Intelligent early warning and control method and system for combustion instability of gas turbine

Abstract

The invention discloses an intelligent early warning and control method and system for combustion instability of a gas turbine. The method comprises the steps that a comprehensive state index is determined according to a current combustion parameter set; determining a dynamic current early warning threshold value; when the comprehensive state index exceeds the current early warning threshold value, early warning is triggered; the current combustion instability type is output; obtaining a plurality of groups of corresponding historical combustion instability category labels and historical optimization adjustment strategies; the gas turbine is controlled based on the current combustion instability category and the plurality of historical combustion instability category tags. According to the invention, through a multi-physics field data collaborative perception mechanism, the problems of early warning lag and false alarm and missing alarm caused by monitoring fragmentation are solved, through dynamically adjusting a current early warning threshold, the problem that a static model cannot adapt to variable working conditions is solved, and through a key feature focusing and executable strategy output mechanism, the real-time performance of the system is improved. The problem that the traditional technology is lack of reason tracing ability is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of industrial intelligence and process control, and in particular to a gas turbine combustion instability intelligent early warning and control method and system. BACKGROUND

[0002] As the core power equipment for modern high-efficiency power generation and industrial driving, the stable operation of the combustion system of the gas turbine is crucial. However, combustion instability phenomena such as combustion oscillation (pressure pulsation), backfire or tendency to extinction are key technical challenges that have long been faced in this field. These instability problems not only significantly reduce combustion efficiency and worsen emission levels, but also cause structural damage to expensive hot end components such as combustion chambers and turbine blades, and even cause serious accidents, directly threatening equipment safety and power plant economic benefits.

[0003] For monitoring and diagnosis of combustion instability, the existing technology mainly adopts the following ways: first, threshold alarm based on operating parameters is the most common means, by monitoring key operating parameters (such as combustion chamber pressure, temperature, fuel flow) in real time, and triggering an alarm when the parameter value exceeds the preset fixed threshold. Second, analysis based on a single physical signal, such as relying only on frequency spectrum analysis of combustion pressure pulsation signal to identify combustion oscillation. Finally, when complex or difficult-to-determine abnormalities occur, human diagnosis is highly dependent, that is, experienced operating engineers or experts make subjective judgments and disposal decisions based on comprehensive information.

[0004] However, the existing technology lacks the collaborative use of multi-source data, making it difficult to fully capture the complexity of the combustion state, resulting in delayed or false alarms. At the same time, static models cannot adapt to changing operating backgrounds and have poor robustness. In addition, the diagnosis results are mostly limited to phenomenon description, making it difficult to trace the root cause and provide executable control strategies. SUMMARY

[0005] The purpose of the present application is to provide a gas turbine intelligent early warning and combustion instability control method and system to overcome the deficiencies of low accuracy, poor adaptability and inability to provide optimization suggestions of the prior art.

[0006] To achieve the above purpose, the following technical solutions are adopted: The present application discloses a gas turbine combustion instability intelligent early warning and control method, comprising the following steps: S1. Obtain a plurality of current combustion parameters and extract features from each current combustion parameter to obtain a current state feature vector corresponding to each current combustion parameter. All current state feature vectors are denoted as a current state feature vector group. The plurality of current combustion parameters include at least two of combustion pressure pulsation signal, casing vibration signal, flame image sequence or fuel composition parameter. S2. input the current state feature vector group into a multi-modal fusion model, output a comprehensive value of the current state feature vector group through attention mechanism weighting, and record the comprehensive value as a comprehensive state index; S3. obtain a current operating background parameter group and input the operating background parameter group into a pre-trained threshold mapping model to obtain a current early warning threshold, the operating background parameter group comprising at least two of an equivalent operating time of a combustion chamber component, a fuel flow index, ambient temperature and humidity, and a unit load change rate; S4. triggering a combustion instability early warning when the comprehensive state index exceeds the current early warning threshold; S5. inputting the current state feature vector group when the combustion instability early warning is triggered into a pre-trained multi-classifier model to output a current combustion instability category; S6. constructing a historical fault case library, the historical fault case library comprising a plurality of historical records, each historical record comprising a corresponding historical combustion instability category label, a historical state feature vector group, a historical operating background parameter group, and a historical optimization adjustment strategy; S7. selecting all historical records under the historical combustion instability category label that is the same as the current combustion instability category from the historical fault case library to obtain a first historical record set; and S8. controlling the gas turbine based on the current operating background parameter group and the historical optimization adjustment strategies contained in the first historical record set.

[0007] Preferably, the pre-trained multi-classifier model in S5 is generated through the following steps: S51. extracting historical state feature vector groups and corresponding historical combustion instability category labels from all historical records in the historical fault case library; S52. training a multi-class machine learning model using the historical state feature vector groups and the historical combustion instability category labels; and S53. deploying the trained model as a pre-trained multi-classifier model.

