Intelligent diagnosis and dynamic control system and method for thermal parameters of an aero-engine combustion chamber

Abstract

This invention relates to the field of aero-engine control technology, and discloses an intelligent diagnosis and dynamic control system and method for thermal parameters of aero-engine combustors. The method includes: performing model-independent element learning training on a distortion evolution prediction network and a causal mapping model to generate element initialization parameters; acquiring real-time data of the compressor outlet section, performing Fourier decomposition to generate intake distortion dynamic evolution characteristic parameters; performing a few-step gradient adaptive update when entering extreme operating conditions; generating a prospective diagnostic result based on the two models after rapid adaptation; calculating the out-of-domain uncertainty index, adaptively allocating feedforward weights and feedback weights, and generating feedforward-feedback fusion adjustment commands; and collecting actual data after executing the commands to perform online adaptive updates.

Description

Technical Field

[0001] This invention relates to the field of aero-engine control technology, specifically to an intelligent diagnostic and dynamic control system and method for thermal parameters of aero-engine combustor. Background Technology

[0002] During flight, the combustion chamber outlet temperature distribution of an aero-engine continuously changes due to the coupled effects of factors such as intake distortion, flight attitude changes, and atmospheric disturbances. Existing technologies employ Fourier decomposition to extract the frequency domain characteristics of the circumferential non-uniformity of the outlet temperature, solve for the flow compensation value of each nozzle based on the spatial transfer relation matrix, and achieve temperature uniformity control through a feedback closed-loop mode of non-uniformity detection, compensation calculation, execution, effect detection, and iterative correction.

[0003] However, the above methods have two shortcomings. First, feedback closed-loop control is limited by the thermal inertia of the combustion chamber, requiring each iteration to wait for the temperature response data after execution. When intake distortion changes continuously on a second-level timescale, the diagnostic results and control compensation always lag behind the distortion changes, leading to a decrease in temperature homogenization. Second, under extreme conditions at the flight envelope boundary, the intake distortion amplitude far exceeds the normal range and the evolution pattern undergoes a qualitative change, causing the distortion prediction network and causal mapping model pre-trained on normal operating condition data to fail. The weighted aggregation-based transfer initialization method tends to the average value of all benchmark networks and loses its specificity when extreme conditions are far from all benchmark regions, and fine-tuning with only a few data points is prone to overfitting, further reducing the reliability of diagnosis and control. Summary of the Invention

[0004] This invention provides an intelligent diagnostic and dynamic control system and method for thermal parameters of aero-engine combustors, which solves the technical problems in related technologies such as the diagnosis of combustor thermal parameters lagging behind changes in operating conditions, insufficient model generalization ability under extreme operating conditions, and difficulty in coordinating feedforward and feedback control.

[0005] This invention discloses an intelligent diagnosis and dynamic control method for thermal parameters of aero-engine combustors, which includes at least the following steps: based on multi-condition data within the flight envelope, performing model-independent meta-learning training on the distortion prediction network and the causal mapping model respectively, and generating meta-initialization parameters of the distortion prediction network and the causal mapping model; Real-time data from the compressor outlet section sensor array is acquired, Fourier decomposition is performed on the circumferential distribution of total pressure and total temperature, the amplitude and phase angle of each harmonic of the intake distortion are extracted, and the amplitude change rate and phase angle rotation rate of each harmonic are calculated to generate dynamic evolution characteristic parameters of intake distortion. When entering the extreme operating condition region, starting from the meta-initialization parameters, a small step gradient adaptive update is performed on the distortion prediction network and the causal mapping model using a small amount of data under this operating condition. The sequence of dynamic evolution characteristic parameters of intake distortion at multiple consecutive time steps is input into the adapted distortion prediction network to generate a forward-looking intake distortion prediction value after one control response cycle. The predicted intake distortion value is input into the adapted causal mapping model to generate the predicted increment of outlet temperature non-uniformity as a prospective diagnostic result of thermal parameters. The predicted increment of the outlet temperature non-uniformity is superimposed with the frequency domain characteristics of the current actual outlet temperature non-uniformity to obtain the estimated value of the outlet temperature non-uniformity. The feedforward compensation value is calculated using the spatial transfer relation matrix. At the same time, the feedback compensation value is calculated based on the frequency domain characteristics of the current actual outlet temperature non-uniformity using the spatial transfer relation matrix. The feedforward compensation value and the feedback compensation value are weighted and fused to generate a feedforward-feedback fusion adjustment command and execute it.

[0006] Furthermore, the model-independent meta-learning training includes: The standardized data is organized into a multi-task learning format according to the operating condition region, with each operating condition region as an independent task. The inner loop performs several steps of gradient updates on the support set of each task to adapt it to that task. The outer loop optimizes the global initialization parameters across all tasks, enabling the few-step gradient adaptation starting from these global initialization parameters to achieve generalization performance across all tasks. The outer loop optimizes the global initialization parameters and also optimizes the task adaptive learning rate for each parameter dimension. The dimension of the task adaptive learning rate is consistent with the dimension of the learnable parameters of the corresponding network, and is used to control the update step of each parameter dimension in the inner loop.

[0007] Furthermore, in the few-step gradient adaptive update, the learning rate for each update step is controlled by the task-adaptive learning rate, specifically: In each adaptation step, the gradient of the loss function is calculated on a small amount of data with the current parameter values, and the parameter values ​​are updated one parameter at a time along the gradient descent direction with the task adaptive learning rate. The loss function of the distortion prediction network is the mean square error between the predicted amplitude and phase angle of each harmonic and the actual value. The loss function of the causal mapping model is the mean square error between the predicted increment of outlet temperature non-uniformity and the actual increment. The number of steps for adaptation is consistent with the number of steps in the loop during the meta-training phase.

