A weight self-adaptive training method of an aero-engine digital twin model

Abstract

The present application belongs to the technical field of deep learning and aero-engine operation maintenance, and particularly relates to a weight adaptive training method for an aero-engine digital twin model. The specific technical solution is as follows: an engine rotating component performance tracking model is built and pre-trained, and the relationship between the model test error and the training data is analyzed; for the transition state data whose test error is obviously higher than the steady state data, an adaptive weight loss function is constructed, the adaptive weight corresponding to each training sample is calculated and saved, and the loss weight corresponding to the transition state data is dynamically adjusted during retraining. By using the adaptive weight loss function, the neural network pays more attention to the samples with greater training difficulty, thereby improving the overall prediction accuracy. The method can effectively solve the problem of overall prediction accuracy decline caused by unbalanced sample quantity.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of deep learning and aero-engine operation maintenance, and particularly relates to a weight adaptive training method of an aero-engine digital twin model. BACKGROUND

[0002] With the continuous development and wide application of artificial intelligence technology, more and more attention has been focused on the performance degradation prediction and health management technology research for aero-engines. Traditional maintenance strategies based on time or number of uses are difficult to accurately predict the operating conditions of engine components. Therefore, people began to use machine learning and data mining technology to analyze and predict the trend of engine component performance changes in order to improve the reliability and production efficiency of equipment.

[0003] In particular, for aero-engines and their internal rotating components, due to their special conditions such as high speed, heavy load, high temperature and high vibration during operation, a certain degree of performance degradation and failure is likely to occur, which brings not small loss to the production of enterprises. Therefore, by using intelligent health management system, collecting and analyzing the data of engine rotating components, using data modeling, fault diagnosis and prediction analysis technology, the real-time monitoring and accurate prediction of the operating conditions can be realized, so as to take corresponding maintenance measures in time, prolong the service life and reduce the maintenance cost.

[0004] With the continuous improvement and development of intelligent technology, the performance degradation prediction and health management technology of aero-engines is also constantly updated and optimized, including using more sensors and data acquisition equipment, introducing more artificial intelligence algorithms and models to improve the prediction accuracy and precision, and establishing a more perfect data management and knowledge base system to improve the reliability and shareability of data.

[0005] These new technologies and methods not only can improve the performance and life of aero-engines, but also can reduce operating costs, improve the competitiveness and innovation ability of enterprises. For example, airlines can monitor and analyze the vibration, temperature, pressure and other parameters of the engine in real time, discover potential problems early, and take corresponding measures to avoid failure. This effectively saves maintenance costs and reduces downtime, and improves the efficiency and service quality of airlines.

[0006] The common performance degradation prediction technologies at present are mainly divided into two categories: model-driven and data-driven.

[0007] Most of the model-driven methods are based on the physical laws of the rotating components of an aero-engine, and physical models representing the degradation and remaining life of the components or systems are constructed based on mechanics and other disciplines. The performance degradation prediction schemes based on mathematical or physical model driving have the following defects: the construction of the model needs to consider various variables and parameters in the system, however, the actual system is often very complex, and there may be some variables and parameters that are difficult to quantify, which leads to the model not being able to fully reflect the actual situation; the accuracy of the model is often affected by various factors such as the parameters selected when modeling the system, the degree of approximation of the model, and measurement errors, which can all lead to a decrease in the prediction accuracy of the model; the model is usually constructed based on the assumption of determinism, but there may be some uncertain factors in the actual system, such as environmental changes, material fatigue, etc., which can cause the model prediction results to be inconsistent with the actual situation; in addition, since there may be some unknown variables and parameters in the actual system, and it is not possible to completely predict the situation, the model may need to be constantly updated and corrected. This requires re-experimentation and data collection, which consumes a lot of time and resources.

[0008] Most of the data-driven methods are constructed by applying deep learning algorithms, which learn the mapping relationship between the performance, control and environmental parameters of the rotating components of an aero-engine during operation by inputting the performance, control and environmental parameters, build its performance digital twin model, track and predict the key performance parameters of the mechanical rotating components, and then realize the performance degradation prediction of the aero-engine and related health management based on it.

