CN102562239A

CN102562239A - Method for monitoring exhaust temperature of aircraft engine

Info

Publication number: CN102562239A
Application number: CN2011103695719A
Authority: CN
Inventors: 丁刚; 雷达; 于广滨; 钟诗胜
Original assignee: Harbin Institute of Technology Shenzhen
Current assignee: Harbin Institute of Technology Shenzhen
Priority date: 2011-11-13
Filing date: 2011-11-13
Publication date: 2012-07-11

Abstract

The invention discloses a method for monitoring the exhaust temperature of an aero-engine. The method comprises the following steps: (1) collecting and storing engine exhaust temperature data; Establish and test the multidimensional training sample set and test sample set of the prediction model and the model input for prediction; (3) establish a prediction model based on convolutional and discrete process neural networks, use the Levenberg-Marquardt algorithm to train the prediction model, test and store Predictive model; (4) Use the predictive model to predict the engine exhaust temperature, and apply it to the monitoring of the state of the aeroengine. The method of the invention is beneficial to improving the status monitoring and maintenance support level of the aeroengine, so as to achieve the purpose of ensuring flight safety.

Description

A method for monitoring the exhaust temperature of an aero-engine

技术领域 technical field

本发明涉及一种航空发动机排气温度的监测方法。The invention relates to a method for monitoring the exhaust temperature of an aeroengine.

背景技术 Background technique

航空发动机是飞机的心脏，其健康状态是影响航空安全的关键因素。排气温度是表征航空发动机健康状态和决定发动机是否可以安全使用的关键参数，是航空公司进行发动机调度和制定合理的发动机检测、维修计划的重要指标，对排气温度进行监视和趋势预测可以保证发动机的安全运行，并且为发动机的维修保障提供决策支持。由于排气温度在随时间不断变化的过程中受到许多复杂非线性时变因素的影响，对其变化趋势进行精确预测是一个实际技术难题，现有技术中只有过程神经网络能够对排气温度中的时间累积效应进行有效处理，而采用过程神经网络对排气温度进行预测时，首先要将采集到的离散排气温度数据拟合成连续函数得到网络输入，然后将输入函数和权函数正交基展开，以简化时间聚合运算中的积分运算。这种处理方法存在的不足是：(1)离散采样数据可能不存在解析函数形式；(2)对输入函数和权函数进行正交基展开时，基函数的个数选择没有理论依据，只能依靠试探确定；(3)拟合离散样本生成输入函数、对输入函数和权函数进行基展开的过程可能造成一定的信息丢失导致精度损失。以上不足导致已有过程神经网络模型的使用过程复杂，且对发动机排气温度的预测难以取得较好效果，从而限制了过程神经网络在航空发动机排气温度监测工程实际中的应用。Aeroengine is the heart of an aircraft, and its health is a key factor affecting aviation safety. Exhaust temperature is a key parameter that characterizes the health status of an aeroengine and determines whether the engine can be used safely. The safe operation of the engine, and provide decision support for the maintenance of the engine. Since the exhaust temperature is affected by many complex nonlinear time-varying factors in the process of changing with time, it is a practical technical problem to accurately predict its change trend. In the prior art, only the process neural network can predict the exhaust temperature. The time accumulation effect of time is effectively processed, and when the process neural network is used to predict the exhaust temperature, the collected discrete exhaust temperature data must first be fitted to a continuous function to obtain the network input, and then the input function and the weight function are orthogonal Basis expansion to simplify integration operations in temporal aggregation operations. The shortcomings of this processing method are: (1) the discrete sampling data may not exist in the form of analytical functions; (2) when performing orthogonal basis expansion on the input function and weight function, there is no theoretical basis for the number of basis functions to be selected, and only Determined by trial and error; (3) The process of fitting discrete samples to generate input functions and performing basis expansion on input functions and weight functions may cause certain information loss and precision loss. The above deficiencies make the use of the existing process neural network model complex, and it is difficult to achieve good results in the prediction of engine exhaust temperature, which limits the application of process neural network in the actual engineering of aeroengine exhaust temperature monitoring.

