Gas turbine control system based on combustion oscillation real-time monitoring and active suppression

Abstract

The invention relates to the technical field of energy conservation and environmental protection, in particular to a gas turbine control system based on combustion oscillation real-time monitoring and active suppression, and the system constructs an upstream disturbance-oscillation response prediction model and a combustion mode fingerprint database double-loop decision-making architecture through multi-source heterogeneous data real-time collection and multi-dimensional intelligent analysis processing. Self-adaptive control parameters are generated and continuously optimized through feedback verification, so that the technical problems of control mismatch, suppression failure and lack of state cognition caused by the fact that a traditional fixed parameter model cannot adapt to dramatic change of physical and chemical properties of fuel are fundamentally solved, predictive suppression and knowledge self-evolution of combustion oscillation are realized, and the method is suitable for popularization and application. And the operation safety and the fuel adaptability of the gas turbine are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of energy conservation and environmental protection technology, and in particular to a gas turbine control system based on real-time monitoring and active suppression of combustion oscillations. Background Technology

[0002] With the global energy structure transitioning towards green and low-carbon technologies, the use of zero-carbon fuels such as hydrogen and ammonia in gas turbines has become an irreversible trend. Combustion oscillation is a key instability issue faced by gas turbines employing lean-burn premixed technology. Existing active suppression technologies achieve oscillation suppression through high-frequency pressure sensor monitoring, fixed-parameter control algorithms, and fuel modulation valve execution. However, these technologies are designed based on the combustion characteristics of natural gas. When fuel switching or blending ratio changes, drastic changes in physicochemical properties such as flame propagation speed and flammability limits cause the fixed-parameter model to fail to adapt online, resulting in control mismatch or even complete failure. Furthermore, existing systems lack the ability to recognize and diagnose fuel type and combustion state, and cannot proactively prevent unknown modal oscillations. Therefore, there is an urgent need for a next-generation intelligent suppression system capable of sensing changes in fuel characteristics online, autonomously recognizing combustion state, and dynamically adjusting control strategies.

[0003] For example, Chinese Patent Publication No. CN103528090A discloses a combustion system including a combustion device and a combustion oscillation suppression system. The combustion oscillation suppression system includes a sensor, a controller, and an airflow loudspeaker. The sensor senses a signal in the combustion device, the controller converts the signal into an execution signal, and the airflow loudspeaker receives the execution signal and generates an acoustic pulsation signal based on the execution signal to act on the combustion device to suppress combustion oscillations in the combustion system. However, this solution still suffers from control mismatch and suppression failure due to the fixed parameter model's inability to adapt to drastic changes in fuel physicochemical properties. Summary of the Invention

[0004] To address this, the present invention provides a gas turbine control system based on real-time monitoring and active suppression of combustion oscillations, in order to overcome the problems of control mismatch and suppression failure caused by the inability of fixed parameter models to adapt to drastic changes in fuel physicochemical properties under varying operating conditions in the prior art.

[0005] To achieve the above objectives, the present invention provides a gas turbine control system based on real-time monitoring and active suppression of combustion oscillations, comprising: The multi-source data acquisition module is used to acquire heterogeneous data from multiple sources. The data preprocessing module is used to perform clock synchronization and signal preprocessing on multi-source heterogeneous data to obtain fused time-series data; The feature acquisition module is used to construct the upstream disturbance-oscillation response prediction model and input the fused time series data into the upstream disturbance-oscillation response prediction model to obtain feedforward prediction features; The feature matching module is used to construct the combustion mode fingerprint database and to input the feedforward prediction features into the combustion mode fingerprint database for pattern matching to obtain control parameters. The parameter execution module is used to control the high-speed fuel modulation valve according to the control parameters; The execution feedback module is used to acquire maintenance feedback data, judge the suppression effectiveness based on the maintenance feedback data, fine-tune the combustion mode fingerprint library based on uncertainty perception based on the suppression effectiveness, and update the upstream disturbance-oscillation response prediction model based on the maintenance feedback data.

[0006] Furthermore, the multi-source data acquisition module acquires multi-source heterogeneous data, specifically including: The multi-source heterogeneous data includes upstream disturbance signals and combustion chamber oscillation signals; Upstream disturbance signals are collected from the upstream premixing section of the combustion chamber using an ultrasonic sensor array. These upstream disturbance signals include flow velocity pulsation, equivalence ratio pulsation, and vorticity pulsation. Pressure oscillation signals are acquired from the combustion chamber wall using a high-frequency dynamic pressure sensor array. The pressure oscillation signals include dynamic pressure amplitude, dominant frequency, and cross-channel phase difference. The heat release rate signal is acquired from the head of the combustion chamber using an optical flame luminescence intensity sensor. The heat release rate signal includes the chemiluminescence intensity of OH* free radicals and the phase difference between the heat release rate and the pressure.

[0007] Furthermore, the data preprocessing module performs clock synchronization and signal preprocessing on the multi-source heterogeneous data, specifically including: The IEEE 1588 precision clock protocol is used to align the timestamps of multi-source heterogeneous data to obtain time-synchronized data. The time synchronization data is bandpass filtered using a hardware filtering circuit to obtain filtered data. The passband of the bandpass filter is 10-5000Hz. The zero-mean normalization method is used to standardize the amplitude of the filtered data to obtain standardized data. Standardized data is output as fused time-series data.

