Aero-engine residual life prediction method fusing physical information

Abstract

The invention discloses a physical information fused aero-engine residual life prediction method, and belongs to the field of neural network residual life prediction. The method comprises the following steps: acquiring sensor data of the whole process from engine simulation operation to fault, and performing standardization and windowing preprocessing; establishing a time-space dual-channel multi-layer perceptron model, extracting time dimension and feature dimension information of the signal, performing self-attention fusion with original data, completing mapping to a life prediction value, and calculating an error with a real life; establishing a physical information neural network model, learning a hidden partial differential equation among a predicted life value, time and a hidden state, and performing error calculation with an equation in an equilibrium state; and performing self-attention fusion with the life error to obtain comprehensive loss so as to continuously train the model until the error is minimum, and finishing final life prediction. Physical information is fused on the basis of the multi-layer perceptron, the complexity of the model is simplified, and the prediction precision of the residual life of the engine is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of life prediction using a neural network model, and more specifically, the present invention proposes a method for predicting the remaining life of an aircraft engine by integrating physical information. Background Art

[0002] With the continuous advancement of science and manufacturing technology, modern industrial equipment is becoming more and more complex. Once the equipment fails, it will cause huge safety risks and property losses. Therefore, scientific planning of maintenance plans for important equipment is crucial to ensure operational safety and improve cost-effectiveness. Traditional post-maintenance methods have certain risks in terms of safety and are difficult to meet the maintenance needs of modern equipment. With the advent of big data and the information age, prediction and health management technology, as an emerging system reliability assurance method, has received widespread attention in many fields. Among them, remaining life prediction, as a key technology in prediction and health management research, is the core of promoting the transformation of equipment maintenance from preventive to predictive. It uses sensor monitoring information to reveal the equipment degradation mechanism and predict the remaining service life of the equipment, providing an important basis for intelligent maintenance decisions.

[0003] In the field of aircraft engines, the remaining life prediction is particularly important. As a highly complex and critical component of an aircraft, the engine's performance will gradually degrade over a long period of operation, eventually leading to failure. Therefore, accurately predicting the remaining life of the engine is crucial to ensuring aviation safety.

[0004] However, with the rapid development of technology, current prediction models generally have high model complexity, including a large number of parameters and multi-level network architecture. This not only requires huge computing resources to build and train the model, but also makes the maintenance, optimization and adjustment of the model extremely difficult. Secondly, due to the diversity and complexity of working conditions in actual application scenarios, the existing models have low predictions 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 based on which key factors and data features. This limits the wide application and in-depth promotion of the model in these key areas. Summary of the invention

[0005] The purpose of the present invention is to solve the above problems and provide a method for predicting the remaining life of an aircraft engine by integrating physical information. To achieve the above purpose, the technical solution adopted by the present invention is:

[0006] A method for predicting the remaining life of an aircraft engine by integrating physical information, firstly obtains the signal generated by the entire process of simulating the operation of the aircraft engine to failure, performs different types of standardization methods for single working conditions and multiple working conditions, and performs window preprocessing; establishes a dual-channel multi-layer perceptron network model and extracts time and space information, and completes the mapping of sensor signals to life prediction values; at the same time, based on the physical information neural network, extracts the physical relationship between the predicted life value and time and the hidden state, uses the life error and physical deviation to reversely optimize the model, and continuously trains the model to complete the life prediction.

[0007] A method for predicting the remaining life of an aircraft engine by integrating physical information, realizing the remaining life prediction function comprises the following steps:

[0008] Step 1: Acquire the signals generated by the entire process of the simulated operation of the aviation turbofan engine until failure, which includes multiple types of sensor data;

[0009] Step 2: Visualize and analyze all sensor data, standardize sensor data collected under a single working condition, and standardize sensor data collected under multiple working conditions;

[0010] Step 3: Use the sliding window algorithm to preprocess the sensor data;

[0011] Step 4: Establish a time-space dual-channel multi-layer perceptron model to extract the signal time and space dimension information, and perform self-attention feature fusion on the time dimension information, feature dimension information and original data information to complete the mapping to the life prediction value;

[0012] Step 5: Establish a physical information neural network model, learn the hidden partial differential equation between the predicted life value and time and the hidden state, and calculate the error between the equation and the equation at the equilibrium state;

[0013] Step 6: Calculate the error between the predicted lifespan and the actual lifespan, and fuse it with the partial differential equation error by self-attention to form a comprehensive loss function for updating the model. Use the comprehensive loss of batch data to continuously update the training model to minimize the error and complete the final lifespan prediction.