[0009] The data-driven performance degradation prediction scheme also has some potential problems and limitations, such as data quality: deep learning models have high requirements for data quality, if the data quality is poor, the prediction results of the model may be affected. Training time: deep learning models usually need a large amount of data and time to train, which can cause the training time of the model to be too long and require a large amount of computing resources. Model complexity: deep learning models usually have high complexity and require a large number of parameters for training, which can cause the model to have problems such as overfitting. Interpretability: deep learning models have poor interpretability and it is difficult to explain the prediction results of the model, so it is difficult to determine the cause of performance degradation. Adaptability to new data: the prediction ability of deep learning models is limited by the quality and quantity of the training data, for new data, the model may need to be retrained to achieve good prediction results. Sample imbalance: in the case of sample imbalance, the model may perform better on the class with more samples and worse on the class with fewer samples, leading to a decrease in overall prediction accuracy. SUMMARY

[0010] In order to solve some defects of the data-driven performance degradation prediction scheme in terms of accuracy, specifically to solve the significant decrease in prediction accuracy of special nodes caused by unbalanced sample quantity, the application provides a weight adaptive training method of an aero-engine digital twin model, mainly including a construction of an adaptive weight loss function for rotating components and a neural network training method with weight insertion in advance.

[0011] To achieve the above-mentioned application purposes, the technical scheme adopted by the application is: a weight adaptive training method of an aero-engine digital twin model, an engine rotating component performance tracking model is built and pre-trained, and the relationship between the model test error and the training data is analyzed; for the transition state data whose test error is obviously higher than the steady state data, an adaptive weight loss function is constructed, the adaptive weight corresponding to each training sample is calculated and saved, and the loss weight corresponding to the transition state data is dynamically adjusted during retraining to obtain the final engine rotating component performance tracking model.

[0012] Preferably, the adaptive weight loss function is constructed under the following conditions:

[0013] 1) The adaptive weight X is a function of itself and the transition state-steady state transition judgment parameter, X t =f(X t-1 , n t ), X t is the adaptive weight of the current time step, X t-1 is the adaptive weight value of the previous time step, n t is the transition state-steady state transition judgment parameter, i.e., the rotating speed;

[0014] 2) When the transition state starts, the adaptive weight gradually increases until the upper limit X max ;

[0015] 3) When the transition state ends and enters the steady state, the adaptive weight gradually decreases until X min =1.

[0016] Preferably, the adaptive weight loss function is constructed based on the idea of recursive sequence.

[0017] Preferably, the adaptive weight loss function is constructed as follows:

[0018] A41, a sequence N is constructed for storing the reciprocal parameter of the aero-engine rotating speed, and the elements N i of the sequence N have a domain of [0, 1];

[0019] A42, suppose a recursive sequence X, each element in the recursive sequence X represents the loss weight corresponding to the current time running data, the elements in the recursive sequence X are real numbers between [1, a], X0=1; and the adjacent elements in the recursive sequence X have a recursive relationship Xt = f(X t-1 , N t );

[0020] A43, construct the recursive relationship X t = f(X t-1 , N t ) and satisfy: when N i far from 0, X monotonically increases, and in the 2a step iteration part, monotonically increases from the lower limit 1 to approach the upper limit a; when N i close to 0, X monotonically decreases, and in the a 2 step iteration part, monotonically decreases from the upper limit a to approach the lower limit 1;

[0021] A44, derive the recursive relationship:

[0022]

[0023] In the formula, the specific values of a1, a2, a3, a4, b1, c1, c2, c3, c4 can be adjusted according to the training parameter characteristics;

[0024] The adaptive weight loss function is as follows:

[0025]

[0026] In the formula, X is the adaptive weight, f(x) is the performance parameter prediction value calculated by the engine rotating part performance tracking model, and y is the true value of the engine performance parameter.

[0027] Preferably, the engine rotating part performance tracking model learns the mapping relationship between the engine control parameters, environmental parameters and state parameters and the engine performance parameters.

