CN105094118A - Airplane engine air compressor stall detection method - Google Patents

Abstract

The invention discloses an airplane engine air compressor stall detection method. The method comprises steps: a sensor is firstly used for collecting pressure signals when the air compressor is in a normal working condition or in a stall condition, and pretreatment is carried out; a negative selection principle in an artificial immune system is then used for building a neural network detector, and the fault detection ability of the detector is improved through training; and finally, stall detection is carried out on the acquired airplane engine air compressor pressure fluctuation signals with a fault label. The negative selection principle in the artificial immune system is used for building the neural network detector, abnormal mode information of the air compressor is stored in the distributed detector through training, fault is found out according to activation of the detector, stall signals can be detected instantly when stall happens, and the detection speed is improved; existing historical data are fully used through training, and the detection rate is improved; and stall of the airplane engine air compressor can be accurately and effectively detected.

Description

A kind of aircraft engine compressor stall fault detection method

Technical field

The present invention relates to a kind of aircraft engine fault detection method, specifically a kind of aircraft engine compressor stall fault detection method.

Background technology

High-speed, multi-stage axial-flow compressor during steady operation, if intergrade generation stall, so directly will cause engine surge under middle high rotating speed.Therefore, the aircraft engine of axial-flow compressor is adopted for major part, under middle high rotating speed stable state, by the detection to stall signal, instruction disappears and breathes heavily system works, to disappear the working method of breathing heavily relative to detection surge signal, the method can prevent compressor pressure ratio and efficiency from occurring greater loss, significant for the stability improving engine operation.

To the detection of aircraft engine compressor stall fault, the performance change mainly showed by stall rear engine is at present differentiated, belong to detection mode afterwards, have comparatively large time delay in time, be difficult to the stall fault finding engine timely and accurately.Therefore current fault detection method can not detect the fault of aircraft engine stall exactly.

Summary of the invention

For above-mentioned prior art Problems existing, the invention provides a kind of aircraft engine compressor stall fault detection method, utilize the reverse side choosing principles of artificial immune system to build Neural Network Detector, by training, the abnormal patterns information of pneumatic plant is stored in the detecting device of distribution, activation according to detecting device finds fault, can detect aircraft engine compressor stall fault accurately and effectively.

To achieve these goals, this aircraft engine compressor stall fault detection method specifically comprises the following steps:

Step 1: obtain aircraft engine pneumatic plant and normally work and the pressure signal under stall condition;

Step 2: utilize reverse side choosing principles to build neural network reverse side preference pattern;

Step 3: build the detecting device be made up of three layers of Feedback Neural Network;

Step 4: determine weight vector by the training of detecting device;

Step 5: by Simulating Test Study, determines the parameter of neural network reverse side preference pattern;

Step 6: carry out stall detection for pressure signal;

Step 7: in testing process, by the exception that S the qualified detecting device that aircraft engine pneumatic plant pressure information to be tested substitution step 6 obtains is used in detected pressures signal, according to the statistical information activating detecting device, judge whether aircraft engine compressor stall fault occurs.

Further, utilize the concrete grammar of reverse side choosing principles structure neural network reverse side preference pattern as follows in described step 2:

Step 2.1: define oneself space and non-own space, oneself space is made up of proper vector during system normal condition, represents with S set elf; Non-own space is made up of the proper vector when system failure or abnormality, represent, and Non_Self is the supplementary set of Self with set Non_Self;

Step 2.2: carry out binary coding to the feature samples of system, form oneself S set elf, represents the normal mode of system, and determines a matching threshold r, when the figure place of two continuous couplings of series winding is more than or equal to r, and two String matching, otherwise do not mate;

Step 2.3: random generation detecting device collection D, each detecting device in D and the sample in Self do not match;

Step 2.4: detect oneself to gather by the matching degree of sample in detecting device and Self and whether change.