[0008] Furthermore, the feature acquisition module constructs an upstream disturbance-oscillation response prediction model, specifically including: Oscillation feature prediction processing is performed on historical fused time series data to determine the equivalent oscillation driving features in the fused time series data. The extracted equivalent oscillation driving features are then used to learn the oscillation response prediction model to form an upstream disturbance-oscillation response prediction model with three-dimensional oscillation prediction feature output. The extraction of the equivalent oscillation driving features involves extracting the set of oscillation response prediction model calculations from the multidimensional matrix data of upstream disturbance data after hydrodynamic analysis, chemical kinetic analysis, and thermo-acoustic coupling analysis. This includes: extracting hydrodynamic features, chemical kinetic features, and thermo-acoustic coupling features from the equivalent oscillation driving features, as well as extracting oscillation prediction quality information from the equivalent oscillation driving features. The extraction of fluid dynamics features from the equivalent oscillation-driven characteristics is achieved by the output set of three lines composed of fluid dynamics analysis layers, including, from left to right: the first line is a velocity fluctuation analysis layer plus a vorticity fluctuation analysis layer; the second line is a turbulence intensity analysis layer plus a coherent structure analysis layer; and the third line is a recirculation region oscillation analysis layer plus a shear layer stability analysis layer. The extraction of chemical dynamics features from the equivalent oscillation-driven characteristics is achieved by the output set of three lines composed of chemical dynamics analysis layers, including, from left to right: the first line is an equivalence ratio fluctuation analysis layer plus a reaction rate analysis layer; the second line is a free radical concentration analysis layer plus a chemical timescale analysis layer; and the third line is a flame surface wrinkling analysis layer plus an autoignition delay analysis layer. The extraction of thermoacoustic coupling features from the equivalent oscillation-driven characteristics is achieved by the output set of three lines composed of thermoacoustic coupling analysis layers, including, from left to right: the first line is a pressure fluctuation phase analysis layer plus a heating release rate fluctuation analysis layer; the second line is an acoustic mode analysis layer plus a flame response function analysis layer; and the third line is a coupling strength analysis layer plus a Rayleigh index analysis layer. The three-dimensional oscillation prediction feature output is implemented by the oscillation prediction feature output module. This module dynamically adjusts the contribution of each input feature through a learnable oscillation prediction generation network, and then performs oscillation prediction feature generation calculation to form a unified three-dimensional oscillation prediction feature output. The extraction of oscillation prediction quality information from the equivalent oscillation driving features is achieved through an oscillation prediction quality assessment mechanism. This mechanism has 12 parallel computing lines, which is greater than the number of model layer computing lines for extracting fluid dynamics features, chemical dynamics features, and thermoacoustic coupling features from the equivalent oscillation driving features.

[0009] Furthermore, the feature acquisition module inputs the fused time series data into the upstream disturbance-oscillation response prediction model to obtain feedforward prediction features, which include the predicted oscillation main frequency f_pred, the predicted amplitude growth trend A_trend, and the predicted mode order m_pred.

[0010] Furthermore, the feature matching module constructs a combustion mode fingerprint database, specifically including: Based on the principle of combustion oscillation, a combustion knowledge graph ontology layer is defined, and entities such as fuel type, operating load, oscillation mode, and control parameter are determined, along with their attribute relationships. These attribute relationships include causal relationships, influence relationships, and inhibition relationships. Entities and relational instances are extracted using offline experimental data and high-precision simulation data to populate the data layer and construct an initial combustion mode fingerprint database. The rule-based inference engine is embedded into the initial combustion mode fingerprint database to obtain the combustion mode fingerprint database.

[0011] Furthermore, the feature matching module inputs the feedforward prediction features into the combustion mode fingerprint database for pattern matching to obtain control parameters, specifically including: The feedforward predicted feature vector F_fusion(t) is input into the combustion modal fingerprint database, and the k-nearest neighbor algorithm is used for pattern matching. The k-nearest neighbor algorithm is set to k=5, and the scalar value davg of the average deviation between the feedforward predicted feature F_fusion(t) and the k most similar records in the combustion modal fingerprint database is obtained. The matching degree M(t) = 1 / (1+davg) is calculated based on davg, and the control parameters are output according to the matching degree, where: When M(t) > 0.85, the corresponding initial set of control parameters [Φ0, G0, f_mod0] is extracted as the control parameters for output; When M(t)≤0.85, a conservative parameter set [Φ0=π, G0=0.5] is used as the control parameter for the output.

[0012] Furthermore, the parameter execution module controls the high-speed fuel modulation valve according to the control parameters, specifically including: The reference phase compensation angle Φ0, reference gain G0, and modulation frequency f_mod0 in the control parameters are converted into valve drive pulse signals to control the high-speed fuel modulation valve.

[0013] Furthermore, the execution feedback module performs uncertainty-aware fine-tuning of the combustion mode fingerprint database, specifically including: Obtain maintenance feedback data, which includes the suppressed amplitude A_final and the prediction error rate Error_rate, and set Error_rate=|f_pred-f_actual| / f_actual; Calculate the suppression efficiency η = (A_peak - A_final) / A_peak × 100% and the prediction accuracy AP = 1 - (∑|f_pred - f_actual|) / NΔf; When η > 85% and AP > 0.8, the maintenance effectiveness is determined to be high, and the feature-parameter record is incrementally added to the combustion mode fingerprint library after being smoothed by Savitzky-Golay filtering. When 70% < η ≤ 85% and 70% < AP ≤ 80%, the maintenance effectiveness is judged to be moderate, and the feature-parameter record features of this time are stored. When η≤70%, uncertainty-aware fine-tuning is triggered, and the fingerprint database records corresponding to the high error range are updated with a learning rate of 5 times. When AP ≤ 70%, uncertainty-aware fine-tuning is triggered, and the fingerprint database records corresponding to the high error range are updated with a learning rate of 5 times.

