Aircraft turbine rotating speed fault prediction method and system based on recurrent neural network

Abstract

The invention belongs to the technical field of intelligent manufacturing, and particularly relates to an aircraft turbine rotating speed fault prediction method and system based on a recurrent neural network, and the method comprises the steps: firstly collecting the actual measurement data of the parameters of an aircraft turbine cooler; the working time of the turbine is converted; standardizing the data; firstly, a recurrent neural network model structure is determined; and initializing model parameters, inputting standard training data to adjust recurrent neural network parameters, and inputting standard test data to obtain a prediction result. Based on the identified flight parameter record item related to the turbine cooler rotating speed fault, the recurrent neural network is used for predicting the turbine cooler rotating speed fault, the health condition of the turbine cooler is monitored, and the problem that there is no aircraft turbine cooler rotating speed fault prediction means at the present stage is solved.

Description

Technical Field

[0001] This application belongs to the field of intelligent manufacturing technology, and particularly relates to an aircraft turbine speed fault prediction method and system based on a recurrent neural network. Background Technique

[0002] So far, with the development of fault diagnosis technology, many methods have been proposed. From the initial analytical model-based methods to the current machine learning-based methods, system fault diagnosis can be completed without too much prior knowledge and accurate system analytical models. Machine learning has a wide range of application spaces, and its applications in the field of fault diagnosis mainly include decision trees, neural networks, and support vector machines, etc.

[0003] Among them, the decision tree is a basic classification and regression method. In the "Fault Diagnosis Method of Power Grid Based on Decision Tree and Pulse Neural Membrane System" in 2020, Liu Wei et al. proposed a power grid fault diagnosis method based on decision tree and fuzzy inference pulse neural network. The results show that this method can still correctly diagnose fault elements when single-type and multi-type fault information is lost.

[0004] Support Vector Machine (SVM) is a supervised learning method based on statistical learning theory. In the "Support Vector Classification for Fault Diagnostics if an Electrical Machine" published in 2002, Poyhonen S et al. applied the support vector machine algorithm to motor fault diagnosis, successfully classifying the healthy power spectrum and fault power of the motor and identifying the faults.

[0005] Convolutional Neural Network (CNN) is a supervised learning method initially proposed by Yann Lecun in 1994. In the "Convolutional Neural Network Based Fault Detection for Rotating Machinery" published in 2016, Olovier Janssens et al. used the convolutional neural network method to automatically learn features for bearing fault detection from data, realizing the fault detection and classification problems of gearboxes. Compared with the fault diagnosis accuracy rate of the random forest classifier-based method, there is a significant improvement.

[0006] Although the theory and methods of machine learning have been relatively mature, there is still no relevant engineering application method in the field of aircraft turbine cooler speed fault prediction. Since an aircraft has a very complex flight parameter recording system with numerous maintenance data records, the data is mainly used for evaluating the health status of the aircraft system. At present, domestic aircraft engineers mainly judge whether the flight parameters are abnormal by manually determining whether the data of each data item exceeds the design technical condition range, and diagnose the health status of the aircraft. Summary of the Invention

[0007] To solve the above problems existing in the prior art, the present invention proposes an aircraft turbine cooler speed fault prediction method based on a recurrent neural network. This method is based on identifying the flight parameter records related to the turbine cooler speed fault, uses a recurrent neural network to predict the turbine cooler speed fault, and monitors the health status of the turbine cooler, solving the problem of no means for predicting the turbine cooler speed fault at the present stage.

[0008] To achieve the above object, the technical solution of the present invention is as follows:

[0009] An aircraft turbine speed fault prediction method based on a recurrent neural network, comprising the following steps:

[0010] Step 1.1, first collect the measured data of the parameters of the aircraft turbine cooler;

[0011] Step 1.2, perform conversion processing on the turbine working time;

[0012] Step 1.3, standardize the data;

[0013] Further, the specific content of step 1.1 is to collect the measured data of the parameters of a single aircraft turbine cooler in the normal working state according to the time series. The parameters include turbine working time, compressor inlet temperature, compressor inlet pressure, compressor outlet temperature, compressor outlet pressure, turbine inlet temperature, turbine inlet pressure, turbine outlet temperature, turbine outlet pressure, and turbine speed.

[0014] Furthermore, the data volume is not less than the start-up data of 8 turbine coolers, and the sampling period is 5 times per second; take no less than the first 6 start-up data in sequence as training data, and no less than the last 2 start-up data as test data.

[0015] Further, the specific content of step 1.2 is as follows: perform the following conversion on the turbine working time, convert the turbine working time into a value in seconds, and perform the following processing. The processing formula is:

[0016] A1 * = A1 × 5

[0017] where A1 is the turbine working time;

[0018] A1 * It is the working time of the processed turbine.

[0019] Furthermore, the specific content of step 1.3 is as follows: Using the Z - normalization method, standardize the data of compressor inlet temperature, compressor inlet pressure, compressor outlet temperature, compressor outlet pressure, turbine inlet temperature, turbine inlet pressure, turbine outlet temperature, and turbine outlet pressure in the training data and test data. The formula is as follows:

[0020]

[0021] where Z is the standardized data;

[0022] x is the value of a specific data;

[0023] μ is the average of this group of data;

[0024] σ is the standard deviation of this group of data;

[0025] And the data after time conversion and standardization are called standard training data and standard test data.

[0026] Step 2: First, determine the structure of the recurrent neural network model; initialize the model parameters, then input the standard training data to adjust the parameters of the recurrent neural network, and then input the standard test data to obtain the prediction result.