Further, the concrete grammar building the detecting device be made up of three layers of Feedback Neural Network in described step 3 is as follows:

Step 3.1: at ground floor, input vector [x ₁, x ₂..., x _n] in element x _iwith weight vector [w ₁, w ₂..., w _n] in element w _ibetween matching degree d _icomputing formula at each input node place is:

d _i＝(x _i-w _i) ²，i＝1,2,…,N；

Step 3.2:d _ias the input of second layer node, f () is second layer node function, adopts the radial basis function of following form:

$f(x,\mathrm{\δ})=\exp (-\frac{\left|\right|x|{|}^2}{2{\mathrm{\δ}}^2});$

Step 3.3: the output of the second layer again as third layer, i.e. the input of output layer, the expression formula finally exporting y is:

$y=\underset{i=1}{\overset{N}{\Σ}}{v}_{i}f\left(d_i\right)=\underset{i=1}{\overset{N}{\Σ}}{v}_{i}f\[{(x_i-w_i)}^2\];$

Step 3.4: the matching error E obtaining detecting device:

E＝y-λ；

If to any one input vector [x ₁, x ₂..., x _n], matching error E > 0, then judge that this detecting device can not mate the sample in oneself space.

Further, determine that the concrete grammar of weight vector is as follows by the training of detecting device in described step 4:

W and v is made to represent weight vector [w respectively ₁, w ₂..., w _n] and [v ₁, v ₂..., v _n], for given a bit (w ^*, v ^*), in order to increase output error E (w ^*, v ^*), weight w and v should respectively to E (w ^*, v ^*) positive gradient direction changes delta w ^*with Δ v ^*,

${\mathrm{\Δw}}^{*}=\mathrm{\η}{\▿}_{w^{*}}E(w^{*},v^{*})=\mathrm{\μ}\frac{\∂E(w^{*},v^{*})}{\∂w^{*}};$

${\mathrm{\Δv}}^{*}=\mathrm{\η}{\▿}_{v^{*}}E(w^{*},v^{*})=\mathrm{\μ}\frac{\∂E(w^{*},v^{*})}{\∂v^{*}};$

Order with be illustrated respectively in w in kth time iterative process _iand v _ivalue, then

$\begin{array}{c}{w}_i^{k+1}=w_i^{\left(k\right)}+\mathrm{\μ}\frac{\∂E}{\∂w_i^{\left(k\right)}}=w_i^{\left(k\right)}+\mathrm{\μ}\frac{\∂y}{\∂w_i^{\left(k\right)}}\\ =w_i^{\left(k\right)}-2{\mathrm{\μv}}_i^{\left(k\right)}{f}^{\′}\left\{{\[x_i-w_i^{\left(k\right)}\]}^2\right\}\[x_i-w_i^{\left(k\right)}\]\end{array};$

$\begin{array}{c}{v}_i^{k+1}=v_i^{\left(k\right)}+\mathrm{\μ}\frac{\∂E}{\∂v_i^{\left(k\right)}}=v_i^{\left(k\right)}+\mathrm{\μ}\frac{\∂y}{\∂v_i^{\left(k\right)}}\\ =v_i^{\left(k\right)}+\mathrm{\μ}f\left\{{\[x_i-w_i^{\left(k\right)}\]}^2\right\}\end{array}.$

Further, by Simulating Test Study in described step 5, determine that the design parameter of neural network reverse side preference pattern is as follows:

Detecting device number S; Window width N; Matching threshold λ;

As the normal data of verifying smart data analysing method validity, Mackey-Glass chaos time sequence x (t) is produced by nonlinear differential equation below:

$\stackrel{\·}{x}\left(t\right)=\frac{ax(t-\mathrm{\τ})}{1+x^c(t-\mathrm{\τ})}-bx\left(t\right);$

Wherein, a=0.1, b=0.2, c=10;

The concrete span of described detecting device number S, window width N and matching threshold λ is obtained by above-mentioned emulation experiment.

