Gas turbine combustor vibration consistency assessment method based on multi-modal fusion

Abstract

The present invention relates to the technical field of gas turbine vibration assessment, and discloses a gas turbine combustion chamber vibration consistency assessment method based on multimodal fusion, the method comprising acquiring a multidimensional historical combustion vibration data set of the gas turbine combustion chamber; segmenting the data set to form sub-data sets; constructing a multimodal feature vector based on the sub-data sets; constructing a deep feature extraction model to extract deep features of the multimodal feature vector; performing cluster analysis on the deep features to divide the health status level of the combustion chamber vibration to form a multi-level health status level criterion; and determining the health status level of the current combustion chamber vibration based on the real-time status signal of each combustion chamber under the current operating state. The method can more accurately assess the vibration consistency of the gas turbine combustion chamber through multimodal data fusion and deep feature extraction, thereby improving the safety and reliability of gas turbine operation.

Description

Technical Field

[0001] The present invention relates to the technical field of gas turbine vibration assessment, and more particularly to a gas turbine combustion chamber vibration consistency assessment method based on multi-modal fusion. Background Art

[0002] Gas turbines are important power equipment widely used in power generation, shipbuilding, and other fields. The vibration health of gas turbines is crucial for their safe and stable operation. The vibration consistency of the combustor, a core component of the gas turbine, has a crucial impact on the vibration health of the entire unit. While research on gas turbine combustor vibration consistency assessment based on multimodal fusion has made some progress, some shortcomings remain.

[0003] Chinese patent application CN113688531B proposes a gas turbine reliability assessment method that incorporates two-dimensional vibration parameters with measurement errors. This method considers the impact of sensor measurement errors on vibration test data processing and uses a Frank Copula function to correlate two vibration test data sets to assess the reliability of the gas turbine's core rotor system. However, this method primarily focuses on the rotor system, with limited research on combustor vibration consistency assessment. Furthermore, this method primarily relies on vibration signals and lacks consideration of other modal signals, such as temperature. Therefore, the comprehensiveness of the assessment needs to be improved.

[0004] Chinese patent application CN113176094B discloses an online monitoring system for thermoacoustic oscillations in gas turbine combustors. This system utilizes multiple sensors installed on each flame tube to collect real-time data on multiple attributes, including pressure pulsation, vibration acceleration, and infrared high temperature. It then performs multi-domain analysis and provides fault warnings. This system improves the timeliness and reliability of combustor thermoacoustic oscillation monitoring and early warning. However, the system primarily focuses on thermoacoustic oscillation phenomena and has limited ability to assess general vibration issues. Furthermore, the system evaluates each combustor independently, lacking a comprehensive assessment of vibration consistency.

[0005] In summary, the existing technology lacks a method to integrate multimodal information such as vibration and temperature in the gas turbine combustor vibration consistency assessment based on multimodal fusion, and fails to evaluate the vibration consistency of multiple combustors from a holistic perspective. Summary of the Invention

[0006] To overcome the aforementioned shortcomings of the prior art, the present invention provides a gas turbine combustor vibration consistency assessment method based on multimodal fusion. This method acquires vibration and temperature signals from multiple combustors, segments and extracts features from historical data, constructs a deep learning model to mine deep features, and forms multi-level health status criteria, ultimately achieving an accurate assessment of the current combustor vibration consistency. This method is able to fuse multimodal information, deeply mine complex operating condition data, and comprehensively assess the vibration consistency of multiple combustors.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] The gas turbine combustor vibration consistency assessment method based on multi-modal fusion includes:

[0009] Acquire a multi-dimensional historical combustion vibration dataset of a gas turbine combustor; the historical combustion vibration dataset includes vibration signals and temperature signals of M combustors; segment the historical combustion vibration dataset to form n1 sub-datasets; construct a preliminary feature vector based on Mahalanobis distance for each sub-dataset, and construct a multimodal feature vector based on mutual information weighted fusion based on the preliminary feature vector;

[0010] A deep feature extraction model is constructed to extract deep features of multimodal feature vectors. Cluster analysis is performed on the deep features to classify the health status of combustion chamber vibration. The cluster centers of each health status level are calculated to form a multi-level health status criterion.

[0011] The real-time status signal of each combustion chamber under the current operating state is collected, and the health status level of the current combustion chamber vibration is determined based on the real-time status signal and the multi-level health status level criterion.

[0012] Furthermore, the segmenting of the historical combustion vibration data set to form n1 sub-data sets includes:

[0013] Extracting gas turbine operating parameters corresponding to a historical combustion vibration data set, the operating parameters including load and speed;

[0014] A load change threshold δ is set, and a traversal analysis of the historical combustion vibration dataset is performed on the time series. Starting at time t0, if the load change at time t1 relative to the starting time t0 exceeds the load change threshold δ, the operating condition is considered to have switched, and time t1 is marked as the operating condition switching point. Then, starting at time t1 as the new starting point, the marking process of the operating condition switching point is repeated until the end of the historical combustion vibration dataset is reached.

[0015] According to the operating condition switching point, the historical combustion vibration data set is divided into n2 initial operating condition segments;

[0016] Cluster the operating parameters within the initial operating condition segment and divide the historical combustion vibration data set into n1 subdivided operating condition segments;

[0017] Generate a corresponding sub-dataset for each subdivided operating condition segment, and obtain n1 sub-datasets.

[0018] Furthermore, clustering the operating condition parameters in the initial operating condition section includes:

[0019] The operating parameters in the initial operating condition section are used as sample features to form a feature matrix X. The dimension of the feature matrix X is m×n, where m is the number of samples and n is the number of features.

[0020] Set the value range of cluster number K to ,in Indicates rounding down; traverse the cluster number K value and calculate the silhouette coefficient sequence and DBI index sequence;

[0021] Calculate the comprehensive metric index sequence according to the silhouette coefficient sequence and the DBI index sequence, and take the cluster number corresponding to the maximum value in the comprehensive metric index sequence as the optimal cluster number N1;

[0022] Fix the optimal cluster number N1, use the K-Means algorithm to cluster the feature matrix X, obtain n1 subdivided operating conditions and the cluster center of each subdivided operating condition segment, and form the cluster center vector of the subdivided operating condition segment; where N1=n1.

[0023] Furthermore, the constructing of a preliminary feature vector based on Mahalanobis distance includes:

[0024] Calculate the Mahalanobis distance of the vibration signal of each combustion chamber relative to the mean of the sub-data set to form a vibration Mahalanobis distance vector ,in, represents the vibration Mahalanobis distance vector of combustion chamber j in the i-th operating condition, i represents the operating condition number, and j represents the combustion chamber number;

[0025] Calculate the Mahalanobis distance of the temperature signal of each combustion chamber relative to the mean of the sub-dataset to form the temperature Mahalanobis distance vector ;in, represents the temperature Mahalanobis distance vector of combustion chamber j in the i-th operating condition;

[0026] The vibration Mahalanobis distance vector and the temperature Mahalanobis distance vector are combined into a preliminary feature vector according to the combustion chamber sequence number. , , Represents the combustion chamber of the i-th operating condition The initial feature vector of .

[0027] Furthermore, the construction of a multimodal feature vector based on mutual information weighted fusion includes:

[0028] For each combustion chamber, calculate the mutual information matrix of the vibration signal and the temperature signal ; represents the mutual information matrix between the vibration signal and the temperature signal in the i-th operating condition;

[0029] Mutual Information Matrix Perform normalization to obtain the fusion weight matrix ; represents the fusion weight matrix of the i-th operating condition;

[0030] According to the fusion weight matrix Weighted fusion of the i-th working condition segment The vibration Mahalanobis distance vector of the combustion chamber and temperature Mahalanobis distance vector , construct multimodal fusion features ; Indicates the i-th operating condition segment Multimodal fusion features of each combustion chamber;

[0031] The multimodal fusion features of each combustion chamber Composition Multimodal feature vectors of operating conditions .

[0032] Furthermore, the deep feature extraction model includes a convolutional neural network and an autoencoder;

[0033] The deep features of the multimodal feature vector extraction include:

[0034] The multimodal feature vector is input into the convolutional neural network to extract the high-dimensional feature map; the high-dimensional feature map output by the convolutional neural network is input into the autoencoder to learn the globally compressed low-dimensional features; the globally compressed low-dimensional features output by the autoencoder are used as the deep features of the multimodal feature vector.

[0035] Furthermore, the health status levels of the combustion chamber vibration include normal, warning, and abnormal;

[0036] The calculation of the cluster center of each health status level to form a multi-level health status level criterion includes:

[0037] Calculate the coordinates Cn, Cw, Ca of the cluster centers of the three health status levels of normal, warning, and abnormal respectively;

[0038] Calculate the distance between each sample in the normal state cluster and the normal state cluster center coordinate Cn, and record the maximum distance as d1;

[0039] Calculate the distance from each sample in the warning state cluster to the center coordinate Cn of the normal state cluster, and record the minimum distance as d2; let δ1=1 / 2(d1+d2) as the boundary threshold of the normal state;

[0040] Calculate the distance between each sample in the warning state cluster and the coordinate Cw of the warning state cluster center, and record the maximum distance as d3;

[0041] Calculate the distance between each sample in the abnormal state cluster and the center coordinate Cw of the warning state cluster, and record the minimum distance as d4; let δ2=1 / 2(d3+d4) as the boundary threshold of the warning state;

[0042] Calculate the distance from each sample in the abnormal state cluster to the abnormal state cluster center coordinate Ca, record the maximum distance as d5, and set δ3=2×d5 as the boundary threshold of the abnormal state.

[0043] Furthermore, the real-time status signal includes a real-time vibration signal and a real-time temperature signal;

[0044] Determining the current health status level of the combustion chamber vibration based on the real-time status signal and the multi-level health status level criterion includes:

[0045] Determine the operating condition segment to which the current operating state belongs and mark it as the current operating condition segment; calculate the real-time multimodal feature vector of the real-time state signal;

[0046] The real-time multimodal feature vector is input into the deep feature extraction model to obtain the real-time deep feature. The Euclidean distances Dn, Dw, Da between the real-time deep feature and the coordinates Cn, Cw, Ca of the cluster centers of the three health status levels of normal, warning, and abnormal are calculated to determine the health status level of the current combustion chamber vibration.