[0008] Furthermore, the distortion prediction network adopts a long short-term memory network structure, with its input being a sequence of dynamic evolution characteristic parameters of intake distortion over multiple consecutive time steps, and its output being the amplitude and phase angle of each harmonic after one future control response cycle; the causal mapping model adopts a multilayer perceptron structure, with its input being the amplitude and phase angle of each harmonic of intake distortion, and its output being the increment of outlet temperature non-uniformity caused by the distortion change; wherein the output layer of the causal mapping model is a fully connected linear layer, and the output dimension is consistent with the representation dimension of the outlet temperature non-uniformity increment.

[0009] Furthermore, it also includes: Calculate the Mahalanobis distance between the current input data point and the distribution of the meta-training task, and use the Mahalanobis distance as an out-of-domain uncertainty index; wherein, the Mahalanobis distance is calculated as follows: the difference between the normalized intake distortion dynamic evolution feature parameter vector at the current moment and the joint distribution mean vector of the feature parameters of each task in the meta-training stage is taken as the square root after performing a quadratic operation on the inverse matrix of the joint distribution covariance matrix. Based on the extraterritorial uncertainty index, the additional uncertainty interval of the predicted increment of the outlet temperature non-uniformity is estimated, and a prospective diagnostic result of the thermal parameters with confidence level is generated.

[0010] Furthermore, in the weighted fusion, the sum of the feedforward weight and the feedback weight is 1; the feedforward weight is adjusted in the opposite direction by the external uncertainty index, and its value is equal to 1 divided by 1 and the sum of the product of the preset adjustment sensitivity coefficient and the external uncertainty index; wherein the adjustment sensitivity coefficient is a preset positive scalar used to control the decay rate of the external uncertainty index on the feedforward weight.

[0011] Furthermore, it also includes: The feedforward feedback fusion control command is pruned by applying physical feasible domain constraints, which include three boundary conditions: lean flameout limit, rich flame smoke limit, and maximum allowable outlet temperature. When the flow compensation value of any nozzle in the feedforward feedback fusion control command exceeds the feasible range determined by the physical feasible domain constraints, the flow compensation value is pruned to the nearest constraint boundary value to generate a safe feedforward feedback fusion control command.

[0012] Furthermore, it also includes: After executing the feedforward fusion adjustment command, the actual distortion evolution data and temperature response data are collected and standardized; the deviation between the actual temperature change and the predicted outlet temperature non-uniformity increment in the prospective diagnosis results of thermal parameters is calculated; each newly collected data point is added to the support set of the distortion prediction network and the causal mapping model, and the few-step gradient adaptation update is re-executed to generate the online incrementally adapted distortion prediction network and causal mapping model.

[0013] Furthermore, it also includes: The newly acquired distortion evolution data and temperature response data are fed back as new task experience to the outer loop of the model-independent meta-learning training of the distortion prediction network and the causal mapping model, respectively; in the next meta-training cycle, the meta-initialization parameters are updated using the expanded task set containing the new task experience.

[0014] This invention provides an intelligent diagnostic and dynamic control system for thermal parameters of an aero-engine combustion chamber, characterized in that it comprises: The meta-training module is used to perform model-independent meta-learning training on the distortion prediction network and the causal mapping model based on multi-condition data within the flight envelope, and to generate meta-initialization parameters for the distortion prediction network and the causal mapping model. The distortion feature extraction module is used to acquire real-time data from the compressor outlet section sensor array, perform Fourier decomposition on the circumferential distribution of total pressure and total temperature, extract the amplitude and phase angle of each harmonic of the intake distortion, calculate the amplitude change rate and phase angle rotation rate of each harmonic, and generate dynamic evolution characteristic parameters of intake distortion. A rapid adaptation module is used to perform a few-step gradient adaptation update on the distortion prediction network and the causal mapping model when entering an extreme working condition region, starting from the meta-initialization parameters and using a small amount of data under that working condition. The distortion prediction module is used to input the sequence of dynamic evolution characteristic parameters of intake distortion at multiple consecutive time steps into the adapted distortion prediction network to generate a forward-looking intake distortion prediction value after one control response cycle. The prospective diagnostic module is used to input the prospective intake distortion prediction value into the adapted causal mapping model and generate the predicted increment of outlet temperature non-uniformity as the prospective diagnostic result of thermal parameters. The fusion control module is used to superimpose the predicted increment of the outlet temperature non-uniformity with the frequency domain characteristics of the current actual outlet temperature non-uniformity to obtain the estimated value of the outlet temperature non-uniformity. The module calculates the feedforward compensation value through the spatial transfer relationship matrix and calculates the feedback compensation value based on the frequency domain characteristics of the current actual outlet temperature non-uniformity through the spatial transfer relationship matrix. The feedforward compensation value and the feedback compensation value are weighted and fused to generate a feedforward-feedback fusion control command and execute it.

[0015] The present invention has the following beneficial effects.