[0027] Furthermore, step 2 specifically includes the following steps:

[0028] Step 2.1: Determine the specific structure of the recurrent neural network model;

[0029] Step 2.2: Use the training data to train the recurrent neural network model;

[0030] Step 2.3: Determine the data fault judgment threshold;

[0031] Step 2.4: Use the prediction data for prediction.

[0032] Furthermore, the specific content of step 2.1 is as follows: The specific structure of the recurrent neural network model includes an input layer, a hidden layer, and an output layer, and build the recurrent neural network model; The constructed recurrent neural network model adopts 1 input layer, 4 hidden layers, and 1 output layer, and each hidden layer contains 4 neurons.

[0033] Furthermore, the specific content of step 2.2 is as follows:

[0034] Step 2.2.1, Initialize the model: Initialize the parameters of the recurrent neural network, including three weight matrices U, W, V and two bias matrices b and c; Set the number of iterations, learning rate, and parameter roles;

[0035] Step 2.2.2, Forward propagation: Input the standard training data into the recurrent neural network model. Through the forward propagation algorithm, obtain the predicted value of the recurrent neural network under the initial model parameters, and use the difference between the predicted value and the actual turbine speed of the standard training data to adjust the model parameters;

[0036] Step 2.2.3, Backward propagation: Perform backward propagation calculation on the recurrent neural network model. Calculate the error by comparing the previous turbine speed prediction output with the actual turbine speed, and then use the gradient descent method to iterate the model parameters according to the error to further optimize the parameters of the recurrent neural network model, including three weight matrices U, W, V and two bias matrices b and c;

[0037] Step 2.2.4, Iterate repeatedly to determine the final parameters.

[0038] Furthermore, the specific steps of Step 2.2.2 are as follows:

[0039] 2.2.2.1 Calculate the hidden state h(t) of the model at time t. h(t) is obtained from the input x(t) and the hidden state h(t - 1) at the previous time. The formula is as follows:

[0040] h(t) = σ(U x(t) + W h(t - 1) + b)

[0041] Where the activation function σ is the ReLU function, the bias matrix b is the bias of the linear relationship, the weight matrices U and W are the linear relationship parameters of the recurrent neural network model, x(t) represents the input of the training sample at time t, and h(t - 1) represents the hidden state of the model at time t - 1;

[0042] Step 2.2.2.2, Use the hidden state h(t) calculated by the above formula to calculate the output o(t) of the model at time t. The formula is as follows

[0043] o(t) = V h(t) + c

[0044] Where the weight matrix V and the bias matrix c are both parameters of the recurrent neural network model.

[0045] Furthermore, the specific steps of Step 2.2.3 are as follows:

[0046] Step 2.2.3.1, Use the loss function to calculate the difference L between the predicted speed and the actual speed of the model. The formula is as follows:

[0047]

[0048] Where \(L(t)\) is the loss function, \(t\) represents the time, and \(n\) represents the final time. represents the predicted output of the turbine speed at time \(t\), and \(y(t)\) represents the actual output of the turbine speed at time \(t\);

[0049] Step 2.2.3.2, the specific formulas for calculating the weight matrix \(V\) and the bias matrix \(c\) are as follows:

[0050]

[0051] Step 2.2.3.3, calculate the weight matrices \(W\), \(U\) and the bias matrix \(b\), and the formulas are as follows respectively:

[0052]

[0053] Where \(\delta(t)\) represents the gradient of the hidden state at position \(t\), and the function \(diag\) represents taking the diagonal elements of the matrix. represents the predicted output of the turbine speed at time \(t\), and \(y(t)\) represents the actual output of the turbine speed at time \(t\).

[0054] Furthermore, Step 2.2.4 specifically includes the following steps:

[0055] Step 2.2.4.1, re-enter the adjusted parameter recurrent neural network model with the same standard training data;

[0056] Step 2.2.4.2, repeat Step 2.2.2 and Step 2.2.3 to adjust the parameters until the required number of training times is reached, and then determine the model parameters;

[0057] Step 2.2.4.3, determine the final model parameters including the weight matrices \(U\), \(W\), \(V\) and the bias matrices \(b\) and \(c\), and use the recurrent neural network model at this time as the turbine speed prediction model.

[0058] Still further, the specific content of Step 2.3 is as follows:

[0059] Step 2.3.1, use the standard test data as the input, apply the turbine speed prediction model to predict the turbine speed, and calculate the turbine speed residual \(Q\), and the calculation formula is as follows:

[0060]

[0061] represents the predicted output of the turbine speed at time \(t\), and \(y(t)\) represents the actual output of the turbine speed at time \(t\);

[0062] Step 2.3.2: Evaluate the accuracy of the turbine speed prediction model through residuals. When the accuracy meets the requirements, use the final model parameters output in Step 2.2.4.3 and the turbine speed prediction model as the subsequent turbine speed prediction model parameters and prediction model; otherwise, repeat Step 2.2 until the residual accuracy meets the requirements.

[0063] Step 2.3.3: Calculate the standard deviation σ1 and the mean value of the turbine speed residual Q as the threshold for determining whether the predicted data is faulty in the subsequent process.

[0064] Step 2.3.4: When the predicted data exceeds 5 times the standard deviation σ1 for 10 consecutive seconds, that is, when it exceeds the range, it is determined that the turbine speed of the data is abnormal.

[0065] Furthermore, the specific content of Step 2.4 is as follows: First, process the predicted data according to Step 1.2 and Step 1.3, then input it into the model obtained in Step 2.3.2 to obtain the predicted speed. After that, calculate the turbine speed residual according to Step 2.3.1, determine whether there are points exceeding the fault determination standard according to Step 2.3.3, and output the determination result of whether the data is faulty.

