Gas generator set performance optimization method and system based on deep learning

Abstract

The invention relates to the technical field of gas power generation, and particularly discloses a gas generator set performance optimization method and system based on deep learning, and the method comprises the steps: collecting the initial state data flow of a set in real time after receiving a starting instruction, and introducing a deep learning algorithm to carry out the deep feature extraction of the initial state data flow of the set, according to the method, a nonlinear coupling effect among multi-dimensional state parameters in the transient starting process of a gas generator set is captured, gas starting initial condition characterization is constructed, and then the current gas starting initial condition characterization is inquired and matched with a historical successful starting case library, so that the successful starting condition characterization of the gas generator set is obtained. And the control strategy of the extracted historical case most similar to the current condition is used as a starting strategy, and the starting process of the gas generator set is guided. According to the method, the state characteristics of the gas generator set in the initial starting stage can be effectively revealed, and the new starting process is guided by fully utilizing the historical accumulated success experience, so that the intelligent control of the starting process is realized, and the starting success rate is improved.

Description

Technical Field

[0001] The present application relates to the field of gas-fired power generation technology, and more specifically, to a method and system for optimizing the performance of a gas-fired power generation unit based on deep learning. Background Art

[0002] Gas-fired generator sets, as crucial energy conversion equipment, play a core role in power generation, industrial drive systems, and other fields. Their operational performance is directly related to energy efficiency, equipment lifespan, and system reliability and safety. Especially during the startup phase, due to the complex system state, drastic changes in operating parameters, and susceptibility to various factors such as environmental conditions, gas-fired generator sets can experience problems such as prolonged startup times, increased fuel consumption, excessive component stress, and even startup failures. These long-term issues can seriously impact the overall performance and economic benefits of the unit.

[0003] Currently, mainstream gas-fired generator set performance optimization solutions focus on parameter tuning during steady-state operation. However, there remains a significant technological gap in intelligent control of transient startup processes, relying heavily on expert experience and fixed parameter curves set manually. However, the transient startup process of gas-fired generator sets is influenced by multiple factors and is highly nonlinear, time-varying, and uncertain. Traditional optimization methods based on fixed parameter curves struggle to cope with the multi-dimensional parameter coupling of gas-fired generator sets, and are prone to control strategy lag or insufficient adaptability, leading to ignition delays, combustion oscillations, and other issues. This not only reduces startup success rates but can also induce thermal stress accumulation, impacting equipment life.

[0004] Therefore, an optimized gas generator set performance optimization method and system based on deep learning is expected. Summary of the Invention

[0005] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides a method and system for optimizing the performance of a gas generator set based on deep learning, which collects the initial state data stream of the unit in real time after receiving the startup instruction, and introduces a deep learning algorithm to perform deep feature extraction on the initial state data stream of the unit, so as to capture the nonlinear coupling effect between the multi-dimensional state parameters during the transient startup of the gas generator set, and construct a characterization of the initial conditions for gas startup. Then, by querying and matching the current initial condition characterization of gas startup with the historical successful startup case library, the control strategy of the historical case that is most similar to the current situation is extracted as the startup strategy to guide the startup process of the gas generator set. This method can effectively reveal the state characteristics of the gas generator set in the initial startup stage, and make full use of the successful experience accumulated in history to guide the new startup process, thereby realizing intelligent control of the startup process and improving the startup success rate.

[0006] According to one aspect of the present application, a gas generator set performance optimization method based on deep learning is provided, which includes: After receiving the start command, the sensor network collects real-time initial state data streams, including the temperature of each temperature measurement point, the pressure of each pressure measurement point, the rotation speed, the ambient temperature, the atmospheric pressure, the humidity, and the downtime; Inputting the real-time initial state data stream into a trained deep learning feature extraction model to obtain a gas startup initial condition deep feature encoding vector; Use deep feature-based case retrieval to match the historical case that is most similar to the deep feature encoding vector of the gas startup initial condition from the historical successful startup case library; The control strategy of the most similar historical case is used as the starting strategy.

[0007] According to another aspect of the present application, a gas generator set performance optimization system based on deep learning is provided, which includes: An initial state data acquisition module is used to collect a real-time initial state data stream through a sensor network after receiving a start-up instruction. The real-time initial state data stream includes the temperature of each temperature measurement point, the pressure of each pressure measurement point, the rotational speed, the ambient temperature, the atmospheric pressure, the humidity, and the downtime duration; An initial condition feature extraction module, configured to input the real-time initial state data stream into a trained deep learning feature extraction model to obtain a gas startup initial condition deep feature encoding vector; A case retrieval and matching module is used to use a case retrieval based on deep features to match a historical case that is most similar to the deep feature encoding vector of the gas startup initial condition from a historical successful startup case library; The startup strategy generation module is used to use the control strategy of the most similar historical case as the startup strategy.

