CN103953490A - Implementation method for monitoring status of hydraulic turbine set based on HLSNE - Google Patents

Abstract

An implementation method for monitoring the status of a hydraulic turbine set based on HSLNE comprises the following steps of (1) signal detection and acquiring, namely collecting a set of vibration signals which can comprehensively reflect vibration anomalies of different noise sources by utilizing a vibration sensor on the hydraulic turbine set; (2) feature extraction, namely calculating an optional linear projection matrix A by using the HLSNE, and performing feature extraction on the vibration signals according to the linear projection matrix A; (3) status identification, namely identifying whether the status of the vibration signals subjected to feature extraction is normal or abnormal; (4) status analysis, namely adopting a nearest neighbor classifier to judge, classify and analyze a fault source of abnormal vibration signals; (5) result output, namely presenting a diagnosis decision according to results of the status analysis. The beneficial effect of the implementation method is mainly reflected in that a heavy-tailed linear stochastic neighbor embedding analysis method (HLSNE) is applied to the status monitoring of the hydraulic turbine set, so that multi-parameter setting and adjusting are feasible in a gradient optimization process by using a fast descent method in the actual application process, parameters do not need to be adjusted manually, and the efficiency and robustness of the hydraulic turbine set status monitoring process are effectively improved.

Description

Water wheels unit status monitoring implementation method based on HLSNE

Technical field

The present invention relates to a kind of water wheels unit status monitoring implementation method and system based on HLSNE.

Background technique

Along with the enforcement of six categories of small rural projects, small power station's proportion in electric power energy structure increases gradually, structure as the Hydropower Unit of small power station's production process nucleus equipment is increasingly sophisticated, integrated degree is more and more higher, between different parts, dynamic behavior influences each other, interacts, unit vibration problem becomes increasingly conspicuous, and the impact that the safe and stable operation of electrical network is caused also highlights day by day.Can the safety that the operation health status of Hydropower Unit is not only related to hydroelectric power plant is also directly connected to hydroelectric power plant to power grid security, reliable electric power is provided economically.Therefore, Hydropower Unit is carried out to monitoring running state, guarantee hydroelectric power unit safety, reliable, stable operation, performance maximum generation benefit, tool is of great significance.But, existing technology is extracted difficulty and mostly is higher-dimension, non-linear and comprise the aspects such as bulk redundancy information and all have certain deficiency solving Hydropower Unit oscillating signal sample characteristics, cannot meet the concrete application of reality, be necessary to carry out more deep research.

In recent years, dimensionality reduction technology plays a part more and more important in fields such as pattern recognition, image retrieval and computer visions.The handled initial data in above-mentioned field often dimension is very high, has brought high computation complexity to subsequent operation, the problems such as large buffer memory and algorithm performance decay.Dimensionality reduction technology or be called feature extracting method by higher-dimension is inputted data projection to significant low n-dimensional subspace n with reduce data redudancy and alleviate " dimension disaster " problem.In mode identification procedure, especially data-oriented differentiate concrete application time, tend to face that input characteristic parameter dimension is too high, data band is made an uproar, the problem such as singular point disturbance or number of training disappearance, input sample is carried out to dimensionality reduction effectively and not only can make up above-mentioned defect, also there is a series of advantage: low-dimensional distributed flow shape that can mining data, and the higher-dimension distributed architecture of true and reliable reflection data; In the limited application process of computing capability, can reduce assessing the cost of discrimination model; There is stronger robustness and survivability; Greatly Lifting scheme recognition performance etc.Therefore dimensionality reduction technology is applied to Hydropower Unit oscillating signal sample process and there is extremely important far reaching significance so that it is carried out to monitoring running state.

Summary of the invention

Utilize fast-descending method to carry out needing manually to adjust in gradient optimization procedure the not validity of parameter in order to overcome existing water wheels unit status monitoring implementation method, the invention provides a kind of water wheels unit status monitoring implementation method based on HLSNE, the multi-parameter that adopts simple fixed point method of iteration to realize in gradient optimization procedure is set and is adjusted, and has effectively improved efficiency and the robustness of water wheels unit status monitoring process.

