CN111291918A - Rotating machine degradation trend prediction method based on stationary subspace exogenous vector autoregression - Google Patents

Abstract

The invention discloses a method for predicting the degradation trend of rotating machinery by means of autoregressive external source vector in stationary subspace. The time and frequency domain degradation feature quantities are extracted and the high-dimensional degradation index vector group is obtained through feature fusion; then the high-dimensional degradation index vector group in the time and frequency domains is subjected to the second stationary subspace decomposition and difference operation to extract weak degradation indicators. The stationary component is used as the degradation index of rotating machinery; the degradation index is subjected to stationarity test, impulse response analysis, and the endogenous and exogenous variables and model order are determined, and the vector autoregressive model parameters are determined by maximum likelihood estimation. Estimates of degradation trends under the forecast starting point. The degradation trend prediction model obtained by the invention not only has good generalization ability under small sample learning, but also has rapid calculation and strong interpretability.

Description

Degradation Trend Prediction Method of Rotating Machinery Based on Exogenous Vector Autoregression in Stationary Subspace

technical field

The invention relates to the technical field of degradation trend prediction in rotating machinery equipment, and is a method for predicting the degradation trend of rotating machinery based on stationary subspace exogenous vector autoregression. Vibration degradation index and a method for predicting degradation trend through exogenous vector autoregression.

Background technique

Continued aging and increasing demand in the operation of rotating machinery call for more advanced failure prediction and health management techniques, in which degradation trend prediction plays a crucial role in the complex engineering systems that assemble rotating machinery. Accurate degradation trend prediction methods can provide machine status information and health status in advance for predictive maintenance, thereby avoiding the risk of sudden downtime and equipment failure, and improving the overall profitability of industrial production.

In order to effectively solve the problem of predicting the degradation trend of rotating machinery, methods based on physical and mechanical models have been widely studied, but such methods require the support of perfect prior knowledge, which is difficult to meet and develop in some complex systems. As another emerging technology, data-driven solutions are widely used in lithium-ion battery failure prediction and health management technology, large-scale industrial process monitoring and rotating machinery failure prediction and health management technology, which can directly generate degradation or life through sampling data. Models flexibly model unknown objects without prior knowledge. It offers the possibility of accurate prediction and evaluation in complex situations such as variable operating conditions and system-level prediction. Artificial intelligence technology led by deep learning has developed rapidly in recent years, which has greatly promoted the promotion of data-driven methods in the field of rotating machinery degradation trend prediction. Such technologies have excellent feature learning and representation capabilities, and most prediction tasks are performed in Performs well with ample and complete labeled data. Unfortunately, this learning-generative model-based prediction architecture is more or less dependent on the quality and quantity of learning samples, and the prediction results are limited by the similarity between the learning and testing datasets. That is to say, this method may fail for some high-end application scenarios of "insufficient data". Although transfer learning techniques can alleviate this problem to a certain extent, the lack of an interpretable modeling process and a sound mathematical and statistical foundation to support it is still a thorny problem.

In addition to the above-mentioned data-driven research based on artificial intelligence technology, data-driven prediction methods based on regression analysis have also been widely studied in the field of rotating machinery degradation trend prediction. sex. However, most regression analysis data-driven forecasting methods have a static structure inside, which greatly limits the extrapolation and generalization from time t to time t+n. Autoregressive theory is widely used in river, geography, economy and other fields as a prediction and forecast model with natural extrapolation structure and perfect mathematical theoretical basis. Such problems usually have a certain periodicity and the data is relatively stable, which is conducive to autoregressive analysis. The degradation process of rotating machinery contains a large number of non-stationary and nonlinear components, which greatly hinders the application and promotion of this theory in this field. Relevant literature reports appear.

In view of the advantages of autoregressive forecasting and forecasting models, as well as the defects existing in the prediction of the degradation trend of rotating machinery, how to deal with non-stationary vibration signals to meet the requirements of autoregressive theory for weak stationarity is a problem that needs to be solved. Starting from the subspace decomposition theory, we perform signal processing on non-stationary and nonlinear vibration degradation signals, and predict the degradation trend through the exogenous vector autoregressive model.