[0047] Furthermore, determining the operating condition section to which the current operating state belongs includes:

[0048] Extract the real-time operating parameters under the current operating state; perform dimensionless processing on the real-time operating parameters and convert them into the real-time operating vector x (c) ; Calculate the Euclidean distance between the real-time working condition vector and the cluster center vector of the subdivided working condition segment; (c) Divide into the subdivided operating condition segment corresponding to the nearest cluster center as the operating condition segment of the current operating state;

[0049] The calculating of the real-time multimodal feature vector of the real-time state signal comprises:

[0050] Calculate the Mahalanobis distance of the real-time vibration signal relative to the mean of the sub-data set of the operating section to form a real-time preliminary feature vector; according to the fusion weight matrix and a real-time preliminary feature vector to construct a real-time multi-modal feature vector based on mutual information weighted fusion.

[0051] Further, the determination of the health status level of the current combustion chamber vibration includes:

[0052] If Dn ≤ δ1 and Dw ≤ Da, it is determined as the normal state;

[0053] If δ1 < Dn ≤ δ2 and Dw ≤ Da, it is determined as the warning state;

[0054] If δ2 < Dn ≤ δ3, it is determined as the abnormal state.

[0055] A gas turbine combustion chamber vibration consistency evaluation system based on multi-modal fusion, which is used to implement the above-mentioned gas turbine combustion chamber vibration consistency evaluation method based on multi-modal fusion. The system includes:

[0056] Feature fusion module: used to obtain the multi-dimensional historical combustion vibration data set of the gas turbine combustion chamber; the historical combustion vibration data set includes the vibration signals and temperature signals of M combustion chambers; segment the historical combustion vibration data set to form n1 sub-data sets; for each sub-data set, construct a preliminary feature vector based on the Mahalanobis distance, and construct a multi-modal feature vector based on mutual information weighted fusion according to the preliminary feature vector;

[0057] Criterion generation module: used to construct a deep feature extraction model to extract the deep features of the multi-modal feature vector; perform clustering analysis on the deep features to divide the health status levels of the combustion chamber vibration; calculate the clustering centers of each health status level to form a multi-level health status level criterion;

[0058] Health level division module: used to collect the real-time status signals of each combustion chamber under the current operating state, and determine the health status level of the current combustion chamber vibration according to the real-time status signals and the multi-level health status level criterion.

[0059] An electronic device includes a memory, a central processing unit, and a computer program stored on the memory and executable on the central processing unit. When the central processing unit executes the computer program, it implements the above-mentioned gas turbine combustion chamber vibration consistency evaluation method based on multi-modal fusion.

[0060] A computer-readable storage medium stores a computer program, and when the computer program is executed, it implements the above-mentioned gas turbine combustion chamber vibration consistency evaluation method based on multi-modal fusion.

[0061] Compared with the prior art, the beneficial effects of the present invention are:

[0062] By fusing multimodal data from vibration and temperature signals and combining it with deep learning to extract deep features, the vibration state of the combustion chamber can be more comprehensively and accurately characterized, improving the accuracy and reliability of the assessment. Adaptive operating condition segmentation and clustering methods can automatically divide different operating conditions based on actual operating data, enabling the assessment method to adapt to changes in vibration characteristics under different operating conditions and showing strong adaptability. By constructing multi-level health status grade criteria, real-time dynamic assessment of the combustion chamber vibration state can be achieved, allowing potential problems to be discovered in a timely manner. Accurate assessment of the combustion chamber vibration consistency can detect abnormal conditions early, effectively reducing gas turbine operating risks and improving equipment safety. Accurate assessment of the vibration state helps optimize maintenance strategies, avoid unnecessary shutdowns and overhauls, and extend the service life of the gas turbine. Accurate judgment of the vibration state can achieve predictive maintenance, reduce unplanned downtime, and reduce maintenance costs. Accurate assessment of vibration consistency helps optimize the working state of the combustion chamber, improve combustion efficiency, and thus enhance the overall efficiency of the gas turbine. Automated data processing and evaluation processes reduce the subjectivity of human judgment and improve the objectivity and consistency of the assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0064] Figure 1 This is a principle flow chart of the gas turbine combustor vibration consistency assessment method based on multi-modal fusion in the present invention;

[0065] Figure 2 This is a flow chart of a method for segmenting a historical combustion vibration data set in a gas turbine combustor vibration consistency assessment method based on multi-modal fusion according to the present invention;

[0066] Figure 3 A flow chart of a method for constructing a preliminary feature vector based on Mahalanobis distance in a gas turbine combustor vibration consistency assessment method based on multimodal fusion according to the present invention;

[0067] Figure 4 A flow chart of a method for constructing a multimodal feature vector based on mutual information weighted fusion in a gas turbine combustor vibration consistency assessment method based on multimodal fusion according to the present invention;

[0068] Figure 5A flow chart of a method for extracting deep features of multimodal feature vectors in a gas turbine combustor vibration consistency assessment method based on multimodal fusion according to the present invention;

[0069] Figure 6 A flow chart of a method for forming a multi-level health status grade criterion in a gas turbine combustor vibration consistency assessment method based on multi-modal fusion according to the present invention;

[0070] Figure 7 This is a flow chart of a method for determining the operating condition section to which the current operating state belongs in the gas turbine combustor vibration consistency assessment method based on multi-modal fusion of the present invention;

[0071] Figure 8 This is a functional module diagram of the gas turbine combustor vibration consistency assessment system based on multi-modal fusion in the present invention. DETAILED DESCRIPTION

[0072] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0073] Example 1

[0074] See also Figure 1 As shown, this embodiment provides a gas turbine combustor vibration consistency assessment method based on multi-modal fusion, including:

[0075] Step S1000: Acquire a multi-dimensional historical combustion vibration dataset of a gas turbine combustor; the historical combustion vibration dataset includes a vibration signal and a temperature signal; segment the historical combustion vibration dataset to form n1 sub-datasets; construct a preliminary feature vector based on Mahalanobis distance for each sub-dataset, and construct a multimodal feature vector based on mutual information weighted fusion based on the preliminary feature vector;

[0076] Furthermore, step S1000 includes:

[0077] Step S1100, obtaining a multi-dimensional historical combustion vibration dataset of a gas turbine combustion chamber; the historical combustion vibration dataset includes vibration signals and temperature signals of M combustion chambers; M is a positive integer;

[0078] Specifically, to obtain data that comprehensively reflects the state of the combustion chamber, precise acquisition of vibration and temperature signals is required. A piezoelectric accelerometer suitable for high-temperature environments is selected, and multiple sensors are arranged on the outer surface of the flame tube end cover of each combustion chamber, evenly distributed circumferentially and perpendicular to the flame tube axis, ensuring that vibration signals can be captured in each combustion chamber. The sensor collects vibration acceleration signals from the combustion chamber structural components, reflecting the vibration state during the combustion process. Vibration signals from each combustion chamber are collected, including acceleration, frequency, and other information. The high temperature and high pressure environment of the combustion process places stringent demands on sensors. The selected piezoelectric accelerometer has excellent high-temperature resistance and frequency response characteristics, and is highly reliable. The accelerometer is placed at the key location of the flame tube—the end cover, which bears the cantilever load of the hot end components and has a significant vibration response. Using a multi-point arrangement, the circumferential and axial vibrations of the combustion chamber can be collected, and spatial distribution characteristics can be obtained.

[0079] Platinum resistance temperature sensors were selected and inserted into the combustion chamber through a dedicated mounting sleeve. The temperature sensors used were high-precision, stable platinum resistors, along with dedicated high-temperature sleeves and a reasonable burial depth, ensuring reliable measurement of the temperature probe in high-temperature airflow. Multiple temperature sensors were placed at the head and tail to obtain temperature distribution along the process, providing a basis for subsequent analysis of the non-uniformity of the hot-end temperature field. The high-frequency sampling and continuous data acquisition of the data acquisition instrument can obtain dynamic characteristic data of the combustion chamber at different times, providing data support for exploring deep-level time-frequency domain characteristics. The rational arrangement of sensors, coupled with high-quality data acquisition equipment, ultimately formed a comprehensive and detailed multi-dimensional time series data set, which is an important foundation for vibration consistency assessment.

[0080] Step S1200 , segmenting the historical combustion vibration data set to form n1 sub-data sets;

[0081] Furthermore, if Figure 2 As shown, step S1200 includes:

[0082] Step S1210, extracting gas turbine operating parameters corresponding to the historical combustion vibration data set, wherein the operating parameters include load and speed;

[0083] Step S1220: Set a load change threshold δ and perform a time series traversal analysis on the historical combustion vibration dataset. Starting at time t0, if the load change at time t1 relative to the starting time t0 exceeds the load change threshold δ, it is considered that the operating condition has switched, and time t1 is marked as the operating condition switching point. The above process is then repeated, starting at time t1, until the end of the historical combustion vibration dataset is reached.

[0084] Step S1230, dividing the historical combustion vibration data set into n2 initial operating condition segments according to the operating condition switching point;

[0085] Step S1240 , clustering the operating condition parameters in the initial operating condition segment, and dividing the historical combustion vibration data set into n1 subdivided operating condition segments;

[0086] Furthermore, step S1240 includes:

[0087] Step S1241: The operating parameters in the initial operating condition segment are used as sample features to form a feature matrix X. The dimension of the feature matrix X is m×n, where m is the number of samples and n is the number of features.

[0088] Step S1242, set the value range of the cluster number K to ,in Indicates rounding down; traverse the cluster number K value and calculate the silhouette coefficient sequence and DBI index sequence;

[0089] Step S1243, calculating a comprehensive metric index sequence based on the silhouette coefficient sequence and the DBI index sequence, and taking the number of clusters corresponding to the maximum value in the comprehensive metric index sequence as the optimal number of clusters N1;

[0090] Step S1244: fix the optimal cluster number N1, cluster the feature matrix X using the K-Means algorithm, obtain n1 subdivided operating condition segments and the cluster center of each subdivided operating condition segment, and form the cluster center vector {C1, C2, ..., C n1}; Where N1=n1, C1 is the cluster center of the first subdivided operating condition segment, C2 is the cluster center of the second subdivided operating condition segment, and C n1 is the cluster center of the n1th subdivided operating condition segment.

[0091] Step S1250 : Generate a corresponding sub-data set for each subdivided operating condition segment, and obtain n1 sub-data sets.

