CN116028865A

CN116028865A - Aeroengine fault diagnosis method based on feature amplification

Info

Publication number: CN116028865A
Application number: CN202310018871.5A
Authority: CN
Inventors: 林琳; 何文辉; 付松; 童昌圣; 祖立争
Original assignee: Harbin Institute of Technology
Current assignee: Harbin Institute of Technology
Priority date: 2023-01-06
Filing date: 2023-01-06
Publication date: 2023-04-28

Abstract

An aeroengine fault diagnosis method based on feature amplification belongs to the technical field of aeroengine fault diagnosis. Aiming at the problems that the significance of the fault signal characteristics of the aero-engine is gradually reduced, and useful information in actual operation and maintenance data of the engine is difficult to fully extract, so that the fault diagnosis accuracy is affected. Performing high-dimensional feature amplification on an original sample to obtain a sample after feature amplification; carrying out normalization processing, and constructing a training sample set by the normalized samples; respectively setting different labels for a normal state sample and a fault sample in a training sample set; training the fault diagnosis network by adopting a training sample set, and obtaining the trained fault diagnosis network after reaching the preset iteration times; acquiring operation data of the aeroengine, and processing the operation data to obtain normalized data to be diagnosed; and inputting the normalized data to be diagnosed into a fault diagnosis network after training to obtain an aeroengine fault diagnosis result. The method is used for fault diagnosis of the aero-engine.

Description

Aeroengine fault diagnosis method based on feature amplification

Technical Field

The invention relates to an aeroengine fault diagnosis method based on feature amplification, and belongs to the technical field of aeroengine fault diagnosis.

Background

Aeroengines are the "heart" of an aircraft, as a primary source of power and air entraining devices, the health of which is critical to the safety and economy of the aircraft's flight. With the reduction of the number of accidents, aeroengines have higher reliability and lower failure rates. However, due to the complex structure of the aero-engine, the working environment is severe, and the problem of faults in working is unavoidable. And because of the characteristics of multiple types, complex internal rules, high diagnosis difficulty and the like of the aero-engine, the fault diagnosis of the aero-engine is always a challenge.

In order to accurately diagnose faults of the aero-engine, the air path fault diagnosis methods of the engines by various airlines and scientific research institutions are intensively studied, and the methods can be mainly classified into a model-based method and a data-driven method at present. The model-based method can detect sudden faults relatively accurately, experience and historical data are not needed, and the detection result has interpretability. However, due to the complexity of engine systems, the difficulty of building accurate models becomes greater and greater; secondly, a great deal of priori knowledge is required for establishing an aircraft engine diagnosis model, and the structure and system design parameters of the engine are required to be accurately mastered, and are always kept secret, so that the difficulty for establishing an accurate engine fault diagnosis model by an airline company is great. Along with the gradual application of digital, networked and intelligent technologies in the industrial field, industrial data with more comprehensive types and huge quantity is more and more easily collected and stored, a foundation is laid for a data-driven fault diagnosis method, and more data-driven methods are applied to fault diagnosis at present. Compared with the method based on the model, the method based on the data driving has the advantages of no need of establishing a mathematical physical model of the engine, no influence of complexity of the engine, no need of a great deal of priori knowledge and experience, and good nonlinear processing capability. Therefore, the data-driven fault diagnosis method based on the monitoring data completely is more practical in the field of aeroengine fault diagnosis. With the rapid development of machine learning methods, particularly deep learning methods, data-driven-based methods have gradually become the mainstream of aeroengine fault diagnosis.

The data-driven method has high requirements on data samples, and the ideal state is that the data samples have sufficient samples and enough labels, so that the data-driven method has good separability; the actual running data of the engine often has the characteristics of time sequence, nonlinearity, difficulty in dividing and the like. Some deep learning algorithms currently have better performance in this field: the method comprises a convolutional neural network, a long-term and short-term memory network, a deep confidence network, a YOLOv3, a pattern recognition algorithm and the like, which can well mine abstract features in time sequence signals and can improve the accuracy of fault recognition. However, as the reliability of the aero-engine increases, the feature significance of the fault signal decreases gradually, and it is important how to make full use of the useful information in the actual operation and maintenance data of the engine to improve the accuracy of fault diagnosis.

Disclosure of Invention

Aiming at the problems that the significance of the fault signal characteristics of the aero-engine is gradually reduced, and useful information in actual operation and maintenance data of the engine is difficult to fully extract, and further the fault diagnosis accuracy is affected, the invention provides a fault diagnosis method of the aero-engine based on characteristic amplification.