[0016] Model-independent meta-learning training is performed on both the distortion evolution prediction network and the causal mapping model. The obtained meta-initialization parameters are located in regions of the parameter space that are sensitive to various working conditions. When entering extreme working conditions, a few-step gradient adaptation update is performed starting from the meta-initialization parameters. This requires less data and has a clear convergence direction, ensuring that the two core diagnostic components remain effective under extreme conditions. This overcomes the problems of weighted aggregation initialization methods tending to the mean and losing specificity, as well as overfitting caused by fine-tuning with a small amount of data.

[0017] By embedding the rapidly adapted distortion evolution prediction network and causal mapping model into the feedforward prediction pathway, the feedforward flow compensation value is generated in advance at the current moment, and the feedforward feedback fusion adjustment command is issued. There is no need to wait for the return of the combustion chamber thermal response data before iterative correction, thus avoiding the double lag in diagnosis and control caused by the gradual accumulation of waiting delays for each feedback iteration due to the thermal inertia of the combustion chamber.

[0018] Through the external uncertainty index Driving feedforward weights With feedback weights The adaptive allocation automatically reduces the feedforward weight and increases the feedback weight in the initial rapid adaptation phase, causing the control to revert to a robust mode dominated by actual temperature feedback. As extreme operating condition data is accumulated through online point-by-point learning, the feedforward weight gradually increases, and the control smoothly transitions to a feedforward-dominated mode. Combined with the pruning of feedforward-feedback fusion control commands by the constraints of the physical feasible region, the flow compensation values ​​of each nozzle are always limited to a safe operating range between the lean-burn flameout limit and the rich-burn smoke limit, ensuring the reliability of thermal parameter control across the entire flight envelope. Attached Figure Description

[0019] Figure 1 This is a flowchart of the intelligent diagnosis and dynamic control method for thermal parameters of aero-engine combustion chamber provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the meta-training task partitioning and data scale provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the change of the loss function during the few-step gradient adaptation process provided in an embodiment of the present invention; Figure 4 This is a prospective diagnostic result provided by an embodiment of the present invention: a schematic diagram of temperature non-uniformity in each circumferential position segment; Figure 5 This is a schematic diagram comparing some of the nozzle feedforward feedback fusion adjustment commands provided in an embodiment of the present invention; Figure 6 This is a schematic diagram illustrating the change in temperature non-uniformity prediction deviation during the online adaptation process provided in an embodiment of the present invention; Figure 7This is a schematic diagram illustrating the changes in out-of-domain uncertainty and feedforward weights during the online adaptation process provided in this embodiment of the invention; Figure 8 This is a schematic diagram of the standardized comparison of the dynamic evolution characteristic parameters of intake distortion provided in the embodiments of the present invention; Figure 9 This is a schematic diagram of the distribution of the prospective diagnostic uncertainty interval provided in an embodiment of the present invention. Detailed Implementation

[0020] Example 1; During flight, the combustion chamber outlet temperature distribution of an aero-engine continuously changes due to the coupled effects of factors such as intake distortion, flight attitude changes, and atmospheric disturbances. Existing technologies employ Fourier decomposition to extract the frequency domain characteristics of the circumferential non-uniformity of the outlet temperature, solve for the flow compensation value of each nozzle based on the spatial transfer relation matrix, and achieve temperature uniformity control through a feedback closed-loop mode of "detecting non-uniformity → calculating compensation → execution → detecting effect → iterative correction".

[0021] In scenarios where extreme operating conditions at the flight envelope boundary coexist with dynamic changes in intake distortion, the aforementioned methods have two shortcomings. First, feedback closed-loop control is limited by the thermal inertia of the combustion chamber, requiring each iteration to wait for the temperature response data to return after execution. When intake distortion changes continuously on a second-level timescale, the diagnostic results and control compensation always lag behind the distortion changes, resulting in a decrease in temperature homogenization. Second, under extreme operating conditions, the amplitude of intake distortion far exceeds the normal range, and the evolution pattern undergoes a qualitative change, causing the distortion prediction network and causal mapping model pre-trained on normal operating condition data to fail. The transfer initialization method based on weighted aggregation tends to average all benchmark networks when extreme operating conditions are far from all benchmark regions, losing its specificity, and fine-tuning with only a few data points is prone to overfitting.

[0022] This embodiment provides a method for intelligent diagnosis and dynamic control of thermal parameters in an aero-engine combustor. The executing entity in this embodiment is the aero-engine full-authority digital electronic controller (FAD) or an airborne computing platform connected to it. This airborne computing platform has the capability to acquire real-time data from compressor outlet cross-section sensors and is communicatively connected to the flow regulation actuators of each nozzle in the combustor.

[0023] A method for intelligent diagnosis and dynamic control of thermal parameters of an aero-engine combustor according to an embodiment of the present invention, such as... Figure 1 As shown, it includes the following steps: Step 1: Perform meta-training on the distortion prediction network and causal mapping model based on multi-condition data within the flight envelope to generate their respective meta-initialization parameters.

[0024] Acquire the inlet distortion time-series data and corresponding outlet temperature response data for each operating condition region within the flight envelope. Perform Z-score normalization on the above data to eliminate the influence of differences in the dimensions of various feature parameters on the training process. Organize the normalized data into a multi-task learning format according to the operating condition region, with each operating condition region as an independent task.