[0014] In the step 2, when the sensor data collected under multiple working conditions are standardized by category, each sensor data is first classified according to the 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 For each category of data, n c is the number of data in each category, μc and σ c are the feature mean and variance of samples in each category, The data are classified after normalization.

[0017] In step 3, when the sliding window algorithm is used to perform window preprocessing on the 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 performing window processing, the starting point of the training data sequence is used as the initial window position, and the window is moved in sequence according to the set sliding step size until the entire range of the sensor data sequence is covered;

[0019] The lifespan labels are generated through the training data, and the lifespan label data is also processed by windowing, and the last lifespan value of each window is selected as the lifespan of the training sample in the window;

[0020] The training data and life values ​​in each window are taken as a data sample group, which will be used for subsequent analysis and processing of engine-related characteristics.

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

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

[0023] The original data x i Divide into multiple batches, and standardize each batch data using the following formula;

[0024]

[0025] Among them, i is the number of data columns, μi is the characteristic mean of the sample, is the characteristic variance of the sample, and ∈ is a small constant used to prevent the denominator from being zero. is the standardized data.

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

[0027]

[0028] Among them, k is the feature dimension, j is time, ReLU is the activation function, and z ikhis the state of the intermediate hidden layer after a linear change with an activation function. ijh and b 1j It is a trainable parameter that helps the model better fit the data;

[0029] Perform dropout processing on the hidden layer state, that is, randomly discard neurons with a certain probability to prevent the model from over-relying on a specific neuron;

[0030]

[0031] Among them, p is the probability of randomly discarding neurons.

[0032] Project the time feature data after Dropout to the original feature space, and convert the dimension to perform residual connection with the original information;

[0033]

[0034]

[0035] Among them, T represents dimension transformation, W ihj and b 2j is a trainable parameter, x' ijk is the time feature data after full connection and transposition, For the original data, It is the final time feature data after the residual connection, that is, the data output after passing through the T-MLP model.

[0036] For spatial dimension information, that is, the characteristic dimension information of the sensor, the fully connected layer with the ReLU activation function is also used as the basis to build the F-MLP model for the spatial dimension. The basic calculation process is the same as the T-MLP model, except that the spatial dimension k is transformed;

[0037]

[0038]

[0039]

[0040]

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

[0042] Use SoftMax to build a feature fusion layer and calculate the original data separately Time dimension information and feature dimension information The importance ratios of the features are used as their respective weights for feature fusion.

[0043] First, Concatenate into a vector X, and map X to an intermediate variable through a fully connected layer;

[0044] Z=XW+b

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

[0046]

[0047] w i =[w1, w2, w3] are the weights of the original data, the time dimension information, and the feature dimension information, respectively, and satisfy w1+w2+w3=1. These weights represent their respective importance in feature fusion.

[0048] The original data Time dimension information and feature dimension information Perform weighted summation and merge into a new feature vector h;

[0049]

[0050] The new feature vector h is transformed into a two-dimensional tensor after the 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] Among them, W fc and b fc is a learnable parameter, d represents the length of the flattened feature, and r is the remaining service life predicted by the model.

[0053] In step 5, a physical information neural network model is established to learn the hidden partial differential equation between the predicted life value, time and hidden state, and to calculate the error between the equation and the equation in the equilibrium state. Mainly through:

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

[0055] Physical information neural network, learning mathematical relationship models corresponding to actual physical processes;

[0056]

[0057] Among them, α and β are learnable parameters. By training the neural network, the equation F(r, t, h) is made as close to 0 as possible to obtain a stable state that satisfies the laws of physics.