[0008] Compared with the existing technology, the gas generator set performance optimization method and system based on deep learning provided by this application, after receiving the start-up instruction, collects the initial state data stream of the unit in real time, and introduces a deep learning algorithm to perform deep feature extraction on the initial state data stream of the unit, so as to capture the nonlinear coupling effect between the multi-dimensional state parameters during the transient start-up process of the gas generator set, and construct a gas start-up initial condition representation. Then, by querying and matching the current gas start-up initial condition representation with the historical successful start-up case library, the control strategy of the historical case most similar to the current situation is extracted as the start-up strategy to guide the start-up process of the gas generator set. This method can effectively reveal the state characteristics of the gas generator set in the initial stage of startup, and make full use of the successful experience accumulated in history to guide the new startup process, thereby realizing intelligent control of the startup process and improving the startup success rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] The above and other purposes, features, and advantages of the present application will become more apparent through a more detailed description of the embodiments of the present application in conjunction with the accompanying drawings. The accompanying drawings are intended to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 This is a flowchart of a gas generator set performance optimization method based on deep learning according to an embodiment of the present application.

[0011] Figure 2 This is a data flow diagram of a gas generator set performance optimization method based on deep learning according to an embodiment of the present application.

[0012] Figure 3 This is a flowchart of sub-step S2 of the gas generator set performance optimization method based on deep learning according to an embodiment of the present application.

[0013] Figure 4 This is a flowchart of sub-step S22 of the gas generator set performance optimization method based on deep learning according to an embodiment of the present application.

[0014] Figure 5 This is a flowchart of sub-step S222 of the gas generator set performance optimization method based on deep learning according to an embodiment of the present application.

[0015] Figure 6 4 is a block diagram of a gas generator set performance optimization system based on deep learning according to an embodiment of the present application. DETAILED DESCRIPTION

[0016] As used in this application and the claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not intended to refer to the singular but may include the plural. Generally speaking, the terms "comprises" and "include" only indicate the inclusion of the steps and elements specifically identified, and these steps and elements do not constitute an exclusive list. A method or apparatus may also include other steps or elements.

[0017] Although the present application makes various references to certain modules in the system according to embodiments of the present application, any number of different modules can be used and run on the user terminal and / or server. The modules are illustrative only, and different aspects of the system and method can use different modules.

[0018] Flowcharts are used in this application to illustrate the operations performed by the systems according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, the various steps may be processed in reverse order or simultaneously, as needed. Furthermore, other operations may be added to these processes, or one or more operations may be removed from these processes.

[0019] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described herein.

[0020] It is worth noting that in this application, all actions to obtain data are carried out in compliance with the relevant data protection laws and policies of the country where they are located and with the authorization given by the owner of the corresponding device.

[0021] In response to the technical problems described in the above background technology, this application proposes a gas generator set performance optimization method based on deep learning. After receiving the startup instruction, the method collects the initial state data stream of the unit in real time, and introduces a deep learning algorithm to perform deep feature extraction on the initial state data stream of the unit to capture the nonlinear coupling effect between the multi-dimensional state parameters during the transient startup of the gas generator set, and construct a gas startup initial condition representation. Then, by querying and matching the current gas startup initial condition representation with the historical successful startup case library, the control strategy of the historical case most similar to the current situation is extracted as the startup strategy to guide the startup process of the gas generator set. This method can effectively reveal the state characteristics of the gas generator set in the initial startup stage, and make full use of the successful experience accumulated in history to guide the new startup process, thereby realizing intelligent control of the startup process and improving the startup success rate.

[0022] Figure 1 This is a flowchart of a gas generator set performance optimization method based on deep learning according to an embodiment of the present application. Figure 2 : is a data flow diagram of a gas generator set performance optimization method based on deep learning according to an embodiment of the present application. Figure 1 and Figure 2As shown, the gas generator set performance optimization method based on deep learning includes the following steps: S1, after receiving the start-up instruction, collecting real-time initial state data stream through the sensor network, the real-time initial state data stream including the temperature of each temperature measurement point, the pressure of each pressure measurement point, the rotation speed, the ambient temperature, the atmospheric pressure, the humidity and the downtime; S2, inputting the real-time initial state data stream into the trained deep learning feature extraction model to obtain the deep feature coding vector of the gas start-up initial condition; S3, using the case retrieval based on deep features to match the historical case that is most similar to the deep feature coding vector of the gas start-up initial condition from the historical successful start-up case library; S4, using the control strategy of the most similar historical case as the start-up strategy.

[0023] In the aforementioned deep learning-based gas generator set performance optimization method, step S1, after receiving a startup command, collects a real-time initial state data stream via a sensor network. This data stream includes the temperature at each temperature measurement point, the pressure at each pressure measurement point, the rotational speed, the ambient temperature, the atmospheric pressure, the humidity, and the downtime duration. It should be understood that the transient startup process of a gas generator set is influenced by a complex coupling of multiple factors, including the device's own state (e.g., internal temperature, pressure distribution, downtime duration, rotational speed) and environmental conditions (e.g., ambient temperature, atmospheric pressure, and humidity), which together determine the specific initial state of the current unit. Therefore, in order to accurately capture and quantify the actual operating environment and its own working conditions of the unit at the current startup moment, this application is based on sensor network technology. By deploying sensors in various key parts and environments of the unit, various key parameters reflecting the initial state of the unit are collected in real time, including the temperature of various temperature measurement points (such as the turbine inlet temperature, exhaust temperature and other key hotspots), the pressure of various pressure measurement points (such as compressor outlet pressure, combustion chamber pressure, etc.), the current speed of the unit, and the ambient temperature, atmospheric pressure, humidity and other environmental parameters of the environment. At the same time, combined with the recorded shutdown duration information, a comprehensive real-time initial state data stream is formed to provide basic data input for subsequent intelligent analysis and strategy formulation, ensuring that the startup strategy formulation of the gas generator set is adapted to the current actual initial state.