[0025] The distortion evolution prediction network and the causal mapping model were respectively incorporated into a model-independent meta-learning framework for meta-training. The distortion evolution prediction network adopted a long short-term memory network structure, with its input being a sequence of dynamic evolution feature parameters of intake distortion over multiple consecutive time steps, and its output being the amplitude and phase angle of each harmonic after one future control response cycle. The causal mapping model adopted a multilayer perceptron structure, with its input being the amplitude and phase angle of each harmonic of intake distortion, and its output being the increment of outlet temperature non-uniformity caused by the distortion change. Both models used the mean squared error loss function and were trained using the Adam optimization algorithm. The output layer of the causal mapping model was a fully connected linear layer, and the output dimension was consistent with the representation dimension of the outlet temperature non-uniformity increment.

[0026] The model-independent meta-learning framework's inner loop performs several gradient updates on the support set of each task to adapt it to that task, while the outer loop optimizes the globally initialized parameters across all tasks. This ensures that the few-step gradient adaptation starting from these globally initialized parameters achieves superior generalization performance across all tasks. After meta-training, the meta-initialized parameters for the distortion prediction network are generated. Meta-initialization parameters of the causal mapping model ,in These are the global initialization values ​​for all learnable parameters of the distortion evolution prediction network. These are the global initialization values ​​for all learnable parameters of the causal mapping model.

[0027] It should be noted that the outer loop of the model-independent meta-learning framework optimizes not only the global initialization parameters, but also the task-adaptive learning rate used for each adaptive update step. . It is a scalar vector with a parameter-wise dimension, whose dimension is consistent with the learnable parameter dimension of the corresponding network. It controls the update step size of each parameter dimension in the inner loop, so that different parameter dimensions can obtain independent update speed in the few-step adaptation.

[0028] Step 2: Obtain real-time data from the compressor outlet section sensor array, perform Fourier decomposition, and generate intake distortion dynamic evolution characteristic parameters.

[0029] Real-time data from the circumferential total pressure and total temperature sensor array at the compressor outlet section are acquired. Fourier decomposition is performed on the circumferential distributions of total pressure and total temperature to extract the amplitude and phase angle of each harmonic of the intake distortion. Based on this, the amplitude change rate and phase angle rotation rate of each harmonic are calculated to generate dynamic evolution characteristic parameters of the intake distortion. The generated dynamic evolution characteristic parameters of the intake distortion are standardized using the same Z-score standardization parameters as in step 1 to ensure that the real-time characteristic parameters are consistent with the data distribution in the meta-training stage.

[0030] It should be noted that the dynamic evolution characteristic parameters of intake distortion include the amplitude, phase angle, amplitude change rate, and phase angle rotation rate of each harmonic. The amplitude change rate is the ratio of the difference in amplitude of the same harmonic at adjacent sampling times to the sampling interval, and the phase angle rotation rate is the ratio of the difference in phase angle of the same harmonic at adjacent sampling times to the sampling interval. These change rate parameters characterize the dynamic evolution trend of the circumferential distribution of intake distortion over time, providing temporal gradient information for the subsequent distortion evolution prediction network.

[0031] Step 3: Perform a few-step gradient adaptation update on the distortion prediction network and causal mapping model based on the meta-initialization parameters to generate a rapidly adapted distortion prediction network and causal mapping model.

[0032] When the flight enters the extreme operating condition region, a small amount of initially usable distortion evolution data and temperature response data under this condition are acquired, and normalization processing is performed using the same Z-score normalization parameters as in step 1. (The remaining text appears to be incomplete and requires further context.) and Starting with the aforementioned limited data, a few-step gradient adaptive update is performed. The learning rate for each update step is determined by the task-adaptive learning rate. Control. After the update is complete, a distortion prediction network and causal mapping model that have adapted quickly to extreme conditions are generated.

[0033] It should be noted that the specific process of the above-mentioned few-step gradient adaptation update is as follows: In each adaptation step, the gradient of the loss function is calculated on a small amount of data with the current parameter values, and the learning rate is adjusted along the gradient descent direction using the task-adaptive learning rate. Parameter values ​​are updated dimensionally. The loss function of the distortion evolution prediction network is the mean square error between the predicted amplitude and phase angle of each harmonic and the actual value. The loss function of the causal mapping model is the mean square error between the predicted increment of outlet temperature non-uniformity and the actual increment. The number of adaptation steps is consistent with the number of steps in the inner loop of the meta-training phase, ranging from 1 to 5 steps.

[0034] Step 4: Input the sequence of dynamic evolution characteristic parameters of intake distortion into the distortion prediction network after rapid adaptation to generate prospective intake distortion prediction values.

[0035] The sequence of dynamic evolution characteristic parameters of intake distortion at multiple consecutive time steps is input into the distortion evolution prediction network after rapid adaptation. The amplitude and phase angle of each harmonic of the intake distortion at the compressor outlet section are predicted after one control response cycle, generating forward-looking intake distortion prediction values.

[0036] It should be noted that the aforementioned control response period refers to the time elapsed from the issuance of the nozzle flow regulation command to the observable change in the combustion chamber outlet temperature distribution. This control response period is determined by the thermal inertia characteristics of the combustion chamber. The prediction time domain of the forward-looking intake distortion prediction value covers a complete control response period, enabling step 5 to obtain an estimate of the future distortion state at the current moment.

[0037] Step 5: Input the prospective intake distortion prediction value into the rapidly adapted causal mapping model to generate prospective diagnostic results for thermal parameters.

[0038] By inputting the forward-looking intake distortion prediction value into the rapidly adapted causal mapping model, the incremental change in outlet temperature non-uniformity caused by the distortion change is predicted after one future control response cycle, generating forward-looking diagnostic results for thermal parameters.

