CN112613227B - Model for predicting remaining service life of aero-engine based on hybrid machine learning - Google Patents

Abstract

The invention belongs to the technical field of aero-engine fault prediction and health management, and discloses a hybrid machine learning-based aero-engine residual service life prediction model, in particular to a hybrid machine learning-based model SGBRT for timely predicting the residual service life of an aero-engine; the model combines a self-organizing mapping network and a gradient lifting regression tree algorithm, and can predict the residual service life of the aircraft engine through the following steps: firstly, the model uses a self-organizing mapping network to cluster an original sample set into clusters; and then respectively constructing a gradient lifting regression tree for each cluster so as to predict the remaining service life of the aircraft engine. The method not only can better predict the residual service life of the aircraft engine, but also reveals the intrinsic characteristics of the degradation data of the aircraft engine.

Description

Model for predicting remaining service life of aero-engine based on hybrid machine learning

Technical Field

The invention belongs to the technical field of aero-engine fault prediction and health management, and particularly relates to a model for predicting the remaining service life of an aero-engine based on hybrid machine learning.

Background

Aircraft engines are one of the most critical parts of an aircraft and are highly complex systems. Aircraft engines typically operate for long periods of time under severe conditions of high temperature, high pressure, high speed and high load, and are therefore prone to failure. Failure of an aircraft engine can have catastrophic consequences and therefore requires very high reliability and safety. Furthermore, the maintenance costs of an aircraft engine are very high. For the management of aircraft engines, airlines are under various pressures, including ensuring the safety and reliability of the engines, avoiding engine failures during operation and reducing the maintenance costs of the engines. Engine fault Prediction and Health Management (PHM) is an effective solution. The PHM technology of the aircraft engine is an important means for promoting the improvement of maintenance modes and improving the safety, reliability and economic bearing capacity of the engine. As a key core technology of PHM, prediction aims to predict (RUL) of a component or system and provide support for operation planning and maintenance decisions.

RUL prediction can predict the time from an operational state to a failure state of a device at a particular time. Existing RUL prediction methods can be divided into two categories: model-based methods and data-driven based methods. Model-based approaches are often more accurate if the degradation of a complex system is modeled accurately. But physical model-based methods require a large amount of a priori knowledge. Recently, data-driven based approaches have received increasing attention. The method based on data driving does not need to know the detailed operation mechanism of a mechanical system, only needs to collect some data from the system, and can identify the condition of the system according to algorithms such as artificial intelligence and the like. There are many methods and models available for data-driven RUL prediction.

In 2015, Nieto et al proposed a RUL prediction model of an aircraft engine based on a support vector machine in Hybrid pseudo-based method for detecting the remaining useful life of the aircraft engine and evaluating the parameters in the support vector machine by using a particle swarm optimization algorithm.

With the advent of large data sets with excellent explicit labels, artificial intelligence methods have been widely used in RUL prediction. Among them, neural networks are one of the most commonly used algorithms.

In 2018, Li et al propose a new method for data prediction using a deep convolutional neural network in "learning using free estimation in modeling using radial convolution neural networks". Experiments were performed on the C-MAPSS data set to demonstrate the effectiveness of the method. However, most deep learning approaches have no effective mechanism to adaptively weight input features when processing multi-feature data. In 2020, Liu et al, in "Remaining using free prediction using a novel feature based end-to-end prediction", proposed a novel end-to-end RUL prediction method based on feature attention. The proposed feature attention mechanism is directly applied to the input data, and can dynamically focus more attention on more important features in the training process.

In RUL prediction, time series analysis is also a common prediction method and is relatively mature. In this predictive approach, the general idea is to predict engine performance and health parameters by single or multiple steps using sensor data as a time series until a set fault threshold is reached.

In 2019, Miao et al propose a dual-task long and short term memory network for the degradation assessment and RUL prediction of an aircraft engine in Joint learning of degradation assessment and a road prediction. This makes the evaluation and prediction results more reliable and accurate, thereby improving reliability and safety of operation and reducing maintenance costs. However, the conventional LSTM network only uses the learned features for regression or classification at the last time step, but actually the learned features at other time steps also contribute to some extent.

