CN108875841A - A kind of pumped storage unit vibration trend forecasting method - Google Patents

Abstract

The invention discloses a vibration trend prediction method of a pumped-storage unit, which first acquires the historical and real-time data of the non-stationary vibration of the unit online, and then transmits the data to the user terminal, and then uses empirical wavelet decomposition to analyze the vibration signal in the time-frequency domain, Then extract the comprehensive features of energy entropy and singular value, then discretize the signal feature data set according to certain rules, and then carry out correlation analysis with the operating conditions of the unit, use the Apriori algorithm to mine frequent items, and analyze the temporal and spatial correlation between data features and unit failures The safe operation area of the unit is divided through the correlation analysis results, and finally the time series model is constructed, and the time series trend prediction method is used to predict its future development trend in a limited time, and then the unit’s operating status trend is predicted and evaluated. Condition-based maintenance provides technical support. The invention not only can accurately predict the trend, but also has the advantages of comprehensive evaluation indexes and convenient evaluation.

Description

A Vibration Trend Prediction Method for Pumped Storage Units

technical field

The invention relates to a vibration trend prediction method of a pumped storage unit.

Background technique

The pump turbine unit of the pumped storage power station has complex working conditions, frequent start and stop, many vibration parts and wide distribution, and many influencing factors (including hydraulic factors, mechanical factors and electromagnetic factors). The trend prediction method widely used at present is to obtain the historical data trend chart of the vibration state of the unit through the online monitoring system, subjectively analyze the possible operating state of the unit in the future, or use the least square method to achieve regression fitting on the characteristic parameters, but the vibration signal is generally non-linear. Linear and non-stationary, so traditional mathematical statistics methods are difficult to achieve accurate trend prediction of its state. Due to the improper selection of the weight of the AHP, it often cannot meet the requirements of the consistency of the judgment matrix, and cannot obtain the evaluation index reflecting the overall state of the pumped storage unit. Therefore, the existing vibration trend prediction methods for pumped storage units have the problems of inaccurate trend prediction and incomplete evaluation indicators.

Contents of the invention

The object of the present invention is to provide a vibration trend prediction method of a pumped storage unit. The invention not only can accurately predict the trend, but also has the advantage of comprehensive evaluation indexes.

The technical solution of the present invention: a method for predicting the vibration trend of a pumped-storage unit, comprising the following steps:

a1. On-line data acquisition. By adopting the 485 communication protocol, the history and real-time data of the non-stationary vibration of the unit are obtained online from the pumped-storage power station vibration monitoring and monitoring system;

b1. Data transmission, converting the historical and real-time data of unit non-stationary vibration obtained online into unit non-stationary vibration signals, and sequentially passing through the network isolation device, WEB server, hardware firewall and power plant LAN to the user terminal;

c1. Signal analysis. For the non-stationary vibration signal of the unit, the user terminal uses empirical wavelet decomposition to analyze the non-stationary vibration signal of the unit in time-frequency domain;

d1, feature extraction, on the basis of using empirical wavelet decomposition to analyze the non-stationary vibration signal of the unit in the time-frequency domain, extract the comprehensive features describing the vibration state of the unit from the non-stationary vibration signal of the unit, and obtain energy entropy and singular value;

e1. Correlation and quantitative analysis of vibration. For the data set of energy entropy and singular value, the vibration signal is decomposed by Ensemble Empirical Mode Decomposition (EEMD) or Improved Empirical Wavelet Decomposition (EWT) to obtain a series of single frequency After discretizing the components, perform correlation analysis with the operating conditions of the unit, use the Apriori algorithm to mine frequent items, analyze the time-space correlation between the vibration data characteristics of the unit and the failure of the unit, and divide the safe operation area of the unit through the correlation analysis results;

f1. Construct time series decomposition model, multiple linear regression model and ARMA model, use time series trend prediction method to predict the development trend in a limited time in the future, and then predict and evaluate the trend of unit operation status, and provide technology for the implementation of unit condition maintenance support.