[0039] In this embodiment of the application, in order to quantify the reliability of the prospective diagnostic results of thermal parameters when they deviate from the training distribution under extreme operating conditions, the Mahalanobis distance between the current input data point and the meta-training task distribution is also calculated, and this Mahalanobis distance is used as an out-of-domain uncertainty index. . The calculation formula is:

[0040] in This is the standardized vector of dynamic evolution characteristic parameters of intake distortion at the current moment. This is the joint mean vector of the feature parameters of each task during the meta-training phase. This is the joint distribution covariance matrix of the feature parameters of each task during the meta-training phase. for The inverse matrix, Indicates transpose. Based on the external uncertainty index. An additional uncertainty interval estimate is applied to the predicted increment of outlet temperature non-uniformity, generating prospective diagnostic results for thermal parameters with confidence level annotations. (Outside-domain uncertainty index) The larger the value, the further the current operating condition deviates from the original training distribution, the wider the uncertainty range of the predicted increment, and the lower the confidence level.

[0041] Step 6: Based on the prospective diagnostic results of thermal parameters and the current actual outlet temperature non-uniformity data, generate feedforward feedback fusion control commands.

[0042] The predicted increment of outlet temperature non-uniformity is superimposed with the frequency domain characteristics of the current actual outlet temperature non-uniformity to obtain the estimated outlet temperature non-uniformity at the future target time. Based on the spatial transfer relation matrix, the feedforward flow compensation value for each nozzle is calculated from the estimated outlet temperature non-uniformity. Simultaneously, based on the frequency domain characteristics of the current actual outlet temperature non-uniformity, the feedback compensation value is calculated using the spatial transfer relation matrix. The feedforward compensation value and the feedback compensation value are weighted and fused to generate a feedforward-feedback fusion control command.

[0043] It should be noted that the feedforward weights in the above weighted fusion are... With feedback weights The allocation satisfies ,in For feedforward weights, Feedforward weights. From the external uncertainty index Reverse adjustment, specifically, feedforward weights Determine as follows:

[0044] in To adjust the sensitivity coefficient, a preset positive scalar is used to control the out-of-domain uncertainty index. Feedforward weights The decay rate. When the out-of-domain uncertainty index At lower levels, feedforward weights The level is relatively high, and the feedforward feedback fusion control command mainly relies on forward-looking intake distortion prediction values ​​for feedforward compensation; when the external uncertainty index is relatively high... When the value is high, the feedforward weight Reduce, feedback weight Accordingly, the feedforward feedback fusion adjustment command relies more on the current actual temperature feedback data for compensation.

[0045] In this embodiment, to ensure that the feedforward feedback fusion control command does not exceed the safe operating boundary of the combustion chamber, based on step 6, a physical feasible domain constraint is applied to the fused feedforward feedback fusion control command for trimming. The physical feasible domain constraint includes three boundary conditions: lean flameout limit, rich flame smoke limit, and maximum permissible outlet temperature. When the flow compensation value of any nozzle in the fused feedforward feedback fusion control command exceeds the feasible range determined by the above constraints, the flow compensation value is trimmed to the nearest constraint boundary value, generating a safe feedforward feedback fusion control command.

[0046] Step 7: After executing the adjustment command, collect actual data, perform online adaptive updates based on the deviation, and generate the incrementally adapted distortion prediction network and causal mapping model.

[0047] After executing the feedforward-feedback fusion adjustment command, actual distortion evolution data and temperature response data are collected, and normalization processing is performed using the same Z-score normalization parameters as in step 1. The deviation between the actual temperature change and the predicted increment of outlet temperature non-uniformity in the prospective diagnosis results of thermal parameters is calculated. Each newly collected data point is synchronously added to the support set of the distortion evolution prediction network and the causal mapping model, and a few-step gradient adaptation update is re-executed to achieve online point-by-point learning, generating the online incrementally adapted distortion evolution prediction network and causal mapping model.

[0048] In this embodiment of the application, in order to make the meta-initialization parameters and The system continuously absorbs distribution information from extreme operating conditions and feeds newly acquired distortion evolution data and temperature response data as new task experience into the outer loop of the model-independent meta-learning frameworks of both the distortion evolution prediction network and the causal mapping model. In the next meta-training cycle, the meta-initialization parameters are updated using an expanded task set containing new task experience. and This method ensures that when encountering similar or identical extreme operating conditions, the initialization parameters already incorporate previous adaptation experience, reducing the amount of data and gradient steps required for adaptation. Simultaneously, it records the changes in prediction deviation before and after each online adaptation, generating a record of the accuracy assessment of thermal parameter diagnosis.

[0049] This implementation performs model-independent meta-learning training on both the distortion evolution prediction network and the causal mapping model to obtain globally optimal initialization parameters across different operating conditions. When the flight enters an extreme operating condition region, a few-step gradient adaptation update is performed starting from the meta-initialization parameters. Since the meta-initialization parameters are located in a region of the parameter space that is sensitive to various operating conditions, a few-step gradient update can move the parameters to an effective fit position along the target task direction. Compared to the weighted aggregation initialization method, which tends to the average value of all benchmark networks and loses its specificity when extreme operating conditions are far from all benchmark regions, and the overfitting problem caused by fine-tuning the aggregated network with only a few data points, the meta-learning adaptation process starts from globally sensitive initialization, requires less data, and has a clear convergence direction, thus ensuring the effectiveness of the two core diagnostic components, the distortion evolution prediction network and the causal mapping model, under extreme operating conditions.