[0024] During implementation, to fully monitor the initial status of a gas-fired generator set, the sensor network must be deployed across multiple key system modules, including but not limited to the compressor, combustion chamber, turbine, exhaust system, and lubrication and cooling system. In the compressor, temperature and pressure measurement points must be installed at the inlet, outlet, and intermediate stages to accurately capture the thermodynamic changes in the airflow during the intake compression process. In the combustion chamber, the pressure fluctuations of the combustion gas and the temperature distribution in the flame zone must be measured to assess combustion stability. In the turbine, the focus is on the changing trends of the turbine inlet temperature, exhaust temperature, and turbine blade surface temperature, which help determine thermal stress accumulation. Furthermore, corresponding temperature and pressure sensors must be installed in the exhaust duct to monitor exhaust emissions. For the lubrication and cooling system, parameters such as lubricating oil temperature, pressure, and cooling water flow are equally important. These data reflect the thermal equilibrium state of the unit during shutdown and are crucial for determining whether the equipment is ready for startup.

[0025] In addition to the unit's internal parameters, monitoring environmental conditions is equally crucial. Ambient temperature directly affects intake air density, which in turn affects compressor output and combustion efficiency. Atmospheric pressure affects air supply, and is particularly sensitive at high altitudes. Changes in air humidity significantly impact combustion stability and nitrogen oxide production levels. Therefore, environmental monitoring sensors should be strategically placed around the unit to ensure real-time acquisition of these three key environmental parameters and integrate them into the overall state awareness system.

[0026] To ensure the reliability and accuracy of data collection, the design and selection of sensor networks must fully consider the complex working conditions of industrial sites. First, sensors should have good anti-interference capabilities and be able to operate stably in harsh environments such as high temperature, high pressure, and vibration. Second, the data acquisition frequency should be high enough to ensure that key dynamic change information is not missed during transient startup. Third, different types of sensors should maintain good time synchronization to avoid data misalignment caused by sampling time differences. In addition, considering that gas-fired generator sets often operate in locations far away from the main control room, the sensor network must also be equipped with remote communication capabilities to transmit the collected data in real time to the central control system via wired or wireless means for subsequent processing and analysis.

[0027] During the specific implementation process, when the gas generator set receives the start-up command, the sensor network immediately enters the activation state and begins to continuously collect various parameters. At this time, the system will automatically trigger a full data reading process to obtain all initial state information including various temperature measurement points and pressure measurement points, and combine the environmental parameters and downtime information provided by the environmental sensors and historical record modules to form a complete initial state data set. Since the startup process of the gas generator set is highly dynamic, a certain sliding window mechanism must be adopted during the acquisition process to cache and compare historical data within several time periods before startup to assist in identifying whether the current state is within the normal range. For example, if the value of a key temperature measurement point fluctuates abnormally in a short period of time, it may indicate that there is a potential failure risk in the equipment. The system can make an early warning or delay the startup decision based on the set threshold.

[0028] During the entire initial state data stream acquisition process, special attention must be paid to the integration and consistency of multi-source heterogeneous data. Due to the diverse sensor types and different interface protocols, the collected data may have format differences, accuracy deviations, or even missing data. To this end, the system should have a built-in data cleaning and correction module to filter and normalize the raw data. A baseline model should be established based on historical data to correct or remove outliers. Furthermore, a unified timestamp mechanism should be established to ensure that data from different sensors remains consistent across time, facilitating subsequent correlation analysis and modeling.

[0029] In the aforementioned deep learning-based gas generator set performance optimization method, step S2 involves inputting the real-time initial state data stream into a trained deep learning feature extraction model to obtain a deep feature encoding vector for the gas startup initial conditions. Specifically, due to the complex nonlinear coupling relationships between the multidimensional parameters (such as temperature, pressure, and speed) collected by sensors during gas generator set startup (e.g., the mutual influence between temperature and pressure, the coupling effect between the potential impact of humidity on combustion efficiency and downtime), traditional linear models or manual feature engineering methods struggle to effectively decouple these high-order interactions. Furthermore, factors such as sensor noise, environmental disturbances, and equipment aging can cause data distribution shifts, and directly using raw data to construct a representation of the current gas startup state is prone to introducing redundancy or noise interference. Therefore, in order to extract the deep state representation of the current gas startup initial conditions from the high-dimensional heterogeneous real-time initial state data stream, this application introduces a deep learning algorithm, which performs nonlinear mapping on the real-time initial state data stream through the trained deep learning feature extraction model, so as to utilize the powerful feature learning ability of the deep learning network, decouple the high-order nonlinear relationship in the multi-source data, and generate a low-dimensional feature representation with strong generalization ability, providing a highly discriminative state encoding for the subsequent startup strategy formulation. Figure 3Flowchart of sub-step S2 of the gas generator set performance optimization method based on deep learning according to an embodiment of the present application. Figure 3 As shown, the step S2 includes the following steps: S21, inputting the real-time initial state data stream into the trained deep learning feature extraction model to obtain the initial gas start-up initial condition deep feature coding vector; S22, performing feature expression robustness optimization on the initial gas start-up initial condition deep feature coding vector to obtain the gas start-up initial condition deep feature coding vector.