In the aforementioned vibration trend prediction method of a pumped-storage unit, the time series trend prediction method in step f1 includes the following steps:

a2. Extract the time series of unit state parameters;

b2. Use empirical wavelet decomposition to decompose the time series of unit state parameters into subsequences of different decomposition domains, and obtain the stationary fluctuation item and the nonlinear trend item;

c2. Use the AR prediction model for the stationary fluctuation item to obtain signal A, and use the least squares support vector machine prediction model for the non-linear trend item to obtain signal B;

d2. Reconstruct the signals of signal A and signal B to realize the prediction of unit state trend.

In the aforementioned method for predicting the vibration trend of pumped storage units, in step b1, data transmission can also convert the historical and real-time data of unit non-stationary vibration obtained online into unit non-stationary vibration signals, and then pass through the network isolation device, WEB The server, hardware firewall and Alibaba Cloud are delivered to the APP on the mobile terminal.

In the aforementioned vibration trend prediction method of a pumped-storage unit, the prediction model of the least squares support vector machine is expressed as

Among them X is the input vector, n is a constant, k is a certain value in 1～n, m is the space dimension, τ is the delay time constant.

In the aforementioned vibration trend prediction method for pumped storage units, the AR model is expressed as

In the formula: is the autoregressive parameter, i=1...p, p is the order; a _t is the white noise, which means the residual.

In the aforementioned vibration trend prediction method of pumped-storage units, the order of the model is determined by the AIC criterion when the AR model is modeled, and the parameters of the model are estimated.

In the aforementioned vibration trend prediction method of a pumped-storage unit, the method for reconstructing the signal in step d2 includes the following steps:

a3. Perform multi-layer wavelet decomposition on the time series of the state parameters of the unit, and obtain the sub-sequences of each layer describing the trend item and the fluctuation item:

X=a _k +d ₁ +d ₂ +...+d _k

In the formula, a _k is the trend item, d _i is the fluctuation item, i=1, 2, ..., k is the number of wavelet decomposition layers;

b3. Establish an LSSVM prediction model for the trend item a _k , perform model training and new value prediction, and accurately evaluate the results;

c3. Perform AR modeling for each fluctuation item d _i , i=1, 2, ..., k, and evaluate the accuracy of the results;

d3. After obtaining the prediction results of the trend item and each fluctuation item, perform superposition calculation to obtain the prediction sequence of the original vibration sequence Calculate the error index of the measured value and the predicted value, and evaluate the accuracy;

e3. Evaluate the accuracy of the prediction result of the original vibration sequence, analyze the degree of deviation between the prediction result and the actual value, and use the following common error indicators to evaluate the prediction effect.

In the aforementioned vibration trend prediction method of a pumped-storage unit, the commonly used error indicators include

Average relative error: and

root mean square error:

Where k is a value from 1 to N, is the average value.

In the aforementioned vibration trend prediction method of a pumped storage unit, the signal analysis shows that for a signal f(t), EWT decomposes it into N+1 modal functions f _k (t) of single frequency components.

where the AM-FM signal of f _k (t) can be defined as:

f _k (t)＝F _k (t)cos(φ _k (t))

Among them, k is a value from 0 to N, F _k is a frequency domain function, and φ _k is an angle vector.

All non-stationary signal decomposition methods hope to decompose the f _k (t) of the AM-FM component from the original signal. EWT extracts the AM-FM components of the signal by constructing a wavelet filter by adaptively dividing the Fourier spectrum. Assume that the Fourier support [0, π] is split into N consecutive parts. Each segment is represented by ∧ _n :

Λ _n = [ω _n-1 , ω _n ], n=1, 2, ..., N

Among them, ω _n is the boundary of each section, and ω _n can be selected in many ways. Generally, the end point between two adjacent maximum points in the Fourier spectrum is taken, and ω ₀ =0, ω _n =π. With ω _n as the center, a transition section with width T _n =2τ _n is defined. According to the construction method of Meyer wavelet, a bandpass filter is constructed for each ∧ _n , and its empirical wavelet function and the empirical scaling function defined as:

in,

Assuming that the Fourier transform is recorded as F[·], and the Fourier inverse transform is recorded as F ^-1 [·]; then the empirical wavelet transform is defined according to the solution method of the traditional wavelet transform, and the detailed coefficients of the empirical wavelet transform are determined by the empirical wavelet function The inner product of Ψ _n (t) and signal f(t) is obtained:

The empirical wavelet approximation coefficients are obtained from the inner product of the signal f(t) and the scaling function Ψ ₁ (t):

in, and are the Fourier transform of empirical wavelet function and scaling function; and are the complex conjugates of the empirical wavelet function and scaling function, respectively.