[0014] Furthermore, the execution feedback module updates the upstream disturbance-oscillation response prediction model based on maintenance feedback data, specifically including: When Error_rate > 20%, the upstream disturbance-oscillation response prediction model is deemed invalid, and the model is updated. The model update includes: A 5x learning rate was applied during the subsequent 24 hours of online training of the upstream perturbation-oscillation response prediction model, and an elastic weight consolidation algorithm was used to prevent catastrophic forgetting, while freezing parameters in the non-error interval. The online training data was set to the actual test samples within the high error range of the past hour, and a small batch online training strategy with a batch size of 64 was adopted, with 5 epochs trained per cycle. When Error_rate≤20%, the upstream disturbance-oscillation response prediction model is deemed valid, and no model update is performed.

[0015] Compared with existing technologies, the beneficial effects of this invention are as follows: By acquiring multi-source heterogeneous data in real time, performing multi-dimensional intelligent analysis and processing, constructing a dual-loop decision architecture of an upstream disturbance-oscillation response prediction model and a combustion mode fingerprint database, and generating adaptive control parameters that are continuously optimized through feedback verification, this invention fundamentally solves the technical problems of control mismatch, suppression failure, and lack of state awareness caused by the inability of traditional fixed-parameter models to adapt to drastic changes in fuel physicochemical properties. It achieves predictive suppression and knowledge self-evolution of combustion oscillations, significantly improving the operational safety and fuel adaptability of gas turbines. The system comprehensively and accurately acquires multi-source heterogeneous data through a multi-source data acquisition module, providing a complete and reliable data foundation for intelligent control under highly time-varying conditions with zero-carbon fuels. Furthermore, the system constructs an upstream disturbance-oscillation response prediction model through a feature acquisition module, using a long short-term memory network to predict the oscillation frequency, amplitude growth trend, and spatial mode order 200ms in advance, achieving a paradigm shift from "passive response" to "feedforward intervention." Even under extreme conditions such as drastic changes in hydrogen fuel flame speed, it can accurately capture oscillation precursors. The system also utilizes special... The matching module constructs a combustion mode fingerprint database, linking the physicochemical properties, operating loads, oscillation modes, and optimal control parameters of fuels such as hydrogen / ammonia through a knowledge graph. This upgrades the architecture from a "single model" to a "queryable knowledge base," enabling the system to identify fuel switching online and dynamically invoke matching strategies, preventing fixed parameters from failing due to fuel mismatch. The system also uses a parameter execution module to directly drive a high-speed fuel modulation valve based on the reference phase compensation angle, reference gain, and modulation frequency. This converts control parameters into millisecond-level valve drive pulses, achieving active suppression through "fire-to-fire" by adjusting secondary fuel flow, ensuring precise execution of oscillation suppression commands under the acoustic time delay constraints of the combustion chamber. Furthermore, the execution feedback module intelligently judges the system's effectiveness based on prediction accuracy and suppression efficiency, fine-tuning the combustion mode fingerprint database with uncertainty perception and applying a 5x learning rate update to high-error intervals. This forms a self-learning closed-loop optimization system, allowing the system to continuously evolve its understanding of new oscillation modes of zero-carbon fuels during continuous operation. Fuel adaptability is improved from the traditional "offline recalibration" to "online self-evolution." Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the gas turbine control system based on real-time monitoring and active suppression of combustion oscillations in this embodiment. Detailed Implementation

[0017] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0018] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0019] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0020] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0021] Please see Figure 1 The diagram shown is a structural schematic of the gas turbine control system based on real-time monitoring and active suppression of combustion oscillations in this embodiment. The system includes: The multi-source data acquisition module is used to acquire heterogeneous data from multiple sources. The data preprocessing module is used to perform clock synchronization and signal preprocessing on multi-source heterogeneous data to obtain fused time-series data. The data preprocessing module is connected to the multi-source data acquisition module. The feature acquisition module is used to construct the upstream disturbance-oscillation response prediction model and input the fused time series data into the upstream disturbance-oscillation response prediction model to obtain feedforward prediction features. The feature acquisition module is connected to the data preprocessing module. The feature matching module is used to construct the combustion mode fingerprint database and to input the feedforward prediction features into the combustion mode fingerprint database for pattern matching to obtain control parameters. The feature matching module is connected to the feature acquisition module. The parameter execution module is used to control the high-speed fuel modulation valve according to the control parameters, and is connected to the feature matching module; The execution feedback module is used to acquire maintenance feedback data, judge the suppression effectiveness based on the maintenance feedback data, fine-tune the combustion mode fingerprint library based on uncertainty perception based on the suppression effectiveness, and update the upstream disturbance-oscillation response prediction model based on the maintenance feedback data. The execution feedback module is connected to the parameter execution module.