[0066] An aircraft turbine speed fault prediction system based on a recurrent neural network, including a data processing module and a recurrent neural network recognition module. The data processing module executes the content of Step 1.1 - Step 1.3; the recurrent neural network recognition module executes the content of Step 2.

[0067] The advantages of this application are as follows:

[0068] 1. Based on the identified flight parameter records related to the turbine cooler speed fault, this invention uses a recurrent neural network to predict the turbine cooler speed fault, monitor the health status of the turbine cooler, and solve the problem of no means for predicting the turbine cooler speed fault at the current stage. In practical applications, it has been proven that this method can predict abnormal speed of the turbine before the turbine cooler speed fault occurs. The fault prediction model provides a basis for the aircraft turbine cooler to carry out condition-based maintenance in advance, reducing the possibility of flight mission termination or flight accidents caused by in-air reports of the aircraft turbine cooler.

[0069] 2. Based on the identified flight parameter records related to the turbine cooler speed fault, this invention uses a recurrent neural network to predict the turbine cooler speed fault, monitor the health status of the turbine cooler, and solve the problem of no means for predicting the turbine cooler speed fault at the current stage. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 It is a schematic flow diagram of the present invention.

[0071] Figure 2It is a flowchart for training a recurrent neural network model. Specific implementation manners

[0072] To make the objectives, technical solutions and advantages of the embodiments of the invention clearer, the technical solutions in the embodiments of the invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the invention. Apparently, the described embodiments are some but not all of the embodiments of the invention. Components of the embodiments of the invention generally described and illustrated in the figures herein can be arranged and designed in a variety of different configurations.

[0073] Therefore, the following detailed description of the embodiments of the invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the invention without creative efforts fall within the scope of protection of the invention.

[0074] It should be noted that like reference numerals and letters denote like items in the following figures, and thus, once an item is defined in one figure, it does not require further definition and explanation in subsequent figures.

[0075] In the description of the invention, it should be noted that the orientation or positional relationship indicated by the terms "upper", "vertical", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings, or the orientation or positional relationship when the product of the invention is normally placed, or the orientation or positional relationship commonly understood by those skilled in the art. It is only for the convenience of describing the invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus should not be construed as a limitation of the invention. In addition, the terms "first", "second", etc. are only used for descriptive distinction and should not be construed as indicating or implying relative importance.

[0076] The aircraft surface feature segmentation method based on contour constraint optimization of the present invention completes the segmentation task of the target in the image based on a deep learning network. The feature extraction backbone network is used to learn the feature information in the image, and then based on this feature information, the outer contour constraint of the target is fitted, and the target to be segmented in the image is initially segmented. Finally, the target contour constraint is used to optimize the segmentation result to achieve high-precision segmentation of each target instance in the image.

[0077] Embodiment 1

[0078] A method for predicting aircraft turbine speed faults based on a recurrent neural network includes the following steps:

[0079] Step 1.1, first collect the measured data of the parameters of the aircraft turbine cooler;

[0080] Step 1.2, perform conversion processing on the turbine working time;

[0081] Step 1.3, standardize the data;

[0082] The specific content of Step 1.1 is to collect the measured data of the parameters of a single aircraft turbine cooler in a normal working state according to the time series. The parameters include turbine working time, compressor inlet temperature, compressor inlet pressure, compressor outlet temperature, compressor outlet pressure, turbine inlet temperature, turbine inlet pressure, turbine outlet temperature, turbine outlet pressure, and turbine speed.

[0083] The data volume is not less than the start-up data of 8 turbine coolers, and the sampling period is 5 times per second; take no less than the first 6 start-up data in sequence as training data, and no less than the last 2 start-up data as test data.

[0084] Furthermore, the specific content of Step 1.2 is as follows: perform the following conversion on the turbine working time to facilitate using the turbine working time as an ID sequence for subsequent calculations when training the neural network model.

[0085] Convert the turbine working time into a value in seconds and perform the following processing. The processing formula is:

[0086] A1 * = A1 × 5

[0087] where A1 is the turbine working time;

[0088] A1 * is the processed turbine working time.

[0089] The specific content of Step 1.3 is as follows: Use the Z-standardization method to standardize the compressor inlet temperature, compressor inlet pressure, compressor outlet temperature, compressor outlet pressure, turbine inlet temperature, turbine inlet pressure, turbine outlet temperature, and turbine outlet pressure data in the training data and test data. The formula is as follows:

[0090]

[0091] where Z is the standard data;

[0092] x is the value of a specific data;

[0093] μ is the average of this group of data;

[0094] σ is the standard deviation of this group of data;

[0095] And the data after time conversion and standardization are called standard training data and standard test data.

[0096] Step 2: First, determine the structure of the recurrent neural network model; initialize the model parameters, then input the standard training data to adjust the parameters of the recurrent neural network, and then input the standard test data to obtain the prediction results.

[0097] The specific steps of Step 2 are as follows:

[0098] Step 2.1: Determine the specific structure of the recurrent neural network model;

[0099] Step 2.2: Use the training data to train the recurrent neural network model;

[0100] Step 2.3: Determine the data fault judgment threshold;

[0101] Step 2.4: Use the prediction data for prediction.

[0102] The specific content of Step 2.1 is as follows: The specific structure of the recurrent neural network model includes an input layer, a hidden layer, and an output layer, and the recurrent neural network model is built; the constructed recurrent neural network model adopts 1 input layer, 4 hidden layers, and 1 output layer, and the number of neurons in each hidden layer is 4.

