Contrast analysis-based vibration data anomaly detection method for gas turbine generator set

Abstract

The present application relates to the technical field of gas turbine generator, and particularly relates to a vibration data anomaly detection method for gas turbine generator unit based on contrast analysis. The method comprises: continuously collecting continuous vibration data of the gas turbine generator unit; using an improved mean absolute deviation method to identify noise points; using background contrast detection to identify continuous noise; using a threshold selection method based on quantile plot to separate noise and outliers; and using the data after removing noise points and continuous noise to calculate the anomaly degree of the original data to complete anomaly detection. The present application effectively distinguishes noise and outliers while retaining actual anomaly data, avoids mistaking outliers as noise and filtering them out, ensures that the control system can respond to real abnormal conditions in time, and improves the reliability and accuracy of the system.

Description

Technical Field

[0001] This invention relates to the field of gas turbine generator technology, and in particular to a method for detecting abnormal vibration data of gas turbine generator sets based on contrast analysis. Background Technology

[0002] In modern power generation environments, vibration sensors are a core component of gas turbine generator condition monitoring systems. By collecting and analyzing vibration data from gas turbine generator sets, early fault detection and health monitoring can be achieved, reducing unplanned downtime and supporting decision-making in production control systems. However, this vibration data is susceptible to various interference factors, such as environmental noise, electromagnetic interference, and noise from mechanical vibrations. Noise interference is particularly pronounced during wireless transmission, leading to a significant deterioration in data quality. Low-quality data affects the performance of statistical learning and machine learning models, thereby impacting the accuracy and reliability of decision-making.

[0003] While noise and outliers may share similarities in their statistical distributions, their actual impacts are vastly different: noise is typically meaningless interference, while outliers serve as early warning signals of potential faults. Misclassifying outliers as noise and filtering them out could lead to missed critical fault warnings, resulting in delayed equipment maintenance or even serious production accidents. Therefore, accurately distinguishing between noise and outliers is crucial for ensuring system reliability in the condition monitoring of gas turbine generator sets.

[0004] Traditional data processing methods often neglect the importance of noise reduction and filtering of unreliable data. While this may improve processing speed, it also leads to noise and anomaly data mixing, reducing the accuracy of fault detection and anomaly monitoring. This can result in biased analysis results, frequent false alarms, and unnecessary downtime, causing economic losses and production risks for enterprises.

[0005] Noise in vibration data can be categorized into point noise (isolated anomalous data points) and continuous noise (continuous fluctuations over a period of time). Due to the inherent nonlinear characteristics of the system, this noise is difficult to distinguish from actual anomalies, making it challenging to effectively identify using traditional anomaly detection methods. However, in-depth analysis of the characteristics of noise and outliers reveals that noise data typically exhibits irregular patterns and deviates significantly from normal data, providing a basis for developing effective noise detection algorithms. Summary of the Invention

[0006] This invention discloses a method for detecting abnormal vibration data of gas turbine generator sets based on contrast analysis. The specific method is as follows: continuously collect continuous vibration data of gas turbine generator sets;

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

[0008] Continuous noise is identified using background contrast detection.

[0009] A threshold selection method based on quantile maps is used to separate noise and outliers;

[0010] Anomaly detection is completed by calculating the degree of anomaly in the original data after removing noise points and continuous noise.

[0011] Furthermore, the continuous vibration data is divided into sliding windows with L time points as a group, i.e.

[0012] Furthermore, an 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 weights are determined using a linearly decreasing weighting 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 rate at which the weight decreases. a and b are set empirically.

[0018] To ensure that the weights are non-negative and their sum is a constant, normalization is used.

[0019] Calculate ω for the i-th window i The absolute deviations of the mean are as follows:

[0020]

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

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

[0023] Define noise fraction For window differences D i WMAD (Absolute Deviation from Global Mean)global The ratio:

[0024]

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

[0026]

[0027] Where N is the number of windows;

[0028] The difference D between the two windows is updated using Mahalanobis distance. i Each window is converted into a multidimensional feature vector, including: mean, standard deviation, and WMAD;

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

[0030] F i =[μ i ,σ i WMAD i ]

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

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

[0033]

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

[0035]

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

[0037] Define noise fraction For window differences D i ′ and global Mahalanobis distance D′ global The ratio:

[0038]

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

[0040]

[0041] Furthermore, continuous noise is identified using background contrast detection, as detailed below:

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

[0043] Using the background model window ω′ respectively i mean μ b and standard deviation σ b Construct two background model datasets with the same number of time windows as the original dataset;