Further, the concrete grammar of stall detection is carried out for pressure signal in described step 6 as follows:

Step 6.1: in training flow process, carry out windowing process to pressure signal, by the window do not overlapped each other, the width of window is N, and time series signal is decomposed into different time series sections, as the input mode vector of detecting device;

Step 6.2: by reverse side selection course, produces an initial S detecting device, then by the training of detecting device, finally obtains S qualified detecting device;

Step 6.3: for detected pressure signal time series, fault detect rate η is:

$\mathrm{\η}=\frac{M}{M+L}\×100\%;$

Wherein M is the correct detecting device number activated, and L is the detecting device number of erroneous activation.

Compared with prior art, first this aircraft engine compressor stall fault detection method utilizes sensor collection pneumatic plant normally to work and the pressure signal under stall condition, and carries out pre-service; Then utilize the reverse side choosing principles of artificial immune system to build Neural Network Detector, and improve the fault-detecting ability of detecting device by training; Finally stall detection is carried out to the aircraft engine pneumatic plant pressure fluctuation signal with faulty tag gathered.The present invention utilizes the reverse side choosing principles of artificial immune system to build Neural Network Detector, by training, the abnormal patterns information of pneumatic plant is stored in the detecting device of distribution, activation according to detecting device finds fault, stall signal can be detected the instantaneous of stall generation, improve the speed of detection; And existing historical data can be utilized fully by training, improve verification and measurement ratio; And then aircraft engine compressor stall fault can be detected accurately and effectively.

Accompanying drawing explanation

Fig. 1 is the structural representation of Neural Network Detector in the present invention;

Fig. 2 is stall fault detect schematic flow sheet in the present invention;

Fig. 3 be in the present invention detecting device number to the influence curve figure of fault detect rate;

Fig. 4 be in the present invention window width to the influence curve figure of fault detect rate;

Fig. 5 mates the influence curve figure of threshold values to fault detect rate in the present invention;

Fig. 6 be in the present invention fault detect rate with the change curve of train epochs.

Embodiment

Below in conjunction with accompanying drawing, the present invention will be further described.

High-speed, multi-stage axial-flow compressor during steady operation, if intergrade generation stall, so directly will cause engine surge under middle high rotating speed.Therefore, the aircraft engine of axial-flow compressor is adopted for major part, under middle high rotating speed stable state, by the detection to stall signal, instruction disappears and breathes heavily system works, to disappear the working method of breathing heavily relative to detection surge signal, the method can prevent compressor pressure ratio and efficiency from occurring greater loss, significant for the stability improving engine operation.

To the detection of aircraft engine compressor stall fault, the performance change mainly showed by stall rear engine is at present differentiated, belong to detection mode afterwards, have comparatively large time delay in time, be difficult to the stall fault finding engine timely and accurately.When stall appears in pneumatic plant, the pressure surge that its stall air mass causes, will change the pattern feature of original signal.Therefore, stall detects can be converted into a kind of abnormality detection problem, and that is, the change of testing pattern in characteristic time sequence, carries out differentiation that is normal and abnormal patterns.

The present invention utilizes reverse side choosing principles to build Neural Network Detector, finds the abnormity point of signal by activating detecting device, and then the stall fault of detection pneumatic plant.Test findings shows, utilizes this method can detect compressor stall fault exactly.

Convenient in order to discuss problem, first the present invention defines oneself space and non-own space: oneself space is made up of proper vector during system normal condition, represents with S set elf; Non-own space is made up of the proper vector when system failure or abnormality, represent, and Non_Self is the supplementary set of Self with set Non_Self.

The basic step of negative selection algorithm is: first carry out binary coding to the feature samples of system, form oneself S set elf, represents the normal mode of system, and determine a matching threshold r, when the figure place of two continuous couplings of series winding is more than or equal to r, two String matching, otherwise do not mate; Secondly, random generation detecting device collection D, each detecting device in D can not match with the sample in Self; Detect oneself to gather finally by the matching degree of sample in detecting device and Self and whether change.In this process simulation immune system, antibody is to the identifying of antigen.

