Combustion evaluation method and system based on physical enhancement residual and working condition decoupling

Abstract

The application discloses a gas turbine combustion evaluation method and system based on physical enhancement residual and working condition decoupling, and relates to the technical field of gas turbine monitoring and diagnosis. The method comprises the following steps: collecting multi-source heterogeneous monitoring data of a gas turbine, cleaning the data, and dividing stable combustion segments; performing working condition clustering on the segments by taking load as a core parameter, constructing a combustion state evaluation background of different working conditions, and realizing working condition decoupling; constructing a physical enhancement auto-encoder model integrating data reconstruction and physical mechanism constraints, training the model by using historical health data to establish a baseline combustion mode; inputting a segment to be evaluated into the trained model, calculating the total loss residual of the model output and actual characteristics, mapping the total loss residual into a combustion health index, combining time series analysis to analyze the change trend of the combustion health index, and outputting a combustion state evaluation result. The application enhances the robustness of the model through physical mechanism, eliminates the interference of differences in operating conditions through working condition decoupling, and significantly improves the combustion state monitoring precision and early fault recognition ability under variable working conditions.

Description

Technical Field

[0001] This invention relates to the field of gas turbine monitoring and diagnostic technology, specifically to a gas turbine combustion evaluation method and system based on physical enhancement of residuals and decoupling of operating conditions. Background Technology

[0002] As a core power equipment in fields such as power, aviation, and shipbuilding, the combustion state of a gas turbine's combustion chamber directly determines the overall operating efficiency, stability, and service life of the unit. However, the combustion process involves complex physicochemical reactions and is affected by multiple factors, including load fluctuations, changes in fuel quality, environmental disturbances, and component aging and degradation. Consequently, the combustion state exhibits significant time-varying and nonlinear characteristics. In actual operation, typical precursors to failures such as combustion instability, excessive acceleration, and abnormal dynamic pressure are highly likely to occur. If the trend of combustion state degradation cannot be detected and accurately identified in a timely manner over time, it may lead to serious safety accidents such as combustion chamber erosion, excessive rotor vibration, or even unplanned shutdowns, causing significant economic losses and safety hazards.

[0003] Currently, the combustion status assessment methods for gas turbines still suffer from the following significant technical shortcomings in engineering applications: (1) Existing methods mostly rely on monitoring data at a single moment for independent judgment, lacking the mining and utilization of the evolution law of combustion status in the time dimension, making it difficult to capture early degradation trends; (2) Anomaly judgment generally adopts a fixed threshold alarm mechanism, which cannot adaptively adjust the judgment criteria according to changes in operating conditions, resulting in a high false alarm rate and false negative rate under varying operating conditions, making it difficult to achieve early warning and accurate location of combustion anomalies. Therefore, there is an urgent need for a combustion status assessment method that can take into account both the mining of degradation characteristics in the time dimension and adaptive judgment under varying operating conditions. Summary of the Invention

[0004] The technical problem to be solved by this invention is how to accurately quantify and evaluate the combustion state of a gas turbine under varying operating conditions.

[0005] This invention solves the above-mentioned technical problems through the following technical means: a gas turbine combustion evaluation method based on physical enhancement residual and decoupling of operating conditions, comprising:

[0006] S1. Collect multi-source heterogeneous monitoring data during the operation of the gas turbine, clean the multi-source heterogeneous monitoring data, and divide it into steady-state combustion segments; S2. Based on load data as operating condition parameters, cluster analysis and standardization processing are performed on the steady-state combustion segment to construct a combustion state evaluation background for different operating conditions. S3. Construct a physical enhancement autoencoder evaluation model that integrates data reconstruction and physical mechanism constraints. Use historical health data to train the evaluation model with the goal of minimizing the total loss. The total loss includes a reconstruction loss term and a physical constraint term. This is to establish a benchmark combustion mode corresponding to the sub-conditions. S4. Map the segment to be evaluated to the combustion state evaluation background of the sub-condition to determine the corresponding target condition interval, input the segment to be evaluated into the trained model corresponding to the target condition interval, calculate the total loss residual between the model output and the segment to be evaluated, map the total loss residual into the combustion health index and output the evaluation result.