[0030] Specifically, step S21 involves inputting the real-time initial state data stream into a trained deep learning feature extraction model to obtain a deep feature encoding vector for the initial gas startup initial conditions. In one specific example of this application, the trained deep learning feature extraction model is an initial state feature extractor based on a multi-layer perceptron. Specifically, this application utilizes a multi-layer perceptron (MLP) model to construct the initial state feature extractor. Leveraging the MLP's powerful nonlinear mapping capabilities, the model performs feature abstraction on the real-time initial state data stream via a deep neural network to obtain a deep feature encoding vector for the initial gas startup initial conditions. Specifically, the MLP model, trained on a large amount of labeled historical data, effectively learns the complex nonlinear relationships between multidimensional state parameters during the gas generator set startup process and extracts key features that reflect the unit's true startup state. In practical applications, the MLP's input layer receives standardized multidimensional sensor data (i.e., the real-time initial state data stream). The hidden layers, consisting of multiple fully connected layers, use ReLU activation functions to learn the nonlinear relationships between the multidimensional parameters layer by layer. The output layer generates a fixed-dimensional deep feature encoding vector for the initial gas startup initial conditions through dimensionality reduction. Based on this, the original data stream is mapped to the latent space through the multi-layer perceptron model, which preliminarily captures the correlation and coupling characteristics between multi-dimensional parameters, thereby providing a basic state representation for the subsequent startup strategy formulation.

[0031] Specifically, the step S22 is to optimize the robustness of feature expression of the initial gas start initial condition deep feature coding vector to obtain the gas start initial condition deep feature coding vector. It should be understood that the present application takes into account the possible aliasing of sensor noise (such as the instantaneous drift of the pressure sensor) and minor modes unrelated to the start control in the real-time initial state data stream, and the control strategy of the gas turbine start-up process has a strong selective dependence on key state parameters (such as hot component temperature, combustion chamber pressure, etc.). However, the MLP model is sensitive to noise, resulting in the initial gas start initial condition deep feature coding vector still containing redundant modes or abnormal components, and direct use for case matching may lead to distortion of similarity measurement. Therefore, in order to further improve the robustness of feature expression of the initial gas start initial condition deep feature coding vector, the present application uses a joint optimization framework based on feature decomposition and distillation enhancement to perform feature decomposition, key feature distillation enhancement based on context-related topological structure, and reconstruction optimization on the initial gas start initial condition deep feature coding vector, thereby effectively filtering out noise components and enhancing key discriminant information therein to obtain the gas start initial condition deep feature coding vector. Among them, Figure 4 FIG is a flowchart of sub-step S22 of the gas generator set performance optimization method based on deep learning according to an embodiment of the present application. Figure 4 As shown, the step S22 includes the steps of: S221, performing feature decomposition based on one-dimensional convolution coding on the initial gas start initial condition deep feature coding vector to obtain a set of initial gas start initial condition deep feature local coding vectors; S222, based on the context-associated topological structure of the set of initial gas start initial condition deep feature local coding vectors, performing feature expression enhancement coding on each initial gas start initial condition deep feature local coding vector in the set of initial gas start initial condition deep feature local coding vectors to obtain a set of gas start initial condition deep feature local enhancement coding vectors; S223, performing feature expression enhancement reconstruction on the set of gas start initial condition deep feature local enhancement coding vectors to obtain the gas start initial condition deep feature coding vector.

[0032] More specifically, the step S221 is expressed as follows: in, represents the initial gas startup initial condition deep feature encoding vector, Indicates based on One-dimensional convolution operation of the convolution kernel, is the scale of the one-dimensional convolution kernel, represents the set of local encoding vectors of the initial gas startup initial condition deep features, 、 、 and They represent the first, second, and third in the set of the initial gas startup initial condition deep feature local encoding vectors. and Initial gas startup initial condition deep feature local encoding vector, The number of local encoding vectors of initial condition deep features for the initial gas startup.

[0033] Specifically, the initial gas startup initial condition deep feature encoding vector is processed using one-dimensional convolutional coding, extracting different local structural information from the complex real-time initial state data stream. The overall representation of the initial gas startup initial condition deep feature encoding vector is converted into a distributed local representation, decoupling the various patterns hidden within the initial gas startup initial condition deep feature encoding vector. This helps capture the local patterns between multidimensional state parameters during the transient startup of the gas generator set. The resulting set of initial gas startup initial condition deep feature local encoding vectors provides rich and structured local features for subsequent operations such as feature expression enhancement encoding based on context-related topological structures, helping to more accurately construct a representation of the gas startup initial condition.

