A method for identifying combustion abnormality of a gas turbine based on pressure pulsation

Abstract

This invention provides a method for identifying combustion anomalies in gas turbines based on pressure pulsation, belonging to the field of gas turbine operation monitoring and fault diagnosis technology. It significantly alleviates or solves the problems of existing technologies that rely primarily on manual experience, fixed thresholds, or single feature quantities for judgment, resulting in insufficient combustion state characterization and low accuracy and reliability. This invention collects pressure pulsation monitoring data during combustion chamber operation, preprocesses it, and extracts time-series window features. This data is then correlated with combustion state labels for corresponding time periods in historical operating data to construct a sample set. A combustion anomaly identification model is trained, and the probability of combustion anomaly at the current moment and the alarm result are output. This method is beneficial for more comprehensively characterizing the combustion state change process, improving the accuracy, reliability, and online application effectiveness of combustion anomaly identification.

Description

Technical Field

[0001] This invention belongs to the field of gas turbine operation monitoring and fault diagnosis technology, specifically relating to a method for identifying abnormal combustion in gas turbines based on pressure pulsation. Background Technology

[0002] Gas turbines, as highly efficient and clean power generation equipment, are widely used in power generation, industrial drives, and integrated energy utilization. With the increasing demands for high efficiency, low emissions, and high reliability in energy systems, gas turbines are increasingly being used in large-capacity generator sets, peak-shaving units, and combined cycle systems. The combustion chamber, as a key component in gas turbines for mixing and burning fuel and air to release heat energy, directly affects the output power, combustion efficiency, emission levels, and overall operational safety of the gas turbine. In actual operation, the combustion chamber is constantly exposed to a complex environment of high temperature, high pressure, and strong disturbances. The combustion process is easily affected by factors such as fuel fluctuations, air distribution changes, load variations, and changes in component conditions, leading to problems such as combustion instability and abnormally increased pressure pulsations. If combustion anomalies are not identified and addressed in a timely manner, they can cause minor issues such as unit operation fluctuations, decreased combustion efficiency, and increased pollutant emissions; in severe cases, they may induce intensified combustion oscillations, abnormally high component heat loads, and even shutdown accidents. Therefore, effective monitoring and timely identification of the combustion status of gas turbines are of great significance.

[0003] In the field of gas turbine combustion status monitoring, pressure pulsation signals are one of the important pieces of information reflecting the dynamic changes in the combustion process. By placing pulsating pressure sensors at corresponding measuring points in the combustion chamber, pressure fluctuations during combustion can be acquired in real time, and the normality of the combustion state can be analyzed accordingly. Currently, monitoring methods for combustion anomalies typically include manual judgment based on experience, judgment methods based on fixed alarm thresholds, and anomaly identification methods relying on single statistical features or simple rules. In recent years, with the development of data analysis technology, a technical approach has also emerged that utilizes historical operating data to build models to assist in the identification of combustion anomalies.

[0004] However, existing technologies still have significant shortcomings. First, current methods often focus on judging instantaneous pressure pulsation amplitude or single alarm thresholds, failing to fully reflect the continuous evolution characteristics of combustion states over time, resulting in insufficient ability to distinguish between short-term disturbances and persistent anomalies. Second, the actual operating conditions of gas turbines are complex, with pressure pulsation signals varying under different loads, turbine models, and operating stages. Relying solely on empirical thresholds or single feature quantities easily leads to false alarms and missed alarms. Third, while some existing methods introduce data modeling approaches, their utilization of real historical operating data from in-service units is insufficient, failing to construct a targeted identification link based on the temporal characteristics of pressure pulsations, resulting in limited applicability and generalization ability of the model engineering. Furthermore, in actual operation, the number of combustion anomaly samples is usually far less than the number of normal samples. If specific processing is not carried out for sample imbalance, temporal sequence relationships, and the requirements of online rolling identification, the accuracy and reliability of anomaly identification results will also be affected.

[0005] To address this, a method for identifying combustion anomalies in gas turbines based on pressure pulsation is proposed. Summary of the Invention

[0006] The present invention aims to solve at least one of the technical problems existing in the prior art, and provides a method for identifying abnormal combustion in gas turbines based on pressure pulsation.