[0103] The specific content of Step 2.2 is as follows:

[0104] Step 2.2.1: Initialize the model: Initialize the parameters of the recurrent neural network, including three weight matrices U, W, V and two bias matrices b and c; set the number of iterations, the learning rate, and the parameter role;

[0105] Step 2.2.2: Forward propagation: Input the standard training data into the recurrent neural network model, and through the forward propagation algorithm, obtain the prediction value of the recurrent neural network under the initial model parameters, and use it to calculate the difference from the actual turbine speed of the standard training data to adjust the model parameters;

[0106] Step 2.2.3: Backward propagation: Perform backward propagation calculation on the recurrent neural network model, calculate the error by comparing the previous turbine speed prediction output with the actual turbine speed, and then use the gradient descent method to iterate the model parameters according to the error to further optimize the parameters of the recurrent neural network model, including three weight matrices U, W, V and two bias matrices b and c;

[0107] Step 2.2.4: Repeatedly iterate to determine the final parameters.

[0108] The specific steps of Step 2.2.2 are as follows:

[0109] 2.2.2.1 Calculate the hidden state h(t) of the model at time t. h(t) is obtained from the input x(t) and the hidden state h(t-1) at the previous time. The formula is as follows:

[0110] h(t) = σ(Ux(t) + Wh(t - 1) + b)

[0111] Where the activation function σ is the ReLU function, the bias matrix b is the bias of the linear relationship, and the weight matrices U and W are the linear relationship parameters of the recurrent neural network model. x(t) represents the input of the training sample at time t, and h(t - 1) represents the hidden state of the model at time t - 1;

[0112] Step 2.2.2.2: Calculate the output o(t) of the model at time t using the hidden state h(t) calculated by the above formula. The formula is as follows

[0113] o(t) = Vh(t) + c

[0114] Where the weight matrix V and the bias matrix c are both parameters of the recurrent neural network model.

[0115] The specific steps of Step 2.2.3 are as follows:

[0116] Step 2.2.3.1: Calculate the difference L between the predicted speed and the actual speed of the model using the loss function. The formula is as follows:

[0117]

[0118] Where L(t) is the loss function, t represents time, and n represents the final time, represents the predicted output of the turbine speed at time t, and y(t) represents the actual output of the turbine speed at time t;

[0119] Step 2.2.3.2: The specific formulas for calculating the weight matrix V and the bias matrix c are as follows:

[0120]

[0121] Step 2.2.3.3: Calculate the weight matrices W, U and the bias matrix b. The formulas are as follows respectively:

[0122]

[0123] Where δ(t) represents the gradient of the hidden state at position t, and the function diag represents taking the diagonal elements of the matrix, represents the predicted output of the turbine speed at time t, and y(t) represents the actual output of the turbine speed at time t.

[0124] Step 2.2.4 specifically includes the following steps:

[0125] Step 2.2.4.1: Re - input the same standard training data into the recurrent neural network model with adjusted parameters;

[0126] Step 2.2.4.2: Repeat Steps 2.2.2 and 2.2.3 to adjust the parameters until the required number of training times is reached, and then determine the model parameters.

[0127] Step 2.2.4.3: Determine the final model parameters including the weight matrices U, W, V and the bias matrices b and c, and use the recurrent neural network model at this time as the turbine speed prediction model.

[0128] The specific content of Step 2.3 is as follows:

[0129] Step 2.3.1: Take the standard test data as the input, apply the turbine speed prediction model to predict the turbine speed, and calculate the turbine speed residual Q. The calculation formula is as follows:

[0130]

[0131] represents the predicted output of the turbine speed at time t, and y(t) represents the actual output of the turbine speed at time t;

[0132] Step 2.3.2: Evaluate the accuracy of the turbine speed prediction model through the residual. When the accuracy meets the requirements, use the final model parameters and the turbine speed prediction model output in Step 2.2.4.3 as the subsequent turbine speed prediction model parameters and prediction model. Otherwise, repeat Step 2.2 until the residual accuracy meets the requirements.

[0133] Step 2.3.3: Calculate the standard deviation σ1 and the mean value of the turbine speed residual Q as the threshold for subsequent judgment of whether the prediction data is faulty.

[0134] Step 2.3.4: When the prediction data exceeds 5 times the standard deviation σ1 for 10 consecutive seconds, that is, when it exceeds the range, it is determined that the turbine speed of the data is abnormal.

[0135] The specific content of Step 2.4 is as follows: First, process the prediction data according to Steps 1.2 and 1.3, then input it into the model obtained in Step 2.3.2. After obtaining the predicted speed, calculate the turbine speed residual according to Step 2.3.1, and determine whether there are points exceeding the fault judgment standard according to Step 2.3.3, and output the judgment result of whether the data is faulty.

[0136] Embodiment 2

[0137] An aircraft turbine speed fault prediction system based on a recurrent neural network, including a data processing module and a recurrent neural network (RNN) recognition module. The data processing module executes the content of Steps 1.1 - 1.3;

[0138] Step 1.1: First, collect the measured data of the aircraft turbine cooler parameters;

[0139] Step 1.2, perform conversion processing on the turbine working time;

[0140] Step 1.3, standardize the data;

[0141] The specific content of Step 1.1 is to collect the measured data of the parameters of a single aircraft turbine cooler in a normal working state according to the time series. The parameters include turbine working time, compressor inlet temperature, compressor inlet pressure, compressor outlet temperature, compressor outlet pressure, turbine inlet temperature, turbine inlet pressure, turbine outlet temperature, turbine outlet pressure, and turbine speed.

