Gas turbine generator set vibration data anomaly detection method based on contrast analysis

Abstract

The invention relates to the technical field of gas turbine generators, in particular to a gas turbine generator set vibration data anomaly detection method based on contrast analysis. The method comprises the following steps: continuously acquiring continuous vibration data of the gas turbine generator set; identifying noise points by adopting an improved mean absolute deviation method; adopting background contrast detection to identify continuous noise; noise and abnormal values are separated according to a threshold selection method based on a quantile map; and according to the data after the noise points and the continuous noise are removed, calculating the anomaly degree of the original data to complete anomaly detection. Noise and abnormal values are effectively distinguished while actual abnormal data are reserved, the situation that the abnormal values are mistakenly considered as noise and filtered out is avoided, it is ensured that a control system can respond to real abnormal conditions in time, and the reliability and accuracy of the system are improved.

Description

Technical Field

[0001] The invention relates to the technical field of gas turbine generators, and in particular to a method for detecting anomaly in vibration data of a gas turbine generator set based on contrast analysis. Background Art

[0002] In modern power production environments, vibration sensors are one of the core components of gas turbine generator condition monitoring systems. By collecting and analyzing vibration data from gas turbine generator sets, early fault detection and health status monitoring can be achieved, unplanned downtime can be reduced, and support can be provided for decision-making in production control systems. However, these vibration data are susceptible to a variety of interference factors, such as environmental noise, electromagnetic interference, and clutter in mechanical vibration. Especially during wireless transmission, noise interference is particularly evident, resulting in a significant decrease in data quality. Low-quality data will affect the performance of statistical learning and machine learning models, and thus affect the accuracy and reliability of decision-making.

[0003] Although noise and outliers may have similarities in statistical distribution, their actual impacts are quite different: noise is usually meaningless interference information, while outliers are early warning signals of potential failures. If outliers are misjudged as noise and filtered out, critical fault warnings may be missed, resulting in equipment failure to be maintained in a timely manner, or even causing serious production accidents. Therefore, in the condition monitoring of gas turbine generator sets, accurately distinguishing between noise and outliers is crucial to ensure system reliability.

[0004] Traditional data processing methods often ignore the importance of denoising and filtering unreliable data. Although the processing speed is improved, it also makes noise and abnormal data unclear, reducing the accuracy of fault detection and abnormal monitoring. This may lead to biased analysis results, frequent false alarms and unnecessary downtime, causing economic losses and production risks to enterprises.

[0005] The noise in vibration data can be divided into point noise (isolated abnormal data points) and continuous noise (continuous fluctuations over a period of time). Due to the inherent nonlinear characteristics of the system, these noises are difficult to distinguish from actual anomalies, and traditional anomaly detection methods are difficult to effectively identify. However, through in-depth analysis of noise and outlier characteristics, it is found that noise data usually exhibits irregular patterns and deviates greatly from normal data, which provides a basis for the development of effective noise detection algorithms. Summary of the invention

[0006] The present invention discloses a method for detecting abnormal vibration data of a gas turbine generator set based on contrast analysis, and the specific method is as follows: continuously collecting continuous vibration data of the gas turbine generator set;

[0007] The improved mean absolute deviation method is used to identify noise points;

[0008] Use background contrast detection to identify continuous noise;

[0009] Separate noise and outliers based on quantile map-based threshold selection method;

[0010] Based on the data after removing noise points and continuous noise, the abnormality degree of the original data is calculated to complete anomaly detection.

[0011] Furthermore, the continuous vibration data is divided into a group of sliding windows at every L time points, that is,

[0012] Furthermore, the improved mean absolute deviation method is used to identify noise points. The specific method is as follows:

[0013] ω of the i-th window i The weighted average of the data is:

[0014]

[0015] in, The weight is determined by the linear decreasing weight method. For each window, the weight is calculated using the following formula:

[0016] W k =ab(j-iL)

[0017] Where j is the position of the current data point in the time series, i is the index of the current window, L is the window size, a is the maximum weight, and b controls the speed at which the weight decreases. a and b are set empirically.