SUMMARY OF THE INVENTION

The purpose of the present invention is to propose a method for predicting the degradation trend of rotating machinery for stationary subspaces-vector autoregressive with exogenous terms (SSVARX). Using two stationary subspace decomposition methods and difference operations, the weak stationary components in the degradation indices in the time domain and frequency domain are extracted as the degradation indices of rotating machinery, respectively. and the model order, then determine the vector autoregressive model parameters through maximum likelihood estimation, and finally estimate the degradation trend of rotating machinery under different prediction starting points. It effectively overcomes the problems of weak generalization ability, long calculation time and "black box effect" of rotating machinery degradation trend prediction under the current small sample, and has important economic and social value.

According to a method for predicting the degradation trend of rotating machinery with stationary subspace exogenous vector autoregression, the method includes the following steps:

Step 1: Use the vibration accelerometer to collect multi-channel signals for the sensitive degradation position of the rotating machinery, and perform wavelet noise reduction on the collected vibration signals to eliminate high-frequency components in the original signals;

Step 2: Perform the first stationary subspace decomposition of the denoised multi-channel vibration signal to synthesize multi-channel degradation and damage information and extract the weak stationary component of vibration, mainly through stationary subspace analysis (SSA) algorithm and The vibration signal after multi-channel noise reduction is realized, namely:

表示通道一采集的所有观测值；

和

分别表示分解后的振动弱平稳与非平稳分量。in the formula

Indicates all observations collected by channel 1;

and

represent the weakly stationary and non-stationary components of the decomposed vibration, respectively.

Step 3: Extract the time-domain and frequency-domain degradation features from the generated weakly stationary components of vibration, and perform feature fusion on the matrix formed by the degradation features in each domain through principal component analysis to generate a time-domain and frequency-domain high-dimensional degradation index vector group, It includes the following specific steps: extraction of degradation features in time domain and frequency domain, and feature fusion based on principal component analysis,

标准差：

平方根振幅：

绝对平均值：

偏度：

峰度：

方差：

最大值：DF₈＝max|x(n)|、最小值：DF₉＝min|x(n)|、峰均值：DF₁₀＝DF₈-DF₉、均方根：

波形指数：

峰值指数：

脉冲指数：

裕度指数：

偏度指数：

和峰度指数：

频域特征提取采用的统计学参数公式如下：

以及

其中y(k)是给定信号的快速傅里叶频谱，f_k则对应于第k个频谱的频率值，DF₁₈在频域上反映振动能量，DF₁₉～DF₂₁、DF₂₃和DF₂₇～DF₃₀描述频谱的集中和离散程度，DF₂₂和DF₂₄～DF₂₆表示主频带的位置变化。Step 3.1: The statistical parameter formula used in time domain feature extraction is as follows: Average value:

Standard deviation:

Square root amplitude:

Absolute mean:

Skewness:

Kurtosis:

variance:

Maximum value: DF ₈ =max|x(n)|, minimum value: DF ₉ =min|x(n)|, peak mean value: DF ₁₀ =DF ₈ -DF ₉ , root mean square:

Waveform Index:

Peak Index:

Pulse Index:

Margin Index:

Skewness Index:

and the kurtosis index:

The statistical parameter formula used in frequency domain feature extraction is as follows:

as well as

where y(k) is the fast Fourier spectrum of a given signal, f _k corresponds to the frequency value of the k-th spectrum, DF ₁₈ reflects vibrational energy in the frequency domain, DF ₁₉ ～DF ₂₁ , DF ₂₃ and DF ₂₇ ~DF ₃₀ describes the degree of concentration and dispersion of the spectrum, DF ₂₂ and DF ₂₄ ~ DF ₂₆ represent the positional changes of the main frequency band.

