Method for predicting flame intensity in heavy gas turbine combustion chamber based on attention mechanism

Abstract

The application discloses a heavy gas turbine combustor flame intensity prediction method based on an attention mechanism, and the method comprises the following steps: collecting multi-source physical quantity data in the operation process of a gas turbine, and constructing a multi-dimensional time sequence input containing physical feature coupling relationship; performing feature fusion processing on the multi-dimensional time sequence data, extracting temperature difference features, pulsation frequency band features and sliding window data partitioned in time sequence, and forming a standardized structured data set; adopting a multi-layer gated recurrent unit (GRU) to construct an encoder, performing deep time sequence feature extraction, and generating encoder output features; constructing a decoder with an attention mechanism based on the encoder output features, dynamically weighting and focusing key time steps and key channel features through an Additive Attention mechanism, and fusing attention output and supervision label features; performing classification processing on the decoder output through a multi-layer fully connected network, outputting flame state prediction results, and realizing real-time identification of whether a flame exists.

Description

Technical Field

[0001] This invention relates to the field of gas turbine combustor flame intensity prediction technology, and in particular to a method for predicting the flame intensity of a heavy-duty gas turbine combustor based on an attention mechanism. Background Technology

[0002] Heavy-duty gas turbines, as core equipment in energy and power systems, are widely used in industries such as power, metallurgy, and petrochemicals. With the accelerated transformation of industry towards intelligent manufacturing, combustion chamber flame intensity monitoring technology has evolved from traditional optical sensors (such as ultraviolet / infrared detectors) to deep learning modeling that integrates multiple physical quantities. Specifically, existing technologies utilize the collaborative operation of multi-source signals such as speed, temperature, and pressure to construct a comprehensive monitoring architecture covering data acquisition, feature extraction, and state recognition. Among these, LSTM-based temporal modeling methods and Transformer architectures have become mainstream technical approaches, but a systematic modeling scheme specifically for the physical characteristics of gas turbine combustion chambers has yet to be developed.

[0003] However, existing flame intensity identification methods directly model using a single physical quantity and lack a dynamic focusing mechanism for the heterogeneous characteristics of multi-channel data. This may lead to the dilution of feature information or the loss of key temporal features. Specifically, under non-steady-state conditions such as ignition / extinguishing, traditional methods struggle to effectively capture the coupling relationship between the energy distribution and temperature difference characteristics (such as the difference between compressor outlet temperature and exhaust temperature) in the combustion pulsation frequency band due to the lack of a collaborative system between the encoder-decoder architecture and the attention mechanism, resulting in insufficient model generalization ability. Consequently, existing technologies suffer from high misjudgment rates (experimental data show that misjudgment rates are generally higher than 3%) and response delays (typical delays reaching [missing information]) in dynamic load switching scenarios. Defects such as these directly affect combustion stability control and emission optimization, thus hindering the improvement of the intelligence level of gas turbines. Summary of the Invention

[0004] The main objective of this invention is to provide a method for predicting the flame intensity of a heavy-duty gas turbine combustor based on an attention mechanism.

[0005] Another objective of this invention is to provide a flame intensity prediction device for the combustion chamber of a heavy-duty gas turbine based on an attention mechanism.

[0006] The third objective of this invention is to provide a computer device.

[0007] A fourth objective of this invention is to provide a non-transitory computer-readable storage medium.

[0008] To achieve the above objectives, a first aspect of the present invention proposes a method for predicting the flame intensity of a heavy-duty gas turbine combustor based on an attention mechanism, comprising: S1 collects multi-source physical quantity data during the operation of the gas turbine, including speed, compressor outlet temperature, exhaust temperature and combustion chamber pulsation value, and constructs a multi-dimensional time series input containing physical feature coupling relationships; S2, perform feature fusion processing on the multidimensional time series data to extract temperature difference features, pulsation frequency band features and sliding window data partitioned by time order, forming a standardized structured dataset; S3, an encoder is constructed using a multi-layer gated recurrent unit (GRU) to perform deep temporal feature extraction on the structured dataset and generate encoder output features; S4. Construct a decoder with an attention mechanism based on the encoder output features. Dynamically weight and focus on key time steps and key channel features through the AdditiveAttention mechanism, and fuse the attention output with the supervision label features to guide the prediction process. S5 uses a multi-layer fully connected network to classify the decoder output and output flame state prediction results, enabling real-time identification of the presence of flames.

[0009] In one embodiment of the present invention, the step of performing feature fusion processing on the multidimensional time series data to extract temperature difference features, pulsation frequency band features, and sliding window data partitioned by time order to form a standardized structured dataset includes: S21, calculate the difference between the compressor outlet temperature and the exhaust temperature as the temperature difference feature, and segment the time series using a sliding window with a window length of [value missing]. And the step size is ,in and Adjustments were made based on the response characteristics of the actual test data; S22, divide the combustion chamber pulsation value into frequency bands and extract... to The energy distribution characteristics of each frequency band are analyzed, and the energy of each frequency band is normalized to... Interval.

[0010] In one embodiment of the present invention, the step of constructing an encoder using a multi-layer gated recurrent unit (GRU) to perform deep temporal feature extraction on the structured dataset and generate encoder output features includes: S31, setting encoder to include A layer of GRU unit, where the hidden state dimension of each layer of GRU is... ,in and Determined through response patterns based on actual experimental data; S32, perform gradient clipping on the encoder output features, with a clipping threshold of [value missing]. And it uses a random discarding technique with a discard probability of . .

[0011] In one embodiment of the present invention, the step of constructing a decoder with an attention mechanism based on the encoder output features, and dynamically weighting and focusing on key time steps and key channel features through an Additive Attention mechanism, includes: S41, Calculate attention weights At that time, use the decoder's current hidden state. encoder history status The dot product is given by the formula: ; S42 concatenates the attention output with the supervised label features, and generates a fused feature vector through a fully connected layer. ,in For attention-weighted features, These are label features.

[0012] In one embodiment of the present invention, the step of classifying the decoder output through a multi-layer fully connected network and outputting a flame state prediction result includes: S51 uses the Softmax activation function to normalize the output of the fully connected layer, resulting in the output probability distribution. ,in This is the weight matrix. It is the bias vector; S52, Set classification threshold ,when If the flame is present, it is determined to be present; otherwise, it is determined to be extinguished.

[0013] In one embodiment of the present invention, it further includes: S6 performs condition balancing on the training samples, adjusting the sample weights according to the condition categories such as ignition, acceleration, and shutdown. ,in , Operating condition category The number of samples, This represents the total number of samples.