In 2020, Chen et al propose an attention-based deep learning framework in Machine learning vision an attention based deep learning approach. Long-short term memory networks are used to learn features from raw sensory data. At the same time, the proposed attention mechanism can learn the importance of features and time steps and assign more weight to more important features. But for aircraft engines, the data set is limited. Therefore, it is necessary to introduce an ensemble learning method.

The ensemble learning method is a very general method, and can be applied to many application programs. In 2019, Li et al propose a prediction method based on ensemble learning in the Degradation modeling and remaining using prediction of aircraft engines to model the Degradation process and predict the RUL of an aircraft engine. The ensemble learning algorithm combines multiple base learners to achieve better prediction performance. And (3) distributing the optimal weight for the basic learner by using a particle swarm optimization algorithm and a sequential quadratic optimization method. In 2017, Zhang et al put forward an integration method of a multi-target deep belief network in Multi object deep belief networks for remaininging user interest in prognosics, which is used for predicting the RUL of an aircraft engine. A multi-objective evolutionary algorithm is applied to train a deep belief network with two conflicting objectives (accuracy and diversity). The trained deep belief networks are then merged to form an integrated model of the final RUL prediction.

The existing research methods are mainly based on supervised learning algorithms. The prediction result is excessively dependent on the history data, and thus it is difficult to ensure accuracy and efficiency of the prediction result. Clustering (self-organizing map networks), an unsupervised learning algorithm, can provide insight into individual relationships that may be missing from the data and identify some more meaningful relationships from the data. Deep learning based approaches are hampered by the lack of data on aircraft engines. Ensemble learning has a great advantage in small sample learning, and therefore, a new ensemble learning method (gradient boosting regression tree) is considered to be used instead of the deep learning method which is widely used at present.

Disclosure of Invention

The invention provides a novel SGBRT (data-driven modeling, fuzzy neural network) based model, which combines a Self-organizing mapping Map (SOM) with a Gradient Boosting Regression Tree (GBRT) to predict the RUL (residual-Boosting Regression Tree) of an aircraft engine. First, the original sample set is clustered into clusters using the SOM network. Each cluster was then individually built into a regression model (GBRT) to predict RUL.

The technical scheme of the invention is as follows:

an aircraft engine RUL prediction model based on hybrid machine learning comprises the following steps:

establishing a hybrid machine learning model combining a Self-organizing mapping network (SOM) and a Gradient Boosting Regression Tree (GBRT) so as to predict the Remaining service Life (RUL) of the aircraft engine; the SOM self-organizes and adaptively changes network parameters and structures by automatically discovering the intrinsic laws and basic attributes in the sample; the hybrid machine learning model is regarded as a modified version of SOM, and data points mapped to neurons after a standard training process are reserved to construct GBRT; establishing a GBRT for each neuron;

the hybrid machine learning model SGBRT is divided into four layers: the system comprises an input layer, an SOM layer, a regression layer and an output layer; in the training stage, inputting the trained feature vectors into an input layer, and then clustering input original data in an SOM layer; in a regression layer, constructing GBRTs for the sub data sets obtained by clustering respectively; in the testing stage, inputting the characteristic vectors of the test set into an input layer, and then judging which type the characteristic vectors belong to in an SOM layer; next, in the regression layer, inputting the feature vectors into the GBRT of the construction class; through the steps, the predicted value is output to an output layer;

(1) after the SOM layer receives the training data, the neurons calculate the distance between the training data and the weight vectors carried by the neurons, and the neurons with the closest distance become competitive winners and are called winning neurons; the weight vectors of the winning neuron and its neighbors will then be adjusted so that the distance of these weight vectors from the current input sample is reduced; this process is iterated continuously until convergence; the method comprises four steps of initialization, a competition process, a cooperation process and weight adjustment;

(1.1) assume that the input layer feature vector x is written as: x ═ X _i I ═ 1., K }, where there are K eigenvectors; the connection weight between input cell i and neuron j is written as m _ji ＝{m _ji :j＝1,.., N; 1., K }, wherein the total number of neurons is N; selecting a smaller random initial value for the network weight;

(1.2) for each input feature vector, the neurons calculating their respective discrimination function values, the winning neuron being the particular neuron having the smallest discrimination function value; the discriminant function is defined as a feature vector x and a weight vector m for each neuron j _ji The squared euclidian distance between, i.e.:

namely, the winning nerve unit is the neuron with the weight vector closest to the characteristic vector;

