Training of an egt prediction model, method, device, medium and product for prediction of egt

Abstract

The application discloses a kind of EGT prediction model training, EGT prediction method, equipment, medium and product.The training method includes: collecting cruise message data of aero-engine, constructs standard time series training sample set;Standard time series training sample is input into full connection neural network in sequence to obtain EGT predicted value by training;According to the number of output layer nodes, full connection neural network is split into multiple sub neural networks;According to current standard EGT label in current standard time series training sample and current EGT predicted value, the gradient vector of each sub neural network back propagation is spliced into complete Jacobian matrix;According to complete Jacobian matrix, the adjustment factor and momentum item of previous round transmission, the parameter update of full connection neural network in current round and the transmission of next round model training initial value are carried out, until aero-engine EGT prediction model is obtained, and the stability of EGT prediction model convergence is improved in the embodiment technical scheme.

Description

Technical Field

[0001] This invention relates to the field of aircraft engine health management and preventive maintenance technology, and in particular to the training of an EGT (Exhaust Gas Temperature) prediction model, EGT prediction methods, equipment, media and products. Background Technology

[0002] Aero engines are the core power components of aircraft, and the degradation of their aero-engine performance is a critical issue during engine operation, directly impacting fuel efficiency, on-wing service life, and flight safety. Engine torque factor (EGT) is a key parameter characterizing the health of the engine's aero-engine aero-engine's aero-engine components. Performance degradation caused by contamination and wear of these components directly manifests as a continuous increase in EGT and a continuous decrease in EGT margin. Engine washing is a core preventative maintenance method for restoring aero-engine aero-engine performance and reducing EGT; the timing of washing directly determines its effectiveness and operational economics.

[0003] Currently, the industry mainly uses two methods to determine the timing of engine water washing: one is an experience-based method based on fixed flight cycles and fixed calendar periods. This method cannot adapt to the performance degradation differences of different engines and operating conditions, and is prone to problems such as washing too early, increasing unnecessary maintenance costs and downtime, or washing too late, leading to irreversible engine performance degradation. The other method is a method based on a fixed EGT threshold. This method can only make a post-event judgment after the engine performance has reached the degradation threshold, and cannot predict the performance degradation trend in advance, making it difficult to achieve a proactive and accurate decision on the optimal water washing time. Summary of the Invention

[0004] This invention provides an EGT prediction model training method, device, medium, and product for EGT prediction. It constructs an EGT prediction model for aero-engines using an improved LM (Levenberg-Marquardt) algorithm based on a dual adaptive adjustment mechanism of gradient descent direction and convergence step size, thereby achieving high-precision time-series prediction of aero-engine EGT.

[0005] According to one aspect of the present invention, a method for training an EGT prediction model for an aero-engine is provided, the method comprising: Collect cruise message data from ACARS (Aircraft Communications Addressing and Reporting System) messages of aircraft engines during their service life, and construct a standard time-series training sample set; The standard time-series training samples include a first number of gas path parameter groups and a second number of standard EGT labels sorted by acquisition time, where the first number and the second number are both integers greater than 1. The current standard time-series training samples are sequentially obtained from the standard time-series training sample set and input into the pre-built fully connected neural network for the current round of model training to obtain a second number of EGT prediction values, wherein the number of output layer nodes of the fully connected neural network matches the second number. Based on the number of output layer nodes in the fully connected neural network, the fully connected neural network is divided into a second number of sub-neural networks with single-node outputs; Based on the second number of current standard EGT labels and the second number of current EGT predictions in the current standard time-series training samples, the gradient vector of backpropagation of each sub-neural network is solved in turn, and the solved gradient vectors are concatenated into a complete Jacobian matrix. Based on the complete Jacobian matrix, the adjustment factor and momentum term passed from the previous round, a dual adaptive adjustment algorithm of gradient descent direction and convergence step size is used to calculate the current parameter adjustment vector, the current adjustment factor and the current momentum term in the current round. Adjust the vector based on the current parameters, perform a round of parameter updates on the fully connected neural network, and use the current adjustment factor and the current momentum term as the initial values ​​for the next round of model training; Return to the operation of sequentially obtaining the current standard time series training samples in the standard time series training sample set until the end of the iteration condition is met, and obtain the EGT prediction model of the aero-engine.

[0006] According to another aspect of the present invention, a method for predicting the EGT of an aero-engine is also provided, the method comprising: The cruise message data in the ACARS message of the target aero-engine to be predicted is acquired in real time, and a prediction input sequence is constructed based on the cruise message data, wherein the prediction input sequence includes a first number of target gas path parameter groups sorted according to the acquisition time. The first number of target air path parameter groups are input into the aero-engine EGT prediction model trained by the method described in any one of the embodiments of the present invention; Obtain the target EGT prediction value at the second number of future time periods output by the aero-engine EGT prediction model; Based on the predicted EGT values ​​for each target and the factory-baseline EGT values ​​matched with the target aero-engine, calculate the EGT margin values ​​for a second number of future times. If a preset number of consecutive EGT margin values ​​are all less than the preset water washing trigger threshold, then the target aero-engine is determined to have reached the optimal water washing time.

[0007] According to another aspect of the present invention, a training apparatus for an aero-engine EGT prediction model is also provided, the apparatus comprising: The training sample set construction module is used to collect cruise message data from ACARS messages of aero engines during their service life and construct a standard time-series training sample set. The standard time-series training samples include a first number of gas path parameter groups and a second number of standard EGT labels sorted by acquisition time, where the first number and the second number are both integers greater than 1. The EGT prediction value acquisition module is used to sequentially acquire the current standard time series training samples in the standard time series training sample set, input them into the pre-built fully connected neural network for the current round of model training, and obtain a second number of EGT prediction values, wherein the number of output layer nodes of the fully connected neural network matches the second number. The network splitting module is used to split a fully connected neural network into a second number of sub-neural networks with single-node outputs, based on the number of output layer nodes included in the fully connected neural network. The Jacobian matrix concatenation module is used to solve the gradient vector of the backpropagation of each sub-neural network in turn based on the second number of current standard EGT labels and the second number of current EGT predictions in the current standard time series training samples, and concatenate the solved gradient vectors into a complete Jacobian matrix. The current parameter calculation module is used to calculate the current parameter adjustment vector, current adjustment factor, and current momentum term in the current round based on the complete Jacobian matrix, the adjustment factor and momentum term passed from the previous round, and the dual adaptive adjustment algorithm of gradient descent direction and convergence step size. The parameter update module is used to adjust the vector according to the current parameters, perform a round of parameter update for the fully connected neural network, and use the current adjustment factor and the current momentum term as the initial values ​​for the next round of model training. The return execution module is used to return the operation of sequentially obtaining the current standard time series training samples in the standard time series training sample set until the end iteration condition is met, and the aero-engine EGT prediction model is obtained.

[0008] According to another aspect of the present invention, an apparatus for predicting the EGT of an aero-engine is also provided, the apparatus comprising: The prediction input sequence construction module is used to acquire cruise message data in the ACARS message of the target aero-engine to be predicted in real time, and construct a prediction input sequence based on the cruise message data, wherein the prediction input sequence includes a first number of target gas path parameter groups sorted according to the acquisition time. The air path parameter group input module is used to input the first number of target air path parameter groups into the aero-engine EGT prediction model trained by the method described in any one of the embodiments of the present invention; The prediction value acquisition module is used to acquire the target EGT prediction value at a second number of future times output by the aero-engine EGT prediction model. The EGT margin calculation module is used to calculate the EGT margin value for a second number of future times based on the predicted EGT values ​​of each target and the factory-baseline EGT value matched with the target aero-engine. The optimal water washing timing prediction module is used to determine that the target aero-engine has reached the optimal water washing timing if all EGT margin values ​​are less than the preset water washing trigger threshold.

[0009] According to another aspect of the present invention, an electronic device is also provided, the electronic device comprising: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform a training method for an aero-engine EGT prediction model as described in any embodiment of the present invention, or to perform a prediction method for an aero-engine EGT as described in any embodiment of the present invention.

[0010] According to another aspect of the present invention, a computer-readable storage medium is also provided, the computer-readable storage medium storing computer instructions, the computer instructions being configured to cause a processor to execute and implement, either, a training method for an aero-engine EGT prediction model as described in any embodiment of the present invention, or a prediction method for an aero-engine EGT as described in any embodiment of the present invention.