[0014] To achieve the above objectives, a second aspect of the present invention provides a heavy-duty gas turbine combustor flame intensity prediction device based on an attention mechanism, comprising: The multi-source data acquisition module is used to collect multi-source physical quantity data during the operation of the gas turbine, including speed, compressor outlet temperature, exhaust temperature and combustion chamber pulsation value, and to construct a multi-dimensional time series input containing physical feature coupling relationships; The feature fusion processing module is used to perform feature fusion processing on the multidimensional time series data, extract temperature difference features, pulsation frequency band features and sliding window data partitioned by time order, and form a standardized structured dataset. The encoder construction module is used to construct an encoder using a multi-layer gated recurrent unit (GRU), perform deep temporal feature extraction on the structured dataset, and generate encoder output features. The decoder construction module is used to construct a decoder with an attention mechanism based on the encoder output features. It dynamically weights and focuses on key time steps and key channel features through the Additive Attention mechanism, and fuses the attention output with the supervision label features to guide the prediction process. The classification processing module is used to classify the decoder output through a multi-layer fully connected network, output flame state prediction results, and realize real-time identification of the presence of flames.

[0015] To achieve the above objectives, a third aspect of this application provides a computer device comprising a processor and a memory; wherein the processor runs a program corresponding to the executable program code stored in the memory, for implementing a method for predicting the flame intensity of a heavy-duty gas turbine combustor based on an attention mechanism as described in the first aspect embodiment.

[0016] To achieve the above objectives, a fourth aspect of this application provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements an attention mechanism-based method for predicting the flame intensity of a heavy-duty gas turbine combustor as described in the first aspect embodiment.

[0017] The embodiments of the present invention have the following beneficial effects: The embodiments of the present invention can effectively integrate multi-source physical quantity characteristics and dynamically focus on key information, significantly improving the accuracy and robustness of flame intensity prediction in the combustion chamber of heavy-duty gas turbines under complex operating conditions. At the same time, it has edge deployment capability to realize real-time flame status identification and early warning. Attached Figure Description

[0018] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart illustrating a method for predicting the flame intensity of a heavy-duty gas turbine combustor based on an attention mechanism, provided in an embodiment of the present invention; Figure 2 A schematic diagram of a heavy-duty gas turbine combustor flame intensity prediction model based on an attention mechanism, provided in an embodiment of the present invention; Figure 3A system diagram of a heavy-duty gas turbine combustor flame intensity prediction model based on an attention mechanism, provided for an embodiment of the present invention; Figure 4 This is a structural diagram of a heavy-duty gas turbine combustor flame intensity prediction device based on an attention mechanism, provided in an embodiment of the present invention. Detailed Implementation

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0021] The following description, with reference to the accompanying drawings, describes a method and apparatus for predicting the flame intensity of a heavy-duty gas turbine combustor based on an attention mechanism, according to an embodiment of the present invention.

[0022] Example 1 This embodiment provides a method for predicting the flame intensity of a heavy-duty gas turbine combustor based on an attention mechanism. For example... Figure 1 As shown, the method includes the following steps: S1 collects multi-source physical quantity data during the operation of the gas turbine, including speed, compressor outlet temperature, exhaust temperature and combustion chamber pulsation value, and constructs a multi-dimensional time series input containing physical feature coupling relationships.

[0023] Specifically, in some implementations, the data input module of this invention constructs a multi-dimensional time series input containing the coupling relationships of physical characteristics by collecting multi-source physical quantity data during the operation of the gas turbine. Specifically, the key physical quantities collected in this step include the gas turbine rotational speed. (Unit: r / min) Compressor outlet temperature (Unit: °C) Exhaust temperature (Unit: °C) and combustion chamber pulsation value (Unit: Pa or dB) These parameters reflect the mechanical operating state, thermodynamic cycle characteristics and combustion dynamic behavior of the gas turbine, respectively.

[0024] In this embodiment of the invention, raw data is collected in real time through a distributed sensor network and transmitted to a data processing platform via industrial communication protocols (such as Modbus TCP and OPC UA). To enhance feature representation capabilities, this invention introduces a heuristic physical feature fusion strategy in the data preprocessing stage, such as constructing temperature difference features. To reflect changes in combustion efficiency; the combustion chamber pulsation signal is divided into frequency bands, and the energy distribution characteristics of specific frequency bands (such as 50–200 Hz) are extracted. This is used to capture combustion instabilities. Furthermore, a sliding window sampling method is employed to divide the raw time-series data into fixed-length window sequences, with a window length of... Typically set to 100–200 sampling points, step size The number of sampling points is 20–50 to balance temporal continuity and computational efficiency.

[0025] Furthermore, the sampling frequency The standard setting is typically 100 Hz–1 kHz, depending on the sensor type and system response requirements. Normalization is achieved using the Z-score normalization method. ,in The mean, The standard deviation is used to ensure that data from all channels are modeled on a uniform scale.

[0026] In practical applications, this step is suitable for flame state recognition tasks under typical operating conditions of gas turbines, such as ignition, acceleration, load changes, and flameout. By fusing multi-source physical quantities and constructing coupled features, it provides structured and physically meaningful inputs for the subsequent encoder-decoder architecture and attention mechanism, thereby improving the model's ability to perceive and recognize changes in flame intensity.

[0027] S2, perform feature fusion processing on the multidimensional time series data to extract temperature difference features, pulsation frequency band features and sliding window data partitioned by time order, forming a standardized structured dataset.

[0028] Specifically, in the technical solution of this invention, feature fusion processing of the multidimensional time series data is a key preliminary step in constructing a high-precision flame intensity prediction model. This step transforms the raw sensor data into a standardized structured dataset with physical meaning and temporal structure by extracting temperature difference features, pulsation frequency band features, and sliding window data partitioned by time sequence, thereby providing high-quality input for subsequent deep learning modeling.

[0029] In this embodiment of the invention, the feature fusion processing is first based on multiple physical quantities during the operation of the gas turbine, including rotational speed, compressor outlet temperature, exhaust temperature, and combustion chamber pulsation signals. Specifically, the temperature difference feature is constructed using the difference between the compressor outlet temperature and the exhaust temperature. This feature can reflect the changing trend of combustion efficiency and enhance the model's sensitivity to combustion conditions. Furthermore, the pulsation frequency band feature is extracted through frequency domain analysis, dividing the combustion chamber pulsation signal into multiple frequency bands (such as the low-frequency band). Mid-frequency band High frequency band The study also calculated the energy distribution of each frequency band to capture the manifestations of combustion instability at different frequencies.

[0030] Furthermore, to preserve the temporal evolution characteristics, a sliding window method is used to segment the original time series for sampling. The window length is typically set to... Each time step, step size This ensures the continuity of the time series and the stability of the model input. The data within each window undergoes standardization (such as Z-score normalization) to eliminate the dimensional differences between signals from different sensors, forming a unified feature vector input format.

[0031] In practical applications, this step is suitable for flame state recognition tasks under non-steady-state conditions such as gas turbine ignition, acceleration, and shutdown. By fusing physically meaningful features, the model can more accurately capture key changes in the combustion process, improving recognition robustness. Simultaneously, the construction of the structured dataset provides a unified input interface for the subsequent GRU encoder and attention mechanism decoder, serving as a fundamental step for end-to-end modeling and edge deployment.

[0032] Furthermore, S2 includes: S21, calculate the difference between the compressor outlet temperature and the exhaust temperature as the temperature difference feature, and segment the time series using a sliding window with a window length of [value missing]. And the step size is ,in and Adjustments are made based on the response characteristics of actual test data.