The invention relates to an aeroengine fault diagnosis method based on feature amplification, which comprises the following steps of,

step one: performing high-dimensional feature amplification on the original sample to obtain a sample after feature amplification;

step two: carrying out normalization treatment on the samples after feature amplification, and constructing a training sample set by the normalized samples; respectively setting different labels for a normal state sample and a fault sample in a training sample set;

step three: training the fault diagnosis network by adopting a training sample set, and obtaining the trained fault diagnosis network after reaching the preset iteration times;

step four: collecting operation data of the aeroengine, and performing high-dimensional feature amplification in the first step and normalization in the second step to obtain normalized data to be diagnosed; and inputting the normalized data to be diagnosed into a fault diagnosis network after training to obtain an aeroengine fault diagnosis result.

According to the aeroengine fault diagnosis method based on feature amplification, in the first step, an original sample is extracted from engine gas path performance monitoring data.

According to the aeroengine fault diagnosis method based on feature amplification, the method for performing high-dimensional feature amplification on the original sample in the first step comprises a high-dimensional mapping method based on a polynomial kernel explicit mapping function, and specifically comprises the following steps:

using polynomial kernel function K _p For the original sample x _i And original sample x _j Performing inner product operation, x _i ≠x _j ：

Phi in _p Representing a mapping function corresponding to the polynomial kernel function, wherein r and d are undetermined parameters in the polynomial kernel function respectively;

wherein the ith original sample x _i ＝[x _i,1 ,x _i,2 ,…,x _i,n ]The jth original sample x _j ＝[x _j,1 ,x _j,2 ,…,x _j,n ]N represents the dimension of the samples, the total number of original samples being N, i=1, 2,3, … … N, j=1, 2,3, … … N;

φ _p (x _i ) Is of the dimension of

To reduce phi _p (x _i ) And preserve the original sample x _i Selecting r=1, d=2, then Φ _p (x _i ) The expression of (2) is:

will phi _p (x _i ) Constant term 1 and coefficient in the expression of (2)

Discard, marked as phi _p ′(x _i )：

Will phi _p ′(x _i ) As a post-feature amplification sample.

According to the aeroengine fault diagnosis method based on feature amplification, the method for performing high-dimensional feature amplification on an original sample in the first step comprises a high-dimensional space modeling method based on experience, and specifically comprises the following steps:

extracting the previous m of the failure of the failed engine k from the OEM data according to the failure time of the failed engine k _k Exhaust temperature deviation value DEGT, high-pressure rotor rotating speed deviation value DN2, fuel flow deviation value DFF and exhaust temperature margin variation EGTM of each flight cycle to obtain an original sample set A _k ：

In the I-th original sample x _I The method comprises the following steps:

x _I ＝[DEGT _I ,DN2 _I ,DFF _I ,EGTM _I ]，I＝1,2,3,……，m _k ；m _k is an integer greater than 10;

the long-term memory network LSTM is adopted for the original sample x _I Performing smoothing denoising treatment to obtain smoothed sample data; carrying out K-step differential calculation on the smoothed sample data, wherein the value of K is 9, and obtaining a sample set A after characteristic amplification _k ′：

Where DEGT 'represents the value of DEGT after being denoised by LSTM smoothing, DN2' represents the value of DN2 after being denoised by LSTM smoothing, DFF 'represents the value of DFF after being denoised by LSTM smoothing, EGTM' represents the value of EGTM after being denoised by LSTM smoothing.

According to the aeroengine fault diagnosis method based on feature amplification, the normalized sample is also used for constructing a test sample set, and the distribution ratio of the training sample set to the test sample set is 4:1.

According to the aeroengine fault diagnosis method based on feature amplification, sample data input by a fault diagnosis network each time are selected in a training sample set through a sliding window.

According to the aeroengine fault diagnosis method based on feature amplification, the fault diagnosis network is a convolutional neural network CNN, a long and short term memory network LSTM, a time convolutional network TCN or a depth residual error contraction network DRSN-CW.

According to the aeroengine fault diagnosis method based on feature amplification, the learning rate of the fault diagnosis network is set to be 0.001, an Adam optimizer is adopted for updating weights, and L2 regularization is used for reducing overfitting; the attenuation coefficient of L2 regularization was 0.0001, batch size 10.

The invention has the beneficial effects that: the method improves the classification effect of fault diagnosis by carrying out characteristic amplification on the original data. One is based on polynomial kernel explicit mapping function, the method does not need to carry out very complex operation, and each dimension has obvious mapping relation with the original dimension after high-dimensional mapping. The other is a high-dimensional space modeling method based on experience, which firstly adopts LSTM to carry out smooth denoising treatment on the original data, and then takes the data after difference as new amplified characteristics. The fault diagnosis method fully utilizes the useful information in the actual operation and maintenance data of the engine, thereby improving the accuracy of fault diagnosis.