[0011] According to another aspect of the present invention, a computer program product is also provided, including a computer program that, when executed by a processor, implements the steps of the method as described in any embodiment of the present invention.

[0012] The technical solution of this invention involves collecting cruise message data from ACARS messages of an aero-engine during its service life, constructing a standard time-series training sample set, and inputting it into a pre-constructed fully connected neural network for model training. During each training round, the fully connected neural network is divided into a second number of sub-neural networks with single-node outputs, based on the number of output layer nodes included in the fully connected neural network. The gradient vectors of each sub-neural network are sequentially calculated and concatenated to form a complete Jacobian matrix. Based on the complete Jacobian matrix, the adjustment factor and momentum term passed from the previous round, a dual adaptive adjustment algorithm of gradient descent direction and convergence step size is used to calculate the current gradient descent direction and convergence step size for each new round. The method involves adjusting the parameter vector, the current adjustment factor, and the current momentum term. Based on the current parameter adjustment vector, a round of parameter updates is performed on the fully connected neural network. The current adjustment factor and the current momentum term are used as the initial values ​​for the next round of model training until the EGT prediction model is obtained. This technique significantly improves the convergence stability and global optimization capability of the aero-engine EGT prediction model, making the scheme more suitable for the engineering requirements of real-time on-wing prediction and online judgment of engine water washing timing for aero-engines. It completes the full-process adaptation for the nonlinear and small-sample characteristics of EGT time series data, effectively achieving high-precision prediction of aero-engine EGT and providing reliable quantitative data support for engine water washing timing judgment.

[0013] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart of a training method for an EGT prediction model of an aero-engine provided according to an embodiment of the present invention; Figure 2 This is a flowchart of a method for predicting the EGT of an aero-engine according to an embodiment of the present invention; Figure 3 This is a spline curve plot of the raw exhaust temperature data for a certain civil passenger aircraft engine. Figure 4 This is a graph of engine exhaust temperature data after smoothing and normalization. Figure 5This is a schematic diagram illustrating the splitting of the Jacobian matrix in a fully connected neural network, as applicable to an embodiment of the present invention. Figure 6 This is a comparison chart of the loss descent curves of the DAF-LM algorithm in this embodiment of the invention with other optimization algorithms; Figure 7 This is a comparison chart of the prediction evaluation metrics of the DAF-LM algorithm and other optimization algorithms in this embodiment of the invention.

[0016] Figure 8 This is a graph comparing the prediction results of the DAF-LM algorithm and other optimization algorithms with the actual data in an embodiment of the present invention.

[0017] Figure 9 This is a comparison chart of the prediction curves and actual curves of the fully connected neural network trained based on the DAF-LM algorithm in this embodiment of the invention with those of other mainstream prediction models.

[0018] Figure 10 This is a schematic diagram of the structure of a training device for an EGT prediction model of an aero-engine provided according to an embodiment of the present invention; Figure 11 This is a schematic diagram of the structure of a predictive device for EGT of an aero-engine according to an embodiment of the present invention; Figure 12 This is a schematic diagram of the structure of an electronic device that implements the training method for the EGT prediction model of an aero-engine, or the prediction method for the EGT of an aero-engine, according to embodiments of the present invention. Detailed Implementation

[0019] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0020] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0021] Figure 1 The flowchart below illustrates a training method for an aero-engine EGT prediction model according to an embodiment of the present invention. This embodiment is applicable to situations where an aero-engine EGT prediction model is trained by splitting the Jacobian matrix and using an improved LM algorithm with a dual adaptive adjustment mechanism of gradient descent direction and convergence step size. This method can be executed by a training device for the aero-engine EGT prediction model, which can be implemented in hardware and / or software and is generally configured in an electronic device with data processing capabilities.

[0022] Correspondingly, such as Figure 1 As shown, the method includes: S110: Collect cruise message data from ACARS messages of the aircraft engine during its service life and construct a standard time-series training sample set.

[0023] The standard time-series training samples include a first number of gas path parameter groups and a second number of standard EGT labels sorted by acquisition time. Both the first and second numbers are integers greater than 1.

[0024] In this embodiment, ACARS messages refer to aircraft operational data messages collected from aero engines, automatically pushed by an aircraft (typically a civil airliner) during flight to airlines, maintenance personnel, or ground monitoring centers via air-to-ground radio or satellite. Service life refers to the entire lifecycle of the engine, from installation and delivery to overhaul and eventual scrapping. Therefore, the data collected during service life is actual measured data from high-altitude flight. Cruise message data refers to ACARS data packets collected and transmitted by the aircraft when it enters its steady-state cruise range.

[0025] Each cruise message mainly includes: timestamp information, EGT data, and a set of gas path parameters associated with the EGT data. The gas path parameter set includes parameters such as low-pressure rotor speed, high-pressure rotor speed, fuel flow rate, total atmospheric temperature, flight altitude, lubricating oil pressure, lubricating oil temperature, and Mach number.

[0026] The standard time-series training sample set includes multiple standard time-series training samples. Each standard time-series training sample includes a first number of gas path parameter groups and a second number of standard EGT tags sorted by acquisition time.

[0027] One standard time-series training sample is used to label the air path parameter set collected by the aero-engine at the first number of acquisition times, and the EGT data at the next second number of future times (i.e., the standard EGT label).

[0028] In a specific example, assume the aircraft collects and sends ACARS messages every ten minutes. The first quantity is 4, and the second quantity is 2. ACARS messages from time t1 to t6 can be obtained in the order of their collection time extension, with a ten-minute interval between each pair of collections. That is, t1 is ten minutes earlier than t2, and t2 is ten minutes earlier than t3. Then, the gas path parameter group 1 collected at time t1, the gas path parameter group 2 collected at time t2, the gas path parameter group 3 collected at time t3, the gas path parameter group 4 collected at time t4, the EGT data 1 collected at time t5, and the EGT data 2 collected at time t6 can be used to construct a standard time-series training sample.

[0029] Accordingly, in an optional implementation of this embodiment, collecting cruise message data from ACARS messages during the service life of the aircraft engine to construct a standard time-series training sample set may include: S1101. Construct an EGT time-series prediction dataset based on cruise message data in ACARS messages during the service life of the aircraft engine.

[0030] Each EGT time-series prediction data includes multiple feature sets sorted by acquisition time. Each feature set includes EGT data matching the set acquisition time and a set of gas path parameters associated with that EGT data.

[0031] Specifically, multiple EGT time-series prediction data can be obtained based on ACARS messages from the aero engines of multiple civil airliners during their service life, and these multiple EGT time-series prediction data can be combined to form an EGT time-series prediction dataset. For example, an EGT time-series prediction data X specifically includes: a set of 600 feature values ​​collected from the aero engines of civil airliner A over 600 consecutive acquisition times.

[0032] Based on the above embodiments, an EGT time-series prediction dataset is constructed using cruise message data from ACARS messages during the service life of the aero-engine, which may further include: Based on the cruise message data in the ACARS messages of the aero-engine during its service life, the original EGT time series prediction dataset is obtained; the exponential smoothing algorithm is used to smooth the various feature values ​​in the original EGT time series prediction dataset, and the smoothed feature values ​​are normalized to obtain the EGT time series prediction dataset.

[0033] In this optional implementation, the EGT time-series raw prediction dataset includes multiple EGT time-series raw prediction datasets. Continuing the previous example, suppose we first construct a time-series raw prediction dataset based on 600 feature values ​​collected over 600 consecutive acquisition times. Specifically, this time-series raw prediction dataset contains 600 EGT data points, 600 low-pressure rotor speeds, and so on. For each type of data, smoothing processing can be performed first based on a predefined exponential smoothing algorithm.

[0034] Specifically, it can be calculated according to the preset exponential smoothing formula: The raw data acquired at acquisition time t is calculated for a specific type of data x (e.g., EGT data or low-pressure rotor speed). Data after smoothing ,in, The preset smoothing coefficient, This refers to the smoothed data of a specific type of data x at the (t-1)th acquisition time.

[0035] Taking the smoothing of 600 EGT data points as an example, the EGT data at the earliest acquisition time can be directly initialized as the first smoothed data. Then, based on the first smoothed data and the above calculation formula, the smoothing results of the remaining 599 EGT data points can be determined one by one to finally complete the smoothing of various feature values ​​in the original EGT time series prediction dataset.