[0007] Furthermore, S2 includes: Based on the number of steady-state combustion segments Dynamically configure the number of clusters in the K-Means clustering algorithm : like <First preset sample threshold> If not, then the clustering operation will not be performed, let The entire operating space is divided into single working condition zones; When the first preset sample threshold ≤ Second preset sample threshold season The operating space is dynamically divided into low-load and high-load zones; when ≥Second preset sample threshold season The operating space is dynamically divided into low-load, medium-load, and high-load zones; Wherein, the first preset sample threshold and the second preset sample threshold Adaptive settings are made based on the gas turbine model and the amount of historical data. Before performing clustering, the load data of the steady-state combustion segment are standardized using a standardization algorithm based on the quantile scaling principle. After completing dynamic clustering, each steady-state combustion segment is mapped to the corresponding load range to construct a combustion state evaluation background for different operating conditions.

[0008] Furthermore, the multi-source heterogeneous monitoring data includes: load data, combustion chamber vibration acceleration data, and combustion chamber dynamic pressure data.

[0009] Furthermore, the formula for calculating the total loss function of the physical enhancement autoencoder evaluation model is as follows:

[0010] in, To reconstruct the loss term, For physical constraint terms, To reconstruct the weights, For physical weights.

[0011] Furthermore, the calculation formula for the physical constraint term is as follows:

[0012] in, The coupling operator for acceleration and pressure is calculated as follows:

[0013] in, For the mutual power of acceleration and dynamic pressure, The average energy of the acceleration signal. The average energy of the dynamic pressure signal; The physical reference mapping function is calculated using the following formula:

[0014] in, , and The coefficients of the curve polynomial, For load.

[0015] Furthermore, the formula for calculating the combustion health index is as follows:

[0016] in, The largest reconstructed residual in historical health data, This is the preset reliability correction factor.

[0017] Furthermore, based on the combustion health index, corresponding health levels are divided, specifically as follows: Let the threshold for the combustion health index be... , , and ,and ; when If so, the health level is judged to be excellent; when If so, the health level is determined to be good; when If this occurs, the health level is determined to be average, and a warning is triggered. when If the condition is met, the health level is determined to be poor, and an equipment inspection command is generated. when If the condition is found to be dangerous, the equipment will be shut down for protection.

[0018] This invention also provides a gas turbine combustion evaluation system based on physical enhancement of residuals and decoupling of operating conditions, comprising: The data acquisition module is used to collect multi-source heterogeneous monitoring data during the operation of the gas turbine, clean the multi-source heterogeneous monitoring data, and divide it into steady-state combustion segments; The operating condition clustering module is used to perform cluster analysis and standardization processing on the steady-state combustion segment based on load data as operating condition parameters, and to construct a combustion state evaluation background for different operating conditions. The model building and training module is used to build a physical enhancement autoencoder evaluation model that integrates data reconstruction and physical mechanism constraints. It uses historical health data to train the evaluation model with the goal of minimizing the total loss, which includes a reconstruction loss term and a physical constraint term, in order to establish a benchmark combustion mode corresponding to the sub-condition. The state assessment and output module is used to map the segment to be assessed to the combustion state evaluation background of the sub-condition to determine the corresponding target condition interval, input the segment to be assessed into the trained model corresponding to the target condition interval, calculate the total loss residual between the model output and the segment to be assessed, map the total loss residual into the combustion health index and output the assessment result.

[0019] The present invention also provides a processing device, including at least one processor and at least one memory communicatively connected to the processor, wherein: the memory stores program instructions executable by the processor, and the processor can execute the above-described method steps by calling the program instructions.

[0020] The present invention also provides a computer-readable storage medium storing computer instructions that cause the computer to perform the above-described method steps.

