Unmanned aerial vehicle aero-engine RUL prediction method based on multi-working-condition state monitoring data

Abstract

The invention discloses an unmanned aerial vehicle (UAV) aero-engine RUL prediction method based on multi-working-condition state monitoring data, and relates to the field of UAVs, and the method comprises the steps: recording historical original data of a state monitoring sensor and a corresponding historical RUL in a normal operation state of a target engine; historical data is subjected to data preprocessing and divided into a training set and a test set, so that model learning data features can be predicted conveniently; a CNN-BiLSTM neural network hybrid model based on Bayesian optimization is constructed; a training set data training model is imported, after training is completed, test set data is input to evaluate the performance of the model, and evaluation indexes of the model are RMSE values and Score values; and recording original data of the state monitoring sensor in the current normal operation state of the target engine, and inputting the data into the trained prediction model to realize RUL prediction of the target engine.

Description

Technical Field

[0001] The invention relates to the field of unmanned aerial vehicles (UAVs), and more particularly to a method for predicting the RUL of an UAV aero-engine based on multi-operating condition monitoring data. Background Art

[0002] As the "heart" of aircraft, aircraft engines are cutting-edge products in the manufacturing industry. Therefore, the development level of aircraft engines, especially the development level of UAV aircraft engines, is a concrete carrier of a country's scientific and technological level and industrial development. In the actual operation of aircraft engines, various components will inevitably degrade under the influence of various factors. When the degradation accumulates to a certain critical point, the aircraft engine will fail or even fail to function, which poses great risks to the environment, personnel safety, economy and other aspects. In order to prevent common or unknown failures of aircraft engines and improve the safety and reliability of aircraft engines, people have adopted the Prognostics Health Management (PHM) technology. PHM technology is divided into two parts: fault prediction and health management. The core of health management is the remaining useful life (RUL) prediction technology.

[0003] At present, there are three types of RUL technologies for aircraft engines based on the prediction principle: model-based RUL prediction, experience-based RUL prediction, and data-driven RUL prediction. The model-based RUL prediction method establishes a model of the aircraft engine and infers the RUL of the aircraft engine based on the working principle of the aircraft engine; the experience-based RUL prediction method uses the previously recorded equipment history information to infer the probability of equipment failure and predicts the RUL of the aircraft engine based on the probability; the data-driven RUL prediction method uses deep learning technology to process data based on equipment history and real-time collected status data, and infers the RUL of the aircraft engine through data processing.

[0004] Although the above three methods can realize the prediction of engine life, there are still obvious shortcomings: the model-based RUL prediction method needs to collect a large amount of structural, material and other characteristic information of the equipment, combine a large number of complex operating mechanisms of aircraft engines and establish an accurate mathematical and physical model of the equipment; although the experience-based RUL prediction method only requires the historical fault information of the equipment and the prediction process is relatively simple, the prediction accuracy of RUL is low; the data-driven method often uses a single neural network model, and the hyperparameters of the model also need to be manually debugged many times, and the RUL is often predicted under the premise of a single working condition, which does not conform to the actual working state of aircraft engines under multiple working conditions. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention proposes a UAV aviation engine RUL prediction method based on multi-operating condition monitoring data, which is used to solve the problems of single prediction model, single operating condition, and manual debugging of model hyperparameters mentioned in the background technology.

[0006] The steps of the technical solution adopted by the present invention are specifically as follows:

[0007] The RUL prediction method of UAV aircraft engine based on multi-condition status monitoring data includes the following steps:

[0008] Step 1: Record the historical raw data and corresponding historical RUL of the condition monitoring sensor of the target engine under normal operating conditions;

[0009] Step 2: Preprocess the historical data and divide it into training set and test set;

[0010] Step 3: Construct a CNN-BiLSTM neural network hybrid model based on Bayesian optimization;

[0011] Step 4: Input the training set data to train the model, input the test set data to evaluate the model, and use the RMSE value and Score value as evaluation parameters to evaluate the model performance;

[0012] Step 5: Record the raw data of the condition monitoring sensor of the target engine under the current normal operating state and input the data into the trained model. The model outputs the predicted result of the RUL of the target engine.