[0007] This invention provides a method for identifying combustion anomalies in gas turbines based on pressure pulsation, comprising the following steps: S1: Collect combustion pressure pulsation monitoring data of the gas turbine combustor during operation; S2: Preprocess the combustion pressure pulsation monitoring data to obtain a pressure pulsation time sequence data. S3: Based on a time window of a preset length, feature extraction is performed on the pressure pulsation time-series data sequence to obtain feature parameters characterizing changes in combustion state; S4: Associate the feature parameters with the combustion status labels of the corresponding time periods in the historical operation data to construct a combustion anomaly identification sample set; S5: Train a combustion anomaly identification model based on the aforementioned combustion anomaly identification sample set; S6: Input the feature parameters corresponding to the time to be identified into the combustion anomaly identification model, and output the combustion anomaly probability at the current time; S7: Based on the comparison result between the combustion abnormality probability and the preset judgment threshold, determine whether the current time is in a combustion abnormality state, and output an alarm signal when it is determined that the current time is in the combustion abnormality state.

[0008] Furthermore, in step S1, a pulsating pressure sensor installed at a measuring point on the outer wall of the gas turbine combustion chamber is used to collect the combustion pressure pulsation monitoring data.

[0009] Specifically, the sampling frequency of the pulsating pressure sensor is set to 10kHz to 50kHz according to the requirements for collecting pressure pulsation signals from the gas turbine combustor; the range of the pulsating pressure sensor is selected according to the historical pressure pulsation amplitude range of the target gas turbine combustor, so that the collected pressure pulsation signal is within the effective measurement range.

[0010] Specifically, in step S2, the preprocessing includes one or more of the following: outlier removal, missing value processing, time-order sorting, and data segmentation.

[0011] Preferably, in step S3, the characteristic parameter includes at least one of the following: Pressure pulsation change, which is the difference between the pressure pulsation value at the current sampling point and the pressure pulsation value at the previous sampling point; The pressure pulsation rate, which is the ratio of the pressure pulsation change to the time interval between adjacent sampling points; and The moving average of pressure pulsation is a local mean obtained by averaging the pressure pulsation values ​​of a preset number of consecutive sampling points with the current sampling point as the endpoint.

[0012] Specifically, in step S3, the feature extraction further includes extracting multi-scale moving average features, which are multiple moving averages calculated based on different window lengths; the different window lengths include multiple continuous sampling point windows of different lengths.

[0013] Furthermore, in step S4, the samples are labeled according to the original combustion abnormality alarm records during the unit's historical operation. Samples that trigger combustion abnormality alarms are marked as combustion abnormality states, and samples that do not trigger combustion abnormality alarms are marked as normal combustion states.

[0014] Furthermore, between step S4 and step S5, the method further includes: dividing the combustion anomaly identification sample set into a training set and a test set according to chronological order, in order to avoid information from future moments being leaked to historical moments.

[0015] Furthermore, in step S5, when the number of abnormal combustion samples in the training set is less than the number of normal samples, the abnormal combustion samples in the training set are oversampled, and when training the abnormal combustion identification model, the abnormal combustion samples are assigned a higher sample weight than the normal samples, so as to improve the detection capability of the abnormal combustion identification model for abnormal combustion states.

[0016] Specifically, step S6 is performed using a real-time scrolling recognition method, which includes: initializing a fixed-length sliding window and caching the pressure pulsation data of the most recent consecutive sampling points; when a new pressure pulsation sampling value is received, the sampling value is written into the sliding window and the earliest sampling value entering the window is deleted; based on the updated sliding window, the feature parameters corresponding to the current moment are extracted, and the feature parameters are input into a pre-trained combustion anomaly recognition model to output the combustion anomaly probability at the current moment.