[0028] Preferably, it comprises the following steps:

[0029] A01, normalizing and classifying the aero-engine data;

[0030] A02, build an aero-engine rotating part performance tracking model from engine environmental parameters, control parameters and state parameters to performance parameters, and use the data processed in step A01 for pre-training to obtain a pre-training model;

[0031] A03, error analysis is performed on the pre-training model obtained in step A02 to determine the test error between the transient state data and the steady state data. When the test error of the two is less than m, the final engine rotating part performance tracking model is obtained; when the test error of the two is greater than or equal to m, step A04 is entered;

[0032] A04, determine the adaptive weight loss function construction condition and determine the adaptive weight X calculation formula;

[0033] A05、According to the operation mechanism of the aero-engine, an adaptive weight loss function is constructed:

[0034]

[0035] In the formula, X is the adaptive weight, f(x) is the predicted value of the performance parameter calculated by the engine rotating part performance tracking model, and y is the true value of the engine performance parameter.

[0036] A06、According to the adaptive weight loss function in step A05, the pre-trained model is subjected to secondary optimization training.

[0037] Preferably, in step A01, the engine data is divided into environmental parameters, control parameters, state parameters and performance parameters, the environmental parameters include inlet air temperature and air pressure height, the control parameters include inlet guide vane angle position, nozzle position and throttle lever position, the state parameters include high and low pressure rotor speed and oil pressure, and the performance parameters include thrust and specific fuel consumption.

[0038] Preferably, the step A06 comprises:

[0039] A61、The adaptive weight of each training sample is calculated, and the one-to-one corresponding training sample and adaptive weight are saved;

[0040] A62、The upper limit and lower limit of the adaptive weight are set according to the adaptive weight calculation formula;

[0041] A63、The training sample and the adaptive weight are substituted into the engine rotating part performance tracking model, and secondary training is started to obtain the final engine rotating part performance tracking model.

[0042] Correspondingly, an electronic device comprises:

[0043] One or more processors;

[0044] A storage device for storing one or more programs;

[0045] When the one or more programs are executed by the one or more processors, the one or more processors implement the weight adaptive training method of the aero-engine digital twin model.

[0046] Correspondingly, a computer readable medium stores a computer program, and the computer program is executed by a processor to implement the weight adaptive training method of the aero-engine digital twin model.

[0047] Compared with the prior art, the present application has the following beneficial effects:

[0048] 1. The application proposes a self-adaptive weight loss function construction method based on the operation mechanism of the rotating parts of an aero-engine, which analyzes the relationship between the model test error and the training data; for the transition state data whose test error is obviously higher than the steady state data, a self-adaptive weight loss function is constructed, the weight corresponding to each training sample is calculated and saved, and the loss weight corresponding to the transition state data is dynamically adjusted during retraining. This method can effectively solve the problem of overall prediction accuracy decline caused by unbalanced sample quantity. By using the self-adaptive weight loss function, the neural network pays more attention to the samples that are difficult to train, thereby improving the overall prediction accuracy. The method proposed in the application can significantly reduce the peak error of special data nodes without significantly affecting the average error of the test set. The overall prediction accuracy can be maintained while more accurately processing special data nodes, improving the stability and reliability of the system.

[0049] 2. The application also proposes a neural network training method based on a self-adaptive weight loss function, i.e. a deep learning neural network training method with weight insertion in advance. This method inserts the self-adaptive weight for special data nodes in advance, so that the neural network can quickly adapt to the importance of different samples at the early stage of training, and gradually reduce the amplitude of weight adjustment at the later stage of training, thereby avoiding overfitting. At the same time, this method can also improve the accuracy of the performance degradation prediction model. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 Figure 1 is a schematic diagram of the performance tracking model structure of the rotating parts of an aero-engine;

[0051] Figure 2 Figure 1 is a schematic diagram of the performance tracking model structure of the rotating parts of an aero-engine;

[0052] Figure 3 Figure 1 is a schematic diagram of the performance tracking model structure of the rotating parts of an aero-engine;

[0053] Figure 4 Figure 1 is a schematic diagram of the performance tracking model structure of the rotating parts of an aero-engine; DETAILED DESCRIPTION

[0054] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only a part of the embodiments of the application, not all the embodiments. If not specifically indicated, the technical means used in the embodiments are conventional means familiar to those skilled in the art.