[0044] Assuming the original dataset has I time windows, then the background mean dataset is obtained as M = [μ]. b1 ,μ b2 ,...,μ bI ], where μ b1 =μ b2 =...=μ bI =μ b Background standard deviation dataset S = [σ] b1 ,σ b2 ,...,σ bI ], where σ b1 =σ b2 =...=σ bI =σ b Let the mean of the background mean dataset M be denoted as . variance is The mean of the background standard deviation dataset S is variance is

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

[0046]

[0047] Assuming the original dataset has I time windows, then calculate the mean of the mean values ​​of all windows in the background model. Variance of the mean of all windows in the background model and the mean of the standard deviations of all windows in the background model. Variance of standard deviation of all windows in the background model

[0048] Calculate ω for each sliding window in the original dataset. i The average value ωm i and standard deviation ωs i :

[0049]

[0050] Calculate the Mahalanobis distance between the mean of each sliding window in the original dataset and the mean of the background model, and the Mahalanobis distance between the standard deviation of each sliding window in the original dataset 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 for that window.

[0053]

[0054] Furthermore, a threshold selection method based on quantile maps is used to separate noise and outliers, as follows:

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

[0056] The X-axis and Y-axis are defined mathematically as follows:

[0057]

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

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

[0060] By comparing the noise scores S of the sliding window k The multiple 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 using the following formula:

[0062] S k ∈{S i |fS k >θ·max{S i}}

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

[0064] Use percentiles to dynamically adjust the threshold:

[0065] θ = percentile(S,p)

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

[0067] Two noise window sets Dω k and Bω k It can be calculated using the mean absolute deviation method and background contrast score, respectively. and The sets are then merged to obtain the final noise window labeling result:

[0068] F=Dω k ∪Bω k

[0069] Furthermore, the degree of anomaly in the original data is calculated to complete anomaly detection. The specific method is as follows:

[0070] In noise-removed time series data, each normal data interval will vary within different time windows. The general equation for this is:

[0071]

[0072] in, This represents the lower bound of the normal range for data collected by the j-th sensor within the i-th window. Let represent the upper bound of the normal range; then the original formula for calculating the degree of abnormality is:

[0073]

[0074] Since the calculation of anomaly degree is closely related to data distribution, when anomaly results are binary classified as 0 or 1, normal data near the anomaly boundary is falsely reported as anomaly. Based on the membership degree of the underlying data distribution, fuzzy theory is introduced to rewrite the anomaly degree calculation formula as follows:

[0075]

[0076] Construct a Bayesian network, where the anomaly degree DecT(i,j,t) of each sensor j is used as a node in the Bayesian network, and define an output node fault to indicate whether an anomaly was detected;

[0077] For each sensor node j, define a conditional probability table P(DecT(i,j,t)|status), where status is normal or abnormal; for the output node fault, define a conditional probability table P(fault|AllS), where AllS is all sensor nodes.

[0078] Using the structure and conditional probability table of the Bayesian network, calculate the joint probability distribution P(all nodes), and based on the joint probability distribution, calculate the probability of the output node being faulty:

[0079] P(fault=true|AllS)

[0080] Data anomaly detection can be achieved by setting a threshold τ;

[0081]

[0082] Due to the adoption of the above technical solutions, the present invention has the following beneficial effects:

[0083] 1. While retaining actual abnormal data, it effectively distinguishes between noise and outliers, preventing outliers from being mistaken for noise and filtered out. This ensures that the control system can respond to real abnormal situations in a timely manner, improving the reliability and accuracy of the system.

[0084] 2. By combining sliding window technology, data can be processed in real time in the data stream, adapting to the needs of online monitoring, timely detection and handling of anomalies, and improving the system's response speed and decision-making efficiency.

[0085] 3. The noise scoring calculation method, which combines the mean absolute deviation method with background contrast, can detect not only isolated point noise but also continuous noise that fluctuates over a period of time, thus improving the comprehensiveness and robustness of noise detection.

[0086] 4. By introducing fuzzy evidence theory and combining it with Bayesian networks, the problem of inconsistent calculation of data with different attributes is eliminated, and the unified processing of heterogeneous data is realized. This enhances the universality and flexibility of the method, enabling the system to better adapt to complex and diverse industrial environments.

[0087] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0088] The accompanying drawings of this invention are described below.

[0089] Figure 1This is a schematic diagram of the overall process of the present invention.