The technical solution adopted for the present invention to solve the technical problems is:

A water wheels unit status monitoring implementation method based on HLSNE, comprises the following steps:

Step 1, signal measures: utilize the vibration transducer in Hydropower Unit to gather one group of oscillating signal that can more comprehensively reflect different noise source abnormal vibrations;

1.1 initialize signal collections: Hydropower Unit operating conditions complexity, be subject to environmental disturbances large, comprehensive under existing experimental condition, effective noise source abnormal vibrations sample is difficult to obtain.Therefore, the present invention, according to the spectral characteristic of Hydropower Unit roadability and each noise source of obtaining through sampling analysis, builds one group of Hydropower Unit abnormal vibrations simulate signal;

1.2 Signal Pretreatment: the time-domain signal of sensor collection is carried out to Wavelet Denoising Method, and time-frequency domain conversion and the amplitude of extracting in appropriate frequency form input sample object.

Step 2, feature extraction: utilize HLSNE to calculate optimum linear projection matrix A, according to linear projection matrix A, oscillating signal is carried out to feature extraction, wherein A is a matrix, be the linear relationship of former space high dimensional data corresponding to the low dimension data in subspace, making former space high dimensional data is n d dimensional vector X={x ₁, x ₂..., x _n, x _nrepresent n high dimensional data sample, the low dimension data in subspace is n r (r<<d) dimensional vector Y={y ₁, y ₂..., y _n, y _nrepresent n low-dimensional data-mapping, A is the linear projection matrix of a r × d, meets y _i=Ax _ilinear relationship, wherein i gets 1～n;

2.1 determine sample matrix X=[x ₁, x ₂..., x _n], set variance parameter λ;

2.2 calculate Euclidean distance between two between input sample according to X; Calculate joint probability P according to formula (1) _ij, P _ij

Represent x in former space _iselect x _jprobability as neighbour:

$P_{\mathrm{ij}}=\frac{\exp (-{\left|\right|x_i-x_j\left|\right|}^2/{2\mathrm{\&lambda;}}^2)}{{\mathrm{\&Sigma;}}_{a\&NotEqual;b}\exp (-{\left|\right|x_a-x_b\left|\right|}^2/{2\mathrm{\&lambda;}}^2)}---\left(1\right)$

Wherein λ is the variance parameter of corresponding Gaussian function, and in addition, i, j, a, b are subscript parameters, get 1～n;

2.3 passing through types (2) calculate joint probability Q _ij, Q _ijrepresent y in subspace _iand y _jbetween similarity:

$Q_{\mathrm{ij}}=\frac{{(1+{\left|\right|{\mathrm{Ax}}_i-{\mathrm{Ax}}_j\left|\right|}^2)}^{-1}}{{\mathrm{\&Sigma;}}_{a\&NotEqual;b}{(1+|\left|{{\mathrm{Ax}}_a-{\mathrm{Ax}}_b\left|\right|}^2\right)}^{-1}}---\left(2\right)$

Wherein i, j, a, b are subscript parameters, get 1～n;

The heavy-tailed function S of 2.4 calculating (|| Ax _i-Ax _j|| ²)=(1+||Ax _i-Ax _j|| ²) ^-1;

2.5 according to formula (3) compute gradient dC (A)/d (A):

$\begin{array}{c}\frac{\mathrm{dC}\left(A\right)}{d\left(A\right)}=2\underset{\mathrm{ij}}{\mathrm{\&Sigma;}}(\frac{1}{{\mathrm{\&Sigma;}}_{a\&NotEqual;b}{q}_{\mathrm{ab}}}-\frac{P_{\mathrm{ij}}}{q_{\mathrm{ij}}})\&CenterDot;q_{\mathrm{ij}}\&CenterDot;(-\frac{h\left({\left|\right|{\mathrm{Ax}}_i-{\mathrm{Ax}}_j\left|\right|}^2\right)}{q_{\mathrm{ij}}})\left(A(x_i-x_j){(x_i-x_j)}^T\right)\\ =2\underset{\mathrm{ij}}{\mathrm{\&Sigma;}}(P_{\mathrm{ij}}-Q_{\mathrm{ij}})S\left({\left|\right|{\mathrm{Ax}}_i-{\mathrm{Ax}}_j\left|\right|}^2\right)\left(A(x_i-x_j){(x_i-x_j)}^T\right)\end{array}---\left(3\right)$