[0092] Specifically, gas turbines undergo start-up and shutdown, variable operating conditions, and other processes during actual operation. The vibration characteristics of the combustion chamber vary significantly under different operating conditions. To improve the pertinence of subsequent evaluations, data from different operating conditions needs to be separated. First, key parameters reflecting the status of the gas turbine, such as load and speed, are extracted based on the operating records. Load refers to the active power of the generator set driven by the gas turbine and is a key parameter reflecting the output of the gas turbine. Speed ​​refers to the rotational speed of the gas turbine rotor, usually expressed in revolutions per minute (RPM), and is another important parameter reflecting the operating status of the gas turbine. By analyzing the operating records, a series of load and speed data corresponding to the historical vibration data can be obtained, providing a basis for the classification of operating conditions.

[0093] Considering the load fluctuations and measurement errors in actual operation, instead of regarding minor load changes as a working condition switch, a threshold value δ for load change is set. Only when the load change exceeds this threshold value is it considered that a significant working condition switch has occurred, and it is marked as a working condition switch point. The threshold value δ can be set according to experience, such as 5% of the rated load. Taking the load as an example, when the load change exceeds the preset threshold value such as 5%, it can be determined as a significant working condition switch. Based on this until the end of the data, several switching moments can be obtained. Segmenting the dataset with the switching moments as the boundaries can initially achieve the separation of data under different working conditions.

[0094] Segmenting the data according to the load change threshold δ results in n2 initial working condition segments. This division method mainly considers the impact of load mutations and cuts the data with the load switching point as the boundary. However, in actual operation, even if the load change is small, changes in other working condition parameters such as rotational speed may also lead to differences in combustion states. To further refine the division of working conditions, it is necessary to comprehensively consider the similarity of each working condition parameter, which requires performing clustering analysis based on the initial segmentation. Clustering analysis is an unsupervised learning method that can automatically divide data into different groups according to the similarity of sample features, making the similarity within the group high and the difference between groups large. Taking the working condition parameters such as load and rotational speed within each initial working condition segment as sample features, through clustering, samples with more similar working conditions can be automatically grouped into one category, obtaining a more detailed division of working conditions. The number of refined working conditions n1 formed by clustering is usually determined by the distribution characteristics of the data itself and may be greater than or less than the initial segmentation number n2. When the sample features of different initial working condition segments are very close, they are likely to be grouped into the same category, resulting in n1 < n2; on the contrary, when there are obvious differences within the same initial working condition segment, it may be subdivided into multiple working conditions, resulting in n1 > n2.

[0095] Taking the working condition parameters such as load and rotational speed within the initial working condition segment as sample features, a feature matrix X is constructed, and necessary dimensionless processing is performed on the feature matrix X. Organizing the multivariate features of each sample into a matrix form is convenient for unified algebraic operations and transformations. Different features often have different dimensions and numerical ranges. To eliminate the influence brought by this difference, it is necessary to perform dimensionless processing on the features, such as maximum-minimum normalization, Z-score standardization, etc. The silhouette coefficient measures the similarity of a sample to its belonging class and the difference from other nearest classes, with a value range between [-1, 1], and the larger the better. The DBI value range is [0, +∞), which comprehensively considers the compactness within the cluster and the separation between clusters. The smaller the DBI value, the better the clustering effect. Set the value range of K to [2, ], traverse each K value, calculate the corresponding silhouette coefficient and DBI index, and select the optimal number of clusters N1. To comprehensively consider the silhouette coefficient and DBI index, a combined metric function can be constructed, taking a weighted average of the two indicators. The weights can be adjusted according to actual needs to calculate a comprehensive metric index. The sum of the weights of the two indicators is equal to 1. A larger weight for the silhouette coefficient indicates a greater emphasis on cluster compactness, while a larger weight for the DBI index indicates a greater emphasis on cluster separation. The optimal number of clusters is selected as the number that maximizes the comprehensive metric index.

[0096] Fix the optimal number of clusters, N1, and use the K-Means algorithm to perform clustering training on the feature matrix X. Initialize N1 cluster centers, and iteratively optimize to minimize the cost function, continuously updating the cluster centers and sample class labels until convergence. Treat each cluster as a separate operating condition, resulting in n1 sub-conditions, where n1 = N1. Extract the time periods within each sub-condition, aggregate the vibration and temperature data within these time periods, and generate a sub-dataset corresponding to that condition.

[0097] Through the above clustering steps, the initial n2 coarse operating condition segments are ultimately refined into n1 more refined operating condition categories, each forming a subset of the dataset. This clustering subdivision can comprehensively characterize the similarities of operating conditions in the high-dimensional feature space, compensating for the shortcomings of single feature division and making the resulting operating condition subsets more accurate and reliable. The silhouette coefficient and DBI index can be used to quantitatively evaluate the clustering quality under different cluster numbers, providing an objective basis for operating condition clustering. Compared with manually setting the number of clusters, this data-driven optimization method can adaptively discover the inherent cluster structure of the data, avoiding subjectivity and blindness.

[0098] Clustering and subdividing operating conditions improves the pertinence of subsequent modeling. Different operating conditions have different data characteristics, and the factors that affect the combustion state vary. After subdividing the operating conditions, the unique influencing mechanism of each operating condition can be explored, which helps to improve the accuracy of the evaluation model. It reduces modeling complexity. If all operating condition data are modeled together, the model needs to have strong expressive capabilities to depict complex data distributions, which often comes at the cost of model complexity. After subdividing the operating conditions, the data distribution within each sub-condition is relatively simple, the requirements for model capabilities are reduced, and it is easier to obtain an ideal fitting effect. It is conducive to knowledge accumulation. The natural division of operating conditions obtained through clustering reflects the inherent laws of the actual operation of the gas turbine. By analyzing the characteristics of different operating conditions, we can summarize their commonalities and individualities, formulate targeted operation and maintenance strategies, and accumulate valuable engineering experience.

[0099] For example, the historical data of a gas turbine is initially divided into five operating conditions. Then, the load, speed and other data within each operating condition are extracted and normalized to form a feature matrix with 1000 samples and 4 features. The K-Means algorithm is then used to cluster the matrix, with K ranging from 2 to 10, and the silhouette coefficient and DBI index are calculated. The results show that when K=7, the silhouette coefficient reaches a maximum value of 0.68 and the DBI index reaches a minimum value of 1.21, so the optimal number of clusters is determined to be 7. All samples are divided into 7 categories, each corresponding to a subdivided operating condition, and all vibration and temperature data within these 7 operating conditions are extracted to form 7 sub-datasets.

[0100] Step S1300 , constructing a preliminary feature vector based on Mahalanobis distance for each sub-dataset;

[0101] Furthermore, if Figure 3 As shown, step S1300 includes:

[0102] Step S1310: For each sub-data set, calculate the Mahalanobis distance of each combustion chamber vibration signal relative to the mean of the sub-data set to form a vibration Mahalanobis distance vector ,in, represents the vibration Mahalanobis distance vector of combustion chamber j in the i-th operating condition, i represents the operating condition number, and j represents the combustion chamber number;

[0103] Step S1320: For each sub-data set, calculate the Mahalanobis distance of each combustion chamber temperature signal relative to the mean of the sub-data set to form a temperature Mahalanobis distance vector ;in, represents the temperature Mahalanobis distance vector of combustion chamber j in the i-th operating condition;

[0104] Step S1330: The vibration Mahalanobis distance vector and the temperature Mahalanobis distance vector are combined into a preliminary feature vector according to the combustion chamber sequence number. , , Represents the preliminary eigenvector of combustion chamber j in the i-th operating condition.

[0105] Specifically, in step S1300, for each sub-dataset, appropriate features are designed to characterize the degree to which each combustion chamber's state deviates from its normal value. Mahalanobis distance is an effective anomaly detection method. By calculating the weighted distance between a sample point and the feature mean, it measures the degree to which a point deviates from the distribution center. First, for each combustion chamber's vibration data, the Mahalanobis distance is calculated relative to the mean of all samples in the sub-dataset, forming a one-dimensional feature vector reflecting the degree of vibration deviation. Similarly, for each combustion chamber's temperature data, a Mahalanobis distance feature vector is calculated. The vibration and temperature Mahalanobis distances corresponding to the same combustion chamber are combined to form a two-dimensional feature vector. This feature vector characterizes the degree of anomaly for each combustion chamber under various operating conditions, from both the vibration and temperature perspectives, providing a basis for subsequent vibration consistency assessment. Compared to traditional distance metrics such as Euclidean distance, Mahalanobis distance considers the correlation between different state parameters. A larger distance value indicates a higher probability that a sample point deviates from normal operating conditions. Combining the vibration and temperature Mahalanobis distances into a unified feature vector provides a more comprehensive reflection of the combustion chamber's health, avoiding misjudgments caused by single physical quantities. It can be seen that the construction of Mahalanobis distance features can effectively screen out abnormal data that deviates from normal operating conditions and is an important means to achieve preliminary fault detection.

[0106] Step S1400 : For each sub-dataset, construct a multimodal feature vector based on mutual information weighted fusion according to the preliminary feature vector.

[0107] Furthermore, if Figure 4 As shown, step S1400 includes:

[0108] Step S1410: For each combustion chamber, calculate the mutual information matrix of the vibration signal and the temperature signal ; represents the mutual information matrix between the vibration signal and the temperature signal in the i-th operating condition;

[0109] Step S1420: Mutual Information Matrix Perform normalization to obtain the fusion weight matrix ; represents the fusion weight matrix of the i-th operating condition;

[0110] Step S1430: According to the fusion weight matrix Weighted fusion of the i-th working condition segment The vibration Mahalanobis distance vector of the combustion chamber and temperature Mahalanobis distance vector , construct multimodal fusion features ; Indicates the i-th operating condition segment Multimodal fusion features of each combustion chamber;

[0111] Step S1440: Fusion of multimodal features of each combustion chamber Composition Multimodal feature vectors of operating conditions , , M is the number of combustion chambers, Indicates the The multi-modal characteristic vector of the first combustion chamber in each operating section, Indicates the The multi-modal characteristic vector of the second combustion chamber in the working section, Indicates the The multi-modal characteristic vector of the Mth combustion chamber in the operating section.

