Gas turbine unit data abnormity identification method based on group decision

Abstract

The invention relates to the technical field of gas turbine unit data optimization processing, and provides a gas turbine unit data anomaly identification method based on group decision, which comprises the following steps: step 1, preprocessing original operation data; 2, performing data detection by adopting multiple detection methods to obtain an abnormal score of each detection method; 3, adopting a voting mechanism to judge abnormal points; and step 4, outputting all abnormal points which are finally judged. The detection method comprises but not limited to a standard deviation method, an isolated forest method, a DBSCAN clustering method, an OCSVM method, an LOF method, a CBLOF method, a KNN method and an HBOS method. According to the method, a plurality of different abnormal data identification methods can be integrated, the idea of majority voting is used, an abnormal value is selected by a proposed group decision-making method, and the abnormal value can be better detected.

Description

Technical Field

[0001] The present invention relates to the technical field of gas turbine unit data optimization processing, and in particular to a method for identifying gas turbine unit data anomalies based on group decision-making. Background Art

[0002] To accurately understand the operating status of gas turbines, sensors are typically installed on various components to collect operational data. As gas turbine operating data continues to accumulate, the data extracted from these large samples can provide strategic support for power plant operations and maintenance. Operations and maintenance personnel collect data from various sensors, but this data can be subject to deviations due to sensor performance and environmental factors. Analyzing this discrepant information can lead to strategic errors and financial consequences for the power plant, making the processing of outlier data particularly important.

[0003] To achieve better results, choosing the right data processing method is crucial. Data sets collected by sensors often contain a large number of outliers due to factors like harsh environmental interference and hardware failures. The theory of abnormal data cleaning was first proposed in the field of statistics. The main approach to abnormal data cleaning is to detect outliers using detection algorithms and then appropriately process them based on actual business needs. Numerous scholars, both domestically and internationally, have conducted research on related algorithms, which can be categorized into: algorithms based on statistical analysis, probability density, distance, clustering, data streams, and machine learning.

[0004] The existing outlier detection algorithm based on machine learning cannot have an accurate evaluation index. The existence of outliers will cause the difference between adjacent time series data to be too large, which will affect the accuracy of the simulation model of the key parts of the heavy-duty gas turbine and the overall performance, and cannot detect outliers well. Summary of the Invention

[0005] The present invention mainly solves the technical problems that a single method in the prior art is insufficient in identification and a probability-based identification method is prone to misjudgment. A method for identifying abnormal data of a gas turbine unit based on group decision-making is proposed. The method integrates a variety of different abnormal data identification methods and uses the idea of ​​majority voting. The proposed group decision-making method selects abnormal values, which can better detect abnormal values.

[0006] The present invention provides a method for identifying abnormal data of a gas turbine unit based on group decision-making, which includes the following steps:

[0007] Step 1: preprocess the original running data;

[0008] Step 2: Use multiple detection methods to detect data and obtain the anomaly score of each detection method;

[0009] Step 3: Use voting mechanism to determine outliers;

[0010] Step 4: Output all abnormal points finally determined.

[0011] Preferably, the operating data includes but is not limited to: collected time series, output power, speed, compressor outlet temperature, adjustable guide vane angle, compressor inlet temperature, turbine 1st stage wheel rear temperature, turbine 1st stage wheel front temperature, turbine 2nd stage wheel rear temperature 1, turbine 2nd stage wheel front temperature, turbine 3rd stage wheel rear temperature, turbine 3rd stage wheel front temperature, gas turbine outlet temperature, and exhaust temperature difference.

[0012] Preferably, the detection method includes but is not limited to: standard deviation method, isolation forest method, DBSCAN clustering method, OCSVM method, LOF method, CBLOF method, KNN method, and HBOS method.

[0013] Preferably, the standard deviation method assumes that the sample follows a normal distribution or an approximately normal distribution. Based on the knowledge related to the normal distribution, the probability that the value is distributed in the interval [μ-3σ, μ+3σ] is 99.73%, and the value outside this interval is considered to be an outlier.

[0014] The isolation forest method interprets abnormal data as points that are sparsely distributed and far away from the dense group. It isolates the low-density points by randomly splitting the feature space of the data.

[0015] The DBSCAN clustering method first finds all the core points in the sample, then ignores all the core points. The remaining non-core points in the neighborhood of the core points are called boundary points, and the rest are outliers.