[0021] The advantages of this invention are as follows: By collecting full-cycle load and combustion characteristic data of gas turbines, using load as a discriminant parameter, a dynamic clustering algorithm is used to divide typical operating condition intervals and construct a steady-state combustion benchmark manifold for each operating condition. The interaction characteristics of acceleration and dynamic pressure are extracted within the same load zone. The deviation of the combustion system from the benchmark manifold is quantified by calculating the nonlinear reconstruction residual between actual operating data and the benchmark manifold. Combined with time series analysis, a full-cycle assessment of the evolution of combustion state from steady state to instability is achieved. This invention effectively eliminates the direct interference of load fluctuations on combustion health evaluation, strengthens the causal correlation between evaluation indicators and physical processes such as combustion backfire and thermoacoustic oscillation, and possesses both higher evaluation accuracy and early warning capability, providing reliable technical support for the safe operation of gas turbines. Attached Figure Description

[0022] Figure 1 This is a flowchart of the gas turbine combustion evaluation method based on physical enhancement residual and decoupling of operating conditions according to Embodiment 1 of the present invention; Figure 2 This is a 3D clustering diagram of combustion conditions under different loads in Embodiment 1 of the present invention; Figure 3(a) is a trend graph of the combustion health index of the method of Embodiment 1 of the present invention under low load; Figure 3(b) is a trend chart of the combustion health index of the method of Embodiment 1 of the present invention under medium load; Figure 3(c) is a trend graph of the combustion health index of the method of Embodiment 1 of the present invention under high load. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] like Figure 1 As shown, the gas turbine combustion evaluation method based on physical enhancement residuals and decoupling from operating conditions includes: S1. Collect multi-source heterogeneous monitoring data during the operation of the gas turbine, clean the multi-source heterogeneous monitoring data, and divide it into steady-state combustion segments.

[0025] Specifically, a communication connection is established with the gas turbine's control system (TCS) via a data communication interface to acquire multi-source heterogeneous monitoring data during gas turbine operation. This multi-source heterogeneous monitoring data includes at least: load data characterizing output power, multi-channel vibration acceleration data of the combustion chamber, and multi-channel dynamic pressure data of the combustion chamber. The acquired multi-source heterogeneous monitoring data is then cleaned to remove transient invalid data during gas turbine startup, shutdown, and low-load phases, obtaining valid monitoring data. For example, a load filtering threshold of 100 (adjustable according to the specific turbine model) is set to filter out invalid low-load data with load values ​​less than 100, retaining valid monitoring data under high-load conditions to eliminate interference periods caused by combustion instability.

[0026] Based on a preset time window (e.g., 5 minutes), the effective monitoring data is divided into multiple consecutive time windows. For each time window, the load fluctuation coefficient within that window is determined. The load fluctuation coefficient is calculated as follows: the standard deviation and mean of the load data within that window are obtained, and the ratio of the standard deviation to the mean of the load data is multiplied by 100 to obtain the load fluctuation coefficient.

[0027] The number of valid data points within the window is verified, and the continuous time period corresponding to the time window where the load fluctuation coefficient is less than or equal to the first preset threshold (e.g., 5) and the number of valid data points is greater than or equal to the second preset threshold (e.g., 5) is marked as a steady-state combustion segment.

[0028] Extract the reference time corresponding to the steady-state combustion segment, and assign a time label to each steady-state combustion segment in combination with the time window offset; the time label includes the specific time and the month to which it belongs.

[0029] Based on the time label, the steady-state combustion segments and their corresponding vibration acceleration data and dynamic pressure data are sliced ​​and integrated to construct a steady-state sample set with a time dimension.

[0030] S2. Based on load data as operating condition parameters, cluster analysis and standardization are performed on steady-state combustion segments to construct a combustion state evaluation background for different operating conditions.

[0031] Specifically, using load data as the core operating condition parameter, the load mean of each steady-state combustion segment is extracted. To reduce the interference of abnormal load fluctuations on cluster centers during gas turbine peak shaving, a standardization algorithm based on quantile scaling is used to standardize the load mean preprocessing. This embodiment preferentially uses the RobustScaler algorithm to improve the robustness of operating condition division; such as Figure 2 As shown, after standardization, based on the number of steady-state combustion segments... Dynamically configure the number of clusters in the K-Means clustering algorithm The specific rule is: when the number of steady-state combustion fragments When no clustering operation is performed, the entire operating space is divided into a single operating condition area; when At that time, the operating space is dynamically divided into two zones: low load and high load; when The operating space is dynamically divided into three independent intervals: low load, medium load, and high load. During K-Means clustering, to improve the stability of the clustering results and avoid getting trapped in local optima, a multiple random initialization strategy is adopted, with a fixed random seed ensuring reproducibility. In the specific implementation, the number of random initializations can be set to 30. After completing the dynamic clustering, each steady-state combustion segment is mapped to its corresponding load interval, thereby constructing a combustion state evaluation background for different operating conditions and suppressing the nonlinear interference introduced by load variations on the combustion state assessment.