Step 3.2: The extracted degradation features in time domain and frequency domain respectively form feature matrices DF _1-17 and DF _18-30 for feature fusion by principal component analysis;

Step 3.3: Use principal component analysis to perform feature fusion on DF _1-17 and DF _18-30 , and extract principal components that satisfy the predetermined contribution rate to form a high-dimensional degradation index vector group, namely

以及

分别表示特征向量，

以及

分别表示符合预定贡献度值的时域、频域主成分，即高维退化指标向量组。in the formula

as well as

are the feature vectors, respectively,

as well as

respectively represent the time-domain and frequency-domain principal components conforming to the predetermined contribution value, that is, the high-dimensional degradation index vector group.

Step 4: Perform the second stationary subspace decomposition on the high-dimensional degradation index vector group in the time domain and frequency domain, extract the weakly stationary component in the degradation index vector group, and perform the differential operation based on spectrum analysis to generate the time domain, The frequency domain degradation index, the basic steps can be summarized as follows:

Step 4.1: Perform the second stationary subspace decomposition on the high-dimensional degradation index vector group through the stationary subspace decomposition algorithm, namely:

表示来自第一主元分量的所有时刻退化指标向量，

和

表示时域高维退化指标向量组的弱平稳与非平稳成分；

和

表示频域高维退化指标向量组的弱平稳与非平稳成分。in the formula

represents the degradation index vector at all times from the first principal component,

and

Represents the weakly stationary and non-stationary components of the time-domain high-dimensional degradation index vector group;

and

Represents the weakly stationary and non-stationary components of a high-dimensional degradation index vector group in the frequency domain.

Step 4.2: According to the generated high-dimensional degradation index vector group in the time domain and frequency domain, calculate the required differential step size through spectrum analysis. Degradation indicator to input the next degradation trend prediction model.

Step 5: Finally, input the time domain and frequency domain degradation indicators into the external vector autoregressive model to predict the degradation trend of the rotating machinery. It includes five parts: defining exogenous and endogenous variables, inputting the stationarity test of the degradation index, determining the model order, impulse response analysis, parameter estimation and degradation trend prediction. The basic steps can be summarized as follows:

Step 5.1: Define the weak stationary component of the high-dimensional degradation index vector group without significant fluctuations in each degradation stage in the life cycle as endogenous variables, and the rest are exogenous variables to assist in the prediction of degradation trend;

Step 5.2: Test for stationarity by augmenting Dickey Fuller's method to detect whether the selected input quantity has a unit root: H ₀ is assumed to be as follows, namely:

H ₀ : y _t =c+y _t-1 +β ₁ Δy _t-1 +β ₂ Δy _t-2 +...+β _p Δy _p-1 +ε _t

H ₁ is assumed to be as follows, namely:

H ₀ : y _t =c+dt+θy _t-1 +β ₁ Δy _t-1 +β ₂ Δy _t-2 +…+β _p Δy _p-1 +ε _t

In the formula, θ<1, [β ₁ ,…,β _p ] and d are the coefficients of the regression term and the trend term, respectively, ε _t represents the random error, and c is the constant term. Then, the range of the order interval with stationary characteristics is preliminarily determined by this method: first, determine the interval without a unit root, that is, reject all suitable intervals of order H ₀ , and define li and _ui as the upper and lower bounds of the _ith interval, respectively ; Then perform the intersection operation on all intervals, namely: [l ₁ ,…,u ₁ ]∪,…,∪[l _i ,…,u _i ]∪,…,∪[l _n ,…,u _n ], thus Determines the range of order intervals with stationary properties.

Step 5.3: Calculate the value of the Akaike information criterion in the order interval with stationary characteristics, namely:

Search for the order corresponding to the minimum value of the above formula, which is defined as the model order;

Step 5.4: Use impulse response analysis to further analyze the endogenous variables, the impact of the disturbance of the time domain and frequency domain degradation indicators through the vector autoregressive model lag structure, and check the impact of the shock on all endogenous variables from a short-term or long-term perspective;

Step 5.5: Use maximum likelihood estimation to estimate the parameters of the exogenous vector autoregressive model, and then use the model to predict the degradation trend under different starting points to test the generalization ability of the model.