[0018] Ensure that weights are non-negative and sum to a constant, by normalizing them;

[0019] Calculate the i-th window ω i The mean absolute deviation of is as follows:

[0020]

[0021] Compare current windowω i Its left neighbor window ω il The difference between windows is defined as D i is the absolute value of the difference between the mean absolute deviations of the two windows:

[0022] D i =|WMAD i -WMAD i-1 |

[0023] Defining the Noise Fraction is the difference between windows D i The absolute deviation from the global mean WMADglobal Ratio of:

[0024]

[0025] Among them, WMAD global is the average of the mean absolute deviations over all windows:

[0026]

[0027] Where N is the number of windows;

[0028] Use Mahalanobis distance to update the difference D between the two windows i , convert each window into a multi-dimensional feature vector, including: mean, standard deviation and WMAD;

[0029] For the i-th window ω i , construct the feature vector F i :

[0030] F i =[μ i ,σ i ,WMAD i ]

[0031] In the formula, σ i is the standard deviation of window i;

[0032] For the current window ω i Calculate the Mahalanobis distance D between the left neighbor window and the i ′:

[0033]

[0034] Where S is the covariance matrix, estimated by the eigenvectors of all windows:

[0035]

[0036] is the average value of all window feature vectors, and N is the number of windows;

[0037] Defining the Noise Fraction is the difference between windows D i ′ and the global Mahalanobis distance D′ global Ratio of:

[0038]

[0039] Among them, D′ global is the average of the mean absolute deviations over all windows:

[0040]

[0041] Furthermore, background contrast detection is used to identify continuous noise. The specific method is as follows:

[0042] A clean and noise-free dataset D′ is selected based on domain expert knowledge and divided into a series of sliding windows D′={ω′1,ω′2,...,ω′ according to the window size L. n}, and calculate the noise-free data set D′ for each window ω′ i The mean μ′ i and variance Σ′ i ;

[0043] The background model window ω′ i The mean μ b and standard deviation σ b Construct two background model data sets with the same number of time windows as the original data set;

[0044] Assuming that the number of time windows in the original data set is I, the background mean data set M = [μ b1 ,μ b2 ,...,μ bI ], where μ b1 =μ b2 =... =μ bI =μ b ; Background standard deviation data set S = [σ b1 ,σ b2 ,...,σ bI ], where σ b1 =σ b2 =... =σ bI =σ b . Let the mean of the background mean data set M be The variance is The mean of the background standard deviation dataset S is The variance is

[0045] Calculate the background model for each window ω′ i The mean value ωm′ i and standard deviation ωs′ i :

[0046]

[0047] Assume that the number of time windows in the original data set is I, and then calculate the mean of all window means of the background model The variance of the mean of all windows of the background model And the mean of the standard deviation of all windows of the background model Variance of the standard deviation of all windows of the background model

[0048] Calculate each sliding window ω of the original data set i The average value of ωm i and standard deviation ωs i :

[0049]

[0050] Calculate the Mahalanobis distance between the mean of each sliding window of the original data set and the mean of the background model, as well as the Mahalanobis distance between the standard deviation of each sliding window of the original data set and the standard deviation of the background model:

[0051]

[0052] The sum of the mean Mahalanobis distance and variance Mahalanobis distance between each window and the background model is the background contrast noise score of the window.

[0053]

[0054] Furthermore, the noise and outliers are separated according to the threshold selection method based on the quantile map. The specific method is as follows:

[0055] Sort the noise scores of all sliding windows and compare them with the quantiles of the theoretical distribution of noise, and draw a QQ graph, where the X-axis represents the quantile of the theoretical distribution and the Y-axis represents the actual observed noise score;

[0056] The mathematical definitions of the X-axis and Y-axis are as follows:

[0057]

[0058] Among them, Φ -1 is the inverse cumulative probability distribution function of the standard normal distribution, satisfying zero mean and unit variance; the parameter n represents the number of data points; S(i) is the noise fraction calculated previously;

[0059] Select the point with the largest deviation on the QQ graph as the threshold of the noise window;

[0060] By comparing the noise score S of the sliding window k Multiples of the maximum noise score θ·max{S i} to determine whether to mark it as a noise window;

[0061] The noise window can be selected by the following formula:

[0062] S k ∈{S i |fS k >θ·max{S i}}

[0063] Among them, S i is the noise score, θ is a deviation coefficient that can be adjusted dynamically according to demand, and f is the linear distribution function of the theoretical value;

[0064] Use percentiles to dynamically adjust thresholds:

[0065] θ=percentile(S,p)

[0066] Where S is the noise fraction of all windows and p is the percentile;

[0067] Two noise window sets Dω k and Bω k It can be scored by the mean absolute deviation method and background contrast score respectively. and Get, merge the sets to get the final noise window marking result:

[0068] F=Dω k ∪Bω k .