[0055] The specific training mechanism of the traditional deep neural network needs to randomly sort the training set, randomly extract training data input into the prediction model and calculate the loss each time, and adjust the node weight parameters of the prediction model based on the loss back propagation. Therefore, during training, the data cannot be input in the actual running time sequence of the rotating part, and the loss weight X cannot be updated and iterated according to time.

[0056] In order to solve the above problems, based on the adaptive weight loss function for the rotating part of the aero-engine described above, a neural network training method with weight insertion in advance is proposed. The weight of the known loss function only plays a role in training when the loss is calculated, and its dynamic change process and structure will not change during the training process. Therefore, the weight calculation process and the training process can be separated. Before starting the training, the weight is calculated according to the weight calculation formula and saved. In the actual training process, the pre-saved weight parameters are directly called, which optimizes the training logic, reduces the training time, and facilitates the adjustment of hyperparameters and the debugging of training code in the actual training process.

[0057] As shown in Figures 1-2 The core idea of the weight adaptive training method of the aero-engine digital twin model is: a weight adaptive training method of an aero-engine digital twin model, a performance tracking model of an engine rotating part is built and pre-trained, and the relationship between the test error of the model and the training data is analyzed; for the transition state data whose test error is obviously higher than that of the steady state data, an adaptive weight loss function is constructed, the weight corresponding to each training sample is calculated and saved, and the loss weight corresponding to the transition state data is dynamically adjusted during retraining to obtain the final engine rotating part performance tracking model.

[0058] Further, the weight adaptive training method of the aero-engine digital twin model comprises the following method steps:

[0059] A01, normalizing and classifying the aero-engine data.

[0060] Specifically, the engine data is divided into environmental parameters, control parameters, state parameters and performance parameters, the environmental parameters include inlet air temperature and air pressure height, the control parameters include inlet guide vane angle position, nozzle position and throttle lever position, the state parameters include high and low pressure rotor speed and oil pressure, and the performance parameters include thrust and specific fuel consumption.

[0061] A02, an aero-engine rotating part performance tracking model from the engine environmental parameters, control parameters and state parameters to the performance parameters is built, and the data processed by step A01 is used for pre-training to obtain a pre-trained model.

[0062] Specifically, according to the training idea of a traditional neural network regression model, an aero-engine rotating component performance tracking model is built and trained, and a specific model structure is as shown in FIG. 1. Figure 1 The mapping relationship between the engine control parameters, the environmental parameters, the state parameters and the engine performance parameters is learned.

[0063] A03, error analysis is performed on the pre-trained model obtained in step A02 to determine the test error between the transition state data and the steady state data. When the test error of the two is less than m, the final engine rotating component performance tracking model is obtained; when the test error of the two is greater than or equal to m, step A04 is entered. Here, the value range of m can be 1%-3%.

[0064] Specifically, the aero-engine rotating component performance tracking model (pre-trained model) obtained by pre-training is analyzed, the relationship between the test error of the aero-engine rotating component performance tracking model and the training data is compared, and the error source is analyzed in depth. It is determined whether there is a significant difference between the test data errors corresponding to the transition state and the steady state.

[0065] A04, the adaptive weight loss function construction condition is determined, and the adaptive weight X calculation formula is determined.

[0066] In view of the phenomenon that the transition state data error is significantly higher than the steady state data, an adaptive weight loss function for the rotating component is constructed, the loss weight corresponding to the transition state data is dynamically adjusted during training, that is, by increasing the loss weight of the data corresponding to the transition state, the prediction accuracy of the prediction model for the transition state is improved, and the peak error is reduced.

[0067] Specifically, on the basis of the traditional deep neural network training plan, the calculation method of the loss function is changed, and an adaptive weight X is added to the loss. The adaptive weight loss function construction condition is as follows:

[0068] 1) The adaptive weight X is a function of itself and the transition state-steady state transition judgment parameter, X t =f(X t-1 ,n t ),X t is the adaptive weight of the current time step, X t-1 is the adaptive weight value of the previous time step, and n t is the transition state-steady state transition judgment parameter, i.e., the rotating speed.

