A method for predicting the remaining life of an aero-engine by fusing physical information

Abstract

The application discloses a kind of aero-engine remaining life prediction methods of fusing physical information, belong to neural network remaining life prediction field.Step as follows: obtaining the sensor data of engine simulation operation to failure whole process and carrying out standardization and window pre-processing;Establish time-space dual-channel multilayer perception machine model, extract the time dimension and feature dimension information of signal, and with original data self-attention fusion, complete to the mapping of life prediction value, and calculate with real life error;Establish physical information neural network model, learn the hidden partial differential equation between predicted life value and time and hidden state, and carry out error calculation with the equation when in equilibrium state;With life error self-attention fusion as comprehensive loss to continuously train model to error minimum, complete final life prediction.The application fuses physical information on the basis of multilayer perception machine, simplifies the complexity of model, improves the prediction accuracy of engine remaining life.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of life prediction using neural network models, and more particularly, the present application provides an aero-engine residual life prediction method fusing physical information. BACKGROUND

[0002] With the continuous progress of science and manufacturing technology, modern industrial equipment is becoming increasingly complex, and once the equipment fails, it will cause great safety risks and property losses. Therefore, scientifically planning the maintenance plan of important equipment is crucial to ensure safe operation and improve cost-effectiveness. The traditional after-maintenance method has certain risks in terms of safety, and it is difficult to meet the maintenance needs of modern equipment. With the advent of the big data and information era, predictive and health management technology as an emerging system reliability guarantee method has received widespread attention in many fields. Among them, residual life prediction as a key technology in the research of predictive and health management is the core of promoting the transformation of equipment maintenance from prevention to prediction. It uses sensor monitoring information to reveal the degradation mechanism of the equipment and predicts the residual service life of the equipment, providing an important basis for intelligent maintenance decision-making.

[0003] In the field of aero-engines, residual life prediction is particularly important. As a highly complex and critical component of an aircraft, the engine will gradually degrade in performance over time, eventually leading to failure. Therefore, accurately predicting the residual life of the engine is crucial to ensure aviation safety.

[0004] However, with the rapid development of technology, the current prediction model generally has high model complexity, containing a large number of parameters and multi-level network architecture. This not only leads to the need for huge computing resources for model construction and training, but also makes it extremely difficult to maintain, optimize and adjust the model. Secondly, due to the diversity and complexity of the working conditions in actual application scenarios, the existing model has low prediction accuracy for complex working conditions. Finally, the internal decision-making mechanism of the existing model is like a "black box". Even if the model can give a prediction result, it cannot provide a clear and intuitive explanation of how the result is obtained and which key factors and data features are used. This limits the widespread application and in-depth promotion of the model in these key fields. SUMMARY

[0005] The purpose of the present application is to solve the above problems, and a residual life prediction method for an aero-engine fusing physical information is provided. To achieve the above purpose, the technical solution adopted by the present application is as follows:

[0006] The application discloses a method for predicting the residual life of an aero-engine by fusing physical information.

[0007] The method for predicting the residual life of an aero-engine by fusing physical information comprises the following steps of realizing the residual life prediction function.

[0008] Step 1: acquiring signals generated in the whole process of simulating the running of an aero-engine to failure, wherein the signals contain various types of sensor data;

[0009] Step 2: performing visual analysis on all the sensor data, performing standardization on the sensor data collected in a single working condition, and performing category standardization on the sensor data collected in multiple working conditions;

[0010] Step 3: performing windowed preprocessing on the sensor data by using a sliding window algorithm;

[0011] Step 4: establishing a time-space dual-channel multilayer perception machine model, extracting signal time and space dimension information, and performing self-attention feature fusion on the time dimension information, feature dimension information and original data information to complete mapping to a life prediction value;

[0012] Step 5: establishing a physical information neural network model, learning a hidden partial differential equation between a predicted life value, time and a hidden state, and performing error calculation on the equation and an equation in a balanced state;

[0013] Step 6: calculating the error between a predicted life and an actual life, performing self-attention fusion on the error and a partial differential equation error into a comprehensive loss function of an updated model, and constantly updating a training model by using the comprehensive loss of batch data so as to minimize the error and complete final life prediction.

[0014] In step 2, when performing category standardization on the sensor data collected in multiple working conditions, first, each sensor data is classified according to a working condition category, and the data in each category is independently standardized by the following calculation principle.

[0015]

[0016] In the formula, x c,i is the data in each category, n c is the number of data in each category, and μc and σ c are the feature mean and variance of the samples in each class respectively, is the data after class normalization.

[0017] In step 3, when the sliding window algorithm is used for windowing preprocessing of sensor data, the window width of the sliding window algorithm is set to 30 data sampling points, and the window sliding step is set to 1 data sampling point.