[0112] Specifically, in step S1400, in order to fully explore the intrinsic relationship between the vibration signal and the temperature signal, a multimodal fusion feature based on mutual information weighted fusion is further constructed. Mutual information can measure the correlation between two random variables. The larger the mutual information value, the stronger the correlation between the two variables. First, the mutual information between the vibration signal and the temperature signal in each working condition is calculated to form a mutual information matrix. The elements in the matrix Describes the The accelerometer and To facilitate subsequent fusion, the mutual information matrix is ​​normalized so that the sum of its elements is 1, and the fusion weight matrix is ​​obtained. . Normalized weights Reflects the acceleration-temperature sensor pair ( , ) is important. On this basis, the vibration Mahalanobis distance vector of each combustion chamber is calculated. and temperature Mahalanobis distance vector Perform weighted fusion. When fusion is performed, Each component and Each component of is combined in pairs and then multiplied by the corresponding fusion weight , and finally superimpose and sum to obtain the scalarized multimodal fusion features The fusion features of all combustion chambers are summarized to form the multimodal feature vector of the working condition. . The fusion feature integrates the complementary information of vibration and temperature, overcomes the limitations of a single modality, and can more comprehensively reflect the state of the combustion chamber. The introduction of mutual information-driven fusion weights allows sensor pairs with strong correlation to obtain higher weights in the fusion, highlighting the role of fault-related modalities, which has physical significance. The adaptive weight calculation method can adjust the weights in real time according to the dynamic changes of the data, thereby improving the robustness of the algorithm. Multimodal fusion reduces the redundancy and interference of information, obtains a more refined and high-level state representation, and is the key to achieving high-precision vibration consistency assessment.

[0113] Step S2000: constructing a deep feature extraction model to extract deep features of the multimodal feature vector; performing cluster analysis on the deep features to classify the health status of the combustion chamber vibration; calculating the cluster center of each health status level to form a multi-level health status level judgment criteria; the health status level of the combustion chamber vibration includes normal, warning, and abnormal;

[0114] Furthermore, step S2000 includes:

[0115] Step S2100: constructing a deep feature extraction model to extract deep features of the multimodal feature vector;

[0116] Furthermore, if Figure 5 As shown, step S2100 includes:

[0117] Step S2110: The deep feature extraction model includes a convolutional neural network and an autoencoder, and the multimodal feature vector is input into the convolutional neural network to extract a high-dimensional feature map;

[0118] Step S2120: inputting the high-dimensional feature map output by the convolutional neural network into the autoencoder to learn global compressed low-dimensional features;

[0119] Step S2130: Using the globally compressed low-dimensional features output by the autoencoder as deep features of the multimodal feature vector.

[0120] Specifically, in step S2100, a combined structure of a convolutional neural network and an autoencoder is used to construct an end-to-end deep feature extraction model. First, the multimodal feature vector obtained in step S1400 is input into the convolutional neural network. The convolutional neural network can efficiently extract local features of the data through local connections and weight sharing. Through operations such as convolution and pooling, the convolutional neural network converts the initial multimodal fusion features into increasingly abstract feature maps layer by layer. The shallower feature maps extract low-order statistics of vibration and temperature signals, while the deeper feature maps contain more complex and global features. The advantage of the convolutional neural network is that it gradually abstracts the data through a hierarchical structure and automatically learns the most valuable key features for state discrimination. These local features comprehensively reflect the detailed information of the combustion state.

[0121] The high-dimensional feature map output by the convolutional neural network is then input into an autoencoder network. An autoencoder is an unsupervised feature learning method consisting of two parts: an encoder and a decoder. The encoder maps the input data into a low-dimensional feature space, while the decoder attempts to reconstruct the original data from the compressed features. By minimizing the reconstruction error, the autoencoder learns an efficient compressed representation of the input data, eliminating redundancy and noise. In this step, the encoder compresses the convolutional feature map to a lower dimension, extracting the global features of the data. These global features summarize the main patterns of the combustion state in the most concise way. The decoder is used to constrain the encoder to learn meaningful features, making the compressed features highly interpretable.

[0122] Finally, the low-dimensional features output by the autoencoder are used as the deep feature representation of the multimodal feature vector. These deep features inherit the local feature extraction capabilities of convolutional neural networks and the global compression capabilities of autoencoders, comprehensively characterizing the inherent laws of combustion chamber vibration. Let M represent the number of combustion chambers and S represent the dimension of the deep features. The dimension of the deep features F output in step S2100 is M×S.

[0123] The construction of the deep feature extraction model is primarily based on data-driven end-to-end learning, with model parameters iteratively optimized using optimization algorithms such as stochastic gradient descent. Given a large number of multimodal fusion feature samples, the model can adaptively learn the optimal feature transformation from the data by minimizing loss functions such as reconstruction error. This data-driven learning paradigm overcomes the limitations of manually designed features, does not rely on expert experience and prior knowledge, and can discover key patterns inherent in the data. By extracting local features through convolutional neural networks and learning global compressed features with autoencoders, it is possible to characterize the multi-level features of the combustion state at different scales, preserving detailed information while revealing global patterns. Compared with traditional manual features, features extracted by deep learning are often more discriminative, robust, and adaptable, laying a solid foundation for subsequent state clustering.

[0124] In summary, step S2100 uses a deep learning model to extract highly refined deep features from multimodal fusion features. These deep features summarize the intrinsic patterns of combustion chamber vibration in the most concise and discriminative way, reflecting the advantages of learning from multi-source data. Deep feature extraction is the key to knowledge discovery and intelligent decision-making, and directly determines the accuracy of subsequent state clustering. The state criteria constructed based on deep features can quickly and accurately assess the health level of the combustion chamber and guide operation and maintenance personnel to take targeted maintenance measures. It can be seen that deep feature learning is of great significance in combustion chamber vibration monitoring and represents a new direction for intelligent fault diagnosis and health management.

[0125] Step S2200 , cluster analysis is performed on the deep features of the multimodal feature vector to classify the health status levels of the combustion chamber vibration; the health status levels of the combustion chamber vibration include normal, warning, and abnormal.

[0126] Specifically, in step S2200, the deep feature vectors are divided into several state clusters using the cluster analysis method, and these clusters are further classified into three health status levels. Cluster analysis is a typical unsupervised learning method that does not require prior labeling of data. Instead, it automatically divides data by maximizing the similarity within the cluster and the difference between clusters. In this step, K-means, as the most classic clustering algorithm, is used to characterize the distribution pattern of deep features. K-means can discover natural cluster structures in the feature space by optimizing the sum of the squares of the distances between samples and cluster centers. By iteratively optimizing the cluster centers and cluster properties of samples, K-means can discover meaningful state patterns from deep features, and each cluster corresponds to a typical combustion chamber working state.

[0127] A key issue in the K-means algorithm is the selection of the number of clusters. Different numbers of clusters lead to different clustering results. Too small a number of clusters may overlook certain important structures in the data, while too large a number of clusters may divide essentially similar samples into different clusters, reducing the generalizability of the clustering. In practical applications, it is often necessary to optimize the number of clusters. In step S2220, the silhouette coefficient and the Davies-Bouldin Index (DBI) are used as evaluation metrics for clustering performance. The silhouette coefficient measures the similarity of a sample to its cluster and the difference from other clusters, ranging from -1 to 1. A larger value indicates a better clustering effect. The DBI index comprehensively considers the compactness within a cluster and the degree of separation between clusters. A smaller value indicates a higher clustering quality. By optimizing the silhouette coefficient and DBI index through cross-validation, the optimal number of clusters can be selected from the data, avoiding overfitting local features and underfitting global structure.

[0128] After obtaining the optimal clustering results, these clusters need to be further interpreted as different health status levels. Based on expert knowledge and historical experience, clusters can be divided into three types: normal, warning, and abnormal. The combustion chamber near the center of the normal cluster has the most stable operating state, and its vibration characteristics are close to those of most samples, representing a healthy baseline state. Clusters that are far away from the center of the normal cluster but have not yet reached the level of failure have vibration levels that are higher than normal but still within the controllable range and are marked as warning status. For samples that are far away from the normal cluster and close to the center of the abnormal cluster, their vibration patterns are already obviously abnormal and can be judged as a fault state, requiring timely maintenance. By reasonably dividing the status levels, a refined assessment of the health level of the combustion chamber can be achieved, providing a quantifiable basis for status monitoring, fault warning, and maintenance decisions.

[0129] In summary, step S2200 constructs state clusters reflecting the typical working conditions of the combustion chamber in the deep feature space through cluster analysis and state division. These clusters summarize the key patterns in massive historical data in a data-driven manner, overcoming the subjectivity of manually defining state boundaries. Through the interpretation and association of state clusters, complex sensor signals can be intuitively mapped to health levels that are easy to make decisions, making state monitoring more intelligent. It is worth mentioning that the cluster analysis in this step is performed on the basis of deep features. The combination of deep feature learning and state clustering can maximize the intrinsic value of the data and form an interpretable and operational combustion chamber health assessment system. The multi-level state division based on clustering provides important support for subsequent hierarchical diagnosis and predictive maintenance, and represents a new way of equipment health management.

[0130] Step S2300, calculating the cluster center of each health status level to form a multi-level health status level criterion;

[0131] Furthermore, if Figure 6 As shown, step S2300 includes:

[0132] Step S2310 , respectively calculating the coordinates Cn, Cw, and Ca of the cluster centers of the three health status levels of normal, warning, and abnormal;

[0133] Step S2320, calculating the distance between each sample in the normal state cluster and the normal state cluster center coordinate Cn, and recording the maximum distance as d1;

[0134] Step S2330, calculating the distance between each sample in the warning state cluster and the center coordinate Cn of the normal state cluster, and recording the minimum distance as d2;

[0135] Step S2340, let δ1=1 / 2(d1+d2) as the boundary threshold of the normal state;

[0136] Step S2350, calculating the distance between each sample in the warning state cluster and the coordinate Cw of the center of the warning state cluster, and recording the maximum distance as d3;

[0137] Step S2360, calculating the distance between each sample in the abnormal state cluster and the center coordinate Cw of the warning state cluster, and recording the minimum distance as d4;

[0138] Step S2370: Set δ2=1 / 2(d3+d4) as the boundary threshold of the warning state;

[0139] Step S2380 , calculate the distance from each sample in the abnormal state cluster to the abnormal state cluster center coordinate Ca, record the maximum distance as d5, and set δ3=2×d5 as the boundary threshold of the abnormal state.

