Gas turbine exhaust gas dispersion anomaly early warning method and device

Abstract

The application discloses a kind of gas turbine exhaust dispersion abnormality early warning method and device, belong to gas turbine state monitoring and fault early warning technical field.Through obtaining old unit operation data and new unit real-time monitoring data, simulation fault data is generated based on component mechanism model and is fused with real fault data, and source domain data set and target domain data set are constructed;Limit learning machine is used to map two-domain data to unified feature space, align by minimizing joint distribution difference, and obtain reconstructed features that eliminate individual differences;Use the labeled source domain reconstruction feature to train the fault classification model, determine the adaptive early warning threshold based on the output probability distribution statistical characteristics of normal samples;Target domain reconstruction feature is input into model to obtain abnormal score, and early warning is triggered when the score continuously exceeds threshold value.The application solves the problem that individual differences of new and old units make it difficult to set adaptive threshold, and realizes accurate and reliable abnormal early warning.

Description

Technical Field

[0001] This invention relates to the field of gas turbine condition monitoring and fault early warning technology, and in particular to a method and device for early warning of abnormal exhaust dispersion in gas turbines. Background Technology

[0002] Gas turbines are core equipment for power peak shaving, industrial drive, and marine propulsion, and their operational reliability directly affects system safety and economic benefits. Exhaust gas dispersion (ETD), as a key indicator characterizing the uniformity of the combustion chamber outlet temperature field, can keenly reflect the early health degradation of combustion chamber components. Anomaly early warning based on ETD is a core means to achieve condition-based maintenance and predictive maintenance.

[0003] Currently, data-driven fault warning methods have become a research hotspot, but when faced with a fleet consisting of multiple units, they have revealed fundamental shortcomings such as isolated modeling dimensions, difficulty in reusing historical knowledge, and limited warning capabilities due to sample size.

[0004] Existing methods mainly fall into two categories. The first category focuses on time-series data modeling for individual generating units, employing methods such as Long Short-Term Memory (LSTM) networks to predict trends or reconstruct anomalies in ETD (Effective Time Defects). This method is effective when there is sufficient data for a single unit, but it treats each unit as an information silo, severing the commonalities in fault mechanisms among units of the same model. Valuable fault knowledge accumulated by older units cannot be reused across units, and new units, due to the lack of fault samples, cannot train effective models and can only rely on conservative fixed thresholds, frequently falling into the dilemma of false negatives or false alarms.

[0005] The second approach attempts to eliminate data distribution differences between units by introducing transfer learning. Researchers utilize domain adaptation techniques to minimize the distribution differences between the source domain (older units) and the target domain (newer units). However, existing transfer learning methods have significant limitations: first, they are not optimized for the slow degradation characteristics of ETD (Extended Time To Difference), making it difficult to capture its temporal drift; second, they often employ complex deep networks, which face challenges such as limited computing resources and poor interpretability in field deployment; and third, most only align the overall distribution (marginal distribution), ignoring the fine-grained distribution (conditional distribution) within the same fault category, resulting in insufficient fault identification capabilities.

[0006] Furthermore, the source domain itself may suffer from sparse fault samples, limiting the effectiveness of direct transfer. Traditional methods either rely on expert experience to set rules or simply oversample the data; the former struggles to cover complex fault modes, while the latter is prone to overfitting.

[0007] In summary, existing technologies suffer from low early warning accuracy in the initial stages of new unit commissioning due to isolated modeling, lack of optimization for ETD in the transfer framework, difficulty in model deployment, and sparse source domain samples, making it difficult to achieve fleet-level intelligent operation and maintenance. There is an urgent need for an early warning framework that can uniformly characterize common degradation mechanisms and individual distribution differences, accurately transfer knowledge from old units to new units, and supplement sparse samples through mechanistic models. Summary of the Invention

[0008] The present invention aims to at least partially solve one of the technical problems in the related art.

[0009] To address this issue, this invention discloses a method for early warning of abnormal exhaust dispersion in gas turbines. By acquiring operating data from older units and real-time monitoring data from newer units, simulated fault data is generated based on a gas turbine component mechanism model and fused with real fault data to generate enhanced samples. The operating data from older units containing these enhanced samples is constructed as a source domain dataset, and the real-time monitoring data from newer units is constructed as a target domain dataset. An extreme learning machine is used to map the source and target domain datasets to a unified feature space, and feature distribution alignment is performed by minimizing the joint distribution difference between the two domains, resulting in source and target domain reconstruction features that eliminate individual differences between units. A fault classification model is trained using the source domain reconstruction features with health status labels, and an adaptive warning threshold is determined based on the statistical characteristics of the output probability distribution of normal samples in the source domain under the fault classification model. The target domain reconstruction features are input into the fault classification model to obtain an anomaly score. When the anomaly score continuously exceeds the adaptive warning threshold, an abnormal exhaust dispersion warning is triggered. This invention solves the problem of difficulty in adaptively setting warning thresholds due to individual differences between older and newer units, achieving accurate knowledge transfer and reliable early warning across units.

[0010] Another objective of this invention is to provide an early warning device for abnormal exhaust dispersion of gas turbines.