[0033] Specifically, in the technical solution of this invention, calculating the difference between the compressor outlet temperature and the exhaust temperature as a temperature difference feature, and segmenting the time series using a sliding window, is one of the key steps in realizing the fusion of multi-source physical quantities and time series modeling. This step is executed in the data preprocessing module, and its technical implementation principle is based on the dynamic response characteristics of the temperature field during gas turbine combustion. It aims to extract features with physical meaning and diagnostic sensitivity, providing high-quality input data for subsequent deep learning models.

[0034] In this embodiment of the invention, the compressor outlet temperature With exhaust temperature These are two important thermodynamic parameters during the operation of a gas turbine, and their difference is... It can reflect the intensity change of energy release inside the combustion chamber. This temperature difference characteristic has good physical interpretability and exhibits high sensitivity under non-steady-state conditions such as ignition, flameout, and sudden load changes, and can effectively capture abnormal fluctuations in the combustion state.

[0035] Furthermore, to adapt to the input requirements of deep learning models for time series data, this invention employs a sliding window method to segment the time series. The length of the sliding window... With step size It is adjusted based on the response characteristics of actual test data, usually The range of values ​​is Each time step for Each time step, the specific value needs to be combined with the sampling frequency of the gas turbine. With the dynamic response time constant of the combustion process Perform configuration optimization. For example, if the sampling frequency is... The response time constant is ,but Can be set to Each time step corresponds to A time window is set to ensure that the model can capture instantaneous changes in the combustion state.

[0036] In practical applications, this step is typically deployed in the online monitoring system of gas turbines, working in conjunction with multi-channel sensors (such as infrared flame detectors, pressure sensors, and temperature sensors) to construct multi-dimensional time-series input vectors. Through a sliding window mechanism, the system can continuously generate time-series segments, providing structured input for the subsequent encoder-attention decoder model, thereby achieving real-time prediction of flame intensity and anomaly detection.

[0037] Furthermore, this step enhances the model's ability to perceive changes in combustion state by introducing physically meaningful temperature difference features. Simultaneously, the time-series segmentation processing using a sliding window improves the model's accuracy in depicting temporal evolution trends. This method demonstrates good robustness and adaptability under multiple operating conditions and unsteady-state operation, providing a solid data foundation for the intelligent diagnosis and control of heavy-duty gas turbine combustion systems.

[0038] S22, divide the combustion chamber pulsation value into frequency bands and extract... to The energy distribution characteristics of each frequency band are analyzed, and the energy of each frequency band is normalized to... Interval.

[0039] Specifically, in the technical solution of this invention, the combustion chamber pulsation value is divided into frequency bands and extracted. to Analyzing the energy distribution characteristics of the frequency band is one of the key preprocessing steps for flame state identification. This step aims to extract physically meaningful frequency domain features from the raw pulsating signal, thereby enhancing the model's ability to perceive dynamic changes in the combustion process.

[0040] In this embodiment of the invention, this step first involves frequency domain analysis of the combustion chamber pulsation signal using either Fast Fourier Transform (FFT) or Short-Time Fourier Transform (STFT). Considering the non-stationary characteristics of the gas turbine combustion process, the STFT method is preferred, using a sliding window to segment the signal. The window length is typically set to... Step size is This process balances time and frequency resolution. After transformation, the signal is mapped to a time-frequency graph, where the horizontal axis represents time, the vertical axis represents frequency, and the amplitude represents energy density.

[0041] Furthermore, based on the typical frequency response characteristics of the combustion chamber pulsation signal, the frequency domain is divided into... Each frequency band, for example , , Each frequency band corresponds to a specific physical phenomenon in the combustion process (such as combustion instability, flame propagation frequency, etc.). Within each frequency band, its energy value is calculated. The power spectral density (PSD) integration method is typically used, i.e.:

[0042] in This is the power spectral density function within this frequency band. Subsequently, the energy of each frequency band is normalized to... The normalization interval is defined to eliminate the impact of amplitude differences between different frequency bands on the model input. The normalization formula is:

[0043] in and These represent the minimum and maximum energy values ​​across all frequency bands. The normalized frequency band energy distribution, as part of the model input, together with time-domain features such as rotational speed and temperature, forms a multi-dimensional temporal feature vector, which is then input into the encoder module for deep modeling.

[0044] In this embodiment of the invention, a multi-channel sensor data acquisition module, typically deployed in the online monitoring system of a gas turbine, is used to perform real-time frequency band division and energy extraction of the combustion chamber pulsation signal. The technical advantages are: frequency band division effectively separates characteristic frequency components closely related to the flame state; normalization improves the model's robustness and generalization ability to input features, providing a structured and standardized data foundation for feature focusing in subsequent attention mechanisms.

[0045] S3. An encoder is constructed using a multi-layer gated recurrent unit (GRU) to perform deep temporal feature extraction on the structured dataset and generate encoder output features.

[0046] Specifically, in the technical solution of this invention, constructing an encoder using a multi-layer gated recurrent unit (GRU) to perform deep temporal feature extraction on the structured dataset and generate encoder output features is one of the key steps in realizing the prediction of flame intensity in the combustion chamber of a heavy-duty gas turbine. This step, by constructing a multi-layer GRU structure, performs deep modeling on the structured temporal data composed of multiple source physical quantities, thereby extracting high-dimensional feature representations with physical meaning and temporal dependence.

[0047] In this embodiment of the invention, the encoder module is composed of several layers of stacked GRU units. Each GRU layer controls the flow and forgetting of information through a gating mechanism (update gate and reset gate), thereby effectively modeling long-term time-series dependencies. Specifically, the input data is a multi-dimensional time series after sliding window processing, and its dimensions include, but are not limited to, engine speed, compressor outlet temperature, exhaust temperature, and combustion chamber pulsation value. In some implementations, the length of the input sequence can be set to... Each time step contains [number] time steps. A physical characteristic, forming a shape as The input tensor. The hidden layer dimension of GRU is typically set to... The number of layers can be set to This enhances the model's ability to capture complex temporal patterns.

[0048] Furthermore, the GRU uses the Sigmoid activation function for both its update and reset gates, and the Tanh function is used to compute the hidden states. The Adam optimizer is used during model training, with a learning rate set to... This is combined with gradient clipping (clipnorm = 1.0) to prevent gradient explosion. Furthermore, to improve generalization, a dropout mechanism can optionally be introduced between GRU layers, with a dropout rate set to... .

[0049] This encoder module is primarily used to process multi-channel sensor data from gas turbines under non-steady-state conditions such as ignition, acceleration, and shutdown. Through time-series modeling using a multi-layer GRU, it can effectively extract key features during the combustion state evolution process, providing high-quality feature representations for subsequent attention mechanism decoding.