To illustrate the effectiveness of the method of the present invention. And the engine data set and the public data set are respectively subjected to experimental verification, and two comparison methods for high-dimensional mapping by machine learning are added. Experimental results show that the Scheme1 and the Scheme2 obtain better classification results than the original data in a plurality of evaluation indexes, and are superior to two machine learning methods. On the engine dataset, three metrics of both Scheme1 and Scheme2 in the TCN deep learning model were greater than 90%. On a common dataset, three metrics of both Scheme1 and Scheme2 in TCN deep learning model were greater than 97%. Even the Accuracy of the Scheme1 in the TCN deep learning model reaches 99.5%. The Accuracy of the Scheme2 in the TCN deep learning model reaches 98.3%. In addition, according to the application of the minimum norm method in two data, the data can be found to have certain orthogonality after high-dimensional mapping if the orthogonality of the original data is strong. This shows that the data still has certain orthogonality after high-dimensional mapping, and the classification result of fault diagnosis can be effectively improved.

Drawings

FIG. 1 is a schematic diagram of an aircraft engine gas path parameter acquisition and conversion process; n in the figure ₁ Equivalent to N1, N in the specification ₂ Equivalent to N2 in the specification;

FIG. 2 is an example CNR report of a typical gas path failure;

FIG. 3 is a schematic diagram of the trend of variation of the DEGT when an EGT indication fault occurs in an aero-engine; flight cycle in the figure represents the Flight cycle;

FIG. 4 is a schematic diagram of the trend of EGTM change when an EGT indication fault occurs in an aero-engine;

fig. 5 is a schematic diagram of a trend of change in DN2 when an EGT indication fault occurs in an aero-engine;

FIG. 6 is a graph showing the trend of DFF change when an EGT indication fault occurs in an aircraft engine;

FIG. 7 is a general frame diagram of an aircraft engine fault diagnosis method based on feature augmentation according to the present invention;

FIG. 8 is a schematic diagram of an LSTM cell structure;

FIG. 9 is a schematic diagram of an engine fault gas circuit parameter fit;

FIG. 10 is a schematic drawing of a sliding window extraction sample;

FIG. 11 is a flow chart of an engine gas circuit fault diagnosis of the method of the present invention;

FIG. 12 is a bar graph corresponding to Accuracy (%) of Table 3;

fig. 13 is a bar graph corresponding to Precision (%) of table 3;

FIG. 14 is a bar graph corresponding to F1-score (%) of Table 3;

FIG. 15 is a bar graph corresponding to the calculation time(s) of Table 3; calculating time in the figure represents calculation time;

FIG. 16 is a bar graph corresponding to Accuracy (%) of Table 4;

fig. 17 is a bar graph corresponding to Precision (%) of table 4;

FIG. 18 is a bar graph corresponding to F1-score (%) of Table 4;

FIG. 19 is a bar graph corresponding to the calculation time(s) of Table 4;

FIG. 20 is a bar graph corresponding to Accuracy (%) of Table 5;

fig. 21 is a bar graph corresponding to Precision (%) of table 5;

FIG. 22 is a bar graph corresponding to F1-score (%) of Table 5;

FIG. 23 is a bar graph corresponding to the calculation time(s) of Table 5.

Detailed Description

The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

It should be noted that, without conflict, the embodiments of the present invention and features of the embodiments may be combined with each other.

The invention is further described below with reference to the drawings and specific examples, which are not intended to be limiting.

The invention provides an aeroengine fault diagnosis method based on feature amplification, which is shown in the accompanying drawings from 1 to 7,

The first step is that the original sample is extracted from the monitoring data of the engine gas path performance.

The civil aviation engine studied in this embodiment mainly consists of 6 main components such as a FAN (FAN), a low-pressure compressor (Low Pressure Compressor, LPC), a high-pressure compressor (High Pressure Compressor, HPC), a Combustor (CC), a low-pressure turbine (Low Pressure Turbine, LPT), and a high-pressure turbine (High Pressure Turbine, HPT), as shown in fig. 1, and Other data are represented by Other in fig. 1. The ground monitoring system monitors the air path performance of the aeroengine mainly by monitoring parameters such as exhaust temperature EGT, fuel flow FF, low-pressure rotor rotating speed N1, high-pressure rotor rotating speed N2 and the like. In the actual process, an original manufacturer (Original Equipment Manufacturer, OEM) of the engine will use the thrust set point, environmental factors and working condition information to convert the originally monitored gas path parameters into a baseline deviation value to monitor the performance state of the engine, namely: OEM data.