[0022] Specifically, this system is applied to the field of active suppression of combustion oscillations in gas turbines, particularly those using zero-carbon fuels such as hydrogen and ammonia. Through real-time acquisition of multi-source heterogeneous data, multi-dimensional intelligent analysis and processing, and the construction of a dual-loop decision architecture consisting of an upstream disturbance-oscillation response prediction model and a combustion mode fingerprint database, adaptive control parameters are generated and continuously optimized through feedback verification. This fundamentally solves the technical problems of control mismatch, suppression failure, and lack of state awareness caused by the inability of traditional fixed-parameter models to adapt to drastic changes in fuel physicochemical properties. It achieves predictive suppression and knowledge self-evolution of combustion oscillations, significantly improving the operational safety and fuel adaptability of gas turbines. The system comprehensively and accurately acquires multi-source heterogeneous data through a multi-source data acquisition module, providing a complete and reliable data foundation for intelligent control under highly time-varying conditions with zero-carbon fuels. Furthermore, the system constructs an upstream disturbance-oscillation response prediction model through a feature acquisition module, using a long short-term memory network to predict the oscillation frequency, amplitude growth trend, and spatial mode order 200ms in advance, achieving a paradigm shift from "passive response" to "feedforward intervention." Even under extreme conditions such as drastic changes in hydrogen fuel flame speed, it can accurately detect... The system detects early signs of oscillations. Furthermore, it constructs a combustion mode fingerprint database through a feature matching module, linking the physicochemical properties, operating loads, oscillation modes, and optimal control parameters of fuels such as hydrogen / ammonia using a knowledge graph. This upgrades the architecture from a "single model" to a "queryable knowledge base," enabling the system to identify fuel switching online and dynamically invoke matching strategies, preventing fixed parameters from failing due to fuel mismatch. Additionally, the system uses a parameter execution module to directly drive a high-speed fuel modulation valve based on the reference phase compensation angle, reference gain, and modulation frequency, converting control parameters into millisecond-level valve drive pulses. By adjusting the secondary fuel flow, the system achieves active suppression of oscillations through "fire-to-fire" action, ensuring that the oscillation suppression command is precisely executed under the acoustic time delay constraint of the combustion chamber. The system also uses an execution feedback module to intelligently judge the effectiveness of the system based on prediction accuracy and suppression efficiency, performs uncertainty perception fine-tuning on the combustion mode fingerprint database, and applies a 5x learning rate update to the high error range, forming a closed-loop optimization system with self-learning capabilities. This allows the system to continuously evolve its understanding of new oscillation modes of zero-carbon fuels during continuous operation, improving fuel adaptability from the traditional "offline recalibration" to "online self-evolution".

[0023] Specifically, the multi-source data acquisition module acquires multi-source heterogeneous data, including: The multi-source heterogeneous data includes upstream disturbance signals and combustion chamber oscillation signals; Upstream disturbance signals are collected from the upstream premixing section of the combustion chamber using an ultrasonic sensor array. These upstream disturbance signals include flow velocity pulsation, equivalence ratio pulsation, and vorticity pulsation. Pressure oscillation signals are acquired from the combustion chamber wall using a high-frequency dynamic pressure sensor array. The pressure oscillation signals include dynamic pressure amplitude, dominant frequency, and cross-channel phase difference. The heat release rate signal is acquired from the head of the combustion chamber using an optical flame luminescence intensity sensor. The heat release rate signal includes the chemiluminescence intensity of OH* free radicals and the phase difference between the heat release rate and the pressure.

[0024] Specifically, the velocity pulsation refers to the instantaneous fluctuation in the velocity of the fuel-air mixture within the upstream premixing section of the combustion chamber; the equivalence ratio pulsation refers to the instantaneous deviation in the fuel-air mixing ratio within the upstream premixing section of the combustion chamber; the vorticity pulsation refers to the instantaneous change in the rotational angular velocity of fluid particles within the upstream premixing section of the combustion chamber; the dynamic pressure amplitude refers to the peak intensity of the instantaneous deviation of the pressure oscillation within the combustion chamber from the average pressure; the dominant frequency refers to the frequency component with the highest energy proportion in the pressure oscillation power spectral density; and the interchannel phase difference refers to the phase difference at different spatial locations. The pressure signal measured by the dynamic pressure sensor has a phase angle difference at the dominant frequency. The OH* radical chemiluminescence intensity refers to the photon radiation intensity of excited-state hydroxyl radicals (OH) at the characteristic wavelength of 308nm during combustion. This intensity is proportional to the chemical reaction rate and heat release rate, and is a direct optical characterization of flame heat release pulsation. Coupled analysis with the pressure signal can quantify the thermoacoustic coupling strength. The heat release rate-pressure phase difference refers to the phase angle corresponding to the time delay between the flame heat release rate pulsation (characterized by the OH* chemiluminescence intensity) and the combustion chamber pressure oscillation at the dominant frequency f_d.

[0025] Specifically, the data preprocessing module performs clock synchronization and signal preprocessing on multi-source heterogeneous data, including: The IEEE 1588 precision clock protocol is used to align the timestamps of multi-source heterogeneous data to obtain time-synchronized data. The time synchronization data is bandpass filtered using a hardware filtering circuit to obtain filtered data. The passband of the bandpass filter is 10-5000Hz. The zero-mean normalization method is used to standardize the amplitude of the filtered data to obtain standardized data. Standardized data is output as fused time-series data.

[0026] Specifically, the IEEE 1588 precision clock protocol refers to an industrial Ethernet clock synchronization standard protocol based on a master-slave clock mechanism, which achieves sub-microsecond time synchronization through network message interaction. By timestamping data packets and compensating for time deviations, it ensures that all sensors and controllers in a multi-source heterogeneous data acquisition system are sampled synchronously under a unified time reference. The hardware filtering circuit refers to a dedicated signal processing circuit composed of operational amplifiers, resistors, capacitors, and other analog electronic components. In this embodiment, a fourth-order Butterworth topology is used to implement an active bandpass filter with an adjustable cutoff frequency under FPGA control. The zero-mean normalization method refers to a digital signal preprocessing algorithm. This algorithm eliminates sensor bias and slow drift by subtracting the mean and unifies the dimensions and dynamic range of different sensor signals by dividing by the standard deviation, making the signals of each channel comparable. This provides standardized input for subsequent feature extraction and fusion, and avoids one channel dominating the algorithm decision due to excessively large range.