[0017] The beneficial effects of this invention are as follows: This invention collects pressure pulsation monitoring data during the operation of a gas turbine combustor, extracts temporal feature parameters characterizing changes in combustion state, constructs a combustion anomaly identification sample set by combining combustion state labels from historical operating data, and trains a combustion anomaly identification model to output the probability of combustion anomaly at the current moment and make an alarm determination. Compared with existing technologies, this invention can more comprehensively characterize the combustion state change process, improving the accuracy and reliability of combustion anomaly identification. Furthermore, by dividing the training and test sets chronologically and performing oversampling and weight enhancement processing on anomaly samples, it helps improve the model's ability to detect anomalies. In addition, this invention can continuously output the probability of combustion anomalies in real-time and provide timely alarms, demonstrating good engineering applicability and online application value. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the steps of a gas turbine combustion anomaly identification method based on pressure pulsation, according to a specific embodiment of the present invention. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] like Figure 1 As shown in the figure, a method for identifying combustion anomalies in a gas turbine based on pressure pulsation, provided by a specific embodiment of the present invention, includes the following steps: S1: Collect combustion pressure pulsation monitoring data of the gas turbine combustor during operation; S2: Preprocess the combustion pressure pulsation monitoring data to obtain a pressure pulsation time series data sequence; S3: Based on a time window of preset length, feature extraction is performed on the pressure pulsation time series data sequence to obtain feature parameters characterizing changes in combustion state; S4: Associate the feature parameters with the combustion status labels of the corresponding time periods in the historical operation data to construct a combustion anomaly identification sample set; S5: Training a combustion anomaly identification model based on a combustion anomaly identification sample set; S6: Input the feature parameters corresponding to the time to be identified into the combustion anomaly identification model and output the combustion anomaly probability at the current time; S7: Based on the comparison between the probability of abnormal combustion and the preset judgment threshold, determine whether the current time is in an abnormal combustion state, and output an alarm signal when it is determined that the current time is in an abnormal combustion state.

[0021] In one implementation, the above steps S1 to S7 constitute a complete combustion anomaly identification process and are executed sequentially in chronological order to identify the combustion state of the gas turbine.

[0022] Preferably, the present invention includes a model building stage and an online identification stage; wherein, in the model building stage, steps S1 to S5 are executed to complete the establishment of a combustion anomaly identification model, and in the online identification stage, steps S1, S2, S3, S6 and S7 are executed and cyclically at preset time intervals to achieve continuous online monitoring of the combustion state.

[0023] Furthermore, the preset judgment threshold in step S7 can be determined based on one or more of the historical operating data statistics of the target gas turbine, the original combustion abnormality alarm records, and the preset combustion abnormality alarm thresholds in the unit control system, and set after verification through a test set.

[0024] Based on the above basic implementation method, in step S1, a pulsating pressure sensor installed at a measuring point on the outer wall of the gas turbine combustor is used to collect the combustion pressure pulsation monitoring data. The pulsating pressure sensor is fixed at the corresponding measuring point using a special mounting fixture. The sampling frequency of the pulsating pressure sensor is set to 10kHz to 50kHz according to the requirements for collecting the pressure pulsation signal of the gas turbine combustor. The range of the pulsating pressure sensor is selected according to the historical pressure pulsation amplitude range of the target gas turbine combustor, so that the collected pressure pulsation signal is within the effective measurement range. Being within the effective measurement range means that the collected pressure pulsation signal is within the rated range of the pulsating pressure sensor, and the corresponding signal amplitude is within the interval that can be stably identified by the pulsating pressure sensor and output as an effective measurement result. Preferably, the effective measurement range is the linear response range of the pulsating pressure sensor or a reliable measurement range confirmed by calibration. By keeping the pressure pulsation signal within the effective measurement range, it is beneficial to reduce the impact of sensor saturation, distortion, or decreased measurement accuracy on subsequent feature extraction and combustion anomaly identification results.

[0025] Furthermore, the pressure pulsation data collected by the pulsation pressure sensor is transmitted to the data processing unit or monitoring terminal in the order of timestamps for subsequent data storage, preprocessing, and identification analysis.

[0026] Furthermore, during normal operation of the gas turbine, the pulsating pressure sensor continuously collects pressure pulsation signals to form continuous pressure pulsation time-series data, providing a data basis for subsequent rolling identification.

[0027] In one specific implementation, in step S2, the preprocessing includes one or more of the following: outlier removal, missing value handling, time-order sorting, and data segmentation.

[0028] In this embodiment, outlier removal is used to eliminate abnormal sampling points that exceed the range of the pulsating pressure sensor or whose change relative to adjacent sampling points exceeds a preset anomaly threshold, thereby reducing the impact of abnormal noise on subsequent feature extraction and model training. In this invention, the anomaly threshold characterizes the upper limit of the allowable pressure pulsation change amplitude between adjacent sampling points. When the change amplitude of a sampling point relative to its adjacent sampling points exceeds the anomaly threshold, the sampling point is determined to be an abnormal sampling point. The anomaly threshold can be determined based on the statistical results of the historical operating data of the target gas turbine. Preferably, it can be set based on the distribution range, mean, standard deviation, quantile interval, or a combination thereof of the change amplitude of adjacent sampling points under normal operating conditions; it can also be corrected in conjunction with sensor measurement accuracy, sampling frequency, and combustion chamber operating conditions. By setting the anomaly threshold, it is beneficial to eliminate abnormal sampling points caused by transient interference, acquisition noise, or sensor malfunction, while retaining effective pressure pulsation data that reflects the actual combustion state changes.