[0013] Preferably, the specific steps of step 1 include:

[0014] The flight altitude, throttle lever angle, Mach number, fan inlet temperature, low-pressure compressor exhaust temperature, high-pressure compressor exhaust temperature, low-pressure turbine exhaust temperature, fan inlet pressure, bypass duct total pressure, high-pressure compressor outlet total pressure, uncorrected fan speed, uncorrected core engine speed, engine pressure ratio, high-pressure compressor outlet static pressure, fuel flow and high-pressure compressor outlet total pressure ratio, fan corrected speed, core engine corrected speed, bypass ratio, combustion chamber gas ratio, bleed air enthalpy, fan speed command value, fan correction speed command value, high-pressure turbine cooling air flow, and low-pressure turbine cooling air flow are recorded as the historical raw data of the UAV engine sensor and the corresponding historical RUL.

[0015] Preferably, the specific steps of step 2 include:

[0016] Step 2.1: Perform Max-Min normalization on the historical raw sensor data acquired by the sensor;

[0017] Step 2.2: Perform variational mode decomposition on the data processed in step 2.1;

[0018] Step 2.3: Use a sliding window to divide the data processed in step 2.2 and the corresponding RUL into windows;

[0019] Step 2.4: Perform random ordering operation on the data processed in step 2.3;

[0020] Step 2.5: Divide the data processed in step 2.4 into training set and test set.

[0021] Preferably, the specific steps of step 3 include:

[0022] Step 3.1: Build a CNN neural network;

[0023] Step 3.2: Build a BiLSTM neural network;

[0024] Step 3.3: Build a CNN-BiLSTM neural network hybrid model;

[0025] Step 3.4: Set the early stop callback function;

[0026] Step 3.5: Set the Bayesian optimizer parameters;

[0027] Step 3.6: Input the training set and test set into the model, start the Bayesian optimizer and call the early stopping callback function set in step 3.4 to perform hyperparameter search;

[0028] Step 3.7: Record the best hyperparameters output by the Bayesian optimizer and reconstruct the CNN-BiLSTM neural network hybrid model based on the best hyperparameters.

[0029] Preferably, the optimization object of the Bayesian optimizer in step 3.4 includes:

[0030] The number of CNN layers, the number of filters in the CNN layer, the size of the convolution kernel, the number of neurons in the BiLSTM layer, the number of neurons in the fully connected layer, the Dropout rate of the CNN layer, the Dropout rate of the Dropout layer after the BiLSTM layer, and the learning rate of the model.

[0031] Preferably, in step 4, the specific calculation formula for the evaluation parameter RMSE value and Score value of the CNN-BiLSTM neural network hybrid model based on Bayesian optimization is as follows:

[0032]

[0033]

[0034] Where, d i is the difference between the predicted RUL value and the true value of the i-th engine in the test set.

[0035] Compared with the prior art, the present invention has the following advantages:

[0036] The invention proposes a method for predicting the RUL of a UAV aeroengine based on multi-condition status monitoring data. After acquiring the status detection data of the UAV aeroengine sensor, the data is subjected to Max-Min normalization. This processing method can effectively reduce the influence of interference information on the data, realize dimensionless data, standardize the amplitude range of different sensor data, and realize dimensionless comparison between different sensor data. After completing the Max-Min normalization, the data is subjected to variational modal decomposition, and the signal is decomposed into three modal components with fixed frequency bandwidths. Each modal component corresponds to a fixed specific frequency and amplitude in the signal. This processing method can separate the degradation information of the aeroengine from the information on the change in normal operating parameters, greatly reduce the interference of interference information on real information, and characterize the degradation process of the aeroengine with three different periodic scales. Using a sliding window to divide the data can increase the utilization rate of the data, and realize The data is enhanced now; the random disorder operation can effectively reduce the logic of the data in each window, suppress the overfitting of the model during training, and improve the generalization ability of the model; the CNN-BiLSTM neural network hybrid model consists of a CNN layer and a BiLSTM layer; the CNN layer can extract data feature information, and the convolution operation can achieve effective filtering; the BiLSTM layer can retain long-term degradation features while learning the current degradation features, thereby improving the prediction accuracy and generalization ability of the model; the Bayesian optimization method has efficient global search capabilities and robustness, takes less time, and saves computing resources; the early stopping rule can assist the Bayesian optimization. When the number of iterations increases and the monitoring indicators do not change significantly, the optimization is stopped, further reducing time and saving computing resources; the two parameters of RMSE value and Score value can digitize the prediction accuracy of the model and quantitatively reflect the accuracy of the prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 The present invention is a flow chart of the method.