[0011] To achieve the above objectives, the present invention provides a method for early warning of abnormal exhaust dispersion in gas turbines, comprising:

[0012] The system acquires operating data of old units and real-time monitoring data of new units, generates simulated fault data based on the mechanism model of gas turbine components, and merges the simulated fault data with real fault data to generate enhanced samples. The operating data of old units containing enhanced samples is used to construct the source domain dataset, and the real-time monitoring data of new units is used to construct the target domain dataset. Extreme learning machine is used to map the source domain dataset and the target domain dataset to a unified feature space. The mapped features are then aligned by minimizing the joint distribution difference between the two domains, resulting in source domain reconstruction features and target domain reconstruction features that eliminate individual differences among the units. A fault classification model is trained using source domain reconstruction features with health status labels, and an adaptive warning threshold is determined based on the statistical characteristics of the output probability distribution of normal samples in the source domain under the fault classification model. The reconstructed features of the target domain are input into the fault classification model to obtain an anomaly score. When the anomaly score continues to exceed the adaptive warning threshold, an abnormal exhaust dispersion warning is triggered.

[0013] In one embodiment of the present invention, generating an enhanced sample includes: Based on the thermodynamic principles of gas turbine gas path, a component-level model is constructed. By adjusting the health parameters of the combustion chamber, the fault response of exhaust dispersion is simulated, and an initial simulation fault sample set is generated. Using scarce real fault samples and an initial set of simulated fault samples as input, a conditional deep convolutional generative adversarial network (GAN) consisting of a generator and a discriminator is trained. The generator receives a random noise vector and a fault category label, and concatenates or channels-wise before inputting them into the generator network. After multiple layers of deconvolution and activation function mapping, it generates realistic fault samples corresponding to the fault categories. The discriminator receives the samples and their corresponding category labels, determines whether the input samples are real or generated, and verifies the consistency of the categories. Through adversarial training, the generator learns the noise distribution of real data and the physical laws of simulated data, resulting in high-fidelity augmented samples. The enhanced samples were added to the old unit operation data to obtain a complete source domain dataset.

[0014] In one embodiment of the present invention, generating an initial simulated fault sample set includes: Mathematical models of key components of gas turbines are established based on thermodynamic and kinetic principles, and the range of variation of health parameters for combustion chamber carbon deposits, flame tube cracks, and fuel nozzle blockage is set. Multiple sets of simulation calculations were performed within the range of health parameter variations to obtain time-series response data of exhaust dispersion and related measurement point parameters under different fault modes; The time-series response data is discretized according to a preset sampling frequency to form an initial simulated fault sample containing fault category labels, which is then summarized to form an initial simulated fault sample set.

[0015] In one embodiment of the present invention, mapping to a unified feature space and performing distribution alignment includes: Randomly generate connection weights and biases between the input layer and hidden layer of the extreme learning machine, calculate the output matrices of the source domain dataset and the target domain dataset in the hidden layer, and concatenate the two output matrices into the total hidden layer output matrix. A marginal distribution maximum mean difference matrix is ​​constructed to measure the difference in marginal distributions, and a pseudo-label iterative optimization mechanism is introduced to construct a conditional distribution maximum mean difference matrix for each health status category. Construct an objective function that includes the output weight norm, source domain data reconstruction error, and joint distribution difference, and solve the optimal output weight matrix by minimizing the objective function; The source domain reconstruction features and target domain reconstruction features that eliminate distribution differences are calculated using the optimal output weight matrix.

[0016] In one embodiment of the present invention, constructing the conditional distribution maximum mean difference matrix includes: In the first iteration, only the maximum mean difference matrix of the marginal distribution is used to obtain the preliminary mapping matrix and calculate the preliminary pseudo-labels of the target domain data; starting from the second iteration, for each health status category, source domain samples belonging to the corresponding category and target domain samples with corresponding pseudo-labels are selected. Based on the selected samples, construct the maximum mean difference matrix of the conditional distribution for the corresponding category only to align the conditional distributions of the source and target domains under the same fault category; and obtain the joint distribution difference metric matrix by weighted summation of the maximum mean difference matrix of the conditional distributions of all categories and the maximum mean difference matrix of the marginal distributions.

[0017] In one embodiment of the present invention, training a fault classification model and determining an adaptive early warning threshold includes: The source domain reconstructed features that eliminate distribution differences are used as input, and the corresponding health status labels are used as targets. A ridge regression classifier is trained to learn the mapping relationship from a unified feature space to failure modes, and the classifier output weight matrix is ​​obtained. Input the normal sample set in the source domain into the trained ridge regression classifier to obtain a set of abnormal scores under normal conditions; A confidence level is set, and the quantile of the corresponding confidence level in the abnormal score set is defined as the adaptive warning threshold, so that only a preset proportion of normal samples have abnormal scores exceeding the adaptive warning threshold.

[0018] In one embodiment of the present invention, triggering an abnormal exhaust dispersion warning includes: The system monitors the abnormal score sequence of new unit samples in real time and determines whether the abnormal score at the current moment is greater than or equal to the adaptive warning threshold. If it exceeds the threshold, the system counts the cumulative number of times the abnormal score exceeds the threshold within a preset time window. When the cumulative number reaches a preset number threshold, the system determines that it is an abnormal state and triggers a warning. Otherwise, the system continues to monitor the data at the next moment.

[0019] In one embodiment of the present invention, it further includes: As new units accumulate reliable data confirmed by maintenance personnel, this reliable data is added to the source domain dataset to achieve dynamic expansion of the source domain. The two steps of feature mapping and distribution alignment and fault classification model training are re-executed according to the preset cycle. The mapping matrix of the extreme learning machine and the output weights of the fault classification model are updated so that the model can continuously adapt to the personalized characteristics of the new unit.