The original signal can be reconstructed from the empirical wavelet function and scaling function:

Among them, * means convolution; ∧ means Fourier transform.

From this, the mode f _k (t) decomposed by the empirical wavelet can be obtained:

In the aforementioned vibration trend prediction method of a pumped storage unit, the feature extraction is assuming that the original signal is subjected to empirical wavelet transformation to obtain a series of modes, c _s (t), s=1, 2,..., K, K is the number of decomposed layers, E ₁ , E ₂ ,..., E _K is the energy value corresponding to each layer mode. The calculation process of the energy feature is as follows:

First calculate the energy signature of each layer mode:

Then calculate the total energy value for all modes:

Finally calculate the energy entropy feature;

The singular value features of all modal matrixes will also be extracted after the signal empirical wavelet transform. Singular value features are a series of eigenvectors obtained by performing singular value decomposition on a matrix. According to the principle of matrix theory, a singular value is a value that reflects the inherent properties of a matrix, and this value is stable. At the same time, singular values are scale invariant and rotation invariant. Therefore, the singular value feature is a very reliable evaluation index to distinguish different faults.

After the original signal undergoes empirical wavelet transform, the energy of each layer decreases gradually from the first layer to the last layer. When the current Q-layer mode can almost contain most of the energy information of the original signal, the former Q-layer mode is taken. state to construct the initial characteristic matrix A:

A=[c ₁ ,c ₂ ,...,c _Q ] ^T

Assume that the size of the real matrix A is PQ and satisfy the following conditions:

A＝ ^U∧VT

Where U and V are orthogonal matrices of size QQ and PP respectively, ∧ is a fault feature matrix of size QP, and all the constituent elements σ _i (i=1, 2,..., Q)(Q≤ P) Arranged from low to high according to the number of modal layers. These failure matrix elements are some singular values of matrix A. Therefore, the vector [σ ₁ , σ ₂ , . . . , σ _Q ] is taken as another fault feature.

Compared with the prior art, the present invention improves the existing vibration trend prediction method of the pumped-storage unit, by establishing a trend prediction model reflecting the key characteristic parameters of the vibration state of the unit, optimizing the prediction algorithm from multiple angles, and constructing a combination based on time series using the wavelet transform theory to extract the detailed features of the signal, decomposing the vibration state parameters of the unit into nonlinear trend items and stationary fluctuation items, respectively using least squares support vector machine (LSSVM) theory and autoregressive (AR ) model to predict the trend, and finally use the addition principle to reconstruct the signal to realize the trend prediction of the vibration state parameters of the unit. The vibration trend prediction of the pumped storage unit is more accurate and the evaluation index is more comprehensive. In addition, the present invention applies the above-mentioned model to the on-line unit state monitoring system, and under the original framework of the pumped-storage power station unit state monitoring system, by adding some equipment and software interfaces (mainly including WEB servers, horizontal isolation devices, hardware firewalls and Network switch), transmit the data to the mobile terminal equipped with TN8000APP, realize the monitoring and analysis of the unit status through the mobile terminal, and push the functions of early warning and alarm and fault information in real time, monitor the unit status through the mobile terminal TN8000APP, and It can automatically analyze the vibration trend of the unit, which is convenient for the operation and maintenance personnel of the pumped storage power station to evaluate the status of the unit. Therefore, the present invention not only can accurately predict the trend, but also has the advantages of comprehensive evaluation indicators and convenient evaluation.

Description of drawings

Fig. 1 is the vibration trend prediction model of the pumped-storage unit of the present invention;

Fig. 2 is the flowchart of Apriori algorithm;

Figure 3 is a schematic diagram of the combined forecasting model.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings and embodiments, but not as a basis for limiting the present invention.