[0027] Specifically, the feature acquisition module constructs an upstream disturbance-oscillation response prediction model, including: Oscillation feature prediction processing is performed on historical fused time series data to determine the equivalent oscillation driving features in the fused time series data. The extracted equivalent oscillation driving features are then used to learn the oscillation response prediction model to form an upstream disturbance-oscillation response prediction model with three-dimensional oscillation prediction feature output. The extraction of the equivalent oscillation driving features involves extracting the set of oscillation response prediction model calculations from the multidimensional matrix data of upstream disturbance data after hydrodynamic analysis, chemical kinetic analysis, and thermo-acoustic coupling analysis. This includes: extracting hydrodynamic features, chemical kinetic features, and thermo-acoustic coupling features from the equivalent oscillation driving features, as well as extracting oscillation prediction quality information from the equivalent oscillation driving features. The extraction of fluid dynamics features from the equivalent oscillation-driven characteristics is achieved by the output set of three lines composed of fluid dynamics analysis layers, including, from left to right: the first line is a velocity fluctuation analysis layer plus a vorticity fluctuation analysis layer; the second line is a turbulence intensity analysis layer plus a coherent structure analysis layer; and the third line is a recirculation region oscillation analysis layer plus a shear layer stability analysis layer. The extraction of chemical dynamics features from the equivalent oscillation-driven characteristics is achieved by the output set of three lines composed of chemical dynamics analysis layers, including, from left to right: the first line is an equivalence ratio fluctuation analysis layer plus a reaction rate analysis layer; the second line is a free radical concentration analysis layer plus a chemical timescale analysis layer; and the third line is a flame surface wrinkling analysis layer plus an autoignition delay analysis layer. The extraction of thermoacoustic coupling features from the equivalent oscillation-driven characteristics is achieved by the output set of three lines composed of thermoacoustic coupling analysis layers, including, from left to right: the first line is a pressure fluctuation phase analysis layer plus a heating release rate fluctuation analysis layer; the second line is an acoustic mode analysis layer plus a flame response function analysis layer; and the third line is a coupling strength analysis layer plus a Rayleigh index analysis layer. The three-dimensional oscillation prediction feature output is implemented by the oscillation prediction feature output module. This module dynamically adjusts the contribution of each input feature through a learnable oscillation prediction generation network, and then performs oscillation prediction feature generation calculation to form a unified three-dimensional oscillation prediction feature output. The extraction of oscillation prediction quality information from the equivalent oscillation driving features is achieved through an oscillation prediction quality assessment mechanism. This mechanism has 12 parallel computing lines, which is greater than the number of model layer computing lines for extracting fluid dynamics features, chemical dynamics features, and thermoacoustic coupling features from the equivalent oscillation driving features.

[0028] Specifically, the historical fusion time-series data refers to the fusion time-series data acquired in the past.

[0029] Specifically, the feature acquisition module inputs the fused time series data into the upstream disturbance-oscillation response prediction model to obtain feedforward prediction features, which include the predicted oscillation main frequency f_pred, the predicted amplitude growth trend A_trend, and the predicted mode order m_pred.

[0030] Specifically, the predicted oscillation dominant frequency refers to the predicted value of the dominant combustion oscillation frequency that will appear within the next 200ms time window, output by the upstream disturbance-oscillation response prediction model every 10ms. The predicted amplitude growth trend is a dimensionless scalar output by the upstream disturbance-oscillation response prediction model, which quantitatively characterizes the exponential growth rate of the combustion oscillation amplitude within the next 200ms. The predicted mode order refers to the discrete classification label output by the upstream disturbance-oscillation response prediction model, which predicts the spatial acoustic mode type of the dominant oscillation within the next 200ms.

[0031] Specifically, the feature matching module constructs a combustion mode fingerprint database, including: Based on the principle of combustion oscillation, a combustion knowledge graph ontology layer is defined, and entities such as fuel type, operating load, oscillation mode, and control parameter are determined, along with their attribute relationships. These attribute relationships include causal relationships, influence relationships, and inhibition relationships. Entities and relational instances are extracted using offline experimental data and high-precision simulation data to populate the data layer and construct an initial combustion mode fingerprint database. The rule-based inference engine is embedded into the initial combustion mode fingerprint database to obtain the combustion mode fingerprint database.

[0032] Specifically, the definition of combustion oscillation principle refers to the theoretical framework and mathematical criterion set for constructing the ontology layer of the combustion knowledge graph; the fuel type entity refers to the node type in the data layer of the combustion modal fingerprint knowledge graph, used to instantiate the physicochemical properties of different fuels; the operating condition load entity refers to the node type in the data layer of the combustion modal fingerprint knowledge graph, used to instantiate the operating boundary conditions of the gas turbine; the oscillation mode entity refers to the core event nodes in the combustion modal fingerprint knowledge graph that need to be identified and suppressed; and the control parameter entity refers to the solution in the combustion modal fingerprint knowledge graph used to suppress oscillations. In the case node, the causal relationship refers to the directed causal edge type defined in the ontology layer of the combustion mode fingerprint knowledge graph, with the direction being fuel type entity → oscillation mode entity or operating condition load entity → oscillation mode entity. The influence relationship refers to the reverse influence edge of the oscillation mode entity on the system performance index in the knowledge graph, with the direction being oscillation mode entity → equipment health entity. The inhibition relationship refers to the active intervention edge of the control parameter entity on the oscillation mode entity in the knowledge graph, with the direction being control parameter entity → oscillation mode entity. The rule-based reasoning engine refers to the logical reasoning kernel embedded in the combustion mode fingerprint database, such as "IF (fuel type = H2_100) AND (load > 80%) AND (upstream disturbance energy > threshold) THEN prediction mode = Mode_T2_HF, confidence level = 0.85".