[0020] In one embodiment of the present invention, constructing a source domain dataset and a target domain dataset includes: The source domain is defined as older units of the same model that have been in operation for a long time and have accumulated operational data including normal operation and known precursory data of various combustor faults. The corresponding dataset is represented as follows:

[0021] in, It is a d-dimensional feature vector containing key measurement parameters such as exhaust dispersion ETD, compressor outlet pressure, fuel flow rate, and exhaust temperature; The target domain is defined as newly commissioned units with scarce historical data, and the corresponding dataset is represented as follows:

[0022] The initial phase only included unlabeled real-time monitoring data; The z-score normalization method was used to standardize all the original measurement point data in the source and target domains to eliminate the influence of dimensions. For the exhaust dispersion index, its statistical characteristics within the sliding window were calculated to enhance the characterization ability of degradation trend, and the source domain feature matrix and the target domain feature matrix were formed respectively.

[0023] This invention also proposes a gas turbine exhaust dispersion abnormality early warning device, comprising: The data augmentation module is used to acquire operating data of old units and real-time monitoring data of new units, generate simulated fault data based on the mechanism model of gas turbine components, and fuse simulated fault data with real fault data to generate augmented samples. The operating data of old units containing augmented samples is used to construct the source domain dataset, and the real-time monitoring data of new units is used to construct the target domain dataset. The feature alignment module is used to map the source domain dataset and the target domain dataset to a unified feature space using the extreme learning machine, and to perform distribution alignment of the mapped features by minimizing the joint distribution difference of the two domain data, so as to obtain source domain reconstruction features and target domain reconstruction features that eliminate the individual differences of the unit. The threshold calibration module is used to train a fault classification model using source domain reconstruction features with health status labels, and to determine an adaptive warning threshold based on the statistical characteristics of the output probability distribution of normal samples in the source domain under the fault classification model. The early warning module is used to input the reconstructed features of the target domain into the fault classification model to obtain an anomaly score. When the anomaly score continues to exceed the adaptive early warning threshold, an abnormal exhaust dispersion early warning is triggered.

[0024] This invention discloses a method and apparatus for early warning of abnormal exhaust dispersion in gas turbines. It generates simulated fault data based on a component mechanism model and fuses it with real fault data to construct a source domain dataset containing enhanced samples. An extreme learning machine is used to map the source and target domain data to a unified feature space and minimize joint distribution differences, eliminating individual differences between units. Furthermore, a fault classification model is trained based on the reconstructed features of the source domain. An adaptive early warning threshold is determined according to the statistical characteristics of the probability distribution of normal sample outputs, and an early warning is triggered when the abnormal score of the reconstructed features in the target domain continuously exceeds the threshold. This invention effectively solves the problems of existing technologies where the early warning threshold is difficult to adaptively set due to individual differences between new and old units, and historical fault knowledge cannot be reused across units. It achieves an integrated solution for the entire process from data augmentation, feature alignment, threshold calibration to abnormal early warning, significantly improving the accuracy and reliability of early warnings and enhancing the adaptability and engineering practical value of the method in fleet-level intelligent operation and maintenance scenarios.

[0025] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0026] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a gas turbine exhaust dispersion anomaly early warning method according to an embodiment of the present invention; Figure 2 This is a general flowchart of the gas turbine exhaust dispersion anomaly early warning method according to an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the differences in data distribution and alignment effect between the source and target domains according to an embodiment of the present invention. Figure 4 This is a comparison chart of the early warning accuracy of newly built units according to the method of the present invention; Figure 5 This is a diagram illustrating the adaptive threshold early warning effect based on reconstructed features according to an embodiment of the present invention. Figure 6 This is a schematic diagram of a gas turbine exhaust dispersion abnormality early warning device according to an embodiment of the present invention. Detailed Implementation

[0027] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0028] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0029] The following description, with reference to the accompanying drawings, describes a method and apparatus for early warning of abnormal exhaust dispersion in gas turbines according to an embodiment of the present invention.

[0030] like Figure 1 As shown, a method for early warning of abnormal exhaust dispersion in gas turbines includes the following steps: S1. Obtain the operating data of the old unit and the real-time monitoring data of the new unit, generate simulation fault data based on the mechanism model of the gas turbine components, and fuse the simulation fault data with the real fault data to generate enhanced samples. Construct the old unit operating data containing the enhanced samples into the source domain dataset and the real-time monitoring data of the new unit into the target domain dataset. S2 uses an extreme learning machine to map the source domain dataset and the target domain dataset to a unified feature space, and performs distribution alignment on the mapped features by minimizing the joint distribution difference of the two domain data, thus obtaining source domain reconstruction features and target domain reconstruction features that eliminate individual differences of the unit. S3, use source domain reconstruction features with health status labels to train a fault classification model, and determine the adaptive warning threshold based on the statistical characteristics of the output probability distribution of normal samples in the source domain under the fault classification model. S4, input the target domain reconstruction features into the fault classification model to obtain an anomaly score, and trigger an exhaust dispersion anomaly warning when the anomaly score continues to exceed the adaptive warning threshold.

[0031] Specifically, this invention addresses the core shortcomings of current gas turbine exhaust dispersion early warning systems, namely the scarcity of historical data and fault samples for newly built units, and the inability of traditional models to effectively reuse knowledge from older units. It proposes a new method for early warning of abnormal gas turbine exhaust dispersion. Traditional techniques suffer from isolated modeling objects and insensitivity to differences in data distribution between different units, resulting in the inability to reuse valuable historical fault knowledge and poor early warning performance for new units. To address this, this invention constructs an integrated architecture encompassing source-target domain definition, deep reconstruction of health indicators, MMD distribution alignment, ELM cross-domain classification, and adaptive threshold early warning. Based on domain adaptive theory and extreme learning machines, it achieves efficient transfer of fault knowledge from older units to new units, providing a core algorithm engine for accurate early warning during the initial commissioning phase of new units.