[0029] Furthermore, missing value processing can be implemented by interpolating adjacent sampling points, filling in the mean, or removing missing data fragments; time order sorting is used to ensure that the pressure pulsation time series data are arranged in the order of collection; data segmentation is used to divide the continuous pressure pulsation time series data according to a preset time length or a preset number of sampling points, so as to carry out subsequent time series window construction and feature extraction.

[0030] In another specific embodiment, in step S3, the feature parameters include at least one of the following: Pressure pulsation change, which is the difference between the pressure pulsation value at the current sampling point and the pressure pulsation value at the previous sampling point; pressure pulsation change rate, which is the ratio of the pressure pulsation change to the time interval between adjacent sampling points; and pressure pulsation moving average, which is the local mean obtained by averaging the pressure pulsation values ​​of a preset number of consecutive sampling points with the current sampling point as the endpoint.

[0031] In this embodiment, the above-mentioned feature parameters together constitute the feature vector of the combustion anomaly identification model to be input, which is used to characterize the change law of combustion state in the current moment and the preceding time period; if the feature parameters are normalized or standardized, the corresponding parameters are determined based on the training set samples and applied to the test set samples and the feature parameters input in the online identification stage.

[0032] Specifically, the pressure pulsation change is used to reflect the instantaneous fluctuation between adjacent sampling times, the pressure pulsation change rate is used to reflect the speed of pressure pulsation changes, and the pressure pulsation moving average is used to reflect the average fluctuation level within a local time range.

[0033] Furthermore, to reduce the impact of differences in dimensions and numerical ranges among different feature parameters on the model training results, the feature parameters can be normalized or standardized after feature extraction. The normalized or standardized parameters are determined based on the feature parameters of the training set and applied to the feature parameters input in the test set and the online recognition stage, so as to avoid information from future moments being leaked to historical moments.

[0034] In another specific embodiment, in step S3, feature extraction further includes extracting multi-scale moving average features, which are multiple moving averages calculated based on different window lengths; different window lengths include multiple continuous sampling point windows of different lengths.

[0035] Furthermore, multi-scale moving average features are used to simultaneously characterize short-term rapid fluctuations and continuous changes over longer time scales, thereby improving the ability to depict the evolution of combustion anomalies.

[0036] Specifically, different window lengths can be preset according to the gas turbine operating conditions, sampling frequency and duration of combustion anomalies, and at least two consecutive sampling point windows of different lengths are used to calculate the corresponding moving averages to obtain multi-scale time series features.

[0037] In another specific embodiment, in step S4, the samples are labeled according to the original combustion abnormality alarm records during the historical operation of the unit. The samples that trigger the combustion abnormality alarm are marked as combustion abnormal state, and the samples that do not trigger the combustion abnormality alarm are marked as normal combustion state.

[0038] Furthermore, the original combustion anomaly alarm records can be obtained from the unit control system, monitoring system, or operation log, and are used to reflect the actual combustion status during the corresponding time period.

[0039] Specifically, after aligning the feature parameters with the original combustion anomaly alarm records within the corresponding time period, the feature parameter samples corresponding to the time period that triggered the combustion anomaly alarm are marked as combustion anomaly states, and the feature parameter samples corresponding to the time period that did not trigger the combustion anomaly alarm are marked as normal combustion states, thereby forming labeled samples that can be used for model training.

[0040] In another specific embodiment, between step S4 and step S5, the method further includes: dividing the combustion anomaly identification sample set into a training set and a test set according to the chronological order, so as to avoid information from future moments being leaked to historical moments.

[0041] Furthermore, when dividing the data according to chronological order, samples corresponding to earlier time periods are divided into the training set, and samples corresponding to later time periods are divided into the test set, to ensure that the model training and model validation processes conform to the temporal causal relationship.

[0042] Furthermore, the training set is used to train the combustion anomaly recognition model, and the test set is used to verify the performance of the combustion anomaly recognition model, thereby improving the reliability of the model in real-world operating scenarios.