Example. A vibration trend prediction method for pumped storage units, as shown in Figure 1, the prediction method mainly includes six steps, namely: online data acquisition, data transmission, signal analysis, feature extraction, vibration correlation and quantitative analysis, and trend prediction.

Firstly, the first step is online data acquisition. Through related interface design and configuration, historical and real-time data of unit non-stationary vibration are obtained online from vibration monitoring and monitoring systems of a pumped-storage power station for signal analysis, feature extraction, status evaluation and fault analysis. Diagnostics provide information resources.

The second step of data transmission is to convert the historical and real-time data of the non-stationary vibration of the unit obtained online into the non-stationary vibration signal of the unit, and then transmit it to the user terminal through the network isolation device, WEB server, hardware firewall and power plant LAN; The acquired historical and real-time data of unit non-stationary vibration are converted into unit non-stationary vibration signals, which are sequentially transmitted to the APP on the mobile terminal through the network isolation device, WEB server, hardware firewall and Alibaba Cloud.

The third step is signal analysis. For the non-stationary vibration signal of a pumped-storage unit, empirical wavelet decomposition (EWT) is used to analyze the vibration signal in the time-frequency domain.

For a signal f(t), EWT decomposes it into N+1 mode functions f _k (t) of single frequency components.

where the AM-FM signal of f _k (t) can be defined as:

f _k (t)＝F _k (t)cos(φ _k (t))

Among them, k is a value from 0 to N, F _k is a frequency domain function, and φ _k is an angle vector.

All non-stationary signal decomposition methods hope to decompose the f _k (t) of the AM-FM component from the original signal. EWT extracts the AM-FM components of the signal by constructing a wavelet filter by adaptively dividing the Fourier spectrum. Assume that the Fourier support [0, π] is split into N consecutive parts. Each segment is represented by ∧ _n :

∧ _n = [ω _n-1 , ω _n ], n = 1, 2, ..., N

Among them, ω _n is the boundary of each section, and ω _n can be selected in many ways. Generally, the end point between two adjacent maximum points in the Fourier spectrum is taken, and ω ₀ =0, ω _n =π. With ω _n as the center, a transition section with width T _n =2τ _n is defined. According to the construction method of Meyer wavelet, a bandpass filter is constructed for each ∧ _n , and its empirical wavelet function and the empirical scaling function defined as:

in,

Assuming that the Fourier transform is recorded as F[·], and the Fourier inverse transform is recorded as F ^-1 [·]; then the empirical wavelet transform is defined according to the solution method of the traditional wavelet transform, and the detailed coefficients of the empirical wavelet transform are determined by the empirical wavelet function The inner product of Ψ _n (t) and signal f(t) is obtained:

The empirical wavelet approximation coefficients are obtained from the inner product of the signal f(t) and the scaling function Ψ ₁ (t):

in, and are the Fourier transform of empirical wavelet function and scaling function; and are the complex conjugates of the empirical wavelet function and scaling function, respectively.

The original signal can be reconstructed from the empirical wavelet function and scaling function:

Among them, * means convolution; ∧ means Fourier transform.

From this, the mode f _k (t) decomposed by the empirical wavelet can be obtained:

The fourth step is feature extraction. On the basis of signal analysis, comprehensive features such as energy entropy and singular value are extracted to describe the vibration state of the unit.

The entropy feature is an index to measure uncertainty and uncertainty. When a fault occurs, the fault part will have an impact, so the frequency response of the vibration signal relative to the vibration signal will also change, and the distribution and size of the energy will also change. Firstly, a series of modes are obtained after the signal is transformed by empirical wavelet, and then energy characteristics and energy entropy characteristics are calculated for these modes. Compared with the fault signal, the normal working signal has stronger certainty and stability, so the entropy value of the signal of the normal working condition is larger than the fault signal.