Wherein h (|| Ax _i-Ax _j|| ²(the 1+||Ax of)=- _i-Ax _j|| ²) ^-2, in addition, i, j are subscript parameters, get 1～n;

2.6 pass through fixed point alternative manner by the A in formula (3) ^mbe updated to A ^m+1:

Make dC (A)/d (A)=0, can be further converted to by simple method of iteration adjustment type (2):

$\begin{array}{c}2\underset{\mathrm{ij}}{\mathrm{\&Sigma;}}(P_{\mathrm{ij}}-Q_{\mathrm{ij}})S{\left(\right||{\mathrm{Ax}}_i-{\mathrm{Ax}}_j|\left|\right)}^2\left(A(x_i-x_j){(x_i-x_j)}^T\right)=0\\ A=A\underset{\mathrm{ij}}{\mathrm{\&Sigma;}}{Q}_{\mathrm{ij}}S{\left(\right||{\mathrm{Ax}}_i-{\mathrm{Ax}}_j|\left|\right)}^2(x_i-x_j){(x_i-x_j)}^T\\ A=A\left(\underset{\mathrm{ij}}{\mathrm{\&Sigma;}}{Q}_{\mathrm{ij}}S{\left(\right||{\mathrm{Ax}}_i-{\mathrm{Ax}}_j|\left|\right)}^2(x_i-x_j){(x_i-x_j)}^T\right)\\ {\left(\underset{\mathrm{ij}}{\mathrm{\&Sigma;}}{P}_{\mathrm{ij}}S{\left(\right||{\mathrm{Ax}}_i-{\mathrm{Ax}}_j|\left|\right)}^2(x_i-x_j){(x_i-x_j)}^T\right)}^{-1}\end{array}---\left(4\right)$

For making to express conveniently, following two auxiliary variables of definition:

$C=\underset{\mathrm{ij}}{\mathrm{\&Sigma;}}{P}_{\mathrm{ij}}S{\left(\right||{\mathrm{Ax}}_i-{\mathrm{Ax}}_j|\left|\right)}^2(x_i-x_j){(x_i-x_j)}^T---\left(5\right)$

$D=\underset{\mathrm{ij}}{\mathrm{\&Sigma;}}{Q}_{\mathrm{ij}}S{\left(\right||{\mathrm{Ax}}_i-{\mathrm{Ax}}_j|\left|\right)}^2(x_i-x_j){(x_i-x_j)}^T---\left(6\right)$

Can be further converted to by formula after simple calculations (6):

A＝ADC ^-1 (7)

Adopt subsequently line search method optimization formula (7): choosing the direction of search is p _m=ADC ^-1-A, p _m

Meet B _mp _m=-g _m, wherein g _mgradient equations while being the m time iteration; B _mbeing a positive definite matrix, is in order to ensure p _mbe the direction of search that functional value is declined, meet p _m ^tg _m<0.Choose the step-length α that meets Wolfe condition _m>0, to obtain the value A of projection matrix A after next iteration ^m+1=A ^m+ α _mp _m.

Step 3, state recognition: the oscillating signal after feature extraction is carried out to the normal or abnormal identification of state;

Step 4, state analysis: adopt nearest neighbor classifier to carry out identification and classification analysis to the source of trouble under abnormal transient vibration signal;

Step 5, result output: propose diagnosis decision-making according to state analysis result.