[0008] Preferably, the pre-trained threshold mapping model in S3 is generated through the following steps: S31. collecting historical operating background parameter groups and corresponding manually labeled comprehensive state index values at critical instability moments; S32. training a regression model with the historical operating background parameter groups as input and the corresponding comprehensive state index values at the critical instability moments as output labels; and S33. deploying the trained regression model as a pre-trained threshold mapping model.

[0009] Preferably, S8 is specifically: S81. obtaining a historical operating background parameter group corresponding to each historical record from the first historical record set; S82. determining the similarity of each historical operating background parameter group to the current operating background parameter group, and screening the first historical record set according to the similarity to obtain a candidate historical record subset; S83. determining a set of key feature identifiers corresponding to the current combustion instability category by using a pre-trained feature importance analysis model; S84. selecting K historical state feature vectors from each historical state feature vector group of the candidate historical record subset according to the key feature identifiers, to obtain a historical key feature vector subgroup; S85. selecting K current state feature vectors from the current state feature vector group at the time of triggering the combustion instability warning according to the key feature identifiers, to obtain a current key feature vector subgroup; S86. sorting the candidate historical record subset according to the similarity between each historical key feature vector subgroup and the current key feature vector subgroup, and taking the historical optimization adjustment strategy in the candidate historical record at the top of the sorting as a control instruction to control the gas turbine.

[0010] Preferably, the candidate historical record subset is sorted according to the similarity between each historical key feature vector subgroup and the current key feature vector subgroup, specifically: The comprehensive similarity between the historical key feature vector subgroup of each historical record in the candidate historical record subset and the current key feature vector subgroup is calculated, and the candidate historical record subset is sorted in descending order according to the comprehensive similarity.

[0011] Preferably, the pre-trained feature importance analysis model is generated by the following steps: Extracting historical state feature vector groups corresponding to all historical records under the same historical combustion instability category label from the historical fault case library to obtain a historical vector set under each historical combustion instability category label; Using a random forest model to analyze the association strength between each historical state feature vector in each historical vector set and the cause of combustion instability; According to the association strength ranking, selecting the top K historical state feature vectors for each historical combustion instability category label, and recording as a set of key feature identifiers; Solidifying each set of key feature identifiers and its corresponding historical combustion instability category label to the feature importance analysis model.

[0012] Preferably, each historical record further includes a historical cause analysis text; Correspondingly, after S86, it further includes: outputting the historical cause analysis text in the candidate historical record at the top of the sorting.

[0013] Preferably, after S8, it further includes: S91. recording the optimization adjustment strategy actually executed by the user and the confirmed fault cause; S92. Add the current state feature vector group, the current operating background parameter group, the current combustion instability category, the executed optimization adjustment strategy and the fault cause as a new historical record to the historical fault case library; S93. Train the multi-modal fusion model, the multi-classifier model and the feature importance analysis model using the updated historical fault case library.

[0014] The application further discloses a gas turbine intelligent early warning and combustion instability control system, which comprises a processor and a memory.

[0015] Compared with the prior art, the application has the following beneficial effects: (1) The application solves the problems of early warning lag and false alarm and omission caused by monitoring fragmentation through a multi-physical field data collaborative perception mechanism, and solves the problem that a static model cannot adapt to variable working conditions by dynamically adjusting the threshold for triggering combustion instability early warning. Furthermore, the application solves the problem of lack of optimization suggestions in traditional technologies through a key feature focusing and executable strategy output mechanism; (2) Furthermore, the application breaks through the limitation of traditional phenomenon alarm by using a pre-trained multi-classifier model to identify specific instability modes (such as oscillation, backfire and flameout) through an operating background perception and historical record driven layered diagnosis architecture. The application further constructs a historical fault case library, and performs two-stage case matching (first mode screening and then key feature comparison) in combination with current operating background parameters, so as to ensure that a diagnosis result matched with the current working condition mechanism can be output under different loads and environmental conditions, and the environmental adaptability of the diagnosis system is greatly improved; (3) The application solves the problems of shallow diagnosis results and lack of root cause tracing ability in traditional diagnosis through a key feature focusing and executable strategy output mechanism. Specifically, the application extracts a key feature subset strongly associated with instability causes based on a feature importance analysis model, and focuses on mechanism similarity analysis in case matching. Finally, an optimization adjustment strategy (such as a quantitative adjustment instruction of a guide vane angle / fuel valve opening) verified by history is output, so as to provide a directly operable control basis for an operator, realize a leap from “phenomenon alarm” to “root cause disposal”, and effectively shorten the fault disposal time; (4) The application solves the problem of insufficient diagnosis capability of the system when facing new fault modes through a closed-loop feedback and model self-evolution mechanism. The confirmed fault cause and adopted strategy of the user are added to the historical fault library as a new case, and incremental training of the multi-modal fusion model, the classifier and the feature analysis model is driven, so that the system continuously learns new fault modes and optimization strategies, gradually improves the early warning and diagnosis precision, and forms a positive cycle of “data accumulation → model optimization → capability upgrading”. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some of the embodiments of the present application, and therefore should not be considered as limiting the scope. For those skilled in the art, other related drawings can also be obtained without creative labor.