[0050] The technical advantage of this step lies in its ability, through a deep GRU structure, to capture the potential dynamic evolution patterns within the input sequence, thereby generating discriminative encoded features. Compared to traditional RNNs, the GRU structure offers significant advantages in training stability and convergence speed; compared to LSTMs, it has fewer parameters, higher computational efficiency, and is better suited for edge deployment requirements. Therefore, this step provides a solid foundation for the subsequent introduction of attention mechanisms and flame intensity prediction, significantly improving the model's temporal modeling capabilities and recognition accuracy.

[0051] Furthermore, S3 includes: S31, setting encoder to include A layer of GRU unit, where the hidden state dimension of each layer of GRU is... ,in and The response pattern is determined through actual experimental data.

[0052] Specifically, in the technical solution of the present invention, the encoder includes... A layer of GRU unit, where the hidden state dimension of each layer of GRU is... This is one of the key steps in achieving deep temporal modeling and feature extraction. This step involves constructing a multi-layer gated recurrent unit (GRU) structure to abstract and semantically encode the multi-dimensional time-series data collected during gas turbine operation layer by layer, thereby extracting deep feature representations with physical meaning and temporal dependence.

[0053] In this embodiment of the invention, the number of encoder layers With the hidden state dimension of each layer It was determined through optimization based on the response patterns of actual experimental data. Specifically, The value range is typically between 2 and 4 layers to balance the model's expressive power and computational complexity; while The dimensions are adjusted based on the complexity of the input features and the convergence stability of the model, and are generally set between 64 and 256. By analyzing the flame intensity signal response under different operating conditions (such as ignition, acceleration, shutdown, etc.), the optimal combination of layers and dimensions can be determined to ensure that the model can effectively capture the dynamic evolution of the combustion state.

[0054] Furthermore, the GRU encoder controls the flow and forgetting of information through gating mechanisms (update gate and reset gate), exhibiting stronger gradient stability and training efficiency compared to traditional RNNs. The output of each GRU layer serves as the input to the next layer, forming a hierarchical feature representation. Finally, the hidden state sequence output by the encoder is fed into the attention decoder module for weighted focusing at key time steps.

[0055] This step has significant technical value in practical applications, especially in flame intensity identification tasks in the combustion chambers of heavy-duty gas turbines. Through proper setting... and The model can more accurately model the nonlinear temporal relationships in the combustion process, improving the robustness and generalization ability of flame state recognition. Furthermore, the lightweight structure of GRU facilitates the deployment of the model on edge computing devices, meeting the dual demands of real-time performance and computing resources in industrial settings.

[0056] S32, perform gradient clipping on the encoder output features, with a clipping threshold of [value missing]. And it uses a random discarding technique with a discard probability of . .

[0057] Specifically, in the technical solution of this invention, gradient clipping of the encoder output features combined with random dropout technology is one of the key steps to improve the stability and generalization ability of model training. This step is mainly applied in the training stage of deep learning models, especially when processing multi-channel time-series data of gas turbine combustors. It can effectively alleviate the gradient explosion problem and suppress overfitting, thereby improving the recognition accuracy and robustness of the model under complex operating conditions.

[0058] In this embodiment of the invention, gradient clipping applies norm constraints to the gradients calculated during backpropagation to prevent excessively large gradient magnitudes from causing instability in model parameter updates. Specifically, this invention employs L2 norm clipping, setting a clipping threshold of [value missing]. When the L2 norm of the gradient When the threshold is exceeded, the gradient will be scaled proportionally to no more than [a certain value]. The range of , its mathematical expression is:

[0059] in, For the original gradient, This represents the clipped gradient. In some implementations, Can be set to or The specific values ​​are dynamically adjusted based on gradient fluctuations during model training to ensure the stability of the model under different operating conditions.

[0060] Meanwhile, this invention introduces a random dropout technique into the encoder output layer to enhance the model's generalization ability. Dropout forces the model to learn more robust feature representations by randomly "shutting down" a portion of neurons during training. In this invention, the dropout probability of the Dropout operation is... , usually set arrive Between. For example, when In this process, about 30% of the neurons in the encoder output features are randomly discarded in each training iteration, thereby reducing the model's dependence on specific features and improving its adaptability under unknown conditions.

[0061] This step is mainly applied to the training phase of the gas turbine combustor flame intensity recognition model. Due to the characteristics of gas turbine operating data such as high noise, multiple channels, and long time series, the combination of gradient clipping and Dropout can effectively deal with gradient instability and overfitting problems during the training process. Especially under non-steady-state conditions such as ignition, flameout, and sudden load changes, it helps the model converge quickly and maintain a high recognition accuracy.

[0062] Furthermore, gradient pruning ensures the numerical stability of the model during backpropagation, avoiding training failures caused by gradient explosion; while Dropout, by introducing randomness, enhances the model's tolerance to noise and interference, improving its robustness in real-world industrial environments. The synergistic effect of these two mechanisms provides a reliable training guarantee for the deep learning model based on GRU and attention mechanisms in this invention, and is a crucial technical support for achieving a flame recognition accuracy of over 97%.

[0063] S4. Based on the encoder output features, construct a decoder with an attention mechanism, dynamically weight and focus on key time steps and key channel features through the AdditiveAttention mechanism, and fuse the attention output with the supervision label features to guide the prediction process.

[0064] Specifically, in the technical solution of this invention, the step of "constructing a decoder with an attention mechanism based on the encoder output features, dynamically weighting and focusing on key time steps and key channel features through the Additive Attention mechanism, and fusing the attention output with the supervision label features to guide the prediction process" is one of the core steps in achieving high accuracy and strong generalization ability of the flame intensity recognition model. This step, by introducing an Additive Attention mechanism, dynamically weights the temporal features output by the encoder during the decoding process, thereby focusing on the time steps and channel features that are crucial for flame state recognition, improving the model's adaptability and recognition sensitivity to complex working conditions.

[0065] In this embodiment of the invention, the decoder consists of an embedding layer, multiple GRU units, and an additive attention module. The encoder outputs a feature sequence. ,in Indicates the first The hidden state at each time step. The total number of time steps. To hide the state dimension. The decoder at each step Generate the current hidden state and use it as a query vector Input to the additive attention module. Attention weights. Calculated using the following formula:

[0066] in, It is the attention weight matrix. It is a bias vector. It is the attention vector. Through this mechanism, the model can adaptively weight each time step of the encoder output, thereby extracting the feature segments that are most informative for the current decoding task.

[0067] Furthermore, the attention-weighted context vector With supervised label features The data is fused and used as input to the decoder at the next time step. This fusion method enhances the model's ability to perceive the target output by introducing label information, realizing a teacher forcing strategy, which helps improve training stability and prediction accuracy.

[0068] In this embodiment of the invention, the dimension of the attention module The attention vector is typically aligned with the encoder output dimension to ensure information alignment. With weight matrix The model learns automatically during training using the backpropagation algorithm, eliminating the need for manual configuration. Furthermore, to improve the model's deployment efficiency on edge devices, the attention calculation process employs low-rank approximation or quantization strategies for optimization, ensuring a recognition accuracy of over 97% even with limited computing resources.