因此，急需一种实用而有效的方法对排气温度进行监视和趋势预测，从而提高航空发动机状态监视和维修保障水平，从而保证飞行安全。Therefore, there is an urgent need for a practical and effective method to monitor and predict the trend of exhaust temperature, so as to improve the status monitoring and maintenance support level of aero-engines, so as to ensure flight safety.

发明内容 Contents of the invention

本发明的目的是为了提供一种能够对航空发动机排气温度随时间变化的过程进行监视和趋势预测的方法。The purpose of the present invention is to provide a method capable of monitoring and predicting the trend of the process of the exhaust temperature of the aeroengine changing with time.

为达到以上目的，本发明的方法具有如下步骤：To achieve the above object, the method of the present invention has the following steps:

步骤(1)：通过安装在航空发动机上的温度传感器采集发动机排气温度数据，并且通过飞机通信寻址与报告系统传输到地面的发动机性能监控中心进行预处理并且存储；Step (1): collect the engine exhaust temperature data through the temperature sensor installed on the aeroengine, and transmit it to the engine performance monitoring center on the ground through the aircraft communication addressing and reporting system for preprocessing and storage;

步骤(2)：从发动机性能监控中心数据库提取指定时段的发动机排气温度历史数据，构成一维的排气温度时间序列，在此基础上构建用于建立和测试预测模型的多维训练样本集、测试样本集以及用于预测的预测模型输入，其具体方法为：采用排气温度时间序列的前m个值对第m+1个值进行预测，使用排气温度历史数据构成一个m列的变量矩阵X，将变量矩阵X的每一列视为同一个变量在不同时刻的取值，从而变量矩阵X每一列的数据可视为时间段[0，T]内对应变量的L个离散采样，采用前L行作为输入就可以模拟过程式输入，根据以上原理可以构建样本对{d_i，x_i}，其中x_i为L×m的矩阵，对应变量矩阵X中的前L行，其中x_i的每一列对应一个过程式输入，d_i为对应目标值。对于长度为N的排气温度时间序列，可以构建N-m-L+1个样本对，将构建的样本对按照一定的比例划分为训练样本集和测试样本集；Step (2): Extract the historical data of engine exhaust temperature for a specified period from the database of the engine performance monitoring center to form a one-dimensional exhaust temperature time series. On this basis, construct a multi-dimensional training sample set for establishing and testing the prediction model, The test sample set and the prediction model input for prediction, the specific method is: use the first m values of the exhaust temperature time series to predict the m+1th value, and use the exhaust temperature historical data to form an m-column variable Matrix X, each column of the variable matrix X is regarded as the value of the same variable at different times, so the data of each column of the variable matrix X can be regarded as L discrete samples of the corresponding variable in the time period [0, T], using The first L rows can be used as input to simulate procedural input. According to the above principles, a sample pair {d _i , xi _} can be constructed, where _xi is an L×m matrix, corresponding to the first L rows in the variable matrix X, where _xi Each column of corresponds to a procedural input, and d _i is the corresponding target value. For the exhaust temperature time series with length N, Nm-L+1 sample pairs can be constructed, and the constructed sample pairs can be divided into training sample set and test sample set according to a certain ratio;

步骤(3)：建立基于卷积和离散过程神经网络的预测模型，采用步骤(2)建立的训练样本集对预测模型进行训练，然后采用测试样本集对预测模型进行测试，确定合适的预测模型参数，存储预测模型，其具体方法为：建立发动机排气温度预测的卷积和离散过程神经网络预测模型，其拓扑结构为一个输入层、一个隐含层和一个输出层，模型采用向量输入以模型过程式输入，采用卷积和运算处理时间累积效应。输入步骤(2)生成的排气温度历史数据就可以得到排气温度的预测值，其预测值由如下的卷积和离散过程神经网络预测模型决定：Step (3): Establish a prediction model based on convolutional and discrete process neural networks, use the training sample set established in step (2) to train the prediction model, and then use the test sample set to test the prediction model to determine the appropriate prediction model Parameters, storage prediction model, the specific method is: build the convolution and discrete process neural network prediction model of engine exhaust temperature prediction, its topological structure is an input layer, a hidden layer and an output layer, the model uses vector input to The model is procedurally input, and convolution and operations are used to handle the cumulative effect of time. The predicted value of the exhaust temperature can be obtained by inputting the exhaust temperature historical data generated in step (2), which is determined by the following convolutional and discrete process neural network prediction models:

$y the y = = {Σ Σ}_{j j = = 11}^{m m} {v v}_{j j} f f (({Σ Σ}_{i i = = 11}^{n no} {ω ω}_{ji the ji} * * {x x}_{i i} + + {θ θ}_{11}^{j j})) + + {θ θ}_{22} = = {Σ Σ}_{j j = = 11}^{m m} {v v}_{j j} f f (({Σ Σ}_{i i = = 11}^{n no} {Σ Σ}_{l l = = 11}^{L L} {ω ω}_{ji the ji} ((l l)) {x x}_{i i} ((l l)) + + {θ θ}_{11}^{j j})) + + {θ θ}_{22}$

其中，*表示卷积运算，ω_ji为连接隐层第j个神经元与输入层第i个输入单元的权向量，x_i为长度为L的输入向量，为输入层第i个输入单元；v_j为连接隐层第j个神经元和输出神经元的权值；

为隐层第j个神经元的激励阈值；θ₂为输出层激励阈值，根据Levenberg-Marquardt训练算法采用训练样本对预测模型进行训练，并采用测试样本对预测模型的预测效果进行测试。如果测试效果满足要求，进行下一步骤，否则，重复进行步骤(3)直到获得满意的预测效果，存储最终得到的预测模型；Among them, * represents the convolution operation, ω _ji is the weight vector connecting the j-th neuron of the hidden layer and the i-th input unit of the input layer, x _i is the input vector of length L, and is the i-th input unit of the input layer; v _j is the weight connecting the jth neuron of the hidden layer and the output neuron;

is the excitation threshold of the jth neuron in the hidden layer; _θ2 is the excitation threshold of the output layer. According to the Levenberg-Marquardt training algorithm, the prediction model is trained with training samples, and the prediction effect of the prediction model is tested with test samples. If the test effect meets the requirements, proceed to the next step, otherwise, repeat step (3) until a satisfactory prediction effect is obtained, and store the final prediction model;

步骤(4)：输入步骤(2)中采用发动机排气温度历史数据生成的用于预测的模型输入样本，使用预测模型对未来一段时间内的发动机排气温度进行预测，并且将其应用于航空发动机实际状态监控，包括判断排气温度变化趋势，监测排气温度是否有突变。Step (4): Input the model input samples for prediction generated by using the historical data of engine exhaust temperature in step (2), use the prediction model to predict the engine exhaust temperature in a period of time in the future, and apply it to aviation Monitoring of the actual state of the engine, including judging the trend of the exhaust temperature and monitoring whether there is a sudden change in the exhaust temperature.

本发明具有如下的有益效果：The present invention has following beneficial effect:

通过本发明的方法，可以对航空发动机排气温度时变过程进行精确建模，从而能够有效处理排气温度时间序列的时间累积效应，因此具有更高的预测精度，相较于已有方法对发动机排气温度的预测具有更好的适应性，从而有利于从而提高航空发动机状态监视和维修保障水平，从而达到保证飞行安全的目的。Through the method of the present invention, the time-varying process of the exhaust temperature of the aeroengine can be accurately modeled, so that the time accumulation effect of the exhaust temperature time series can be effectively dealt with, so it has higher prediction accuracy, compared with the existing method for The prediction of engine exhaust temperature has better adaptability, which is conducive to improving the status monitoring and maintenance support level of aeroengines, so as to achieve the purpose of ensuring flight safety.

附图说明 Description of drawings

图1是卷积和离散过程神经元示意图，Figure 1 is a schematic diagram of convolutional and discrete process neurons,

图2是卷积和离散过程神经网络示意图，Figure 2 is a schematic diagram of convolutional and discrete process neural networks,

图3是航空发动机排气温度监测方法流程图。Fig. 3 is a flowchart of a method for monitoring exhaust gas temperature of an aero-engine.