The beneficial effects of the present invention are:

1. The SSVARX degradation trend prediction method proposed by the present invention is the first attempt to apply the vector autoregressive theory to the field of rotating machinery, and provides a perfect mathematical theoretical basis for data-driven prediction and prediction research;

2. The degradation trend prediction method proposed by the present invention not only has a natural extrapolation structure, fast calculation speed, but also provides high-precision prediction results in the case of small samples, which is very suitable for some high-end application scenarios with scarce state data;

3. The degradation trend prediction method proposed in the present invention provides a feasible way for the non-stationary vibration signal to be converted into a weakly stationary degradation index, and at the same time considers the potential causal relationship and relationship between endogenous variables from different fields, so as to further Improve the accuracy of degradation trend prediction.

Description of drawings

Fig. 1 is the implementation flow chart of the method of the present invention.

Fig. 2 is a multi-channel vibration original signal collected in the present invention.

Figure 3 shows the first stationary subspace decomposition components obtained by the method of the present invention.

FIG. 4 is the time domain and frequency domain feature extraction of the weak stationary component of the vibration signal in the present invention.

FIG. 5 shows HDDIVs in time domain and frequency domain in the present invention.

Fig. 6 is the degradation index after the second stationary subspace decomposition and difference operation obtained by the method of the present invention.

FIG. 7 is an impulse response analysis result between degradation indexes in the present invention.

FIG. 8 is the prediction result of the degradation trend of the rolling bearing under different actual prediction points obtained by the method of the present invention.

FIG. 9 is a comparison diagram of the degradation trend prediction result obtained by the method of the present invention and the comparative tests A, B, and C. FIG.

Detailed ways

The present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments.

A method for predicting the degradation trend of rotating machinery by exogenous vector autoregression in stationary subspace is shown in Figure 1. The steps can be summarized as follows:

Step 1. In this example, HRB6308 rolling bearing is used, and the ABLT-1A bearing life-enhancing testing machine is used to conduct a full-life fatigue accelerated test. Firstly, two-channel signal acquisition is carried out on the sensitive degradation position of rotating machinery through PCB 608A11 vibration accelerometer and National Instruments 9234 data acquisition card. The original signal is shown in Figure 2, and the collected vibration signal is subjected to wavelet noise reduction to eliminate high-frequency components in the original signal. ;

Step 2. Perform the first stationary subspace decomposition of the multi-channel vibration signal after noise reduction through a stationary subspace analysis (stationary subspace analysis, SSA) algorithm and the multi-channel vibration signal after noise reduction, and synthesize the multi-channel degradation and damage information And extract the weak stationary component of vibration, see Figure 3 for details. which is:

表示通道一采集的所有观测值；

和

分别表示分解后的振动弱平稳与非平稳分量。in the formula

Indicates all observations collected by channel 1;

and

represent the weakly stationary and non-stationary components of the decomposed vibration, respectively.

Step 3: Extract 17 time-domain DF _1-17 and 13 frequency-domain degradation features DF _18-30 from the generated weakly stationary components of vibration, and the normalized features are shown in Figure 4 . Then, through principal component analysis, the matrix composed of degenerate features of each domain, namely: [DF ₁ , DF ₂ , ..., DF ₁₇ ] and [DF ₁₈ , DF ₁₉ , ..., DF ₃₀ ], perform feature fusion, where The contribution rate threshold is set to 98% to ensure that degradation and damage information is covered to a greater extent. Figure 5 shows the generated high-dimensional degradation indicator vectors (HDDIVs) in time domain and frequency domain. Specifically, it includes the following steps: extraction of degradation features in time and frequency domains, and feature fusion based on principal component analysis,