[0142] The amount of data is not less than the start-up data of 8 turbine coolers, and the sampling period is 5 times per second; take no less than the first 6 start-up data in sequence as training data, and no less than the last 2 start-up data as test data.

[0143] The specific content of Step 1.2 is as follows: perform the following conversion on the turbine working time to facilitate subsequent calculations by using the turbine working time as an ID sequence when training the neural network model.

[0144] Convert the turbine working time into a value in seconds and perform the following processing. The processing formula is:

[0145] A1 * = A1 × 5

[0146] where A1 is the turbine working time;

[0147] A1 * is the processed turbine working time.

[0148] The specific content of Step 1.3 is as follows: use the Z-standardization method to standardize the compressor inlet temperature, compressor inlet pressure, compressor outlet temperature, compressor outlet pressure, turbine inlet temperature, turbine inlet pressure, turbine outlet temperature, and turbine outlet pressure data in the training data and test data. The formula is as follows:

[0149]

[0150] where Z is the standard data;

[0151] x is the value of a specific data;

[0152] μ is the average of this group of data;

[0153] σ is the standard deviation of this group of data;

[0154] And the data after time conversion and standardization are called standard training data and standard test data.

[0155] The recurrent neural network recognition module executes the content of step 2:

[0156] Step 2: In the recurrent neural network recognition module, first determine the structure of the recurrent neural network model; initialize the model parameters, then input the standard training data to adjust the recurrent neural network parameters, and then input the standard test data to obtain the prediction result.

[0157] The specific steps of step 2 are as follows:

[0158] Step 2.1: Determine the specific structure of the recurrent neural network model;

[0159] Step 2.2: Use the training data to train the recurrent neural network model;

[0160] Step 2.3: Determine the data fault judgment threshold;

[0161] Step 2.4: Use the prediction data for prediction.

[0162] The specific content of step 2.1 is as follows: The specific structure of the recurrent neural network model includes an input layer, a hidden layer, and an output layer, and the recurrent neural network model is built; the constructed recurrent neural network model uses 1 input layer, 4 hidden layers, and 1 output layer, and the number of neurons in each hidden layer is 4.

[0163] The specific content of step 2.2 is as follows:

[0164] Step 2.2.1: Initialize the model: Initialize the recurrent neural network parameters, including three weight matrices U, W, V and two bias matrices b and c; set the number of iteration rounds, learning rate, and parameter roles;

[0165] Step 2.2.2: Forward propagation: Input the standard training data into the recurrent neural network model, and through the forward propagation algorithm, obtain the prediction value of the recurrent neural network under the initial model parameters, and use the difference between the predicted value and the actual turbine speed of the standard training data to adjust the model parameters;

[0166] Step 2.2.3: Backward propagation: Perform backward propagation calculation on the recurrent neural network model, calculate the error by comparing the previous turbine speed prediction output with the actual turbine speed, and then use the gradient descent method to iterate the model parameters according to the error to further optimize the recurrent neural network model parameters, including three weight matrices U, W, V and two bias matrices b and c;

[0167] Step 2.2.4: Repeatedly iterate to determine the final parameters.

[0168] The specific steps of step 2.2.2 are as follows:

[0169] 2.2.2.1 Calculate the hidden state h(t) of the model at time t. h(t) is obtained from the input x(t) and the hidden state h(t-1) at the previous time, and its formula is as follows:

[0170] h(t) = σ(U x(t) + W h(t-1) + b)

[0171] Where the activation function σ is the ReLU function, the bias matrix b is the bias of the linear relationship, the weight matrices U and W are the linear relationship parameters of the recurrent neural network model, x(t) represents the input of the training sample at time t, and h(t-1) represents the hidden state of the model at time t-1;

[0172] Step 2.2.2.2, use the hidden state h(t) calculated by the above formula to calculate the output o(t) of the model at time t, and the formula is as follows

[0173] o(t) = V h(t) + c

[0174] Where the weight matrix V and the bias matrix c are both parameters of the recurrent neural network model.

[0175] The specific steps of the said Step 2.2.3 are as follows:

[0176] Step 2.2.3.1, use the loss function to calculate the difference L between the predicted speed and the actual speed of the model, and the formula is as follows:

[0177]

[0178] Where L(t) is the loss function, t represents time, n represents the final time, represents the predicted output of the turbine speed at time t, and y(t) represents the actual output of the turbine speed at time t;

[0179] Step 2.2.3.2, the specific formulas for calculating the weight matrix V and the bias matrix c are as follows:

[0180]

[0181] Step 2.2.3.3, calculate the weight matrices W, U and the bias matrix b, and the formulas are as follows respectively:

[0182]

[0183]

[0184] Where δ(t) represents the gradient of the hidden state at position t, and the function diag represents taking the diagonal elements of the matrix, represents the predicted output of the turbine speed at time t, and y(t) represents the actual output of the turbine speed at time t.

[0185] Step 2.2.4 specifically includes the following steps:

[0186] Step 2.2.4.1, re-enter the adjusted parameter recurrent neural network model with the same standard training data;

[0187] Step 2.2.4.2, repeat Steps 2.2.2 and 2.2.3 to adjust the parameters until the required number of training times is reached, and then determine the model parameters;

[0188] Step 2.2.4.3, determine that the final model parameters include the weight matrices U, W, V and the bias matrices b and c, and use the recurrent neural network model at this time as the turbine speed prediction model.