Assuming that the original signal is subjected to empirical wavelet transformation to obtain a series of modes, c _s (t), s=1, 2, ..., K, K is the number of decomposed layers, E ₁ , E ₂ , ..., E _K is each The energy value corresponding to the layer mode. The calculation process of the energy feature is as follows:

First calculate the energy signature of each layer mode:

Then calculate the total energy value for all modes:

Finally calculate the energy entropy feature;

The singular value features of all modal matrixes will also be extracted after the signal empirical wavelet transform. Singular value features are a series of eigenvectors obtained by performing singular value decomposition on a matrix. According to the principle of matrix theory, a singular value is a value that reflects the inherent properties of a matrix, and this value is stable. At the same time, singular values are scale invariant and rotation invariant. Therefore, the singular value feature is a very reliable evaluation index to distinguish different faults.

After the original signal undergoes empirical wavelet transform, the energy of each layer decreases gradually from the first layer to the last layer. When the current Q-layer mode can almost contain most of the energy information of the original signal, the former Q-layer mode is taken. state to construct the initial characteristic matrix A:

A=[c ₁ ,c ₂ ,...,c _Q ] ^T

Assume that the size of the real matrix A is PQ and satisfy the following conditions:

A＝ ^U∧VT

Where U and V are orthogonal matrices of size QQ and PP respectively, ∧ is a fault feature matrix of size QP, and all the constituent elements σ _i (i=1, 2,..., Q)(Q≤P) Arranged from low to high according to the number of modal layers. These failure matrix elements are some singular values of matrix A. Therefore, the vector [σ ₁ , σ ₂ , . . . , σ _Q ] is taken as another fault feature.

The fifth step is the correlation and quantitative analysis of vibration. The vibration signal characteristic data set of a pumped storage power station is discretized according to certain rules and then correlated with the operating conditions of the unit. The Apriori algorithm is used to mine frequent items to analyze the vibration of the unit. The spatio-temporal correlation between the data characteristics and the unit failure, and the safe operation area of the unit are divided through the correlation analysis results.

Among them, the Apriori algorithm is an effective data association analysis method for mining hidden information of data and discovering frequent itemsets. It discovers frequent itemsets by limiting candidate generation. The flow chart is shown in Figure 2. The steps of using the Apriori algorithm to mine the discretized vibration data of the pumped storage unit are as follows: First, through scanning the established pumped storage unit vibration signal characteristics and operating state analysis database, the number of occurrences of the unit’s operating state quantity is accumulated, and collected Find the set of frequent 1-itemsets for the state items that meet the set minimum support, and denote it as L1. Then, according to the Apriori property, use L1 to find out the set L2 of frequent 2-itemsets, use L2 to find out L3, and so on, until no more frequent k-itemsets can be found. Through the correlation analysis between the swing signal and the state of the unit, the safe operation area of the unit is obtained, which provides support for the identification of the safe area of the unit operation.

The sixth step is trend forecasting, constructing time series decomposition model, multiple linear regression model, and ARMA model, using time series trend forecasting method to predict its future development trend within a limited time, and then predicting and estimating the trend of unit operating status. Unit condition maintenance provides technical support.

The sixth step of unit vibration state trend prediction process is as follows: 1) Extract the time series of unit state parameters; 2) Wavelet decomposition decomposes the time series of unit state parameters into different decomposition domains (non-linear trend items and stationary fluctuation items) 3) The AR prediction model is used for the stationary fluctuation item, and the least squares support vector machine prediction model is used for the non-linear trend item; 4) The reconstructed signal realizes the unit state trend prediction. The specific implementation is as follows:

1) Construct input vector based on time series modeling of least squares support vector machine:

The original time series is a set of one-dimensional measured values. Before establishing the LSSVM model, it is necessary to reconstruct the phase space of the time series, and obtain its corresponding phase space matrix as the input vector of the LSSVM model.

Among them X is the input vector, n is a constant, k is a certain value in 1～n, m is the space dimension, τ is the delay time constant.

2) AR modeling

The stochastic difference equation (AR model) of the linear regression model is expressed as:

In the formula: is the autoregressive parameter, i=1...p, p is the order; a _t is the white noise, which means the residual.

When building an AR model, it is necessary to determine the order of the model and estimate the parameters of the model. A common criterion for determining the order of the model is the AIC criterion:

In the formula: p is the order of the model; N is the number of data; is the prediction error at different levels.

In the modeling process, the order of the model should be given first, and the parameters of the model should be estimated to obtain models of each order. Finally, the order corresponding to the first minimum value of AIC(p) should be taken to determine the optimal order of the model. Finally, it can be Determine the AR model.