[0016] The OCSVM method places all samples on one side and maximizes the distance from the origin by finding a hyperplane in the data feature space. When testing data, samples that appear on the side close to the origin are considered abnormal data.

[0017] The LOF method identifies outliers by calculating the ratio of the local density of a data point to the local density of its neighbors;

[0018] The CBLOF,first clusters the data and then calculates the anomaly score based on the cluster size,to which each data point belongs and the distance from the cluster center;

[0019] The KNN method identifies outliers by calculating the distance between a data point and its k nearest neighbors;

[0020] The HBOS method calculates the anomaly score of each data point by constructing a histogram of features.

[0021] Preferably, step 3 includes the following process:

[0022] If a data point is identified as an outlier by method A, it is assigned a score of 1, meaning that method A has voted for it. If it is not identified as an outlier by method A, it is assigned a score of 0, meaning that method A considers it normal and does not vote for it. After processing the data point using all eight methods, the votes received are tallied. If the cumulative number of votes for the data point exceeds a preset threshold, the point is ultimately identified as an outlier.

[0023] The present invention provides a method for identifying abnormalities in gas turbine unit data based on group decision-making, which addresses the phenomenon that a large number of abnormal values ​​are contained in the data set collected by the sensor due to interference from harsh environments, hardware equipment failures, etc. It is crucial to reduce the data that affects the operation monitoring, performance alarms and predictive modeling of key components and effectively handle abnormal values. The present invention integrates a variety of different abnormal data identification methods, and combines them with the actual unit operating conditions. The group decision-making method proposed uses the idea of ​​majority voting to select abnormal values, which is similar to the "voting" method of machine learning. It can better detect abnormal values ​​and improve data governance effects. The group decision-making method integrates the results of multiple detectors through a voting mechanism to improve the accuracy and robustness of abnormal data identification; the present invention can dynamically adjust the voting threshold according to the distribution of the data to improve the adaptability of the model. The present invention can also introduce integrated learning methods, such as Boosting or Bagging, to further improve the robustness and accuracy of the model. For large-scale data sets, parallel computing technology can be used to accelerate the training and prediction process of the model.

[0024] The present invention uses eight different methods to identify the same set of data, adding a column at the end of the overall data set as a marker column (data type int, initial value 0). If the method determines that a point is an outlier, the marker column of that point is increased by 1. Finally, if the marker column is greater than 4, the point is determined to be an outlier. This group decision method combines conceptually different machine learning detection methods and uses majority voting to determine outliers. This method integrates different machine learning methods, can balance the weaknesses of each machine learning method, and improve the accuracy of recognition.

[0025] The data processed by the present invention meets the basic requirements of the time series model, laying the foundation for further data analysis, and subsequently providing clean and effective data for gas turbine performance operation monitoring and performance warning strategy modeling, thereby improving model accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1This is a flow chart for implementing a method for identifying abnormal gas turbine unit data based on group decision-making provided by the present invention;

[0027] Figure 2 It is the layout of gas turbine and related sensor locations;

[0028] Figure 3 It is the statistics of abnormal data identified by each method;

[0029] Figure 4 It is a line graph of the processed data set and the original result;

[0030] Figure 5 is the label value after processing the data;

[0031] Figure 6 This is the time required to run 100,000+ data levels. DETAILED DESCRIPTION

[0032] To make the technical problems solved, the technical solutions adopted, and the technical effects achieved by the present invention more clearly apparent, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention. It should also be noted that, for ease of description, the accompanying drawings only illustrate portions relevant to the present invention, rather than all of the contents.

[0033] like Figure 1 As shown, an embodiment of the present invention provides a method for identifying abnormal data of a gas turbine unit based on group decision-making, which includes the following processes:

[0034] Step 1: Preprocess the original running data.

[0035] The operating data include but are not limited to: collected time series (Time), output power (DWATT), speed (TNH), compressor outlet temperature (CTD), adjustable guide vane angle (CSGV), compressor inlet temperature (CTIM), turbine 1st stage wheel after temperature (TTWS1AO1, TTWS1AO2), turbine 1st stage wheel inlet temperature (TTWS1FI1, TTWS1FI2), turbine 2nd stage wheel after temperature 1 (TTWS2AO1, TTWS2AO2), turbine 2nd stage wheel inlet temperature (TTWS2FI1, TTWS2FI2), turbine 3rd stage wheel after temperature (TTWS3AO1, TTWS3AO2), turbine 3rd stage wheel inlet temperature (TTWS3FI1, TTWS3FI2), gas turbine outlet temperature (TTXM), exhaust temperature difference (TTXSP1, TTXSP2, TTXSP3).