[0189] The specific content of Step 2.3 is as follows:

[0190] Step 2.3.1, use the standard test data as the input, apply the turbine speed prediction model to predict the turbine speed, and calculate the turbine speed residual Q. The calculation formula is as follows:

[0191]

[0192] represents the predicted output of the turbine speed at time t, and y(t) represents the actual output of the turbine speed at time t;

[0193] Step 2.3.2, evaluate the accuracy of the turbine speed prediction model through the residual. When the accuracy meets the requirements, use the final model parameters and the turbine speed prediction model output in Step 2.2.4.3 as the subsequent turbine speed prediction model parameters and prediction model. Otherwise, repeat Step 2.2 until the residual accuracy meets the requirements;

[0194] Step 2.3.3, calculate the standard deviation σ1 and the mean value of the turbine speed residual Q as the threshold for subsequent judgment of whether the prediction data is faulty;

[0195] Step 2.3.4, when the prediction data exceeds 5 times the standard deviation σ1 continuously for 10S, that is, exceeds the range, it is determined that the turbine speed of the data is abnormal.

[0196] The specific content of Step 2.4 is: first process the prediction data according to Steps 1.2 and 1.3, then input it into the model obtained in Step 2.3.2. After obtaining the predicted speed, calculate the turbine speed residual according to Step 2.3.1, determine whether there are points exceeding the fault judgment standard according to Step 2.3.3, and output the judgment result of whether the data is faulty.

[0197] Embodiment 3

[0198] Step 1 In the data processing module, first collect the measured data of the relevant parameters of the aircraft turbine cooler, convert the turbine working time, and then standardize some of the data.

[0199] The specific steps in the above Step 1 include the following steps:

[0200] 1.1 Collect the measured data of the relevant parameters of a single turbine cooler of a certain CB aircraft in the normal working state according to the time series, including the start data of 8 turbine coolers. The parameters include turbine working time, compressor inlet temperature, compressor inlet pressure, compressor outlet temperature, compressor outlet pressure, turbine inlet temperature, turbine inlet pressure, turbine outlet temperature, turbine outlet pressure, and turbine speed. The sampling period is 5 times per second; use the first 6 start data as training data and the last 2 as test data.

[0201] 1.2 Conversion of turbine working time: Convert the turbine working time (A1) as follows to facilitate subsequent calculations with the turbine working time as an ID sequence when training the neural network model.

[0202] Convert the parameter turbine working time into a value in seconds and perform the following processing. The processing formula is:

[0203] A1 * = A1 × 5

[0204] where A1 is the turbine working time;

[0205] A1 * is the processed turbine working time;

[0206] 1.3 Use the Z-standardization method to standardize the compressor inlet temperature, compressor inlet pressure, compressor outlet temperature, compressor outlet pressure, turbine inlet temperature, turbine inlet pressure, turbine outlet temperature, and turbine outlet pressure data in the above training data and test data. The formula is as follows

[0207]

[0208] where Z is the standard data;

[0209] x is the value of a specific data;

[0210] μ is the average of this group of data;

[0211] σ is the standard deviation of this group of data;

[0212] The data after time conversion and standardization are called standard training data and standard test data.

[0213] Step 2 In the recurrent neural network recognition module, first determine the structure of the recurrent neural network model, initialize the model parameters, then input the processed training data to adjust the recurrent neural network parameters, and then input the test data to obtain the prediction result.

[0214] The specific steps in the above Step 2 include the following steps:

[0215] 2.1 Determine the specific structure of the recurrent neural network model, including the input layer, hidden layer, and output layer, and build the recurrent neural network model; the constructed recurrent neural network model uses 1 input layer, 4 hidden layers, and 1 output layer, and each hidden layer contains 4 neurons;

[0216] 2.2 The training process of the recurrent neural network model, the specific steps are as follows:

[0217] 2.2.1 Initialize the model: Use the random normal method to initialize the recurrent neural network parameters, including three weight matrices U, W, V and two bias matrices b and c; set the number of iterations to 250 and the learning rate to 0.005; use the converted turbine working time as the ID sequence, and use the compressor inlet temperature, compressor inlet pressure, compressor outlet temperature, compressor outlet pressure, turbine inlet temperature, turbine inlet pressure, turbine outlet temperature, and turbine outlet pressure as independent variables, and use the turbine speed as the dependent variable.

[0218] 2.2.2 Forward propagation: Input the standard training data into the recurrent neural network model, and through the forward propagation algorithm, obtain the predicted value of the turbine speed under the initial model parameters, and use it to calculate the difference from the turbine cooler speed of the standard training data to adjust the model parameters. The specific steps are as follows:

[0219] 2.2.2.1 Calculate the hidden state h(t) of the model at time t. h(t) can be obtained from the input x(t) and the hidden state h(t - 1) at the previous moment. The formula is as follows:

[0220] h(t) = σ(U x(t)+W h(t - 1)+b)

[0221] Where the activation function σ is the ReLU function, the bias matrix b is the bias of the linear relationship, the weight matrices U and W are the linear relationship parameters of the recurrent neural network model, x(t) represents the input of the training sample at time t, and h(t - 1) represents the hidden state of the model at time t - 1.