[0069] In practical applications, this step is particularly suitable for flame state identification tasks under non-steady-state conditions such as gas turbine ignition, flameout, and sudden load changes. Through the attention mechanism, the model can effectively capture weak signals before changes in combustion state, thereby achieving early warning and state discrimination. Simultaneously, this step provides a unified feature fusion and attention computation framework for subsequent multi-task modeling (such as flame intensity regression and anomaly detection), enhancing the system's scalability and industrial adaptability.

[0070] Furthermore, S4 includes: S41, Calculate attention weights At that time, use the decoder's current hidden state. encoder history status The dot product is given by the formula: .

[0071] Specifically, in the flame intensity recognition model of the present invention, attention weights are calculated. This is one of the core steps in achieving key feature focusing. This step utilizes the hidden state of the decoder at the current time step. State at a historical moment in the encoder A dot product operation is performed to quantify the correlation between the current decoding state and the historical encoding state. Specifically, the formula for calculating the attention weights is:

[0072] in, The decoder is in the first... The hidden state vector at each time step Indicates the encoder at the 1st The hidden state vector at each time step This represents the total number of time steps output by the encoder. Essentially, this formula is a Softmax function used to normalize the dot product result into a probability distribution, allowing the model to dynamically focus on the encoding state most relevant to the current decoding state.

[0073] In this embodiment of the invention, this step is typically embedded after the GRU layer of the decoder, and a weighted summation of the encoder output is achieved by constructing an attention matrix. The dot product operation is performed in a high-dimensional vector space, and its magnitude reflects the semantic similarity between the current decoding state and the historical encoding state. This is achieved through an exponential function. The dot product result is non-linearly mapped and then normalized to ensure the attention weights are optimized. The sum of these sums is 1, thus forming an effective attention distribution.

[0074] Furthermore, the hidden state dimension of the encoder and decoder is typically set to 128 or 256 to balance model expressiveness and computational efficiency. The calculation of attention weights is performed independently at each time step, therefore its time complexity is O(n log n). ,in The input sequence length is specified. In practical applications, this step is typically deployed on edge computing devices, requiring the model to maintain recognition accuracy while having low computational latency. Therefore, quantization or pruning strategies can be optionally employed in the implementation to further optimize performance.

[0075] This step plays a crucial role in identifying flame intensity in the combustion chamber of heavy-duty gas turbines. By dynamically focusing the encoded state of key time steps, the model can effectively capture precursor features of flame state changes, significantly improving robustness and accuracy, especially under non-steady-state conditions such as ignition and flameout. Furthermore, this attention mechanism provides interpretability to the model, facilitating subsequent feature backtracking and diagnostic analysis within industrial control systems, thereby enhancing the system's intelligence and operational safety.

[0076] S42 concatenates the attention output with the supervised label features, and generates a fused feature vector through a fully connected layer. ,in For attention-weighted features, These are label features.

[0077] Specifically, in the technical solution of this invention, the step "concatenates the attention output with the supervision label features and generates a fused feature vector through a fully connected layer" is described. This is a crucial step in realizing feature fusion and model guidance in the flame state recognition model. This step involves extracting key temporal features through an attention mechanism. With supervised label features The features are concatenated to construct a fusion feature vector with stronger semantic expressive power. This improves the model's recognition accuracy and robustness under complex working conditions.

[0078] In this embodiment of the invention, attention output It is obtained by weighted aggregation of the multi-channel temporal features output by the encoder through the Additive Attention mechanism in the decoder, and its dimension is usually 1. ,in Indicates the number of time steps. Representing feature dimensions. Supervised label features. It is a pre-processed flame intensity signal, its dimension is ,in This indicates the number of categories or feature dimensions of the label. (Through...) Operation, will and The features are concatenated along the feature dimension to obtain the fused feature vector. Its dimensions are This splicing operation not only preserves the attention mechanism's ability to focus on key features, but also guides the model's prediction process by introducing label information, a so-called "forced teaching" strategy.

[0079] In this step It is typically set between 64 and 128, depending on the number of hidden units in the GRU layer of the encoder; Generally, it's 1 or 2, corresponding to binary (flame present / absent) or multi-category tasks, respectively. (After splicing...) As input to the fully connected layer, the fully connected layer typically consists of 2 to 3 layers, with 64 to 256 neurons per layer. The activation function can be either ReLU or Tanh to enhance non-linear expressive power. The output layer uses either the Softmax or Sigmoid function to achieve classification or regression prediction of the flame state.

[0080] Furthermore, this step is primarily applied to flame status identification tasks in heavy-duty gas turbine combustors, particularly under non-steady-state conditions such as ignition, flameout, and load changes. By fusing temporal features extracted through an attention mechanism with supervisory label information, the model's sensitivity to flame intensity changes and response speed are improved. This method is suitable for deployment in edge computing devices or industrial control systems, enabling second-level inference and real-time flame status monitoring.

[0081] This step significantly enhances the model's ability to discriminate flame states by introducing joint modeling of supervised label features and attention outputs. In actual testing, this fusion strategy helps improve the model's accuracy to over 97% and accelerates the convergence speed of the training process. Furthermore, this step provides a unified feature fusion interface for subsequent multi-task modeling (such as flame intensity regression and anomaly detection), enhancing the model's scalability and industrial adaptability.

[0082] S5 uses a multi-layer fully connected network to classify the decoder output and output flame state prediction results, enabling real-time identification of the presence of flames.

[0083] Specifically, in the technical solution of the present invention, the step of "classifying the decoder output through a multi-layer fully connected network, outputting the flame state prediction result, and realizing real-time identification of whether a flame exists" is a key link in the entire model inference process. Its technical implementation is based on the classification mechanism in deep learning, combined with the high-dimensional feature vector output by the decoder, to complete the binary classification or soft classification judgment of the flame state.

[0084] In this embodiment of the invention, the decoder output is a sequence of temporal feature vectors, typically with a dimension of . ,in Indicates the time step. This represents the feature dimension of the decoder output. To achieve real-time flame state recognition, this invention employs a multi-layer fully connected network (FCN) to perform nonlinear mapping and classification on the final state vector output by the decoder (such as the hidden state at the last time step). This fully connected network typically consists of 2 to 3 fully connected layers, each followed by a ReLU activation function, and optionally introduces a Dropout mechanism to prevent overfitting. The final output layer uses the Sigmoid function for binary classification, with output values ​​ranging from [value missing]. , which represents the probability that a flame exists at the current moment.

[0085] Furthermore, the input dimension of a fully connected network is The output dimension is The intermediate layer can be set as or The specific parameters can be optimized based on actual working condition data. The cross-entropy loss function is used during model training. ,in The label represents the actual flame status (0 indicates no flame, 1 indicates flame). To predict probabilities for the model. In actual deployment, model inference latency is controlled within... It meets the second-level response requirement of the gas turbine control system for real-time identification of flame status.