[0036] Clean and standardize the original operation data to ensure data quality and consistency.

[0037] Step 2: Detect data using multiple detection methods and obtain anomaly scores for each method. The following is a systematic analysis and applicability assessment of anomalous data detection methods for heavy-duty gas turbine sensors, taking into account the operational characteristics of gas turbines (high temperature and pressure, multi-sensor coupling, high-dimensional data, and strong noise interference).

[0038] The detection method includes but is not limited to: standard deviation method, isolation forest method, DBSCAN (Density-Based Spatial Clustering of Applications with Noise) clustering method, OCSVM (One Class SVM) method, LOF (Local Outlier Factor) method, CBLOF (Cluster-based Local Outlier Factor) method, KNN (k-Nearest Neighbors) method and HBOS (Histogram-based Outlier Detection) method.

[0039] The standard deviation method (3σ method) assumes that the sample follows a normal distribution or an approximate normal distribution. Based on the knowledge related to the normal distribution, the probability that the value is distributed in the interval [μ-3σ, μ+3σ] is 99.73%, and the value outside this interval is considered to be an outlier. However, the data is required to follow a normal distribution or an approximate normal distribution. The advantage of the 3σ method is that it is simple to calculate and is suitable for rapid initial screening of single parameters (such as speed, stable temperature) under steady-state conditions. The limitation is that the variable operating condition data of heavy-duty gas turbines often show a non-normal distribution (such as the start-up and shutdown process), resulting in a high false alarm rate; it cannot handle the coupling relationship of multiple sensors (such as the dynamic correlation between exhaust temperature and load).

[0040] The Isolation Forest method understands abnormal data as points that are sparsely distributed and far away from high-density groups. By randomly dividing the feature space of the data, it is easy to isolate points with lower density (i.e., outliers). This method does not require labeled samples for training, but the features need to be continuous. The advantage of the Isolation Forest method is that it does not require labeled data and is suitable for scenarios where fault samples are scarce; it can efficiently process high-dimensional continuous parameters (such as gas turbine power and multi-channel temperature data). The limitation is that it is sensitive to feature correlation and may misjudge strongly coupled parameters (such as compressor inlet and outlet pressures).

[0041] The DBSCAN (Density-Based Spatial Clustering of Applications with Noise) clustering method is a density-based spatial clustering algorithm. It first finds all the core points in the sample, then ignores all the core points. The remaining non-core points in the neighborhood of the core points are called boundary points, and the rest are outliers. It is particularly suitable for processing data sets with irregular shapes and uneven density, but it requires high computer computing power. The DBSCAN clustering method can identify clusters of arbitrary shapes and is applicable to nonlinear data (such as vibration signal spectra and acoustic emission waveforms during gas turbine operation); and it can automatically eliminate instantaneous interference noise (such as sensor jumps caused by electromagnetic interference). The limitations are high computational complexity and poor real-time performance (large-scale data requires distributed computing).

[0042] The OCSVM (One Class SVM) method finds a hyperplane in the data feature space, placing all samples on one side and maximizing its distance from the origin. When testing data, samples that appear close to the origin are considered abnormal data. However, this method requires high data dimensions or no assumptions about the distribution of the relevant data. The OCSVM method only requires normal data training and is suitable for rare fault scenarios. Kernel functions (such as RBF) can model nonlinear relationships (such as the exhaust temperature distribution of gas turbines). However, its limitation is the high computational overhead of high-dimensional data (requiring GPU acceleration).

[0043] The LOF (Local Outlier Factor) method identifies outliers by calculating the ratio of the local density of a data point to the local density of its neighbors. It can identify local outliers and is suitable for data sets with uneven density. It does not require many requirements for data distribution, but the computational complexity is relatively high. The advantages of the LOF method are that it can adapt to data with uneven density (such as temperature gradients under partial load) and can detect local sensor failures (such as single-point failures in turbine exhaust arrays). The limitation is that the computational complexity is high (O(n 2 )), requires approximate algorithm (such as kd tree) optimization.