[0032] S3. Construct a Physics-Enhanced Autoencoder (PE-AE) evaluation model that integrates data reconstruction and physical mechanism constraints. Use historical health data to train the evaluation model with the goal of minimizing the total loss, which includes a reconstruction loss term and a physical constraint term; to establish a benchmark combustion mode corresponding to the sub-condition.

[0033] Specifically, unlike traditional data-driven models, this method constructs a physically enhanced autoencoder evaluation model PE-AE, embedding the physical laws of thermoacoustic coupling into the model training process to solve the problem of inaccurate evaluation under varying operating conditions.

[0034] The PE-AE model adopts a symmetrical fully connected encoder-decoder structure.

[0035] Input layer: The input data dimension is set to 5, which is composed of a 3D acceleration signal (ACC) and a 2D dynamic pressure signal (HUM), i.e., Input Dimension=[ACC×3, HUM×2].

[0036] Network structure: The encoder consists of three fully connected layers with 5, 8, and 4 neurons, respectively, used to progressively reduce the dimensionality of the high-dimensional input and extract deep coupling features; the decoder also consists of three fully connected layers with 4, 8, and 5 neurons, respectively, used to reconstruct the original acceleration and dynamic pressure signals from the feature space.

[0037] Activation function: All hidden layers use the ReLU function to introduce non-linear expressiveness and alleviate the gradient vanishing problem.

[0038] Data preprocessing: Before inputting the data into the model, the Min-Max Scaling method is used to normalize all input feature data to the [0, 1] interval to eliminate the negative impact of different physical units on model training.

[0039] (1) The training objective of PE-AE is to minimize the total loss function. The calculation formula is as follows:

[0040] in, To reconstruct the weights, we control the model's ability to reproduce the original combustion features. If the weights are too large, the model will become a regular autoencoder, easily ignoring weak physical evolution signals. These are physical weights that control the model's sensitivity to whether the coupling relationship between acceleration and pressure conforms to physical benchmarks. If... Larger values ​​will enhance the model's ability to capture physical deviations such as tempering or oscillations.

[0041] This is a reconstruction loss term used to quantify the model's ability to map and restore the original input combustion feature data.

[0042] Specifically, the autoencoder maps the high-dimensional combustion feature vector to a low-dimensional latent space through the encoder, and then the decoder recovers the original features. The reconstruction loss term is calculated using the mean squared error (MSE), and its formula is as follows:

[0043] in, The dimension of the input feature vector. For the first The original input feature values ​​of each channel, including the mean acceleration or the mean dynamic pressure, etc. The corresponding feature values ​​output by the model reconstruction.

[0044] The physical constraint term is defined as a coupling operator. Mapping function with reference Sum of squared residuals:

[0045] The physical constraint term essentially transforms the physical laws of thermo-acoustic coupling in the combustion system into mathematical supervision signals for a deep learning model. It aims to embed the thermo-acoustic coupling laws in combustion dynamics into the parameter optimization process of a deep neural network in the form of mathematical penalty terms.

[0046] In the actual operation of gas turbines, the vibration and pressure signals of the combustion chamber exhibit strong nonlinear characteristics due to the influence of environmental parameters such as load changes and ambient temperature. Traditional pure data-driven models often struggle to distinguish between normal operating condition fluctuations and potential combustion anomalies.

[0047] The physical constraint term, by introducing a physical baseline mapping function built based on historical health data, provides the model with a dynamically changing physical distribution baseline that varies with the load. Its functions are as follows: Operating condition decoupling constraint: It is mandatory that the feature vectors reconstructed by the model under different power levels must conform to the preset physical mapping logic, thereby achieving active decoupling of combustion characteristics and load fluctuations in the feature extraction stage.

[0048] Physical consistency supervision: By calculating the sum of squared residuals between the coupling operator generated by the model reconstruction and the benchmark mapping function, it is ensured that the autoencoder does not deviate from the essential physical law of thermo-acoustic coupling during the process of compressing and reconstructing information.