[0140] Specifically, the core of step S2300 is to use the clustering method of unsupervised learning to automatically discover the intrinsic cluster structure of samples in the deep feature space, and use the cluster center as a typical representative of the health state, and the cluster boundary as the criterion threshold for state division. K-means clustering is a typical distance-based partitioning clustering algorithm that achieves clustering by minimizing the sum of squares of the distances from the samples to the cluster center. When clustering, the Euclidean distance between each sample and the cluster center is calculated, and the samples are divided into the clusters with the closest distance. When the center is updated, the geometric center coordinates of the samples in each cluster are recalculated as the new cluster center. The above process is repeated continuously, so that the cluster center gradually stabilizes and can better represent the characteristic distribution of each type of sample.

[0141] This step clusters historical samples in the deep feature space into three categories, corresponding to the three health states of the combustion chamber: normal, warning, and abnormal. In the normal state, the combustion chamber parameters are stable and the operating conditions are close to those of most samples. In the warning state, the combustion chamber begins to fluctuate slightly, but has not yet reached an abnormal level. In the abnormal state, the combustion chamber deviates significantly from normal conditions and requires attention. These three typical states are automatically identified through clustering, and the coordinates of their cluster centers are used as representative features of the state. When a newly collected sample is close to a cluster center, it can be judged that the sample is likely to be in the healthy state represented by the center.

[0142] To quantitatively delineate the boundaries between different states, it's necessary to estimate appropriate thresholds based on clustering. The normal state boundary threshold, δ1, is calculated as the average of the maximum distance d1 within the normal cluster and the minimum distance d2 from the warning cluster to the normal center. Samples falling within the radius of δ1 tend to be classified as normal. The warning state boundary threshold, δ2, is calculated as the average of the maximum distance d3 within the warning cluster and the minimum distance d4 from the abnormal cluster to the warning center. The abnormal state boundary threshold, δ3, is calculated as twice the maximum distance d5 within the abnormal cluster. Samples with a distance from the abnormal center exceeding δ3 often indicate serious faults. This threshold selection method, based on clustering results, fully considers both intra-cluster compactness and inter-cluster separation, effectively defining the effective boundaries between different states.

[0143] This judgment method does not require manual setting of state thresholds. Instead, it automatically learns the distribution characteristics of typical states based on the clustering structure of historical data, which is adaptive. Secondly, using the cluster center as a state model can reduce the influence of individual extreme samples and improve the robustness of the judgment. Thirdly, the fusion takes into account the compactness within the cluster and the degree of separation between clusters, which can better balance the classification accuracy and generalization performance. Finally, by comparing the distances to the three cluster centers, the fuzzy interval caused by a single threshold division can be avoided, making the state judgment more refined.

[0144] Step S3000: collect the real-time status signal of each combustion chamber under the current operating state, and determine the health status level of the current combustion chamber vibration according to the multi-level health status level criterion.

[0145] Furthermore, step S3000 includes:

[0146] Step S3100: collecting real-time status signals of each combustion chamber under the current operating state, wherein the real-time status signals include real-time vibration signals and real-time temperature signals; determining the operating section to which the current operating state belongs, marking it as the current operating section; and calculating a real-time multimodal feature vector of the real-time status signals;

[0147] Furthermore, if Figure 7 As shown, step S3100 includes:

[0148] Step S3110, extracting real-time operating parameters under the current operating state;

[0149] Step S3120: Perform dimensionless processing on the real-time operating condition parameters and convert them into a real-time operating condition vector x (c) ;

[0150] Step S3130: Calculate the cluster center vectors {C1, C2, ..., C n1}Euclidean distance of each cluster center in the cluster to obtain the distance vector ; Represents the Euclidean distance between the real-time working condition vector and the cluster center of the first subdivided working condition segment, represents the Euclidean distance between the real-time working condition vector and the cluster center of the second subdivided working condition segment, Represents the Euclidean distance between the real-time working condition vector and the cluster center of the n1th subdivided working condition segment;

[0151] Step S3140: The real-time working condition vector x (c) Divide into the subdivided operating condition segment corresponding to the nearest cluster center as the operating condition segment of the current operating state;

[0152] Step S3150 , calculating the Mahalanobis distance of the real-time vibration signal relative to the mean of the sub-data set of the corresponding operating section to form a real-time preliminary feature vector;

[0153] Step S3160, according to the fusion weight matrix and real-time preliminary feature vectors, and construct a real-time multimodal feature vector based on mutual information weighted fusion.

[0154] Specifically, step S3100 is the online monitoring phase, which detects changes in the combustion chamber's vibration state in real time. To achieve this, the unit must be equipped with piezoelectric accelerometers and platinum resistance temperature sensors, identically configured as those used during the offline training phase. The sensor installation location, quantity, model, and specifications must remain consistent with those used during the offline phase to ensure comparability of the collected data.

[0155] In addition, it is also necessary to automatically determine the type of operating condition of the current signal based on the real-time operating parameters of the gas turbine, such as load and speed, using the operating condition classification rules summarized in the offline stage, and match the corresponding fusion weights and evaluation thresholds for subsequent feature calculations. First, extract the key parameters reflecting the current state of the gas turbine, such as load and speed, to form real-time operating condition parameters. Since the dimensions and numerical ranges of these parameters may differ from the training data in the offline stage, they need to be dimensionlessly processed to uniformly map each component to the interval [0,1] to obtain the real-time operating condition vector x (c)The specific method of dimensionless processing needs to be consistent with the offline stage to ensure that the vector x (c) Compare with the cluster center vectors of each subdivided working condition in the same scale space. In the offline stage, the K-Means algorithm is used to obtain n1 cluster center vectors (see step S1244). Each cluster center represents the typical position of a subdivided working condition in the high-dimensional feature space. Therefore, the vector x can be calculated. (c) The Euclidean distance from each cluster center is used to measure the closeness between the current working condition and each subdivided working condition. The smaller the distance, the more similar the current working condition parameters are to the characteristic distribution of the subdivided working condition. Traverse all n1 subdivided working conditions and calculate their distance from x (c) The Euclidean distance of . Finally, the subdivided working condition corresponding to the nearest cluster center is selected as the working condition segment to which the current operating state belongs. Once the working condition segment to which the state belongs is determined, the evaluation model for the working condition established in the offline stage can be called to perform an online evaluation of the current vibration consistency level. It can be seen that the above working condition matching process avoids the secondary clustering of real-time data, makes full use of the working condition division knowledge summarized in the offline stage, and realizes rapid working condition identification by comparing the distance between the normalized vector and the cluster center, which greatly reduces the computational complexity of the online stage. At the same time, the use of Euclidean distance, a simple and intuitive similarity measure, is easy to implement in engineering, and is consistent with the clustering method in the offline stage, maintaining consistency. In practical applications, the Euclidean distance can also be weighted and corrected to highlight the influence of certain important working condition parameters and further improve the accuracy of working condition division. In general, step S3200 provides an effective way to achieve real-time working condition division based on clustering results, laying the foundation for subsequent feature extraction, consistency evaluation and other links.

[0156] Then, the feature fusion method given in steps S1300 and S1400 is used to calculate the real-time multimodal feature vector of the currently collected real-time state signal. First, the Mahalanobis distance of the real-time vibration signal and real-time temperature signal of each combustion chamber relative to the mean of the operating condition segment data set to which it belongs is calculated to form a preliminary feature vector. Then, the mutual information fusion weight matrix calculated in the offline stage is used to perform weighted fusion on the vibration features and temperature features in the real-time preliminary feature vector to obtain the real-time multimodal fusion features of each combustion chamber under the current operating state. Finally, the multimodal fusion features of each combustion chamber are combined in sequence to form a multimodal feature vector that characterizes the current operating state of the entire combustion system. This series of calculations can be automatically completed with the help of a trained deep learning model to achieve rapid characterization and evaluation of the combustion state.

[0157] Through the preprocessing and feature fusion of real-time signals, the obtained multi-modal feature vectors can comprehensively reflect the dynamic changes of combustion chamber vibration. On the one hand, with reasonable sensor layout and strict data acquisition, the consistency and comparability of monitoring data and training data can be ensured; on the other hand, with automatic working condition judgment and dynamic loading of fusion weights, the online evaluation is given the ability to adapt to working conditions, ensuring the reliability and continuity of the evaluation process. Inputting the monitoring data into the trained deep learning model for inference can quickly calculate the degree of deviation of the current operating state from the normal mode, timely detect the abnormal signs of the combustion system, and provide a basis for fault warning and operation and maintenance decision-making.

[0158] Step S3200: Input the real-time multi-modal feature vectors into the deep feature extraction model to obtain real-time deep features; calculate the Euclidean distances Dn, Dw, Da between the real-time deep features and the coordinates Cn, Cw, Ca of the clustering centers of the three health state levels of normal, warning, and abnormal, and determine the health state level of the current combustion chamber vibration.

[0159] If Dn ≤ δ1 and Dw ≤ Da, it is determined to be in a normal state;

[0160] If δ1 < Dn ≤ δ2 and Dw ≤ Da, it is determined to be in a warning state;

[0161] If δ2 < Dn ≤ δ3, it is determined to be in an abnormal state;

[0162] If Dn > δ3, it is determined to be in a severe fault state.

[0163] Specifically, in step S3200, the vibration and temperature signals of the combustion chamber collected in real time are passed through the same feature construction steps as in the offline stage to form real-time multi-modal feature vectors. Then, the real-time multi-modal feature vectors are input into the trained deep feature extraction model to obtain the real-time deep features after dimensionality reduction. The real-time deep features are a high-level abstraction of the original high-dimensional features and can better reflect the essential characteristics of the data. To judge the health state level of the current combustion chamber, it is necessary to calculate the Euclidean distances between the real-time deep features and the clustering centers of the three health levels. The Euclidean distance is a commonly used similarity measure, and the smaller the distance, the closer the feature is to the center of the class. Cn, Cw, and Ca represent the coordinates of the clustering centers of the normal, warning, and abnormal levels respectively, and Dn, Dw, and Da represent the Euclidean distances from the real-time deep features to the three centers. δ1, δ2, and δ3 are the boundary thresholds of the three levels.