In negative selection algorithm, produce the key that suitable detecting device is algorithm, this depends on three fundamentals: the generating algorithm of the coding of detecting device, matched rule and detecting device.But the negative selection algorithm of no matter any improvement is only the formation efficiency and the covering power that improve detecting device, and the structure of detecting device itself, still by binary numeral coded representation, does not thus break away from its intrinsic limitation.Neural network has very strong self study, self-organization and adaptive ability as a kind of important computational intelligence method, is incorporated in negative selection algorithm below by neural network, utilizes neural network to construct detecting device, fundamentally addresses these problems.

As depicted in figs. 1 and 2, this aircraft engine compressor stall fault detection method specifically comprises the following steps:

Step 1: obtain aircraft engine pneumatic plant and normally work and the pressure signal under stall condition;

Step 2: utilize reverse side choosing principles to build neural network reverse side preference pattern;

Step 3: build the detecting device be made up of three layers of Feedback Neural Network;

Step 4: determine weight vector by the training of detecting device;

Step 5: by Simulating Test Study, determines the parameter of neural network reverse side preference pattern;

Step 6: carry out stall detection for pressure signal;

Step 7: in testing process, by the exception that S the qualified detecting device that aircraft engine pneumatic plant pressure information to be tested substitution step 6 obtains is used in detected pressures signal, according to the statistical information activating detecting device, judge whether aircraft engine compressor stall fault occurs.

Further, utilize the concrete grammar of reverse side choosing principles structure neural network reverse side preference pattern as follows in described step 2:

Step 2.1: define oneself space and non-own space, oneself space is made up of proper vector during system normal condition, represents with S set elf; Non-own space is made up of the proper vector when system failure or abnormality, represent, and Non_Self is the supplementary set of Self with set Non_Self;

Step 2.2: carry out binary coding to the feature samples of system, form oneself S set elf, represents the normal mode of system, and determines a matching threshold r, when the figure place of two continuous couplings of series winding is more than or equal to r, and two String matching, otherwise do not mate;

Step 2.3: random generation detecting device collection D, each detecting device in D and the sample in Self do not match;

Step 2.4: detect oneself to gather by the matching degree of sample in detecting device and Self and whether change.

Further, the concrete grammar building the detecting device be made up of three layers of Feedback Neural Network in described step 3 is as follows:

Step 3.1: at ground floor, input vector [x ₁, x ₂..., x _n] in element x _iwith weight vector [w ₁, w ₂..., w _n] in element w _ibetween matching degree d _icomputing formula at each input node place is:

d _i＝(x _i-w _i) ²，i＝1,2,…,N；

Step 3.2:d _ias the input of second layer node, f () is second layer node function, adopts the radial basis function of following form:

$f(x,\mathrm{\δ})=\exp (-\frac{\left|\right|x|{|}^2}{2{\mathrm{\δ}}^2});$

Step 3.3: the output of the second layer again as third layer, i.e. the input of output layer, the expression formula finally exporting y is:

$y=\underset{i=1}{\overset{N}{\Σ}}{v}_{i}f\left(d_i\right)=\underset{i=1}{\overset{N}{\Σ}}{v}_{i}f\[{(x_i-w_i)}^2\];$

Step 3.4: the matching error E obtaining detecting device:

E＝y-λ；

If to any one input vector [x ₁, x ₂..., x _n], matching error E > 0, then judge that this detecting device can not mate the sample in oneself space.