[0222] 2.2.2.2 Use the hidden state h(t) calculated by the above formula to calculate the output o(t) of the model at time t. The formula is as follows

[0223] o(t) = V h(t)+c

[0224] Among them, both the weight matrix V and the bias matrix c are parameters of the recurrent neural network model

[0225] 2.2.3 Backpropagation: Perform backpropagation calculation on the recurrent neural network model. Calculate the error by comparing the previous predicted output of the turbine cooler speed with the actual turbine cooler speed, and then use the gradient descent method to iterate the model parameters according to the error to further optimize the recurrent neural network model parameters, including three weight matrices U, W, V and two bias matrices b and c. The specific steps are as follows:

[0226] 2.2.3.1 Use the loss function to calculate the difference L between the predicted speed of the model and the actual speed. The formula is as follows:

[0227]

[0228] Where L(t) is the loss function, t represents the time, n represents the final time, represents the predicted output of the turbine speed at time t, and y(t) represents the actual output of the turbine speed at time t.

[0229] 2.2.3.2 Calculate the weight matrix V and the bias matrix c. The specific formulas are as follows:

[0230]

[0231] 2.2.3.3 Calculate the weight matrices W, U and the bias matrix b. The formulas are as follows respectively:

[0232]

[0233] Where δ(t) represents the gradient of the hidden state at position t, and the function diag represents taking the diagonal elements of the matrix, represents the predicted output of the turbine speed at time t, and y(t) represents the actual output of the turbine speed at time t.

[0234] 2.2.4 Repeated iteration to determine the final parameters: The steps to determine the final parameters are as follows:

[0235] 2.2.4.1 Re-enter the adjusted recurrent neural network model with the same standard training data.

[0236] 2.2.4.2 Repeat steps 2.2.2 and 2.2.3 to adjust the parameters until the required number of training times is reached, and then determine the model parameters.

[0237] 2.2.4.3 Determine the final model parameters including the weight matrices U, W, V and the bias matrices b and c, and use the recurrent neural network model at this time as the turbine speed prediction model.

[0238] 2.3 Determine the threshold for data fault judgment. The steps for determining the threshold are as follows:

[0239] 2.3.1 Take the standard test data as input, apply the turbine speed prediction model to predict the turbine speed, and calculate the turbine speed residual. The calculation formula is as follows:

[0240]

[0241] represents the predicted output of the turbine speed at time t, and y(t) represents the actual output of the turbine speed at time t.

[0242] 2.3.2 Calculate the standard deviation σ1 = 0.097 and the mean of the turbine speed residual Q Meet the accuracy requirements, and use it as the threshold for subsequent judgment of whether the predicted data is faulty. Take the final model parameters output in 2.2.4.3 and the turbine speed prediction model as the subsequent turbine speed prediction model parameters and prediction model.

[0243] 2.3.3 When the residual of the predicted data exceeds [-0.396, 0.574] continuously for 10S (i.e., 50 consecutive data), it is determined that the turbine speed of the data is abnormal.

[0244] 2.4 Extract the data of the turbine cooler speed failure and the first four turbine cooler startups before the failure of a certain aircraft. After processing according to steps 1.2 and 1.3 first, input the model obtained in step 2.3.2. After obtaining the predicted speed, calculate the turbine speed residual according to 2.3.1. Through data analysis, it is found that there are points where the data exceeds the threshold range [-0.396, 0.574] continuously for 10S in the third, fourth, and fifth turbine operation data. Enter the residual data into the MATLAB software and draw a time series diagram, and it can be significantly seen that there is an expanding trend in the out-of-tolerance data of the third, fourth, and fifth times.

Claims

1. A method for predicting aircraft turbine speed faults based on a recurrent neural network, characterized in that: The steps include: Step 1.1, first collect the measured data of aircraft turbine cooler parameters; Step 1.2, converting the turbine working time; Step 1.3, standardize the data; Step 2, first determine the recurrent neural network model structure; initialize the model parameters, then input standard training data to adjust the recurrent neural network parameters, and then input standard test data to obtain the prediction results.

2. The aircraft turbine speed fault prediction method based on recurrent neural network according to claim 1 is characterized in that: The specific content of step 1.1 is to collect the measured data of the parameters of a single aircraft turbine cooler under normal working conditions in time series, and the parameters include turbine working time, compressor inlet temperature, compressor inlet pressure, compressor outlet temperature, compressor outlet pressure, turbine inlet temperature, turbine inlet pressure, turbine outlet temperature, turbine outlet pressure, and turbine speed.

3. The aircraft turbine speed fault prediction method based on recurrent neural network according to claim 2 is characterized in that: The data volume is no less than 8 turbine cooler startup data, and the sampling period is 5 times / second; the first no less than 6 startup data are taken in sequence as training data, and the last no less than 2 startup data are taken as test data.

4. The aircraft turbine speed fault prediction method based on recurrent neural network according to claim 1 is characterized in that: The specific content of step 1.2 is: convert the turbine working time as follows, convert the turbine working time into a value in seconds, and perform the following processing. The processing formula is: A1 * =A1×5 Where A1 is the turbine working time; A1 * It is the working time of the turbine after treatment.

5. The aircraft turbine speed fault prediction method based on recurrent neural network according to claim 1, characterized in that: The specific content of step 1.3 is: using the Z normalization method, the compressor inlet temperature, compressor inlet pressure, compressor outlet temperature, compressor outlet pressure, turbine inlet temperature, turbine inlet pressure, turbine outlet temperature, and turbine outlet pressure data in the training data and the test data are normalized, and the formula is as follows: Where Z is the standard data; x is the value of a specific data; μ is the mean of the data set; σ is the standard deviation of the data set; The time-converted and standardized data are called standard training data and standard test data.

6. The aircraft turbine speed fault prediction method based on recurrent neural network according to claim 1, characterized in that: The step 2 specifically includes the following steps: Step 2.1, determine the specific structure of the recurrent neural network model; Step 2.2, use the training data to train the recurrent neural network model; Step 2.3, determine the data fault judgment threshold; Step 2.4, use the prediction data to make predictions.