[0017] Figure 1 The flow chart of the present application. DETAILED DESCRIPTION

[0018] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the following will combine the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. The components of the embodiments of the present application described and shown in the drawings can be arranged and designed in various different configurations.

[0019] Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. All other embodiments obtained by those skilled in the art based on the embodiments in the present application without creative labor are within the scope of the present application.

[0020] It should be noted that: similar numbers and letters represent similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0021] In the description of the embodiments of the present application, it should be noted that if the terms "upper", "lower", "horizontal", "inner" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, or the orientation or positional relationship when the product of the present application is usually placed, and are only for the convenience of describing the present application and simplifying the description, and therefore cannot be understood as indicating or implying that the indicated device or element must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application. In addition, the terms "first", "second" and the like are only used to distinguish the description, and cannot be understood as indicating or implying relative importance.

[0022] In addition, if the term "horizontal" appears, it does not mean that the component must be absolutely horizontal, but can be slightly inclined. For example, "horizontal" only means that its direction is relatively more horizontal than "vertical", and does not mean that the structure must be completely horizontal, but can be slightly inclined.

[0023] In the description of the embodiments of the present application, it also needs to be explained that, unless explicitly specified and limited, if the terms "arrange", "install", "connect", "connect" appear, they should be understood in a broad sense, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected, it can be mechanically connected, or it can be electrically connected, it can be directly connected, or indirectly connected through an intermediate medium, it can be the communication inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0024] The present application will be further described in detail below in combination with the drawings: Referring to Figure 1 The present application discloses an intelligent early warning and control method for combustion instability of a gas turbine, comprising the following steps: S1. obtaining a plurality of current combustion parameters and performing feature extraction on each current combustion parameter to obtain a current state feature vector corresponding to each current combustion parameter, and all current state feature vectors are denoted as a current state feature vector group, wherein the plurality of current combustion parameters include at least two of a combustion pressure pulsation signal, a casing vibration signal, a flame image sequence or a fuel composition parameter; The present application can capture early signs of combustion instability more accurately and earlier than single signal or simple parameter combination by deeply fusing multi-physical field data such as operating parameters, vibration, acoustics and flame images, and dynamically weighting by using attention mechanism.

[0025] S1 comprises the following steps: S11. obtaining a current combustion parameter group; The multi-physical field data related to the state of the combustion system of the gas turbine are synchronously collected on an industrial data platform, and the multi-physical field data at least include a combustion pressure pulsation signal, a casing vibration signal, a flame image sequence and operating parameters including fuel composition and flow. These data have different time resolutions and need to be synchronized through a high-speed collection interface.

[0026] S12. performing feature extraction on each combustion parameter in the current combustion parameter group to generate a corresponding current state feature vector, and combining all current state feature vectors into a current state feature vector group; The multi-physical field data are processed for feature extraction to generate state feature vectors for representing the state of the combustion system in different physical dimensions, for example, the processing of the flame image sequence includes extracting the spatiotemporal evolution dynamic features of the flame core area.

[0027] S2. inputting the current state feature vector group into a multi-modal fusion model, outputting a comprehensive value of the current state feature vector group through attention mechanism weighting, and denoted as a comprehensive state index.

[0028] The application dynamically weights combustion parameters under different physical field characteristics through the attention mechanism of the multi-modal fusion model. The dynamic weighting means that the weight is adjusted in real time according to the current running background state parameter, for example, the weight proportion of the pressure pulsation feature is automatically increased when the fuel composition suddenly changes. The multi-modal fusion model finally outputs a comprehensive state index that can quantitatively represent the degree of combustion instability. The index is a single continuous value (range 0~1). Further, the dynamic weighting means that the weights of the combustion parameters under each physical field characteristic (such as pressure, vibration, and flame) are not the same, and more importantly, these weights are not fixed but always change in real time and dynamically according to the current running background and data input. For example, under a certain running background, the model may automatically give a higher weight to the flame feature, and under another background, it may consider the pressure pulsation feature to be more critical. Therefore, in the application, the weights of the combustion parameters have the characteristic of changing with time.

[0029] The "comprehensive state index" is a single, continuous numerical value (for example, 0.75). The role of the multi-modal deep fusion model is to receive a high-dimensional "state feature vector" containing a plurality of feature values from multiple physical fields as input, and then output a single index value that can comprehensively and quantitatively represent the current combustion instability risk level through its internal complex fusion and calculation.