具体实施方式 Detailed ways

本实施方式的方法步骤如下：The method steps of this embodiment are as follows:

(1)通过安装在航空发动机上的温度传感器采集发动机排气温度数据，并且通过飞机通信寻址与报告系统传输到地面的发动机性能监控中心进行预处理并且存储；(1) Collect the engine exhaust temperature data through the temperature sensor installed on the aeroengine, and transmit it to the engine performance monitoring center on the ground through the aircraft communication addressing and reporting system for preprocessing and storage;

(2)从发动机性能监控中心数据库提取指定时段的发动机排气温度历史数据，构成一维的排气温度时间序列，在此基础上构建用于建立和测试预测模型的多维训练样本集、测试样本集以及用于预测的预测模型输入，(2) Extract the historical data of engine exhaust temperature for a specified period from the database of the engine performance monitoring center to form a one-dimensional exhaust temperature time series. On this basis, construct a multi-dimensional training sample set and test sample for establishing and testing the prediction model set and the predictive model input for prediction,

记由发动机排气温度历史数据构成的时间序列为

选取前m个数据构成一个输入向量对第m+1个数据进行预测，可以得到以下的矩阵：Record the time series composed of historical data of engine exhaust temperature as

Select the first m data to form an input vector to predict the m+1th data, and the following matrix can be obtained:

$X x = = [\begin{matrix} {x x}_{m m}^{11},, {x x}_{m m - - 11}^{11},, {x x}_{m m - - 22}^{11},, . . . . . . \\ {x x}_{m m}^{22},, {x x}_{m m - - 11}^{22},, {x x}_{m m - - 22}^{22},, . . . . . . \\ \begin{matrix} . . & . . & . . \\ . . & . . & . . \\ . . & . . & . . \end{matrix} \\ {x x}_{m m}^{R R},, {x x}_{m m - - 11}^{R R},, {x x}_{m m - - 22}^{R R},, . . . . . . \end{matrix}] - - - - - - ((11))$

其中

将变量矩阵X的每一列视为同一个变量在不同时刻的取值，从而变量矩阵X每一列的数据可视为时间段[0，T]内对应变量的L个离散采样，采用前L行作为输入就可以模拟过程式输入，根据以上原理可以构建样本对{d_i，x_i}，其中x_i为L×m的矩阵，对应变量矩阵X中的前L行，其中x_i的每一列对应一个过程式输入，d_i为对应目标值。对于长度为N的排气温度时间序列，可以构建N-m-L+1个样本对。按照以上的方法构建训练样本集、测试样本集以及用于预测的预测模型输入；in

Each column of the variable matrix X is regarded as the value of the same variable at different times, so the data of each column of the variable matrix X can be regarded as L discrete samples of the corresponding variable in the time period [0, T], using the first L rows As the input, the procedural input can be simulated. According to the above principles, the sample pair {d _i , xi _} can be constructed, where _xi is an L×m matrix, corresponding to the first L rows in the variable matrix X, and each column of _xi Corresponding to a procedural input, d _i is the corresponding target value. For the exhaust gas temperature time series with length N, Nm-L+1 sample pairs can be constructed. Construct the training sample set, test sample set and prediction model input for prediction according to the above method;

(3)建立基于卷积和离散过程神经网络的预测模型，采用步骤(2)建立的训练样本集对预测模型进行训练，然后采用测试样本集对预测模型进行测试，确定合适的预测模型参数，存储预测模型，(3) set up the predictive model based on convolution and discrete process neural network, adopt the training sample set that step (2) establishes to predictive model is trained, adopt test sampled set to test predictive model then, determine suitable predictive model parameter, store predictive models,

具体方法为：将如附图1所示的卷积和离散过程神经元按照一定的拓扑结构组合成卷积和离散过程神经网络预测模型。卷积和离散过程神经元通过卷积和运算来实现对时间累积效应的处理。由卷积和运算的原理可知

其中

为卷积计算的结果，记

为其运算结果的第L个元素，则对于第i个输入向量x_i与其对应权值的卷积和运算有：The specific method is as follows: the convolution and discrete process neurons shown in Figure 1 are combined into a convolution and discrete process neural network prediction model according to a certain topology. Convolution and Discrete Process Neurons process the cumulative effects of time through convolution and sum operations. From the principle of convolution and operation, we can know that

in

For the result of convolution calculation, record

is the Lth element of its operation result, then for the convolution sum operation of the i-th input vector x _i and its corresponding weight:

${K K}_{i i} ((&CircleTimes; &CircleTimes;)) | | L L = = {x x}_{i i} * * {w w}_{i i} = = {Σ Σ}_{l l = = 11}^{L L} {x x}_{i i} ((l l)) {w w}_{i i} ((L L - - l l + + 11)) - - - - - - ((22))$

显然

的值与输入向量x_i的每一个元素都有关系，因此取

为卷积和离散过程神经网络时间聚合运算的结果。记ω_i(l)＝w_i(L－l+1)，显然ω_i亦为长度为L的向量，则(2)式可以写为：obviously

The value of is related to each element of the input vector x _i , so take

Results of temporal aggregation operations for convolutional and discrete process neural networks. Note ω _i (l) = w _i (L－l+1), obviously ω _i is also a vector with length L, then formula (2) can be written as:

${K K}_{i i} ((&CircleTimes; &CircleTimes;)) | | L L = = {Σ Σ}_{l l = = 11}^{L L} {x x}_{i i} ((l l)) {ω ω}_{i i} ((l l)) - - - - - - ((33))$

因此，卷积和离散过程神经元的输入输出映射关系由以下方程决定：Therefore, the input-output mapping relationship of convolutional and discrete process neurons is determined by the following equation:

$y the y = = f f (({Σ Σ}_{i i = = 11}^{n no} {x x}_{i i} * * {ω ω}_{i i} + + θ θ)) = = f f (({Σ Σ}_{i i = = 11}^{n no} {Σ Σ}_{l l = = 11}^{L L} {x x}_{i i} ((l l)) {ω ω}_{i i} ((l l)) + + θ θ)) - - - - - - ((44))$

式中，x_i是长度为L的输入向量，ω_i为权向量，其长度也为L，θ为激励阈值。*表示卷积运算，f(·)为激励函数，为sigmoid函数。In the formula, x _i is the input vector with length L, ω _i is the weight vector, and its length is also L, and θ is the excitation threshold. *Indicates convolution operation, f( ) is the activation function, which is the sigmoid function.

建立拓扑结构如附图2所示卷积和离散过程神经网络预测模型，具体为含有n个输入单元，为n个长度为L的向量；隐层含有m个卷积和离散过程神经元，激励函数为sigmoid函数；输出层有一个传统神经元，其激励函数为线性函数，预测模型的预测结果由以下方程决定：Establish the topological structure as shown in Figure 2. The convolution and discrete process neural network prediction model specifically contains n input units, which are n vectors of length L; the hidden layer contains m convolution and discrete process neurons, and the excitation The function is a sigmoid function; the output layer has a traditional neuron, and its activation function is a linear function, and the prediction result of the prediction model is determined by the following equation:

$y the y = = {Σ Σ}_{j j = = 11}^{m m} {v v}_{j j} f f (({Σ Σ}_{i i = = 11}^{n no} {ω ω}_{ji the ji} * * {x x}_{i i} + + {θ θ}_{11}^{j j})) + + {θ θ}_{22} = = {Σ Σ}_{j j = = 11}^{m m} {v v}_{j j} f f (({Σ Σ}_{i i = = 11}^{n no} {Σ Σ}_{l l = = 11}^{L L} {ω ω}_{ji the ji} ((l l)) {x x}_{i i} ((l l)) + + {θ θ}_{11}^{j j})) + + {θ θ}_{22} - - - - - - ((55))$

式中，*表示卷积运算，ω_ji为连接隐层第j个神经元与输入层第i个输入单元的权向量，x_i为长度为L的输入向量，为输入层第i个输入单元；v_j为连接隐层第j个神经元和输出神经元的权值；

为隐层第j个神经元的激励阈值；θ₂为输出层激励阈值。In the formula, * represents the convolution operation, ω _ji is the weight vector connecting the jth neuron of the hidden layer and the i-th input unit of the input layer, x _i is the input vector of length L, and is the i-th input unit of the input layer ;v _j is the weight connecting the jth neuron of the hidden layer and the output neuron;

is the excitation threshold of the jth neuron in the hidden layer; θ ₂ is the excitation threshold of the output layer.