Technical conceive of the present invention: by existing water wheels unit monitoring present situation being investigated in real time and to existing monitoring method comparative analysis, by modern information technologies such as Internet of Things, embedded scm and mobile communication, water wheels unit status monitoring implementation method and system based on HLSNE are designed.Mainly comprise the following aspects content:

1) on-the-spot design of node: on-the-spot node is mainly responsible for the various signals in Real-time Collection and transmission water wheels unit running process, receive data acquisition command and some other configuration order from monitoring station, the data that collect are uploaded to monitoring station by RS-485 bus simultaneously, realize supervisory function bit;

2) design of monitoring station: monitoring station is a high performance process control machine IPC, connects by RS-485 bus the on-the-spot node being distributed in everywhere downwards.The monitoring of software of upper-position unit plays central role in whole condition monitoring and failure diagnosis system, mainly completes: the various parameters while gathering hydraulic turbine operation; The working method of on-the-spot node is set; Show the overall situation and the local state of water wheels unit in modes such as numeral, figures; Data store and management; Data analysis and fault diagnosis.The various data of utilizing monitoring system to collect, adopt various intellectual technologies, extract the various characteristic informations in data, thereby obtain the sign relevant to fault, utilize sign to carry out fault diagnosis; 3) communicating by letter of monitoring station and on-the-spot node: monitoring station receives the data of on-the-spot node by RS-485 bus, adopts principal and subordinate's working method, and monitoring station is main frame, and on-the-spot node is slave.

The invention has the beneficial effects as follows: 1) HLSNE dimension-reduction algorithm embeds random neighbour and on the basis of analytical method, introduces linear projective transformation matrix and generate corresponding subspace data, effectively solve random neighbour and embed " the sample exterior problem " of analytical method; 2) HLSNE is described the t random neighbour that distributes and is embedded the heavy-tailed feature of analytical method by distribute negative cost function that random neighbour embeds analytical method of t, contribute to choose suitable parameter and reach best cabrage, and carry out gradient optimization procedure by simple fixed point method of iteration, effectively improve the t random neighbour that distributes and embed efficiency and the robustness of analytical method iteration optimal process.The present invention has designed a kind of water wheels Unit State Monitor System of novelty, taking equipment running status as basic forecast equipment state development trend, can be in time, targetedly equipment is overhauled, not only improve the availability of equipment, also effectively reduce recondition expense.This system realizes easily, and cost is lower, is easy to promote.

Brief description of the drawings

Fig. 1 is the inventive method flow chart.

Fig. 2 is the structural drawing that uses the system of the inventive method.

Fig. 3 is the Technology Roadmap that uses the system of the inventive method.

Embodiment

The invention will be further described below.

A water wheels unit status monitoring implementation method based on HLSNE, comprises the following steps:

Step 1, signal measures: utilize the vibration transducer in Hydropower Unit to gather one group of oscillating signal that can more comprehensively reflect different noise source abnormal vibrations;

1.1 initialize signal collections: Hydropower Unit operating conditions complexity, be subject to environmental disturbances large, comprehensive under existing experimental condition, effective noise source abnormal vibrations sample is difficult to obtain.Therefore, the present invention, according to the spectral characteristic of Hydropower Unit roadability and each noise source of obtaining through sampling analysis, builds one group of Hydropower Unit abnormal vibrations simulate signal;

1.2 Signal Pretreatment: the time-domain signal of sensor collection is carried out to Wavelet Denoising Method, and time-frequency domain conversion and the amplitude of extracting in appropriate frequency form input sample object.

Step 2, feature extraction: utilize HLSNE to calculate optimum linear projection matrix A, according to linear projection matrix A, oscillating signal is carried out to feature extraction, wherein A is a matrix, be the linear relationship of former space high dimensional data corresponding to the low dimension data in subspace, making former space high dimensional data is n d dimensional vector X={x ₁, x ₂..., x _n, x _nrepresent n high dimensional data sample, the low dimension data in subspace is n r (r<<d) dimensional vector Y={y ₁, y ₂..., y _n, y _nrepresent n low-dimensional data-mapping, A is the linear projection matrix of a r × d, meets y _i=Ax _ilinear relationship, wherein i gets 1～n;

2.1 determine sample matrix X=[x ₁, x ₂..., x _n], set variance parameter λ;

2.2 calculate Euclidean distance between two between input sample according to X; Calculate joint probability P according to formula (1) _ij, P _ij

Represent x in former space _iselect x _jprobability as neighbour:

$P_{\mathrm{ij}}=\frac{\exp (-{\left|\right|x_i-x_j\left|\right|}^2/{2\mathrm{\&lambda;}}^2)}{{\mathrm{\&Sigma;}}_{a\&NotEqual;b}\exp (-{\left|\right|x_a-x_b\left|\right|}^2/{2\mathrm{\&lambda;}}^2)}---\left(1\right)$

Wherein λ is the variance parameter of corresponding Gaussian function, and in addition, i, j, a, b are subscript parameters, get 1～n;

2.3 passing through types (2) calculate joint probability Q _ij, Q _ijrepresent y in subspace _iand y _jbetween similarity:

$Q_{\mathrm{ij}}=\frac{{(1+{\left|\right|{\mathrm{Ax}}_i-{\mathrm{Ax}}_j\left|\right|}^2)}^{-1}}{{\mathrm{\&Sigma;}}_{a\&NotEqual;b}{(1+|\left|{{\mathrm{Ax}}_a-{\mathrm{Ax}}_b\left|\right|}^2\right)}^{-1}}---\left(2\right)$

Wherein i, j, a, b are subscript parameters, get 1～n;

The heavy-tailed function S of 2.4 calculating (|| Ax _i-Ax _j|| ²)=(1+||Ax _i-Ax _j|| ²) ^-1;

2.5 according to formula (3) compute gradient dC (A)/d (A):

$\begin{array}{c}\frac{\mathrm{dC}\left(A\right)}{d\left(A\right)}=2\underset{\mathrm{ij}}{\mathrm{\&Sigma;}}(\frac{1}{{\mathrm{\&Sigma;}}_{a\&NotEqual;b}{q}_{\mathrm{ab}}}-\frac{P_{\mathrm{ij}}}{q_{\mathrm{ij}}})\&CenterDot;q_{\mathrm{ij}}\&CenterDot;(-\frac{h\left({\left|\right|{\mathrm{Ax}}_i-{\mathrm{Ax}}_j\left|\right|}^2\right)}{q_{\mathrm{ij}}})\left(A(x_i-x_j){(x_i-x_j)}^T\right)\\ =2\underset{\mathrm{ij}}{\mathrm{\&Sigma;}}(P_{\mathrm{ij}}-Q_{\mathrm{ij}})S\left({\left|\right|{\mathrm{Ax}}_i-{\mathrm{Ax}}_j\left|\right|}^2\right)\left(A(x_i-x_j){(x_i-x_j)}^T\right)\end{array}---\left(3\right)$

Wherein h (|| Ax _i-Ax _j|| ²(the 1+||Ax of)=- _i-Ax _j|| ²) ^-2, in addition, i, j are subscript parameters, get 1～n;

2.6 pass through fixed point alternative manner by the A in formula (3) ^mbe updated to A ^m+1:

Make dC (A)/d (A)=0, can be further converted to by simple method of iteration adjustment type (2):

$\begin{array}{c}2\underset{\mathrm{ij}}{\mathrm{\&Sigma;}}(P_{\mathrm{ij}}-Q_{\mathrm{ij}})S{\left(\right||{\mathrm{Ax}}_i-{\mathrm{Ax}}_j|\left|\right)}^2\left(A(x_i-x_j){(x_i-x_j)}^T\right)=0\\ A=A\underset{\mathrm{ij}}{\mathrm{\&Sigma;}}{Q}_{\mathrm{ij}}S{\left(\right||{\mathrm{Ax}}_i-{\mathrm{Ax}}_j|\left|\right)}^2(x_i-x_j){(x_i-x_j)}^T\\ A=A\left(\underset{\mathrm{ij}}{\mathrm{\&Sigma;}}{Q}_{\mathrm{ij}}S{\left(\right||{\mathrm{Ax}}_i-{\mathrm{Ax}}_j|\left|\right)}^2(x_i-x_j){(x_i-x_j)}^T\right)\\ {\left(\underset{\mathrm{ij}}{\mathrm{\&Sigma;}}{P}_{\mathrm{ij}}S{\left(\right||{\mathrm{Ax}}_i-{\mathrm{Ax}}_j|\left|\right)}^2(x_i-x_j){(x_i-x_j)}^T\right)}^{-1}\end{array}---\left(4\right)$