[0049] Amplification of abnormal features: When the system undergoes evolution such as backfire or thermoacoustic oscillation, the interaction characteristics of the signals violate the physical consistency constraints mastered by the model during the training phase, which leads to a nonlinear surge in the proportion of physical constraint terms in the total loss function.

[0050] (2) Coupling Operator

[0051] To quantify the reconstructed acceleration With dynamic pressure The physical correlation. Define the energy cross-correlation operator:

[0052] in, The cross-power of acceleration and dynamic pressure measures the correlation between the signals in the time domain. This value increases significantly if the two signals are phase-synchronized and have strong energy interaction. This directly corresponds to the energy transfer efficiency in thermoacoustic coupling. The average energy of the acceleration signal is represented by the second moment of the signal, reflecting the total power level of the body vibration within this time segment. The average energy of the dynamic pressure signal represents the second moment of the acoustic pulsation signal, reflecting the level of thermal energy release fluctuations in the acoustic field inside the combustion chamber. This value characterizes the physical efficiency of the conversion of combustion pulsation energy into machine characteristics and is the core characteristic of thermo-acoustic coupling strength. The coupling strength is limited to the interval [-1, 1] by using the square root term of the denominator. This eliminates the influence of the absolute signal amplitude (i.e., decouples the amplitude fluctuations caused by the load size) and only reflects the physical tightness of the thermo-acoustic coupling.

[0053] (3) Physical reference mapping function Construction For different load ranges, using historical data from normal operation, a polynomial regression was used to fit the relationship curve between load and coupling strength:

[0054] in, , and The coefficients of the curve polynomial, For load. , and The coefficients were obtained by training on historical health data using the least squares method. This function provides physical standard values ​​for PE-AE at different load points.

[0055] In actual operation of gas turbines, vibration and pressure signals exhibit strong nonlinear characteristics. To ensure the model both grasps the basic distribution and conforms to physical laws, the following training strategy is adopted: The Adam optimizer is selected. The initial learning rate is set to 0.01. The total number of training epochs is set to 150. In the initial stage of model training, the following settings are implemented: =1, =0.1. This allows the model to grasp the basic distribution of combustion characteristics, i.e., prioritizes reducing the reconstruction loss term. To address the need for early warning, gradually increase... The proportion of this. This way, even if the signal amplitude doesn't change much, as long as its coupling characteristics violate thermoacoustic physics, the total loss function L will jump drastically, thus improving the sensitivity of anomaly detection. In the initial training phase, you can first set... =0, allowing the model to quickly grasp the basic working condition distribution; to be continued After stabilizing, gradually increase the size. Physical constraints are introduced for fine-tuning.

[0056] S4. Map the segment to be evaluated to the combustion state evaluation background of the sub-condition to determine the corresponding target condition interval. Input the segment to be evaluated into the trained model corresponding to the target condition interval. Calculate the total loss residual between the model output and the segment to be evaluated. Map the total loss residual to the combustion health index and output the evaluation result.

[0057] Specifically, in the model inference stage, the combustion health index of the real-time segment is determined by the total loss generated by PE-AE in the physical residual-based health quantification assessment.

[0058] When a flashback occurs in a generator unit, the amplitude of a single signal may still be within the normal range, but the phase / energy relationship between acceleration and pressure will deviate significantly from the physical reference mapping function due to the change in flame position. Therefore, by calculating the total loss... After performing reverse normalization, a quantified combustion health index can be obtained. The formula for calculating the combustion health index is:

[0059] in, The largest reconstructed residual in historical health data, This is a preset reliability correction coefficient used to balance the random fluctuations in residuals caused by occasional noise in the combustion chamber, and to avoid non-physical negative values ​​in the health index; in this embodiment, The value range is 1.2-1.8.

[0060] Health Level Classification: Based on the combustion health index value, the combustion state of gas turbines is divided into 5 quantitative health levels. The specific classification criteria are as follows: excellent: This indicates that the combustion state was excellent during this period, and all indicators were at their optimal levels.

[0061] good: This indicates that the combustion status was good during this period, with no obvious abnormalities and only slight fluctuations in some indicators.