[0069] 2) When the transition state starts, the adaptive weight gradually increases until the upper limit X max .

[0070] 3) When the transition state ends and enters the steady state, the adaptive weight gradually decreases until X min= 1. Further, an adaptive weight loss is constructed by using the idea of recursive sequence, and the construction process is as follows:

[0071] A41, a sequence N is constructed for storing the reciprocal parameter of the aero-engine speed, and the elements N i The domain is [0, 1].

[0072] A42, suppose a recursive sequence X, each element in which represents the loss weight corresponding to the current time running data, the elements in the recursive sequence X are real numbers between [1, a], X0=1; and the adjacent elements in the recursive sequence X have a recursive relationship X t = f(X t-1 , N t ).

[0073] A43, the recursive relationship X t = f(X t-1 , N t ) is constructed and satisfies: when N i far away from 0, X monotonically increases, and in the 2a step iteration part, monotonically increases from the lower limit 1 to the upper limit a; when N i close to 0, X monotonically decreases, and in the a 2 step iteration part, monotonically decreases from the upper limit a to the lower limit 1.

[0074] A44, the recursive relationship is obtained:

[0075]

[0076] In the formula, the specific values of a1, a2, a3, a4, b1, c1, c2, c3, c4 can be adjusted according to the training parameter characteristics.

[0077] A05, according to the running mechanism of aero-engine, an adaptive weight loss function is constructed:

[0078]

[0079] In the formula, X is the adaptive weight, f(x) is the performance parameter prediction value calculated by the engine rotating part performance tracking model, and y is the real value of the engine performance parameter.

[0080] A06, according to the adaptive weight loss function in step A05, the pre-trained model is trained again. The step A06 includes:

[0081] A61, the adaptive weight of each training sample is calculated, and the corresponding training sample and adaptive weight are saved.

[0082] A62, the upper limit and lower limit of the adaptive weight are set according to the adaptive weight calculation formula.

[0083] A63, the training sample and the adaptive weight are substituted into the engine rotating component performance tracking model, secondary training is started, and a final engine rotating component performance tracking model is obtained.

[0084] The application further discloses an electronic device, comprising: one or more processors; a storage device for storing one or more programs; and when the one or more programs are executed by the one or more processors, the one or more processors implement the weight adaptive training method of the aero-engine digital twin model. The electronic device in the embodiment can include but is not limited to mobile terminals such as mobile phones, notebook computers, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablets), PMPs (portable multimedia players), vehicle terminals (such as vehicle navigation terminals), and the like, and fixed terminals such as digital TVs, desktop computers, and the like.

[0085] The application further discloses a computer readable medium, wherein the readable medium stores a computer program, and the computer program is executed by a processor to implement the weight adaptive training method of the aero-engine digital twin model. Embodiments of the present disclosure include a computer program product comprising a computer program carried on a non-transitory computer readable medium, and the computer program comprises a program code for executing the flowchart Figures 1-2 The program code of the method shown.

[0086] It should be noted that the computer readable medium of the present disclosure can be a computer readable signal medium or a computer readable storage medium or any combination of the two. The computer readable storage medium may, for example, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or component, or any combination of the above. More specific examples of the computer readable storage medium can include, but are not limited to, an electrical connection with one or more conductive wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0087] In this disclosure, a computer readable storage medium can be any tangible medium that contains or stores a program used by or in connection with an instruction execution system, apparatus, or device. In this disclosure, a computer readable signal medium can include a propagated data signal with computer readable program code embodied therein, for use by or in connection with an instruction execution system, apparatus, or device. The propagated data signal can take any number of forms, including, but not limited to, electro-magnetic, optical, or any suitable combination thereof. A computer readable signal medium can be any computer readable medium that is not a computer readable storage medium and that can communicate, propagate or transport program code for use by or in connection with an instruction execution system, apparatus, or device. Program code embodied on a computer readable medium can be transmitted using any appropriate medium, including but not limited to wireless, wire line, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

[0088] In some embodiments, the client, server, or both can communicate using any current known or future developed network protocol, such as HTTP (HyperText Transfer Protocol), and can be interconnected with any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include local area networks ("LAN"), wide area networks ("WAN"), the Internet, and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any current known or future developed networks.