[0033] Specifically, the feature matching module inputs the feedforward predicted features into the combustion mode fingerprint database for pattern matching to obtain control parameters, including: The feedforward predicted feature vector F_fusion(t) is input into the combustion modal fingerprint database, and the k-nearest neighbor algorithm is used for pattern matching. The k-nearest neighbor algorithm is set to k=5, and the scalar value davg of the average deviation between the feedforward predicted feature F_fusion(t) and the k most similar records in the combustion modal fingerprint database is obtained. The matching degree M(t) = 1 / (1+davg) is calculated based on davg, and the control parameters are output according to the matching degree, where: When M(t) > 0.85, the corresponding initial set of control parameters [Φ0, G0, f_mod0] is extracted as the control parameters for output; When M(t)≤0.85, a conservative parameter set [Φ0=π, G0=0.5] is used as the control parameter for the output.

[0034] Specifically, the k-nearest neighbor algorithm refers to the lazy learning classification algorithm used in the matching process of the combustion modality fingerprint database, and the calculation formula for davg is: , where di is the Mahalanobis distance between the current feature vector and the i-th nearest neighbor record.

[0035] Specifically, the parameter execution module controls the high-speed fuel modulation valve according to control parameters, including: The reference phase compensation angle Φ0, reference gain G0, and modulation frequency f_mod0 in the control parameters are converted into valve drive pulse signals to control the high-speed fuel modulation valve.

[0036] Specifically, the reference phase compensation angle refers to the reference phase angle used to compensate for the time delay of the combustion chamber thermoacoustic path, the reference gain refers to the ratio used to quantify the secondary fuel modulation amplitude relative to the main fuel flow rate, and the modulation frequency refers to the real-time frequency command that drives the high-speed fuel modulation valve.

[0037] Specifically, the execution feedback module performs uncertainty-aware fine-tuning of the combustion mode fingerprint database, including: Obtain maintenance feedback data, which includes the suppressed amplitude A_final and the prediction error rate Error_rate, and set Error_rate=|f_pred-f_actual| / f_actual; Calculate the suppression efficiency η = (A_peak - A_final) / A_peak × 100% and the prediction accuracy AP = 1 - (∑|f_pred - f_actual|) / NΔf; When η > 85% and AP > 0.8, the maintenance effectiveness is determined to be high, and the feature-parameter record is incrementally added to the combustion mode fingerprint library after being smoothed by Savitzky-Golay filtering. When 70% < η ≤ 85% and 70% < AP ≤ 80%, the maintenance effectiveness is judged to be moderate, and the feature-parameter record features of this time are stored. When η≤70%, uncertainty-aware fine-tuning is triggered, and the fingerprint database records corresponding to the high error range are updated with a learning rate of 5 times. When AP ≤ 70%, uncertainty-aware fine-tuning is triggered, and the fingerprint database records corresponding to the high error range are updated with a learning rate of 5 times.

[0038] Specifically, in this embodiment, after the combustion oscillation suppression cycle ends, the execution feedback module acquires maintenance feedback data from the time-series measurements returned by the high-frequency dynamic pressure sensor array, the optical flame intensity sensor, and the high-speed fuel modulation valve. The feature-parameter record refers to structured data generated when the suppression effectiveness is determined to be "high," used for incrementally updating the combustion mode fingerprint database. The Savitzky-Golay refers to the Savitzky-Golay filter, a digital filtering algorithm that smooths the feature-parameter record. The high error range refers to the range in the execution feedback module where the prediction error rate (Error_rate) > 1. The automatically calibrated operating condition feature subspace at 20% specifically involves aggregating the feedforward prediction features F_fusion(t) of continuously high-error samples using the DBSCAN density clustering algorithm, recalculating the mean vector μ_error and covariance radius r_error of the cluster centers, and thus deducing the boundary of the abnormal region relative to the full operating condition feature space. This is used to locate the systematic deviation source between the combustion mode fingerprint database and the upstream disturbance-oscillation response prediction model under a specific fuel-load combination, achieving targeted model correction. The application of a 5x learning rate update to the fingerprint database records corresponding to the high-error interval refers to adjusting the learning rate in uncertainty-aware fine-tuning. The parameter optimization mechanism that aggressively corrects knowledge entering the high error range specifically involves increasing the learning rate of fingerprint database record weight updates from the baseline value η_normal=0.001 to η_aggressive=0.005, recalculating the correlation strength gradient between the oscillation mode entity and the control parameter entity, and under the constraints of update amplitude (single G0 adjustment ≤ ±15%) and weight upper limit protection, thus deriving a knowledge coverage process that is 5 times faster than conservative updates. This is used to quickly eliminate prediction failures caused by model mismatch under highly time-varying conditions such as hydrogen fuel, while ensuring the controllability of risks in aggressive updates through rollback mechanisms and A / B testing.