For making to express conveniently, following two auxiliary variables of definition:

$C=\underset{\mathrm{ij}}{\mathrm{\&Sigma;}}{P}_{\mathrm{ij}}S{\left(\right||{\mathrm{Ax}}_i-{\mathrm{Ax}}_j|\left|\right)}^2(x_i-x_j){(x_i-x_j)}^T---\left(5\right)$

$D=\underset{\mathrm{ij}}{\mathrm{\&Sigma;}}{Q}_{\mathrm{ij}}S{\left(\right||{\mathrm{Ax}}_i-{\mathrm{Ax}}_j|\left|\right)}^2(x_i-x_j){(x_i-x_j)}^T---\left(6\right)$

Can be further converted to by formula after simple calculations (6):

A＝ADC ^-1 (7)

Adopt subsequently line search method optimization formula (7): choosing the direction of search is p _m=ADC ^-1-A, p _mmeet B _mp _m=-g _m, wherein g _mgradient equations while being the m time iteration; B _mbeing a positive definite matrix, is in order to ensure p _mbe the direction of search that functional value is declined, meet p _m ^tg _m<0.Choose the step-length α that meets Wolfe condition _m>0, to obtain the value A of projection matrix A after next iteration ^m+1=A ^m+ α _mp _m.

Step 3, state recognition: the oscillating signal after feature extraction is carried out to the normal or abnormal identification of state;

Step 4, state analysis: adopt nearest neighbor classifier to carry out identification and classification analysis to the source of trouble under abnormal transient vibration signal;

Step 5, result output: propose diagnosis decision-making according to state analysis result.

Water wheels Unit State Monitor System structure:

Various parameters when hydraulic turbine operation, as vibratory output, electric parameters, water levels of upstream and downstream, power loss, pressure, temperature etc., by being distributed in on-the-spot multiple data acquisition node Real-time Collections, then be sent to upper-position unit monitoring and fault diagnosis system by RS-485 bus, and be kept at the position that database specifies.The ST-DP type earthquake low-frequency shock transducer that vibration transducer selects Sang Tuo Institute for Applied Technology, Beijing to produce, throw sensor is selected CWY-DO series current vortex sensor, and pressure transducer is selected AK-1 type strain type oscillatory pressure pick-up.The important content of water wheels unit status monitoring to the collection of oscillating signal with analyzing.The vibration of water turbine has larger difference compared with the vibration of other machineries, and except wanting the mechanical property of taking into account system itself, the flow passage components of water turbine and electromagnetic force also can make water turbine produce larger vibration.System is divided into 3 levels, i.e. the on-site data gathering node of process layer, the monitoring station of monitor layer and server, the engineer station etc. of station level from vertical looking up.At process layer, in order to improve the real-time of data transmission, according to the quantity of water wheels unit by RS-485 bus sectionalization, a network segment of a water turbine composition, various status parameters while gathering hydraulic turbine operation by on-the-spot node, and data are sent to the monitoring station of monitor layer, the running state of a water wheels unit of a monitoring station charge of overseeing.Then by switching Ethernet, the running state of all water wheels units of whole hydroelectric power plant is all sent to the server of station level, issue the operation information of water wheels unit to outside by Internet by the Web server that is positioned at station level, and realize remote monitoring and the fault diagnosis based on Internet.

The design of water wheels unit state monitoring station:

The monitoring station that is positioned at monitor layer is a high performance process control machine IPC, connects the on-the-spot node being distributed in everywhere downwards by RS-485 bus, is upwards connected with Web server, database server etc. by switching Ethernet.The developing instrument of monitoring station is Visual C++ and SQL Server.The monitoring of software of upper-position unit plays central role in the monitoring of whole water wheels unit special topic, and the function completing comprises:

1. the various parameters while gathering hydraulic turbine operation, comprise quantity of state, electric parameters etc.

2. the working method of on-the-spot node is set.As the port number, the sampling period etc. that gather.

3. show the overall situation and the local state of water wheels unit in modes such as numeral, figures.Draw real-time curve chart, history graphs, frequency curve chart etc.Can be with the form display alarm information of sound, color.