[0018] When windowing processing is performed, the initial window position is the starting point of the training data sequence, and the window is moved in turn according to the set sliding step until all ranges of the sensor data sequence are covered.

[0019] By training data to make life labels, the life label data is also windowed, and the last life value of each window is selected as the life of the training sample of the window.

[0020] The training data and life value in each window are a data sample group, and these sample groups will be used for subsequent analysis and processing of engine-related characteristics.

[0021] In step 4, a time-space dual-channel multi-layer perception model is established to extract signal time and space dimension information, and to perform self-attention feature fusion on time dimension information, feature dimension information and original data information to complete the mapping to the life prediction value. Mainly through:

[0022] A T-MLP model for time dimension is built based on a fully connected layer with a ReLU activation function;

[0023] The original data x i is divided into multiple batches, and each batch of data is standardized by the following formula:

[0024]

[0025] where i is the number of data columns, μi is the feature mean of the samples, is the feature variance of the samples, and ∈ is a very small constant to prevent the denominator from being 0. is the standardized data.

[0026] The original data time and feature dimension are converted to match the shape of the model, and input into the model with an activation function to extract time information;

[0027]

[0028] where k is the feature dimension, j is the time, ReLU is the activation function, and z ikhis the intermediate hidden layer state after a linear transformation with an activation function. W ijh and b 1j are trainable parameters to help the model fit the data better;

[0029] Dropout is applied to the hidden layer state, which means randomly dropping neurons with a certain probability to prevent the model from over-reliance on a specific neuron;

[0030]

[0031] where p is the probability of randomly dropping neurons.

[0032] The time feature data after Dropout is projected into the original feature space, while the dimension is converted and connected with the original information by residual connection;

[0033]

[0034]

[0035] where T represents the dimension conversion, W ihj and b 2j are trainable parameters, x’ ijk is the time feature data after full connection and transposition, is the original data, is the final time feature data after residual connection, i.e. the data output after T-MLP model.

[0036] For spatial dimension information, i.e. the feature dimension information of the sensor, a F-MLP model for spatial dimension is also built based on the full connection layer with ReLU activation function. The basic calculation process of T-MLP model is the same, except that the spatial dimension k is transformed;

[0037]

[0038]

[0039]

[0040]

[0041] where x” ijk is the output feature information after 3 layers of full connection, represents the spatial feature information output by the final F-MLP model after residual connection with the original time feature information.

[0042] SoftMax is used to build a feature fusion layer to calculate the original data time dimension information and feature dimension information The importance ratio is taken as the respective weight for feature fusion.

[0043] Firstly, the is concatenated into a vector X, which is mapped to an intermediate variable through a fully connected layer;

[0044] Z=XW+b

[0045] Wherein, the dimension of Z is n x 3, for each sample (each row in the batch), Z contains three elements, which correspond to the scores of the original data, time dimension information and feature dimension information after mapping. The weight vector obtained after SofttMax.

[0046]

[0047] w i =[w1, w2, w3] are the weight of the original data, the weight of the time dimension information and the weight of the feature dimension information, and satisfy w1+w2+w3=1, these weights represent the importance ratio of each in feature fusion.

[0048] The original data time dimension information and feature dimension information are weighted and summed to form a new feature vector h;

[0049]

[0050] The new feature vector h becomes a two-dimensional tensor after Flatten operation, which is input into a fully connected layer with an output dimension of 1;

[0051] r=h flat *W fc +b fc , W fc ∈R d×1 , b fc ∈R 1

[0052] Wherein, W fc and b fc are learnable parameters, d represents the length of the flattened features, and r is the predicted remaining useful life of the model.

[0053] In step 5, the physical information neural network model is established, the hidden partial differential equation between the predicted life value and time and hidden state is learned, and the equation is calculated with the equation at the equilibrium state. Mainly through:

[0054] solving the first derivative of the life prediction value r with respect to time t, and the first, second and higher order partial derivatives of the life prediction value r with respect to the hidden state h of the fusion data;

[0055] a physical information neural network learning a mathematical relationship model corresponding to an actual physical process;

[0056]

[0057] wherein a and β are learnable parameters, and the equation F(r, t, h) is made as close to 0 as possible through training the neural network to obtain a stable state that satisfies the physical law;

[0058] The root mean square loss of the equation F learned by each batch of samples and the equation F(0) is calculated as one of the model loss functions:

[0059] loss 1i=RMSE(F,F(0))

[0060] In step 5, the error between the predicted life and the true life is calculated, and is fused with the partial differential equation error to obtain an integrated loss function for updating the model, and the integrated loss of the batch data is used to continuously update the training model to minimize the error and complete the final life prediction. Mainly through:

[0061] The root mean square error of the life prediction value r obtained in step 4 and the true life value ture_r is calculated as another one of the model loss functions;

[0062] loss2 i =RMSE(r, ture_r)

[0063] The life loss is fused with the partial differential equation loss loss1 i obtained in step 5 through SoftMax to construct an integrated loss function loss;

[0064]

[0065] loss=α i *loss1 i +β i *loss2 i

[0066] Based on the integrated loss loss obtained from each batch of samples, the back propagation algorithm is performed, and the model parameters are updated and adjusted according to the gradient descent algorithm;

[0067] This process is repeated until the loss value of the model reaches a preset minimum threshold or satisfies the convergence condition, thereby completing the final life prediction model training, so that the model has high-precision life prediction capability. Attached Figure Description

[0068] Figure 1 This is a flowchart illustrating the present invention;

[0069] Figure 2 This is a data visualization diagram of multi-condition data before and after standardization in an embodiment of the present invention;

[0070] Figure 3 The sliding window algorithm of this invention performs windowing preprocessing on the data;

[0071] Figure 4 This is a prediction effect diagram of an embodiment of the present invention. Detailed Implementation

[0072] This invention proposes a method for predicting the remaining life of aero-engines by integrating physical information. The invention will be described in detail below with reference to the accompanying drawings and text. It is important to note that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0073] like Figure 1 As shown, a method for predicting the remaining life of an aero-engine by fusing physical information includes the following steps:

[0074] Step 1: Acquire the signals generated during the entire simulated operation of the aero-engine from start to finish, including data from various types of sensors. Divide the dataset into a training subset and a test subset. The training subset records the sampled values ​​of multiple sensors throughout the complete cycle of the aero-engine from normal operation to failure, while the test subset contains the sensor sampled values ​​and corresponding remaining lifespan up to a certain point in time before the failure.

[0075] Step 2: Perform visual analysis on all sensor data, standardize sensor data collected under single operating conditions, and standardize the categories of sensor data collected under multiple operating conditions. Figure 2 This is a data visualization diagram of multi-condition data before and after category standardization.

[0076] Step 3: Perform windowing preprocessing on the sensor data using the sliding window algorithm;

[0077] The sliding window algorithm is configured with a window width of 30 data sampling points and a window sliding step of 1 data sampling point. Both the data and lifetime labels are processed using windowing. Figure 3 As shown.

[0078] Step 4: Establish a time-space dual-channel multilayer perceptron model, extract the time and space dimension information of the signal, perform self-attention feature fusion on the time dimension information, feature dimension information and raw data information, and complete the mapping to lifetime prediction value.

[0079] Based on the full connection layer with ReLU activation function, a T-MLP model for time dimension is built to extract time feature data Time feature data The calculation formula is:

[0080]

[0081]

[0082]

[0083] Wherein, k is the feature dimension, j is the time, ReLU is the activation function, W and b are trainable parameters, and T represents the dimension conversion. The time dimension information

[0084] For the spatial dimension information, a F-MLP model for spatial dimension is built based on the full connection layer with ReLU activation function, the time dimension information is transformed for the spatial dimension, and the feature dimension information is learned through the F-MLP model The calculation formula is:

[0085]

[0086]

[0087]

[0088]

[0089] Wherein, x" ijk is the output feature information after 3 layers of full connection, represents the spatial feature information output by the final F-MLP model after residual connection with the original time feature information.

[0090] A feature fusion layer is built by using SoftMax, the importance ratio of the original data Time dimension information And the feature dimension information Is calculated respectively, and is used as the respective weight for feature fusion, which is fused into a new feature vector h.

[0091]

[0092] Wherein, w1, w2, w3 are the weight of the original data, the weight of the time dimension information and the weight of the feature dimension information respectively. h is input into a full connection layer with an output dimension of 1, and the predicted remaining useful life value r is obtained.

[0093]

[0094] where W fc and b fc are learnable parameters, d denotes the length of flattened features, and r is the remaining useful life predicted by the model.

[0095] Step 5, establish a physical information neural network model, learn the hidden partial differential equation between the predicted life value and time and hidden state, and calculate the error between the equation and the equation at the equilibrium state;

[0096] Solve the first derivative of the life prediction value r with respect to time t, and the first, second, and high-order partial derivatives of the life prediction value r with respect to the hidden state h of the fused data;

[0097] Physical information neural network, learn the mathematical relationship model corresponding to the actual physical process;

[0098]

[0099]

[0100] Calculate the root mean square loss of the equation F learned by each batch sample and the equation F(0) as one of the model loss functions.