[0032] Furthermore, the present invention mainly involves the following steps: First, to address the potential shortage of fault samples in the source domain, a data augmentation module based on a mechanistic model and coupled with numerical simulation is introduced as a preliminary optimization step. When samples of a specific type of fault are also scarce in older units, a component-level model is constructed based on the thermodynamic principles of the gas turbine gas path. By adjusting health parameters, the fault response of the ETD (Electronic Tolerance Dispute) is simulated, generating simulated fault data. Subsequently, a conditional deep convolutional generative adversarial network is used to fuse the simulated data with real small sample data, generating a large number of high-fidelity augmented samples that integrate mechanistic characteristics and actual noise characteristics. These augmented samples are then added to the source domain dataset, fundamentally solving the problem of sparse source domain samples.

[0033] Secondly, source and target domain data construction and preprocessing are performed. Addressing the issue that traditional methods cannot directly reuse knowledge from older units, a cross-unit data definition and processing method for gas turbine ETD indicators is proposed. Older units, after data augmentation and accumulation of operational data including normal operation and known precursory data for various combustor faults, are defined as the source domain; newer units with scarce historical data and an urgent need to establish early warning models are defined as the target domain. The source domain dataset is represented as follows: ,in The multi-dimensional time-series characteristic matrix of the source unit includes ETD and its related parameters, such as exhaust temperature, combustion chamber pressure, and fuel flow rate; These are the corresponding health status labels. The target domain dataset is represented as follows: In the initial stage, only unlabeled real-time monitoring data was available. This method clarifies the data composition of the two domains and the task objectives of transfer learning, laying a data foundation for subsequent distribution alignment and knowledge transfer.

[0034] Next, cross-domain feature reconstruction and distribution alignment based on Extreme Learning Machine (ELM) are performed. To address the inconsistency in data distribution between new and old generating units caused by installation deviations and differences in measurement points, a joint distribution-adaptive ELM feature reconstruction method is proposed. Its core lies in leveraging the powerful mapping capability of a single hidden-layer feedforward network to learn a feature transformation function, making the transformed source domain data and target domain data as close as possible in distribution, while preserving the key attributes and structure of the source domain data.

[0035] To achieve this goal, this invention introduces Maximum Mean Difference (MMD) as a metric for distribution discrepancy. MMD is defined as the distance between the means of data in two domains after mapping the original data to the regenerating kernel Hilbert space. By minimizing MMD, the distribution discrepancy between the two domains is effectively reduced. This not only aligns marginal distributions but also iteratively aligns conditional distributions by assigning pseudo-labels to the target domain data, achieving Joint Distribution Fitting (JDTELM). This maps data from different units to a unified feature space, eliminating distribution shifts caused by individual differences.

[0036] Furthermore, in a unified feature space, a ridge regression classifier is trained using labeled source domain reconstructed features to construct a mapping rule from the unified feature space to fault modes. Real-time data from new generating units is mapped to the unified space in the same way and input into the classifier to obtain the fault prediction probability at the current moment. To achieve accurate early warning, a fixed threshold is abandoned, and a statistically-based adaptive threshold method is adopted. Based on the predicted probability distribution of various normally operating samples in the source domain, the quantile corresponding to the confidence level is set as the early warning threshold. When the fault prediction probability of a new generating unit sample continuously exceeds this threshold, an anomaly warning is triggered.

[0037] This invention constructs an end-to-end cross-unit knowledge transfer early warning framework through the synergy of the above-mentioned steps. Verification has shown that, compared to LSTM models using only limited data from new units and ELM models without domain adaptation, the method of this invention significantly improves key indicators such as early warning accuracy, false alarm rate, and early fault warning time, providing a reliable technical tool for lean operation and maintenance of gas turbine fleets.

[0038] Furthermore, such as Figure 2 As shown, the specific implementation process of the method in this embodiment of the invention revolves around data augmentation, data preprocessing, cross-domain feature reconstruction and distribution alignment, migration early warning model construction, and adaptive threshold early warning. Combined with the stringent requirements for accuracy and practicality in gas turbine fleet operation and maintenance, it forms a technical solution that can be implemented in engineering. The specific steps are as follows: Step 1: To address the extreme scarcity of samples for certain specific fault types in the source domain, a data augmentation module based on a mechanistic model and mathematical simulation is introduced to improve the robustness of the transfer model. First, mathematical models of key components of the gas turbine are established based on thermodynamic and kinetic principles. By adjusting the health parameters of the combustion chamber, the ETD response under different fault modes is simulated, generating an initial set of simulated fault samples. Then, data augmentation is performed using a conditional deep convolutional generative adversarial network (CDCGAN). This augments the scarce real-world fault samples. With simulation samples As input, a conditional deep convolutional generative adversarial network (GAN) is trained. This model consists of a generator and a discriminator. The generator takes a noise vector and a fault category label as input and outputs realistic fault samples; the discriminator judges the realism and category of the input samples. Through adversarial training, the generator learns the noise distribution of real data and the physical laws of simulated data, generating a large number of high-fidelity, diverse augmented samples. Finally, the enhanced samples Supplement to source domain dataset In the middle, a complete A migration early warning model with better ability to identify fault types.

[0039] Step 2 addresses the issue of differences in data distribution between new and old units and the inability of traditional methods to directly reuse historical knowledge. It constructs source and target domain datasets and performs data cleaning, normalization, and feature engineering on the data to lay a data foundation for subsequent transfer learning.