4. data store and management.For the ease of to data analysis, the mass data that on-the-spot node is collected deposits database in, and access control is set.Data are divided into real time data and historical data, and these data both can be used as accident or trouble analysis use, also can be the operation of water wheels set optimization foundation is provided.

5. data analysis and fault diagnosis.The various data of utilizing monitoring system to collect, adopt various intellectual technologies, extract the various characteristic informations in data, thereby obtain the sign relevant to fault, utilize sign to carry out fault diagnosis.

6. the Operational Limits of water wheels unit is uploaded to the server of station level, and receive the various configuration orders from station level.

Claims (1)

1. the water wheels unit status monitoring implementation method based on HLSNE, comprises the following steps:

Step 1, signal measures: utilize the vibration transducer in Hydropower Unit to gather one group of oscillating signal that can more comprehensively reflect different noise source abnormal vibrations;

1.1 initialize signal collections: Hydropower Unit operating conditions complexity, be subject to environmental disturbances large, comprehensive under existing experimental condition, effective noise source abnormal vibrations sample is difficult to obtain.Therefore, the present invention, according to the spectral characteristic of Hydropower Unit roadability and each noise source of obtaining through sampling analysis, builds one group of Hydropower Unit abnormal vibrations simulate signal;

1.2 Signal Pretreatment: the time-domain signal of sensor collection is carried out to Wavelet Denoising Method, and time-frequency domain conversion and the amplitude of extracting in appropriate frequency form input sample object;

Step 2, feature extraction: utilize HLSNE to calculate optimum linear projection matrix A, according to linear projection matrix A, oscillating signal is carried out to feature extraction, wherein A is a matrix, be the linear relationship of former space high dimensional data corresponding to the low dimension data in subspace, making former space high dimensional data is n d dimensional vector X={x ₁, x ₂..., x _n, x _nrepresent n high dimensional data sample, the low dimension data in subspace is n r (r<<d) dimensional vector Y={y ₁, y ₂..., y _n, y _nrepresent n low-dimensional data-mapping, A is the linear projection matrix of a r × d, meets y _i=Ax _ilinear relationship, wherein i gets 1～n;

2.1 determine sample matrix X=[x ₁, x ₂..., x _n], set variance parameter λ;

2.2 calculate Euclidean distance between two between input sample according to X; Calculate joint probability P according to formula (1) _ij, P _ijrepresent x in former space _iselect x _jprobability as neighbour:

$P_{\mathrm{ij}}=\frac{\exp (-{\left|\right|x_i-x_j\left|\right|}^2/{2\mathrm{\&lambda;}}^2)}{{\mathrm{\&Sigma;}}_{a\&NotEqual;b}\exp (-{\left|\right|x_a-x_b\left|\right|}^2/{2\mathrm{\&lambda;}}^2)}---\left(1\right)$

Wherein λ is the variance parameter of corresponding Gaussian function, and in addition, i, j, a, b are subscript parameters, get 1～n;

2.3 passing through types (2) calculate joint probability Q _ij, Q _ijrepresent y in subspace _iand y _jbetween similarity:

$Q_{\mathrm{ij}}=\frac{{(1+{\left|\right|{\mathrm{Ax}}_i-{\mathrm{Ax}}_j\left|\right|}^2)}^{-1}}{{\mathrm{\&Sigma;}}_{a\&NotEqual;b}{(1+|\left|{{\mathrm{Ax}}_a-{\mathrm{Ax}}_b\left|\right|}^2\right)}^{-1}}---\left(2\right)$

Wherein i, j, a, b are subscript parameters, get 1～n;

The heavy-tailed function S of 2.4 calculating (|| Ax _i-Ax _j|| ²)=(1+||Ax _i-Ax _j|| ²) ^-1;

2.5 according to formula (3) compute gradient dC (A)/d (A):