[0034] Figure 5 Flowchart of sub-step S222 of the gas generator set performance optimization method based on deep learning according to an embodiment of the present application. Figure 5 As shown, the step S222 includes the steps of: S2221, calculating the feature correlation factor between any two initial gas start initial condition deep feature local coding vectors in the set of the initial gas start initial condition deep feature local coding vectors to obtain a gas start initial condition deep feature local correlation topology matrix composed of multiple feature correlation factors; S2222, inputting the gas start initial condition deep feature local correlation topology matrix into a gated mask function to obtain a gas start initial condition deep feature local fine-grained correlation mask topology matrix; S2223, based on the gas start initial condition deep feature local fine-grained correlation mask topology matrix, performing feature structure feedback modulation on each initial gas start initial condition deep feature local coding vector in the set of the initial gas start initial condition deep feature local coding vector to obtain a set of the gas start initial condition deep feature local enhancement coding vectors.

[0035] In a specific example of the present application, step S2221 is expressed as follows: in, The first one in the set of the initial gas startup initial condition deep feature local encoding vectors is represented Initial gas startup initial condition deep feature local encoding vector, represents transpose, Indicates the calculation of the 2-norm of the vector, represents the bandwidth parameter, represents the exponential function with base e, express and The characteristic correlation factor between the initial condition depth characteristics of the gas start-up is the first local correlation topological matrix. The element value at position.

[0036] Specifically, the interrelationships between the local encoding vectors of the deep features of the initial gas start-up conditions in the feature space are further quantified to capture the intrinsic correlations between multidimensional state parameters. Furthermore, the correlations and geometric relationships of the local features are explicitly modeled by constructing a local correlation topology matrix of the deep features of the initial gas start-up conditions to address the complexity of parameter coupling during gas generator set startup. Based on this, the resulting local correlation topology matrix of the deep features of the initial gas start-up conditions clearly demonstrates the correlation strength and structural relationships between the local encoding vectors of the deep features of the initial gas start-up conditions, providing a structured correlation information foundation for subsequent context-based feature expression enhancement.

[0037] In a specific example of the present application, step S2222 is expressed as follows: in, represents the gated mask weight matrix, Represents the local correlation topology matrix of the deep features of the initial conditions of gas startup, represents the gated mask bias matrix, represents the sigmoid activation function, Represents the local fine-grained correlation mask topology matrix of the deep features of the gas startup initial conditions.

[0038] Specifically, the gated mask function, an adaptive mechanism, dynamically filters and modulates the correlation strengths between features using noise or non-critical correlations in the local correlation topology matrix of the deep features of the initial gas startup conditions. This emphasizes key correlations and suppresses irrelevant or minor connections. The resulting fine-grained local correlation mask topology matrix of the deep features of the initial gas startup conditions can adjust the original correlation strengths based on the actual importance of feature correlations. This filter out noise interference while retaining strong correlations between key features, forming a fine-grained mask structure that focuses on core correlations, thereby focusing on important contextual information for subsequent feature processing.

[0039] In particular, considering that the geometric relationship distribution pattern of the local correlation topology matrix of the gas start initial condition deep feature on the implicit low-dimensional topological configuration will be nonlinear and not fully saturated, the global topological correlation distribution paradigm of the local correlation topology matrix of the gas start initial condition deep feature will be limited in the topological steady-state solution space due to the nonlinear coupling characteristics, and this will become more significant due to the polarization effect enhancement mechanism of the gated mask function, affecting the fine geometric configuration characterization efficiency of the local fine-grained correlation mask topology matrix of the gas start initial condition deep feature. Based on this, in a preferred example of the present application, the step S2223 includes: first, performing local topological structure balanced optimization on the local fine-grained correlation mask topology matrix of the gas start initial condition deep feature to obtain an optimized local fine-grained correlation mask topology matrix of the gas start initial condition deep feature.

[0040] Specifically, for the gas startup initial condition deep feature local fine-grained correlation mask topology matrix Each eigenvalue of ,Introducing the gradient vector correction factor for the local geometric mismatch compensation mechanism, thereby achieving multi-scale balance of topological relations: in, express Middle Rank Elements of the column, Represents the local fine-grained correlation mask topology matrix of the deep features in the gas startup initial condition middle The corresponding gradient vector correction factor, express Middle Rank Elements of the column, It means partial derivative.

[0041] Then, the gradient vector correction factor As an exogenous control parameter, the gas startup initial condition deep feature local fine-grained correlation mask topology matrix Statistical balance adjustment of each eigenvalue in: in, is the local fine-grained correlation mask topology matrix of the gas startup initial condition deep feature The eigenmean of all eigenvalues of , represents the dynamic gain parameter, Represents the optimization of the gas startup initial condition deep feature local fine-grained correlation mask topology matrix Rank Elements of a column.