(1.3) there is a topological neighborhood similar to neurobiology among neurons in the SOM; s _ji And if the lateral distance between the neuron j and the input unit i on the neuron grid is the lateral distance, the topological neighborhood is taken as:

wherein I (x) is the input unit index of the winning neuron; the function is largest among the winning neurons and is symmetric about the winning neurons; when the distance reaches infinity, it decays monotonically to zero; sigma is the effective width of the topological neighborhood, and the size of the topological neighborhood shrinks along with time;

(1.4) the SOM must contain some adaptive or learning process by which the output nodes self-organize to form a feature map between the inputs and outputs; not only can the winning neuron get weight updates, its neighbors will also update their weights; the weight updating method comprises the following steps:

Δm _ji ＝η(t)·T _j,I(x) (t)·(x _i -m _ji )

wherein eta (t) is the learning rate, and t is the cycle number; the update applies to all trained feature vectors X in multiple iterations; the effect of each learning weight update is to win the nervesWeight vector m of element and its neighbors _ji Moving to the feature vector X; iterating through the process may order the topology of the network;

(2) for each clustered cluster in the regression layer, constructing GBRT to predict the RUL of the aircraft engine; GBRT is the Boosting type of the ensemble learning algorithm;

(2.1) the Boosting framework is trained separately using multiple sets of base models, and the results of all base models are linearly combined to obtain a more reliable prediction, as shown in the following equation:

wherein h is _d (x) Represents a base model; d is the number of the basic models, and D is the total number of the basic models; the training target of the overall model is to enable the predicted value F (x) to approach the true value y, enable each basic model to respectively bear department prediction tasks by using the idea of a greedy algorithm, and pay attention to the error generated by each basic model:

F _d (x)＝F ^d-1 (x)+h _d (x)

(2.2) fitting formula of inverse gradient by introducing arbitrary loss function L:

wherein H is the cycle number, and H is the maximum cycle number;

(2.3) GBRT is a continuously developed ensemble learning model, and the basic function of the GBRT is a tree structure; the predicted output using E-tree function accumulation is shown in the formula:

wherein e represents an index of the tree; Γ is the number of leaves on the tree; each F corresponds to an independent tree structure; the values of the leaf node regions are estimated using a linear search, the loss function is minimized, and then the regression tree is updated.

The invention has the beneficial effects that: the present invention proposes a hybrid model, called SGBRT, which combines SOM and GBRT to predict RUL for an aircraft engine. First, the method clusters the original sample set into clusters by using SOM. Then, GBRT was constructed separately for each cluster to predict RUL. Regression analysis based on data partitioning can not only provide more accurate predictions, but can also reveal intrinsic connections between data and extract meaningful information.

Drawings

FIG. 1 is a topology of a SOM in accordance with an embodiment of the present invention;

FIG. 2 is a diagram of a gradient boosting regression tree GBRT algorithm according to an embodiment of the present invention;

fig. 3 is a schematic diagram of a hybrid machine learning model SGBRT algorithm according to an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail with reference to the following embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

The following detailed description of the principles of the invention is provided in connection with the accompanying drawings.

The SOM automatically searches for intrinsic rules and intrinsic attributes in the sample, and changes network parameters and structures in a self-organizing and self-adaptive manner. It is proposed herein that the hybrid model is considered an improved version of the SOM, which is a standard training process that combines the GBRT regression model with the SOM more deeply. The difference from the traditional SOM method is that the data mapped to neurons is preserved to build the GBRT regression model. Each neuron builds a GBRT regression model. The system architecture is shown in fig. 3.

The proposed hybrid model SGBRT is divided into four layers: the system comprises an input layer, an SOM layer, a regression layer and an output layer; in the training stage, inputting the trained feature vectors into an input layer, and then clustering input original data in an SOM layer; in a regression layer, constructing sub data sets obtained by clustering into GBRTs respectively; in the testing stage, inputting the characteristic vectors of the test set into an input layer, and then judging which type the characteristic vectors belong to in an SOM layer; next, in the regression layer, inputting the feature vector into the GBRT of the construction class; through the steps, the predicted value is output to an output layer;