[0089] The computer readable medium described above can be included in the electronic device described above; alternatively, the computer readable medium can exist as a separate entity, which is not incorporated in the electronic device.

[0090] Example 1 is directed to adaptive weight loss function construction for aero-engine rotating parts

[0091] S01: Normalize and preprocess the aero-engine data and classify it, and the specific classification results are as follows:

[0092] Environmental parameters: inlet air temperature, air pressure height, etc.

[0093] Control parameters: inlet guide vane angle position, nozzle position, throttle lever position, etc.

[0094] State parameters: high and low pressure rotor speed, oil pressure, etc.

[0095] Performance parameters: thrust, specific fuel consumption.

[0096] S02: build an aero-engine rotating component performance tracking model from environmental parameters, control parameters and state parameters to performance parameters, and use the data after step S01 preprocessing to pre-train the model to obtain a pre-trained model.

[0097] S03: error analysis on the pre-trained model obtained in step S02, due to the characteristics of aero-engine operation data, i.e. the amount of steady-state data is significantly greater than the amount of transition-state data, the error between the two is compared and analyzed. Determine whether there is a significant difference between the test data errors corresponding to the transition state and the steady state. According to the analysis result, determine whether it is necessary to use the method of constructing an adaptive weight loss function to perform secondary optimization training on the model:

[0098] If the error difference between the two is large, go to step S04;

[0099] If the error difference between the two is small, the pre-trained model is used as the final aero-engine rotating component performance tracking model.

[0100] S04: According to the analysis of the training results, it can be concluded that the peak error of the aero-engine rotating component performance tracking model mostly occurs in the transition state operation and a certain time after its end, and the prediction error for the transition state is generally higher than that for the steady-state data. According to the operation mechanism of the aero-engine, an adaptive weight loss function is constructed to meet the following conditions:

[0101] 1) The adaptive weight X is a function of itself and the transition state-steady state transition judgment parameter, X t =f(X t-1 ,n t )。X t is the adaptive weight of the current time step; X t-1 is the adaptive weight value of the previous time step; n t is the transition state-steady state transition judgment parameter, i.e. the speed at the current time (this embodiment takes a mechanical rotating component as an example). That is, the adaptive weight X t is updated and iterated based on the adaptive weight value X t-1 of the previous time step according to the speed n t at the current time.

[0102] 2) When the transition state starts, the adaptive weight gradually increases until the upper limit X max .

[0103] 3) When the transition state ends and enters the steady state, the adaptive weight gradually decreases until X min =1.

[0104] S05: According to the requirements proposed in S04, according to the specific operation mechanism of the aero-engine, the rotating speed is taken as the basis for judging whether the aero-engine enters the transition state, specifically, when the rotating speed derivative of the aero-engine significantly increases, for example, after the maximum and minimum normalization of the rotating speed derivative of the aero-engine, when the value increases from less than 0.01 to more than 0.1, it can be considered that there is a significant increase, and it can be considered that the aero-engine enters the transition state.

[0105] Adopt the idea of recursive sequence in mathematics to construct an adaptive weight loss function:

[0106] 1) Construct a sequence N to store the maximum and minimum normalized aero-engine rotating speed derivative parameters, and the domain of the elements N[i] of the sequence N is [0, 1].

[0107] 2) Assume that there is a recursive sequence X, each element in which represents the loss weight corresponding to the current time running data, the elements in the sequence X are real numbers between 1 and 10, including 1 and 10, and X[0] is x0=1. At the same time, adjacent elements in the array X have a recursive relationship X t =f(X t-1 , N t ), where N t is a certain element in the array N shown above.