[0038] Figure 2 An example diagram of the variational mode decomposition of the fan inlet temperature.

[0039] Figure 3 This is the structure diagram of the CNN neural network.

[0040] Figure 4 This is the BiLSTM neural network structure diagram.

[0041] Figure 5 This is a structural diagram of the CNN-BiLSTM neural network hybrid model of the present invention.

[0042] Figure 6 Flowchart for building the hybrid CNN-BiLSTM neural network model with optimal hyperparameters.

[0043] Figure 7 Prediction flow chart of the CNN-BiLSTM neural network hybrid model with the best hyperparameters. DETAILED DESCRIPTION

[0044] In order to enable persons skilled in the relevant art to more clearly understand the technical solutions and advantages of the present invention, the present invention is further clearly and completely described below in conjunction with the accompanying drawings.

[0045] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by technicians in the technical field to which the present invention belongs. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0046] The RUL prediction method of UAV aircraft engine based on multi-condition status monitoring data includes the following steps:

[0047] Step 1: Record the historical raw data of the condition monitoring sensor and the corresponding historical RUL under the normal operating state of the target engine.

[0048] The data records of step 1 are flight altitude, throttle lever angle, Mach number, fan inlet temperature, low-pressure compressor exhaust temperature, high-pressure compressor exhaust temperature, low-pressure turbine exhaust temperature, fan inlet pressure, bypass duct total pressure, high-pressure compressor outlet total pressure, uncorrected fan speed, uncorrected core engine speed, engine pressure ratio, high-pressure compressor outlet static pressure, fuel flow and high-pressure compressor outlet total pressure ratio, fan corrected speed, core engine corrected speed, bypass ratio, combustion chamber gas ratio, bleed air enthalpy, fan speed command value, fan correction speed command value, high-pressure turbine cooling air flow, and low-pressure turbine cooling air flow as the historical raw data of the aviation UAV engine sensor and the corresponding historical RUL.

[0049] Step 2: Preprocess the historical data and divide it into training set and test set;

[0050] Step 2.1: Perform Max-Min normalization on the historical raw sensor data acquired by the sensor;

[0051] Step 2.2: Perform variational mode decomposition on the data processed in step 2.1;

[0052] Step 2.3: Use a sliding window to divide the data processed in step 2.2 and the corresponding RUL into windows;

[0053] Step 2.4: Perform random ordering operation on the data processed in step 2.3;

[0054] Step 2.5: Divide the data processed in step 2.4 into training set and test set.

[0055] The Max-Min normalization in step 2.1 is to use Python 3.9 in the Pycharm platform to call the MinMaxScaler function of the sklearn library to perform Max-Min normalization on the historical raw data of the aviation drone engine sensor, so that the distribution of each data is transformed into a distribution with a mean of 0 and a standard deviation of 1. The specific calculation formula of Max-Min normalization is:

[0056]

[0057] In the above formula, x norm is the sample data after Max-Min normalization, x is the original data of the sensor, and x max is the maximum value of the original data of the same dimension of the sensor, x min It is the minimum value of the original data of the same dimension of the sensor.

[0058] The variational mode decomposition in step 2.2 calls the VMD function in the vmdpy library to decompose the sensor signal into three modal components with fixed frequency and bandwidth. The frequency of each modal component is automatically determined by the variational mode decomposition. The specific calculation formula of variational mode decomposition is:

[0059]

[0060] In the above formula, u k is the kth modal component after variational mode decomposition, w k is the center frequency of the kth mode component after variational mode decomposition, “*” is the convolution operation, δ(t) is the Dirac function, is the gradient solution, and f(t) is the original sensor signal.