According to the customer notification report (Customer Notification Report, CNR) fed back to the airline company by the OEM and the combined literature, the OEM manufacturer monitors the engine air path performance mainly by using four air path parameters, namely, an exhaust temperature deviation value DEGT, an exhaust temperature margin variation (Exhaust Gas Temperature Margin, EGTM), a high-pressure rotor Speed deviation value (Delta Core Speed, DN 2), a Fuel Flow deviation value (Delta Fuel Flow, DFF), and the like. CNR records the abnormal time and possible cause of civil aviation engine, and OEM judges the abnormal time of engine to be issued to the airline company through analysis of own case base and knowledge base. CNR has a certain time lag, cannot be timely fed back to the airlines, and also generates service charges in the process. Therefore, it is important that the airline company can make a judgment on the health state of the engine according to the air path parameters in time.

Fig. 2 is an example of CNR reporting for a typical gas path failure. Fig. 2 is only schematic, and thus a detailed description of specific english in the drawing is not necessary. From fig. 2, it can be found that the relevant information such as the model number, the arrangement position and the report of the engine is indicated in the box 1, so that the subsequent sorting work is facilitated; in block 2, the change in response parameters upon failure is indicated; in block 3, specifying the fault type of the OEM based on the diagnostic knowledge base analysis; in the box 4, the starting and ending time point of abnormal change of the state parameter is indicated, so that the client can find the specific parameter value of the corresponding date in the database to perform more visual judgment, and meanwhile, the box 4 is also the key data for collecting the fault sample in the embodiment.

According to the engine maintenance report and the CNR report, the air circuit faults of the aeroengine collected at present comprise: EGT indicates a fault, TAT indicates a fault, and high pressure turbine blade erosion faults. According to the analysis of the fault data corresponding to the three types of fault types, when the engine has a gas circuit fault, the DEGT, EGTM, DFF gas circuit parameters, DN2 gas circuit parameters and other 4 gas circuit parameters can change to a certain extent.For further explanation, one of the fault types is selected as an example, and the changes of the four gas path parameters when the fault occurs are analyzed. Fig. 3 to 6 show the variation of four gas path parameters when an EGT indication fault occurs in an aeroengine, wherein T ₁ The moment of the actual abnormality of the engine is indicated; t (T) ₂ Indicating the moment at which the engine is actually diagnosed as faulty; the point A is the value of each gas path parameter when the engine starts to be abnormal; the point B is the value of each gas path parameter when the engine is diagnosed as a fault; delta T represents T ₁ Time and T ₂ Number of cycles of flight cycle at intervals between moments. From the figure, it can be seen that from T ₁ From moment to T ₂ At the moment, when the engine fails, four gas path parameters change to a certain extent, such as: the tendency of DEGT decreases, the tendency of EGTM increases, and the tendency of DN2 and DFF increases before decreasing.

As can be seen from the analysis of fig. 3 to 6, when the air circuit of the aeroengine fails, the failure type and the change of the 4 air circuit parameters such as DEGT, EGTM, DFF and DN2 generate a certain association relation. Therefore, the change of four gas path parameters can be utilized to carry out gas path multi-fault classification.

Fig. 7 is a schematic diagram of the method of the present invention, consisting essentially of three parts: (1) Performing high-dimensional mapping on the original fault sample by using two characteristic amplification technologies; (2) Analyzing the orthogonality strength of the fault sample subjected to feature amplification in a theoretical layer; (3) And (5) training the deep learning model by utilizing the fault sample set after feature amplification to realize fault diagnosis. The detailed description is as follows: firstly, for obtaining the actual normal and fault data of the engine, considering the problem of high nonlinearity difficulty of the original monitoring data, two feature amplification technologies (a polynomial kernel explicit mapping method and an empirical method) proposed by the embodiment are utilized to map the original sample of the engine into a high-dimensional space so as to promote the separability of the original sample. And then, judging whether the orthogonality of the data after the high-dimensional mapping is weakened or not by a minimum relative norm method through the sample data after the feature amplification. The two amplification methods provided by the embodiment are explained from the theoretical level, so that the orthogonality of each dimension of the data after feature amplification is strong, and each dimension has strong independent information features. And finally, carrying out normalization processing on the sample data after feature amplification, taking the sample data as the input of a deep learning model, carrying out fault diagnosis, and judging whether the input data is normal data or certain type of fault data. And finally, inputting the sample data subjected to the feature amplification and the original data into the same deep learning model in an experimental part, and reflecting the effectiveness of the feature amplification method provided by the method by using a plurality of evaluation indexes.

Further, the method for performing high-dimensional feature amplification on the original sample in the first step includes two methods:

the first is a high-dimensional mapping method based on a polynomial kernel explicit mapping function, wherein the polynomial kernel function refers to a kernel function expressed in a polynomial form. The core idea is to map sample data in high dimensions so that data that is otherwise linearly inseparable is linearly separable. The method comprises the following steps:

φ _p (x _i ) Is of the dimension of

according to the pair phi _p (x _i ) Phi is observed by _p (x _i ) One of the constants is 1, and the constant term does not contain useful information, so that the constant term can be omitted. Coefficients when normalizing samples

The presence or absence of (1) does not affect the processed data, and therefore the rejection coefficient +.>

Obtaining phi' _p (x _i )，φ′ _p (x _i ) Compared with phi _p (x _i ) It is more concise:

phi 'is set' _p (x _i ) As a post-feature amplification sample.