(1) after the SOM layer receives the training data, the neurons calculate the distance between the training data and the weight vectors carried by the neurons, and the neurons with the closest distance become competitive winners and are called winning neurons; the weight vectors of the winning neuron and its neighboring neurons will then be adjusted such that the distance of these weight vectors from the current input sample is reduced; this process is iterated continuously until convergence; the topological structure of the SOM is shown in figure 1 and comprises four steps of initialization, a competition process, a cooperation process and weight adjustment;

(1.1) assume that the input layer feature vector x is written as: x ═ X _i I 1.., K }, where there are K eigenvectors; the connection weight between input cell i and neuron j is written as m _ji ＝{m _ji J ═ 1., N; 1.., K }, wherein the total number of neurons is N; selecting a smaller random initial value for the network weight;

(1.2) for each input feature vector, the neurons calculating their respective discrimination function values, the winning neuron being the particular neuron having the smallest discrimination function value; the discriminant function is defined as a feature vector x and a weight vector m for each neuron j _ji The squared euclidian distance between, i.e.:

namely the winning nerve unit is the neuron with the weight vector closest to the characteristic vector;

(1.3) there is a topological neighborhood similar to neurobiology among neurons in the SOM; s. the _ji And if the lateral distance between the neuron j and the input unit i on the neuron grid is the lateral distance, the topological neighborhood is taken as:

wherein I (x) is the input unit index of the winning neuron; the function is largest among the winning neurons and is symmetric about the winning neurons; when the distance reaches infinity, it decays monotonically to zero; sigma is the effective width of the topological neighborhood, and the size of the topological neighborhood shrinks along with time;

(1.4) the SOM must contain some adaptive or learning process by which the output nodes self-organize to form a feature map between the inputs and outputs; not only can the winning neuron get weight updates, its neighbors will also update their weights; the weight updating method comprises the following steps:

Δm _ji ＝η(t)·T _j,I(x) (t)·(x _i -m _ji )

wherein eta (t) is the learning rate, and t is the cycle number; the update applies to all trained feature vectors X in multiple iterations; the effect of each learning weight update is to update the weight vector m of the winning neuron and its neighbors _ji Moving to the feature vector X; iterating through the process may order the topology of the network;

(2) for each clustered cluster in the regression layer, constructing GBRT to predict the RUL of the aircraft engine; GBRT is the Boosting type of the ensemble learning algorithm, and the training process thereof is shown in FIG. 1;

(2.1) the Boosting framework is trained separately using multiple sets of base models, and the results of all base models are linearly combined to obtain a more reliable prediction, as shown in the following equation:

wherein h is _d (x) Representing a base model; d is the number of the basic models, and D is the total number of the basic models; the training goal of the ensemble model is to approximate the predicted values F (x) to the true values y, each elementary using the idea of a greedy algorithmThe models respectively undertake department prediction tasks and pay attention to errors generated by each basic model:

F _d (x)＝F ^d-1 (x)+h _d (x)

(2.2) fitting formula of inverse gradient by introducing arbitrary loss function L:

wherein H is the cycle number, and H is the maximum cycle number;

(2.3) GBRT is a continuously developed ensemble learning model, and the basic function of the GBRT is a tree structure; the predicted output using E-tree function accumulation is shown in the formula:

wherein e represents an index of the tree; Γ is the number of leaves on the tree; each F corresponds to an independent tree structure; the values of the leaf node regions are estimated using a linear search, the loss function is minimized, and then the regression tree is updated.

In summary, the following steps:

in order to accurately predict the RUL of an aircraft engine, a hybrid ensemble learning based model SGBRT is proposed in the present invention that combines SOM with GBRT to predict the RUL of an aircraft engine. First, the method clusters the original sample set into clusters using the SOM network. GBRT is then constructed separately for each cluster to predict the RUL of the aircraft engine. The proposed hybrid machine learning-based model not only can better predict the RUL of an aircraft engine, but also reveals the intrinsic characteristics of aircraft engine degradation data.

The foregoing shows and describes the general principles and broad features of the present invention and advantages thereof. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, which are described in the specification and illustrated only to illustrate the principle of the present invention, but that various changes and modifications may be made therein without departing from the spirit and scope of the present invention, which fall within the scope of the invention as claimed. The scope of the invention is defined by the appended claims and equivalents thereof.