Portfolio Predictive Modeling

Assuming that the state quantity sequence of the hydroelectric unit is X _t ={X ₁ …X _n }, the predictive modeling steps are as follows:

1. Utilize wavelet transform to carry out multi-layer decomposition to obtain the subsequences of each layer describing the trend item and the fluctuation item:

X=a _k +d ₁ +d ₂ +...+d _k

II. In the formula, a _k is the trend item, d _i is the fluctuation item, i=1, 2, ..., k is the number of wavelet decomposition layers.

III. Establish an LSSVM prediction model for the trend item a _k , perform model training and new value prediction, and accurately evaluate the results.

IV. Perform AR modeling for each fluctuation item d _i , i=1, 2, . . . , k, and evaluate the accuracy of the results.

V. After obtaining the prediction results of each partial sequence, perform superposition calculation to obtain the prediction sequence of the original vibration sequence Calculate the error index of the measured value and the predicted value to evaluate the accuracy. The implementation process of the high-combination forecasting model is shown in Figure 3.

VI. Evaluate the prediction performance of the results, that is, evaluate the accuracy of the prediction results, analyze the degree of deviation between the prediction results and the actual value, and use the following common error indicators to evaluate the prediction effect:

VII. Average relative error:

VIII. Root mean square error:

Where k is a value from 1 to N, is the average value.

The mobile terminal APP in the embodiment of the present invention can monitor and analyze the status of the unit through the mobile device, and simultaneously push the alarm, early warning and equipment failure information to the mobile device in real time.

In order to realize the mobile application, related equipment was added on the basis of the original unit status monitoring system of the pumped storage power station, mainly including WEB server, horizontal isolation device, hardware firewall and network switch in Zone III. The WEB server is deployed in the safe area III, and is isolated from the II area through the one-way network isolation device, and the data is transmitted from the II area to the WEB server for management and storage. The WEB server is connected to the local area network through the firewall, and the data is released to the local area network. Users on the local area network can monitor and analyze the data through special software. The WEB server is connected to the Internet through the firewall (using a dedicated line or through a local area network), connected to the dedicated server for mobile applications deployed on Alibaba Cloud, and publishes status data, alarms, and early warning information to mobile terminals through the server. Mobile terminal users can pass Dedicated APP to browse and analyze data. The data is stored on the WEB server and published through a dedicated server on Alibaba Cloud, which does not store any data. In order to ensure data security, the data transmission process is encrypted, and the data is decrypted and displayed on the mobile terminal. In order to ensure the security of the WEB server, the WEB server is only connected to the mobile application server on the Internet (realized through the firewall security policy), and connections from other computers will be considered illegal and rejected.

Mobile terminal APP functions mainly include the following aspects:

1) Real-time monitoring. TN8000APP can monitor vibration, swing, axial displacement, pressure pulsation, air gap, partial discharge and working condition parameters in real time. The system provides graphs, numerical tables, bar graphs, waveform graphs, and spectrum graphs A variety of visual charts and charts are available for users to choose, and the status of the unit can be monitored from different angles and levels, and the operating status of the unit can be grasped anytime and anywhere.

2) Data analysis and trend prediction. The system predicts and analyzes the status of the unit through the vibration trend prediction model, which is convenient for the operation and maintenance personnel of the pumped storage power station to evaluate the status of the unit.

3) Event push, the system can push the early warning and alarm information to the mobile terminal in real time, reminding the user to pay attention to the status change of the unit in time to prevent accidents. The system also provides the function of querying historical event information.

4) System self-diagnosis. Through the mobile APP, the status of the unit status monitoring system itself can be monitored in real time, including sensor status, acquisition device status, system network status, software running status, etc., and defects in the system operation process can be found in time to ensure the long-term stability of the system. ,Stable operation.