To phi _p (x _i ) Phi 'obtained by modification' _p (x _i ) Substantially reserve phi _p (x _i ) So still phi 'is taken as the mapping form of' _p (x _i ) Is seen as a form of polynomial core explicit mapping function. The polynomial kernel explicit mapping function mentioned later in this embodiment is specifically referred to as phi _p ′(x _i ). It is known from analysis of the gaussian kernel explicit mapping function and the polynomial kernel explicit mapping function that the polynomial kernel explicit mapping function is composed of the original data, the square of the original data, and the mutual dot product of a part of the original data. The polynomial kernel explicit mapping space is based on the original space to amplify features, and is different from the structure of Gaussian kernel explicit mapping space.

The following minimum relative norm method is used to prove whether the orthogonality of each feature of the sample after feature amplification is weakened in a high-dimensional mapping space:

let the matrix form of the original sample be a: a= (DEGT, DN2, DFF, EGTM).

The DEGT, DN2, DFF, and EGTM are column vectors. Orthogonalizing A by column according to Gram-Schmit orthogonalization method. The orthogonalization formula is expressed as:

DEGT″＝DEGT

replacing corresponding variables with x and y, wherein < x, y > represents the inner product of x and y; the DEGT ', DN2', DFF ', EGTM' are orthogonalized column vectors and intersect one another. And then a matrix b= [ DEGT ", DN 2", DFF ", EGTM" ] is obtained.

The 1-norms of the column vectors of matrices a and B are found respectively, component vector C = [ | DEGT|| ₁ ,||DN2|| ₁ ,||DFF|| ₁ ,||EGTM|| ₁ ]Sum vector d= [ DEGT' | ₁ ,||DN2″|| ₁ ,||DFF″|| ₁ ,||EGTM″|| ₁ ]。||.|| ₁ Representing the 1-norm of the vector. Calculation of K _f Is the value of (1): k (K) _f The method is characterized by comprising the following steps of (1) evaluating the orthogonality strength index among the dimensions of data:

defining a lower threshold K ₁ And an upper threshold value K ₂ ，K ₁ ＜K ₂ . If K _f ≤K ₁ Then there is serious complex linearity among the column vectors of the matrix a, i.e. orthogonality among the column vectors is weak; if K ₁ ≤K _f ≤K ₂ Combining other discriminant criteria, and further analyzing; if K _f ≥K ₂ There is some orthogonality between the column vectors.

K ₁ 、K ₂ 、K _f The values of (2) are shown in Table 1.

Table 1 table of minimum relative norm values

From Table 1, it can be seen that K _f >K ₂ It is explained that certain orthogonality exists between the original sample data column vectors, namely, each dimension of the original data has better independent information.

Calculating K of sample after feature amplification mapped by Gaussian kernel explicit mapping function by using minimum relative norm method _f Obtaining K _f Has a value of 1.3X10 ^-4 Less than the lower threshold K ₁ . It is explained that the orthogonality of the features in the gaussian kernel explicit mapping space becomes weaker, which weakens the strength of the individual features to some extent. It is explained that gaussian kernel explicit mapping functions are not suitable for high-dimensional mapping of engine data.

Calculating K of sample after feature amplification obtained after high-dimensional mapping of polynomial core explicit mapping function _f Calculated K _f Is 0.12, greater than the upper threshold K ₂ The method has the advantages that the polynomial kernel explicit mapping function is adopted, and certain orthogonality exists among all features after high-dimensional mapping, so that useful information in an original feature space can be utilized in a subsequent classification process after high-dimensional mapping of data.

The second method is a high-dimensional space modeling method based on experience, after the polynomial kernel and Gaussian kernel explicit mapping function are analyzed, the dimension of the original sample can be analyzed first, and if the dimension of the original sample has certain orthogonality, the original sample plus amplified new features can be adopted to form a high-dimensional space so as to improve the accuracy of fault diagnosis. Thus, this embodiment proposes a new feature augmentation method, which is called an empirically based high-dimensional spatial modeling method. As can be seen from the analysis of fig. 3 to 6, when the air circuit fault occurs in the aero-engine, the fault type and the trend changes of the 4 air circuit parameters such as DEGT, EGTM, DFF, DN2 and the like generate a certain association relation. Thus, the trend change of the 4 gas path parameters can be considered as a new feature of the amplification. Considering that the original data varies drastically, the differential term of the original data is selected as a possible amplification feature in this embodiment. If the original data is directly differenced, errors may exist. Since LSTM has a very good smooth denoising effect. Therefore, the original data is subjected to smoothing denoising processing, and a difference term is constructed in the smoothed denoising data, and then a high-dimensional space is formed together with the original dimension.