[0042] In this way, under the action of the exogenous control parameter as a high-order gradient, the nonlinear adaptive convergence of the geometric correlation distribution under the mean field is reversely promoted (i.e., the gradient sensitivity threshold decreases), thereby compensating for the fragmentation of local topological units caused by the enhancement of correlation polarization through statistical resonance compensation under the mean field, thereby improving the substantial geometric fine-grained correlation structure expression effect of the local fine-grained correlation mask topology matrix of the deep characteristics of the initial conditions of gas startup.

[0043] Then, each initial gas start initial condition deep feature local encoding vector in the set of the initial gas start initial condition deep feature local encoding vector and the optimized gas start initial condition deep feature local fine-grained association mask topology matrix are input into the feature dense feedback enhancement unit to obtain the set of the gas start initial condition deep feature local enhancement encoding vector, which is expressed as follows: in, Represents the local fine-grained correlation mask topology matrix of the deep feature of the optimized gas startup initial condition, represents the feature reinforcement weight matrix, represents dot product, represents the matrix multiplication operation, is a nonlinear activation function, express The characteristic scale value of The first one in the set of local enhanced encoding vectors of the deep features of the initial conditions of gas startup is represented A local enhanced encoding vector of the deep features of the initial conditions of gas startup.

[0044] Specifically, by leveraging the feature association weight information carried by the optimized local fine-grained association mask topology matrix of the deep features of the initial gas start conditions, the local encoding vectors of the initial gas start conditions deep features are subjected to neighborhood association-based information fusion and reinforcement. Through the processing mechanism of the feature-intensive feedback reinforcement unit, the feature information of each local encoding vector of the initial gas start conditions deep features and its strongly associated vectors is weightedly aggregated or interactively transferred. Based on this, the generated set of local reinforced encoding vectors of the deep features of the initial gas start conditions fully absorbs the effective information of the associated features during the intensive feedback process guided by the optimized local fine-grained association mask topology matrix of the deep features of the initial gas start conditions, eliminating the ambiguity or one-sidedness that may exist in a single feature, thereby more accurately characterizing the complex state characteristics of the initial gas start conditions.

[0045] More specifically, in a specific example of the present application, step S223 includes: inputting the set of local enhanced coding vectors of the gas startup initial condition deep feature into a feature expression enhancement and reconstruction module based on the Transformer architecture to obtain the gas startup initial condition deep feature coding vector, which is expressed as follows: in, represents the set of local enhanced encoding vectors of the deep features of the initial conditions of gas startup, 、 and The first, second and third vectors in the set of local enhanced encoding vectors representing the initial condition of gas startup are Gas start initial condition deep feature local enhanced encoding vector, represents the feature reconstruction operation, 、 and represent the query matrix, key matrix and value matrix respectively, 、 and denote the query embedding matrix, key embedding matrix and value embedding matrix respectively, represents the normalized exponential function, Represents the deep feature encoding vector of the gas startup initial condition.

[0046] That is, with the help of the global information integration capability of the self-attention mechanism in the Transformer architecture, the set of locally enhanced local reinforcement coding vectors of the deep features of the gas start-up initial conditions is modeled with long-distance dependencies across features, and the criticality of each local reinforcement coding vector of the deep features of the gas start-up initial conditions from a global perspective is dynamically captured. The generated deep feature coding vector of the gas start-up initial conditions not only retains the detailed information of the parameters of each dimension, but also highlights the features and their association patterns that play a key role in the startup process through dynamic weight allocation, providing high-quality input for subsequent case retrieval based on deep features that can accurately characterize the complex state of the startup initial conditions, thereby improving the accuracy and adaptability of the control strategy for matching historical successful cases.

[0047] In the above-mentioned deep learning-based gas generator performance optimization method, step S3 uses deep feature-based case retrieval to match the historical case most similar to the deep feature encoding vector of the gas startup initial condition from a historical successful startup case library. It should be understood that the optimal control strategy for the transient startup process of a gas generator set is highly dependent on its initial state. Therefore, in order to determine the optimal control strategy that matches the current unit startup initial state, the present application pre-constructs a historical successful startup case library. The historical successful startup case library stores multiple historical cases that successfully started and performed well under different initial conditions. Each case is associated with a historical case gas startup initial condition deep feature encoding vector and a detailed startup process control strategy record. The historical case gas startup initial condition deep feature encoding vector is a representation of the gas startup initial condition obtained by performing the same deep learning feature extraction process on the initial state data stream collected during the historical startup process. During the case retrieval stage, the present application calculates the cosine similarity between the historical case gas startup initial condition deep feature encoding vector associated with each historical case in the historical successful startup case library and the gas startup initial condition deep feature encoding vector corresponding to the current unit to be started, using this as a measure of similarity between the cases. The value range of cosine similarity is [-1,1]. The closer the value is to 1, the more consistent the directions between the two vectors are, that is, the higher the similarity. Based on the calculated cosine similarity, this application uses the historical case corresponding to the maximum cosine similarity as the optimal reference for the current unit startup control. In this way, by utilizing the directional consistency of deep features to capture the coordinated change pattern of multi-dimensional state parameters, even if there are differences in the absolute values of some parameters, essentially similar operating states can still be accurately identified to ensure the physical adaptability of the control strategy.