[0040] Step 2.1: Define the source and target domains. Define older generating units of the same model, with a long operating history and a wealth of accumulated operational and fault data, as the source domain; its dataset is denoted as... ,in It is a d-dimensional feature vector containing key measurement parameters such as exhaust dispersion ETD, compressor outlet pressure, fuel flow rate, and exhaust temperature; This is tagged with corresponding health status labels such as normal, combustion chamber carbon buildup, and flame tube cracks. Newly commissioned units with limited historical data are defined as the target domain, and their dataset is denoted as [database name missing]. The initial phase only included unlabeled real-time monitoring data.

[0041] Step 2.2: Outlier removal and missing value imputation are performed on all original measurement point data. The z-score normalization method is used to standardize the source and target domain data respectively to eliminate the influence of dimensions. For the ETD index, its statistical characteristics within the sliding window are calculated to enhance its ability to represent degradation trends, constructing an enhanced feature set that better reflects the degradation process. The source domain feature matrix is ​​denoted as... The feature matrix of the target domain is denoted as .

[0042] Step 3: To address the issue of inconsistent data distribution between new and old units, construct a Joint Distribution Adaptive Extreme Learning Machine (JDTELM) model. This model learns a nonlinear mapping from the original feature space to a unified feature space, mapping the data from the source and target domains to the unified feature space while minimizing the distribution differences between the two domains.

[0043] Step 3.1: Randomly initialize the ELM hidden layer parameters. According to ELM theory, randomly generate the connection weights between the input layer and the hidden layer. and bias ,in , This represents the number of hidden layer nodes. For the input sample... Its hidden layer output vector is: (1) in The activation function is used. The hidden layer output matrices of the source and target domain data are calculated separately. and ,in The Behavior , The Behavior Concatenate the two matrices to form .

[0044] Step 3.2: Construct the marginal distribution MMD matrix Based on the definition of MMD, construct an MMD matrix to measure the differences in marginal distributions. Its element definition is as follows: (2) in This is the index of the spliced ​​sample.

[0045] Step 3.3: Iteratively construct the MMD matrix to align conditional distributions. To align conditional distributions within the same fault category, pseudo-label iterative optimization is introduced. In the first iteration, only... A preliminary mapping matrix is ​​obtained, and preliminary pseudo-labels for the target domain data are calculated. Starting from the second iteration, for each category... Construct the MMD matrix for each category. To align conditional distributions within the same category, Definition and Similar, but only for categories. The sample is used for calculation.

[0046] Step 3.4: Solve the JDTELM optimization problem. Construct the objective function, which aims to minimize the norm of the output weights, the reconstruction error of the source domain data, and the joint distribution difference. Its expression is: (3) in For the extended input matrix, It is an all-zero matrix, used to constrain the target domain from participating in the reconstruction error term; This is the penalty matrix, used to constrain the reconstruction error of the source domain data; This is a regularization parameter that balances the distribution alignment term with model complexity. Total number of health status categories; Represents the trace of a matrix; The output weight matrix to be solved is obtained by... Taking the derivative and setting it to zero yields the analytical solution. When the number of hidden layer nodes... Not greater than the total number of samples At that time, output weights The calculation formula is: (4) in It is an identity matrix.

[0047] Step 3.5: Calculate the reconstructed cross-domain features. Utilize the solved optimal output weights. Calculate the aligned features of the source and target domains: (5) at this time and It lies in a unified feature space that eliminates distributional differences.

[0048] Step S4: In a unified feature space, a ridge regression classifier is trained using the reconstructed features from the labeled source domain, and an adaptive early warning threshold is constructed for the ETD data of the new unit to achieve real-time monitoring and early warning of abnormal states of the new unit.

[0049] Step 4.1: Train the ridge regression classifier. As input, As the objective, a ridge regression classifier is trained, and its output weights are calculated using the following formula: (6) in The regularization coefficient is... This is the identity matrix. The classifier learned the mapping from a unified feature space to failure modes. It achieves a mapping from a unified feature space to fault modes, for any input features Its fault prediction probability vector is .

[0050] Step 4.2: Online monitoring and feature mapping of new units. For each new real-time sample of the new unit... First, the JDTELM model from step 3 is used to map it to a unified feature space, resulting in... Then Input the data into the classifier to obtain the probability vector of each fault type. Take the maximum probability value. As an anomaly score at the current moment.

[0051] Step 4.3: Construct an adaptive warning threshold. This is done using the aligned source domain normal sample set obtained in Step 2.5. The data is then input into a classifier to obtain a set of anomaly scores under normal conditions. Set a confidence level. The warning threshold Defined as of Quantiles: That is, only Normal samples will have abnormal scores exceeding this threshold, ensuring that the false alarm rate is controllable.

[0052] Step 4.4: Real-time monitoring of abnormal scoring sequences in new unit samples Combinatorial logic is used to improve the robustness of the early warning system. Specifically, during the early warning process, if the current new unit has an abnormal score... Furthermore, if the number of times the abnormal score exceeds the threshold reaches four or more within the past 10 minutes, it is judged as an abnormal state, triggering an alert. This combined logic can effectively avoid false alarms caused by single-point noise, while ensuring timely response to continuous anomalies.

[0053] Step 5: Model Iteration and Incremental Update Optimization. After the new unit has been running for a period of time, it accumulates a certain amount of operational data. This new data can be used to incrementally update the model. First, add some reliable data accumulated by the new unit and confirmed by maintenance personnel to the source domain dataset to achieve dynamic expansion of the source domain. Periodically (e.g., quarterly), re-execute steps 3 and 4 to update the JDTELM mapping matrix. Ridge Regression Classifier Update the model to continuously adapt it to the unique characteristics of new units.