The method comprises the following steps:

obtaining the failure time of the failed engine k through the CNR report and the maintenance report, and extracting the previous m of the failure of the failed engine k from OEM data according to the failure time of the failed engine k _k Exhaust temperature deviation value DEGT, high-pressure rotor rotating speed deviation value DN2, fuel flow deviation value DFF and exhaust temperature margin variation EGTM of each flight cycle to obtain an original sample set A _k ：

In the I-th original sample x _I The method comprises the following steps:

the long-term memory network LSTM is adopted for the original sample x _I Performing smoothing denoising treatment to obtain smoothed sample data; the LSTM deep learning network can extract effective time sequence information from a long-span time sequence, and can be better suitable for processing time sequence information. There are 3 gates within the LSTM cell: forget gate, input gate and output gate. Forgetting gate and input gate control neuronal state C _t The output gate is formed by integrating the input h _t-1 Current input x _t Neuron State C _t Co-determined knotAs a result, the LSTM cell structure is schematically shown in FIG. 8. FIG. 8 is a graph illustrating a smoothing of fault gas path parameters for an engine. F in FIG. 8 _t Outputting a result for the forget gate; sigma represents a sigmoid function; i.e _t Outputting a result for the input gate; tanh represents a tanh function; o (O) _t The result is output for the output gate.

Because individual differences exist in the civil aviation engines in actual operation and maintenance, in order to weaken the influence of the individual differences on the smooth denoising of the gas path parameters, the independent gas path parameters of a single engine are subjected to smooth denoising treatment. It can be seen from an examination of fig. 9 that the fluctuation width of the smoothed value is smaller than that of the original value.

And carrying out K-step difference calculation on the smoothed sample data, wherein the difference can play a role of a gentle time sequence. But inevitably causes a certain degree of information loss in the fitting and differential calculation of the data. The present embodiment therefore retains the original data and uses the differentiated data as a new feature of the amplification. According to analysis of CNR reports of OEM manufacturers, 10 cycles can be used as interval segments of fault indication data to meet most fault diagnosis requirements. So that the value of K is selected to be 9, and a sample set A after the characteristic amplification is obtained _k ′：

Sample set A after feature amplification _k ' on the basis of preserving the original data, the original data is amplified from 4 dimensions to 8 dimensions, and the low-dimensional data is mapped into the high-dimensional space. To illustrate the orthogonality between features after high-dimensional mapping, a minimum relative norm method is still used for verification. By calculation of K _f Is 0.42, greater than the upper threshold K ₂ The method for modeling the high-dimensional mapping space based on experience is described to enable each feature to have information independence。

Still further, the normalized sample is also used to construct a test sample set, and the distribution ratio of the training sample set to the test sample set is 4:1.

Still further, sample data input by the fault diagnosis network each time is selected in the training sample set through a sliding window. From analysis of the engine dataset and the common dataset, both datasets are typically multi-dimensional time series data. In order to better capture the sequence characteristics, the present embodiment adopts a sliding window method to extract samples of each input of the fault diagnosis network, as shown in fig. 10. 4 is the dimension of the engine dataset and n represents the dimension after the high-dimensional mapping. When using polynomial kernel explicit mapping functions, n=14; when empirically, n=8; let the length of the sliding window be l and the step size of the sliding window be S. When the start point of the sliding window reaches column q, the start point is pi and the end point is pi+l-1. After the data segment of the current window is obtained, the sliding window is advanced by step S. At this time, the start point and the end point of the sliding window become pi+s and pi+l+s-1. The window is continuously slid forward in the time series data, and a data segment is continuously generated. In this embodiment, l is 10 and s is 5.

As examples, the fault diagnosis network is a convolutional neural network CNN, a long short-term memory network LSTM, a time convolutional network TCN, or a depth residual contraction network DRSN-CW. The four fault diagnosis network structure parameters are shown in table 2. Abbreviations in table 2 are as follows, batch normalization (Batch Normalization, BN), modified linear units (Rectifier linear unit, relu), global average pooling (Global Average Pooling, GAP), full connected layer (Fully Connected layer, FC). 2/4 in Table 2 indicates that the value is 2 when it is classified into two categories; when the number is multiple, the value is 4. The learning rate of the four network models was set to 0.001. Adam optimizers are employed for updating weights when training the four network models, and L2 regularization is used to reduce overfitting. The attenuation coefficient of L2 regularization was 0.0001, batch size 10.