[0036] After smoothing each type of data, normalization can be further performed on each type of data. Taking EGT data as an example again, assuming 600 EGT data points have already been smoothed, then the minimum-maximum normalization formula can be used: Calculate the smoothed data at acquisition time t. Corresponding normalization results ,in, The minimum value among 600 smoothed EGT data points. This represents the maximum value among 600 smoothed EGT data points. By performing the above operations on each type of smoothed data, the final EGT time series prediction dataset can be obtained.

[0037] S1102. In each EGT time series prediction data in the EGT time series prediction dataset, extract multiple data subsequences of target data lengths according to the time series, wherein the target data length is the sum of the first quantity and the second quantity.

[0038] In this embodiment, after determining the first quantity N1 and the second quantity N2, the sum of N1 + N2, N, can be calculated as the target data length. Then, a non-repeating subsequence of length N can be extracted from each EGT time-series prediction data. For example, when N is 6, a data subsequence can include six feature sets obtained from collecting data on the aero-engines of a designated civil aircraft at six consecutive acquisition times.

[0039] S1103. In each data subsequence, retain the gas path parameter group in the first number of feature value sets and retain the EGT data as standard EGT labels in the second number of feature value sets to obtain the standard time series training sample set.

[0040] S120. Sequentially obtain the current standard time-series training samples from the standard time-series training sample set, input them into the pre-built fully connected neural network for the current round of model training, and obtain the second number of EGT prediction values, wherein the number of output layer nodes of the fully connected neural network matches the second number.

[0041] In this embodiment, the fully connected neural network model architecture can be selected according to actual needs. For example, a four-layer fully connected neural network framework can be used, with the number of neurons (or nodes) in the input layer consistent with the first number length in the standard time-series training samples, for example, it can be set to 9. In this case, it is necessary to obtain the gas path parameter sets under 9 consecutive sampling times. Then, the number of neurons in the first hidden layer and the second hidden layer can be set according to actual needs. For example, the number of neurons in the first hidden layer can be set to 18, and the number of neurons in the second hidden layer can be set to 6, etc.

[0042] The number of neurons (or nodes) in the output layer is matched to a second set of time steps to output the EGT predictions for those second set of time steps. In a specific example, assuming the second set of time steps is 3, the time interval between any two time steps corresponds to the sampling time interval between any two pairs of gas path parameter groups input into the fully connected neural network, for example, 10 minutes. In this case, the fully connected neural network can obtain the EGT predictions at T+10 minutes, T+20 minutes, and T+30 minutes for the input of the current standard time-series training samples (9 sets of gas path parameter groups collected at 9 consecutive sampling times, with the last set of gas path parameter groups collected at time T).

[0043] Furthermore, the activation function in this fully connected neural network can be the Sigmoid function, the loss function can be RMSE (Root Mean Square Error), and the optimizer can be the dual adaptive adjustment algorithm for gradient descent direction and convergence step size proposed in this embodiment of the invention, hereinafter referred to as the DAF-LM (Dual Adaptive Fast Levenberg-Marquardt) algorithm.

[0044] S130. Based on the number of output layer nodes included in the fully connected neural network, the fully connected neural network is divided into a second number of sub-neural networks with single-node outputs.

[0045] In this embodiment, the output layer of a multi-output fully connected neural network can be split based on the number of output layer nodes, while maintaining the sharing of the network backbone structure and parameters, generating a corresponding number of single-output sub-neural networks. That is, all sub-neural networks completely reuse the network structure and weight parameters of the original fully connected neural network's input layer, first hidden layer, and second hidden layer. After splitting, a second number of structurally identical, output-independent single-node output sub-neural networks are obtained. Each sub-neural network is responsible for EGT prediction at only one time step, and all sub-networks share the hidden layer feature extraction capability, with only the output layer nodes being independent. Through this setup, the computational complexity of high-dimensional matrices can be effectively reduced.

[0046] Continuing the previous example, when the second number is 3, the first sub-neural network retains only the first output node of the output layer to output the EGT prediction value at T+10 minutes; the second sub-neural network retains only the second output node of the output layer to output the EGT prediction value at T+20 minutes; and the third sub-neural network retains only the third output node of the output layer to output the EGT prediction value at T+30 minutes.

[0047] S140. Based on the second number of current standard EGT labels and the second number of current EGT prediction values ​​in the current standard time-series training samples, solve the gradient vector of backpropagation for each sub-neural network in sequence, and concatenate the solved gradient vectors into a complete Jacobian matrix.

[0048] In an optional implementation of this embodiment, based on a second number of current standard EGT labels and a second number of current EGT prediction values ​​in the current standard time-series training samples, the gradient vector of the backpropagation of each sub-neural network is solved sequentially, and the solved gradient vectors are concatenated into a complete Jacobian matrix, which may include: S1401. In multiple sub-neural networks, the target neural network is obtained sequentially.

[0049] S1402. Calculate the target residual value of the target neural network based on the current EGT prediction value that matches the target neural network and the current standard EGT label.

[0050] In a specific example, suppose target neural network 1 is used to predict the EGT prediction value at the first time step. Then, it can obtain the current standard EGT label that matches the first time step from the current standard time-series training samples. And from the second number of current EGT predictions, obtain the current EGT prediction that matches the first time step. Then, using the formula: The target residual value of target neural network 1 is calculated. The target residual value Equivalent to setting a loss function (For example, RMSE) The derivative of the output y, that is: .

[0051] S1403. Based on the target residual value, use the backpropagation chain derivative algorithm to solve for the partial derivatives of the target loss function used by the fully connected neural network with respect to all weight parameters and bias parameters of the target neural network.

[0052] Specifically, after obtaining the target residual value, the gradient (or partial derivative) can be calculated first for each weight and bias in the output layer. For example, the output layer weights... The gradient can be expressed as: Output layer bias The gradient can be expressed as: ,in, This is the input vector for the output layer of the target neural network 1. After the output layer calculation is completed, the error signal can be backpropagated using the chain rule. Combining the derivative properties of the sigmoid activation function, the weight gradient and bias gradient of each non-output layer can be calculated step by step.

[0053] S1404. Organize all the partial derivatives obtained by the solution into a single-row gradient vector corresponding to the target neural network according to the preset parameter sorting rules.

[0054] In a specific example, if the loss function under the target neural network is defined as Then it can be calculated to be of the form: The single-row gradient vector. , … Let m be the weights in a fully connected neural network. , … Let s be the s biases contained in a fully connected neural network.

[0055] S1405. Determine whether the processing of all sub-neural networks is complete: if yes, proceed to S1406; otherwise, return to S1401.

[0056] S1406. According to the arrangement order of each node in the output layer of the fully connected neural network, the single-row gradient vectors of all sub-neural networks are concatenated to obtain a complete Jacobian matrix covering all weights and all bias parameters of the fully connected neural network.

[0057] Continuing the previous example, assuming the fully connected neural network includes n output layer nodes, it can be further divided into n sub-neural networks, ultimately yielding n single-row gradient vectors. Correspondingly, the complete Jacobian matrix J covering all weights and bias parameters of the fully connected neural network can be expressed as: in, , … Let be the loss function for n sub-neural networks.

[0058] S150. Based on the complete Jacobian matrix, the adjustment factor and momentum term passed from the previous round, the current parameter adjustment vector, the current adjustment factor and the current momentum term in the current round are calculated using a dual adaptive adjustment algorithm of gradient descent direction and convergence step size.

[0059] In an optional implementation of this embodiment, based on the complete Jacobian matrix, the adjustment factor and momentum term passed from the previous round, a dual adaptive adjustment algorithm of gradient descent direction and convergence step size is used to calculate the current parameter adjustment vector, the current adjustment factor, and the current momentum term in the current round. Specifically, this may include: S1501, Obtain the current regulation factor Among them, the current round k Initialize it to the adjustment factor passed down in the previous round k-1.

[0060] S1502, according to the formula The initial parameter adjustment vector for the current round k is calculated. .

[0061] in, For the residual vector, n is the second quantity. , For the j-th current EGT prediction value, For the j-th current standard EGT tag, J is the complete Jacobian matrix, and G is the approximate Hessian matrix. In addition, the parameter adjustment vector It can be understood as a one-dimensional row (column) vector. This parameter adjusts each vector element in the vector, corresponding to a weight or bias in a fully connected neural network.