Claims (10)

1. A vibration trend prediction method for a pumping unit is characterized by comprising the following steps:

a1, acquiring online data, namely acquiring historical and real-time data of non-stationary vibration of the unit online from a pumped storage power station vibration monitoring and monitoring system by adopting a 485 communication protocol;

b1, data transmission, converting the history and real-time data of the non-stationary vibration of the unit acquired on line into a non-stationary vibration signal of the unit, and transmitting the non-stationary vibration signal to a user terminal through a network isolation device, a WEB server, a hardware firewall and a power plant local area network in sequence;

c1, analyzing signals, namely, performing time-frequency domain analysis on the non-stationary vibration signals of the unit by the user terminal through empirical wavelet decomposition on the non-stationary vibration signals of the unit;

d1, extracting features, namely extracting comprehensive features describing the vibration state of the unit from the non-stationary vibration signals of the unit on the basis of performing time-frequency domain analysis on the non-stationary vibration signals of the unit by using empirical wavelet decomposition to obtain an energy entropy and singular values;

e1, association and quantitative analysis of vibration, decomposing a vibration signal for a data set of energy entropy and singular value by using Ensemble Empirical Mode Decomposition (EEMD) or empirical wavelet decomposition (EWT) to obtain a series of single-frequency components, performing discretization treatment and then performing association analysis on the vibration signal and the operation condition of the unit, performing frequent item mining by using an Apriori algorithm, analyzing the time-space correlation between the vibration data characteristics of the unit and the fault of the unit, and dividing a safe operation area of the unit according to the association analysis result;

f1, constructing a time series decomposition model, a multiple linear regression model and an ARMA model, predicting the development trend within limited time in the future by adopting a time series trend prediction method, further predicting and evaluating the unit operation state trend, and providing technical support for implementing unit state overhaul.

2. The method of claim 1, wherein the method comprises the steps of: the time series trend prediction method in the step f1 comprises the following steps:

a2, extracting a time sequence of the state parameters of the unit;

b2, decomposing the time sequence of the unit state parameters into subsequences of different decomposition domains by using empirical wavelet decomposition to obtain a fluctuation term of stationarity and a nonlinear trend term;

c2, obtaining a signal A by adopting an AR prediction model for a fluctuation item of stationarity, and obtaining a signal B by adopting a prediction model of a least square support vector machine for a nonlinear trend item;

d2, reconstructing the signal A and the signal B to realize the prediction of the unit state trend.

3. The method of claim 1, wherein the method comprises the steps of: in the step b1, data transmission may further convert the history of the non-stationary vibration of the unit and real-time data acquired online into a non-stationary vibration signal of the unit, and the non-stationary vibration signal is transmitted to the APP on the mobile terminal through the network isolation device, the WEB server, the hardware firewall and the ali cloud in sequence.

4. The method of claim 2, wherein the method comprises the steps of: the prediction model of the least square support vector machine is expressed as

Where X is the input vector, n is a constant, k is one of values 1-n, m is the spatial dimension, and τ is the delay time constant.

5. The method of claim 2, wherein the method comprises the steps of: the AR model is represented as

In the formula:is an autoregressive parameter, i is 1 … p, and p is an order; a is_tIs white noise, representing the residual.

6. The method of claim 5, wherein the method comprises the steps of: and determining the order of the model through an AIC (automatic aided objective) criterion and performing parameter estimation on the model when the AR model is modeled.

7. The method of claim 2, wherein the method comprises the steps of: method for reconstructing a signal in step d2, comprising the steps of:

a3, performing multilayer wavelet decomposition on the time sequence of the unit state parameters to obtain each layer of subsequence describing a trend term and a fluctuation term:

X＝a_k+d₁+d₂+…+d_k

in the formula, a_kAs a trend term, d_iFor the fluctuation term, i is 1, 2, …, k is the wavelet decomposition layer number;

b3 for trend item a_kEstablishing an LSSVM prediction model, carrying out model training and new value prediction, and accurately evaluating the result;

c3, for each fluctuation term d_iThe results are evaluated in terms of accuracy by performing AR modeling on the i-1, 2, … and k respectively;

d3, obtaining the prediction results of the trend item and each fluctuation item, and then carrying out superposition calculation to obtain the prediction sequence of the original vibration sequenceCalculating error indexes of the measured value and the predicted value, and performing precision evaluation;

e3, evaluating the precision of the prediction result of the original vibration sequence, analyzing the deviation degree of the prediction result and the actual value, and adopting the following common error indexes for evaluating the prediction effect.