[0030] The application unifies feature extraction and attention weighting fusion of multi-source heterogeneous data such as combustion pressure, vibration, flame image, and operating parameters, generates a comprehensive state index, and overcomes the defect that single signal analysis cannot fully depict the combustion state. At the same time, the warning threshold is dynamically adjusted in combination with the operating background parameters, so that the system can still accurately capture the early instability signs under complex working conditions such as fuel change and equipment aging, significantly improving the warning sensitivity and accuracy.

[0031] S3. Obtain the current operating background parameter set and input the operating background parameter set into the pre-trained threshold mapping model to obtain the current warning threshold. The current operating background parameter set includes at least two of the equivalent running time of the combustion chamber component since the last combustion chamber component replacement, the current fuel calorific value index, the environment temperature and humidity, and the current load change rate of the unit. These parameters provide key context information for model adaptive adjustment.

[0032] In the present application, the "combustion parameter" is a high-frequency, real-time parameter signal that directly reflects the combustion state, and the "operating background parameter" is a low-frequency, slowly changing parameter that represents the operating environment, which can be pre-divided by those skilled in the art.

[0033] The pre-trained threshold mapping model in S3 is generated by the following steps: S31. Collect historical operating background parameter sets and their corresponding manually labeled comprehensive state index values of the instability critical moment; S32. Training a regression model with the historical operating background parameter set as input and the comprehensive state index value corresponding to the instability critical moment as output label; S33. Deploying the trained regression model as a pre-trained threshold mapping model.

[0034] When applied online, the threshold mapping model dynamically calculates the warning threshold according to the current operating background parameter set, for example, when the fuel calorific value index decreases by 10%, the threshold is correspondingly adjusted by 15%. When the comprehensive state index exceeds the current warning threshold, the combustion instability warning is triggered, and the comparison process is continuously performed in real time.

[0035] The attention mechanism in the multi-modal fusion model of the application can adaptively adjust the warning threshold of the comprehensive state index according to the current operating background state parameter of the gas turbine; thereby dynamically associating the warning threshold with the gas turbine operating background (equipment aging, fuel change, environmental conditions, etc.), effectively reducing the false positive rate and false negative rate under variable operating conditions, and enhancing the robustness of the warning system; further, the application can also be adjusted according to the current operating background parameter set to focus on the physical field characteristics with stronger indication of combustion instability under different operating backgrounds, which makes the generated comprehensive state index more intelligent and targeted.

[0036] Specifically, adaptive adjustment means that the current operating background state parameter is composed of multiple parameters (such as load, fuel calorific value, equipment operating hours, etc.), and different combinations of operating background parameters will produce different values at different times, and different combinations of operating background parameters with different values correspond to different and unique warning thresholds. With the continuous change of the operating background parameters of the gas turbine, the warning threshold calculated by the threshold mapping function or threshold mapping model also changes in real time and continuously, thereby dynamically adapting to the current operating condition of the unit.

[0037] S4. When the comprehensive state index exceeds the current warning threshold, triggering the combustion instability warning; The comparison between the comprehensive state index and the dynamically changing warning threshold is completed at every calculation instant. That is, the comparison is a continuous, frame-by-frame comparison process: at time t1, the system calculates the comprehensive state index S1 at that time and the dynamic warning threshold T1 at that time, and compares them; at the next time t2, the system calculates new S2 and new T2 for comparison. Although the warning threshold is changing, at any given time point, it is a definite value that can be compared with the comprehensive state index calculated at that time point. S5. Inputting the current state feature vector set when the combustion instability warning is triggered into the pre-trained multi-classifier model, and outputting the current combustion instability category; When the combustion instability early warning is triggered, the current state feature vector group generated in step S1 is comprehensively analyzed to identify a specific mode of combustion instability (referred to as a combustion instability category) including at least one of combustion oscillation, backfire or tendency to extinction.

[0038] The pre-trained multi-classifier model in S5 is generated by the following steps: S51. extracting the historical state feature vector group and the corresponding historical combustion instability category label in all historical records from the historical failure case library; The label refers to the quantity artificially labeled by a person when training a model to learn the mapping relationship between the input and output of the model, and in S51, the classification result obtained by the person pre-classifying the combustion instability of the historical record. S52. Using the historical state feature vector group and the historical combustion instability category label, a multi-class machine learning model is trained; S53. The trained model is deployed as a pre-trained multi-classifier model.