[0164] If the distance Dn from the real-time deep feature to the normal center does not exceed δ1, and at the same time the distance Dw to the warning center is less than the distance Da to the abnormal center, it can be determined that the current combustion chamber is in a normal state. This means that the current feature highly coincides with the normal mode, the combustion chamber operates smoothly, and all indicators are within a reasonable range.

[0165] If the real-time depth signature deviates from the normal center by more than δ1 but less than δ2, and the distance Dw to the warning center is less than the distance Da to the abnormal center, a warning state is determined. In this case, while the combustion chamber has experienced some degree of abnormality, it has not yet reached a level that seriously affects safety. This warrants attention and enhanced monitoring, but generally does not require emergency shutdown.

[0166] If the real-time depth signature deviates from the normal center by more than δ2 but no more than δ3, it is considered abnormal. This indicates a significant abnormality in the combustion chamber. While not yet causing permanent damage, it could deteriorate into a malfunction if not addressed promptly. Close monitoring is required, along with comprehensive analysis and analysis based on other information, and appropriate measures should be taken to prevent escalation.

[0167] If the distance Dn that the real-time depth signature deviates from the normal center exceeds the maximum threshold δ3, a serious fault is highly suspected. At this point, the combustion chamber's state is completely outside the normal range. If emergency measures are not taken immediately, an accident is likely to occur, resulting in serious consequences such as equipment damage and safety incidents.

[0168] The above discrimination method draws heavily on the principles of pattern recognition. Using distance metrics, it matches real-time features with standard patterns to automatically identify the closest health level. Distance thresholds are set to differentiate severity levels. δ1, δ2, and δ3 can be considered the boundaries between normal, warning, and abnormal levels. Setting multiple thresholds allows for a more detailed depiction of health status evolution, enabling early detection of abnormalities and providing a basis for subsequent maintenance decisions.

[0169] Compared to traditional single-threshold evaluation, this health assessment method provides a more comprehensive and detailed picture of the combustion chamber status, enabling "graded early warning and early detection." The progression from normal to warning to abnormal reflects the escalating severity of the problem. Multi-level evaluation vividly reveals the development trend of the fault, enabling personnel to take targeted preventive and response measures, ensuring proactive and effective response. Furthermore, this method leverages the complex data mining capabilities of deep learning to uncover deep features and patterns that are difficult to detect manually. Compared to shallow models, deep models have stronger feature representation and classification capabilities, can handle high-dimensional and nonlinear operating data, and have wider applicability. Because real-time status recognition is performed online, this method places high demands on the algorithm's timeliness. To achieve this, the model design can balance accuracy and speed, such as using lightweight networks with fewer parameters. Furthermore, computational efficiency can be improved through algorithm optimization and high-performance hardware. In summary, this online assessment method enables automatic grading of the combustion chamber's health status and real-time early warning, significantly facilitating risk control and equipment management. Through continuous monitoring and trend analysis, we can understand the dynamic changes in equipment health, gain insight into equipment degradation processes, and shift from passive response to proactive health management. This has important theoretical and practical value for improving the safety, reliability, economy, and environmental performance of gas turbines and promoting the advancement of intelligent operation and maintenance technologies.

[0170] Step S3300: Output the evaluation result according to the determined health status level, trigger the corresponding early warning information or fault alarm, and update the historical combustion vibration data.

[0171] Specifically, timely feedback of the combustion chamber health assessment results to relevant personnel is a key link in the practical effect of this method. To this end, the assessment system needs to be designed with a user-friendly human-computer interaction interface to display the health status of each combustion chamber in an intuitive and eye-catching manner. Commonly used information presentation forms include lists, charts, three-dimensional visualizations, etc., which can be flexibly selected according to the spatial geometric distribution of the assessment object. At the same time, to facilitate subsequent traceability and auditing, the results of each assessment must be bound to necessary information such as the combustion chamber number and assessment timestamp, and then packaged and persistently stored in the assessment record database to form a complete assessment report.

[0172] Example 2

[0173] This embodiment, based on the first embodiment, provides a gas turbine combustor vibration consistency assessment method based on multi-modal fusion, including:

[0174] Step S1243, calculating a comprehensive metric index sequence based on the silhouette coefficient sequence and the DBI index sequence, and taking the number of clusters corresponding to the maximum value in the comprehensive metric index sequence as the optimal number of clusters N1;

[0175] Calculating the comprehensive metric index sequence according to the silhouette coefficient sequence and the DBI index sequence includes:

[0176] ;

[0177] in:

[0178] : Comprehensive metric index, which is the number of clusters The function has a value range of The larger it is, the better the clustering effect is.

[0179] : Number of clusters, which is how the data is divided into categories. The value range is ,in is the sample size, Indicates rounding down.

[0180] :When the number of clusters is The silhouette coefficient when .

[0181] :When the number of clusters is DBI index at that time.

[0182] : Weight coefficient of silhouette coefficient.

[0183] : The weight coefficient of the DBI index, satisfying The larger it is, the more emphasis is placed on the cohesion of the category; The larger the value, the more emphasis is placed on the separation of categories.

[0184] : The power exponent of the silhouette coefficient, used to adjust the influence of the silhouette coefficient. When , the difference in silhouette coefficient is magnified; , the difference in silhouette coefficient is reduced.

[0185] and It needs to be set and adjusted by those skilled in the art according to the needs of actual problems. , that is, equal weight.

[0186] It is selected by those skilled in the art based on the distribution characteristics of the data and the sensitivity to the silhouette coefficient. , which is the squared silhouette coefficient.

[0187] : exponential function with the base e of natural logarithm as base;

[0188] In this formula, the numerator transforms the silhouette coefficient to a power, introducing the parameter , the influence of the silhouette coefficient can be flexibly adjusted. When , the difference of silhouette coefficient is magnified, making the evaluation of clustering quality pay more attention to the cohesion of categories; when , the difference in silhouette coefficient is reduced, making the evaluation smoother.

[0189] In the numerator, the DBI index is negatively exponentially transformed , mapping it to The negative exponential transformation can amplify the difference of DBI index, making the evaluation of clustering quality pay more attention to the separation of categories.

[0190] In the denominator, introduce the normalization factor , the comprehensive measurement index The value range is limited to Normalization processing makes the comprehensive metric index under different data sets and different parameter settings comparable, which facilitates the evaluation and selection of the optimal number of clusters.

[0191] Through power transformation and negative exponential transformation, the sensitivity of the comprehensive metric index to the differences between the silhouette coefficient and the DBI index is enhanced, making the evaluation of clustering quality more comprehensive and accurate.

[0192] Normalization processing makes the value range of the comprehensive measurement index fixed in There is a unified interpretation and comparison standard within the framework, which facilitates evaluation and selection under different data sets and different parameter settings. When the optimal number of clusters N1 is taken, the numerator is equal to the denominator; in other cases, . The closer it is to 1, the better the clustering effect; The closer it is to 0, the worse the clustering effect is.

[0193] In summary, through power transformation, negative exponential transformation, and normalization, this formula enhances the sensitivity of the comprehensive metric index to cluster quality differences and the accuracy of its evaluation, while also providing a unified interpretation and comparison standard. This formula, based on the existing silhouette coefficient and DBI index, introduces normalization, providing a more reliable and effective quantitative basis for operating condition classification and offering new insights and methods for evaluating other clustering problems.

[0194] Example 3

[0195] This embodiment provides a gas turbine combustion chamber vibration consistency assessment system based on multi-modal fusion on the basis of embodiment 1, such as Figure 8 Shown, including:

[0196] Feature fusion module: used to obtain a multi-dimensional historical combustion vibration dataset of a gas turbine combustor; the historical combustion vibration dataset includes vibration signals and temperature signals of M combustors; the historical combustion vibration dataset is segmented to form n1 sub-datasets; for each sub-dataset, a preliminary feature vector based on Mahalanobis distance is constructed; based on the preliminary feature vector, a multimodal feature vector based on mutual information weighted fusion is constructed;

[0197] Criteria generation module: used to build a deep feature extraction model to extract deep features of multimodal feature vectors; perform cluster analysis on the deep features to classify the health status of the combustion chamber vibration; calculate the cluster center of each health status level to form a multi-level health status level criterion; the health status level of the combustion chamber vibration includes normal, warning and abnormal;

[0198] Health level classification module: used to collect the real-time status signal of each combustion chamber under the current operating state, and determine the health status level of the current combustion chamber vibration based on the real-time status signal and multi-level health status level criteria.

[0199] The feature fusion module includes: a data acquisition unit, a data segmentation unit, a preliminary feature construction unit and a multimodal feature construction unit;

[0200] The data acquisition unit is used to acquire a multi-dimensional historical combustion vibration data set of a gas turbine combustion chamber; the historical combustion vibration data set includes vibration signals and temperature signals of M combustion chambers; M is a positive integer;

[0201] The data segmentation unit is used to segment the historical combustion vibration data set to form n1 sub-data sets;

[0202] The preliminary feature construction unit is used to construct a preliminary feature vector based on Mahalanobis distance for each sub-data set;

[0203] The multimodal feature construction unit is used to construct a multimodal feature vector based on mutual information weighted fusion according to the preliminary feature vector for each sub-data set.

[0204] The segmentation processing of the historical combustion vibration data set to form n1 sub-data sets includes:

[0205] Step S1210, extracting gas turbine operating parameters corresponding to the historical combustion vibration data set, wherein the operating parameters include load and speed;

[0206] Step S1220: Set a load change threshold δ and perform a time series traversal analysis on the historical combustion vibration dataset. Starting at time t0, if the load change at time t1 relative to the starting time t0 exceeds the load change threshold δ, it is considered that the operating condition has switched, and time t1 is marked as the operating condition switching point. The above process is then repeated, starting at time t1, until the end of the historical combustion vibration dataset is reached.

[0207] Step S1230, dividing the historical combustion vibration data set into n2 initial operating condition segments according to the operating condition switching point;

[0208] Step S1240 , clustering the operating condition parameters in the initial operating condition segment, and dividing the historical combustion vibration data set into n1 subdivided operating condition segments;

[0209] Step S1250 : Generate a corresponding sub-data set for each subdivided operating condition segment, and obtain n1 sub-data sets.