[0043] In one specific implementation, in step S5, when the number of abnormal combustion samples in the training set is less than the number of normal samples, the abnormal combustion samples in the training set are oversampled, and when training the abnormal combustion identification model, the abnormal combustion samples are assigned a higher sample weight than the normal samples, so as to improve the detection capability of the abnormal combustion identification model for abnormal combustion states.

[0044] Specifically, oversampling is only applied to combustion anomaly samples in the training set, and not to samples in the test set, in order to avoid affecting the objectivity of the model validation results.

[0045] Furthermore, after the model training is completed, the performance of the combustion anomaly identification model can be evaluated using recall rate, F1 score, or other evaluation metrics used to characterize the detection effect of minority class samples, so as to verify the model's ability to identify combustion anomalies.

[0046] In one specific implementation, step S6 is performed using a real-time rolling recognition method, which includes: initializing a fixed-length sliding window and caching pressure pulsation data of the most recent consecutive sampling points; when a new pressure pulsation sampling value is received, writing the sampling value into the sliding window and deleting the earliest sampling value that entered the sliding window; extracting the feature parameters corresponding to the current moment based on the updated sliding window, and inputting the feature parameters into a pre-trained combustion anomaly recognition model to output the combustion anomaly probability at the current moment.

[0047] In this embodiment, the real-time rolling identification method enables the calculation of the probability of combustion anomaly based on the latest time window after each new pressure pulsation sampling value is received, thereby achieving continuous dynamic identification of the combustion state.

[0048] Specifically, when the probability of abnormal combustion at the current moment is greater than the preset judgment threshold, an alarm signal can be sent to the monitoring terminal, control system or operation and maintenance platform to prompt the operators that the current combustion status is abnormal and take corresponding measures.

[0049] In a specific embodiment, taking an in-service gas turbine as an example, the gas turbine combustion anomaly identification method based on pressure pulsation provided by the present invention will be described. Pressure pulsation measuring points are pre-set on the outer wall of the gas turbine combustion chamber. Pulsation pressure sensors are installed at these measuring points, and the sensors are fixed to the corresponding measuring points on the outer wall of the combustion chamber using a special mounting fixture. These sensors are used to continuously collect combustion pressure pulsation monitoring data during combustion chamber operation. In this embodiment, the sampling frequency of the pulsation pressure sensor is set to 20kHz, which falls within the aforementioned sampling frequency range of 10kHz to 50kHz. Simultaneously, the range of the pulsation pressure sensor is selected based on the historical pressure pulsation amplitude range of the gas turbine combustion chamber, ensuring that the collected pressure pulsation signal is within the effective measurement range. Historical combustion pressure pulsation data of the in-service gas turbine for four consecutive days is collected as raw sample data, resulting in a total of 73,380 sampling points. After the raw sample data is transmitted to the data processing unit, it is first preprocessed. The preprocessing process includes outlier removal, missing value handling, time-sequence sorting, and data segmentation. Specifically, sampling points that exceed the range of the pulsating pressure sensor or whose variation relative to adjacent sampling points exceeds a preset anomaly threshold are identified as outliers and removed. For a small number of missing points, imputation using adjacent sampling points is used to complete the data. Then, all sampled data are reordered according to timestamps to form a continuous pressure pulsation time-series data sequence. After preprocessing, sample construction and feature extraction are performed on the pressure pulsation time-series data sequence based on a sliding time window. In this embodiment, a sliding time window is formed by 10 consecutive sampling points, with the current sampling point as the window endpoint. The following feature parameters are extracted within each sliding time window: (1) Pressure pulsation change, which is the difference between the pressure pulsation value at the current sampling point and the pressure pulsation value at the previous sampling point; (2) Pressure pulsation rate, which is the ratio of the pressure pulsation change to the time interval between adjacent samplings; (3) Pressure pulsation moving average, which is the local mean obtained by averaging the pressure pulsation values ​​of the current sampling point and its four consecutive sampling points with the current sampling point as the endpoint; (4) Multi-scale moving average characteristics, namely, two moving averages calculated based on 5 consecutive sampling points and 10 consecutive sampling points respectively, are used to simultaneously characterize short-term fluctuations and changes over a longer time scale.

[0050] This forms a set of feature parameters corresponding to each time point, which serves as the input features for subsequent model training and online recognition.