[0054] This invention provides a complete implementation plan encompassing data augmentation, data construction, cross-domain feature reconstruction, migration early warning model construction, and adaptive threshold setting. It addresses the source domain sample sparsity problem through data augmentation via a combination of digital and analog models. Key technologies such as JDTELM joint distribution adaptation, iterative pseudo-label optimization, and adaptive early warning based on quantile thresholds eliminate data distribution shifts between old and new units caused by installation deviations and measurement point differences. This enables efficient reuse and accurate early warning of exhaust dispersion fault knowledge accumulated in older units for new units. The resulting structured early warning report can be seamlessly integrated into existing asset health management platforms, providing quantitative and interpretable technical support for accurate early warning in the initial commissioning phase of new units, risk-based differentiated maintenance decisions, and overall fleet reliability improvement. This completes the technical chain from multi-source data input from gas turbines to cross-unit ETD anomaly early warning and knowledge reuse, providing a reliable technical solution for gas turbine power generation groups to achieve fleet-level intelligent operation and maintenance.

[0055] Furthermore, to fully verify the technical effectiveness and engineering applicability of the present invention, a multi-source data fusion test environment for gas turbine fleets was constructed and systematic verification was carried out. The test data covers three years of historical operating data from the same type of gas turbine units of a power generation group, including key measurement parameters such as exhaust gas dispersion (ETD), compressor outlet pressure, fuel flow rate, and exhaust temperature. The sampling frequency is once per hour, with a total of 12 features. The data includes 5 old units with an operating life of over 10 years and 3 newly commissioned units. Addressing the scarcity of fault-related samples in the source domain, step 1, data augmentation, was first performed: 570 simulated samples were generated annually based on the mechanistic model and input together with the real scarce samples into CDCGAN to generate high-fidelity augmented samples. After augmentation, the total number of source domain samples covers simulated samples and fault samples accumulated from old units, covering four health states: normal, combustion chamber carbon buildup, flame tube cracks, and fuel nozzle blockage, totaling over 12,000 tagged samples. New units only include normal operating data and a small amount of unlabeled real-time monitoring data. Data from the first two years were used as the source domain training set, and data from the third year were used as the source domain validation set. Data from the first year after the new unit was put into operation was used as the target domain test set. Domain experts labeled actual fault events to verify the early warning effect. The experiment verified the technical effectiveness based on core indicators such as early warning accuracy, false alarm rate, fault early warning time, cross-domain generalization ability, and computational efficiency. The following is a detailed explanation with reference to the attached figures.

[0056] like Figure 2 The diagram illustrates the cross-unit feature reconstruction and transfer learning process based on JDTELM. In the test, the number of hidden layer nodes in the JDTELM model was set to 200, and the regularization parameters γ=10 and λ=10. - ³. After iterative optimization, the source domain reconstruction features Reconstructing features with the target domain The high similarity within a unified space allows ridge regression classifiers trained on source domain labels to be directly applied to the target domain. Experiments show that the JDTELM model has low training time and low time to map a single new sample, meeting the real-time requirements of engineering applications.

[0057] like Figure 3The figures show the differences in data distribution between the source and target domains and the alignment effect of JDTELM. The left figure (a) shows the ETD feature distribution of the old and new units after data augmentation in the original feature space. It is evident that there is a significant distribution shift due to factors such as installation deviations and measurement point characteristics. The feature distribution of the same type of fault samples differs significantly across different units, resulting in poor direct migration performance. The right figure (b) shows the distribution of source and target domain data in a unified feature space after JDTELM feature reconstruction. At this point, the edge distributions of the two domains and the conditional distributions of the same fault category are well aligned, with the scatter points almost completely overlapping. Quantitative calculations show that the MMD distance between the source and target domains in the original space is 0.87, while after JDTELM alignment, it decreases to 0.12, a reduction of 86.2%. This demonstrates that the method effectively eliminates cross-unit data distribution differences, laying a feature foundation for subsequent accurate early warning.

[0058] like Figure 4 As shown, the early warning accuracy of various technical solutions on new generating units is clearly compared. Four methods were used for comparison: Method 1 is an LSTM time-series prediction model using only a small amount of data from new generating units; Method 2 is an ELM classification model without domain adaptation; Method 3 is a DTELM method that only aligns edge distributions; and Method 4 is the JDTELM method of this invention. On a test set containing four health states (normal + three faults), the average accuracy and standard deviation of 10 independent experiments were: LSTM 68.2% ± 3.5%, ELM 72.5% ± 3.2%, DTELM 83.7% ± 2.8%, and JDTELM 91.3% ± 2.1%. The method of this invention improves accuracy by 7.6 percentage points compared to DTELM, 18.8 percentage points compared to ELM, and 23.1 percentage points compared to LSTM. Meanwhile, the false alarm rate decreased from 31.5% for LSTM to 6.8% for JDTELM, and the early fault warning time increased from an average of 3.2 days for LSTM to 8.5 days, providing maintenance personnel with a valuable processing window. This demonstrates that the present invention significantly improves the accuracy, timeliness, and reliability of cross-unit fault warnings by jointly aligning edge distribution and conditional distribution.

[0059] like Figure 5 The figure shows the effect of adaptive threshold early warning based on reconstructed features. The blue solid line in the figure represents the abnormal score of a certain gas turbine in the new unit over 200 consecutive days. The curve shows the adaptive warning threshold (95th percentile, θ=0.31) calculated based on the statistical characteristics of normal samples in the source domain (red dashed line), and the traditional fixed threshold (θ=0.6) (green dotted line). According to test data, the unit exhibited early abnormal signs (slight carbon buildup in the combustion chamber) between days 80 and 120, which developed into a persistent abnormality (increased carbon buildup) after day 120. As shown in the graph, the traditional fixed threshold caused multiple false alarms (green circles) during the normal phase (0-80 days) due to its low threshold, while it resulted in missed alarms (orange squares) during the persistent abnormal phase due to its high threshold. The adaptive threshold of this invention is dynamically set based on the distribution of normal samples in the source domain, resulting in no false alarms during the normal phase. On day 125, it accurately triggered a warning (red triangle) after meeting the combined logic of "exceeding the threshold 4 times within 10 minutes," successfully capturing the early abnormal signs. Statistically, the adaptive threshold method had a false alarm rate of only 6.8% and a missed alarm rate of 4.2% on the test set, while the fixed threshold method had a false alarm rate as high as 28.3% and a missed alarm rate of 21.5%. This method effectively balances early warning sensitivity and robustness through probabilistic threshold setting and combinational logic judgment.