[0039] Specifically, the execution feedback module updates the upstream disturbance-oscillation response prediction model based on maintenance feedback data, including: When Error_rate > 20%, the upstream disturbance-oscillation response prediction model is deemed invalid, and the model is updated. The model update includes: A 5x learning rate was applied during the subsequent 24 hours of online training of the upstream perturbation-oscillation response prediction model, and an elastic weight consolidation algorithm was used to prevent catastrophic forgetting, while freezing parameters in the non-error interval. The online training data was set to the actual test samples within the high error range of the past hour, and a small batch online training strategy with a batch size of 64 was adopted, with 5 epochs trained per cycle. When Error_rate≤20%, the upstream disturbance-oscillation response prediction model is deemed valid, and no model update is performed.

[0040] Specifically, the elastic weight consolidation algorithm refers to a parameter importance constraint mechanism applied when the execution feedback module updates the upstream disturbance-oscillation response prediction model online, in order to prevent the model from forgetting the knowledge of natural gas combustion it has already mastered in the process of learning new operating conditions such as hydrogen fuel. The catastrophic forgetting refers to the unexpected knowledge overwriting phenomenon caused by the new gradient completely covering the old parameter space when the weights of the upstream disturbance-oscillation response prediction model are fine-tuned using the traditional online training method.

[0041] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A gas turbine control system based on real-time monitoring and active damping of combustion oscillations, characterized by, include: The multi-source data acquisition module is used to acquire heterogeneous data from multiple sources. The data preprocessing module is used to perform clock synchronization and signal preprocessing on multi-source heterogeneous data to obtain fused time-series data; The feature acquisition module is used to construct the upstream disturbance-oscillation response prediction model and input the fused time series data into the upstream disturbance-oscillation response prediction model to obtain feedforward prediction features; The feature matching module is used to construct the combustion mode fingerprint database and to input the feedforward prediction features into the combustion mode fingerprint database for pattern matching to obtain control parameters. The parameter execution module is used to control the high-speed fuel modulation valve according to the control parameters; The execution feedback module is used to acquire maintenance feedback data, judge the suppression effectiveness based on the maintenance feedback data, fine-tune the combustion mode fingerprint library based on uncertainty perception based on the suppression effectiveness, and update the upstream disturbance-oscillation response prediction model based on the maintenance feedback data.

2. The gas turbine control system based on real-time monitoring and active suppression of combustion oscillations according to claim 1, characterized in that, The multi-source data acquisition module acquires multi-source heterogeneous data, specifically including: The multi-source heterogeneous data includes upstream disturbance signals and combustion chamber oscillation signals; Upstream disturbance signals are collected from the upstream premixing section of the combustion chamber using an ultrasonic sensor array. These upstream disturbance signals include flow velocity pulsation, equivalence ratio pulsation, and vorticity pulsation. Pressure oscillation signals are acquired from the combustion chamber wall using a high-frequency dynamic pressure sensor array. The pressure oscillation signals include dynamic pressure amplitude, dominant frequency, and cross-channel phase difference. The heat release rate signal is acquired from the head of the combustion chamber using an optical flame luminescence intensity sensor. The heat release rate signal includes the chemiluminescence intensity of OH* free radicals and the phase difference between the heat release rate and the pressure.

3. The gas turbine control system based on real-time monitoring and active suppression of combustion oscillations according to claim 2, characterized in that, The data preprocessing module performs clock synchronization and signal preprocessing on multi-source heterogeneous data, specifically including: The IEEE 1588 precision clock protocol is used to align the timestamps of multi-source heterogeneous data to obtain time-synchronized data. The time synchronization data is bandpass filtered using a hardware filtering circuit to obtain filtered data. The passband of the bandpass filter is 10-5000Hz. The zero-mean normalization method is used to standardize the amplitude of the filtered data to obtain standardized data. Standardized data is output as fused time-series data.

4. The gas turbine control system based on real-time monitoring and active suppression of combustion oscillations according to claim 3, characterized in that, The feature acquisition module constructs an upstream disturbance-oscillation response prediction model, specifically including: Oscillation feature prediction processing is performed on historical fused time series data to determine the equivalent oscillation driving features in the fused time series data. The extracted equivalent oscillation driving features are then used to learn the oscillation response prediction model to form an upstream disturbance-oscillation response prediction model with three-dimensional oscillation prediction feature output. The extraction of the equivalent oscillation driving features involves extracting the set of oscillation response prediction model calculations from the multidimensional matrix data of upstream disturbance data after hydrodynamic analysis, chemical kinetic analysis, and thermo-acoustic coupling analysis. This includes: extracting hydrodynamic features, chemical kinetic features, and thermo-acoustic coupling features from the equivalent oscillation driving features, as well as extracting oscillation prediction quality information from the equivalent oscillation driving features. The extraction of fluid dynamics features from the equivalent oscillation-driven characteristics is achieved by the output set of three lines composed of fluid dynamics analysis layers, including, from left to right: the first line is a velocity fluctuation analysis layer plus a vorticity fluctuation analysis layer; the second line is a turbulence intensity analysis layer plus a coherent structure analysis layer; and the third line is a recirculation region oscillation analysis layer plus a shear layer stability analysis layer. The extraction of chemical dynamics features from the equivalent oscillation-driven characteristics is achieved by the output set of three lines composed of chemical dynamics analysis layers, including, from left to right: the first line is an equivalence ratio fluctuation analysis layer plus a reaction rate analysis layer; the second line is a free radical concentration analysis layer plus a chemical timescale analysis layer; and the third line is a flame surface wrinkling analysis layer plus an autoignition delay analysis layer. The extraction of thermoacoustic coupling features from the equivalent oscillation-driven characteristics is achieved by the output set of three lines composed of thermoacoustic coupling analysis layers, including, from left to right: the first line is a pressure fluctuation phase analysis layer plus a heating release rate fluctuation analysis layer; the second line is an acoustic mode analysis layer plus a flame response function analysis layer; and the third line is a coupling strength analysis layer plus a Rayleigh index analysis layer. The three-dimensional oscillation prediction feature output is implemented by the oscillation prediction feature output module. This module dynamically adjusts the contribution of each input feature through a learnable oscillation prediction generation network, and then performs oscillation prediction feature generation calculation to form a unified three-dimensional oscillation prediction feature output. The extraction of oscillation prediction quality information from the equivalent oscillation driving features is achieved through an oscillation prediction quality assessment mechanism. This mechanism has 12 parallel computing lines, which is greater than the number of model layer computing lines for extracting fluid dynamics features, chemical dynamics features, and thermoacoustic coupling features from the equivalent oscillation driving features.