[0061] Taking the decomposition of fan intake temperature as an example, the modal components of the fan intake temperature after variational mode decomposition are as follows: Figure 2 As shown;

[0062] In step 2.3, the window division uses a window with a length of window_length and a step size of 1 to divide the historical data of each aircraft engine. The specific calculation formula for the value of window_length is:

[0063] window_length=min(20,min_num_machine)

[0064] In the above formula, window_length is the value of the sliding window size, min_num_machine is the number of data of the engine with the smallest number of recorded data among all recorded engine data, and the min() function takes the minimum value from "20" and min_num_machine as the function value.

[0065] The random shuffling operation in step 2.4 calls the random.permutation function in the numpy library to randomly shuffle each window divided in step 2.3 and disrupt the order of the data in the window, thereby suppressing overfitting during subsequent model training.

[0066] In step 2.5, the data set is divided into training set and test set by calling the train_test_split function of the scikit-learn library, where test_size is set to 0.2, that is, the training set accounts for 80% of the total data set and the test set accounts for 20% of the total data set.

[0067] Step 3: Construct a CNN-BiLSTM neural network hybrid model based on Bayesian optimization;

[0068] Step 3.1: Build a CNN neural network;

[0069] Step 3.2: Build a BiLSTM neural network;

[0070] Step 3.3: Build a CNN-BiLSTM neural network hybrid model;

[0071] Step 3.4: Set the early stop callback function;

[0072] Step 3.5: Set the Bayesian optimizer parameters;

[0073] Step 3.6: Input the training set and test set into the model, start the Bayesian optimizer and call the early stopping callback function set in step 3.4 to perform hyperparameter search;

[0074] Step 3.7: Record the best hyperparameters output by the Bayesian optimizer and reconstruct the CNN-BiLSTM neural network hybrid model based on the best hyperparameters.

[0075] Each CNN layer in the CNN neural network of step 3.1 consists of 1 1D convolutional layer, 1 BN layer, 1 1D maximum pooling layer and 1 Dropout layer. There is a flattening layer at the end of all CNN layers.

[0076] In the CNN neural network, the number of layers, number of filters, convolution kernel size, and Dropout rate of the CNN layer are all determined by the Bayesian optimizer. The convolution kernel of the convolution layer is a one-dimensional convolution kernel, the activation function is "Relu", the pooling method is maximum pooling, and the pooling window size is 2.

[0077] The structure of CNN neural network is as follows Figure 3 shown.

[0078] The specific calculation formula for extracting feature information at the CNN layer is:

[0079]

[0080] Where n = 1, 2, ..., N, m is the number of information features, T is the length of the time series processed by the convolution kernel once, σ is the activation function, W is the weight, and b is the bias.

[0081] The BiLSTM neural network in step 3.2 has only one BiLSTM layer, where the BiLSTM layer consists of a forward LSTM, a backward LSTM, and an activation function. Both the forward LSTM and the backward LSTM are hidden parameters and do not need to be manually added or set.

[0082] In the BiLSTM neural network, the number of neurons in the BiLSTM layer is determined by the Bayesian optimizer, and the activation function is "Relu".

[0083] The BiLSTM neural network structure is as follows Figure 4 shown.

[0084] The specific calculation formula of the BiLSTM layer is:

[0085]

[0086] In the formula, f is the LSTM unit function, g is the activation function, and w is x is the corresponding weight.

[0087] The CNN-BiLSTM neural network hybrid model in step 3.3 consists of a CNN neural network, a BiLSTM neural network, 1 fully connected layer, 1 Dropout layer, and 1 fully connected layer.

[0088] In the CNN-BiLSTM neural network hybrid model of step 3.3, the number of neurons in the fully connected layer following the BiLSTM neural network, the Dropout rate of the Dropout layer after the BiLSTM layer, and the learning rate of the model are all determined by the Bayesian optimizer. The activation function of the fully connected layer following the BiLSTM neural network is "Relu", and the number of neurons in the fully connected layer at the end of the model is "1", and the activation function is "linear".

[0089] The structure of the CNN-BiLSTM neural network hybrid model is as follows Figure 5 shown.