[0051] In this embodiment, during the sample labeling process, samples for each time period are labeled based on the original combustion anomaly alarm records from the unit's historical operation. Specifically, the feature parameter samples corresponding to the time periods that triggered the original combustion anomaly alarm are marked as combustion anomaly states, while the feature parameter samples corresponding to the time periods that did not trigger the original combustion anomaly alarm are marked as normal combustion states. To facilitate model training, in this embodiment, normal combustion states are recorded as 0, and combustion anomaly states are recorded as 1. After labeling is completed, a combustion anomaly identification sample set is obtained. This sample set is then divided into a training set and a test set according to chronological order, with the first 80% of the samples used as the training set and the last 20% used as the test set, to avoid information from future moments being leaked to historical moments. Considering that the number of combustion anomaly samples in the training set is significantly less than the number of normal samples, oversampling is performed on the combustion anomaly samples in the training set, and the combustion anomaly samples are assigned higher sample weights than normal samples when training the model, in order to improve the model's ability to detect combustion anomaly states. Subsequently, a supervised learning classification model is used to train the training set. In this embodiment, the supervised learning classification model is a support vector classification model, which learns the mapping relationship between the feature parameters of the training set and the corresponding labels, thereby establishing a combustion anomaly identification model. After the model training is completed, the test set is input into the combustion anomaly identification model for verification. Based on the probability output result of the combustion anomaly identification model or the output result after probability mapping processing, the combustion anomaly probability corresponding to each test sample is obtained. Then, a preset judgment threshold is used to determine whether the test sample belongs to the combustion anomaly state. In this embodiment, the preset judgment threshold is set to 0.3. When the combustion anomaly probability of a certain test sample is greater than 0.3, the test sample is determined to be in the combustion anomaly state at the corresponding time. When the combustion anomaly probability is less than or equal to 0.3, the test sample is determined to be in the normal combustion state at the corresponding time.

[0052] Specifically, in the online application phase, step S6 is executed by calling the pre-trained combustion anomaly identification model. Specifically, firstly, a fixed-length sliding window is initialized and the pressure pulsation data of the 10 most recent sampling points are cached; each time a new pressure pulsation sampling value is received, the sampling value is written into the sliding window, and the earliest sampled value entering the sliding window is deleted; then, feature parameters corresponding to the current moment are extracted based on the updated sliding window, and the feature parameters are input into the pre-trained combustion anomaly identification model to output the combustion anomaly probability at the current moment; finally, the combustion anomaly probability at the current moment is compared with a preset judgment threshold. When the combustion anomaly probability is greater than the preset judgment threshold, an alarm signal is output to alert the operator that the current combustion state is abnormal.

[0053] In summary, this embodiment has at least the following technical effects: This invention identifies combustion anomalies based on combustion pressure pulsation monitoring data collected during the operation of in-service gas turbines. The data source used directly corresponds to the actual operating conditions, which can reduce the limitations of relying solely on manual experience or a single alarm threshold, thereby improving the applicability of the combustion anomaly identification method in actual units. By extracting pressure pulsation change, pressure pulsation rate of change, pressure pulsation moving average, and multi-scale moving average features from pressure pulsation time series data, it is possible to simultaneously reflect the instantaneous fluctuation characteristics and the change characteristics at different time scales of the pressure pulsation signal, thereby more comprehensively characterizing the combustion state change process and improving the accuracy of identifying abnormal combustion states. This invention associates feature parameters with combustion state labels for corresponding time periods in historical operation data to construct a sample set, and divides the training set and test set according to the chronological order. This helps to avoid information from future moments being leaked to historical moments, making the model training and verification process more consistent with the time causal relationship in actual operation, thereby improving the reliability of the model in real engineering scenarios. To address the issues of a low proportion of combustion anomaly samples and significant class imbalance in actual operation, this invention performs oversampling on combustion anomaly samples in the training set and assigns higher weights to combustion anomaly samples during model training. This helps to improve the combustion anomaly identification model's attention to minority anomaly samples, thereby enhancing its ability to detect combustion anomalies and reducing the risk of missed detections. The system employs a real-time scrolling identification method, continuously receiving new pressure pulsation sampling values, updating feature parameters, and outputting the current combustion anomaly probability through a sliding window. When the combustion anomaly probability exceeds a preset judgment threshold, an alarm signal is output. Therefore, it can achieve continuous online identification of the combustion status of the gas turbine and issue timely alarms when an anomaly occurs, facilitating operators to take early action.