Further, determine that the concrete grammar of weight vector is as follows by the training of detecting device in described step 4:

W and v is made to represent weight vector [w respectively ₁, w ₂..., w _n] and [v ₁, v ₂..., v _n], for given a bit (w ^*, v ^*), in order to increase output error E (w ^*, v ^*), weight w and v should respectively to E (w ^*, v ^*) positive gradient direction changes delta w ^*with Δ v ^*,

${\mathrm{\Δw}}^{*}=\mathrm{\η}{\▿}_{w^{*}}E(w^{*},v^{*})=\mathrm{\μ}\frac{\∂E(w^{*},v^{*})}{\∂w^{*}};$

${\mathrm{\Δv}}^{*}=\mathrm{\η}{\▿}_{v^{*}}E(w^{*},v^{*})=\mathrm{\μ}\frac{\∂E(w^{*},v^{*})}{\∂v^{*}};$

Order with be illustrated respectively in w in kth time iterative process _iand v _ivalue, then

$\begin{array}{c}{w}_i^{k+1}=w_i^{\left(k\right)}+\mathrm{\μ}\frac{\∂E}{\∂w_i^{\left(k\right)}}=w_i^{\left(k\right)}+\mathrm{\μ}\frac{\∂y}{\∂w_i^{\left(k\right)}}\\ =w_i^{\left(k\right)}-2{\mathrm{\μv}}_i^{\left(k\right)}{f}^{\′}\left\{{\[x_i-w_i^{\left(k\right)}\]}^2\right\}\[x_i-w_i^{\left(k\right)}\]\end{array};$

$\begin{array}{c}{v}_i^{k+1}=v_i^{\left(k\right)}+\mathrm{\μ}\frac{\∂E}{\∂v_i^{\left(k\right)}}=v_i^{\left(k\right)}+\mathrm{\μ}\frac{\∂y}{\∂v_i^{\left(k\right)}}\\ =v_i^{\left(k\right)}+\mathrm{\μ}f\left\{{\[x_i-w_i^{\left(k\right)}\]}^2\right\}\end{array}.$

Further, by Simulating Test Study in described step 5, determine that the design parameter of neural network reverse side preference pattern is as follows:

Detecting device number S; Window width N; Matching threshold λ;

As the normal data of verifying smart data analysing method validity, Mackey-Glass chaos time sequence x (t) is produced by nonlinear differential equation below:

$\stackrel{\·}{x}\left(t\right)=\frac{ax(t-\mathrm{\τ})}{1+x^c(t-\mathrm{\τ})}-bx\left(t\right);$

Wherein, a=0.1, b=0.2, c=10;

The concrete span of described detecting device number S, window width N and matching threshold λ is obtained by above-mentioned emulation experiment.

Concrete, according to equation the value changing τ will obtain different chaos time sequences, and with the time series of τ=30 correspondence for normal condition, with the time series of τ=20 correspondence for stall conditions, such stall fault detection problem is exactly to detect the exception caused by τ=20.Therefore, 500 sample points are gathered respectively from the chaos time sequence of τ=30 and τ=20, as normal sample set and stall fault sample collection, form the training sample set of 1000 sample points, and Neural Network Detector is trained, produce one group of new Mackey-Glass chaos time sequence as suspect signal, for stall fault detect simultaneously.The parameter value of algorithm is: detecting device number S=100; Window width N=10; Matching threshold λ=3.By changing the size of some parameters, keeping its excess-three parameter constant, analyzing the impact of this parameter on stall failure detection result.

As shown in Fig. 3, Fig. 4 and Fig. 5, can find: detecting device number is more by analyzing, represent that the kind of antibody is more, the ability that it catches alloantigen is stronger, just can reach good Detection results when detector number is greater than 100; Window width reflects the detail section of signal, and width is larger, and the schema category comprised will be more, simultaneously also can flood the change of some feature mode, and therefore window width N gets 10 proper; Less matching threshold can not effectively catch the essential characteristic of stall fault mode and produce error detection, larger matching threshold then can make detecting device not be activated and cause leakage detected artifacts, only have and select moderate matching threshold, between λ=3 ~ 3.5, just can obtain good Detection results.