[0044] The CBLOF (Cluster-based Local Outlier Factor) method first clusters the data and then calculates the anomaly score based on the size of the cluster to which each data point belongs and the distance from the cluster center. Combining clustering and distance metrics, it can effectively identify outliers. However, it is sensitive to the choice of clustering algorithm. The CBLOF method combines clustering results to effectively identify group anomalies (such as synchronous drift of multiple sensors); it is suitable for working condition clustering data (parameter groups under different loads). The limitation is that it depends on the quality of clustering, and high-dimensional data must be reduced in dimensionality first (such as PCA); it is sensitive to the distance from the cluster center and has poor effect on non-spherical clusters.

[0045] The KNN (k-Nearest Neighbors) method identifies outliers by calculating the distance between a data point and its k nearest neighbors. It is simple to implement and applicable to a wide range of data types. However, it is sensitive to the choice of k value. The KNN method is simple and intuitive, suitable for offline validation, and can handle mixed data types (such as temperature, pressure, and flow in compressors, combustion chambers, and turbine modules). However, its limitations are that computational efficiency decreases sharply with data volume, and the selection of a k value requires cross-validation (too large a value may result in missed results, while too small a value may result in false positives).

[0046] The HBOS (Histogram-based Outlier Detection) method calculates an anomaly score for each data point by constructing a feature histogram. It boasts fast computation and is suitable for high-dimensional data. However, it is sensitive to the histogram construction method. The HBOS method is extremely fast (O(n)), making it suitable for online monitoring of high-dimensional data. It requires no training and can adapt to real-time updates of data streams. However, its limitation lies in its assumption of feature independence and its neglect of sensor coupling (e.g., the correlation between fuel flow and power).

[0047] The standard deviation method is suitable for continuous parameters that conform to a normal distribution, such as temperature and pressure. It is simple and efficient, but has poor performance on non-normal data. Isolation forest is suitable for high-dimensional sensor data, has fast computation, and requires no distributional assumptions, making it suitable for real-time monitoring. DBSCAN is suitable for spatially clustered operating data and is adept at detecting anomalies in irregular clusters. OCSVM is suitable for scenarios where only normal operating data is available, such as fault detection. LOF and CBLOF are suitable for capturing local density anomalies and are applicable to complex operating modes. KNN is sensitive to distance and is suitable for small-scale, precise detection. HBOS is suitable for rapid analysis of high-dimensional, independent features, such as multi-parameter real-time streams. Gas turbine sensor anomaly detection requires comprehensive consideration of the operating environment (high temperature, high pressure, strong vibration), data characteristics (high dimensionality, multi-source coupling, non-steady state), and engineering constraints (real-time nature, scarce fault samples). A single method cannot fully cover complex fault modes, and a hierarchical or fusion strategy must be constructed based on the method's characteristics. A voting mechanism for multi-source information fusion can effectively improve robustness.

[0048] Step 3: Use voting mechanism to determine outliers.

[0049] Each data point is processed using one of eight methods. If a data point is identified as an outlier by method A, it is assigned a score of 1, meaning that method A has voted for it. If the data point is not identified as an outlier by method A, it is assigned a score of 0, meaning that method A considers the data point normal and does not vote for it. After processing the data point using all eight methods, the votes received are tallied. If the cumulative number of votes for the data point exceeds a preset threshold (e.g., 4 votes), the point is ultimately identified as an outlier.

[0050] Step 4: Output all abnormal points finally determined.

[0051] This step outputs all the abnormal points finally determined, which can be deleted or otherwise processed.

[0052] The present invention uses eight different methods to identify the same set of data, adding a column at the end of the overall data set as a marker column (data type int, initial value 0). If the method determines that a point is an outlier, the marker column of that point is increased by 1. Finally, if the marker column is greater than 4, the point is determined to be an outlier. This group decision method combines conceptually different machine learning detection methods and uses majority voting to determine outliers. This method integrates different machine learning methods, can balance the weaknesses of each machine learning method, and improve the accuracy of recognition.

[0053] In order to make the objects and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the following examples are used to illustrate and explain the present invention, but are not used to limit the scope of the present invention. In order to verify the effect of the group decision-making method for identifying outliers, a complete sample data set of 500 items of a certain model of 300MW gas turbine running in 2013 was used. An interval of 5 minutes was used to write 8 abnormal data detection methods through Python code, and the identified data was marked. Then, the group decision-making method was used to perform the final screening of the abnormal data, thereby improving the accuracy of identifying abnormal data.