[0062] generally: This indicates that the combustion status was generally normal during this period, but there were some issues with large fluctuations in certain indicators, which require attention.

[0063] Difference: This indicates that the combustion state was poor during this period, and there was a significant abnormality that required timely inspection.

[0064] Danger: This indicates that the combustion condition was extremely poor during this period, posing a serious risk of malfunction, and requiring immediate shutdown and maintenance.

[0065] Output results: Output a quantitative assessment sample set containing combustion health index and health level, realizing quantitative grading assessment of combustion status over time.

[0066] The trained PE-AE model is used to perform physical consistency detection on real-time monitoring segments: (1) Calculation of total loss function: Input the segment to be evaluated into the PE-AE model and simultaneously obtain its data reconstruction loss term and physical constraint term. When the combustion undergoes thermoacoustic instability evolution, the physical constraint loss will increase nonlinearly because the signal characteristics violate the thermoacoustic coupling physical law mastered by the model during the training phase.

[0067] (2) Exception triggering mechanism design: Normal state determination: Real-time monitoring of total loss The numerical evolution of . If If the combustion state is within the residual probability distribution range of historical healthy samples (95% confidence interval) and the deviation between the coupling operator and the physical baseline mapping function does not exceed the preset physical constraint boundary, then the combustion state at that moment is determined to be normal.

[0068] Abnormal state identification: If the energy interaction relationship between the current acceleration and dynamic pressure has deviated from the reference manifold of steady-state operation, then this period is determined to be a combustion abnormality point.

[0069] Anomaly Quantification Mapping: Mapping the total loss Mapped to a combustion health index of 0-100, total loss The higher the number, the lower the health index.

[0070] This embodiment focuses on an Ansaldo AE94.3A gas turbine unit that has been in operation for nearly six years. The unit employs an annular combustion chamber structure with 24 uniformly distributed mixing STD burners along its circumference. Under the complex operating environment of long-term high temperature and high pressure, the unit has exhibited typical combustion stability problems, primarily manifested as fluctuations in dynamic pressure (humming) measurement points within the combustion chamber, large cylinder acceleration, excessive NOx emissions, and significant dispersion in exhaust gas temperature. Overhaul of the unit revealed corrosion and scaling on the exterior of the burners and inside the fuel passages.

[0071] The dataset was extracted from the unit's full historical operating data for the entire year preceding the cooling hole blockage event. To cover atmospheric temperature drift and complex operating conditions caused by cross-seasonal operation, typical operating month segments, including April-May, July-August, and October-December 2024, were selected. This long-period sampling strategy ensures that the evaluation model can learn the physical steady-state baseline of the combustion system under different environmental conditions.

[0072] Figures 3(a)-3(c) show the PE-AE model evaluation results for different load ranges. Analysis of the results shows that during the October-December operating period, the combustion health index under high load conditions had already reached or exceeded the threshold of 60, and there was also a significant deterioration trend under low load conditions. At this time, the existing alarm system of the gas turbine had not yet issued a hard alarm for dynamic pressure or exhaust temperature deviations. The high physical constraint terms extracted by the PE-AE model essentially reveal that the acoustic-solid energy distribution ratio inside the combustion chamber had seriously deviated from the healthy manifold defined by the Rayleigh criterion.

[0073] Example 2 Based on Example 1, Example 2 also provides a gas turbine combustion evaluation system based on physical enhancement residuals and decoupling of operating conditions, including: The data acquisition module is used to collect multi-source heterogeneous monitoring data during the operation of the gas turbine, clean the multi-source heterogeneous monitoring data, and divide it into steady-state combustion segments.

[0074] Specifically, the multi-source heterogeneous monitoring data includes: load data, combustion chamber vibration acceleration data, and combustion chamber dynamic pressure data.

[0075] The operating condition clustering module is used to perform cluster analysis and standardization on steady-state combustion segments based on load data as operating condition parameters, and to construct a combustion state evaluation background for different operating conditions.

[0076] Specifically, the dynamic cluster number configuration unit is used to configure the number of steady-state combustion segments. Dynamically configure the number of clusters in the K-Means clustering algorithm : like <First preset sample threshold> If not, then the clustering operation will not be performed, let The entire operating space is divided into single working condition zones.