[0050] Building upon this foundation, this implementation embeds a rapidly adaptable distortion evolution prediction network and a causal mapping model into the feedforward prediction pathway. The distortion evolution prediction network predicts the intake distortion state one control response cycle later, which is then mapped to an outlet temperature non-uniformity increment via the causal mapping model. Based on this, a feedforward flow compensation value is generated in advance at the current moment, and a feedforward feedback fusion control command is issued. This feedforward compensation eliminates the need to wait for combustion chamber thermal response data to return before iterative correction. In scenarios where intake distortion continuously changes on a second-level timescale, it avoids the dual lag in diagnosis and control caused by the gradual accumulation of waiting delays in each feedback iteration due to combustion chamber thermal inertia.

[0051] At the same time, through the external uncertainty index The adaptive allocation of feedforward and feedback weights drives the system to automatically reduce feedforward weights and increase feedback weights when the accuracy of the initial forward-looking diagnostic results of thermal parameters has not yet fully converged. This allows the control system to revert to a robust mode dominated by actual temperature feedback. As online point-by-point learning continuously accumulates extreme operating condition data, the accuracy of the forward-looking diagnostic results of thermal parameters continues to improve, the feedforward weights gradually recover, and the control system smoothly transitions to a feedforward-dominated mode. By combining the physical feasible domain constraint with the pruning of feedforward-feedback fusion control commands, the flow compensation values ​​of each nozzle are always limited to a safe operating range between the lean-burn flameout limit and the rich-burn smoke limit, ensuring both reliability and forward-looking capability of thermal parameter control across the entire flight envelope.

[0052] The following is an example of an application of the present invention, such as... Figure 2-9 As shown, the implementation process is as follows: In a scenario where a certain type of turbofan engine encounters significant intake distortion during high-altitude, high-speed flight, the aircraft rapidly climbs from cruise mode to the extreme operating region at the boundary of its flight envelope. At this point, the intake distortion amplitude at the compressor outlet section increases significantly, and the distortion pattern evolves continuously over several seconds. The onboard computing platform collects real-time data from 24 circumferentially distributed total pressure and total temperature sensors at the compressor outlet section and maintains communication with the flow control actuators of the 12 nozzles in the combustion chamber. The control response cycle is 0.8 seconds, which is the time required from issuing a nozzle flow control command to producing an observable change in the outlet temperature distribution.

[0053] Within the flight envelope, the system is divided into several operating condition regions based on altitude and Mach number, with each region organized as an independent training task. Both the distortion evolution prediction network (long short-term memory network structure) and the causal mapping model (multilayer perceptron structure) are incorporated into a model-independent meta-learning framework for meta-training. The inlet distortion time-series data and outlet temperature response data for each operating condition region are Z-score normalized and then divided into inner and outer loop training based on support and query sets. After meta-training is complete, meta-initialization parameters for the distortion prediction network are generated. and causal mapping model meta-initialization parameters Simultaneously, a task-adaptive learning rate vector is obtained along the parameter-wise dimension. .

[0054] Table 1. Meta-training task division and data scale

[0055] After the aircraft enters the extreme operating condition region, the onboard computing platform continuously collects data from 24 sensors at 20-millisecond sampling intervals. Fourier decomposition is performed on the circumferential distributions of total pressure and total temperature, extracting the amplitude and phase angle of the first three harmonics, and calculating the amplitude change rate and phase angle rotation rate at adjacent sampling times. Taking a certain sampling time as an example, the generated dynamic evolution characteristic parameters of intake distortion are shown in the table below, and normalization processing is performed using the Z-score normalization parameters from step 1.

[0056] Table 2. Dynamic evolution characteristics of intake distortion at a certain sampling time.

[0057] After the aircraft entered extreme operating conditions, only 12 sample points of data were initially available. These were respectively... and Starting with the above 12 sample points, a 3-step gradient adaptation update is performed, with the learning rate for each step being... Parameter-by-parameter control. The table below shows the change of the loss function of the distortion prediction network during the adaptation process, reflecting the characteristic that the meta-initialization parameters start from the globally sensitive position and converge to the effective region of the target task with a few updates.

[0058] Table 3. Changes in loss function during few-step gradient adaptation.

[0059] The sequence of dynamic evolution characteristic parameters of inlet distortion over eight consecutive time steps (covering 160 milliseconds) is input into a rapidly adapted distortion evolution prediction network to predict the amplitude and phase angle of each harmonic at the compressor outlet section after 0.8 seconds, generating forward-looking inlet distortion prediction values. These prediction values ​​are then input into a rapidly adapted causal mapping model to obtain the incremental prediction of outlet temperature non-uniformity. Simultaneously, the current characteristic parameter vector is calculated. The Mahalanobis distance between the joint distribution of the meta-training and the untrained distribution yields the out-of-domain uncertainty index. The formula is:

[0060] in, Let be the mean vector of the joint distribution of the meta-training. This corresponds to the covariance matrix. It is calculated at the current time step. This indicates that there is a certain deviation between the current extreme working conditions and the original training distribution, and the diagnostic results are supplemented with an uncertainty interval of corresponding width.