Further, as shown in Figure 2, the concrete grammar of stall detection is carried out for pressure signal in described step 6 as follows:

Step 6.1: in training flow process, carry out windowing process to pressure signal, by the window do not overlapped each other, the width of window is N, and time series signal is decomposed into different time series sections, as the input mode vector of detecting device;

Step 6.2: by reverse side selection course, produces an initial S detecting device, then by the training of detecting device, finally obtains S qualified detecting device;

Step 6.3: for detected pressure signal time series, fault detect rate η is:

$\mathrm{\η}=\frac{M}{M+L}\×100\%;$

Wherein M is the correct detecting device number activated, and L is the detecting device number of erroneous activation.

As shown in Figure 6, a nearly step analysis can find, fault detect rate η is relevant with train epochs, η increases gradually along with the increase of train epochs, when training beginning, η is 75% (M=40, L=13), and after the training of 800 steps, η just reaches 100% (M=83, L=0), the fault-detecting ability of Neural Network Detector therefore can be improved significantly by training.

In sum, the present invention utilizes the reverse side choosing principles of artificial immune system to build Neural Network Detector, by training, the abnormal patterns information of pneumatic plant is stored in the detecting device of distribution, activation according to detecting device finds fault, stall signal can be detected the instantaneous of stall generation, improve detection speed; And existing historical data can be utilized fully by training, improve verification and measurement ratio; And then detect aircraft engine compressor stall fault accurately and effectively.

Claims (6)

1. an aircraft engine compressor stall fault detection method, is characterized in that,

Specifically comprise the following steps:

Step 1: obtain aircraft engine pneumatic plant and normally work and the pressure signal under stall condition;

Step 2: utilize reverse side choosing principles to build neural network reverse side preference pattern;

Step 3: build the detecting device be made up of three layers of Feedback Neural Network;

Step 4: determine weight vector by the training of detecting device;

Step 5: by Simulating Test Study, determines the parameter of neural network reverse side preference pattern;

Step 6: carry out stall detection for pressure signal;

Step 7: in testing process, by the exception that S the qualified detecting device that aircraft engine pneumatic plant pressure information to be tested substitution step 6 obtains is used in detected pressures signal, according to the statistical information activating detecting device, judge whether aircraft engine compressor stall fault occurs.

2. a kind of aircraft engine compressor stall fault detection method according to claim 1, is characterized in that,

Utilize the concrete grammar of reverse side choosing principles structure neural network reverse side preference pattern as follows in described step 2:

Step 2.1: define oneself space and non-own space, oneself space is made up of proper vector during system normal condition, represents with S set elf; Non-own space is made up of the proper vector when system failure or abnormality, represent, and Non_Self is the supplementary set of Self with set Non_Self;

Step 2.2: carry out binary coding to the feature samples of system, form oneself S set elf, represents the normal mode of system, and determines a matching threshold r, when the figure place of two continuous couplings of series winding is more than or equal to r, and two String matching, otherwise do not mate;

Step 2.3: random generation detecting device collection D, each detecting device in D and the sample in Self do not match;

Step 2.4: detect oneself to gather by the matching degree of sample in detecting device and Self and whether change.

3. a kind of aircraft engine compressor stall fault detection method according to claim 1, is characterized in that,

The concrete grammar building the detecting device be made up of three layers of Feedback Neural Network in described step 3 is as follows:

Step 3.1: at ground floor, input vector [x ₁, x ₂..., x _n] in element x _iwith weight vector [w ₁, w ₂..., w _n] in element w _ibetween matching degree d _icomputing formula at each input node place is:

d _i＝(x _i-w _i) ²，i＝1,2,…,N；

Step 3.2:d _ias the input of second layer node, f () is second layer node function, adopts the radial basis function of following form:

$f(x,\mathrm{\δ})=\exp (-\frac{\left|\right|x|{|}^2}{2{\mathrm{\δ}}^2});$

Step 3.3: the output of the second layer again as third layer, i.e. the input of output layer, the expression formula finally exporting y is:

$y=\underset{i=1}{\overset{N}{\Σ}}{v}_{i}f\left(d_i\right)=\underset{i=1}{\overset{N}{\Σ}}{v}_{i}f\[{(x_i-w_i)}^2\];$

Step 3.4: the matching error E obtaining detecting device:

E＝y-λ；

If to any one input vector [x ₁, x ₂..., x _n], matching error E > 0, then judge that this detecting device can not mate the sample in oneself space.