标准差：

平方根振幅：

绝对平均值：

偏度：

峰度：

方差：

最大值：DF₈＝max|x(n)|、最小值：DF₉＝min|x(n)|、峰均值：DF₁₀＝DF₈-DF₉、均方根：

波形指数：

峰值指数：

脉冲指数：

裕度指数：

偏度指数：

和峰度指数：

频域特征提取采用的统计学参数公式如下：

以及

其中y(k)是给定信号的快速傅里叶频谱，f_k则对应于第k个频谱的频率值，DF₁₈在频域上反映振动能量，DF₁₉～DF₂₁、DF₂₃和DF₂₇～DF₃₀描述频谱的集中和离散程度，DF₂₂和DF₂₄～DF₂₆表示主频带的位置变化。Step 3.1. The statistical parameter formula used in time domain feature extraction is as follows: Average value:

Standard deviation:

Square root amplitude:

Absolute mean:

Skewness:

Kurtosis:

variance:

Maximum value: DF ₈ =max|x(n)|, minimum value: DF ₉ =min|x(n)|, peak mean value: DF ₁₀ =DF ₈ -DF ₉ , root mean square:

Waveform Index:

Peak Index:

Pulse Index:

Margin Index:

Skewness Index:

and the kurtosis index:

The statistical parameter formula used in frequency domain feature extraction is as follows:

as well as

where y(k) is the fast Fourier spectrum of a given signal, f _k corresponds to the frequency value of the k-th spectrum, DF ₁₈ reflects vibrational energy in the frequency domain, DF ₁₉ ～DF ₂₁ , DF ₂₃ and DF ₂₇ ~DF ₃₀ describes the degree of concentration and dispersion of the spectrum, and DF ₂₂ and DF ₂₄ to DF ₂₆ represent the positional changes of the main frequency band.

Step 3.2, the extracted degradation features in time domain and frequency domain respectively form feature matrices DF _1-17 and DF _18-30 for feature fusion by principal component analysis;

Step 3.3. Use principal component analysis to perform feature fusion on DF _1-17 and DF _{18-30 to} extract principal components that satisfy the predetermined contribution rate to form a high-dimensional degradation index vector group, namely

以及

分别表示特征向量，

以及

分别表示符合预定贡献度值的时域、频域主成分，即高维退化指标向量组。in the formula

as well as

are the feature vectors, respectively,

as well as

respectively represent the time-domain and frequency-domain principal components conforming to the predetermined contribution value, that is, the high-dimensional degradation index vector group.

Step 4. Perform the second stationary subspace decomposition on the high-dimensional degradation index vector group in the time domain and frequency domain, extract the weakly stationary components DI _Time and DI _Frequency in the degradation index vector group, and perform the differential operation based on spectrum analysis to generate a weak stationary component. The required time-domain and frequency-domain degradation indicators. The basic steps can be summarized as follows:

Step 4.1. Perform the second stationary subspace decomposition on the high-dimensional degradation index vector group through the stationary subspace decomposition algorithm, namely:

表示来自第一主元分量的所有时刻退化指标向量，

和

表示时域高维退化指标向量组的弱平稳与非平稳成分；

和

表示频域高维退化指标向量组的弱平稳与非平稳成分。in the formula

represents the degradation index vector at all times from the first principal component,

and

Represents the weakly stationary and non-stationary components of the time-domain high-dimensional degradation index vector group;

and

Represents the weakly stationary and non-stationary components of a high-dimensional degradation index vector group in the frequency domain.

Step 4.2. According to the generated high-dimensional degradation index vector group in the time domain and frequency domain, calculate the required differential step size through spectrum analysis. The differential step size is the reciprocal of the frequency corresponding to the main peak of the spectrum analysis, and then perform a differential operation to generate a periodic Degradation indicator and input into the next degradation trend prediction model.