[0070] Furthermore, the abnormality degree of the original data is calculated to complete the abnormality detection. The specific method is as follows:

[0071] In the time series data from which noise has been removed, each normal data interval will change in different time windows, and its general equation can be expressed as:

[0072]

[0073] in, represents the lower bound of the normal interval of data collected by the jth sensor in the i-th window, represents the upper limit of the normal interval; then the original abnormality degree calculation formula is:

[0074]

[0075] Since the calculation of the degree of abnormality is closely related to the data distribution, when the abnormal result is classified as 0 or 1, the normal data close to the abnormal boundary is mistakenly reported as abnormal. Based on the membership of the underlying data distribution, the fuzzy theory is introduced to rewrite the calculation formula of the degree of abnormality as follows:

[0076]

[0077] Construct a Bayesian network, use the abnormality degree DecT(i,j,t) of each sensor j as a node in the Bayesian network, and define an output node fault to indicate whether an abnormality is detected;

[0078] For each sensor node j, a conditional probability table P(DecT(i,j,t)|status) is defined, where status is normal or abnormal; for the output node fault, a conditional probability table P(fault|AllS) is defined, where AllS is all sensor nodes;

[0079] Using the structure of the Bayesian network and the conditional probability table, calculate the joint probability distribution P (all nodes), and calculate the probability of the output node fault based on the joint probability distribution:

[0080] P(fault=true|AllS)

[0081] Data anomaly detection can be achieved by setting the threshold τ;

[0082]

[0083] Due to the adoption of the above technical solution, the present invention has the following beneficial effects:

[0084] 1. While retaining the actual abnormal data, it effectively distinguishes between noise and outliers, avoiding the outliers being mistaken for noise and filtered out, ensuring that the control system can respond to real abnormal situations in a timely manner, and improving the reliability and accuracy of the system.

[0085] 2. Combined with sliding window technology, it can process data in real time in the data stream, adapt to the needs of online monitoring, promptly detect and handle abnormal situations, and improve the system's response speed and decision-making efficiency.

[0086] 3. The noise score calculation method that combines the mean absolute deviation method with background contrast can not only detect isolated point noise, but also detect continuous noise that fluctuates continuously over a period of time, which improves the comprehensiveness and robustness of noise detection.

[0087] 4. The introduction of fuzzy evidence theory combined with Bayesian network eliminates the problem that data with different attributes cannot be calculated uniformly, realizes the unified processing of heterogeneous data, enhances the universality and flexibility of the method, and enables the system to better adapt to complex and diverse industrial environments.

[0088] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be achieved and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] The accompanying drawings of the present invention are as follows.

[0090] Figure 1It is a schematic diagram of the overall process of the present invention.

[0091] Figure 2 The figure is a schematic diagram of the system structure for implementing the present invention. DETAILED DESCRIPTION

[0092] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0093] Embodiment 1:

[0094] A gas turbine generator set vibration data anomaly detection method based on contrast analysis, such as Figure 1 As shown, the specific steps are as follows:

[0095] S1. Continuously collect continuous vibration data of gas turbine generator sets.

[0096] S2. Use the improved mean absolute deviation method to identify noise points.

[0097] In steps S1 and S2, continuous sensing data is collected from the sensor, and each L time points is divided into a group of sliding windows, that is,

[0098] ω of the i-th window i The weighted average of the data is:

[0099]

[0100] in The weight can be determined by the linear decreasing weight method. For each window, the weight can be calculated using the following formula:

[0101] W k =ab(j-iL)

[0102] Where j is the position of the current data point in the time series, i is the index of the current window, L is the window size, a is the maximum weight, usually the weight of the latest data point in the window, and b controls the speed at which the weight is reduced. a and b can be set empirically.

[0103] Ensuring that the weights are non-negative and sum to a constant, such as 1, can be achieved by normalization.

[0104] Then calculate the i-th window ω i The mean absolute deviation of is as follows:

[0105]

[0106] To detect noise, we need to compare the current window ω i Its left neighbor window ω il The difference between windows is defined as Di is the absolute value of the difference between the mean absolute deviations of the two windows:

[0107] D i =|WMAD i -WMAD i-1 |

[0108] Defining the Noise Fraction is the difference between windows D i The absolute deviation from the global mean WMAD global The ratio of

[0109]

[0110] Among them, WMAD global is the average of the mean absolute deviations over all windows:

[0111]

[0112] N is the number of windows.