[0062] In this embodiment, the inventors creatively constructed a quadratic form formula for calculating the parameter adjustment vector (also known as the adaptive step size) by combining the steepest descent method with the standard LM algorithm framework. The residual vector mentioned above... This is the combination of the residual values ​​of each sub-neural network calculated in S140.

[0063] S1503, According to the formula: The current gain rate is calculated. ,in: This represents the parameter vector that has not yet been updated in the current round k. Let be the target loss function of the fully connected neural network. Let be the second-order Taylor expansion of the target loss function.

[0064] in, This can be understood as the parameter vectors passed to the current k round after the model training is completed in the previous round k-1, which includes all the weights and biases in the fully connected neural network.

[0065] Based on the current gain rate As can be seen from the calculation formula, What it actually describes is the loss reduction gain rate. If If the value is greater than 0, the direction of loss reduction is determined to be correct, and the current adjustment factor can be adaptively reduced. This makes the entire algorithm closer to the second-order Gauss-Newton method, thus accelerating the convergence process of the entire model. If If it is less than or equal to 0, then determine the current adjustment factor. Ineffective; the current adjustment factor needs to be increased appropriately. This makes the entire algorithm closer to the first-order gradient descent method, ensuring a stable decrease in the loss value.

[0066] S1504, Determine the current gain rate Is it greater than 0? If yes, then execute S1505; otherwise, execute S1507.

[0067] S1505, Adjust the current adjustment factor according to the first step long adjustment coefficient. Perform a reduction process to obtain the current adjustment factor for the current round, and execute S1506.

[0068] As an example rather than a limitation, the current adjustment factor can be used... Multiply by the first step length adjustment coefficient of 1 / 2 to obtain the current adjustment factor for the current round.

[0069] S1506, According to the formula: The current momentum term under the current round k is calculated. Afterwards, Determined as the current parameter adjustment vector in the current round k Clear the step size adjustment counter and end the calculation process for this round; among which, This refers to the momentum term transferred in the previous round k-1. The preset momentum coefficient.

[0070] In this embodiment, the inventors creatively introduced a step-size momentum-driven mechanism, which dynamically adjusts the step size using historical gradient information. The descent step size is adjusted to effectively reduce loss oscillations.

[0071] Furthermore, this step size adjustment counter is used to record the number of invalid step size adjustments in the current round k, that is, if the current adjustment factor is... Multiple attempts to expand the scope of the measures failed to effectively reduce the losses; that is, If the value is always less than or equal to 0, then a convergence step size algorithm is needed to adjust the step size optimization direction in order to escape the local minimum.

[0072] S1507. Check if the step size adjustment counter has accumulated to the target count value: if not, execute S1508; if yes, execute S1509.

[0073] In this embodiment, when it is necessary to adjust the current adjustment factor again... Before performing the amplification process, it is necessary to first check whether the cumulative number of amplification processes in the step size adjustment counter has exceeded the preset target count value (i.e., the adjustment upper limit), for example, 10 times. If not, the current adjustment factor will continue to be adjusted. Expanded processing.

[0074] S1508. Adjust the current adjustment factor according to the second step size adjustment coefficient. After performing the expansion process and accumulating the step size adjustment counter, return to execute S1501.

[0075] As an example rather than a limitation, the current adjustment factor can be used... Multiply by the second step size adjustment factor 2 to obtain the expanded version. Then, the step size adjustment counter is incremented by 1, and the process returns to execute S1501 to gradually calculate the new current gain rate. .

[0076] S1509, Clear the step size adjustment counter and reset the current momentum term. After setting it to 0, according to the formula: The current parameter adjustment vector under the current round k is calculated. This concludes the calculation process for this round, where I is the identity matrix.

[0077] In this embodiment, after detecting that the step size adjustment counter has accumulated to the target count value, a new formula is selected to be used: To adjust the step size optimization direction, the current parameter adjustment vector under the current round k is directly calculated. Among them, the formulas mentioned above use... This refers to the current adjustment factor obtained after the most recent amplification process when the step-size adjustment counter accumulates to the target count value. At this point, because a new algorithm is used to calculate the current parameter adjustment vector, the momentum term no longer continues to accumulate.

[0078] S160. Adjust the vector according to the current parameters, perform a round of parameter update on the fully connected neural network, and use the current adjustment factor and the current momentum term as the initial values ​​for the next round of model training.

[0079] It is understandable that the current parameter adjustment vector mentioned above is calculated in S1506. Or, calculated in S1509 .

[0080] Specifically, the current parameter adjustment vector represents the direction and magnitude of correction for all weights and bias parameters of the fully connected neural network in the current training round. Based on the current parameter adjustment vector, a round of overall iterative update can be performed on all trainable parameters in the fully connected network, completing parameter optimization for a single training round, gradually converging the model prediction error, and thus gradually improving the EGT time series prediction accuracy.

[0081] S170. Determine if the termination iteration condition is met: if yes, directly identify the currently trained fully connected neural network as the EGT prediction model for aero-engines; otherwise, return to execute S120.

[0082] The termination condition can be the maximum number of iterations or the loss convergence threshold, etc., and this embodiment does not impose any restrictions on it.

[0083] The technical solution of this invention involves collecting cruise message data from ACARS messages of an aero-engine during its service life, constructing a standard time-series training sample set, and inputting it into a pre-constructed fully connected neural network for model training. In each training round, the fully connected neural network is divided into a second number of single-node output sub-neural networks according to the number of output layer nodes. The gradient vectors of each sub-neural network are solved sequentially and then concatenated to form a complete Jacobian matrix. Based on the complete Jacobian matrix, the adjustment factor and momentum term passed from the previous round, a dual adaptive adjustment algorithm of gradient descent direction and convergence step size is used to calculate the current parameter adjustment vector and the current... This approach involves adjusting the adjustment factor and current momentum term; adjusting the vector based on the current parameters; updating the parameters of the fully connected neural network in one round; and using the current adjustment factor and current momentum term as the initial values ​​for the next round of model training until the final EGT prediction model for aero-engines is obtained. This technique significantly improves the convergence stability and global optimization capability of the aero-engine EGT prediction model, making the scheme more suitable for the engineering requirements of real-time on-wing prediction and online judgment of engine washing timing for aero-engines. It achieves full-process adaptation for the nonlinear and small-sample characteristics of aero-engine EGT time-series data, effectively realizing high-precision prediction of aero-engine EGT and providing reliable quantitative data support for engine washing timing judgment.

[0084] Based on the above embodiments, after concatenating the solved gradient vectors into a complete Jacobian matrix, the method may further include: Confidence assessment is performed on each gradient vector in the complete Jacobian matrix; if the confidence of a preset number of gradient vectors in the confidence assessment results is lower than a preset confidence threshold, the momentum term transmitted in the previous round is obtained; if the momentum term transmitted in the previous round is not 0, the momentum term transmitted in the previous round is reduced according to the target reduction factor.

[0085] In this embodiment, after the complete Jacobian matrix is ​​constructed, the gradient of each column (row) can be used as a gradient vector to further evaluate the effectiveness and reliability of each gradient vector and determine whether there are abnormal gradients after the current gradient update.

[0086] Specifically, considering that abrupt gradient changes, excessively large or small gradients, homogenization between gradients, and abnormal residuals all indicate the presence of abnormal gradients, a preset confidence score can be constructed to evaluate the confidence of each column (row) of gradients. Furthermore, if the confidence of multiple columns (e.g., 3 or 5 columns) or multiple rows of gradients is determined to be lower than the preset confidence threshold, it indicates that the overall gradient direction in this round is distorted, the learning information in this round is unreliable, and cannot be directly used to update weights.

[0087] Considering that the momentum term transmitted in the previous round is the accumulated inertia of multiple effective gradients from previous rounds, used to accelerate convergence and smooth gradient oscillations, when the gradient fails in the current round, the new gradient is no longer trusted, and the momentum inertia transmitted in the previous round needs to be corrected. That is, if the momentum term transmitted in the previous round is not zero, it is reduced according to the target reduction factor to eliminate the accumulated inertia.