TABLE 2 deep learning model parameters

The fault diagnosis process and the implementation steps of the invention are as follows:

referring to fig. 11, the method of the present invention comprises the following steps:

firstly, mapping engine gas path state data into a high-dimensional space by adopting a high-dimensional mapping method based on a polynomial kernel explicit mapping function or a high-dimensional space modeling method based on experience;

step 2, carrying out normalization processing on the input of the sample, and reducing interference between parameters caused by different orders of magnitude;

and 3, constructing a training sample set and a test sample set, and setting the label of the normal working state data of the engine to be 0. If the classification is two, the labels of all fault sample data are set to be 1; if the fault type is multi-classified, different labels are set for different fault types. Wherein the allocation ratio of the training set to the test set is 4:1;

step 4, training the LSTM, TCN, CNN or DRSN-CW network by using the training sample set, and keeping the network parameters unchanged after the network training is completed;

and 5, inputting the test set into the trained network model, comparing the classification result with the labels of the test set, and outputting the Accuracy Accurcy, the Precision and the F1 score F1-score of the network model in the test set.

The following were performed to experimentally verify the engine dataset: in order to verify that the high-dimensional mapping method based on the polynomial core explicit mapping function and the high-dimensional space modeling method based on experience can improve the data separability, an actual operation and maintenance data set of an engine and a bearing vibration data set are adopted for verification. For convenience in description of experimental process, a high-dimensional mapping method based on a polynomial kernel explicit mapping function is called Scheme1 (Scheme 1); the empirical-based high-dimensional mapping spatial modeling method is referred to as Scheme2 (Scheme 2). Convolutional self-encoders (CAE) and (AE) are chosen as high-dimensional map contrast methods. Therefore, in experiments, scheme1, scheme2, cae, ae were used as methods for feature amplification, and Raw data represents Raw data without feature amplification.

Since the data after the high-dimensional mapping acquired by CAE and AE are difficult to calculate by a minimum norm method, the orthogonality strength is not analyzed. The experimental results of fault diagnosis using the actual operation and maintenance data set of the engine are shown in tables 3 and 4, and the bar charts corresponding to tables 3 and 4 are fig. 12 to 15 and fig. 16 to 19, respectively.

TABLE 3 accuracy of two classifications of engine operation data

TABLE 4 Multi-class accuracy of Engine operation and maintenance data

As for the task of the classification fault diagnosis, analysis table 3 and fig. 12 to 15 show that the feature amplification methods (Scheme 1 and Scheme 2) according to the present embodiment all obtain classification results superior to other methods on three evaluation indexes (Accuracy, precision, F1-score) of four classifiers (CNN, LSTM, TCN, DRSN-CW). When CNN is used as a classifier, the Scheme2 obtains the optimal classification result, the Scheme1 obtains the second best classification result, and AE obtains the worst classification result; when the LSTM is used as a classifier, the Scheme1 obtains the optimal classification result, the Scheme2 obtains the second best classification result, and the Raw data obtains the worst classification result; when the TCN is used as a classifier, the Scheme2 obtains the optimal classification result, the Scheme1 obtains the second best classification result, and the CAE obtains the worst classification result; when DRSN-CW is used as the classifier, scheme1 obtains the optimal classification result, scheme2 obtains the second best classification result, and AE obtains the worst classification result.

For multi-classification fault diagnosis tasks, analysis of table 4 and fig. 16-19 shows that Scheme1 and Scheme2 are still superior to other methods in classification results. When CNN is used as a classifier, the Scheme2 obtains the optimal classification result, the Scheme1 obtains the second best classification result, and AE obtains the worst classification result; when the LSTM is used as a classifier, the Scheme1 obtains the optimal classification result, the Scheme2 obtains the second best classification result, and the AE obtains the worst classification result; when the TCN is used as a classifier, the Scheme1 obtains the optimal classification result, the Scheme2 obtains the second best classification result, and the CAE obtains the worst classification result; when DRSN-CW is used as the classifier, scheme1 obtains the optimal classification result, scheme2 obtains the second best classification result, and CAE obtains the worst classification result. From the above analysis, it is known that, in both the two-classification and multi-classification experimental results, under the same deep learning model, the classification results of Scheme1 and Scheme2 are superior to the classification results of Raw data, CAE, AE. Even three metrics of Scheme1 and Scheme2 in TCN deep learning model are greater than 90%.

According to the analysis, in the two-classification task and the multi-classification task, no matter which model of the four models is used as a final fault diagnosis classifier, the two feature amplification methods are used for carrying out feature amplification on the original sample, and then the classifier is trained for fault diagnosis, so that the obtained diagnosis effect is obviously better than that of other methods. Experimental results show that the original sample is mapped into a proper high-dimensional space by using the characteristic amplification method provided by the invention, so that the separability of the sample is effectively improved, and the problem of dimension disaster is avoided.