[0090] Figure 2 A schematic diagram of the system structure for implementing the present invention. Detailed Implementation

[0091] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0092] Example 1:

[0093] A method for detecting vibration anomalies in gas turbine generator sets based on contrast analysis, such as... Figure 1 As shown, the specific steps are as follows:

[0094] S1. Continuously collect continuous vibration data of the gas turbine generator set.

[0095] S2. The improved mean absolute deviation method is used to identify noise points.

[0096] In steps S1 and S2, continuous sensing data is collected from the sensor, and each L time points form a group of sliding windows, i.e.

[0097] ω of the i-th window i The weighted average of the data is:

[0098]

[0099] in The weights can be determined using a linearly decreasing weighting method. For each window, the weight can be calculated using the following formula:

[0100] W k =ab(j-iL)

[0101] 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, which is usually the weight of the newest data point in the window, and b controls the rate at which the weight decreases. a and b can be set empirically.

[0102] Ensuring that the weights are non-negative and that their sum is a constant, such as 1, can be achieved through normalization.

[0103] Then calculate ω for the i-th window. i The absolute deviations of the mean are as follows:

[0104]

[0105] 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 D.i This is the absolute value of the difference between the absolute deviations of the means of the two windows:

[0106] D i =|WMAD i -WMAD i-1 |

[0107] Define noise fraction For window differences D i WMAD (Absolute Deviation from Global Mean) global The ratio,

[0108]

[0109] Among them, WMAD global It is the average of the absolute deviations of the means of all windows:

[0110]

[0111] N is the number of windows.

[0112] Considering the correlation between heterogeneous datasets and the variance of different variables, Mahalanobis distance is used to update the difference D between the two windows. i We need to convert each window into a multi-dimensional feature vector. Commonly used features include mean, standard deviation, WMAD, etc. For the i-th window ω i Construct feature vector F i :

[0113] F i =[μ i ,σ i WMAD i ]

[0114] σ i Let be the standard deviation of window i.

[0115] Mahalanobis distance is used to measure the difference between the feature vectors of two windows. For the current window ω i Calculate the Mahalanobis distance D between the window and its left neighbor window. i ′:

[0116]

[0117] Where S is the covariance matrix, which can usually be estimated using the eigenvectors of all windows:

[0118]

[0119] It is the average of the feature vectors of all windows, and N is the number of windows.

[0120] Define noise fraction For window differences D i ′ and global Mahalanobis distance D′ global The ratio,

[0121]

[0122] Among them, D′ global It is the average of the absolute deviations of the means of all windows:

[0123]

[0124] S3. Background contrast detection is used to identify continuous noise.

[0125] The noise score for each sliding window in the previous step can effectively detect point noise, but it cannot detect continuous noise with significant bias. Therefore, to improve the robustness of this method, a background model containing normal windows is created as an aid, and the background contrast is calculated to achieve the detection of continuous noise.

[0126] First, the background model is identified. Based on domain expert knowledge, a clean and noise-free dataset D′ is selected, and it is divided into a series of sliding windows D′={ω′1,ω′2,...,ω′} according to the window size L. n}, and calculate ω′ for each window of the noiseless dataset D′. i The mean μ′ i Sum of variance Σ′ i Since the background model represents the statistical characteristics of noise-free data, the Mahalanobis distance between the relevant statistics of each sliding window of the original dataset and the background model is used as the background contrast noise score.

[0127] First, take the background model window ω′ as an example. i 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 the original dataset has I time windows, the resulting background mean dataset is M = [μ...]. b1 ,μ b2 ,...,μ bI ], where μ b1 =μ b2 =...=μ bI =μ b Background standard deviation dataset S = [σ] b1 ,σ b2 ,...,σ bI ], where σ b1 =σ b2 =...=σ bI =σ bLet the mean of the background mean dataset M be denoted as . variance is The mean of the background standard deviation dataset S is variance is

[0128] First calculate ω′ for each window of the background model. i The mean ωm′ i and standard deviation ωs′ i :

[0129]

[0130] Assuming the original dataset has I time windows, then calculate the mean of the mean values ​​of all windows in the background model. Variance of the mean of all windows in the background model and the mean of the standard deviations of all windows in the background model. Variance of standard deviation of all windows in the background model

[0131] Then calculate ω for each sliding window in the original dataset. i The average value ωm i and standard deviation ωs i :

[0132]

[0133] Calculate the Mahalanobis distance between the mean of each sliding window in the original dataset and the mean of the background model, and the Mahalanobis distance between the standard deviation of each sliding window in the original dataset and the standard deviation of the background model:

[0134]

[0135] Finally, the sum of the mean Mahalanobis distance and variance Mahalanobis distance between each window and the background model gives the background contrast-noise score for that window.