4. a kind of aircraft engine compressor stall fault detection method according to claim 1, is characterized in that,

Determine that the concrete grammar of weight vector is as follows by the training of detecting device in described step 4:

W and v is made to represent weight vector [w respectively ₁, w ₂..., w _n] and [v ₁, v ₂..., v _n], for given a bit (w ^*, v ^*), in order to increase output error E (w ^*, v ^*), weight w and v should respectively to E (w ^*, v ^*) positive gradient direction changes delta w ^*with Δ v ^*,

${\mathrm{\Δw}}^{*}=\mathrm{\η}{\▿}_{w^{*}}E(w^{*},v^{*})=\mathrm{\μ}\frac{\∂E(w^{*},v^{*})}{\∂w^{*}};$

${\mathrm{\Δv}}^{*}=\mathrm{\η}{\▿}_{v^{*}}E(w^{*},v^{*})=\mathrm{\μ}\frac{\∂E(w^{*},v^{*})}{\∂v^{*}};$

Order with be illustrated respectively in w in kth time iterative process _iand v _ivalue, then

$\begin{array}{c}{w}_i^{k+1}=w_i^{\left(k\right)}+\mathrm{\μ}\frac{\∂E}{\∂w_i^{\left(k\right)}}=w_i^{\left(k\right)}+\mathrm{\μ}\frac{\∂y}{\∂w_i^{\left(k\right)}}\\ =w_i^{\left(k\right)}-2{\mathrm{\μv}}_i^{\left(k\right)}{f}^{\′}\left\{{\[x_i-w_i^{\left(k\right)}\]}^2\right\}\[x_i-w_i^{\left(k\right)}\]\end{array};$

$\begin{array}{c}{v}_i^{k+1}=v_i^{\left(k\right)}+\mathrm{\μ}\frac{\∂E}{\∂v_i^{\left(k\right)}}=v_i^{\left(k\right)}+\mathrm{\μ}\frac{\∂y}{\∂v_i^{\left(k\right)}}\\ =v_i^{\left(k\right)}+\mathrm{\μ}f\left\{{\[x_i-w_i^{\left(k\right)}\]}^2\right\}\end{array}.$

5. a kind of aircraft engine compressor stall fault detection method according to claim 1, is characterized in that,

By Simulating Test Study in described step 5, determine that the design parameter of neural network reverse side preference pattern is as follows:

Detecting device number S; Window width N; Matching threshold λ;

As the normal data of verifying smart data analysing method validity, Mackey-Glass chaos time sequence x (t) is produced by nonlinear differential equation below:

$\stackrel{\·}{x}\left(t\right)=\frac{ax(t-\mathrm{\τ})}{1+x^c(t-\mathrm{\τ})}-bx\left(t\right);$

Wherein, a=0.1, b=0.2, c=10;

The concrete span of described detecting device number S, window width N and matching threshold λ is obtained by above-mentioned emulation experiment.

6. a kind of aircraft engine compressor stall fault detection method according to claim 1, is characterized in that,

The concrete grammar of stall detection is carried out for pressure signal as follows in described step 6:

Step 6.1: in training flow process, carry out windowing process to pressure signal, by the window do not overlapped each other, the width of window is N, and time series signal is decomposed into different time series sections, as the input mode vector of detecting device;

Step 6.2: by reverse side selection course, produces an initial S detecting device, then by the training of detecting device, finally obtains S qualified detecting device;

Step 6.3: for detected pressure signal time series, fault detect rate η is:

$\mathrm{\η}=\frac{M}{M+L}\×100\%;$

Wherein M is the correct detecting device number activated, and L is the detecting device number of erroneous activation.