[0088] With the above settings, low-quality gradients can be accurately identified, avoiding invalid parameter updates. In addition, it can prevent the superposition of historical momentum and erroneous gradients, completely solving the problems of model training jitter and loss bounce. Furthermore, the above improvements can be further adapted to the characteristics of nonlinearity, weak regularity and small fluctuations of aero-engine time series data, so that the DAF-LM algorithm can still maintain stable convergence in gradient anomaly rounds without diverging, which is particularly suitable for the on-wing water washing scenario of aero-engines.

[0089] Figure 2 This is a flowchart of a method for predicting the EGT of an aero-engine according to an embodiment of the present invention. This embodiment is applicable to situations where, based on the prediction input sequence of the target aero-engine, the EGT prediction model predicts multiple future EGT prediction values, and based on the above-mentioned EGT prediction values, the optimal water washing time of the target aero-engine is accurately predicted. This method can be executed by an aero-engine EGT prediction device, which can be implemented in hardware and / or software and is generally configured in an electronic device with data processing capabilities.

[0090] Correspondingly, such as Figure 2 As shown, the method includes: S210. Acquire cruise message data from the ACARS message of the target aero-engine to be predicted in real time, and construct a prediction input sequence based on the cruise message data.

[0091] The predicted input sequence includes a first number of target gas path parameter groups sorted by acquisition time.

[0092] In this embodiment, when constructing the prediction input sequence, it is also necessary to smooth and normalize the various data in the cruise message data of the target aero-engine during its service life in the ACARS messages to ensure the consistency of features between the prediction data and the training data.

[0093] S220. Input the first number of target air path parameter groups into the aero-engine EGT prediction model trained by the method as described in any one of the embodiments of the present invention.

[0094] S230. Obtain the target EGT prediction value at the second number of future times output by the EGT prediction model of the aero-engine.

[0095] In a specific example, assuming the first quantity is 9 and the second quantity is 3, the prediction input sequence includes the target gas path parameter sets acquired by the target aero-engine at the nine most recent acquisition times (the most recent acquisition time is T). Based on the above prediction input sequence, the aero-engine EGT prediction model outputs the target EGT prediction values ​​at three consecutive future times (T+10 minutes, T+20 minutes, and T+30 minutes).

[0096] S240. Based on the predicted EGT values ​​of each target and the factory-baseline EGT values ​​matched with the target aero-engine, calculate the EGT margin values ​​for a second number of future time periods.

[0097] Specifically, in obtaining a target EGT prediction value for a given future time T1... Then, combined with the factory-standard EGT value matched to the target aero-engine. It can be done according to the formula: The EGT margin value at the future time T1 is calculated. .

[0098] S250. If a preset number of consecutive EGT margin values ​​are all less than the preset water washing trigger threshold, then it is determined that the target aero-engine has reached the optimal water washing time.

[0099] In this embodiment, if the calculated consecutive preset number of EGT margin values ​​(e.g., all or more than half) are all less than the preset water washing trigger threshold (e.g., 0.5, 0.8 or 1), it is determined that the target aero-engine has reached the optimal water washing time, and then the water washing operation reminder and maintenance suggestions can be output to the aircraft maintenance system.

[0100] The technical solution of this invention adopts a multi-step EGT prediction combined with a multi-time margin joint judgment mechanism to avoid the misjudgment defects of traditional single-point prediction and single judgment, effectively improving the scientificity and accuracy of engine water washing timing determination. At the same time, it also realizes advanced prediction of aero-engine performance and intelligent operation and maintenance decision-making, replacing the traditional extensive mode of periodic water washing. It can accurately match the actual degradation state of the engine, extend the on-wing service life of the engine while reducing ineffective operation and maintenance, and adapt to the engineering requirements of refined in-service monitoring of aero-engines.

[0101] Specific application scenarios In this application scenario, a fully connected neural network based on the DAF-LM optimization algorithm is proposed for determining the timing of water washing for aero-engines. The complete implementation steps are as follows: (Specifically for the No. 1 turbofan engine of a domestic airliner.) Step 1: Data Collection and Preprocessing ACARS cruise message data for the engine from May 21, 2019 to April 4, 2024 were collected throughout its service life. Core parameters were extracted from the messages: low-pressure rotor speed, high-pressure rotor speed, fuel flow, total atmospheric temperature, flight altitude, lubricating oil pressure, lubricating oil temperature, Mach number, and exhaust temperature. After data cleaning to remove invalid messages and outliers, a total of 2,559 valid EGT data were obtained, which were used as the original EGT data.

[0102] The original EGT data was smoothed using an exponential smoothing algorithm, with a smoothing coefficient of... We set the value to 0.3 to eliminate sensor noise and random disturbances; then we use the minimum-maximum normalization method to map all smoothed parameter data to the (0,1) interval, completing the dimensionless processing and reducing the computational complexity of the model.

[0103] Specifically, in Figure 3 The image shows a spline curve of the raw exhaust temperature data for a section of a civil passenger aircraft engine. Figure 4 The image shows a graph of engine exhaust temperature data after smoothing and normalization.

[0104] Step 2: Constructing the Time Series Dataset A sliding window of length 9 was used to decompose the preprocessed time series data into fixed-length subsequences. Each subsequence contained 9 consecutive time series data as input features, and the next 5 EGT data of the subsequence were used as corresponding labels. Finally, 2359 sets of valid samples were constructed.

[0105] The training and test sets were divided in a 9:1 ratio, with 2123 samples used as the training set for training model parameters and 236 samples used as the test set for verifying the model's prediction accuracy and generalization performance.

[0106] Step 3: Construction of a fully connected neural network model based on the DAF-LM optimization algorithm 1. Network structure configuration: A four-layer fully connected neural network is constructed. The number of neurons in the input layer is 9, which is consistent with the length of the sliding window; the number of neurons in the first hidden layer is 18, the number of neurons in the second hidden layer is 6; the number of neurons in the output layer is 5, and the output is the 5-step EGT prediction value; the activation function is Sigmoid, the loss function is the root mean square error (RMSE), and the optimizer is the DAF-LM optimization algorithm proposed in this embodiment of the invention.

[0107] 2. Jacobian Matrix Calculation: The network is split into 5 single-output sub-neural networks with 5 nodes in the output layer. The weights and bias gradients of the output and non-output layers are calculated separately, and then concatenated to obtain the Jacobian matrix of the entire network. This significantly reduces the computational complexity of inverting high-dimensional matrices. Specifically, in Figure 5 The diagram illustrates a method for splitting the Jacobian matrix in a fully connected neural network, as described in an embodiment of the present invention.

[0108] 3. Configuration of core parameters for the DAF-LM optimization algorithm: Initial adjustment factor initial value of momentum term The first step length adjustment factor is 1 / 2, the second step length adjustment factor is 2, and the momentum coefficient... The step size adjustment counter is initialized to 0, and the maximum number of iterations is set. Iteration stopping error threshold .

[0109] 4. Iterative process configuration: The iterative process designed according to the technical solution of this invention completes the iterative logic configuration for model training, realizing dual adaptive adjustment of gradient descent direction and convergence step size, as well as automatic jumping out of local minima.

[0110] Step 4: Model Training and Performance Validation 1. Model Training: Input the training set into the fully connected neural network model based on the DAF-LM optimization algorithm, start iterative training, and adaptively adjust the adjustment factor according to the loss reduction gain rate during the iteration process. With the momentum term, training is terminated when the number of iterations reaches 100 or the loss is lower than the iteration stopping error threshold, and the model weight file is saved.

[0111] 2. Performance Validation: Input the test set into the trained model and output the EGT prediction results, using MAE, MSE, cosine similarity, etc. Four metrics were used to evaluate the model's performance. The model was also compared with traditional LM, RMSP, and AdaBound optimization algorithms, as well as mainstream time series prediction models such as DR-LM-NN, BP-Net, SVR, RFR, LSTM, TCN, TFT, and WaveNet. The experimental results are as follows: (1) In terms of convergence performance, such as Figure 6 As shown, the DAF-LM optimization algorithm proposed in this embodiment of the invention achieves loss convergence after 10 iterations, while the traditional LM algorithm requires 15 steps, the AdaBound algorithm requires 70 steps, and the RMSP algorithm still fails to converge after 100 iterations. Clearly, the DAF-LM algorithm's convergence speed is significantly better than other optimization algorithms.