采用Levenberg-Marquardt训练算法对卷积和离散过程神经网络进行训练。设定网络的学习误差精度、最大迭代次数；初始化网络待训练参数ω_ji，v_j，θ₁，θ₂以及迭代次数q，然后将所有训练样本输入到网络。利用式(6)计算网络的实际输出，然后计算误差e_s＝d_s－y_s，其中y_s表示第S个输入对应的网络输出值，d_s是与其对应的目标值；接下来按下式计算平方和误差SSE：Convolutional and discrete process neural networks were trained using the Levenberg-Marquardt training algorithm. Set the learning error precision and the maximum number of iterations of the network; initialize the network parameters to be trained ω _ji , v _j , θ ₁ , θ ₂ and the number of iterations q, and then input all training samples to the network. Use formula (6) to calculate the actual output of the network, and then calculate the error e _s = d _s -y _s , where y _s represents the network output value corresponding to the Sth input, and d _s is the corresponding target value; then press The formula calculates the sum of square error SSE:

$SSE SSE = = {Σ Σ}_{s the s = = 11}^{S S} {[[{d d}_{s the s} - - {y the y}_{s the s}]]}^{22} = = {Σ Σ}_{s the s = = 11}^{S S} {[[{d d}_{s the s} - - (({Σ Σ}_{j j = = 11}^{m m} {v v}_{j j} f f (({Σ Σ}_{i i = = 11}^{n no} {Σ Σ}_{l l = = 11}^{L L} {ω ω}_{ji the ji} ((l l)) {x x}_{i i} ((l l)) + + {θ θ}_{11}^{j j})) + + {θ θ}_{22}))]]}^{22} - - - - - - ((66))$

按照下式以迭代的方式调整参数矩阵W：Adjust the parameter matrix W iteratively according to the following formula:

$\{\begin{matrix} ΔW ΔW ((q q)) = = - - {[[{J J}^{T T} ((W W ((q q)))) \cdot \cdot J J ((W W ((q q)))) + + μ μ ((q q)) \cdot \cdot I I]]}^{- - 11} \cdot \cdot {J J}^{T T} ((W W ((q q)))) \cdot \cdot E E. ((W W ((q q)))) \\ W W ((q q + + 11)) = = W W ((q q)) + + ΔW ΔW ((q q)) \end{matrix} - - - - - - ((77))$

式中：q为迭代次数；I为单位矩阵；μ为学习速率；W(q)和W(q+1)分别指第q次和第q+1次迭代的权矩阵，ΔW(q)为第q次迭代的权矩阵增量；E(W(q))为第q次迭代的误差矩阵，J(W)为关于W的Jacobi矩阵，J(W(q))为关于W(q)的Jacobi矩阵，J(W)的表达式具体可写为如下形式：In the formula: q is the number of iterations; I is the identity matrix; μ is the learning rate; W(q) and W(q+1) refer to the weight matrix of the qth iteration and the q+1th iteration respectively, and ΔW(q) is The weight matrix increment of the qth iteration; E(W(q)) is the error matrix of the qth iteration, J(W) is the Jacobi matrix about W, and J(W(q)) is about W(q) The Jacobi matrix, the expression of J(W) can be specifically written as the following form:

$J J ((W W)) = = [\begin{matrix} \frac{{&PartialD; &PartialD; e e}_{11}}{{&PartialD; &PartialD; ω ω}_{1111} ((11))},, . . . . . .,, \frac{{&PartialD; &PartialD; e e}_{11}}{{&PartialD; &PartialD; ω ω}_{1111} ((L L))},, . . . . . .,, \frac{{&PartialD; &PartialD; e e}_{11}}{{&PartialD; &PartialD; ω ω}_{mn mn} ((11))},, . . . . . .,, \frac{{&PartialD; &PartialD; e e}_{11}}{{&PartialD; &PartialD; ω ω}_{mn mn} ((L L))},, \frac{{&PartialD; &PartialD; e e}_{11}}{{&PartialD; &PartialD; θ θ}_{11}^{11}},, . . . . . .,, \frac{{&PartialD; &PartialD; e e}_{11}}{{&PartialD; &PartialD; θ θ}_{11}^{m m}},, \frac{{&PartialD; &PartialD; e e}_{11}}{{&PartialD; &PartialD; v v}_{11}},, . . . . . .,, \frac{{&PartialD; &PartialD; e e}_{11}}{{&PartialD; &PartialD; v v}_{m m}},, \frac{{&PartialD; &PartialD; e e}_{11}}{{&PartialD; &PartialD; θ θ}_{22}} \\ \frac{{&PartialD; &PartialD; e e}_{22}}{{&PartialD; &PartialD; ω ω}_{1111} ((11))},, . . . . . .,, \frac{{&PartialD; &PartialD; e e}_{22}}{{&PartialD; &PartialD; ω ω}_{1111} ((L L))},, . . . . . .,, \frac{{&PartialD; &PartialD; e e}_{22}}{{&PartialD; &PartialD; ω ω}_{mn mn} ((11))},, . . . . . .,, \frac{{&PartialD; &PartialD; e e}_{22}}{{&PartialD; &PartialD; ω ω}_{mn mn} ((L L))},, \frac{{&PartialD; &PartialD; e e}_{22}}{{&PartialD; &PartialD; θ θ}_{11}^{11}},, . . . . . .,, \frac{{&PartialD; &PartialD; e e}_{22}}{{&PartialD; &PartialD; θ θ}_{11}^{m m}},, \frac{{&PartialD; &PartialD; e e}_{22}}{{&PartialD; &PartialD; v v}_{11}},, . . . . . .,, \frac{{&PartialD; &PartialD; e e}_{22}}{{&PartialD; &PartialD; v v}_{m m}},, \frac{{&PartialD; &PartialD; e e}_{22}}{{&PartialD; &PartialD; θ θ}_{22}} \\ \begin{matrix} . . & . . & . . & . . & . . & . . & . . & . . & . . & . . & . . & . . & . . & . . \\ . . & . . & . . & . . & . . & . . & . . & . . & . . & . . & . . & . . & . . \\ . . & . . & . . & . . & . . & . . & . . & . . & . . & . . & . . & . . & . . & . . \end{matrix} \\ \frac{{&PartialD; &PartialD; e e}_{S S}}{{&PartialD; &PartialD; ω ω}_{1111} ((11))},, . . . . . .,, \frac{{&PartialD; &PartialD; e e}_{S S}}{{&PartialD; &PartialD; ω ω}_{1111} ((L L))},, . . . . . .,, \frac{{&PartialD; &PartialD; e e}_{S S}}{{&PartialD; &PartialD; ω ω}_{mn mn} ((11))},, . . . . . .,, \frac{{&PartialD; &PartialD; e e}_{S S}}{{&PartialD; &PartialD; ω ω}_{mn mn} ((L L))},, \frac{{&PartialD; &PartialD; e e}_{S S}}{{&PartialD; &PartialD; θ θ}_{11}^{11}},, . . . . . .,, \frac{{&PartialD; &PartialD; e e}_{S S}}{{&PartialD; &PartialD; θ θ}_{11}^{m m}},, \frac{{&PartialD; &PartialD; e e}_{S S}}{{&PartialD; &PartialD; v v}_{11}} . . . . . .,, \frac{{&PartialD; &PartialD; e e}_{S S}}{{&PartialD; &PartialD; v v}_{m m}},, \frac{{&PartialD; &PartialD; e e}_{S S}}{{&PartialD; &PartialD; θ θ}_{22}} \end{matrix}] - - - - - - ((88))$

记

则Jacobi矩阵J(W)中的元素可由下式加以计算：remember

Then the elements in the Jacobi matrix J(W) can be calculated by the following formula:

$\{\begin{matrix} \frac{{&PartialD; &PartialD; e e}_{s the s}}{{&PartialD; &PartialD; ω ω}_{ji the ji} ((l l))} = = - - {v v}_{j j} {f f}^{' '} (({U u}_{j j})) {x x}_{i i} ((l l)) \\ \frac{{&PartialD; &PartialD; e e}_{s the s}}{{&PartialD; &PartialD; θ θ}_{11}^{j j}} = = - - {v v}_{j j} {f f}^{' '} (({U u}_{j j})) \\ \frac{{&PartialD; &PartialD; e e}_{s the s}}{{&PartialD; &PartialD; v v}_{j j}} = = - - f f (({U u}_{j j})) \\ \frac{{&PartialD; &PartialD; e e}_{s the s}}{{&PartialD; &PartialD; θ θ}_{22}} = = - - 11 \end{matrix} - - - - - - ((99))$