[0101] loss 1i=RMSE(F,F(0))

[0102] Step 6, calculate the error between the predicted life and the true life, and fuse it with the partial differential equation error to form an integrated loss function for updating the model, and constantly update the training model using the integrated loss of the batch data to minimize the error and complete the final life prediction.

[0103] Calculate the root mean square error of the life prediction value r obtained in step 4 and the true value ture_r as another one of the model loss functions.

[0104] loss2 i =RMSE(r, ture_r)

[0105] Fuse this life loss with the partial differential equation loss loss1 i obtained in step 5 through SoftMax to construct an integrated loss function loss.

[0106]

[0107]

[0108] loss=α i *loss1i + β i *loss2 i

[0109] The reverse propagation algorithm is performed based on the comprehensive loss loss obtained for each batch of samples, and the model parameters are updated and adjusted according to the gradient descent algorithm;

[0110] This process is repeatedly performed until the loss value of the model reaches a preset minimum threshold or satisfies a convergence condition, thereby completing the final life prediction model training, so that the model has high-precision life prediction capability. The prediction result is as shown in Figure 4 .

Claims

1. A method for predicting the remaining life of an aeroengine by fusing physical information, characterized in that, The method comprises the following steps: Step 1: obtaining signals generated in the whole process of simulating operation to failure of an aero turbofan engine, wherein the signals contain various types of sensor data; Step 2: performing visual analysis on all sensor data, standardizing sensor data collected in a single working condition, and performing category standardization on sensor data collected in multiple working conditions; Step 3: performing window preprocessing on the sensor data by using a sliding window algorithm; Step 4: establishing a time-space dual-channel multi-layer perception model, extracting time and space dimension information of the signals, performing self-attention feature fusion on the time dimension information, feature dimension information and original data information, and completing mapping to a life prediction value; Step 5: establishing a physical information neural network model, learning a hidden partial differential equation between the predicted life value and time and a hidden state, and performing error calculation on the equation and an equation in a balanced state; Step 6: calculating an error between the predicted life and an actual life, and taking the error and a partial differential equation error as a loss function for updating the model, and constantly updating a training model by using the two losses of the batch data, so as to minimize the error and complete final life prediction; Step 5 comprises: (1) solving a first-order derivative of the life prediction value r with respect to time t, and a first-order and high-order partial derivative of the life prediction value with respect to a hidden state h of the fused data; (2) establishing a physical information neural network, learning a mathematical relationship model corresponding to an actual physical process; (3) making the equation F as close to 0 as possible by training the neural network, so as to obtain a stable state satisfying a physical law; (4) taking a root mean square loss of the equation F learned by each batch sample and an equation F(0) as one of the model loss functions; loss1 i = RMSE(F, F(0)); Step 6 comprises: (1) taking a root mean square error of the life prediction value r and a life true value ture_r as another of the model loss functions; loss2 i = RMSE(r,ture_r) (2) fusing the life loss and a partial differential equation loss loss1i obtained in step 5 by using SoftMax, and constructing a comprehensive loss function loss; loss = a i • loss1 i + b i • loss2 i (3) performing a back propagation algorithm based on the comprehensive loss loss obtained by each batch sample, and updating and adjusting model parameters according to a gradient descent algorithm, and constantly repeating the process until the loss value of the model reaches a preset minimum threshold or satisfies a convergence condition, so as to complete training of the final life prediction model, and make the model have high-precision life prediction capability; Step 4 comprises: (1) using a multi-layer perception network framework to build a T-MLP model for the time dimension, extracting time dimension information by inputting original sample data into the T-MLP model, and mapping the time dimension information to a shape same as the original data; (2) using an MLP network framework to build an F-MLP model for the feature dimension, inputting the time dimension information output by the T-MLP model into the F-MLP model to extract feature dimension information, and mapping the feature dimension information to a shape same as the original data; (3) Using SoftMax to build a feature fusion layer, respectively calculating the importance proportion of the original data , time dimension information and feature dimension information , and taking them as respective weights for feature fusion, which is fused into a new feature vector h; wherein, are the weight of the original data, the weight of the time dimension information and the weight of the feature dimension information, respectively; (4) inputting h into a fully connected layer with an output dimension of 1 to obtain a predicted remaining useful life value r, wherein, and are learnable parameters, d denotes the length of flattened features, and r is the remaining useful life predicted by the model.

2. The method of claim 1, wherein, Step 3 comprises: (1) The window width of the sliding window algorithm is set to 30 data sampling points, and the window sliding step is set to 1 data sampling point; (2) When performing window processing, the initial window position is the starting point of the sensor data sequence, and the window is moved according to the set sliding step until it covers all ranges of the sensor data sequence; (3) The sensor data in each window is taken as a data sample, and these samples will be used for subsequent analysis and processing of engine-related characteristics.

Images