Claims (1)

1. A model for predicting the remaining service life of an aircraft engine based on hybrid machine learning is characterized by comprising the following steps:

establishing a hybrid machine learning model combining a self-organizing mapping network and a gradient lifting regression tree to predict the residual service life of the aircraft engine; the SOM self-organizes and adaptively changes network parameters and structure by automatically discovering the intrinsic laws and basic properties in the sample; the hybrid machine learning model is regarded as a modified version of SOM, and data points mapped to neurons after a standard training process are reserved to construct GBRT; establishing a GBRT for each neuron;

the hybrid machine learning model SGBRT is divided into four layers: the system comprises an input layer, an SOM layer, a regression layer and an output layer; in the training stage, inputting the trained feature vectors into an input layer, and then clustering input original data in an SOM layer; in a regression layer, constructing sub data sets obtained by clustering into GBRTs respectively; in the testing stage, inputting the characteristic vectors of the test set into an input layer, and then judging which type the characteristic vectors belong to in an SOM layer; next, in the regression layer, inputting the feature vector into the GBRT of the construction class; through the steps, the predicted value is output to an output layer;

(1) after the SOM layer receives the training data, the neurons calculate the distance between the training data and the weight vectors carried by the neurons, and the neurons with the closest distance become competitive winners and are called winning neurons; the weight vectors of the winning neuron and its neighboring neurons will then be adjusted such that the distance of these weight vectors from the current input sample is reduced; this process is iterated continuously until convergence; the method comprises four steps of initialization, a competition process, a cooperation process and weight adjustment;

(1.1) assume that the input layer feature vector x is written as: x ═ X _i I 1.., K }, where there are K eigenvectors; the connection weight between input cell i and neuron j is written asm _ji ＝{m _ji J ═ 1.., N; 1.., K }, wherein the total number of neurons is N; selecting a smaller random initial value for the network weight;

(1.2) for each input feature vector, the neurons calculating their respective discrimination function values, the winning neuron being the particular neuron having the smallest discrimination function value; the discriminant function is defined as a feature vector x and a weight vector m for each neuron j _ji The squared euclidian distance between, i.e.:

namely the winning nerve unit is the neuron with the weight vector closest to the characteristic vector;

(1.3) there is a topological neighborhood similar to neurobiology among neurons in the SOM; s _ji And if the lateral distance between the neuron j and the input unit i on the neuron grid is the lateral distance, the topological neighborhood is taken as:

wherein I (x) is the input unit index of the winning neuron; the function is largest among the winning neurons and is symmetric about the winning neurons; when the distance reaches infinity, it decays monotonically to zero; sigma is the effective width of the topological neighborhood, and the size of the topological neighborhood shrinks along with time;

(1.4) the SOM must contain some adaptive or learning process by which the output nodes self-organize to form a feature map between the inputs and outputs; not only can the winning neuron get weight updates, its neighbors will also update their weights; the weight updating method comprises the following steps:

Δm _ji ＝η(t)·T _j,I(x) (t)·(x _i -m _ji )

wherein eta (t) is the learning rate, and t is the cycle number; the update applies to all trained feature directions in multiple iterationsAn amount X; the effect of each learning weight update is to update the weight vector m of the winning neuron and its neighbors _ji Moving to the feature vector X; iterating through the process may order the topology of the network;

(2) for each clustered cluster in the regression layer, constructing GBRT to predict the RUL of the aircraft engine; GBRT is the Boosting type of the ensemble learning algorithm;

(2.1) the Boosting framework is trained separately using multiple sets of base models, and the results of all base models are linearly combined to obtain a more reliable prediction, as shown in the following equation:

wherein h is _d (x) Represents a base model; d is the number of the basic models, and D is the total number of the basic models; the training target of the overall model is to enable the predicted value F (x) to approach the true value y, enable each basic model to respectively undertake department prediction tasks by using the idea of a greedy algorithm, and pay attention to errors generated by each basic model:

F _d (x)＝F ^d-1 (x)+h _d (x)

(2.2) fitting formula of inverse gradient by introducing arbitrary loss function L:

wherein H is the cycle number, and H is the maximum cycle number;

(2.3) GBRT is a continuously developing integrated learning model, and the basic function of the GBRT uses a tree structure; the predicted output using E-tree function accumulation is shown in the formula:

wherein e represents an index of the tree; Γ is the number of leaves on the tree; each F corresponds to an independent tree structure; the values of the leaf node regions are estimated using a linear search, the loss function is minimized, and then the regression tree is updated.

Images