[0048] In the aforementioned deep learning-based gas generator set performance optimization method, in step S4, the control strategy of the most similar historical case is used as the startup strategy. That is, empirical knowledge from historical successful startup cases is used to assist in formulating the current gas generator set startup strategy. Key control instructions, such as the control sequence, parameter settings, valve opening, and ignition timing used in the most similar historical case, are directly input into the current unit control system for execution. This approach, by directly utilizing proven strategies from historical experience, significantly shortens control response time. Deep feature matching ensures the physical rationality of the strategy, avoiding combustion instability or thermal stress exceeding limits caused by aggressive or conservative strategies. This enables data-driven startup strategy optimization and improves the startup efficiency and stability of the gas generator set.

[0049] In summary, a gas generator set performance optimization method based on deep learning based on the embodiment of the present application is illustrated, which collects the initial state data stream of the unit in real time after receiving the start-up instruction, and introduces a deep learning algorithm to perform deep feature extraction on the initial state data stream of the unit, so as to capture the nonlinear coupling effect between the multidimensional state parameters during the transient start-up process of the gas generator set, and construct a gas start-up initial condition representation. Then, by querying and matching the current gas start-up initial condition representation with the historical successful start-up case library, the control strategy of the historical case most similar to the current situation is extracted as the start-up strategy to guide the start-up process of the gas generator set. This method can effectively reveal the state characteristics of the gas generator set in the initial stage of startup, and make full use of the successful experience accumulated in history to guide the new startup process, thereby realizing intelligent control of the startup process and improving the startup success rate.

[0050] Furthermore, a gas generator set performance optimization system based on deep learning is also provided.

[0051] Figure 6 FIG is a block diagram of a gas generator set performance optimization system based on deep learning according to an embodiment of the present application. Figure 6As shown, according to an embodiment of the present application, a gas generator set performance optimization system 100 based on deep learning includes: an initial state data acquisition module 110, which is used to collect real-time initial state data streams through a sensor network after receiving a start-up instruction, wherein the real-time initial state data streams include the temperature of each temperature measurement point, the pressure of each pressure measurement point, the rotational speed, the ambient temperature, the atmospheric pressure, the humidity and the downtime; an initial condition feature extraction module 120, which is used to input the real-time initial state data stream into a trained deep learning feature extraction model to obtain a gas start-up initial condition deep feature coding vector; a case retrieval matching module 130, which is used to use a case retrieval based on deep features to match the historical case that is most similar to the gas start-up initial condition deep feature coding vector from a historical successful start-up case library; a start-up strategy generation module 140, which is used to use the control strategy of the most similar historical case as the start-up strategy.

[0052] Here, those skilled in the art will understand that the specific operations of each module in the above-mentioned gas generator performance optimization system based on deep learning have been referred to above. Figures 1 to 5 The description of the deep learning-based gas generator set performance optimization method has been introduced in detail, and therefore, its repeated description will be omitted.

[0053] The basic principles of the present invention have been described above in conjunction with specific embodiments. However, it should be noted that the advantages, strengths, and effects mentioned in the present invention are merely illustrative and non-limiting, and should not be construed as necessarily possessed by each embodiment of the present invention. Furthermore, the specific details of the above embodiments are provided for illustrative purposes and to facilitate understanding, and are not intended to be limiting. These details do not necessarily limit the present invention to being implemented using these specific details.

[0054] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, please refer to the relevant description of other embodiments. In the several embodiments provided by the present invention, it should be understood that the disclosed system and method can be implemented in other ways. For example, the system embodiment described above is only schematic. For example, the unit division is only a logical function division, and there may be other division methods in actual implementation. The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the scheme of this embodiment.

[0055] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be encompassed therein. Any reference to a figure in a claim should not be construed as limiting the claim to which it relates.

[0056] In addition, it is obvious that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units stated in the system claims can also be implemented by one unit through software or hardware.

[0057] Finally, it should be noted that the above description has been provided for the purpose of illustration and description. In addition, the above embodiments are intended only to illustrate the technical solutions of the present invention and are not intended to be limiting. Although the technical solutions may be modified or replaced with equivalents with reference to the preferred embodiments, they do not depart from the spirit and scope of the technical solutions of the present invention.

Claims

1. A gas generator performance optimization method based on deep learning, characterized in that: include: After receiving the start command, the sensor network collects real-time initial state data streams, including the temperature of each temperature measurement point, the pressure of each pressure measurement point, the rotation speed, the ambient temperature, the atmospheric pressure, the humidity, and the downtime; Inputting the real-time initial state data stream into a trained deep learning feature extraction model to obtain a deep feature encoding vector of the gas startup initial condition; Use deep feature-based case retrieval to match the historical case that is most similar to the deep feature encoding vector of the gas startup initial condition from the historical successful startup case library; The control strategy of the most similar historical case is used as the starting strategy.

2. The gas generator set performance optimization method based on deep learning according to claim 1, characterized in that: Inputting the real-time initial state data stream into the trained deep learning feature extraction model to obtain a deep feature encoding vector of the gas startup initial condition, including: Inputting the real-time initial state data stream into a trained deep learning feature extraction model to obtain an initial gas startup initial condition deep feature encoding vector; The initial gas startup initial condition deep feature coding vector is optimized for feature expression robustness to obtain the initial gas startup initial condition deep feature coding vector.