Step 5. Finally, input the degraded indexes in the time domain and frequency domain after the difference into the exogenous vector autoregressive model to predict the degradation trend of the rotating machinery. It includes five parts: defining exogenous and endogenous variables, inputting the stationarity test of the degradation index, determining the model order, impulse response analysis, parameter estimation and degradation trend prediction. The basic steps can be summarized as follows:

Step 5.1. Define the weak stationary component of the high-dimensional degradation index vector group without significant fluctuations in each degradation stage in the life cycle as the endogenous variable, as shown in Figure 6, and the rest are exogenous variables to assist the degradation trend prediction;

Step 5.2. Check stationarity by augmenting the Dickey Fuller method to detect whether the selected input quantity has a unit root: H ₀ is assumed to be as follows, namely:

H ₀ : y _t =c+y _t-1 +β ₁ Δy _t-1 +β ₂ Δy _t-2 +...+β _p Δy _p-1 +ε _t

H ₁ is assumed to be as follows, namely:

H ₀ : y _t =c+dt+θy _t-1 +β ₁ Δy _t-1 +β ₂ Δy _t-2 +…+β _p Δy _p-1 +ε _t

In the formula, θ<1, [β ₁ ,…,β _p ] and d are the coefficients of the regression term and the trend term, respectively, ε _t represents the random error, and c is the constant term. Then, the range of the order interval with stationary characteristics is preliminarily determined by this method: first, determine the interval without a unit root, that is, reject all suitable intervals of order H ₀ , and define li and _ui as the upper and lower bounds of the _ith interval, respectively ; Then perform the intersection operation on all intervals, namely: [l ₁ ,…,u ₁ ]∪,…,∪[l _i ,…,u _i ]∪,…,∪[l _n ,…,u _n ], thus Determines the range of order intervals with stationary properties.

Step 5.3, calculate the Akaike information criterion value in the order interval with stationary characteristics, namely:

The order corresponding to the minimum value of the search formula is defined as the model order, and the unit root test results are shown in Table 1;

Table 1, the unit root test results of the degradation index in the frequency domain

variable t-statistic delay order 1% threshold 5% threshold 10% threshold in conclusion DI_Time -3.6143 60 -2.5695 -1.9415 -1.6166 smooth DI_Frequency -2.9781 60 -2.5695 -1.9415 -1.6166 smooth

Step 5.4. Use impulse response analysis to further analyze the endogenous variables, the impact of the disturbance of the time domain and frequency domain degradation indicators through the vector autoregressive model lag structure, and check the impact of the shock on all endogenous variables from a short-term or long-term perspective, specific results As shown in Figure 7;

Step 5.5, using maximum likelihood estimation to estimate the parameters of the exogenous vector autoregressive model, the stationary subspace exogenous vector autoregressive degradation trend prediction model proposed by the present invention is as follows. Then the model was used to predict the degradation trend under different starting points to test the generalization ability of the model. The prediction results are shown in Figure 8.

Step 6. In order to highlight the effectiveness and necessity of the method of the present invention, a comparative experiment is constructed: A, lacking the first stationary subspace decomposition; B, lacking the second stationary subspace decomposition and C, based on the classical autoregressive model And lack of the quadratic stationary subspace decomposition in the method of the present invention, the three comparative experiments are respectively carried out 20 times to predict the degradation trend under different prediction starting points. It can be seen from Figure 9 and Table 2 that the quadratic stationary subspace decomposition and the vector autoregressive modeling method in the method proposed by the present invention can effectively improve the accuracy of the degradation trend and have significant engineering application value.

Table 2 Comparative test of the method of the present invention

Average forecast error over twenty forecasts Comparative test A 0.2402 Comparative test B 0.1052 Comparative test C 0.4104 method of the invention 0.1038

Step 7. In order to highlight the advantages of the method of the present invention compared with other existing prediction technologies, respectively construct predictions of Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) under different prediction starting points The details are shown in Table 3: the method of the present invention has lower prediction errors under different prediction starting points, and the calculation time is much lower than that of the LSTM and GRU models.

Table 3 Comparison of the method of the present invention and other methods

The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art who is familiar with the technical scope disclosed by the present invention can easily think of changes or substitutions. All should be covered within the protection scope of the present invention.