[0061] Table 4 Prospective Diagnostic Results (Increment in Outlet Temperature Non-uniformity)

[0062] Based on the predicted target value of outlet temperature non-uniformity, the feedforward flow compensation value for each nozzle is solved using the spatial transfer relation matrix; simultaneously, the feedback compensation value is calculated based on the current actual non-uniformity frequency domain characteristics. Adjust the sensitivity coefficient The feedforward weights are:

[0063] in, To adjust the sensitivity coefficient, This is the external uncertainty index. Substituting the values ​​into the expression... Feedback weight After fusion, the flow compensation values ​​of each nozzle are trimmed according to the physical feasible domain constraints to ensure that they do not exceed the lean flameout limit and the rich flame smoke limit.

[0064] Table 5. Partial Nozzle Feedforward Fusion Adjustment Commands

[0065] It should be noted that the fusion compensation value of nozzle 9, 3.80, exceeds the upper limit of 3.50 corresponding to the fuel-rich smoke limit and is therefore trimmed to the constraint boundary value.

[0066] After executing the fusion adjustment command, the airborne computing platform collects actual distortion evolution data and temperature response data, and calculates the deviation between the actual outlet temperature change and the predicted increment. After adding the newly collected data points to the support set, the three-step gradient adaptation update is re-executed, and the prediction bias of the distortion prediction network and the causal mapping model continues to shrink. The newly collected data is synchronously fed back as new task experience to the outer loop of the model-independent meta-learning framework, and updated in the next meta-training cycle. and .

[0067] Table 6 Record of Prediction Deviation Changes During Online Adaptation

[0068] The entire data flow process demonstrates a clear logical chain. Step 1 generates the meta-initialization parameters. and and task-adaptive learning rate This is the starting point for all subsequent steps. The standardized intake distortion dynamic evolution characteristic parameters output in step 2 provide the data foundation for the few-step adaptation in step 3 and also serve as the sequence input for step 4. The fast-adapted model generated in step 3 directly connects to the forward prediction link of steps 4 and 5. The output of step 5... The weight allocation in step 6 is adjusted in real time so that the fusion ratio of feedforward and feedback adapts to the diagnostic confidence level. The actual data collected in step 7 is fed back to the support set in step 3 to form a closed loop of online point-by-point learning. At the same time, new task experience is fed into the meta-learning outer loop, so that the meta-initialization parameters continuously absorb extreme working condition information, ensuring the efficiency of the adaptation process under the same working conditions in the future.

[0069] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for intelligent diagnosis and dynamic control of thermal parameters in an aero-engine combustion chamber, characterized in that, At least the following steps are included: Based on multi-condition data within the flight envelope, model-independent meta-learning training is performed on the distortion prediction network and the causal mapping model respectively to generate meta-initialization parameters for the distortion prediction network and the causal mapping model. Real-time data from the compressor outlet section sensor array is acquired, Fourier decomposition is performed on the circumferential distribution of total pressure and total temperature, the amplitude and phase angle of each harmonic of the intake distortion are extracted, and the amplitude change rate and phase angle rotation rate of each harmonic are calculated to generate dynamic evolution characteristic parameters of intake distortion. When entering the extreme operating condition region, starting from the meta-initialization parameters, a small step gradient adaptive update is performed on the distortion prediction network and the causal mapping model using a small amount of data under this operating condition. The sequence of dynamic evolution characteristic parameters of intake distortion at multiple consecutive time steps is input into the adapted distortion prediction network to generate a forward-looking intake distortion prediction value after one control response cycle. The predicted intake distortion value is input into the adapted causal mapping model to generate the predicted increment of outlet temperature non-uniformity as a prospective diagnostic result of thermal parameters. The predicted increment of the outlet temperature non-uniformity is superimposed with the frequency domain characteristics of the current actual outlet temperature non-uniformity to obtain the estimated value of the outlet temperature non-uniformity. The feedforward compensation value is calculated using the spatial transfer relation matrix. At the same time, the feedback compensation value is calculated based on the frequency domain characteristics of the current actual outlet temperature non-uniformity using the spatial transfer relation matrix. The feedforward compensation value and the feedback compensation value are weighted and fused to generate a feedforward-feedback fusion adjustment command and execute it.

2. The method for intelligent diagnosis and dynamic control of thermal parameters of aero-engine combustion chamber according to claim 1, characterized in that, The model-independent meta-learning training includes: The standardized data is organized into a multi-task learning format according to the operating condition region, with each operating condition region as an independent task. The inner loop performs several steps of gradient updates on the support set of each task to adapt it to that task. The outer loop optimizes the global initialization parameters across all tasks, enabling the few-step gradient adaptation starting from these global initialization parameters to achieve generalization performance across all tasks. The outer loop optimizes the global initialization parameters and also optimizes the task adaptive learning rate for each parameter dimension. The dimension of the task adaptive learning rate is consistent with the dimension of the learnable parameters of the corresponding network, and is used to control the update step of each parameter dimension in the inner loop.

3. The method for intelligent diagnosis and dynamic control of thermal parameters of aero-engine combustors according to claim 2, characterized in that, In the few-step gradient adaptive update, the learning rate for each update step is controlled by the task-adaptive learning rate, specifically: In each adaptation step, the gradient of the loss function is calculated on a small amount of data with the current parameter values, and the parameter values ​​are updated one parameter dimension by one along the gradient descent direction with the task adaptive learning rate. The loss function of the distortion prediction network is the mean square error between the predicted amplitude and phase angle of each harmonic and the actual value. The loss function of the causal mapping model is the mean square error between the predicted increment of outlet temperature non-uniformity and the actual increment. The number of adaptation steps is consistent with the number of steps in the loop during the meta-training phase.