[0136]

[0137] S4. Separate noise and outliers using a threshold selection method based on quantile maps.

[0138] Step S4 uses the mean absolute deviation method and background contrast noise score calculated for each sliding window to separate noise and outliers using a threshold selection method based on quantile maps.

[0139] First, sort the noise scores of all sliding windows and compare them with the quantiles of the theoretical distribution of noise (usually a normal distribution), then plot a QQ plot, where the X-axis represents the quantiles of the theoretical distribution and the Y-axis represents the observed noise scores. The X-axis and Y-axis are mathematically defined as follows:

[0140]

[0141] Where Φ -1 It is the inverse cumulative probability distribution function of the standard normal distribution, satisfying zero mean and unit variance; parameter n represents the number of data points; S(i) is the noise score calculated earlier.

[0142] Ideally, if the data conforms to the theoretical distribution, the data points should be distributed along a straight line; however, if noise or outliers exist, these points will deviate from this 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 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 for the noise window.

[0143] If the noise score of the sliding window exceeds a predetermined threshold, the window is considered a noisy window. Specifically, this is determined by comparing the noise scores S of the sliding windows. k The multiple of the maximum noise score θ·max{S i The selection of a noise window is determined by the following formula:

[0144] S k ∈{S i |fS k >θ·max{S i}}

[0145] Among them, S i It is the noise score, θ is a deviation coefficient that can be dynamically adjusted according to needs, and f is the linear distribution function of the theoretical value.

[0146] To further optimize threshold selection, a dynamic threshold adjustment method can be used. This invention uses percentiles to determine the threshold:

[0147] θ = percentile(S,p)

[0148] Where: S is the noise score for all windows, and p is the percentile (e.g., 95%).

[0149] Two noise window sets Dω k and Bω k It can be calculated using the mean absolute deviation method and background contrast score, respectively. and The sets are then merged to obtain the final noise window labeling result:

[0150] F=Dω k ∪Bω k

[0151] By following the steps above, noise can be effectively separated, thus improving data quality.

[0152] S5. Based on the data after removing noise points and continuous noise, calculate the degree of anomaly in the original data to complete the anomaly detection.

[0153] In this embodiment, anomaly detection of heterogeneous data is achieved within a unified framework by combining fuzzy evidence theory and Bayesian theory. Specifically, in the noise-removed time series data, each normal data interval varies within different time windows, and the general equation for this variation can be expressed as:

[0154]

[0155] in, This represents the lower bound of the normal range for data collected by the j-th sensor within the i-th window. This represents the upper bound of the normal range.

[0156] The original formula for calculating the degree of anomaly is:

[0157]

[0158] Since the calculation of anomaly severity is closely related to data distribution, when anomaly results are binary classified as 0 or 1, normal data near the anomaly boundary is falsely reported as anomalies. Therefore, based on the membership degree of the underlying data distribution, fuzzy theory is introduced to rewrite the anomaly severity calculation formula as follows:

[0159]

[0160] The above formula eliminates the problem of inconsistent calculation for data with different attributes. The magnitude of this value indicates the normal certainty of a given data point; the closer the value is to 1, the higher the certainty that the data point is normal.

[0161] Next, a Bayesian network is constructed, with the anomaly level DecT(i,j,t) of each sensor j serving as a node in the Bayesian network. An output node fault is defined to indicate whether an anomaly was detected.

[0162] For each sensor node j, define a conditional probability table P(DecT(i,j,t)|status), where status can be normal or abnormal. For the output node fault, define a conditional probability table P(fault|AllS), where AllS represents all sensor nodes.

[0163] Using the structure and conditional probability table of the Bayesian network, calculate the joint probability distribution P(all nodes), and based on the joint probability distribution, calculate the probability of the output node being faulty:

[0164] P(fault=true|AllS)

[0165] Data anomaly detection can be achieved by setting a threshold τ.

[0166]

[0167] By combining fuzzy evidence theory with Bayesian networks, the problem of anomaly detection in heterogeneous data can be effectively addressed. This method not only handles uncertainty but also improves the accuracy and robustness of anomaly detection through structured reasoning using Bayesian networks.