8. The method of claim 7, wherein the method comprises the steps of: the common error index comprises

Average relative error:and

root mean square error:

where k is a value from 1 to N,are averages.

9. The method of claim 1, wherein the method comprises the steps of: the signal analysis is expressed as, for a signal f (t), EWT decomposes it into N +1 modal functions f of single frequency components_k(t)。

Wherein f is_kThe AM-FM signal (AM-FM) of (t) may be defined as:

f_k(t)＝F_k(t)cos(φ_k(t))

wherein k is a value of 0 to N, F_kAs a function of the frequency domain, phi_kIs an angular vector.

Decomposition of non-stationary signals is desirable to be able to decompose f of the AM-FM component from the original signal_k(t) of (d). EWT extracts the AM-FM component of a signal by an adaptive split fourier spectrum construction wavelet filter. Suppose that Fourier is supported by [0, π [ ]]Divided into N successive portions. Each section is made of ^_nRepresents:

∧_n＝[ω_n-1，ω_n]，n＝1，2，...，N

wherein, ω is_nFor the boundary of each segment, ω_nThe selection mode is many, and the terminal point between two adjacent maximum value points in the Fourier spectrum is generally taken, and omega₀＝0，ω_nPi. At omega_nDefining a width T for the center_n＝2τ_nThe transition section of (1). For each lambda according to the construction mode of Meyer wavelet_nConstruction ofBandpass filter of empirical wavelet functionAnd empirical scale functionIs defined as:

wherein,

β(x)＝x⁴(35-84x+70x²-20x³)。

suppose the Fourier transform is denoted as F [ ·]Inverse Fourier transform is denoted F^-1[·](ii) a Then defining empirical wavelet transform according to solving mode of traditional wavelet transform, and making detail coefficient of empirical wavelet transform be defined by empirical wavelet function psi_n(t) inner product of signal f (t):

empirical wavelet approximation coefficient is formed by signal f (t) and scale function Ψ₁(t) obtaining an inner product of:

wherein,andfourier transform of an empirical wavelet function and a scale function, respectively;andrespectively, the complex conjugate of the empirical wavelet function and the scale function.

The original signal can be reconstructed from an empirical wavelet function and a scale function:

wherein denotes a convolution; Λ represents the fourier transform.

From this, the mode f decomposed by the empirical wavelet can be obtained_k(t)：

。

10. The method of claim 1, wherein the method comprises the steps of: the feature extraction is represented by assuming that the original signal is subjected to empirical wavelet transform to obtain a series of modes, c_s(t), s ═ 1, 2, …, K, K being the number of layers after decomposition, E₁,E₂,…，E_KIs the energy value corresponding to each layer mode. The energy characteristics are calculated as follows:

firstly, calculating the energy characteristics of each layer of mode:

the total energy value for all modes is then calculated:

finally, calculating energy entropy characteristics;

after signal experience wavelet transformation, singular value characteristics of all modes forming a matrix are extracted. The singular value feature is a series of eigenvectors obtained by performing singular value decomposition on the matrix. According to the matrix theory principle, a singular value is a value that reflects an intrinsic property of a matrix and has stability. Meanwhile, singular values have scale invariance and rotation invariance. Therefore, the singular value characteristic is a very reliable evaluation index for distinguishing different faults

After the original signal is subjected to empirical wavelet transform, the energy of each layer decreases from the first layer to the last layer, and when the current Q layer mode almost can contain most energy information of the original signal, the front Q layer mode is taken to construct an initial characteristic matrix A:

A＝[c₁，c₂，...，c_Q]^T

the size of the real number matrix A is assumed to be PQ, and the following conditions are satisfied:

A＝U∧V^T

wherein U and V are orthogonal matrix of QQ and PP, respectively, Λ is a fault feature matrix of QP, and all constituent elements σ_i(i ═ 1, 2., Q) (Q ≦ P) is arranged in low to high order number of modal layers. These failure matrix elements are some of the singular values of matrix a. Thus will vector [ sigma ]₁，σ₂，...，σ_Q]As another fault feature.