[0060] In summary, the application scenarios of this invention, through key technologies such as JDTELM joint distribution adaptation, iterative pseudo-label optimization, ridge regression classifier, and adaptive threshold warning, achieve a 91.3% accuracy rate for early warning of newly built units, a 23.1 percentage point improvement over the traditional LSTM model; the false alarm rate is reduced to 6.8%, a 21.5 percentage point decrease compared to the fixed threshold method; the early warning time for faults is extended to 8.5 days, a 5.3-day improvement compared to traditional methods; the MMD distance is reduced by 86.2% after cross-domain distribution alignment; the single-sample inference time is less than 1 millisecond, and the model size is less than 10MB after lightweighting. Experimental data fully verify the significant advantages of this invention in solving the problem of small-sample early warning for newly built units, realizing the efficient reuse of fault knowledge from old units to new units, and providing accurate, reliable, and interpretable technical support for intelligent operation and maintenance at the gas turbine fleet level.

[0061] To implement the method of the above embodiments, such as Figure 6 As shown, the present invention proposes a gas turbine exhaust dispersion abnormality early warning device 10, comprising: The data augmentation module 100 is used to acquire the operating data of the old unit and the real-time monitoring data of the new unit, generate simulated fault data based on the mechanism model of the gas turbine components, and fuse the simulated fault data with the real fault data to generate augmented samples. The operating data of the old unit containing the augmented samples is used to construct the source domain dataset, and the real-time monitoring data of the new unit is used to construct the target domain dataset.

[0062] The feature alignment module 200 is used to map the source domain dataset and the target domain dataset to a unified feature space using an extreme learning machine, and to perform distribution alignment on the mapped features by minimizing the joint distribution difference between the two domains, so as to obtain source domain reconstruction features and target domain reconstruction features that eliminate individual differences of the unit.

[0063] The threshold calibration module 300 is used to train a fault classification model using source domain reconstruction features with health status labels, and to determine an adaptive warning threshold based on the statistical characteristics of the output probability distribution of normal samples in the source domain under the fault classification model.

[0064] The early warning module 400 is used to input the target domain reconstruction features into the fault classification model to obtain an anomaly score. When the anomaly score continues to exceed the adaptive early warning threshold, an abnormal exhaust dispersion early warning is triggered.

[0065] The device of this invention effectively solves the problems of existing technologies, such as the inability to reuse historical fault knowledge across units due to individual differences between new and old units, and the difficulty in adaptively setting early warning thresholds. It realizes integrated early warning throughout the entire process from data augmentation, feature alignment, threshold calibration to anomaly early warning, significantly improving the accuracy and reliability of fleet-level exhaust dispersion anomaly early warning, and enhancing the adaptability and practical value of engineering deployment.

[0066] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0067] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for early warning of abnormal exhaust dispersion in gas turbines, characterized in that, Includes the following steps: The system acquires operating data of old units and real-time monitoring data of new units, generates simulated fault data based on the mechanism model of gas turbine components, and merges the simulated fault data with real fault data to generate enhanced samples. The operating data of old units containing enhanced samples is used to construct the source domain dataset, and the real-time monitoring data of new units is used to construct the target domain dataset. Extreme learning machine is used to map the source domain dataset and the target domain dataset to a unified feature space. The mapped features are then aligned by minimizing the joint distribution difference between the two domains, resulting in source domain reconstruction features and target domain reconstruction features that eliminate individual differences among the units. A fault classification model is trained using source domain reconstruction features with health status labels, and an adaptive warning threshold is determined based on the statistical characteristics of the output probability distribution of normal samples in the source domain under the fault classification model. The reconstructed features of the target domain are input into the fault classification model to obtain an anomaly score. When the anomaly score continues to exceed the adaptive warning threshold, an abnormal exhaust dispersion warning is triggered.

2. The method as described in claim 1, characterized in that, Generate enhanced samples, including: Based on the thermodynamic principles of gas turbine gas path, a component-level model is constructed. By adjusting the health parameters of the combustion chamber, the fault response of exhaust dispersion is simulated, and an initial simulation fault sample set is generated. Using scarce real fault samples and an initial set of simulated fault samples as input, a conditional deep convolutional generative adversarial network (GAN) consisting of a generator and a discriminator is trained. The generator receives a random noise vector and a fault category label, and concatenates or channels-wise before inputting them into the generator network. After multiple layers of deconvolution and activation function mapping, it generates realistic fault samples corresponding to the fault categories. The discriminator receives the samples and their corresponding category labels, determines whether the input samples are real or generated, and verifies the consistency of the categories. Through adversarial training, the generator learns the noise distribution of real data and the physical laws of simulated data, resulting in high-fidelity augmented samples. The enhanced samples were added to the old unit operation data to obtain a complete source domain dataset.