The calculation time required after the data high-dimensional mapping is longer than that of the original data in calculation time. This is because the larger the dimension of the sample, the larger the calculation time required in the case where the number of input samples is the same. Meanwhile, since the scheme2 needs to perform smoothing denoising treatment on the data, the calculation time is further increased, but the required maximum time is only 166.3s. The time includes training time and test time of the model. When the time for actually testing the sample is very short after model training is completed, the test time is not more than 5s.

Public dataset experimental verification:

to further illustrate schema 1 and schema 2 dataThe high-dimensional mapping method can effectively improve the classification result of fault diagnosis, and the bearing vibration data set disclosed by the Kassi Chu Da bearing data center is adopted for verification. Orthogonality after Scheme1,Scheme 2,Raw data high-dimensional mapping was analyzed prior to experimental verification. K of Raw data _f 0.99, K after Scheme1 high-dimensional mapping _f 0.89, K after scheme2 high-dimensional mapping _f 0.62, all greater than the upper threshold value of 0.1. The Scheme1,Scheme 2,Raw data high-dimensional mapping is shown to have strong orthogonality of all features. The experimental results of the bearing vibration data set are shown in table 5, and the bar charts corresponding to table 5 are shown in fig. 20 to 23.

TABLE 5 Multi-class accuracy of bearing vibration dataset

From the analysis of table 5, the classification effect of the bearing data set on the four classifiers as a whole was improved compared to the engine data set. This is because the bearing dataset is data measured under laboratory conditions, and the engine dataset is data measured under conditions where the engine is actually operating. The bearing dataset is superior to the engine dataset in terms of data quality. However, as can be seen from the analysis of table 5 and fig. 10, the classification results superior to Raw data, CAE, AE were obtained by the Scheme1 and the Scheme2 on different evaluation indexes of the four classifiers. The method is the same as the conclusion obtained by the actual operation and maintenance data set of the engine, and further illustrates that the high-dimensional mapping method of the Scheme1 and the Scheme2 can effectively improve the classification result of fault diagnosis. Further analysis of Scheme1 and Scheme2 shows that both Scheme1 and Scheme2 amplify features based on the original data, and that each feature retains certain orthogonality after high-dimensional mapping of the data. In the process of mapping the low-dimensional data to the high-dimensional space, if the orthogonality of the original data is strong and the data still has certain orthogonality after the high-dimensional mapping, the classification result of fault diagnosis can be effectively improved after the high-dimensional mapping.

Although the invention herein has been described with reference to particular embodiments, it is to be understood that these embodiments are merely illustrative of the principles and applications of the present invention. It is therefore to be understood that numerous modifications may be made to the illustrative embodiments and that other arrangements may be devised without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that the different dependent claims and the features described herein may be combined in ways other than as described in the original claims. It is also to be understood that features described in connection with separate embodiments may be used in other described embodiments.

Claims

1. A fault diagnosis method of an aeroengine based on feature amplification is characterized by comprising the following steps of,

2. The method for diagnosing an aircraft engine failure based on feature amplification according to claim 1, wherein,

3. The method for diagnosing an aircraft engine fault based on feature amplification according to claim 2, wherein the method for performing high-dimensional feature amplification on the original sample in the first step comprises a high-dimensional mapping method based on a polynomial kernel explicit mapping function, specifically comprising:

φ _p (x _i ) Is of the dimension of

will phi _p (x _i ) Constant term 1 and coefficient in the expression of (2)

Discard, marked as phi' _p (x _i )：

Phi 'is set' _p (x _i ) As a post-feature amplification sample.

4. The method for diagnosing an aircraft engine fault based on feature amplification as recited in claim 2, wherein the method for performing high-dimensional feature amplification on the original sample in the first step includes an empirical high-dimensional spatial modeling method, specifically:

In the I-th original sample x _I The method comprises the following steps:

5. The method for diagnosing an aircraft engine failure based on feature amplification according to claim 3 or 4, wherein,

the normalized sample is also used for constructing a test sample set, and the distribution ratio of the training sample set to the test sample set is 4:1.

6. The feature amplification-based aeroengine fault diagnosis method according to claim 5, wherein sample data input by the fault diagnosis network each time is selected from the training sample set through a sliding window.

7. The method for diagnosing an aircraft engine failure based on feature amplification according to claim 6, wherein,

the fault diagnosis network is a convolutional neural network CNN, a long and short term memory network LSTM, a time convolutional network TCN or a depth residual error contraction network DRSN-CW.

8. The method for diagnosing an aircraft engine failure based on feature amplification according to claim 7, wherein,

the learning rate of the fault diagnosis network is set to be 0.001, an Adam optimizer is adopted for updating the weight, and L2 regularization is used for reducing overfitting; the attenuation coefficient of L2 regularization was 0.0001, batch size 10.