4. The method for intelligent diagnosis and dynamic control of thermal parameters of aero-engine combustors according to claim 1, characterized in that, The distortion prediction network adopts a long short-term memory network structure. Its input is a sequence of dynamic evolution characteristic parameters of intake distortion over multiple consecutive time steps, and its output is the amplitude and phase angle of each harmonic after one future control response cycle. The causal mapping model adopts a multilayer perceptron structure. Its input is the amplitude and phase angle of each harmonic of intake distortion, and its output is the increment of outlet temperature non-uniformity caused by the distortion change. The output layer of the causal mapping model is a fully connected linear layer, and the output dimension is consistent with the representation dimension of the outlet temperature non-uniformity increment.

5. The intelligent diagnosis and dynamic control method for thermal parameters of aero-engine combustion chamber according to claim 1, characterized in that, Also includes: Calculate the Mahalanobis distance between the current input data point and the distribution of the meta-training task, and use the Mahalanobis distance as an out-of-domain uncertainty index; wherein, the Mahalanobis distance is calculated as follows: the difference between the normalized intake distortion dynamic evolution feature parameter vector at the current moment and the joint distribution mean vector of the feature parameters of each task in the meta-training stage is taken as the square root after performing a quadratic operation on the inverse matrix of the joint distribution covariance matrix. Based on the extraterritorial uncertainty index, the additional uncertainty interval of the predicted increment of the outlet temperature non-uniformity is estimated, and a prospective diagnostic result of the thermal parameters with confidence level is generated.

6. The intelligent diagnosis and dynamic control method for thermal parameters of aero-engine combustors according to claim 5, characterized in that, In the weighted fusion, the sum of the feedforward weight and the feedback weight is 1; the feedforward weight is adjusted in the opposite direction by the external uncertainty index, and its value is equal to 1 divided by 1 and the sum of the product of the preset adjustment sensitivity coefficient and the external uncertainty index; The adjustment sensitivity coefficient is a preset positive scalar used to control the decay rate of the out-of-domain uncertainty index on the feedforward weight.

7. The method for intelligent diagnosis and dynamic control of thermal parameters of aero-engine combustors according to claim 1, characterized in that, Also includes: The feedforward feedback fusion control command is pruned by applying physical feasible domain constraints, which include three boundary conditions: lean flameout limit, rich flame smoke limit, and maximum allowable outlet temperature. When the flow compensation value of any nozzle in the feedforward feedback fusion control command exceeds the feasible range determined by the physical feasible domain constraints, the flow compensation value is pruned to the nearest constraint boundary value to generate a safe feedforward feedback fusion control command.

8. The method for intelligent diagnosis and dynamic control of thermal parameters of aero-engine combustors according to claim 1, characterized in that, Also includes: After executing the feedforward fusion adjustment command, actual distortion evolution data and temperature response data are collected and standardized. Calculate the deviation between the actual temperature change and the predicted outlet temperature non-uniformity increment in the prospective diagnosis results of thermal parameters; add each newly acquired data point to the support set of the distortion prediction network and the causal mapping model, re-execute the few-step gradient adaptation update, and generate the online incrementally adapted distortion prediction network and causal mapping model.

9. The intelligent diagnosis and dynamic control method for thermal parameters of an aero-engine combustion chamber according to claim 8, characterized in that, Also includes: The newly acquired distortion evolution data and temperature response data are fed back as new task experience to the outer loop of the model-independent meta-learning training of the distortion prediction network and the causal mapping model, respectively; in the next meta-training cycle, the meta-initialization parameters are updated using the expanded task set containing the new task experience.

10. A system for intelligent diagnosis and dynamic control of thermal parameters of an aero-engine combustor, used to execute the method for intelligent diagnosis and dynamic control of thermal parameters of an aero-engine combustor as described in any one of claims 1 to 9, characterized in that, include: The meta-training module is used to perform model-independent meta-learning training on the distortion prediction network and the causal mapping model based on multi-condition data within the flight envelope, and to generate meta-initialization parameters for the distortion prediction network and the causal mapping model. The distortion feature extraction module is used to acquire real-time data from the compressor outlet section sensor array, perform Fourier decomposition on the circumferential distribution of total pressure and total temperature, extract the amplitude and phase angle of each harmonic of the intake distortion, calculate the amplitude change rate and phase angle rotation rate of each harmonic, and generate dynamic evolution characteristic parameters of intake distortion. A rapid adaptation module is used to perform a few-step gradient adaptation update on the distortion prediction network and the causal mapping model when entering an extreme working condition region, starting from the meta-initialization parameters and using a small amount of data under that working condition. The distortion prediction module is used to input the sequence of dynamic evolution characteristic parameters of intake distortion at multiple consecutive time steps into the adapted distortion prediction network to generate a forward-looking intake distortion prediction value after one control response cycle. The prospective diagnostic module is used to input the prospective intake distortion prediction value into the adapted causal mapping model and generate the predicted increment of outlet temperature non-uniformity as the prospective diagnostic result of thermal parameters. The fusion control module is used to superimpose the predicted increment of the outlet temperature non-uniformity with the frequency domain characteristics of the current actual outlet temperature non-uniformity to obtain the estimated value of the outlet temperature non-uniformity. The module calculates the feedforward compensation value through the spatial transfer relationship matrix and calculates the feedback compensation value based on the frequency domain characteristics of the current actual outlet temperature non-uniformity through the spatial transfer relationship matrix. The feedforward compensation value and the feedback compensation value are weighted and fused to generate a feedforward-feedback fusion control command and execute it.

Images