[0113] Considering the correlation between heterogeneous data sets and the variance of different variables, the Mahalanobis distance is used to update the difference D between the two windows. i We need to convert each window into a multidimensional feature vector. Common features include mean, standard deviation, WMAD, etc. For the i-th window ω i , construct the feature vector F i :

[0114] F i =[μ i ,σ i ,WMAD i ]

[0115] σ i is the standard deviation of window i.

[0116] The Mahalanobis distance is used to measure the difference between the feature vectors of two windows. i Calculate the Mahalanobis distance D between the left neighbor window and the i ′:

[0117]

[0118] where S is the covariance matrix, which can usually be estimated by the eigenvectors of all windows:

[0119]

[0120] is the average of all window feature vectors, and N is the number of windows.

[0121] Defining the Noise Fraction is the difference between windows D i ′ and the global Mahalanobis distance D′ global The ratio of

[0122]

[0123] Among them, D′ global is the average of the mean absolute deviations over all windows:

[0124]

[0125] S3. Use background contrast detection to identify continuous noise.

[0126] The noise score of each sliding window in the previous step can effectively detect point noise, but cannot detect continuous noise with significant deviation. Therefore, in order to improve the robustness of this method, a background model containing a normal window is created as an auxiliary, and the background contrast is calculated to achieve the detection of continuous noise.

[0127] The first step is to confirm the background model. Based on the domain expert knowledge, a clean and noise-free dataset D′ is selected and divided into a series of sliding windows D′={ω′1,ω′2,...,ω′ n}, and calculate the noise-free data set D′ for each window ω′ i The mean μ′ i and variance Σ′ i Because the background model represents the statistical characteristics of noise-free data, the Mahalanobis distance between the correlation statistics of the original data set and each sliding window of the background model is used as the background contrast noise score.

[0128] First, the background model window ω′ i The mean μ b and standard deviation σ b Construct two background model datasets with the same number of time windows as the original dataset. Assuming that the number of time windows of the original dataset is I, the background mean dataset M = [μ b1 ,μ b2 ,...,μ bI ], where μ b1 =μ b2 =... =μ bI =μ b ; Background standard deviation data set S = [σ b1 ,σ b2 ,...,σ bI ], where σ b1 =σ b2 =... =σ bI =σ b. Let the mean of the background mean data set M be The variance is The mean of the background standard deviation dataset S is The variance is

[0129] First calculate each window ω′ of the background model i The mean value ωm′ i and standard deviation ωs′ i :

[0130]

[0131] Assume that the number of time windows in the original data set is I, and then calculate the mean of all window means of the background model The variance of the mean of all windows of the background model And the mean of the standard deviation of all windows of the background model Variance of the standard deviation of all windows of the background model

[0132] Then calculate each sliding window ω of the original data set i The average value of ωm i and standard deviation ωs i :

[0133]

[0134] Calculate the Mahalanobis distance between the mean of each sliding window of the original data set and the mean of the background model, as well as the Mahalanobis distance between the standard deviation of each sliding window of the original data set and the standard deviation of the background model:

[0135]

[0136] Finally, the sum of the mean Mahalanobis distance and variance Mahalanobis distance between each window and the background model is the background contrast noise score of the window.

[0137]

[0138] S4. Separate noise and outliers based on the threshold selection method based on the quantile map.

[0139] Step S4 separates noise and outliers based on the mean absolute deviation method and background contrast noise score calculated for each sliding window and the threshold selection method based on the quantile map.

[0140] First, sort the noise scores of all sliding windows and compare them with the quantiles of the theoretical distribution of noise (usually normal distribution), and draw a QQ graph, where the X-axis represents the quantile of the theoretical distribution and the Y-axis represents the actual observed noise score. The mathematical definitions of the X-axis and Y-axis are as follows:

[0141]

[0142] where Φ -1 is the inverse cumulative probability distribution function of the standard normal distribution, satisfying zero mean and unit variance; the parameter n represents the number of data points; S(i) is the noise fraction calculated previously.

[0143] Ideally, if the data conforms to the theoretical distribution, the data points should be distributed along a straight line; however, if there is noise or outliers, these points will deviate from this straight line on the QQ plot. By analyzing the distribution of data points on the QQ plot, it can be found that some points deviate significantly from the straight line. The sliding windows corresponding to these points are likely to contain noise or outliers, and the point with the largest deviation is selected as the threshold of the noise window.