[0086] This step is primarily used for flame status monitoring during critical operating conditions such as ignition, operation, and shutdown of the gas turbine combustor. By mapping the high-dimensional semantic features output by the decoder to classification results indicating the presence or absence of flame, it can be directly integrated into industrial control systems to achieve early warning of flame extinction and safety interlock control. Furthermore, this classification module can be expanded into a multi-task learning framework, supporting tasks such as flame intensity regression prediction and combustion anomaly detection.

[0087] Furthermore, this step, through end-to-end classification network design, effectively integrates the temporal features extracted by the encoder with the key information focused by the attention mechanism, improving the robustness and accuracy of flame state recognition. In actual testing, the model achieves a recognition accuracy of over 97% under multiple operating conditions, significantly outperforming traditional threshold-based optical flame detection methods, providing reliable technical support for the intelligent operation of gas turbine combustion systems.

[0088] The flame intensity prediction method for heavy-duty gas turbine combustor based on attention mechanism in this invention can effectively integrate the temporal features of multi-source physical quantities, improve the accuracy and robustness of flame state recognition through the GRU encoder and attention decoder structure, adapt to complex working conditions and support real-time deployment of edge devices.

[0089] Furthermore, S5 includes: S51 uses the Softmax activation function to normalize the output of the fully connected layer, resulting in the output probability distribution. ,in This is the weight matrix. This is the bias vector.

[0090] Specifically, in the technical solution of this invention, the step "normalizing the output of the fully connected layer using the Softmax activation function, and the output probability distribution" This step is a crucial step in achieving flame state classification and prediction. By introducing the Softmax function, the high-dimensional feature vector output by the decoder is mapped to an output with probabilistic semantics, thereby providing interpretable classification results for the identification of flame intensity in the combustion chamber of heavy-duty gas turbines.

[0091] In this embodiment of the invention, this step is typically located in the output layer of the model, and its input is the feature vector output by the decoder. After being processed by multiple GRU layers and an attention mechanism, this vector has incorporated temporal information from multiple physical sources (such as rotational speed, temperature, and pulsating pressure) and extracted features strongly correlated with the flame state. The fully connected layer uses a weight matrix... and bias vector right Perform a linear transformation to obtain the unnormalized output. The Softmax function then normalizes this output, converting it into a probability distribution representing the presence or absence of flame. Specifically, if the output dimension is Then the calculation form of Softmax is:

[0092] Furthermore, the weight matrix The dimensions are usually 1 ,in The feature dimension output by the decoder. The number of categories for the flame state (e.g., in binary classification). Bias vector The dimension is This is used to adjust the baseline of the output. In actual training, the parameters of this fully connected layer are optimized using the backpropagation algorithm, and the loss function is usually cross-entropy loss to maximize the consistency between the predicted probability and the true label.

[0093] This step primarily transforms the model's final output into interpretable flame state probabilities, such as under complex operating conditions like ignition, flameout, and load changes. The output value determines whether the flame is currently in a stable state. Combined with actual gas turbine operating data, this step can support second-level inference, meeting the real-time and reliability requirements of industrial control systems.

[0094] The technical effect of this step is that, through Softmax normalization, it ensures that the output value is within the range of... The sum of all probabilities within the specified interval is 1, thus forming an effective probability distribution that facilitates subsequent classification decisions and model evaluation. Furthermore, this step, in close conjunction with the attention mechanism and encoder-decoder structure, enhances the model's responsiveness to key features, improving the robustness and generalization performance of flame recognition.

[0095] S52, Set classification threshold ,when If the flame is present, it is determined to be present; otherwise, it is determined to be extinguished.

[0096] Specifically, in the technical solution of this invention, a classification threshold is set. This is one of the key steps in achieving accurate flame status identification. This step is based on the flame presence probability output by the model. By setting a reasonable threshold, continuous probability values ​​are mapped to binary classification results (flame presence or extinguishment), thereby enabling real-time judgment of the operating status of the combustion chamber of a heavy-duty gas turbine.

[0097] In this embodiment of the invention, the classification threshold The optimization can be determined using the validation set during the model training phase. Specifically, after model training is complete, the validation set data is used... The output is compared with the actual flame state labels. By calculating metrics such as precision, recall, and F1 score at different thresholds, the optimal result is selected. For example, it can be Set as ,when The system determines whether a flame is present or extinguished if the flame is present. This threshold setting must be considered in conjunction with the flame signal distribution characteristics under actual operating conditions to ensure good robustness under different loads, ignition stages, and unsteady operating conditions.

[0098] Furthermore, this step has significant application value in real-world industrial scenarios. In the combustion system of heavy-duty gas turbines, misjudgment of the flame state can lead to malfunctions in the control system, affecting equipment safety and operational efficiency. By setting... This can transform the soft classification results output by the model into hard decisions, facilitating integration with industrial control systems (such as DCS and PLC). In edge deployment scenarios, the computational complexity of this step is extremely low, requiring only a single comparison operation to determine the flame state, thus meeting real-time requirements.

[0099] Optionally, to enhance the model's adaptability under dynamic operating conditions, a dynamic threshold adjustment mechanism can be introduced, adjusting the threshold based on the current operating status (such as speed, load rate, etc.). Fine-tuning can be performed. For example, during the initial ignition phase, the threshold can be appropriately lowered to improve sensitivity, while during the stable operation phase, the threshold can be increased to reduce false alarms.

[0100] In summary, setting a classification threshold This step is not only a crucial part of model output post-processing, but also a key guarantee for achieving high accuracy, high robustness, and low latency in the flame state recognition system. By combining the output characteristics of the deep learning model with actual physical response laws, this step provides reliable technical support for the intelligent sensing and control of heavy-duty gas turbine combustion systems.

[0101] The flame intensity prediction method for heavy-duty gas turbine combustor based on attention mechanism in this invention can effectively integrate the temporal features of multi-source physical quantities, improve the accuracy and robustness of flame state recognition through the GRU encoder and attention decoder structure, adapt to complex working conditions and support real-time deployment of edge devices.

[0102] S6 performs condition balancing on the training samples, adjusting the sample weights according to the condition categories such as ignition, acceleration, and shutdown. ,in , Operating condition category The number of samples, This represents the total number of samples.

[0103] Specifically, in the technical solution of this invention, performing operating condition equalization processing on the training samples is one of the key steps to improve the model's generalization ability and recognition accuracy. The core of this step lies in adjusting the weights of the samples according to different operating condition categories commonly encountered during gas turbine operation, such as ignition, acceleration, and shutdown, to alleviate the model bias problem caused by uneven sample distribution. Specifically, for each operating condition category... Sample weights According to the formula Perform calculations, where This indicates the number of samples under this working condition category. This represents the total number of samples across all operating condition categories.

[0104] In this embodiment of the invention, the first step is to classify the collected multidimensional time series data into operating condition labels, and then classify the samples according to the gas turbine operating status (such as ignition, acceleration, shutdown, etc.). Subsequently, the number of samples under each operating condition category is counted. And calculate its corresponding weight. During model training, this weight is incorporated into the loss function, assigning different levels of influence to samples under different operating conditions. This guides the model to focus more on key operating condition categories with fewer samples during training. For example, when there are fewer samples in the ignition phase, increasing its weight allows the model to more sensitively capture the characteristics of flame intensity changes during ignition.