[0039] The pre-trained multi-classifier model for identifying the specific mode of combustion instability mentioned in S5 can also obtain training data from the historical failure case library: when the case library is constructed, each historical record has been labeled with a specific category label (such as "combustion oscillation", "backfire"), and these labels and the corresponding historical state feature vector group together constitute a supervised learning sample. Each historical record includes: a specific category label of combustion instability, a multi-physical field state feature vector (i.e. a historical state feature vector group), a root cause analysis text, an optimization adjustment strategy, and a running background state parameter (i.e. a historical running background parameter group) at the time of occurrence. This provides a knowledge base for implementing high-level diagnostic reasoning.

[0040] S6. Constructing a historical failure case library, the historical failure case library includes multiple historical records, and each historical record includes: a corresponding historical combustion instability category label, a historical state feature vector group, a historical running background parameter group and a historical optimization adjustment strategy; In the present application, the historical running background parameters and the current running background parameter group include the same types of parameters; correspondingly, the types and determination methods of the parameters in the historical state feature vector group and the current state feature vector group are also the same.

[0041] The present application not only can realize combustion instability alarm, but also can diagnose specific instability modes such as combustion oscillation, backfire and tendency to extinction, and trace the root cause by deep matching with historical records. The process of matching with historical records is as follows: S7. From the historical failure case library, all historical records under the same historical combustion instability category label as the current combustion instability category are filtered from the historical failure case library to obtain a first historical record set; this step solidifies expert knowledge of combustion diagnosis in the model, rules and case library, improves the overall operation and maintenance intelligent level of the power plant, and reduces the dependence on a few experts. S8. Control the gas turbine based on the current operating background parameter group and the historical optimization adjustment strategy contained in the first historical record set.

[0042] S8 specifically is: S81. From the first historical record set, obtain a historical operating background parameter group corresponding to each historical record; S82. Determine the similarity of each historical operating background parameter group to the current operating background parameter group, and filter the first historical record set according to the similarity to obtain a candidate historical record subset; The preset threshold is a value set by a person skilled in the art.

[0043] S83. Determine the key feature identifier set corresponding to the current combustion instability category using a pre-trained feature importance analysis model; The pre-trained feature importance analysis model is generated by the following steps: Extract the historical state feature vector group corresponding to all historical records under the same historical combustion instability category label from the historical failure case library to obtain a historical vector set under each historical combustion instability category label; Specifically, the present application first classifies the historical case library according to the category label; then a machine learning model (such as a random forest) is used to analyze the feature importance for each historical record set (such as all "backfire" records).

[0044] The random forest model is used to analyze the correlation strength between each feature dimension in the historical vector set and the cause of combustion instability; According to the correlation strength, the top K historical state feature vectors are selected for each historical combustion instability category label, and are recorded as a key feature identifier set; Each key feature identifier set and its corresponding historical combustion instability category label is solidified into a feature importance analysis model.

[0045] The present application determines and stores the key feature identifier set under each historical combustion instability category label. In order to directly call when performing step S83.

[0046] The key feature dimension determined by the machine learning model is pre-calculated according to the feature importance analysis of all historical records of the same specific mode in the historical failure case library. This makes the similarity calculation more focused on the features most relevant to the specific failure mode.

[0047] S84. According to the key feature identification, from each historical state feature vector group of the candidate historical record subset, K historical state feature vectors are selected to obtain a historical key feature vector subgroup; S85. According to the key feature identification, from the current state feature vector group when the combustion instability early warning is triggered, K current state feature vectors are selected to obtain a current key feature vector subgroup; According to the similarity of each historical key feature vector subgroup and the current key feature vector subgroup, the candidate historical record subset is sorted, specifically: The comprehensive similarity of the historical key feature vector subgroup and the current key feature vector subgroup of each historical record in the candidate historical record subset is calculated, and the candidate historical record subset is sorted in descending order according to the comprehensive similarity.

[0048] For each historical record, the similarity of each historical key feature vector and each current key feature vector is calculated one by one, and then these individual similarity values are fused into a single, comprehensive similarity score through weighted summation. The weight of each individual similarity value is output by the feature importance analysis model. Finally, the comprehensive similarity score is used to sort all candidate historical record sets S86. According to the similarity of each historical key feature vector subgroup and the current key feature vector subgroup, the candidate historical record subset is sorted, and the historical optimization adjustment strategy in the candidate historical record at the top of the sorting is used as a control instruction to control the gas turbine.

[0049] The historical optimization adjustment strategy includes quantitative adjustment suggestions for one or more parameters (such as guide vane angle, fuel staging valve opening) in the gas turbine control system, making it more operable.

[0050] Each historical record in the historical failure case library also includes a historical cause analysis text; Correspondingly, after S86, it also includes outputting the historical cause analysis text in the candidate historical record at the top of the sorting.

[0051] In other embodiments, the invention also outputs the root cause analysis and verified optimization adjustment strategy of the historical record most similar to the current combustion instability category as an auxiliary diagnostic suggestion for the current combustion instability through a visual configuration interface.