[0054] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for identifying combustion anomalies in gas turbines based on pressure pulsation, characterized in that, Includes the following steps: S1: Collect combustion pressure pulsation monitoring data of the gas turbine combustor during operation; S2: Preprocess the combustion pressure pulsation monitoring data to obtain a pressure pulsation time sequence data. S3: Based on a time window of a preset length, perform feature extraction on the pressure pulsation time-series data sequence to obtain feature parameters characterizing changes in combustion state; S4: Associate the feature parameters with the combustion status labels of the corresponding time periods in the historical operation data to construct a combustion anomaly identification sample set; S5: Train a combustion anomaly identification model based on the aforementioned combustion anomaly identification sample set; S6: Input the feature parameters corresponding to the time to be identified into the combustion anomaly identification model, and output the combustion anomaly probability at the current time; S7: Based on the comparison result between the combustion abnormality probability and the preset judgment threshold, determine whether the current time is in a combustion abnormality state, and output an alarm signal when it is determined that the current time is in the combustion abnormality state.

2. The gas turbine combustion anomaly identification method based on pressure pulsation according to claim 1, characterized in that, In step S1, a pulsating pressure sensor installed at a measuring point on the outer wall of the gas turbine combustion chamber is used to collect the combustion pressure pulsation monitoring data.

3. The gas turbine combustion anomaly identification method based on pressure pulsation according to claim 2, characterized in that, The sampling frequency of the pulsating pressure sensor is set to 10kHz to 50kHz according to the requirements for collecting pressure pulsation signals from the gas turbine combustor; the range of the pulsating pressure sensor is selected according to the historical pressure pulsation amplitude range of the target gas turbine combustor, so that the collected pressure pulsation signal is within the effective measurement range.

4. The gas turbine combustion anomaly identification method based on pressure pulsation according to claim 1, characterized in that, In step S2, the preprocessing includes one or more of the following: outlier removal, missing value handling, time-order sorting, and data segmentation.

5. The gas turbine combustion anomaly identification method based on pressure pulsation according to claim 1, characterized in that, In step S3, the feature parameters include at least one of the following: Pressure pulsation change, which is the difference between the pressure pulsation value at the current sampling point and the pressure pulsation value at the previous sampling point; Pressure pulsation rate of change, wherein the pressure pulsation rate of change is the ratio of the pressure pulsation change to the time interval between adjacent samplings; as well as The moving average of pressure pulsation is a local mean obtained by averaging the pressure pulsation values ​​of a preset number of consecutive sampling points with the current sampling point as the endpoint.

6. The gas turbine combustion anomaly identification method based on pressure pulsation according to claim 1, characterized in that, In step S3, the feature extraction further includes extracting multi-scale moving average features, which are multiple moving averages calculated based on different window lengths; the different window lengths include multiple continuous sampling point windows of different lengths.

7. The gas turbine combustion anomaly identification method based on pressure pulsation according to claim 1, characterized in that, In step S4, the samples are labeled according to the original combustion abnormality alarm records during the historical operation of the unit. The samples that trigger the combustion abnormality alarm are marked as combustion abnormal state, and the samples that do not trigger the combustion abnormality alarm are marked as normal combustion state.

8. The gas turbine combustion anomaly identification method based on pressure pulsation according to claim 1, characterized in that, Between step S4 and step S5, the method further includes: dividing the combustion anomaly identification sample set into a training set and a test set according to the chronological order, so as to avoid information from future moments being leaked to historical moments.

9. The gas turbine combustion anomaly identification method based on pressure pulsation according to claim 8, characterized in that, In step S5, when the number of abnormal combustion samples in the training set is less than the number of normal samples, the abnormal combustion samples in the training set are oversampled, and when training the abnormal combustion identification model, the abnormal combustion samples are assigned a higher sample weight than the normal samples, so as to improve the detection capability of the abnormal combustion identification model for abnormal combustion states.

10. The gas turbine combustion anomaly identification method based on pressure pulsation according to any one of claims 1 to 9, characterized in that: Step S6 is performed using a real-time rolling recognition method, which includes: initializing a fixed-length sliding window and caching the pressure pulsation data of the most recent consecutive sampling points; when a new pressure pulsation sampling value is received, the sampling value is written into the sliding window and the earliest sampling value entering the sliding window is deleted; based on the updated sliding window, the feature parameters corresponding to the current moment are extracted, and the feature parameters are input into a pre-trained combustion anomaly recognition model to output the combustion anomaly probability at the current moment.

Images