[0112] (2) In terms of prediction accuracy, such as Figure 7 As shown, the prediction MAE of the fully connected neural network based on the DAF-LM optimization algorithm is 0.06, the MSE is 0.006, and the cosine similarity is 98.4%. With a score of 0.79, it outperforms other optimization algorithms in all metrics.

[0113] (3) Regarding the fitting effect, such as Figure 8 and Figure 9 As shown, the predicted curve EGT of the fully connected neural network based on the DAF-LM optimization algorithm is compared with the actual data curve. Figure 8 True-data in Figure 9 The predicted data in the model is almost perfectly matched, accurately capturing the long-term decline trend and short-term minor fluctuations of EGT data. It has no lag or over-smoothing issues, and is significantly better than other optimization algorithms and other mainstream prediction models, fully meeting the requirements of water washing timing judgment for EGT prediction accuracy.

[0114] (4) Regarding model solidification, it has been verified that the prediction accuracy of the fully connected neural network model based on the DAF-LM optimization algorithm meets the requirements of engineering applications. Solidifying the DAF-LM neural network model can be used for online EGT prediction and water washing timing judgment of this engine model.

[0115] Step 5: EGT Timing Prediction and Margin Calculation for Aero-engines The engine collects ACARS cruise message data in real time. After smoothing and normalization preprocessing, the data is input into a fully connected neural network model based on the DAF-LM optimization algorithm. The model outputs the EGT prediction values ​​for the next 5 flight cycles, as shown in Table 1. Table 1 The factory-standard EGT value for this engine model The EGT margin value for the corresponding flight cycle was calculated according to the formula, and the results are shown in Table 2: Table 2 Step Six: Intelligent Determination of Engine Water Washing Timing The judgment is based on the preset water washing trigger rules: the model predicts that the EGT margin value of the 3rd flight cycle is 0.8, which is close to the water washing trigger threshold of 0; the EGT margin values ​​of the 3rd, 4th and 5th consecutive flight cycles are all less than the preset water washing trigger threshold (for example, 1), which meets the continuous verification rule. Finally, it is determined that the engine has reached the optimal water washing time, and the water washing operation reminder is immediately output to the airline's maintenance system. At the same time, an engine performance degradation analysis report and maintenance recommendations are provided.

[0116] The proposed DAF-LM optimization algorithm improves the step size selection strategy of the LM algorithm by combining a dual adaptive adjustment mechanism for gradient descent direction and convergence step size with the steepest descent method. It also introduces a momentum-driven mechanism and an optimization direction adjustment mechanism, fundamentally solving the problems of traditional LM algorithms easily getting trapped in local minima and truncation errors leading to incorrect descent directions when processing EGT height nonlinear data. This significantly improves the stability of model convergence and global optimization capability. Experimental results show that the DAF-LM optimization algorithm can complete loss convergence in only about 10 steps, improving the convergence speed by more than 30% compared to the traditional LM algorithm, making it more suitable for engineering requirements such as real-time on-wing prediction of aero-engines and online judgment of water washing timing.

[0117] Furthermore, this embodiment of the invention utilizes a fully connected neural network model trained using the DAF-LM optimization algorithm. This model achieves full-process adaptation to the nonlinear and small-sample characteristics of aero-engine EGT time-series data, realizing high-precision EGT prediction. Experimental results show that compared to traditional LM neural networks, this model reduces MAE by 16.9% and MSE by 18.1%; compared to mainstream time-series prediction models such as LSTM and TCN, it reduces MAE by a maximum of 59.1% and MSE by a maximum of 81.7%, with a cosine similarity of 98.4% and a determination coefficient of... The value reached 0.79, providing reliable quantitative data support for determining the timing of engine water washing.

[0118] Furthermore, this invention deeply integrates high-precision EGT timing prediction technology with engine water washing maintenance scenarios. Based on the EGT margin prediction value, it sets clear water washing trigger thresholds and continuous verification judgment rules, replacing the traditional experience-based and fixed-cycle water washing timing judgment method. This enables accurate prediction of the optimal water washing time for the engine in advance, maximizing the performance recovery effect of water washing, avoiding irreversible degradation of engine performance, effectively reducing airline fuel consumption and maintenance costs, and extending engine on-wing service time.

[0119] Finally, the method of this invention is directly compatible with the measured data collected by the existing ACARS system of civil aircraft, without the need for additional sensors. The data preprocessing process is standardized and the model hyperparameter configuration is clear, which has strong engineering applicability. It can be directly integrated into the existing aero-engine health management system and has broad industry promotion value.

[0120] Figure 10 This is a schematic diagram of the structure of a training device for an aero-engine EGT prediction model provided in an embodiment of the present invention. Figure 10 As shown, the device includes: The training sample set construction module 1010 is used to collect cruise message data in ACARS messages of aero engines during their service life and construct a standard time-series training sample set. The standard time-series training samples include a first number of gas path parameter groups and a second number of standard EGT labels sorted by acquisition time, where the first number and the second number are both integers greater than 1. EGT prediction value acquisition module 1020 is used to sequentially acquire the current standard time series training sample in the standard time series training sample set, input it into the pre-built fully connected neural network for the current round of model training, and obtain a second number of EGT prediction values, wherein the number of output layer nodes of the fully connected neural network matches the second number. The network splitting module 1030 is used to split the fully connected neural network into a second number of sub-neural networks with single-node outputs according to the number of output layer nodes included in the fully connected neural network; The Jacobian matrix concatenation module 1040 is used to solve the gradient vector of the backpropagation of each sub-neural network in turn based on the second number of current standard EGT labels and the second number of current EGT prediction values ​​in the current standard time series training samples, and concatenate the solved gradient vectors into a complete Jacobian matrix. The current parameter calculation module 1050 is used to calculate the current parameter adjustment vector, current adjustment factor and current momentum term in the current round based on the complete Jacobian matrix, the adjustment factor and momentum term passed from the previous round, and the dual adaptive adjustment algorithm of gradient descent direction and convergence step size. The parameter update module 1060 is used to adjust the vector according to the current parameters, perform a round of parameter update on the fully connected neural network, and use the current adjustment factor and the current momentum term as the initial values ​​for the next round of model training. Return to execution module 1070, which is used to return the operation of sequentially obtaining the current standard time series training samples in the standard time series training sample set until the end of iteration condition is met, and the aero-engine EGT prediction model is obtained.

[0121] The technical solution of this invention involves collecting cruise message data from ACARS messages of an aero-engine during its service life, constructing a standard time-series training sample set, and inputting it into a pre-constructed fully connected neural network for model training. In each training round, the fully connected neural network is divided into a second number of single-node output sub-neural networks according to the number of output layer nodes. The gradient vectors of each sub-neural network are solved sequentially and then concatenated to form a complete Jacobian matrix. Based on the complete Jacobian matrix, the adjustment factor and momentum term passed from the previous round, a dual adaptive adjustment algorithm of gradient descent direction and convergence step size is used to calculate the current parameter adjustment vector and the current... This approach involves adjusting the adjustment factor and current momentum term; adjusting the vector based on the current parameters; updating the parameters of the fully connected neural network in one round; and using the current adjustment factor and current momentum term as the initial values ​​for the next round of model training until the final EGT prediction model for aero-engines is obtained. This technique significantly improves the convergence stability and global optimization capability of the aero-engine EGT prediction model, making the scheme more suitable for the engineering requirements of real-time on-wing prediction and online judgment of engine washing timing for aero-engines. It achieves full-process adaptation for the nonlinear and small-sample characteristics of aero-engine EGT time-series data, effectively realizing high-precision prediction of aero-engine EGT and providing reliable quantitative data support for engine washing timing judgment.

[0122] Based on the above embodiments, the training sample set construction module 1010 can be used for: Based on the cruise message data in the ACARS messages of the aero-engine during its service life, an EGT time series prediction dataset is constructed. Each EGT time series prediction dataset includes multiple feature value sets sorted by acquisition time. Each feature value set includes EGT data that matches the set acquisition time and a set of gas path parameters that are associated with the EGT data. In each EGT time series prediction data in the EGT time series prediction dataset, multiple data subsequences of target data length are continuously extracted according to the time series, where the target data length is the sum of the first and second counts; In each data subsequence, the gas path parameter set is retained in the first number of feature value sets, and the EGT data is retained as standard EGT labels in the second number of feature value sets, thus obtaining the standard time series training sample set.