[0144] If the noise score of the sliding window exceeds the determined threshold, the window is considered to be a noisy window. Specifically, by comparing the noise score S of the sliding window k Multiples of the maximum noise score θ·max{S i} to determine whether to mark it as a noise window. The selection of the noise window can be achieved by the following formula:

[0145] S k ∈{S i |fS k >θ·max{S i}}

[0146] Among them, S i is the noise score, θ is a deviation coefficient that can be adjusted dynamically according to demand, and f is the linear distribution function of the theoretical value.

[0147] In order to further optimize the threshold selection, a method of dynamically adjusting the threshold can be used. The present invention uses percentiles to determine the threshold:

[0148] θ=percentile(S,p)

[0149] Where: S is the noise fraction of all windows and p is the percentile (e.g., 95%).

[0150] Two noise window sets Dω k and Bω k It can be scored by the mean absolute deviation method and background contrast score respectively. and Get, merge the sets to get the final noise window marking result:

[0151] F=Dω k ∪Bω k

[0152] Through the above steps, noise can be effectively separated and the quality of data can be improved.

[0153] S5. Based on the data after removing the noise points and continuous noise, the abnormality degree of the original data is calculated to complete the abnormality detection.

[0154] In this embodiment, the anomaly detection of heterogeneous data is realized in a unified framework by combining fuzzy evidence theory with Bayesian theory. The specific method is that each normal data interval in the time series data from which noise has been removed will change in different time windows, and its general equation can be expressed as:

[0155]

[0156] in, represents the lower bound of the normal interval of data collected by the jth sensor in the i-th window, Indicates the upper bound of the normal interval.

[0157] The original abnormality degree calculation formula is:

[0158]

[0159] Since the calculation of the degree of abnormality is closely related to the data distribution, when the abnormal result is classified as 0 or 1, the normal data close to the abnormal boundary is mistakenly reported as abnormal. Therefore, based on the membership of the underlying data distribution, the fuzzy theory is introduced to rewrite the calculation formula of the degree of abnormality as follows:

[0160]

[0161] The above formula eliminates the problem that data with different attributes cannot be calculated uniformly. The size of this value indicates the certainty that a given data point is normal. The closer the value is to 1, the higher the certainty that the data point is normal.

[0162] Next, we construct a Bayesian network, with the abnormality degree DecT(i,j,t) of each sensor j as a node in the Bayesian network, and define an output node fault to indicate whether an abnormality is detected.

[0163] For each sensor node j, a conditional probability table P(DecT(i,j,t)|status) is defined, where status can be normal or abnormal. For the output node fault, a conditional probability table P(fault|AllS) is defined, where AllS is all sensor nodes.

[0164] Using the structure of the Bayesian network and the conditional probability table, calculate the joint probability distribution P (all nodes), and calculate the probability of the output node fault based on the joint probability distribution:

[0165] P(fault=true|AllS)

[0166] Data anomaly detection can be achieved by setting the threshold τ.

[0167]

[0168] By combining fuzzy evidence theory with Bayesian networks, the anomaly detection problem of heterogeneous data can be effectively handled. This method can not only handle uncertainty, but also improve the accuracy and robustness of anomaly detection through structured reasoning of Bayesian networks.

[0169] Embodiment 2:

[0170] A gas turbine generator set vibration data anomaly detection system based on contrast analysis, such as Figure 2 As shown, it mainly includes the following modules:

[0171] Data communication module: responsible for real-time data collection from vibration sensors distributed in the power plant environment, and transmits the data to the central control system securely and quickly through wireless or wired networks. This module has efficient data transmission capabilities and good network adaptability to ensure the real-time and integrity of data. In addition, it also has a data cache function, which can temporarily store data when the network is unstable to ensure that data will not be lost, and ensure the security of data during transmission through data encryption technology.

[0172] Noise data cleaning module: The collected raw data is denoised using the mean absolute deviation method and background contrast method. First, the raw data is preliminarily cleaned to remove obvious noise and outliers; then, the noise is further removed by calculating the mean absolute deviation method score between the sliding window and its adjacent windows, and calculating the background contrast noise score of the sliding window based on the background model. Finally, the threshold of the noise window is determined using the QQ graph, and the noise window is marked to ensure that the subsequent processing module receives high-quality data.