5. The gas turbine control system based on real-time monitoring and active suppression of combustion oscillations according to claim 4, characterized in that, The feature acquisition module inputs the fused time series data into the upstream disturbance-oscillation response prediction model to obtain feedforward prediction features, which include the predicted oscillation main frequency f_pred, the predicted amplitude growth trend A_trend, and the predicted mode order m_pred.

6. The gas turbine control system based on real-time monitoring and active suppression of combustion oscillations according to claim 5, characterized in that, The feature matching module constructs the combustion mode fingerprint database, specifically including: Based on the principle of combustion oscillation, a combustion knowledge graph ontology layer is defined, and entities such as fuel type, operating load, oscillation mode, and control parameter are determined, along with their attribute relationships. These attribute relationships include causal relationships, influence relationships, and inhibition relationships. Entities and relational instances are extracted using offline experimental data and high-precision simulation data to populate the data layer and construct an initial combustion mode fingerprint database. The rule-based inference engine is embedded into the initial combustion mode fingerprint database to obtain the combustion mode fingerprint database.

7. The gas turbine control system based on real-time monitoring and active suppression of combustion oscillations according to claim 6, characterized in that, The feature matching module inputs the feedforward predicted features into the combustion mode fingerprint database for pattern matching to obtain control parameters, specifically including: The feedforward predicted feature vector F_fusion(t) is input into the combustion modal fingerprint database, and the k-nearest neighbor algorithm is used for pattern matching. The k-nearest neighbor algorithm is set to k=5, and the scalar value davg of the average deviation between the feedforward predicted feature F_fusion(t) and the k most similar records in the combustion modal fingerprint database is obtained. The matching degree M(t) = 1 / (1+davg) is calculated based on davg, and the control parameters are output according to the matching degree, where: When M(t) > 0.85, the corresponding initial set of control parameters [Φ0, G0, f_mod0] is extracted as the control parameters for output; When M(t)≤0.85, a conservative parameter set [Φ0=π, G0=0.5] is used as the control parameter for the output.

8. The gas turbine control system based on real-time monitoring and active suppression of combustion oscillations according to claim 7, characterized in that, The parameter execution module controls the high-speed fuel modulation valve according to the control parameters, specifically including: The reference phase compensation angle Φ0, reference gain G0, and modulation frequency f_mod0 in the control parameters are converted into valve drive pulse signals to control the high-speed fuel modulation valve.

9. The gas turbine control system based on real-time monitoring and active suppression of combustion oscillations according to claim 8, characterized in that, The execution feedback module performs uncertainty-aware fine-tuning of the combustion mode fingerprint database, specifically including: Obtain maintenance feedback data, which includes the suppressed amplitude A_final and the prediction error rate Error_rate, and set Error_rate=|f_pred-f_actual| / f_actual; Calculate the suppression efficiency η = (A_peak - A_final) / A_peak × 100% and the prediction accuracy AP = 1 - (∑|f_pred - f_actual|) / NΔf; When η > 85% and AP > 0.8, the maintenance effectiveness is determined to be high, and the feature-parameter record is incrementally added to the combustion mode fingerprint library after being smoothed by Savitzky-Golay filtering. When 70% < η ≤ 85% and 70% < AP ≤ 80%, the maintenance effectiveness is judged to be moderate, and the feature-parameter record features of this time are stored. When η≤70%, uncertainty-aware fine-tuning is triggered, and the fingerprint database records corresponding to the high error range are updated with a learning rate of 5 times. When AP ≤ 70%, uncertainty-aware fine-tuning is triggered, and the fingerprint database records corresponding to the high error range are updated with a learning rate of 5 times.

10. The gas turbine control system based on real-time monitoring and active suppression of combustion oscillations according to claim 9, characterized in that, The execution feedback module updates the upstream disturbance-oscillation response prediction model based on maintenance feedback data, specifically including: When Error_rate > 20%, the upstream disturbance-oscillation response prediction model is deemed invalid, and the model is updated. The model update includes: A 5x learning rate was applied during the subsequent 24 hours of online training of the upstream perturbation-oscillation response prediction model, and an elastic weight consolidation algorithm was used to prevent catastrophic forgetting, while freezing parameters in the non-error interval. The online training data was set to the actual test samples within the high error range of the past hour, and a small batch online training strategy with a batch size of 64 was adopted, with 5 epochs trained per cycle. When Error_rate≤20%, the upstream disturbance-oscillation response prediction model is deemed valid, and no model update is performed.

Images