[0210] The clustering of the operating parameters in the initial operating condition segment and the division of the historical combustion vibration data set into n1 subdivided operating condition segments include:

[0211] Step S1241: The operating parameters in the initial operating condition segment are used as sample features to form a feature matrix X. The dimension of the feature matrix X is m×n, where m is the number of samples and n is the number of features.

[0212] Step S1242, set the value range of the cluster number K to ,in Indicates rounding down; traverse the cluster number K value and calculate the silhouette coefficient sequence and DBI index sequence;

[0213] Step S1243, calculating a comprehensive metric index sequence based on the silhouette coefficient sequence and the DBI index sequence, and taking the number of clusters corresponding to the maximum value in the comprehensive metric index sequence as the optimal number of clusters N1;

[0214] Step S1244: fix the optimal cluster number N1, cluster the feature matrix X using the K-Means algorithm, obtain n1 subdivided operating condition segments and the cluster center of each subdivided operating condition segment, and form the cluster center vector {C1, C2, ..., C n1}; Where N1=n1, C1 is the cluster center of the first subdivided operating condition segment, C2 is the cluster center of the second subdivided operating condition segment, and C n1 is the cluster center of the n1th subdivided operating condition segment.

[0215] For each sub-dataset, constructing a preliminary feature vector based on Mahalanobis distance includes:

[0216] Step S1310: For each sub-data set, calculate the Mahalanobis distance of each combustion chamber vibration signal relative to the mean of the sub-data set to form a vibration Mahalanobis distance vector ,in, represents the vibration Mahalanobis distance vector of combustion chamber j in the i-th operating condition, i represents the operating condition number, and j represents the combustion chamber number;

[0217] Step S1320: For each sub-data set, calculate the Mahalanobis distance of each combustion chamber temperature signal relative to the mean of the sub-data set to form a temperature Mahalanobis distance vector ;in, represents the temperature Mahalanobis distance vector of combustion chamber j in the i-th operating condition;

[0218] Step S1330: The vibration Mahalanobis distance vector and the temperature Mahalanobis distance vector are combined into a preliminary feature vector according to the combustion chamber sequence number. , , Represents the preliminary eigenvector of combustion chamber j in the i-th operating condition.

[0219] The method of constructing a multimodal feature vector based on mutual information weighted fusion for each sub-dataset according to the preliminary feature vector includes:

[0220] Step S1410: For each combustion chamber, calculate the mutual information matrix of the vibration signal and the temperature signal ; represents the mutual information matrix between the vibration signal and the temperature signal in the i-th operating condition;

[0221] Step S1420: Mutual Information Matrix Perform normalization to obtain the fusion weight matrix ; represents the fusion weight matrix of the i-th operating condition;

[0222] Step S1430: According to the fusion weight matrix Weighted fusion of the i-th working condition segment The vibration Mahalanobis distance vector of the combustion chamber and temperature Mahalanobis distance vector , construct multimodal fusion features ; Indicates the i-th operating condition segment Multimodal fusion features of each combustion chamber;

[0223] Step S1440: Fusion of multimodal features of each combustion chamber Composition Multimodal feature vectors of operating conditions , , M is the number of combustion chambers, Indicates the The multi-modal characteristic vector of the first combustion chamber in each operating section, Indicates the The multi-modal characteristic vector of the second combustion chamber in the working section, Indicates the The multi-modal characteristic vector of the Mth combustion chamber in the operating section.

[0224] In the criterion generation module, the construction of a deep feature extraction model to extract deep features of the multimodal feature vector includes:

[0225] Step S2110: The deep feature extraction model includes a convolutional neural network and an autoencoder, and the multimodal feature vector is input into the convolutional neural network to extract a high-dimensional feature map;

[0226] Step S2120: inputting the high-dimensional feature map output by the convolutional neural network into the autoencoder to learn global compressed low-dimensional features;

[0227] Step S2130: Using the globally compressed low-dimensional features output by the autoencoder as deep features of the multimodal feature vector.

[0228] In the criterion generation module, the calculation of the cluster center of each health status level to form a multi-level health status level criterion includes:

[0229] Step S2310 , respectively calculating the coordinates Cn, Cw, and Ca of the cluster centers of the three health status levels of normal, warning, and abnormal;

[0230] Step S2320, calculating the distance between each sample in the normal state cluster and the normal state cluster center coordinate Cn, and recording the maximum distance as d1;

[0231] Step S2330, calculating the distance between each sample in the warning state cluster and the center coordinate Cn of the normal state cluster, and recording the minimum distance as d2;

[0232] Step S2340, let δ1=1 / 2(d1+d2) as the boundary threshold of the normal state;

[0233] Step S2350, calculating the distance between each sample in the warning state cluster and the coordinate Cw of the center of the warning state cluster, and recording the maximum distance as d3;

[0234] Step S2360, calculating the distance between each sample in the abnormal state cluster and the center coordinate Cw of the warning state cluster, and recording the minimum distance as d4;

[0235] Step S2370: Set δ2=1 / 2(d3+d4) as the boundary threshold of the warning state;

[0236] Step S2380 , calculate the distance from each sample in the abnormal state cluster to the abnormal state cluster center coordinate Ca, record the maximum distance as d5, and set δ3=2×d5 as the boundary threshold of the abnormal state.

[0237] The health level classification module includes: a real-time feature extraction unit, a real-time health level classification unit and a result output unit;

[0238] The real-time feature extraction unit is used to collect real-time status signals of each combustion chamber under the current operating state, wherein the real-time status signals include real-time vibration signals and real-time temperature signals; determine the operating condition segment to which the current operating state belongs, and mark it as the current operating condition segment; and calculate the real-time multimodal feature vector of the real-time status signal;

[0239] The real-time health level classification unit is used to input the real-time multimodal feature vector into the deep feature extraction model to obtain the real-time deep feature; calculate the Euclidean distance Dn, Dw, Da between the real-time deep feature and the coordinates Cn, Cw, Ca of the cluster centers of the three health status levels of normal, warning, and abnormal, and determine the health status level of the current combustion chamber vibration;

[0240] The result output unit is used to output the evaluation result according to the determined health status level, trigger corresponding early warning information or fault alarm, and update historical combustion vibration data.

[0241] In the real-time feature extraction unit, determining the operating condition segment to which the current operating state belongs includes:

[0242] Step S3110, extracting real-time operating parameters under the current operating state;

[0243] Step S3120: Perform dimensionless processing on the real-time operating condition parameters and convert them into a real-time operating condition vector x (c) ;

[0244] Step S3130: Calculate the cluster center vectors {C1, C2, ..., C n1}Euclidean distance of each cluster center in the cluster to obtain the distance vector ; Represents the Euclidean distance between the real-time working condition vector and the cluster center of the first subdivided working condition segment, represents the Euclidean distance between the real-time working condition vector and the cluster center of the second subdivided working condition segment, Represents the Euclidean distance between the real-time working condition vector and the cluster center of the n1th subdivided working condition segment;

[0245] Step S3140: The real-time working condition vector x (c)Partition it into the refined working condition segment corresponding to the nearest clustering center as the working condition segment of the current operating state.

[0246] In the real-time feature extraction unit, the calculation of the real-time multi-modal feature vector of the real-time state signal includes:

[0247] Step S3150, calculate the Mahalanobis distance of the real-time vibration signal relative to the mean of the sub-dataset of the working condition segment to form a real-time preliminary feature vector;

[0248] Step S3160, according to the fusion weight matrix and the real-time preliminary feature vector, construct a real-time multi-modal feature vector based on mutual information weighted fusion.

[0249] In the real-time health level division unit, the determination of the health state level of the current combustion chamber vibration includes:

[0250] If Dn ≤ δ1 and Dw ≤ Da, it is determined to be in a normal state;

[0251] If δ1 < Dn ≤ δ2 and Dw ≤ Da, it is determined to be in a warning state;

[0252] If δ2 < Dn ≤ δ3, it is determined to be in an abnormal state;

[0253] If Dn > δ3, it is determined to be in a serious fault state.

[0254] Embodiment 4

[0255] This embodiment discloses an electronic device, which may include one or more processors and one or more memories. Among them, computer-readable code is stored in the memory, and when the computer-readable code is run by one or more processors, it can execute the above-mentioned method for evaluating the vibration consistency of a gas turbine combustion chamber based on multi-modal fusion.

[0256] The methods or systems according to the embodiments of this application can also be implemented using the architecture of an electronic device. The electronic device may include a bus, one or more CPUs, read-only memory (ROM), random access memory (RAM), a communication port connected to a network, input / output components, a hard disk, and the like. A storage device within the electronic device, such as a ROM or hard disk, can store the gas turbine combustor vibration consistency assessment method based on multimodal fusion provided herein. A gas turbine combustion chamber vibration consistency assessment method based on multimodal fusion may, for example, include: obtaining a multi-dimensional historical combustion vibration data set of a gas turbine combustion chamber; the historical combustion vibration data set includes vibration signals and temperature signals of M combustion chambers; segmenting the historical combustion vibration data set to form n1 sub-data sets; for each sub-data set, constructing a preliminary feature vector based on Mahalanobis distance, and based on the preliminary feature vector, constructing a multimodal feature vector based on mutual information weighted fusion; constructing a deep feature extraction model to extract deep features of the multimodal feature vector; performing cluster analysis on the deep features to divide the health status level of the combustion chamber vibration; calculating the cluster center of each health status level to form a multi-level health status level criterion; collecting the real-time status signal of each combustion chamber under the current operating state, and determining the health status level of the current combustion chamber vibration based on the real-time status signal and the multi-level health status level criterion.

[0257] Furthermore, the electronic device may further include a user interface. Of course, the architecture disclosed in the present invention is only exemplary, and when implementing different devices, one or more components in the electronic device disclosed in the present invention may be omitted according to actual needs.

[0258] Example 5

[0259] This embodiment discloses a computer-readable storage medium storing computer-readable instructions. When executed by a processor, the computer-readable instructions can execute the gas turbine combustor vibration consistency assessment method based on multimodal fusion according to an embodiment of this application. The storage medium includes, but is not limited to, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory. Non-volatile memory may include, for example, read-only memory (ROM), a hard disk, or flash memory.