[0123] Based on the above embodiments, the training sample set construction module 1010 can be further used for: Based on the cruise message data in the ACARS messages of the aero-engine during its service life, obtain the raw EGT time series prediction dataset; The exponential smoothing algorithm is used to smooth the feature values ​​in the original EGT time series prediction dataset, and the smoothed feature values ​​are then normalized to obtain the EGT time series prediction dataset.

[0124] Based on the above embodiments, the Jacobian matrix splicing module 1040 can be specifically used for: The target neural network is obtained sequentially from multiple sub-neural networks; Calculate the target residual value of the target neural network based on the current EGT prediction value that matches the target neural network and the current standard EGT label; Based on the target residual value, the partial derivatives of the target loss function used by the fully connected neural network with respect to all weight parameters and bias parameters of the target neural network are solved by the backpropagation chain derivative algorithm. All partial derivatives obtained are organized into a single-row gradient vector corresponding to the target neural network according to a preset parameter sorting rule. The process returns to the operation of the target neural network in sequence within multiple sub-neural networks until all sub-neural networks have been processed. Following the order of nodes in the output layer of the fully connected neural network, the single-row gradient vectors of all sub-neural networks are concatenated to obtain a complete Jacobian matrix covering all weights and bias parameters of the fully connected neural network.

[0125] Based on the above embodiments, the current parameter calculation module 1050 can be further used for: Get the current regulation factor Among them, the current round k Initialize it to the adjustment factor passed down from the previous round k-1; According to the formula The initial parameter adjustment vector for the current round k is calculated. ; in, For the residual vector, n is the second quantity. , For the j-th current EGT prediction value, For the j-th current standard EGT tag, J is the complete Jacobian matrix, and G is the approximate Hessian matrix. ; According to the formula: The current gain rate is calculated. ,in: This represents the parameter vector that has not yet been updated in the current round k. Let be the target loss function of the fully connected neural network. This is the second-order Taylor expansion of the target loss function; If the current gain rate If the value is greater than 0, then the current adjustment factor is adjusted according to the first step's long adjustment coefficient. The reduction process is performed to obtain the current adjustment factor for the current round. According to the formula: The current momentum term under the current round k is calculated. Afterwards, Determined as the current parameter adjustment vector in the current round k Clear the step size adjustment counter and end the calculation process for this round; among which, This refers to the momentum term transferred in the previous round k-1. The preset momentum coefficient; If the current gain rate If the value is less than or equal to 0, the current adjustment factor is adjusted according to the second step size adjustment coefficient. After performing amplification processing and accumulating the step size adjustment counter, the process returns to retrieve the current adjustment factor. Operation; When the step size adjustment counter accumulates to the target count value, if the current gain rate is determined... If it is still less than or equal to 0, clear the step size adjustment counter and reset the current momentum term. After setting it to 0, according to the formula: The current parameter adjustment vector under the current round k is calculated. This concludes the calculation process for this round, where I is the identity matrix.

[0126] Based on the above embodiments, it may further include: a momentum term advance adjustment module, used for: After concatenating the solved gradient vectors into a complete Jacobian matrix, the confidence level of each gradient vector in the complete Jacobian matrix is ​​evaluated. If it is determined that the confidence of a preset number of gradient vectors in the confidence assessment results is lower than the preset confidence threshold, then the momentum term passed in the previous round is obtained. If the momentum term transferred in the previous round is not 0, then the momentum term transferred in the previous round is reduced according to the target reduction factor.

[0127] The training device for the EGT prediction model of aero-engines provided in this embodiment of the invention can execute the training method for the EGT prediction model of aero-engines provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0128] Figure 11This is a schematic diagram of a predictive device for the EGT of an aero-engine, provided as an embodiment of the present invention. Figure 11 As shown, the device includes: The prediction input sequence construction module 1101 is used to acquire cruise message data in the ACARS message of the target aero-engine to be predicted in real time, and construct a prediction input sequence based on the cruise message data, wherein the prediction input sequence includes a first number of target gas path parameter groups sorted according to the acquisition time. The air path parameter group input module 1102 is used to input the first number of target air path parameter groups into the aero-engine EGT prediction model trained by the method described in any one of the embodiments of the present invention; The prediction value acquisition module 1103 is used to acquire the target EGT prediction value at a second number of future times output by the aero-engine EGT prediction model. EGT margin calculation module 1104 is used to calculate the EGT margin value for a second number of future times based on the target EGT prediction value and the factory baseline EGT value matched with the target aero-engine. The optimal water washing timing prediction module 1105 is used to determine that the target aero-engine has reached the optimal water washing timing if a preset number of consecutive preset EGT margin values ​​are all less than a preset water washing trigger threshold.

[0129] The technical solution of this invention adopts a multi-step EGT prediction combined with a multi-time margin joint judgment mechanism to avoid the misjudgment defects of traditional single-point prediction and single judgment, effectively improving the scientificity and accuracy of engine water washing timing determination. At the same time, it also realizes advanced prediction of aero-engine performance and intelligent operation and maintenance decision-making, replacing the traditional extensive mode of periodic water washing. It can accurately match the actual degradation state of the engine, extend the on-wing service life of the engine while reducing ineffective operation and maintenance, and adapt to the engineering requirements of refined in-service monitoring of aero-engines.

[0130] The aircraft engine EGT prediction device provided in this embodiment of the invention can execute the aircraft engine EGT prediction method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0131] The collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0132] Figure 12A schematic diagram of an electronic device 10, which can be used to implement embodiments of the present invention, is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0133] like Figure 12 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0134] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0135] Processor 11 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, central processing unit (CPU), graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as performing the training method for an aero-engine EGT prediction model as described in any embodiment of the present invention, or the prediction method for an aero-engine EGT as described in any embodiment of the present invention.

[0136] That is, a training method for an aero-engine exhaust temperature (EGT) prediction model, including: Collect cruise message data from the Aircraft Communication Addressing and Reporting System (ACARS) messages of aero engines during their service life, and construct a standard time-series training sample set; The standard time-series training samples include a first number of gas path parameter groups and a second number of standard EGT labels sorted by acquisition time, where the first number and the second number are both integers greater than 1. The current standard time-series training samples are sequentially obtained from the standard time-series training sample set and input into the pre-built fully connected neural network for the current round of model training to obtain a second number of EGT prediction values, wherein the number of output layer nodes of the fully connected neural network matches the second number. Based on the number of output layer nodes in the fully connected neural network, the fully connected neural network is divided into a second number of sub-neural networks with single-node outputs; Based on the second number of current standard EGT labels and the second number of current EGT predictions in the current standard time-series training samples, the gradient vector of backpropagation of each sub-neural network is solved in turn, and the solved gradient vectors are concatenated into a complete Jacobian matrix. Based on the complete Jacobian matrix, the adjustment factor and momentum term passed from the previous round, a dual adaptive adjustment algorithm of gradient descent direction and convergence step size is used to calculate the current parameter adjustment vector, the current adjustment factor and the current momentum term in the current round. Adjust the vector based on the current parameters, perform a round of parameter updates on the fully connected neural network, and use the current adjustment factor and the current momentum term as the initial values ​​for the next round of model training; Return to the operation of sequentially obtaining the current standard time series training samples in the standard time series training sample set until the end of the iteration condition is met, and obtain the EGT prediction model of the aero-engine.

[0137] Alternatively, a method for predicting the exhaust gas temperature (EGT) of an aircraft engine includes: The cruise message data in the Aircraft Communication Addressing and Reporting System (ACARS) messages of the target aero-engine to be predicted is acquired in real time, and a prediction input sequence is constructed based on the cruise message data, wherein the prediction input sequence includes a first number of target gas path parameter groups sorted according to the acquisition time. The first number of target air path parameter groups are input into the aero-engine EGT prediction model trained by the method of any one of claims 1-6; Obtain the target EGT prediction value at the second number of future time periods output by the aero-engine EGT prediction model; Based on the predicted EGT values ​​for each target and the factory-baseline EGT values ​​matched with the target aero-engine, calculate the EGT margin values ​​for a second number of future times. If a preset number of consecutive EGT margin values ​​are all less than the preset water washing trigger threshold, then the target aero-engine is determined to have reached the optimal water washing time.