[0105] Furthermore, this weight adjustment mechanism ensures that the model maintains a high recognition accuracy even when facing imbalanced data. In actual tests of this invention, the model processed by this method achieved a flame recognition accuracy of over 97% in multi-condition scenarios, significantly outperforming the model without sample balancing. In addition, this method improves the model's deployment adaptability on edge computing devices because its computational complexity is low, involving only sample count statistics and simple division operations.

[0106] This step is widely applicable to flame status identification tasks in gas turbines during different operating phases (such as startup, shutdown, and load changes). Especially in scenarios involving sudden changes in flame intensity during the initial ignition phase or shutdown process, sample weight adjustment can effectively improve the model's ability to identify these critical moments, thereby enhancing the system's early warning and safety interlock control performance.

[0107] Furthermore, this load balancing step plays a dual role in the overall model architecture, serving both as data preprocessing and model training optimization. By rationally allocating sample weights, it not only improves the model's adaptability to various load conditions but also enhances the attention mechanism's ability to focus on key features, providing a high-quality input foundation for subsequent deep time-series modeling. Therefore, this step has significant technical value in improving model robustness and industrial applicability.

[0108] The method for predicting the flame intensity of a heavy-duty gas turbine combustor based on an attention mechanism in this invention introduces sample weights for different operating conditions. The training samples were balanced, which effectively alleviated the model bias problem caused by uneven sample distribution under different working conditions, and further improved the generalization ability and prediction stability of flame state recognition under key working conditions such as ignition, acceleration and shutdown.

[0109] Example 2 This invention relates to a flame intensity prediction model for heavy-duty gas turbine combustors based on an attention mechanism, such as... Figure 2 and Figure 3 As shown, the model specifically includes: The data input module is used to collect key physical quantity data during the operation of the gas turbine, including speed, compressor outlet temperature, exhaust temperature, combustion chamber pulsation value, etc., to construct a multi-dimensional time series input; the flame intensity signal is used as a supervision label for model training.

[0110] Specifically, multiple key physical quantities during gas turbine operation (such as speed, compressor outlet temperature, exhaust temperature, combustion chamber pulsation signal, etc.) are fused and processed, and temperature difference characteristics (such as the difference between compressor outlet temperature and gas turbine exhaust temperature), frequency band characteristics of combustion chamber pulsation value, and sliding window data sampling method of time sequence partitioning are introduced to form a multi-dimensional time-series feature input sample with physical meaning.

[0111] The data preprocessing module is used to perform feature fusion on the raw data (such as constructing temperature difference features and extracting pulsation features from frequency band division), and to perform standardization and sliding window sampling to form a structured dataset that can be used for training deep learning models.

[0112] The encoder module is used to construct an encoder using a multi-layer gated cyclic unit (GRU) to extract deep temporal semantic features of the input sequence.

[0113] In this embodiment of the invention, a deep temporal feature extraction structure based on a GRU encoder is proposed.

[0114] Specifically, an encoder structure composed of multi-layer gated recurrent units (GRUs) is constructed to effectively model the long-term dependency characteristics of gas turbine operating data, enhance the ability to model dynamic changes, and is more stable and accurate than RNN models, lighter and faster than LSTM models, and more suitable for edge deployment.

[0115] The decoder module with an attention mechanism consists of an embedding layer, multiple GRU layers, and an AdditiveAttention mechanism, enabling the focusing and decoding of key segments from the encoder output. The attention output is fused with label features and then fed into the decoder to guide prediction and implement forced instruction.

[0116] In this embodiment of the invention, an attention decoder structure based on Additive Attention is proposed.

[0117] Specifically, an additive attention mechanism is used to dynamically weight and focus the encoder output, extract feature information in key time periods, and enhance the stability and convergence speed of model training through a forced teaching mechanism.

[0118] Furthermore, an attention query mechanism is proposed to jointly construct the encoder output and decoder state.

[0119] By using the current hidden state of the decoder as the query vector and matching it with the historical state of the encoder as the key-value pair, a dynamic attention matrix is ​​formed for the flame state recognition task, enabling dynamic focusing of information at critical moments.

[0120] The output module is used to classify the decoding results through a multi-layer fully connected network to predict the state of the flame (such as whether a flame exists).

[0121] In this embodiment of the invention, a unified Encoder–Attention–Decoder structure is constructed to realize end-to-end modeling of the entire process from raw gas turbine operating data to flame state output, eliminating the need for traditional expert rules and manual threshold setting steps, and significantly improving the system's automation level.

[0122] Furthermore, the model structure considers inference efficiency, employing methods such as dimensionality enhancement through supervised label embedding layers, replacing LSTM with GRU modules, and adjusting hyperparameters based on actual experimental data. This effectively balances recognition performance with computational resource requirements, supporting deployment in industrial control systems or edge computing devices, and is expected to enable online flame recognition. It also provides a scalable network architecture applicable to scenarios such as continuous flame intensity regression prediction and anomaly detection, offering a unified modeling foundation for subsequent intelligent combustion control and anomaly early warning tasks in gas turbines.

[0123] Through the above methods, the present invention has strong time-series modeling capabilities and key feature focusing capabilities, and is suitable for ignition / quenching process detection and flame stability monitoring tasks, providing technical support for intelligent perception and control of heavy-duty gas turbine combustion systems.

[0124] The embodiments of the present invention have the following significant technical advantages: 1. Improve recognition accuracy and accelerate convergence.

[0125] This invention integrates multiple key operational features (including rotational speed, temperature difference, and pulsating pressure frequency band features), effectively eliminating redundant features and achieving perception of time evolution trends through sliding window modeling. It can accurately identify flame states in multiple operating scenarios (such as ignition, acceleration, and shutdown). Hyperparameter optimization based on actual experimental patterns and practical testing show an expected accuracy rate of over 97%. Furthermore, gradient pruning and random dropout regularization methods are used during training to further accelerate training convergence.

[0126] 2. Possesses the ability to perform time series modeling and focus on critical moments.

[0127] In this embodiment of the invention, the encoder adopts a gated cyclic unit (GRU) structure to effectively model the long-term time-series dependencies during the operation of the gas turbine. The attention mechanism of the decoder further focuses on key information before and after drastic changes in the flame, enabling the model to capture weak but predictive feature signals and significantly enhance the model's early warning capability.

[0128] 3. Achieve key feature fusion and physical constraint mapping.

[0129] In the data preprocessing stage, this invention abandons the traditional method of simultaneously standardizing multiple physical quantities and innovatively adopts heuristic physical feature fusion (such as temperature difference, frequency band energy distribution, etc.) to construct more representative and diagnostically sensitive input variables, thereby enhancing the model's generalization ability and interpretability.

[0130] 4. It adapts to changes in actual working conditions and has strong model transfer capabilities.