[0090] The callback function of step 3.4 sets the monitoring indicator of the callback function to "val_loss" by setting the parameters of the keras.callbacks.EarlyStopping function, and sets the maximum number of iterations to "10" when there is no obvious change in the monitoring indicator.

[0091] The call and parameter setting of the Bayesian optimizer in step 3.5 are implemented by calling the BayesianOptimization function in the keras_tuner library. The objective function of the optimizer is "mae", the maximum number of optimization trials is 50, the number of Bayesian optimization random sampling initial points is 2, the weight factor used in the expected improvement calculation is 0.0001, and the weight factor used in the probability improvement calculation is 2.6.

[0092] Among all the optimization objects of the Bayesian optimizer, the optimization range of the number of CNN layers is 1-3; the optimization range of the number of filters in the CNN layer is 8-128, with a step size of 4; the optimization range of the convolution kernel size is 2-10, the optimization range of the number of neurons in the BiLSTM layer is 8-128, with a step size of 4; the optimization range of the number of neurons in the fully connected layer is 32-128, with a step size of 32; the optimization selection of the Dropout rate of the CNN layer is 0.2, 0.3, 0.4, 0.5; the optimization selection of the Dropout rate of the Dropout layer after the BiLSTM layer is 0.2, 0.3, 0.4, 0.5; the optimization selection of the model learning rate is 0.01, 0.001, 0.0001.

[0093] The startup of the Bayesian optimizer in step 3.6 is set up through the tuner.search function. The training set and test set are input into the tuner.search function. The training round is set to 100. The callback function calls the early stopping callback function in step 3.4.

[0094] The optimal parameters of the Bayesian optimizer in step 3.7 are obtained through the tuner.get_best_hyperparameters function. The number of CNN layers, the number of CNN layer filters, the convolution kernel size, the number of BiLSTM layer neurons, the number of fully connected layer neurons, the Dropout rate of the CNN layer, the Dropout rate of the Dropout layer after the BiLSTM layer, and the learning rate of the model are reset according to the optimal parameters.

[0095] The construction process of the optimal hyperparameter CNN-BiLSTM neural network hybrid model is as follows Figure 6 shown.

[0096] Step 4: Input the training set data to train the model, input the test set data to evaluate the model, and use the RMSE value and Score value as evaluation parameters to evaluate the model performance;

[0097] Step 4.1: Input the training set into the CNN-BiLSTM neural network hybrid model for training;

[0098] Step 4.2: Input the test set into the CNN-BiLSTM neural network hybrid model, and use the RMSE value and Score value to evaluate the prediction effect of the model.

[0099] In step 4.1, set the model's optimizer to "Adam", the loss function to "mse", the training rounds to 100, and the training batch size to 64.

[0100] In step 4.2, the specific calculation formulas for the evaluation parameters RMSE and Score of the CNN-BiLSTM neural network hybrid model based on Bayesian optimization are as follows:

[0101]

[0102]

[0103] Where, d i is the difference between the predicted RUL value and the true value of the i-th engine in the test set.

[0104] The smaller the values ​​of the evaluation parameters RMSE and Score, the higher the prediction accuracy of the model.

[0105] Step 5: Record the raw data of the condition monitoring sensor of the target engine under the current normal operating state and input the data into the trained model, and the model outputs the predicted result of the RUL of the target engine;

[0106] Step 5.1: Record the raw data of the condition monitoring sensor of the target engine under the current normal operating state;

[0107] Step 5.2: Perform Max-Min normalization on the recorded sensor data;

[0108] Step 5.3: Perform variational mode decomposition on the data processed in step 5.2;

[0109] Step 5.4: Use sliding windows to divide the data processed in step 5.3 into windows;

[0110] Step 5.5: Input the data processed in step 5.4 into the trained CNN-BiLSTM neural network hybrid model;

[0111] Step 5.5: CNN-BiLSTM neural network hybrid model outputs predicted RUL;

[0112] The process of model prediction is as follows Figure 7 shown.

[0113] The above implementation modes are only for illustrating the technical concept and features of the present invention, and their purpose is to enable technicians in this field to understand the content of the present invention and implement it. They cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made according to the spirit of the present invention should be included in the protection scope of the present invention.