$\begin{array}{c}\frac{\mathrm{dC}\left(A\right)}{d\left(A\right)}=2\underset{\mathrm{ij}}{\mathrm{\&Sigma;}}(\frac{1}{{\mathrm{\&Sigma;}}_{a\&NotEqual;b}{q}_{\mathrm{ab}}}-\frac{P_{\mathrm{ij}}}{q_{\mathrm{ij}}})\&CenterDot;q_{\mathrm{ij}}\&CenterDot;(-\frac{h\left({\left|\right|{\mathrm{Ax}}_i-{\mathrm{Ax}}_j\left|\right|}^2\right)}{q_{\mathrm{ij}}})\left(A(x_i-x_j){(x_i-x_j)}^T\right)\\ =2\underset{\mathrm{ij}}{\mathrm{\&Sigma;}}(P_{\mathrm{ij}}-Q_{\mathrm{ij}})S\left({\left|\right|{\mathrm{Ax}}_i-{\mathrm{Ax}}_j\left|\right|}^2\right)\left(A(x_i-x_j){(x_i-x_j)}^T\right)\end{array}---\left(3\right)$

Wherein h (|| Ax _i-Ax _j|| ²(the 1+||Ax of)=- _i-Ax _j|| ²) ^-2, in addition, i, j are subscript parameters, get 1～n;

2.6 pass through fixed point alternative manner by the A in formula (3) ^mbe updated to A ^m+1:

Make dC (A)/d (A)=0, can be further converted to by simple method of iteration adjustment type (2):

$\begin{array}{c}2\underset{\mathrm{ij}}{\mathrm{\&Sigma;}}(P_{\mathrm{ij}}-Q_{\mathrm{ij}})S{\left(\right||{\mathrm{Ax}}_i-{\mathrm{Ax}}_j|\left|\right)}^2\left(A(x_i-x_j){(x_i-x_j)}^T\right)=0\\ A=A\underset{\mathrm{ij}}{\mathrm{\&Sigma;}}{Q}_{\mathrm{ij}}S{\left(\right||{\mathrm{Ax}}_i-{\mathrm{Ax}}_j|\left|\right)}^2(x_i-x_j){(x_i-x_j)}^T\\ A=A\left(\underset{\mathrm{ij}}{\mathrm{\&Sigma;}}{Q}_{\mathrm{ij}}S{\left(\right||{\mathrm{Ax}}_i-{\mathrm{Ax}}_j|\left|\right)}^2(x_i-x_j){(x_i-x_j)}^T\right)\\ {\left(\underset{\mathrm{ij}}{\mathrm{\&Sigma;}}{P}_{\mathrm{ij}}S{\left(\right||{\mathrm{Ax}}_i-{\mathrm{Ax}}_j|\left|\right)}^2(x_i-x_j){(x_i-x_j)}^T\right)}^{-1}\end{array}---\left(4\right)$

For making to express conveniently, following two auxiliary variables of definition:

$C=\underset{\mathrm{ij}}{\mathrm{\&Sigma;}}{P}_{\mathrm{ij}}S{\left(\right||{\mathrm{Ax}}_i-{\mathrm{Ax}}_j|\left|\right)}^2(x_i-x_j){(x_i-x_j)}^T---\left(5\right)$

$D=\underset{\mathrm{ij}}{\mathrm{\&Sigma;}}{Q}_{\mathrm{ij}}S{\left(\right||{\mathrm{Ax}}_i-{\mathrm{Ax}}_j|\left|\right)}^2(x_i-x_j){(x_i-x_j)}^T---\left(6\right)$

Can be further converted to by formula after simple calculations (6):

A＝ADC ^-1 (7)

Adopt subsequently line search method optimization formula (7): choosing the direction of search is p _m=ADC ^-1-A, p _mmeet B _mp _m=-g _m, wherein g _mgradient equations while being the m time iteration; B _mbeing a positive definite matrix, is in order to ensure p _mbe the direction of search that functional value is declined, meet p _m ^tg _m<0.Choose the step-length α that meets Wolfe condition _m>0, to obtain the value A of projection matrix A after next iteration ^m+1=A ^m+ α _mp _m;

Step 3, state recognition: the oscillating signal after feature extraction is carried out to the normal or abnormal identification of state;

Step 4, state analysis: adopt nearest neighbor classifier to carry out identification and classification analysis to the source of trouble under abnormal transient vibration signal;

Step 5, result output: propose diagnosis decision-making according to state analysis result.