[0108] 3) Construct the recursive relationship f, that is, X t =f(X t-1 , N t ), and must ensure that the following requirements must be met: when N[i] is far from 0, X is monotonically increasing (the increasing speed is fast first and then slow), and in the 20-step iteration part, it can be monotonically increased from the lower limit 1 to approach the upper limit 10, when N[i] is close to 0, X is monotonically decreasing (the decreasing speed is slow first and then fast), and in the 100-step iteration part, it can be monotonically reduced from the upper limit 10 to approach the lower limit 1.

[0109] 4) According to the above requirements, the recursive relationship f and the adaptive weight loss function MSE are given in the following form:

[0110]

[0111]

[0112] The multiple parameters in the above formula are only for reference, the specific values need to be adjusted according to the training parameter characteristics, and the relationship f that meets the conditions is not unique.

[0113] S06: According to the formula shown in S05, assuming that the speed of an aero-engine increases from 0 to the cruising speed, and after a period of smooth operation, the speed decreases to 0, which is used as an example to calculate and verify the specific changes of the weight, and the weight calculation result of the adaptive weight loss function is shown in the figure Figure 4 .

[0114] S07: Apply the adaptive weight loss function described above to the pre-trained model for secondary optimization training.

[0115] Embodiment 2: A deep learning neural network training method with weight pre-insertion

[0116] This example is mainly aimed at S07 of the process in embodiment 1. In the training scheme of traditional neural network, in order to prevent the model from memorizing the order of training samples and causing overfitting, and improve the generalization ability of the model, it is usually necessary to perform randomization processing on the training set input. If the traditional method is used to calculate the weight at the same time of calculating the loss, it will conflict with the idea of dynamic iteration of the adaptive weight loss function proposed in example one, which uses the value of the previous time step according to the time sequence. The specific embodiment scheme is shown as follows:

[0117] S08: Design an adaptive weight loss function to make different samples have different effects on the training of the model. For example, for some samples, because their noise is larger or difficult to handle, a high weight can be set to force the model to pay more attention to these samples. For details, please refer to embodiment 1.

[0118] S09: After designing the adaptive weight loss function, the adaptive weight of each sample needs to be calculated and saved. Generally, this part of the calculation can be completed offline, as it does not need to access network data. The adaptive weight of each sample can be calculated by traversing the entire data set. The calculation method can be customized according to specific problems, for example, the sample difficulty, sample noise level, etc. can be used as the basis for adaptive weight calculation.

[0119] S10: According to the adaptive weight calculation formula, the upper and lower limits of the adaptive weight can be set to ensure that the adaptive weight value is not too high or too low. After calculating the adaptive weights of all samples, the adaptive weight values need to be saved in a file. In the actual training process, these saved adaptive weight parameters can be directly loaded.

[0120] S11: Before starting the training, the training data, test data and pre-calculated sample adaptive weights need to be prepared. During the training process, the saved sample adaptive weight parameters can be directly called for training. The advantage of this method is that it can greatly improve the training speed and efficiency. In addition, this method can also facilitate effective hyperparameter adjustment and training code debugging.

[0121] S12: In the training process, the sample data and its corresponding adaptive weight need to be input into the neural network together. When calculating the loss function, the sample adaptive weight needs to be multiplied by the loss value to make some samples have greater influence on the training of the model. In the back propagation process, the sample adaptive weight is also needed to update the neural network parameters. It should be noted that when using the adaptive weight loss function, the training data needs to be properly sampled and divided. For example, the entire data set can be divided into several subsets, each subset corresponding to an adaptive weight loss function. At the same time, the sample weight needs to be updated regularly to ensure that the model can learn more comprehensive and accurate features.

[0122] The above-described embodiments are only to describe the preferred modes of the present application, and not to limit the scope of the present application. Without departing from the design spirit of the present application, various modifications, variations, modifications, and replacements of the technical solutions of the present application made by those skilled in the art shall fall within the protection scope determined by the claims of the present application.

