is performed during the offline training process. Input the constructed training sample set. The input data is the training data set with the time stamp removed, and the output is the obtained vibration frequency domain feature, so as to perform real-time estimation on the frequency domain feature. The variable estimation model includes machine learning and deep learning methods such as SVR, GBRT, ANN, LSTM, GRU, or WTS-GRU, and calculate the residual sequence between the actual running value and the model estimated value in the training text set.

4. The design method of an adaptive early warning strategy for the blade tower sweeping fault of a wind turbine set according to claim 1, characterized in that In step S6, training the residual distribution includes: S61: Obtain the training set residual sequence, and acquire the probability distribution model of the residual sequence through the kernel density estimation method, where the probability density at the residual d x is calculated as follows: where h is the bandwidth, N is the number of training samples, and d i is the training residual of the i-th sample; S62: Based on the probability distribution of the obtained training set residual sequence, by setting the confidence level α, the upper threshold upperLimit and the lower threshold lowerLimit of the training set residual are obtained, and the mathematical calculation formula is as follows: Statistically, the maximum number of consecutive occurrences greater than the upper threshold or less than the lower threshold in the training residual sequence is the maximum consecutive overrun count train_max_count. The upper and lower thresholds in the training set, the maximum training overrun parameter, and the trained model are used as the final output results in the training process for subsequent use.

5. The design method of an adaptive early warning strategy for the blade tower sweeping fault of a wind turbine unit according to claim 1, characterized in that In step S7, in the online application stage, obtain the online real-time operation data set as the test set [X test , y test . The parameter selection is the same as that of the training set selection, and input it into the variable estimation model that has been trained to obtain the residual sequence of the actual value of the current operation data set minus the estimated value of the model.

6. The design method of an adaptive early warning strategy for the blade tower sweeping fault of a wind turbine according to claim 5, characterized in that, In step S7, during the application stage, the input is the real-time running second-level data collected by the wind turbine SCADA system. The discrimination criterion is the 0 / 1 sequence of whether the "blade sweeping the tower alarm" is reported at the current moment when there is a continuous overrun. 0 means no early warning, 1 means early warning, and the auxiliary decision-making information corresponding to the real-time running data; Based on the consecutive maximum overrun count train_max_count obtained from the training set as the normal situation, the lower limit of the continuous overrun threshold parameter in the test set is set to train_max_count. The discrimination criterion is whether the real-time overrun degree is greater than the training maximum overrun degree. Finally, the application model outputs the 0 / 1 sequence of real-time alarm, 0 means no alarm, and 1 means alarm.

Images