7. The aircraft turbine speed fault prediction method based on recurrent neural network according to claim 6 is characterized in that: The specific content of step 2.1 is: the specific structure of the recurrent neural network model includes an input layer, a hidden layer and an output layer, and the recurrent neural network model is set up; the constructed recurrent neural network model adopts 1 input layer, 4 hidden layers and 1 output layer, and the number of neurons contained in each hidden layer is 4.

8. The aircraft turbine speed fault prediction method based on recurrent neural network according to claim 6 is characterized in that: The specific content of step 2.2 is: Step 2.2.1, Initialize the model: Initialize the recurrent neural network parameters, including three weight matrices U, W, V and two bias matrices b and c; set the number of iterations, learning rate, and parameter roles; Step 2.2.2, forward propagation: input the standard training data into the recurrent neural network model, and obtain the recurrent neural network prediction value under the initial model parameters through the forward propagation algorithm, which is used to adjust the model parameters by subtracting the actual turbine speed of the standard training data; Step 2.2.3, back propagation: back propagation calculation is performed on the recurrent neural network model. The error is calculated by comparing the previous turbine speed prediction output with the actual turbine speed, and then the model parameters are iterated using the gradient descent method according to the error to further optimize the recurrent neural network model parameters, including three weight matrices U, W, V and two bias matrices b and c; Step 2.2.4, iterate repeatedly to determine the final parameters.

9. The aircraft turbine speed fault prediction method based on recurrent neural network according to claim 8, characterized in that: The specific steps of step 2.2.2 are as follows: 2.2.2.1 Calculate the hidden state h(t) of the model at time t. h(t) is obtained by the input x(t) and the hidden state h(t-1) at the previous moment. The formula is as follows: h(t)=σ(U x(t)+W h(t-1)+b) The activation function σ is the ReLU function, the bias matrix b is the bias of the linear relationship, the weight matrices U and W are the linear relationship parameters of the recurrent neural network model, x(t) represents the input of the training sample at time t, and h(t-1) represents the hidden state of the model at time t-1; Step 2.2.2.2, use the hidden state h(t) calculated by the above formula to calculate the output o(t) of the model at time t, the formula is as follows o(t)=V h(t)+c The weight matrix V and the bias matrix c are both recurrent neural network model parameters.

10. The aircraft turbine speed fault prediction method based on recurrent neural network according to claim 8, characterized in that: The specific steps of step 2.2.3 are: Step 2.2.3.1, use the loss function to calculate the difference L between the model predicted speed and the actual speed, the formula is as follows: Where L(t) is the loss function, t represents the time, and n represents the final time. represents the predicted output of turbine speed at time t, and y(t) represents the actual output of turbine speed at time t; Step 2.2.3.2, calculate the weight matrix V and bias matrix c. The specific formula is as follows: Step 2.2.3.3, calculate the weight matrix W, U and bias matrix b, the formulas are as follows: Where δ(t) represents the gradient of the hidden state at position t, and the function diag represents taking the diagonal elements of the matrix. represents the predicted output of turbine speed at time t, and y(t) represents the actual output of turbine speed at time t.

11. The aircraft turbine speed fault prediction method based on recurrent neural network according to claim 8, characterized in that: Step 2.2.4 specifically includes the following steps: Step 2.2.4.1, use the same standard training data to re-input the recurrent neural network model with adjusted parameters; Step 2.2.4.2, repeat steps 2.2.2 and 2.2.3 to adjust the parameters until the required number of training times is reached, and then determine the model parameters; Step 2.2.4.3, determine the final model parameters including weight matrices U, W, V and bias matrices b and c, and use the recurrent neural network model at this time as the turbine speed prediction model.

12. The aircraft turbine speed fault prediction method based on recurrent neural network according to claim 11, characterized in that: The specific contents of step 2.3 are as follows: Step 2.3.1, using the standard test data as input, applying the turbine speed prediction model to predict the turbine speed, and calculating the turbine speed residual Q, the calculation formula is as follows: represents the predicted output of turbine speed at time t, and y(t) represents the actual output of turbine speed at time t; Step 2.3.2, evaluate the accuracy of the turbine speed prediction model through the residual. When the accuracy meets the requirements, the final model parameters and turbine speed prediction model output by step 2.2.4.3 are used as the subsequent turbine speed prediction model parameters and prediction model. Otherwise, repeat step 2.2 until the residual accuracy meets the requirements. Step 2.3.3, calculate the standard deviation σ1 and mean of the turbine speed residual Q, which will be used as the threshold for determining whether the predicted data is faulty; Step 2.3.4: When the predicted data exceeds 5 times the standard deviation σ1 for 10 consecutive seconds, it is considered as exceeding range, it is determined that the turbine speed is abnormal.

13. The aircraft turbine speed fault prediction method based on recurrent neural network according to claim 12, characterized in that: The specific content of step 2.4 is: first process the predicted data according to steps 1.2 and 1.3, then input the model obtained in step 2.3.2, and after obtaining the predicted speed, calculate the turbine speed residual according to step 2.3.1, and determine whether there are points that exceed the fault judgment standard according to step 2.3.3, and output the judgment result of whether the data is faulty.

14. An aircraft turbine speed fault prediction system based on recurrent neural network, characterized in that: It includes a data processing module and a recurrent neural network recognition module, wherein the data processing module executes the contents of steps 1.1 to 1.3 as claimed in claim 1; and the recurrent neural network recognition module executes the contents of step 2 as claimed in claim 1.

Images