3. The gas generator set performance optimization method based on deep learning according to claim 2, characterized in that: The trained deep learning feature extraction model is an initial state feature extractor based on a multi-layer perceptron.

4. The gas generator set performance optimization method based on deep learning according to claim 3 is characterized in that: Performing feature expression robustness optimization on the initial gas startup initial condition deep feature coding vector to obtain the initial gas startup initial condition deep feature coding vector, including: Performing feature decomposition based on one-dimensional convolution coding on the initial gas startup initial condition deep feature coding vector to obtain a set of initial gas startup initial condition deep feature local coding vectors; Based on the context-associated topological structure of the set of initial gas start-up initial condition deep feature local coding vectors, performing feature expression enhancement coding on each initial gas start-up initial condition deep feature local coding vector in the set of initial gas start-up initial condition deep feature local coding vectors to obtain a set of gas start-up initial condition deep feature local enhancement coding vectors; A set of local enhanced coding vectors of the gas startup initial condition deep feature is subjected to feature expression enhancement reconstruction to obtain the gas startup initial condition deep feature coding vector.

5. The gas generator set performance optimization method based on deep learning according to claim 4 is characterized in that: Based on the context-associated topological structure of the set of initial gas start-up initial condition deep feature local coding vectors, performing feature expression enhancement coding on each initial gas start-up initial condition deep feature local coding vector in the set of initial gas start-up initial condition deep feature local coding vectors to obtain a set of gas start-up initial condition deep feature local enhancement coding vectors, including: Calculating a feature correlation factor between any two initial gas start-up initial condition deep feature local encoding vectors in the set of the initial gas start-up initial condition deep feature local encoding vectors to obtain a gas start-up initial condition deep feature local correlation topology matrix composed of a plurality of feature correlation factors; Inputting the local correlation topology matrix of the gas startup initial condition deep feature into a gated mask function to obtain a local fine-grained correlation mask topology matrix of the gas startup initial condition deep feature; Based on the local fine-grained association mask topology matrix of the gas start-up initial condition deep feature, feature structure feedback modulation is performed on each initial gas start-up initial condition deep feature local coding vector in the set of the initial gas start-up initial condition deep feature local coding vector to obtain the set of the gas start-up initial condition deep feature local enhanced coding vector.

6. The gas generator set performance optimization method based on deep learning according to claim 5, characterized in that: Based on the local fine-grained association mask topology matrix of the gas start initial condition deep feature, feature structure feedback modulation is performed on each initial gas start initial condition deep feature local encoding vector in the set of the initial gas start initial condition deep feature local encoding vectors to obtain the set of the gas start initial condition deep feature local enhanced encoding vectors, including: Performing local topology structure balanced optimization on the gas startup initial condition deep feature local fine-grained correlation mask topology matrix to obtain an optimized gas startup initial condition deep feature local fine-grained correlation mask topology matrix; Each initial gas start-up initial condition deep feature local encoding vector in the set of the initial gas start-up initial condition deep feature local encoding vector and the optimized gas start-up initial condition deep feature local fine-grained association mask topology matrix are input into the feature dense feedback enhancement unit to obtain the set of the gas start-up initial condition deep feature local enhancement encoding vector.

7. The gas generator set performance optimization method based on deep learning according to claim 6, characterized in that: Performing feature expression enhancement reconstruction on a set of local enhanced coding vectors of the gas starting initial condition deep features to obtain the gas starting initial condition deep feature coding vector, including: The set of local enhanced coding vectors of the gas startup initial condition deep features is input into a feature expression enhancement and reconstruction module based on a Transformer architecture to obtain the gas startup initial condition deep feature coding vector.

8. The method for optimizing gas generator set performance based on deep learning according to claim 7, characterized in that: Use deep feature-based case retrieval to match historical cases that are most similar to the deep feature encoding vector of the gas startup initial conditions from the historical successful startup case library, including: Calculate the cosine similarity between the historical case gas startup initial condition deep feature coding vector associated with each historical case in the historical successful startup case library and the gas startup initial condition deep feature coding vector, and take the historical case corresponding to the maximum cosine similarity as the most similar historical case.

9. A gas generator performance optimization system based on deep learning, characterized in that: include: An initial state data acquisition module is used to collect a real-time initial state data stream through a sensor network after receiving a start-up instruction. The real-time initial state data stream includes the temperature of each temperature measurement point, the pressure of each pressure measurement point, the rotational speed, the ambient temperature, the atmospheric pressure, the humidity, and the downtime duration; An initial condition feature extraction module, configured to input the real-time initial state data stream into a trained deep learning feature extraction model to obtain a gas startup initial condition deep feature encoding vector; A case retrieval and matching module is used to use a case retrieval based on deep features to match a historical case that is most similar to the deep feature encoding vector of the gas startup initial condition from a historical successful startup case library; The startup strategy generation module is used to use the control strategy of the most similar historical case as the startup strategy.

Images