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

[0059] loss1i=RMSE(F,F(0))

[0060] In step 5, the error between the predicted life and the actual life is calculated and fused with the partial differential equation error to form a comprehensive loss function for updating the model. The training model is continuously updated using the comprehensive loss of batch data to minimize the error and complete the final life prediction. This is mainly done by:

[0061] Calculate the root mean square error between the predicted life value r obtained in step 4 and the true life value turer as another one in the model loss function;

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

[0063] The lifetime loss is compared with the partial differential equation loss loss1 obtained in step 5 i Through SoftMax fusion, a comprehensive loss function loss is constructed;

[0064]

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

[0066] The back propagation algorithm is performed based on the comprehensive loss obtained from each batch of samples, 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 the preset minimum threshold or meets the convergence condition, thereby completing the final life prediction model training and enabling the model to have high-precision life prediction capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is a schematic diagram of the process of the present invention;

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

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

[0071] Figure 4 It is a diagram of the predicted effect of the implementation mode of the present invention. DETAILED DESCRIPTION

[0072] The present invention proposes a method for predicting the remaining life of an aircraft engine by integrating physical information. The present invention is explained in detail below with reference to the accompanying drawings and text. It is worth noting that the specific implementation described here is only used to explain the present invention and is not used to limit the present invention.

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

[0074] Step 1: Obtain the signals generated by the entire process of simulated operation of the aircraft turbofan engine to failure, which contains multiple types of sensor data. The data set is divided into a training subset and a test subset. The training subset records the sampling values ​​of multiple sensors in the complete cycle of the aircraft engine from normal operation to failure, and the test subset contains the sensor sampling values ​​and corresponding remaining life before a certain time point before the failure;

[0075] Step 2: Visualize and analyze all sensor data, standardize the sensor data collected under a single working condition, and standardize the categories of sensor data collected under multiple working conditions. Figure 2 Data visualization diagram of multi-operating condition data before and after category standardization.

[0076] Step 3: Use the sliding window algorithm to preprocess the sensor data;

[0077] The window width of the sliding window algorithm is set to 30 data sampling points, the window sliding step is set to 1 data sampling point, and both the data and the lifespan label are processed by windowing, such as Figure 3 shown.

[0078] Step 4: Establish a time-space dual-channel multi-layer perceptron model to extract the signal time and space dimension information, perform self-attention feature fusion on the time dimension information, feature dimension information and original data information, and complete the mapping to the life prediction value;

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

[0080]

[0081]

[0082]

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

[0084] For spatial dimension information, we also use the fully connected layer with ReLU activation function as the basis to build an F-MLP model for the spatial dimension, transform the time dimension information into the spatial dimension, and learn the feature dimension information through the F-MLP model. The calculation formula is:

[0085]

[0086]

[0087]

[0088]

[0089] Among them, 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 temporal feature information.

[0090] Use SoftMax to build a feature fusion layer and calculate the original data separately Time dimension information and feature dimension information The importance ratios of the features are used as their respective weights for feature fusion, and are fused into a new feature vector h.

[0091]

[0092] Among them, w1, w2, and w3 are the weights of the original data, the weights of the time dimension information, and the weights of the feature dimension information, respectively. h is input into a fully connected layer with an output dimension of 1 to obtain the predicted remaining useful life value r.

[0093]

[0094] Among them, W fc and b fc is a learnable parameter, d represents the length of the flattened feature, and r is the remaining service 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 the hidden state, and calculate the error between the equation and the equation in the equilibrium state;

[0096] Solve the first-order derivative of the life prediction value r with respect to time t, as well as the first-order, second-order and higher-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, learning mathematical relationship models corresponding to actual physical processes;

[0098]

[0099]

[0100] The root mean square loss between equation F learned for each batch of samples and equation F(0) is calculated as one of the model loss functions.

[0101] loss1i=RMSE(F,F(0))

[0102] Step 6: Calculate the error between the predicted lifespan and the actual lifespan, and fuse it with the partial differential equation error to form a comprehensive loss function for updating the model. Use the comprehensive loss of batch data to continuously update the training model to minimize the error and complete the final lifespan prediction.

[0103] The root mean square error between the predicted life value r obtained in step 4 and the true life value ture_r is calculated as another one in the model loss function.

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

[0105] The lifetime loss is compared with the partial differential equation loss loss1 obtained in step 5 i Through SoftMax fusion, a comprehensive loss function loss is constructed;

[0106]

[0107]

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

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

[0110] This process is repeated until the loss value of the model reaches the preset minimum threshold or meets the convergence condition, thus completing the final life prediction model training and enabling the model to have high-precision life prediction capabilities. Figure 4 shown.