[0173] Data anomaly detection module: Utilizes fuzzy evidence theory combined with Bayesian network to calculate the degree of anomaly of data in each sliding window. Use Bayesian network to explicitly model the dependency between variables, and use fuzzy evidence theory to handle uncertainty to determine whether a data point is abnormal data, and filter out abnormal data based on pre-set thresholds. The detection results are then uploaded to the control system, providing an important basis for intelligent decision-making.

[0174] In the specific application, vibration sensors are distributed around the gas turbine generator set to monitor the operating status of the equipment in real time. Each area is assigned to a corresponding edge computing server to process the data from the vibration sensors. The specific process is as follows:

[0175] Data collection and transmission: Vibration sensors are distributed throughout the power plant environment to monitor various parameters of the gas turbine generator set in real time. The edge computing server is responsible for collecting data from these sensors and transmitting the data to the central control system securely and quickly through an efficient data communication module. The data communication module also has a data cache function that can temporarily store data when the network is unstable to ensure that data is not lost, and uses data encryption technology to ensure the security of data during transmission.

[0176] Data cleaning and denoising: The acquired raw data is first uploaded to the data analysis layer through the data communication module. The data denoising module uses the mean absolute deviation method and background contrast method to clean the data in the system. By calculating the mean absolute deviation method score between the sliding window and its adjacent windows, and calculating the background contrast noise score of the sliding window based on the background model, the noise is further removed, and the threshold of the noise window is determined by the QQ graph, and finally the noise window is marked to ensure that the subsequent processing module receives high-quality data.

[0177] Data anomaly detection: After cleaning, the time series data is calculated through fuzzy evidence theory combined with Bayesian network to determine the abnormal data, and all anomalies are screened according to the pre-set threshold. The detection results are uploaded to the control system as an important basis for intelligent decision-making.

[0178] 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 above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for detecting abnormal vibration data of a gas turbine generator set based on contrast analysis, characterized in that: The specific method is as follows: Continuously collect continuous vibration data of gas turbine generator sets; The improved mean absolute deviation method is used to identify noise points; Use background contrast detection to identify continuous noise; Separate noise and outliers based on quantile map-based threshold selection method; Based on the data after removing noise points and continuous noise, the abnormality degree of the original data is calculated to complete anomaly detection.

2. The method for detecting abnormal vibration data of a gas turbine generator set based on contrast analysis according to claim 1, characterized in that: The continuous vibration data is divided into a group of sliding windows at every L time points, that is, 3. The method for detecting abnormal vibration data of a gas turbine generator set based on contrast analysis according to claim 2, characterized in that: The improved mean absolute deviation method is used to identify noise points. The specific method is as follows: ω of the i-th window i The weighted average of the data is: in, The weight is determined by the linear decreasing weight method. For each window, the weight is calculated using the following formula: W k ab(j-iL) Where j is the position of the current data point in the time series, i is the index of the current window, L is the window size, a is the maximum weight, and b controls the speed at which the weight decreases. a and b are set empirically. Ensure that weights are non-negative and sum to a constant, by normalizing them; Calculate the i-th window ω i The mean absolute deviation of is as follows: Compare current windowω i Its left neighbor window ω il The difference between windows is defined as D i is the absolute value of the difference between the mean absolute deviations of the two windows: D i =|WMAD i -WMAD i-1 | Define the noise fraction S MADi is the difference between windows D i The absolute deviation from the global mean WMAD global Ratio of: Among them, WMAD global is the average of the mean absolute deviations over all windows: Where N is the number of windows; Use Mahalanobis distance to update the difference D between the two windows i , convert each window into a multi-dimensional feature vector, including: mean, standard deviation and WMAD; For the i-th window ω i , construct the feature vector F i : F i =[μ i ,s i ,WMAD i ] In the formula, σ i is the standard deviation of window i; For the current window ω i Calculate the Mahalanobis distance D between the left neighbor window and the i ′: Where S is the covariance matrix, estimated by the eigenvectors of all windows: is the average value of all window feature vectors, and N is the number of windows; Defining the Noise Fraction is the difference between windows D i ′ and the global Mahalanobis distance D′ global Ratio of: Among them, D′ global is the average of the mean absolute deviations over all windows:

4. The method for detecting abnormal vibration data of a gas turbine generator set based on contrast analysis according to claim 2, characterized in that: Background contrast detection is used to identify continuous noise. The specific method is as follows: A clean and noise-free dataset D′ is selected based on domain expert knowledge and divided into a series of sliding windows D′={ω′1,ω′2,...,ω′ according to the window size L. n }, and calculate the noise-free data set D′ for each window ω′ i The mean μ′ i and variance Σ′ i ; The background model window ω′ i The mean μ b and standard deviation σ b Construct two background model data sets with the same number of time windows as the original data set; Assuming that the number of time windows in the original data set is I, the background mean data set M = [μ b1 ,μ b2 ,...,μ bI ], where μ b1 =μ b2 =... =μ bI =μ b ; Background standard deviation data set S = [σ b1 ,σ b2 ,...,σ bI ], where σ b1 =σ b2 =... =σ bI =σ b . Let the mean of the background mean data set M be The variance is The mean of the background standard deviation dataset S is The variance is Calculate the background model for each window ω′ i The mean value ωm′ i and standard deviation ωs′ i : Assume that the number of time windows in the original data set is I, and then calculate the mean of all window means of the background model The variance of the mean of all windows of the background model And the mean of the standard deviation of all windows of the background model Variance of the standard deviation of all windows of the background model Calculate each sliding window ω of the original data set i The average value of ωm i and standard deviation ωs i : Calculate the Mahalanobis distance between the mean of each sliding window of the original data set and the mean of the background model, as well as the Mahalanobis distance between the standard deviation of each sliding window of the original data set and the standard deviation of the background model: The sum of the mean Mahalanobis distance and variance Mahalanobis distance between each window and the background model is the background contrast noise score of the window.

5. The method for detecting abnormal vibration data of a gas turbine generator set based on contrast analysis according to claim 2, characterized in that: The threshold selection method based on the quantile map is used to separate noise and outliers. The specific method is as follows: Sort the noise scores of all sliding windows and compare them with the quantiles of the theoretical distribution of noise, and draw a QQ graph, where the X-axis represents the quantile of the theoretical distribution and the Y-axis represents the actual observed noise score; The mathematical definitions of the X-axis and Y-axis are as follows: Among them, Φ -1 is the inverse cumulative probability distribution function of the standard normal distribution, satisfying zero mean and unit variance; the parameter n represents the number of data points; S(i) is the noise fraction calculated previously; Select the point with the largest deviation on the QQ graph as the threshold of the noise window; By comparing the noise score S of the sliding window k Multiples of the maximum noise score θ·max{S i } to determine whether to mark it as a noise window; The noise window can be selected by the following formula: S k ∈{S i |f-S k >θ·max{S i }} Among them, S i is the noise score, θ is a deviation coefficient that can be adjusted dynamically according to demand, and f is the linear distribution function of the theoretical value; Use percentiles to dynamically adjust thresholds: θ=percentile(S,p) Where S is the noise fraction of all windows and p is the percentile; Two noise window sets Dω k and Bω k It can be scored by the mean absolute deviation method and background contrast score respectively. and Get, merge the sets to get the final noise window marking result: F=Dω k ∪Bω k 。 6. The method for detecting abnormal vibration data of a gas turbine generator set based on contrast analysis according to claim 5, characterized in that: Calculate the abnormality degree of the original data to complete anomaly detection. The specific method is as follows: In the time series data from which noise has been removed, each normal data interval will change in different time windows, and its general equation can be expressed as: in, represents the lower bound of the normal interval of data collected by the jth sensor in the i-th window, represents the upper limit of the normal interval; then the original abnormality degree calculation formula is: Since the calculation of the degree of abnormality is closely related to the data distribution, when the abnormal result is classified as 0 or 1, the normal data close to the abnormal boundary is mistakenly reported as abnormal. Based on the membership of the underlying data distribution, the fuzzy theory is introduced to rewrite the calculation formula of the degree of abnormality as follows: Construct a Bayesian network, use the abnormality degree DecT(i,j,t) of each sensor j as a node in the Bayesian network, and define an output node fault to indicate whether an abnormality is detected; For each sensor node j, a conditional probability table P(DecT(i,j,t)|status) is defined, where status is normal or abnormal; for the output node fault, a conditional probability table P(fault|AllS) is defined, where AllS is all sensor nodes; Using the structure of the Bayesian network and the conditional probability table, calculate the joint probability distribution P (all nodes), and calculate the probability of the output node fault based on the joint probability distribution: P(fault=true|AllS) Data anomaly detection can be achieved by setting the threshold τ;

Images