Claims

1. A weight self-adaptive training method of an aero-engine digital twin model, characterized in that: The engine rotating component performance tracking model is built and pre-trained, and the relationship between the test error of the model and the training data is analyzed; for the transition state data with test error obviously higher than the steady state data, an adaptive weight loss function is constructed, the adaptive weight corresponding to each training sample is calculated and saved, and the loss weight corresponding to the transition state data is dynamically adjusted during re-training to obtain the final engine rotating component performance tracking model; The method comprises the following steps: A01, normalizing and classifying the aero-engine data; A02, building an aero-engine rotating component performance tracking model from engine environment parameters, control parameters and state parameters to performance parameters, and pre-training the model using the data processed in step A01 to obtain a pre-trained model; the aero-engine rotating component performance tracking model learns the mapping relationship between engine control parameters, environment parameters, state parameters and engine performance parameters; A03, error analysis is performed on the pre-trained model obtained in step A02 to determine the test error between the transition state data and the steady state data; when the test error is less than m, the pre-trained model is used as the final engine rotating component performance tracking model; when the test error is greater than or equal to m, step A04 is entered; A04, determining the adaptive weight loss function construction condition and the adaptive weight X calculation formula; The adaptive weight loss function construction condition is as follows: 1) Adaptive weight X is a function of itself and the transition- steady state transition judging parameter, X t = f(X t-1 , n t ), X t is the adaptive weight of the current time step, X t-1 is the adaptive weight value of the previous time step, n t is the transition- steady state transition judging parameter, i.e. the rotational speed; 2) When the transition state begins, the adaptive weight is gradually raised until the upper limit X max ; 3) When the transition ends and the steady state is entered, the adaptive weights are gradually reduced until X min = 1; The adaptive weight loss function construction process is as follows: A41. Constructing a sequence N for storing the reciprocal of the engine speed parameter, whose elements N i The domain is [0, 1]. A42、Suppose a recursive sequence X, each element in the sequence represents the loss weight corresponding to the running data at the current time, the elements in the recursive sequence X are real numbers between the interval [1, a], X0=1; and, adjacent elements in the recursive sequence X have a recursive relationship X t =f(X t-1 , N t ); A43、constructing recursive relation X t = f(X t-1 , N t ) and satisfies: when N i far from 0, X monotonically increases, and in the 2a step iteration part, monotonically increases from the lower limit 1 to approximate the upper limit a; when N i close to 0, X monotonically decreases, and in the a 2 step iteration part, monotonically decreases from the upper limit a to approximate the lower limit 1; A44, the recursive relationship is obtained: In the formula, the specific values of a1, a2, a3, a4, b1, c1, c2, c3 and c4 can be adjusted according to the training parameter characteristics; The adaptive weight loss function is as follows: In the formula, X is the adaptive weight, f(x) is the performance parameter prediction value calculated by the engine rotating component performance tracking model, and y is the true value of the engine performance parameter; A05, according to the running mechanism of the aero-engine, the adaptive weight loss function is constructed: In the formula, X is the adaptive weight, f(x) is the performance parameter prediction value calculated by the engine rotating component performance tracking model, and y is the true value of the engine performance parameter; A06, according to the adaptive weight loss function in step A05, the pre-trained model is subjected to secondary optimization training; The step A06 comprises: A61, calculating the adaptive weight of each training sample and saving the one-to-one corresponding training sample and adaptive weight; A62, setting the upper limit and lower limit of the adaptive weight according to the adaptive weight calculation formula; A63, substituting the training sample and the adaptive weight into the engine rotating component performance tracking model to start secondary training and obtain the final engine rotating component performance tracking model.

2. The weight adaptive training method of an aero-engine digital twin model according to claim 1, characterized in that: In the step A01, the engine data is divided into environment parameters, control parameters, state parameters and performance parameters; the environment parameters include inlet air temperature and air pressure height; the control parameters include inlet guide vane angle position, nozzle position and throttle lever position; the state parameters include high and low pressure rotor speed and oil pressure; and the performance parameters include thrust and specific fuel consumption.

3. An electronic device, comprising: The method comprises: one or more processors; a storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement a weight adaptive training method of an aero-engine digital twin model as claimed in any one of claims 1-2.

4. A computer readable medium storing a computer program, characterized in that: The computer program, when executed by a processor, implements a weight adaptive training method of an aero-engine digital twin model as claimed in any one of claims 1-2.

Images