Claims

1. A method for predicting the remaining life of an aircraft engine by integrating physical information, characterized in that: Based on various sensor signals in aircraft engine operation experiments, a classification and standardization method is adopted for multi-operating condition data to increase the degradation trend of data, and a sliding window algorithm is used to segment continuous signals into representative time segments to enhance the characteristic information of the signal time dimension; the segmented signals are input into a dual-channel multi-layer perceptron network model in batches to extract time and space information, and complete the mapping of sensor signals to life prediction values; at the same time, based on the physical information neural network, the physical relationship between the predicted life value and time and the hidden state is extracted, and the life error and physical deviation are used to reversely optimize the model, and the training model is continuously optimized to complete the life prediction.

2. A method for predicting the remaining life of an aircraft engine by integrating physical information, characterized in that: The following steps are involved: Step 1: Acquire the signals generated by the entire process of the simulated operation of the aviation turbofan engine until failure, which includes multiple types of sensor data; Step 2: Visualize and analyze all sensor data, standardize sensor data collected under a single working condition, and standardize sensor data collected under multiple working conditions; Step 3: Use the sliding window algorithm to preprocess the sensor data; Step 4: Establish a time-space dual-channel multi-layer perceptron model to extract the signal time and space dimension information, perform self-attention feature fusion on the time dimension information, feature dimension information and original data information, and complete the mapping to the life prediction value; Step 5: Establish a physical information neural network model, learn the hidden partial differential equation between the predicted life value and time and the hidden state, and calculate the error between the equation and the equation at the equilibrium state; Step 6: Calculate the error between the predicted lifespan and the actual lifespan, and use it together with the partial differential equation error as the loss function for updating the model. Use these two losses of batch data to continuously update the training model to minimize the error and complete the final lifespan prediction.

3. The method for predicting the remaining life of an aircraft engine integrating physical information according to claim 2, characterized in that: Step 3 includes: (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 starting point of the sensor data sequence is used as the initial window position, and the window is moved in sequence according to the set sliding step size until the entire range of the sensor data sequence is covered; (3) The sensor data in each window is taken as a data sample, which will be used in the subsequent analysis and processing of engine-related characteristics.

4. The method for predicting the remaining life of an aircraft engine integrating physical information according to claim 2, characterized in that: Step 4 includes: (1) A T-MLP model for the time dimension is built using a multi-layer perceptron (MLP) network framework. The time dimension information is extracted by inputting the original sample data into the T-MLP model and mapped to the same shape as the original data. (2) Using the MLP network framework, an F-MLP model is built for the feature dimension. The time dimension information output by the T-MLP model is input into the F-MLP model to extract the feature dimension information and map it to the same shape as the original data. (3) Use SoftMax to build a feature fusion layer, calculate the importance ratios of original data, time dimension information, and feature dimension information respectively, and use them as their respective weights for feature fusion; (4) The fused features are input into a basic MLP network framework and the output dimension is set to 1 to complete the mapping of feature data to life prediction values.

5. The method for predicting the remaining life of an aircraft engine integrating physical information according to claim 2, characterized in that: Step 5 includes: (1) Solve the first-order derivative of the predicted life value r with respect to time t, as well as the first-order and higher-order partial derivatives of the predicted life value with respect to the hidden state h of the fused data; (2) Establish a physical information neural network to learn the mathematical relationship model corresponding to the actual physical process; (3) By training the neural network, the equation F is made as close to 0 as possible to obtain a stable state that satisfies the laws of physics; (4) Calculate the root mean square loss between equation F learned for each batch of samples and equation F(0) as one of the model loss functions. loss1 i =RMSE(F,F(0))。 6. The method for predicting the remaining life of an aircraft engine integrating physical information according to claim 2, characterized in that: Step 6 includes: (1) Use the predicted life value r obtained in step 4 to calculate the root mean square error between the predicted life value ture_r and the true life value as another one in the model loss function; loss2 i =RMSE(r,ture_r) (2) The life loss is combined with the partial differential equation loss loss1 obtained in step 5 i Through SoftMax fusion, a comprehensive loss function loss is constructed; loss=a i *loss1 i +b i *loss2 i (3) The back propagation algorithm is performed based on the comprehensive loss obtained from each batch of samples, and the model parameters are updated and adjusted according to the gradient descent algorithm. This process is repeated until the loss value of the model reaches the preset minimum threshold or meets the convergence condition, thereby completing the final life prediction model training and enabling the model to have high-precision life prediction capabilities.

Images