[0052] After S8, further comprising: S91. Recording the optimization adjustment strategy actually performed by the user and the confirmed failure cause; S92. Adding the current state feature vector group, the current operation background parameter group, the current combustion instability category, the performed optimization adjustment strategy and the failure cause as a new historical record to the historical failure case library; S93. Training the multi-modal fusion model, the multi-classifier model and the feature importance analysis model using the updated historical failure case library. Through the closed loop of user feedback, the system can continuously learn new failure modes, expand the historical failure case library and optimize the internal model, so that the diagnosis capability is continuously enhanced over time.

[0053] The optimization adjustment strategy includes a quantitative adjustment suggestion for one or more parameters (such as guide vane angle, fuel staging valve opening) in the gas turbine control system.

[0054] The user's adoption of the historical optimization adjustment strategy and the final confirmed failure cause are recorded, and this information is used to expand the historical combustion instability failure case database, and to perform online or offline incremental training on the multi-modal deep fusion model and the machine learning model in the hierarchical matching strategy. This gives the system the ability to self-learn and continuously evolve.

[0055] The application also discloses a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to realize the intelligent early warning and control method for combustion instability of a gas turbine.

[0056] The application also discloses an intelligent early warning and combustion instability control system for a gas turbine, which comprises a processor and a memory.

[0057] The following is an embodiment of the application: First stage: synchronous acquisition and feature representation of multi-physical field data On an integrated industrial data platform (or smart power plant data base), a high-speed data acquisition interface is deployed to realize synchronous acquisition of related data of the combustion system of the gas turbine.

[0058] Data sources of the current combustion parameter group: Operation parameter data: from a DCS (Distributed Control System), including but not limited to total fuel flow, fuel flow at each stage, guide vane angle (IGV), compressor outlet pressure (PCD), combustion chamber pressure, turbine inlet temperature (TIT), exhaust gas temperature (TET) and the like, and the acquisition frequency is usually 1-5 seconds.

[0059] Fuel composition data: From online analyzers or batch reports provided by fuel suppliers, key indicators such as heating value, Wobbe index, methane number, etc.

[0060] Combustion pressure pulsation signals: Collected by high-frequency dynamic pressure sensors installed on the combustion chamber wall, usually with a sampling frequency in the kHz range, used to capture the acoustic characteristics of combustion oscillations.

[0061] Casing vibration signals: Collected by acceleration sensors installed on the outer casing of the combustion chamber, usually with a sampling frequency in the kHz range, used to capture the structural vibration response caused by the combustion process.

[0062] Flame image sequence: Collected by industrial endoscopes or high-temperature cameras installed on the observation hole of the combustion chamber, to obtain dynamic video stream of the flame at a certain frame rate (such as 50~200 frames / second).

[0063] Feature extraction process for current combustion parameter group: For operating parameters: Calculate its statistical characteristics (mean, variance, kurtosis, skewness) and trend characteristics in a time window.

[0064] For pressure pulsation and vibration signals: Time domain features: Root mean square value (RMS), peak value, margin, kurtosis, etc.

[0065] Frequency domain features: Get the frequency spectrum by Fast Fourier Transform (FFT), extract the main frequency, secondary main frequency, harmonic frequency and its amplitude, total energy, etc.

[0066] Time-frequency domain features: Through Short-Time Fourier Transform (STFT) or Wavelet Transform, analyze the variation law of frequency with time.

[0067] For flame image sequence: Static features: Extract the average brightness of single frame image, color histogram, flame area, center of gravity position, morphological moment, etc.

[0068] Dynamic features (spatio-temporal evolution features): Analyze continuous multiple frames of images, calculate the flicker frequency of flame core area, swing amplitude, expansion / contraction speed, chaotic characteristics (such as Lyapunov exponent) of centroid motion trajectory, etc.

[0069] All extracted features are combined into a high-dimensional current state feature vector group, which comprehensively represents the multi-physical field state of the current time combustion system.

[0070] Second stage: Context-adaptive early warning model construction and triggering Model construction (offline training phase): Collect a large number of historical records containing normal operation and various combustion instabilities (labeled).

[0071] A multi-modal deep fusion neural network is constructed. The network can adopt a parallel structure, containing multiple input branches that process different types of features (e.g., CNN for image features, LSTM for time-series features, and MLP for scalar features) respectively.

[0072] An attention mechanism module is introduced before the outputs of the branches are fused into the backbone network. This module can learn to assign different attention (weights) to features from different physical fields under different inputs, allowing the model to automatically focus on the most critical information for judging combustion stability.

[0073] The output of the model is a continuous value, i.e., a comprehensive state indicator, ranging for example between 0 and 1, with 0 representing absolute stability and 1 representing severe instability.