[0168] Example 2:

[0169] A system for detecting vibration anomalies in gas turbine generator sets based on contrast analysis, such as... Figure 2 As shown, it mainly includes the following modules:

[0170] Data communication module: This module is responsible for collecting data in real time from vibration sensors distributed throughout the power plant environment and transmitting the data securely and quickly to the central control system via wireless or wired networks. It features efficient data transmission capabilities and good network adaptability, ensuring data real-time performance and integrity. Furthermore, it has a data caching function to temporarily store data when the network is unstable, ensuring no data loss, and uses data encryption technology to guarantee data security during transmission.

[0171] The noisy data cleaning module employs the mean absolute deviation method and background contrast method to denoise the collected raw data. First, the raw data undergoes preliminary cleaning to remove obvious noise and outliers. Then, noise is further eliminated by calculating the mean absolute deviation score between the sliding window and its adjacent windows, and by calculating the background contrast noise score of the sliding window based on a background model. Finally, a QQ plot is used to determine the threshold for noisy windows and to mark these windows, ensuring that subsequent processing modules receive high-quality data.

[0172] The data anomaly detection module utilizes fuzzy evidence theory combined with Bayesian networks to calculate the degree of anomaly in the data within each sliding window. It explicitly models the dependencies between variables using Bayesian networks and handles uncertainty through fuzzy evidence theory to determine whether a data point is an anomaly, filtering out anomalies based on pre-set thresholds. The detection results are then uploaded to the control system, providing crucial information for intelligent decision-making.

[0173] In practical applications, vibration sensors are distributed around the gas turbine generator set to monitor the equipment's operating status 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:

[0174] Data Acquisition and Transmission: Vibration sensors are distributed throughout the power plant environment, monitoring various parameters of the gas turbine generator set in real time. Edge computing servers collect data from these sensors and transmit it securely and quickly to the central control system via an efficient data communication module. The data communication module also features data caching, enabling temporary data storage during network instability to prevent data loss, and employs data encryption technology to ensure data security during transmission.

[0175] Data cleaning and denoising: The acquired raw data is first uploaded to the data analysis layer via the data communication module. The data denoising module cleans the data in the system using a method based on the mean absolute deviation and background contrast. Noise is further removed by calculating the mean absolute deviation score between the sliding window and its adjacent windows, and by calculating the background contrast noise score of the sliding window based on the background model. Finally, a QQ graph is used to determine the threshold of noisy windows, and noisy windows are marked to ensure that subsequent processing modules receive high-quality data.

[0176] Data anomaly detection: After cleaning, the time-series data is used to calculate the degree of anomaly by combining fuzzy evidence theory with Bayesian networks to identify anomalous data. All anomalies are then filtered according to pre-set thresholds. The detection results are uploaded to the control system as an important basis for intelligent decision-making.

[0177] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do 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: Continuous acquisition of continuous vibration data of gas turbine generator sets; An improved mean absolute deviation method is used to identify noise points; Continuous noise is identified using background contrast detection. A threshold selection method based on quantile maps is used to separate noise and outliers; Based on the data after removing noise points and continuous noise, the degree of anomaly in the original data is calculated to complete the anomaly detection. The continuous vibration data is divided into sliding windows with L time points as a group, i.e. 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 weights are determined using a linearly decreasing weighting 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 rate at which the weight decreases. a and b are set empirically. To ensure that the weights are non-negative and their sum is a constant, normalization is used. Calculate ω for the i-th window i The absolute deviations of the mean are as follows: Compare the current window ω i Its left neighbor window ω il The difference between windows is defined as D. i This is the absolute value of the difference between the absolute deviations of the means of the two windows: D i =|WMAD i -WMAD i-1 | Define noise fraction For window differences D i WMAD (Absolute Deviation from Global Mean) global The ratio: Among them, WMAD global It is the average of the absolute deviations of the means of all windows: Where N is the number of windows; The difference D between the two windows is updated using Mahalanobis distance. i Each window is converted into a multidimensional feature vector, including: mean, standard deviation, and WMAD; For the i-th window ω i Construct feature vector F i : F i =[μ i ,s i ,WMAD i ] In the formula, σ i Let i be the standard deviation of window i; For the current window ω i Calculate the Mahalanobis distance D between the window and its left neighbor window. i ′: Where S is the covariance matrix, estimated using the eigenvectors of all windows: It is the average of the feature vectors of all windows, and N is the number of windows; Define noise fraction For window differences D i ′ and global Mahalanobis distance D′ global The ratio: Among them, D′ global It is the average of the absolute deviations of the means of all windows:

2. The method for detecting abnormal vibration data of gas turbine generator sets based on contrast analysis as described in claim 1, characterized in that, Continuous noise is identified using background contrast detection, as detailed below: Based on domain expert knowledge, a clean and noise-free dataset D′ is selected, and it is divided into a series of sliding windows D′={ω1′,ω′2,...,ω′} according to window size L. n }, and calculate ω for each window of the noiseless dataset D′. i The mean μ of ′ i Σ′ and variance i ; Using the background model window ω respectively i The mean μ of ′ b and standard deviation σ b Construct two background model datasets with the same number of time windows as the original dataset; Assuming the original dataset has I time windows, then the background mean dataset is obtained as M = [μ]. b1 ,μ b2 ,...,μ bI ], where μ b1 =μ b2 =...=μ bI =μ b Background standard deviation dataset S = [σ] b1 ,σ b2 ,...,σ bI ], where σ b1 =σ b2 =...=σ bI =σ b Let the mean of the background mean dataset M be denoted as . variance is The mean of the background standard deviation dataset S is variance is Calculate the background model for each window ω i The mean ωm of ′ i ′ and standard deviation ωs i ′: Assuming the original dataset has I time windows, then calculate the mean of the mean values ​​of all windows in the background model. Variance of the mean of all windows in the background model and the mean of the standard deviations of all windows in the background model. Variance of standard deviation of all windows in the background model Calculate ω for each sliding window in the original dataset. i The average value ωm i and standard deviation ωs i : Calculate the Mahalanobis distance between the mean of each sliding window in the original dataset and the mean of the background model, and the Mahalanobis distance between the standard deviation of each sliding window in the original dataset 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 for that window.

3. The method for detecting abnormal vibration data of gas turbine generator sets based on contrast analysis as described in claim 1, characterized in that, A threshold selection method based on quantile maps 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 noise distribution, and plot a QQ plot, where the X-axis represents the quantiles of the theoretical distribution and the Y-axis represents the actual observed noise scores. The X-axis and Y-axis are defined mathematically as follows: Where, Φ -1 It is the inverse cumulative probability distribution function of the standard normal distribution, satisfying zero mean and unit variance; parameter n represents the number of data points; S(i) is the noise score calculated earlier; Select the point with the largest deviation on the QQ graph as the threshold of the noise window; By comparing the noise scores S of the sliding window k The multiple of the maximum noise score θ·max{S i To determine whether to mark it as a noise window; The noise window is selected using the following formula: S k ∈{S i |f-S k >θ·max{S i }} Among them, S i It is the noise score, θ is a deviation coefficient that can be dynamically adjusted according to needs, and f is the linear distribution function of the theoretical value; Use percentiles to dynamically adjust the threshold: θ = percentile(S,p) Where S is the noise score of all windows, and p is the percentile; Two noise window sets Dω k and Bω k It can be calculated using the mean absolute deviation method and background contrast score, respectively. and The sets are then merged to obtain the final noise window labeling result: F=Dω k ∪Bω k 。 4. The method for detecting abnormal vibration data of gas turbine generator sets based on contrast analysis as described in claim 3, characterized in that, Anomaly detection is performed by calculating the degree of anomaly in the original data. The specific method is as follows: In noise-removed time series data, each normal data interval will vary within different time windows. The general equation for this is: in, This represents the lower bound of the normal range for data collected by the j-th sensor within the i-th window. Let represent the upper bound of the normal range; then the original formula for calculating the degree of abnormality is: Since the calculation of anomaly degree is closely related to data distribution, when anomaly results are binary classified as 0 or 1, normal data near the anomaly boundary is falsely reported as anomaly. Based on the membership degree of the underlying data distribution, fuzzy theory is introduced to rewrite the anomaly degree calculation formula as follows: Construct a Bayesian network, where the anomaly degree DecT(i,j,t) of each sensor j is used as a node in the Bayesian network, and define an output node fault to indicate whether an anomaly was detected; For each sensor node j, define a conditional probability table P(DecT(i,j,t)|status), where status is normal or abnormal; for the output node fault, define a conditional probability table P(fault|AllS), where AllS is all sensor nodes. Using the structure and conditional probability table of the Bayesian network, calculate the joint probability distribution P(all nodes), and based on the joint probability distribution, calculate the probability of the output node being faulty: P(fault=true|AllS) Data anomaly detection can be achieved by setting a threshold τ;

Images