[0131] In this embodiment of the invention, by using a sliding window method and a sample equalization mechanism, the model can maintain stable recognition capability under different operating conditions (including partial load, non-steady state, instantaneous fluctuations, etc.), and has good industrial adaptability.

[0132] 5. It has edge inference deployment capability and supports real-time identification.

[0133] In this embodiment of the invention, by lightweighting the model structure (such as replacing LSTM with GRU structure, optimizing attention calculation, and simplifying the output module), the trained model can be deployed to edge computing devices or industrial control systems to achieve flame state inference in seconds, meet the requirements of real-time on-site detection and safety interlock control, and overcome the shortcomings of existing models that rely on GPUs and have high deployment costs.

[0134] 6. It has strong compatibility and scalability, and supports multi-task joint modeling.

[0135] The embodiments of this invention have a flexible structure and can be expanded to tasks such as continuous flame intensity regression prediction and anomaly detection (based on model residuals), providing a unified modeling framework for subsequent intelligent diagnosis of combustion systems.

[0136] Example 3 This invention also provides an attention mechanism-based device for predicting the flame intensity of a heavy-duty gas turbine combustor, such as... Figure 4 As shown, the device 10 includes: The multi-source data acquisition module 100 is used to acquire multi-source physical quantity data during the operation of the gas turbine, including speed, compressor outlet temperature, exhaust temperature and combustion chamber pulsation value, and to construct a multi-dimensional time series input containing physical feature coupling relationships; The feature fusion processing module 200 is used to perform feature fusion processing on the multidimensional time series data, extract temperature difference features, pulsation frequency band features and sliding window data partitioned by time order, and form a standardized structured dataset. The encoder construction module 300 is used to construct an encoder using a multi-layer gated recurrent unit (GRU), perform deep temporal feature extraction on the structured dataset, and generate encoder output features. The decoder construction module 400 is used to construct a decoder with an attention mechanism based on the encoder output features. It dynamically weights and focuses on key time steps and key channel features through the Additive Attention mechanism, and fuses the attention output with the supervision label features to guide the prediction process. The classification processing module 500 is used to classify the decoder output through a multi-layer fully connected network, output flame state prediction results, and realize real-time identification of the presence of flames.

[0137] Example 4 To implement the methods of the above embodiments, the present invention also provides a computer device, which includes a memory and a processor; wherein the processor runs a program corresponding to the executable program code by reading executable program code stored in the memory, so as to implement the various steps of the methods described above.

[0138] Example 5 To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in the foregoing embodiments.

[0139] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0140] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0141] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for predicting the flame intensity of a heavy-duty gas turbine combustor based on an attention mechanism, characterized in that, Includes the following steps: S1 collects multi-source physical quantity data during the operation of the gas turbine, including speed, compressor outlet temperature, exhaust temperature and combustion chamber pulsation value, and constructs a multi-dimensional time series input containing physical feature coupling relationships; S2, perform feature fusion processing on the multidimensional time series data to extract temperature difference features, pulsation frequency band features and sliding window data partitioned by time order, forming a standardized structured dataset; S3, an encoder is constructed using a multi-layer gated recurrent unit (GRU) to perform deep temporal feature extraction on the structured dataset and generate encoder output features; S4. Construct a decoder with an attention mechanism based on the encoder output features. Dynamically weight and focus on key time steps and key channel features through the Additive Attention mechanism, and fuse the attention output with the supervision label features to guide the prediction process. S5 uses a multi-layer fully connected network to classify the decoder output and output flame state prediction results, enabling real-time identification of the presence of flames.

2. The method according to claim 1, characterized in that, The process of feature fusion processing on the multidimensional time series data extracts temperature difference features, pulsation frequency band features, and sliding window data partitioned by time order to form a standardized structured dataset, including: S21, calculate the difference between the compressor outlet temperature and the exhaust temperature as the temperature difference feature, and segment the time series using a sliding window with a window length of [value missing]. And the step size is ,in and Adjustments were made based on the response characteristics of the actual test data; S22, divide the combustion chamber pulsation value into frequency bands and extract... to The energy distribution characteristics of each frequency band are analyzed, and the energy of each frequency band is normalized to... Interval.

3. The method according to claim 1, characterized in that, The encoder, constructed using a multi-layer gated recurrent unit (GRU), performs deep temporal feature extraction on the structured dataset to generate encoder output features, including: S31, setting encoder to include A layer of GRU units, where the hidden state dimension of each layer of GRU is... ,in and Determined through response patterns based on actual experimental data; S32, perform gradient clipping on the encoder output features, with a clipping threshold of [value missing]. And it uses a random discarding technique with a discard probability of . .

4. The method according to claim 1, characterized in that, The step of constructing a decoder with an attention mechanism based on the encoder output features, and dynamically weighting and focusing on key time steps and key channel features through an Additive Attention mechanism, includes: S41, Calculate attention weights At that time, use the decoder's current hidden state. encoder history status The dot product is given by the formula: ; S42 concatenates the attention output with the supervised label features, and generates a fused feature vector through a fully connected layer. ,in For attention-weighted features, These are label features.

5. The method according to claim 1, characterized in that, The process of classifying the decoder output through a multi-layer fully connected network to output flame state prediction results includes: S51 uses the Softmax activation function to normalize the output of the fully connected layer, resulting in the output probability distribution. ,in This is the weight matrix. It is the bias vector; S52, Set classification threshold ,when If the flame is present, it is determined to be present; otherwise, it is determined to be extinguished.

6. The method according to claim 1, characterized in that, Also includes: S6 performs condition balancing on the training samples, adjusting the sample weights according to the condition categories such as ignition, acceleration, and shutdown. ,in , Operating condition category The number of samples, This represents the total number of samples.

7. A heavy-duty gas turbine combustor flame intensity prediction device based on an attention mechanism, characterized in that, include: The multi-source data acquisition module is used to collect multi-source physical quantity data during the operation of the gas turbine, including speed, compressor outlet temperature, exhaust temperature and combustion chamber pulsation value, and to construct a multi-dimensional time series input containing physical feature coupling relationships; The feature fusion processing module is used to perform feature fusion processing on the multidimensional time series data, extract temperature difference features, pulsation frequency band features and sliding window data partitioned by time order, and form a standardized structured dataset. The encoder construction module is used to construct an encoder using a multi-layer gated recurrent unit (GRU), perform deep temporal feature extraction on the structured dataset, and generate encoder output features. The decoder construction module is used to construct a decoder with an attention mechanism based on the encoder output features. It dynamically weights and focuses on key time steps and key channel features through the Additive Attention mechanism, and fuses the attention output with the supervision label features to guide the prediction process. The classification processing module is used to classify the decoder output through a multi-layer fully connected network, output flame state prediction results, and realize real-time identification of the presence of flames.

8. A computer device, characterized in that, Including processor and memory; The processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement the heavy-duty gas turbine combustor flame intensity prediction method based on the attention mechanism as described in any one of claims 1-6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method for predicting the flame intensity of a heavy-duty gas turbine combustor based on an attention mechanism as described in any one of claims 1-6.

Images