[0074] When training the attention mechanism, the current operating background state parameters of the gas turbine can be used as additional conditional inputs, allowing the attention weights to consider the current operating context.

[0075] Early warning trigger (online application stage): The system calculates the current comprehensive state indicator in real time.

[0076] At the same time, the system obtains the current operating background state parameters (such as the number of operating hours since the last maintenance, fuel heat value, current load, etc.).

[0077] Through a pre-set function or small model, an adaptive early warning threshold is dynamically calculated based on the current operating background state parameters. For example, when the equipment is aging or using poor-quality fuel, the early warning threshold can be appropriately lowered to increase sensitivity; while in normal transient processes such as rapid load changes, the threshold can be appropriately raised to avoid false alarms.

[0078] When the real-time calculated comprehensive state indicator exceeds the dynamic early warning threshold for a certain period of time, the system immediately issues an early intelligent warning.

[0079] Third stage: intelligent diagnosis based on multi-field features and case-based reasoning When the warning is triggered, the system automatically enters the intelligent diagnosis process.

[0080] Instability mode identification: The current state feature vector group at the moment when the warning is triggered is input into a pre-trained multi-classifier model (such as support vector machine, decision tree forest, or a specialized neural network).

[0081] The output of the classifier model is a specific class label of the current combustion instability, such as: "high-frequency combustion oscillation", "low-frequency backfire tendency", or "lean fuel extinction tendency".

[0082] Stage 4: Hierarchical Matching Retrieval and Case Recommendation: First Layer Filtering (Context and Pattern Matching): The system uses the identified specific category label and the current operating context state parameters as primary search conditions to filter in the historical combustion instability fault case database. For example, if the current diagnosis is "high-frequency combustion oscillation", only match in records that are also labeled as "high-frequency combustion oscillation" in history. At the same time, records with similar operating contexts (such as load range, fuel type) can be prioritized for matching.

[0083] Second Layer Matching (Key Feature Sub-vector Similarity Calculation): For the candidate historical record set filtered by the first layer, the system calls a pre-computed list of key feature sub-vectors and their weights related to the current specific pattern. This list is derived from machine learning feature importance analysis on all cases of the same pattern in the case library.

[0084] The system only calculates the weighted similarity (such as weighted Euclidean distance or cosine similarity) between the current state feature vector group and the feature vectors of candidate cases on these key feature sub-vectors.

[0085] This hierarchical and focused matching on key feature sub-vectors significantly improves the accuracy and efficiency of diagnosis, as it is looking for similarities in the underlying mechanisms rather than just surface phenomena.

[0086] Diagnosis Suggestion Generation and Output: The system ranks the candidate cases according to the similarity scores calculated in the second layer matching, and recommends the Top-N (for example N=1~3) cases with the highest similarity.

[0087] In the user's monitored visual configuration interface, the recommended cases are displayed in clear cards or pop-up windows. The display content includes: Fault pattern, occurrence time, operating context of historical records.

[0088] Root cause analysis text: For example, "fuel nozzle blockage causes local fuel over-concentration" or "control system parameter misadjustment causes self-excited oscillation".

[0089] Proven historical optimization adjustment strategies: For example, "suggest checking and cleaning the No. 2 fuel nozzle" or "suggest reducing the gain of control parameter X by 5%".

[0090] (Optional) The key feature curve comparison chart of the current working condition and the recommended case can be displayed side by side to assist the operator in understanding.

[0091] Feedback and Optimization: The system provides interactive functions, allowing the operating personnel to evaluate the diagnostic results (such as "accurate" or "inaccurate"), record the actual measures taken and the final confirmed fault causes.

[0092] These high-quality feedback data are used to supplement and correct the historical fault case library, and as new training samples, the multi-modal fusion model, multi-classifier model, feature importance analysis model, etc. are periodically incrementally trained and optimized to realize continuous iteration of the system.

[0093] The above is only a preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. An intelligent early warning and control method for combustion instability of a gas turbine, characterized in that, Comprise the following steps: S1. Obtain a plurality of current combustion parameters and perform feature extraction on each current combustion parameter to obtain a current state feature vector corresponding to each current combustion parameter, and all current state feature vectors are denoted as a current state feature vector group, wherein the plurality of current combustion parameters comprise at least two of a combustion pressure pulsation signal, a casing vibration signal, a flame image sequence, or a fuel composition parameter; S2. Input the current state feature vector group into a multi-modal fusion model, and output a comprehensive value of the current state feature vector group through attention mechanism weighting, denoted as a comprehensive state index; S3. Obtain a current operating background parameter group and input the operating background parameter group into a pre-trained threshold mapping model to obtain a current warning threshold, wherein the operating background parameter group comprises at least two of an equivalent operating time of a combustion chamber component, a fuel white index, an environmental temperature and humidity, and a unit load change rate; S4. Trigger a combustion instability warning when the comprehensive state index exceeds the current warning threshold; S5. Input the current state feature vector group when the combustion instability warning is triggered into a pre-trained multi-classifier model to output a current combustion instability category; S6. Construct a historical fault case library, wherein the historical fault case library comprises a plurality of historical records, and each historical record comprises a corresponding historical combustion instability category label, a historical state feature vector group, a historical operating background parameter group, and a historical optimization adjustment strategy; S7. From the historical fault case library, select all historical records under the historical combustion instability category label that is the same as the current combustion instability category to obtain a first historical record set; S8. Control the gas turbine based on the current operating background parameter group and the historical optimization adjustment strategies contained in the first historical record set.