3. The method as described in claim 2, characterized in that, Generate an initial set of simulated fault samples, including: Mathematical models of key components of gas turbines are established based on thermodynamic and kinetic principles, and the range of variation of health parameters for combustion chamber carbon deposits, flame tube cracks, and fuel nozzle blockage is set. Multiple sets of simulation calculations were performed within the range of health parameter variations to obtain time-series response data of exhaust dispersion and related measurement point parameters under different fault modes; The time-series response data is discretized according to a preset sampling frequency to form an initial simulated fault sample containing fault category labels, which is then summarized to form an initial simulated fault sample set.

4. The method as described in claim 1, characterized in that, Mapping to a unified feature space and performing distribution alignment includes: Randomly generate connection weights and biases between the input layer and hidden layer of the extreme learning machine, calculate the output matrices of the source domain dataset and the target domain dataset in the hidden layer, and concatenate the two output matrices into the total hidden layer output matrix. A marginal distribution maximum mean difference matrix is ​​constructed to measure the difference in marginal distributions, and a pseudo-label iterative optimization mechanism is introduced to construct a conditional distribution maximum mean difference matrix for each health status category. Construct an objective function that includes the output weight norm, source domain data reconstruction error, and joint distribution difference, and solve the optimal output weight matrix by minimizing the objective function; The source domain reconstruction features and target domain reconstruction features that eliminate distribution differences are calculated using the optimal output weight matrix.

5. The method as described in claim 4, characterized in that, Construct the maximum mean difference matrix of the conditional distribution, including: In the first iteration, only the maximum mean difference matrix of the marginal distribution is used to obtain the preliminary mapping matrix and calculate the preliminary pseudo-labels of the target domain data; starting from the second iteration, for each health status category, source domain samples belonging to the corresponding category and target domain samples with corresponding pseudo-labels are selected. Based on the selected samples, construct the maximum mean difference matrix of the conditional distribution for the corresponding category only to align the conditional distributions of the source and target domains under the same fault category; and obtain the joint distribution difference metric matrix by weighted summation of the maximum mean difference matrix of the conditional distributions of all categories and the maximum mean difference matrix of the marginal distributions.

6. The method as described in claim 1, characterized in that, Training a fault classification model and determining an adaptive warning threshold includes: The source domain reconstructed features that eliminate distribution differences are used as input, and the corresponding health status labels are used as targets. A ridge regression classifier is trained to learn the mapping relationship from a unified feature space to failure modes, and the classifier output weight matrix is ​​obtained. Input the normal sample set in the source domain into the trained ridge regression classifier to obtain a set of abnormal scores under normal conditions; A confidence level is set, and the quantile of the corresponding confidence level in the abnormal score set is defined as the adaptive warning threshold, so that only a preset proportion of normal samples have abnormal scores exceeding the adaptive warning threshold.

7. The method as described in claim 1, characterized in that, Triggering an abnormal exhaust dispersion warning includes: The system monitors the abnormal score sequence of new unit samples in real time and determines whether the abnormal score at the current moment is greater than or equal to the adaptive warning threshold. If it exceeds the threshold, the system counts the cumulative number of times the abnormal score exceeds the threshold within a preset time window. When the cumulative number reaches a preset number threshold, the system determines that it is an abnormal state and triggers a warning. Otherwise, the system continues to monitor the data at the next moment.

8. The method as described in claim 1, characterized in that, Also includes: As new units accumulate reliable data confirmed by maintenance personnel, this reliable data is added to the source domain dataset to achieve dynamic expansion of the source domain. The two steps of feature mapping and distribution alignment and fault classification model training are re-executed according to the preset cycle. The mapping matrix of the extreme learning machine and the output weights of the fault classification model are updated so that the model can continuously adapt to the personalized characteristics of the new unit.

9. The method as described in claim 1, characterized in that, Construct the source domain dataset and the target domain dataset, including: The source domain is defined as older units of the same model that have been in operation for a long time and have accumulated operational data including normal operation and known precursory data of various combustor faults. The corresponding dataset is represented as follows: in, It is a d-dimensional feature vector containing key measurement parameters such as exhaust dispersion ETD, compressor outlet pressure, fuel flow rate, and exhaust temperature; The target domain is defined as newly commissioned units with scarce historical data, and the corresponding dataset is represented as follows: The initial phase only included unlabeled real-time monitoring data; The z-score normalization method was used to standardize all the original measurement point data in the source and target domains to eliminate the influence of dimensions. For the exhaust dispersion index, its statistical characteristics within the sliding window were calculated to enhance the characterization ability of degradation trend, and the source domain feature matrix and the target domain feature matrix were formed respectively.

10. A gas turbine exhaust gas dispersion abnormality early warning device, characterized in that, Includes the following steps: The data augmentation module is used to acquire operating data of old units and real-time monitoring data of new units, generate simulated fault data based on the mechanism model of gas turbine components, and fuse simulated fault data with real fault data to generate augmented samples. The operating data of old units containing augmented samples is used to construct the source domain dataset, and the real-time monitoring data of new units is used to construct the target domain dataset. The feature alignment module is used to map the source domain dataset and the target domain dataset to a unified feature space using the extreme learning machine, and to perform distribution alignment of the mapped features by minimizing the joint distribution difference of the two domain data, so as to obtain source domain reconstruction features and target domain reconstruction features that eliminate the individual differences of the unit. The threshold calibration module is used to train a fault classification model using source domain reconstruction features with health status labels, and to determine an adaptive warning threshold based on the statistical characteristics of the output probability distribution of normal samples in the source domain under the fault classification model. The early warning module is used to input the reconstructed features of the target domain into the fault classification model to obtain an anomaly score. When the anomaly score continues to exceed the adaptive early warning threshold, an abnormal exhaust dispersion early warning is triggered.

Images