[0077] When the first preset sample threshold ≤ Second preset sample threshold season The operating space is dynamically divided into low-load and high-load zones.

[0078] when ≥Second preset sample threshold season The operating space is dynamically divided into low-load, medium-load, and high-load zones.

[0079] Among them, the first preset sample threshold Second preset sample threshold Adaptive settings are made based on the gas turbine model and historical data.

[0080] The preprocessing unit is used to standardize the load data of the steady-state combustion segment using a standardization algorithm based on the quantile scaling principle before performing clustering operations.

[0081] The operating condition mapping and construction unit is used to map each steady-state combustion segment to the corresponding load range after dynamic clustering is completed, and to construct the combustion state evaluation background for each operating condition.

[0082] The model building and training module is used to build a physical enhancement autoencoder evaluation model that integrates data reconstruction and physical mechanism constraints. It uses historical health data to train the evaluation model with the goal of minimizing the total loss, which includes a reconstruction loss term and a physical constraint term, in order to establish a benchmark combustion mode corresponding to the sub-conditions.

[0083] Specifically, the formula for calculating the total loss function of the physical augmented autoencoder evaluation model is as follows:

[0084] in, To reconstruct the loss term, For physical constraint terms, To reconstruct the weights, For physical weights.

[0085] The formula for calculating the physical constraint term is:

[0086] in, The coupling operator for acceleration and pressure is calculated as follows:

[0087] in, For the mutual power of acceleration and dynamic pressure, The average energy of the acceleration signal. This represents the average energy of the dynamic pressure signal.

[0088] The physical reference mapping function is calculated using the following formula:

[0089] in, , and The coefficients of the curve polynomial, For load.

[0090] The state assessment and output module is used to map the segment to be assessed to the combustion state evaluation background of the sub-condition to determine the corresponding target condition range, input the segment to be assessed into the trained model corresponding to the target condition range, calculate the total loss residual between the model output and the segment to be assessed, map the total loss residual into the combustion health index and output the assessment result.

[0091] Specifically, the formula for calculating the combustion health index is as follows:

[0092] in, The largest reconstructed residual in historical health data, This is the preset reliability correction factor.

[0093] The rating unit is used to classify the corresponding health level based on the combustion health index, specifically as follows: Let the threshold for the combustion health index be, , , and ,and .

[0094] when If the test result is positive, the health level is determined to be excellent.

[0095] when If the condition is as described above, the health level is determined to be good.

[0096] when If the condition is not met, the health level will be determined as "normal," and a warning will be triggered.

[0097] when If the condition is not met, the health level is determined to be poor, and an equipment inspection command is generated.

[0098] when If the condition is found to be dangerous, the equipment will be shut down for protection.

[0099] Example 3 Based on Embodiment 1, Embodiment 3 of the present invention also provides a processing device, including at least one processor and at least one memory communicatively connected to the processor, wherein: the memory stores program instructions executable by the processor, and the processor can execute the method steps of Embodiment 1 by calling the program instructions.

[0100] Example 4 Based on Embodiment 1, Embodiment 4 of the present invention also provides a computer-readable storage medium storing computer instructions that cause the computer to perform the steps of the method described in Embodiment 1.

[0101] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A gas turbine combustion evaluation method based on physical enhancement residuals and decoupling of operating conditions, characterized in that, include: S1. Collect multi-source heterogeneous monitoring data during the operation of the gas turbine, clean the multi-source heterogeneous monitoring data, and divide it into steady-state combustion segments; S2. Based on load data as operating condition parameters, cluster analysis and standardization processing are performed on the steady-state combustion segment to construct a combustion state evaluation background for different operating conditions. S3. Construct a physical augmented autoencoder evaluation model that integrates data reconstruction and physical mechanism constraints. Use historical health data to train the evaluation model with the goal of minimizing the total loss. The total loss includes a reconstruction loss term and a physical constraint term. To establish a reference combustion mode corresponding to the aforementioned operating conditions; S4. Map the segment to be evaluated to the combustion state evaluation background of the sub-condition to determine the corresponding target condition interval, input the segment to be evaluated into the trained model corresponding to the target condition interval, calculate the total loss residual between the model output and the segment to be evaluated, map the total loss residual into the combustion health index and output the evaluation result.