2. The intelligent warning and control method for combustion instability of gas turbine according to claim 1, characterized in that, The pre-trained multi-classifier model in S5 is generated by the following steps: S51. Extract the historical state feature vector group and the corresponding historical combustion instability category label from all historical records in the historical fault case library; S52. Train a multi-class machine learning model using the historical state feature vector group and the historical combustion instability category label; S53. Deploy the trained model as a pre-trained multi-classifier model.

3. The intelligent warning and control method for combustion instability of gas turbine according to claim 2, characterized in that, The pre-trained threshold mapping model in S3 is generated by the following steps: S31. Collect historical operating background parameter groups and their corresponding manually labeled comprehensive state index values at instability critical moments; S32. Train a regression model with the historical operating background parameter groups as input and the corresponding comprehensive state index values at the instability critical moments as output labels; S33. Deploy the trained regression model as a pre-trained threshold mapping model.

4. The intelligent warning and control method for combustion instability of gas turbine as claimed in claim 1, wherein, S8 is specifically: S81. From the first historical record set, obtain the historical operating background parameter group corresponding to each historical record; S82. Determine the similarity of each historical operating background parameter group to the current operating background parameter group, and filter the first historical record set according to the similarity to obtain a candidate historical record subset; S83. Using a pre-trained feature importance analysis model, determine the set of key feature identifiers corresponding to the current combustion instability category; S84. Based on the key feature identifier, select K historical state feature vectors from each historical state feature vector group of the candidate historical record subset to obtain a historical key feature vector subgroup; S85. Based on the key feature identifier, select K current state feature vectors from the current state feature vector group when the combustion instability warning is triggered to obtain the current key feature vector subgroup; S86. Based on the similarity between each historical key feature vector subgroup and the current key feature vector subgroup, the candidate historical record subsets are sorted, and the historical optimization adjustment strategy in the candidate historical record with the first place in the sort is used as a control command to control the gas turbine.

5. The intelligent warning and control method for combustion instability of gas turbine as claimed in claim 4, wherein, The candidate historical record subsets are sorted based on the similarity between each historical key feature vector subset and the current key feature vector subset, specifically as follows: Calculate the comprehensive similarity between the historical key feature vector subgroup of each historical record in the candidate historical record subset and the current key feature vector subgroup, and sort the candidate historical record subset in descending order based on the comprehensive similarity.

6. The intelligent warning and control method of combustion instability of gas turbine according to claim 5, characterized in that, The pre-trained feature importance analysis model is generated through the following steps: From the historical fault case library, extract the historical state feature vector group corresponding to all historical records under the same historical combustion instability category label to obtain the historical vector set under each historical combustion instability category label; The random forest model was used to analyze the correlation strength between each historical state feature vector and its corresponding historical combustion instability category in each historical vector set. Based on the correlation strength, the top K historical state feature vectors are selected for each historical combustion instability category label and denoted as the key feature identifier set; Each set of key feature identifiers and its corresponding historical combustion instability category label are integrated into the feature importance analysis model.

7. The intelligent warning and control method of combustion instability of gas turbine as claimed in claim 6, characterized in that, Each historical record also includes a text analyzing the historical reasons; Correspondingly, S86 also includes: outputting the historical reason analysis text of the candidate history in the first sorted list.

8. The intelligent warning and control method for combustion instability of gas turbine as claimed in claim 7, wherein, Following S8, it also includes: S91. Record the optimization and adjustment strategies actually executed by the user and the confirmed causes of the failures; S92. Add the current state feature vector group, the current operating background parameter group, the current combustion instability category, the executed optimization adjustment strategy, and the cause of the failure as new historical records to the historical failure case library; S93. Use the updated historical fault case library to train the multimodal fusion model, multi-classifier model, and feature importance analysis model.

9. An intelligent warning and control system for combustion instability in a gas turbine, comprising a processor and a memory, wherein, When the computer program instructions stored in the memory are executed by the processor, they implement the method as described in any one of claims 1-8.

Images