Claims

1. A UAV aircraft engine RUL prediction method based on multi-condition status monitoring data, characterized in that: The method specifically comprises the following steps: Step 1: Record the historical raw data and corresponding historical RUL of the condition monitoring sensor of the target engine under normal operating conditions; Step 2: Preprocess the historical data and divide it into training set and test set; Step 3: Construct a CNN-BiLSTM neural network hybrid model based on Bayesian optimization; Step 4: Input the training set data to train the model, input the test set data to evaluate the model, and use the RMSE value and Score value as evaluation parameters to evaluate the model performance; Step 5: Record the raw data of the condition monitoring sensor of the target engine under the current normal operating state and input the data into the trained model. The model outputs the predicted result of the RUL of the target engine.

2. The method for predicting RUL of a UAV aircraft engine based on multi-operating condition monitoring data according to claim 1 is characterized in that: In step 1, The historical data required to be recorded for the aviation UAV engine include flight altitude, throttle lever angle, Mach number, fan inlet temperature, low-pressure compressor exhaust temperature, high-pressure compressor exhaust temperature, low-pressure turbine exhaust temperature, fan inlet pressure, total bypass pressure, high-pressure compressor outlet total pressure, uncorrected fan speed, uncorrected core engine speed, engine pressure ratio, high-pressure compressor outlet static pressure, fuel flow and high-pressure compressor outlet total pressure ratio, fan corrected speed, core engine corrected speed, bypass ratio, combustion chamber gas ratio, bleed air enthalpy, fan speed command value, fan correction speed command value, high-pressure turbine cooling air flow, low-pressure turbine cooling air flow and corresponding historical RUL.

3. The method for predicting RUL of a UAV aircraft engine based on multi-condition status monitoring data according to claim 1 is characterized in that: In step 2, the specific steps of preprocessing the historical data are: Step 2.1: Perform Max-Min normalization on the historical raw sensor data acquired by the sensor; Step 2.2: Perform variational mode decomposition on the data processed in step 2.1; Step 2.3: Use a sliding window to divide the data processed in step 2.2 and the corresponding RUL into windows; Step 2.4: Perform random ordering operation on the data processed in step 2.3; Step 2.5: Divide the data processed in step 2.4 into training set and test set.

4. The method for predicting RUL of a UAV aircraft engine based on multi-operating condition monitoring data according to claim 1, characterized in that: In step 3, the CNN-BiLSTM neural network hybrid model based on Bayesian optimization includes a CNN neural network, a BiLSTM neural network and a Bayesian optimizer. The specific steps for constructing the CNN-BiLSTM neural network hybrid model based on Bayesian optimization are: Step 3.1: Build a CNN neural network; Step 3.2: Build a BiLSTM neural network; Step 3.3: Build a CNN-BiLSTM neural network hybrid model; Step 3.4: Set the early stop callback function; Step 3.5: Set the Bayesian optimizer parameters; Step 3.6: Input the training set and test set into the model, start the Bayesian optimizer and call the early stopping callback function set in step 3.4 to perform hyperparameter search; Step 3.7: Record the best hyperparameters output by the Bayesian optimizer and reconstruct the CNN-BiLSTM neural network hybrid model based on the best hyperparameters.

5. The method for predicting RUL of a UAV aircraft engine based on multi-operating condition monitoring data according to claim 4 is characterized in that: In step 3, The optimization objects of the Bayesian optimizer are the number of CNN layers, the number of filters in the CNN layer, the size of the convolution kernel, the number of neurons in the BiLSTM layer, the number of neurons in the fully connected layer, the Dropout rate of the CNN layer, the Dropout rate of the Dropout layer after the BiLSTM layer, and the learning rate of the model.

6. The method for predicting RUL of a UAV aircraft engine based on multi-operating condition monitoring data according to claim 1, characterized in that: In step 4, the specific calculation formulas for the evaluation parameter RMSE value and Score value of the CNN-BiLSTM neural network hybrid model based on Bayesian optimization are as follows: Where, d i is the difference between the predicted RUL value and the true value of the i-th engine in the test set.

Images