2. The gas turbine combustion evaluation method based on physical enhancement residual and decoupling of operating conditions according to claim 1, characterized in that, S2 includes: Based on the number of steady-state combustion segments Dynamically configure the number of clusters in the K-Means clustering algorithm : like <First preset sample threshold> If not, then the clustering operation will not be performed, let The entire operating space is divided into single working condition zones; When the first preset sample threshold ≤ Second preset sample threshold season The operating space is dynamically divided into low-load and high-load zones; when ≥Second preset sample threshold season The operating space is dynamically divided into low-load, medium-load, and high-load zones; Wherein, the first preset sample threshold and the second preset sample threshold Adaptive settings are made based on the gas turbine model and the amount of historical data. Before performing clustering, the load data of the steady-state combustion segment are standardized using a standardization algorithm based on the quantile scaling principle. After completing dynamic clustering, each steady-state combustion segment is mapped to the corresponding load range to construct a combustion state evaluation background for different operating conditions.

3. The gas turbine combustion evaluation method based on physical enhancement residual and decoupling of operating conditions according to claim 1, characterized in that, The multi-source heterogeneous monitoring data includes: load data, combustion chamber vibration acceleration data, and combustion chamber dynamic pressure data.

4. The gas turbine combustion evaluation method based on physical enhancement residual and decoupling of operating conditions according to claim 1, characterized in that, The formula for calculating the total loss function of the physical augmented autoencoder evaluation model is as follows: in, To reconstruct the loss term, For physical constraints, To reconstruct the weights, For physical weights.

5. The gas turbine combustion evaluation method based on physical enhancement residual and decoupling of operating conditions according to claim 4, characterized in that, The formula for calculating the physical constraint term is: in, The coupling operator for acceleration and pressure is calculated as follows: in, For the mutual power of acceleration and dynamic pressure, The average energy of the acceleration signal. The average energy of the dynamic pressure signal; The physical reference mapping function is calculated using the following formula: in, , and The coefficients of the curve polynomial, For load.

6. The gas turbine combustion evaluation method based on physical enhancement residual and decoupling of operating conditions according to claim 1, characterized in that, The formula for calculating the health index of combustion is: in, The largest reconstructed residual in historical health data, This is the preset reliability correction factor.

7. The gas turbine combustion evaluation method based on physical enhancement residual and decoupling of operating conditions according to claim 6, characterized in that, The corresponding health levels are determined based on the aforementioned combustion health index, specifically as follows: Let the threshold for the combustion health index be... , , and ,and ; when If so, the health level is judged to be excellent; when If so, the health level is determined to be good; when If this occurs, the health level is determined to be average, and a warning is triggered. when If the condition is met, the health level is determined to be poor, and an equipment inspection command is generated. when If the condition is found to be dangerous, the equipment will be shut down for protection.

8. A gas turbine combustion evaluation system based on physical enhancement residuals and decoupling of operating conditions, characterized in that, include: The data acquisition module is used to collect multi-source heterogeneous monitoring data during the operation of the gas turbine, clean the multi-source heterogeneous monitoring data, and divide it into steady-state combustion segments; The operating condition clustering module is used to perform cluster analysis and standardization processing on the steady-state combustion segment based on load data as operating condition parameters, and to construct a combustion state evaluation background for different operating conditions. The model building and training module is used to build a physical augmented autoencoder evaluation model that integrates data reconstruction and physical mechanism constraints. The evaluation model is trained using historical health data with the goal of minimizing the total loss, which includes a reconstruction loss term and a physical constraint term. To establish a reference combustion mode corresponding to the aforementioned operating conditions; The state assessment and output module is used to map the segment to be assessed to the combustion state evaluation background of the sub-condition to determine the corresponding target condition interval, input the segment to be assessed into the trained model corresponding to the target condition interval, calculate the total loss residual between the model output and the segment to be assessed, map the total loss residual into the combustion health index and output the assessment result.

9. A processing device, characterized in that, The method includes at least one processor and at least one memory communicatively connected to the processor, wherein the memory stores program instructions executable by the processor, and the processor can execute the method as described in any one of claims 1 to 7 by invoking the program instructions.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause the computer to perform the method as described in any one of claims 1 to 7.

Images