[0138] In some embodiments, the training method for an aero-engine EGT prediction model as described in any one of the embodiments of the present invention, or the prediction method for an aero-engine EGT as described in any one of the embodiments of the present invention, may be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the training method for an aero-engine EGT prediction model as described above, or the prediction method for an aero-engine EGT as described in any one of the embodiments of the present invention, may be performed. Alternatively, in other embodiments, processor 11 may be configured by any other suitable means (e.g., by means of firmware) to perform the training method for an aero-engine EGT prediction model as described in any one of the embodiments of the present invention, or the prediction method for an aero-engine EGT as described in any one of the embodiments of the present invention.

[0139] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0140] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0141] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0142] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0143] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0144] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0145] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0146] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A training method for an aero-engine exhaust temperature (EGT) prediction model, characterized in that, include: Collect cruise message data from the Aircraft Communication Addressing and Reporting System (ACARS) messages of aero engines during their service life, and construct a standard time-series training sample set; The standard time-series training samples include a first number of gas path parameter groups and a second number of standard EGT labels sorted by acquisition time, where the first number and the second number are both integers greater than 1. The current standard time-series training samples are sequentially obtained from the standard time-series training sample set and input into the pre-built fully connected neural network for the current round of model training to obtain a second number of EGT prediction values, wherein the number of output layer nodes of the fully connected neural network matches the second number. Based on the number of output layer nodes in the fully connected neural network, the fully connected neural network is divided into a second number of sub-neural networks with single-node outputs; Based on the second number of current standard EGT labels and the second number of current EGT predictions in the current standard time-series training samples, the gradient vector of backpropagation of each sub-neural network is solved in turn, and the solved gradient vectors are concatenated into a complete Jacobian matrix. Based on the complete Jacobian matrix, the adjustment factor and momentum term passed from the previous round, a dual adaptive adjustment algorithm of gradient descent direction and convergence step size is used to calculate the current parameter adjustment vector, the current adjustment factor and the current momentum term in the current round. Adjust the vector based on the current parameters, perform a round of parameter updates on the fully connected neural network, and use the current adjustment factor and the current momentum term as the initial values ​​for the next round of model training; Return to the operation of sequentially obtaining the current standard time series training samples in the standard time series training sample set until the end of the iteration condition is met, and the aero-engine EGT prediction model is obtained.

2. The method according to claim 1, characterized in that, Cruise message data from ACARS messages of the aircraft engine during its service life were collected to construct a standard time-series training sample set, including: Based on the cruise message data in the ACARS messages of the aero-engine during its service life, an EGT time series prediction dataset is constructed. Each EGT time series prediction dataset includes multiple feature value sets sorted by acquisition time. Each feature value set includes EGT data that matches the set acquisition time and a set of gas path parameters that are associated with the EGT data. In each EGT time series prediction data in the EGT time series prediction dataset, multiple data subsequences of target data length are continuously extracted according to the time series, where the target data length is the sum of the first and second counts; In each data subsequence, the gas path parameter set is retained in the first number of feature value sets, and the EGT data is retained as standard EGT labels in the second number of feature value sets, thus obtaining the standard time series training sample set.

3. The method according to claim 2, characterized in that, Based on cruise message data from ACARS messages during the service life of the aero-engine, an EGT time-series prediction dataset is constructed, which further includes: Based on the cruise message data in the ACARS messages of the aero-engine during its service life, obtain the raw EGT time series prediction dataset; The exponential smoothing algorithm is used to smooth the feature values ​​in the original EGT time series prediction dataset, and the smoothed feature values ​​are then normalized to obtain the EGT time series prediction dataset.

4. The method according to claim 2, characterized in that, Based on the second number of current standard EGT labels and the second number of current EGT predictions from the current standard time-series training samples, the gradient vectors for backpropagation of each sub-neural network are solved sequentially, and the solved gradient vectors are concatenated into a complete Jacobian matrix, including: The target neural network is obtained sequentially from multiple sub-neural networks; Calculate the target residual value of the target neural network based on the current EGT prediction value that matches the target neural network and the current standard EGT label; Based on the target residual value, the partial derivatives of the target loss function used by the fully connected neural network with respect to all weight parameters and bias parameters of the target neural network are solved by the backpropagation chain derivative algorithm. All partial derivatives obtained are organized into a single-row gradient vector corresponding to the target neural network according to a preset parameter sorting rule. The process returns to the operation of the target neural network in sequence within multiple sub-neural networks until all sub-neural networks have been processed. Following the order of nodes in the output layer of the fully connected neural network, the single-row gradient vectors of all sub-neural networks are concatenated to obtain a complete Jacobian matrix covering all weights and bias parameters of the fully connected neural network.

5. The method according to any one of claims 1-4, characterized in that, Based on the complete Jacobian matrix, the adjustment factor and momentum term passed from the previous round, a dual adaptive adjustment algorithm using gradient descent direction and convergence step size is employed to calculate the current parameter adjustment vector, current adjustment factor, and current momentum term for the current round, including: Get the current regulation factor Among them, the current round k Initialize it to the adjustment factor passed down from the previous round k-1; According to the formula The initial parameter adjustment vector for the current round k is calculated. ; in, For the residual vector, n is the second quantity. , For the j-th current EGT prediction value, For the j-th current standard EGT tag, J is the complete Jacobian matrix, and G is the approximate Hessian matrix. ; According to the formula: The current gain rate is calculated. ,in: This represents the parameter vector that has not yet been updated in the current round k. Let be the target loss function of the fully connected neural network. This is the second-order Taylor expansion of the target loss function; If the current gain rate If the value is greater than 0, then the current adjustment factor is adjusted according to the first step's long adjustment coefficient. The reduction process is performed to obtain the current adjustment factor for the current round. According to the formula: The current momentum term under the current round k is calculated. Afterwards, Determined as the current parameter adjustment vector in the current round k Clear the step size adjustment counter and end the calculation process for this round; among which, This refers to the momentum term transferred in the previous round k-1. The preset momentum coefficient; If the current gain rate If the value is less than or equal to 0, the current adjustment factor is adjusted according to the second step size adjustment coefficient. After performing amplification processing and accumulating the step size adjustment counter, the process returns to retrieve the current adjustment factor. Operation; When the step size adjustment counter accumulates to the target count value, if the current gain rate is determined... If it is still less than or equal to 0, clear the step size adjustment counter and reset the current momentum term. After setting it to 0, according to the formula: The current parameter adjustment vector under the current round k is calculated. This concludes the calculation process for this round, where I is the identity matrix.

6. The method according to any one of claims 1-4, characterized in that, After concatenating the solved gradient vectors into a complete Jacobian matrix, the method further includes: Confidence assessment is performed on each gradient vector in the complete Jacobian matrix; If it is determined that the confidence of a preset number of gradient vectors in the confidence assessment results is lower than the preset confidence threshold, then the momentum term passed in the previous round is obtained. If the momentum term transferred in the previous round is not 0, then the momentum term transferred in the previous round is reduced according to the target reduction factor.

7. A method for predicting the exhaust gas temperature (EGT) of an aero-engine, characterized in that, include: The cruise message data in the Aircraft Communication Addressing and Reporting System (ACARS) messages of the target aero-engine to be predicted is acquired in real time, and a prediction input sequence is constructed based on the cruise message data, wherein the prediction input sequence includes a first number of target gas path parameter groups sorted according to the acquisition time. The first number of target air path parameter groups are input into the aero-engine EGT prediction model trained by the method of any one of claims 1-6; Obtain the target EGT prediction value at the second number of future time periods output by the aero-engine EGT prediction model; Based on the predicted EGT values ​​for each target and the factory-baseline EGT values ​​matched with the target aero-engine, calculate the EGT margin values ​​for a second number of future times. If a preset number of consecutive EGT margin values ​​are all less than the preset water washing trigger threshold, then the target aero-engine is determined to have reached the optimal water washing time.

8. An electronic device, characterized in that, The electronic device includes: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the training method of the aircraft engine EGT prediction model as described in any one of claims 1-6, or to perform the aircraft engine EGT prediction method as described in claim 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed by a processor, implement the training method for the aircraft engine EGT prediction model as described in any one of claims 1-6, or implement the aircraft engine EGT prediction method as described in claim 7.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the training method for the EGT prediction model of an aero-engine as described in any one of claims 1-6, or implements the prediction method for the EGT of an aero-engine as described in claim 7.

Images