[0260] In addition, according to an embodiment of the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the present application provides a non-transitory machine-readable storage medium, wherein the non-transitory machine-readable storage medium stores machine-readable instructions, wherein the machine-readable instructions can be executed by a processor to execute instructions corresponding to the method steps provided in the present application, for example: obtaining a multi-dimensional historical combustion vibration dataset of a gas turbine combustion chamber; the historical combustion vibration dataset includes vibration signals and temperature signals of M combustion chambers; segmenting the historical combustion vibration dataset to form n1 sub-datasets; for each sub-dataset, constructing a preliminary feature vector based on Mahalanobis distance, and constructing a multimodal feature vector based on mutual information weighted fusion based on the preliminary feature vector; constructing a deep feature extraction model to extract deep features of the multimodal feature vector; performing cluster analysis on the deep features to divide the health status level of the combustion chamber vibration; calculating the cluster center of each health status level to form a multi-level health status level criterion; collecting real-time status signals of each combustion chamber in the current operating state, and determining the health status level of the current combustion chamber vibration based on the real-time status signal and the multi-level health status level criterion. When the computer program is executed by a central processing unit (CPU), the above functions defined in the method of the present application are performed.

[0261] The methods, systems, and devices of the present application may be implemented in many ways. For example, the methods, systems, and devices of the present application may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above order of steps for the method is for illustration only, and the steps of the method of the present application are not limited to the order specifically described above unless otherwise specified. In addition, in some embodiments, the present application may also be implemented as programs recorded in a recording medium, which include machine-readable instructions for implementing the methods according to the present application. Therefore, the present application also covers recording media that store programs for executing the methods according to the present application.

[0262] In addition, the parts of the above technical solutions provided in the embodiments of the present application that are consistent with the implementation principles of the corresponding technical solutions in the prior art are not described in detail to avoid excessive redundancy.

[0263] The above-described specific embodiments further illustrate the objectives, technical solutions, and beneficial effects of the present invention. It should be understood that the above description is merely a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A gas turbine combustor vibration consistency assessment method based on multi-modal fusion, characterized in that: The method comprises: Acquire a multi-dimensional historical combustion vibration dataset of a gas turbine combustor; the historical combustion vibration dataset includes vibration signals and temperature signals of M combustors; segment the historical combustion vibration dataset to form n1 sub-datasets; The method for segmenting a historical combustion vibration data set comprises: extracting gas turbine operating condition parameters corresponding to the historical combustion vibration data set, identifying operating condition switching points based on the operating condition parameters, and dividing the historical combustion vibration data set into n2 initial operating condition segments based on the operating condition switching points; clustering the operating condition parameters within the initial operating condition segments to divide the historical combustion vibration data set into n1 subdivided operating condition segments; and generating a corresponding subdata set for each subdivided operating condition segment to obtain n1 subdata sets; For each sub-dataset, a preliminary feature vector based on Mahalanobis distance is constructed. Based on the preliminary feature vector, a multimodal feature vector based on mutual information weighted fusion is constructed. M and n1 are both positive integers. A deep feature extraction model is constructed to extract deep features of multimodal feature vectors. Cluster analysis is performed on the deep features to classify the health status of combustion chamber vibration. The cluster centers of each health status level are calculated to form a multi-level health status criterion. The real-time status signals of each combustion chamber under the current operating state are collected. The real-time status signals include real-time vibration signals and real-time temperature signals. The health status level of the current combustion chamber vibration is determined based on the real-time status signals and multi-level health status level criteria.

2. The gas turbine combustor vibration consistency assessment method based on multi-modal fusion according to claim 1 is characterized in that: The operating parameters include load and speed; The method for identifying the operating condition switching point based on the operating condition parameters includes: setting a load change threshold δ, performing a traversal analysis on the historical combustion vibration data set in a time series, taking time t0 as the starting point, and if the load change at time t1 relative to the starting time t0 exceeds the load change threshold δ, it is considered that the operating condition has switched, and time t1 is marked as the operating condition switching point; then, taking time t1 as a new starting point, the marking process of the operating condition switching point is repeated until the end.

3. The gas turbine combustor vibration consistency assessment method based on multi-modal fusion according to claim 2, characterized in that: The clustering of the operating condition parameters in the initial operating condition section includes: The operating parameters in the initial operating condition section are used as sample features to form a feature matrix X. The dimension of the feature matrix X is m×n, where m is the number of samples and n is the number of features. Set the value range of cluster number K to ,in Indicates rounding down; traverse the cluster number K value and calculate the silhouette coefficient sequence and DBI index sequence; Calculate the comprehensive metric index sequence according to the silhouette coefficient sequence and the DBI index sequence, and take the cluster number corresponding to the maximum value in the comprehensive metric index sequence as the optimal cluster number N1; Fix the optimal cluster number N1, use the K-Means algorithm to cluster the feature matrix X, obtain n1 subdivided operating conditions and the cluster center of each subdivided operating condition, and form the cluster center vector of the subdivided operating condition; where N1=n1.

4. The gas turbine combustor vibration consistency assessment method based on multi-modal fusion according to claim 1, characterized in that: The construction of the preliminary feature vector based on Mahalanobis distance includes: Calculate the Mahalanobis distance of the vibration signal of each combustion chamber relative to the mean of the sub-data set to form a vibration Mahalanobis distance vector ,in, represents the vibration Mahalanobis distance vector of combustion chamber j in the i-th operating condition, i represents the operating condition number, and j represents the combustion chamber number; Calculate the Mahalanobis distance of the temperature signal of each combustion chamber relative to the mean of the sub-dataset to form the temperature Mahalanobis distance vector ;in, represents the temperature Mahalanobis distance vector of combustion chamber j in the i-th operating condition; The vibration Mahalanobis distance vector and the temperature Mahalanobis distance vector are combined into a preliminary feature vector according to the combustion chamber sequence. , , Indicates the Combustion chamber in each operating section The initial feature vector of .

5. The gas turbine combustor vibration consistency assessment method based on multi-modal fusion according to claim 4 is characterized in that: The construction of a multimodal feature vector based on mutual information weighted fusion includes: For each combustion chamber, calculate the mutual information matrix of the vibration signal and the temperature signal ; represents the mutual information matrix between the vibration signal and the temperature signal in the i-th operating condition; Mutual Information Matrix Perform normalization to obtain the fusion weight matrix ; represents the fusion weight matrix of the i-th operating condition; According to the fusion weight matrix Weighted fusion of the i-th working condition segment The vibration Mahalanobis distance vector of the combustion chamber and temperature Mahalanobis distance vector , construct multimodal fusion features ; Indicates the i-th operating condition segment Multimodal fusion features of each combustion chamber; The multimodal fusion features of each combustion chamber Composition Multimodal feature vectors of operating conditions .

6. The gas turbine combustor vibration consistency assessment method based on multi-modal fusion according to claim 1, characterized in that: The deep feature extraction model includes a convolutional neural network and an autoencoder; The deep features of the multimodal feature vector extraction include: The multimodal feature vector is input into the convolutional neural network to extract the high-dimensional feature map; the high-dimensional feature map output by the convolutional neural network is input into the autoencoder to learn the global compression low-dimensional features; The global compressed low-dimensional features output by the autoencoder are used as the deep features of the multimodal feature vector.

7. The gas turbine combustor vibration consistency assessment method based on multi-modal fusion according to claim 1, characterized in that: The health state levels of the combustion chamber vibration include normal, warning, and abnormal. Calculating the cluster centers of each health state level to form a multi-level health state level criterion includes: Respectively calculating the coordinates Cn, Cw, and Ca of the cluster centers of the normal, warning, and abnormal health state levels. Calculating the distance from each sample in the normal state cluster to the coordinate Cn of the normal state cluster center, and denoting the maximum distance as d1. Calculating the distance from each sample in the warning state cluster to the coordinate Cn of the normal state cluster center, and denoting the minimum distance as d2; Let δ1 = 1 / 2(d1 + d2) as the boundary threshold of the normal state. Calculating the distance from each sample in the warning state cluster to the coordinate Cw of the warning state cluster center, and denoting the maximum distance as d3. Calculating the distance from each sample in the abnormal state cluster to the coordinate Cw of the warning state cluster center, and denoting the minimum distance as d4; Let δ2 = 1 / 2(d3 + d4) as the boundary threshold of the warning state. Calculating the distance from each sample in the abnormal state cluster to the coordinate Ca of the abnormal state cluster center, and denoting the maximum distance as d5, and let δ3 = 2×d5 as the boundary threshold of the abnormal state.

8. The gas turbine combustor vibration consistency assessment method based on multi-modal fusion according to claim 7, characterized in that: Determining the health state level of the current combustion chamber vibration according to the real-time state signal and the multi-level health state level criterion includes: Determining the working condition section to which the current operating state belongs, marked as the current working condition section; calculating the real-time multimodal feature vector of the real-time state signal. Inputting the real-time multimodal feature vector into the deep feature extraction model to obtain the real-time deep features; calculating the Euclidean distances Dn, Dw, and Da between the real-time deep features and the coordinates Cn, Cw, and Ca of the normal, warning, and abnormal health state level cluster centers, and determining the health state level of the current combustion chamber vibration.

9. The gas turbine combustor vibration consistency assessment method based on multi-modal fusion according to claim 8, characterized in that: Determining the working condition section to which the current operating state belongs includes: Extract the real-time operating parameters under the current operating state; perform dimensionless processing on the real-time operating parameters and convert them into the real-time operating vector x (c) ; Calculate the Euclidean distance between the real-time working condition vector and the cluster center vector of the subdivided working condition segment; (c) Divide into the subdivided operating condition segment corresponding to the nearest cluster center as the operating condition segment of the current operating state; Calculating the real-time multimodal feature vector of the real-time state signal includes: Calculate the Mahalanobis distance of the real-time vibration signal relative to the mean of the sub-data set of the operating section to form a real-time preliminary feature vector; according to the fusion weight matrix and real-time preliminary feature vectors, and construct a real-time multimodal feature vector based on mutual information weighted fusion.

10. The gas turbine combustor vibration consistency assessment method based on multi-modal fusion according to claim 8, characterized in that: Determining the health state level of the current combustion chamber vibration includes: Determining the health state level of the current combustion chamber vibration includes: If Dn ≤ δ1 and Dw ≤ Da, it is determined to be in the normal state. If δ1 < Dn ≤ δ2 and Dw ≤ Da, it is determined to be in the warning state. If δ2 < Dn ≤ δ3, it is determined to be in the abnormal state.

Images