当网络的训练误差满足精度要求或者达到最大迭代次数时候停止网络训练，网络训练结束后采用测试样本对预测模型的预测效果进行测试，如果预测效果满足发动机排气温度预测的实际需求，则保存模型，否则，重复该步骤直到得到满意的模型。When the training error of the network meets the accuracy requirements or reaches the maximum number of iterations, the network training is stopped. After the network training is completed, the test sample is used to test the prediction effect of the prediction model. If the prediction effect meets the actual needs of engine exhaust temperature prediction, the model is saved. , otherwise, repeat this step until a satisfactory model is obtained.

Claims

1. the monitoring method of an aeroengine delivery temperature, it is characterized in that: the method includes the steps of:

Step (1): the temperature transducer through being installed on the aeroengine is gathered the engine exhaust temperature data, and carries out pretreatment and storage through the engine performance Surveillance center that aircraft communication addressing and reporting system are transferred to ground.

Step (2): the engine exhaust temperature historical data of extracting set period from engine performance Surveillance center database; Constitute the delivery temperature time series of one dimension, make up the forecasting model input that is used to set up and test multidimensional training sample set, the test sample book collection of forecasting model and is used to predict on this basis.

Step (3): set up forecasting model based on convolution and departure process neuron network; The said prediction model of training sample set pair that adopts said step (2) to set up is trained; Adopt test sample book set pair forecasting model to test then, confirm suitable prediction model parameters, the storage forecasting model; The concrete grammar of said step (3) is: the convolution and the departure process neural network prediction model that are used for the engine exhaust temperature prediction have an input layer, a hidden layer and an output layer; The delivery temperature historical data that input step (2) generates just can obtain the predicted value of delivery temperature, and its predicted value is determined by following convolution and departure process neural network prediction model:

y = Σ_{j = 1}^{m} v_{j} f (Σ_{i = 1}^{n} ω_{ji} * x_{i} + θ_{1}^{j}) + θ_{2} = Σ_{j = 1}^{m} v_{j} f (Σ_{i = 1}^{n} Σ_{l = 1}^{L} ω_{ji} (l) x_{i} (l) + θ_{1}^{j}) + θ_{2}

Wherein, * representes convolution algorithm, ω _JiFor connecting the weight vector of j neuron of latent layer and i input unit of input layer, x _iFor length is the input vector of L, be i input unit of input layer; v _jFor connecting the weights of j neuron of latent layer and output neuron;

Be j neuronic excitation threshold of latent layer; θ ₂Be the output layer excitation threshold;

Adopt training sample that forecasting model is trained according to the Levenberg-Marquardt training algorithm; And adopt the test sample book in the step (2) that forecasting model is tested; If the test effect meets the demands, carry out next step, otherwise; Repeat step (3) up to obtaining satisfied prediction effect, the forecasting model that storage finally obtains.

Step (4): the model that is used to the predict input sample that adopts the engine exhaust temperature historical data to generate in the input step (2); Use forecasting model that the engine exhaust temperature in following a period of time is predicted; And be applied to the aeroengine condition monitoring; Comprise and judge exhaust gas temperature change trend, whether the monitoring delivery temperature has sudden change.

2. like the monitoring method of right 1 described aeroengine delivery temperature; It is characterized in that: the concrete grammar of said step (2) is: m value predicted m+1 value before adopting the delivery temperature seasonal effect in time series; Use the delivery temperature historical data to constitute the matrix of variables X of m row; Each row of matrix of variables X are regarded as same variable in difference value constantly, thereby the data visualization of each row of matrix of variables X is interior L discrete sampling to strain capacity of time period [0, T]; L is capable before adopting just can the simulation process formula import as input, can make up sample to { d according to above principle _i, x _i, x wherein _iBe the matrix of L * m, the preceding L among the corresponding matrix of variables X is capable, wherein x _iThe corresponding process type input of each row, d _iBeing corresponding target value, is the delivery temperature time series of N for length, and it is right to make up N-m-L+1 sample, according to above method structure training sample set, test sample book collection and the model input that is used to predict.