[0054] The experimental environment for this paper is a Dell PC with a Core i7 processor, 32GB of memory, and Windows 10 system. The programming language is Python. A simulation comparison test was conducted on 500 sample data from a certain model 300MW gas turbine in 2013. The data set contains 500 data items and 22 variables. Figure 2 The basic structure of a single-shaft gas turbine and the general layout of typical sensor locations are shown. The basic structure of a single-shaft gas turbine includes Figure 2 The air intake, compressor, combustion chamber, turbine and exhaust duct components shown are Figure 2 The meanings of the names of the 22 sensors are shown in Table 1.

[0055] Table 1 Analysis and meaning of some variables

[0056]

[0057]

[0058] Figure 5 It shows the distribution of votes obtained by some data points after being processed by the above method and the final total number of votes obtained. Figure 3It is the corresponding line chart, which can more intuitively observe the overall distribution of votes of data points. Figure 4 Taking the compressor outlet temperature (CTD) as an example, the data distribution before and after the above method is compared. The running time for the data level of 500 samples is 7.2s. The running time for the data level of 100,000+ is 163.16s. Figure 6 shown.

[0059] 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, those skilled in the art should understand that modifications to the technical solutions described in the above embodiments, or equivalent replacement of some or all of the technical features therein, do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying abnormal data of a gas turbine unit based on group decision making, characterized in that: The following processes are included: Step 1: preprocess the original running data; Step 2: Use multiple detection methods to detect data and obtain the anomaly score of each detection method; Step 3: Use voting mechanism to determine outliers; Step 4: Output all abnormal points finally determined.

2. The method for identifying abnormal data of a gas turbine unit based on group decision-making according to claim 1 is characterized in that: The operating data includes but is not limited to: collected time series, output power, speed, compressor outlet temperature, adjustable guide vane angle, compressor inlet temperature, turbine 1st stage wheel rear temperature, turbine 1st stage wheel front temperature, turbine 2nd stage wheel rear temperature 1, turbine 2nd stage wheel front temperature, turbine 3rd stage wheel rear temperature, turbine 3rd stage wheel front temperature, gas turbine outlet temperature, and exhaust temperature difference.

3. The method for identifying abnormal data of a gas turbine unit based on group decision-making according to claim 1, characterized in that: The detection method includes, but is not limited to: standard deviation method, isolation forest method, DBSCAN clustering method, OCSVM method, LOF method, CBLOF method, KNN method, and HBOS method.

4. The method for identifying abnormal data of a gas turbine unit based on group decision-making according to claim 3 is characterized in that: The standard deviation method assumes that the sample follows a normal distribution or an approximate normal distribution. Based on the knowledge related to the normal distribution, the probability that the value is distributed in the interval [μ-3σ, μ+3σ] is 99.73%, and the value outside this interval is considered to be an outlier. The isolation forest method interprets abnormal data as points that are sparsely distributed and far away from the dense group. It isolates the low-density points by randomly splitting the feature space of the data. The DBSCAN clustering method first finds all the core points in the sample, then ignores all the core points. The remaining non-core points in the neighborhood of the core points are called boundary points, and the rest are outliers. The OCSVM method places all samples on one side and maximizes the distance from the origin by finding a hyperplane in the data feature space. When testing data, samples that appear on the side close to the origin are considered abnormal data. The LOF method identifies outliers by calculating the ratio of the local density of a data point to the local density of its neighbors; The CBLOF,first clusters the data and then calculates the anomaly score based on the cluster size,to which each data point belongs and the distance from the cluster center; The KNN method identifies outliers by calculating the distance between a data point and its k nearest neighbors; The HBOS method calculates the anomaly score of each data point by constructing a histogram of features.

5. The method for identifying abnormal data of a gas turbine unit based on group decision-making according to claim 4 is characterized in that: Step 3 The following processes are included: If a data point is identified as an outlier by method A, it is assigned a score of 1, meaning that method A has voted for it. If it is not identified as an outlier by method A, it is assigned a score of 0, meaning that method A considers it normal and does not vote for it. After processing the data point using all eight methods, the votes received are tallied. If the cumulative number of votes for the data point exceeds a preset threshold, the point is ultimately identified as an outlier.

Images