Claims (8)

1. a rotating machinery degradation trend prediction method of stationary subspace exogenous vector autoregression, is characterized in that, comprises the following steps: Step 1: Perform multi-channel signal acquisition on the sensitive degradation position of the rotating machinery through the vibration accelerometer, and perform noise reduction processing on the collected vibration signals; Step 2: Perform the first stationary subspace decomposition on the multi-channel vibration signal after noise reduction to synthesize multi-channel degradation and damage information and extract the weak stationary component of vibration; Step 3: Extract the degradation features in the time domain and the frequency domain from the generated weakly stationary components of vibration, and perform feature fusion on the matrix formed by the degradation features in each domain through principal component analysis to generate a high-dimensional degradation index vector group in the time domain and frequency domain; Step 4: Perform the second stationary subspace decomposition on the high-dimensional degradation index vector group in the time domain and frequency domain, extract the weakly stationary component in the degradation index vector group, and perform the differential operation based on spectrum analysis to generate the time domain, Frequency domain degradation index; Step 5: Finally, input the time domain and frequency domain degradation indicators into the external vector autoregressive model to predict the degradation trend of the rotating machinery. 2. a kind of stationary subspace external source vector autoregressive autoregressive rotary machinery degradation trend prediction method according to claim 1, is characterized in that: in the described step 1, the noise reduction processing to the collected vibration signal is through the wavelet drop. Noise removes high frequency components in the original signal. 3. a kind of stationary subspace external source vector autoregressive rotary machinery degradation trend prediction method according to claim 1, is characterized in that: in described step 2, the first stationary subspace decomposition is by stationary subspace analysis algorithm And the realization of multi-channel vibration signal after noise reduction, namely:

in the formula

Indicates all observations collected by channel 1;

and

represent the weakly stationary and non-stationary components of the decomposed vibration, respectively. 4. a kind of rotating machinery degradation trend prediction method of stationary subspace exogenous vector autoregression according to claim 1, is characterized in that: in described step 3, the construction of time domain, frequency domain high-dimensional degradation index vector group comprises The specific steps are as follows: extraction of degradation features in time domain and frequency domain, and feature fusion based on principal component analysis, Step 3.1: The statistical parameter formula used in time domain feature extraction is as follows: Average value:

Standard deviation:

Square root amplitude:

Absolute mean:

Skewness:

Kurtosis:

variance:

Maximum value: DF ₈ =max|x(n)|, minimum value: DF ₉ =min|x(n)|, peak mean value: DF ₁₀ =DF ₈ -DF ₉ , root mean square:

Waveform Index:

Peak Index:

Pulse Index:

Margin Index:

Skewness Index:

and the kurtosis index:

The statistical parameter formula used in frequency domain feature extraction is as follows:

as well as

where y(k) is the fast Fourier spectrum of a given signal, f _k corresponds to the frequency value of the k-th spectrum, DF ₁₈ reflects vibrational energy in the frequency domain, DF ₁₉ ～DF ₂₁ , DF ₂₃ and DF ₂₇ ~DF ₃₀ describes the degree of concentration and dispersion of the spectrum, DF ₂₂ and DF ₂₄ ~ DF ₂₆ represent the positional changes of the main frequency band. Step 3.2: The extracted degradation features in time domain and frequency domain respectively form feature matrices DF _1-17 and DF _18-30 for feature fusion by principal component analysis; Step 3.3: Use principal component analysis to perform feature fusion on DF _1-17 and DF _18-30 , and extract principal components that satisfy the predetermined contribution rate to form a high-dimensional degradation index vector group, namely

in the formula

as well as

are the feature vectors, respectively,

as well as

respectively represent the time-domain and frequency-domain principal components conforming to the predetermined contribution value, that is, the high-dimensional degradation index vector group. 5. The method for predicting the degradation trend of rotating machinery by autoregressive external source vectors in a stationary subspace according to claim 1, wherein the second stationary subspace decomposition in the step 4 is performed by the stationary subspace. The decomposition algorithm and the realization of the high-dimensional degradation index vector group, the basic steps are as follows: Step 4.1: Perform the second stationary subspace decomposition on the high-dimensional degradation index vector group through the stationary subspace decomposition algorithm, namely:

in the formula

represents the degradation index vector at all times from the first principal component,

and

Represents the weakly stationary and non-stationary components of the time-domain high-dimensional degradation index vector group;

and

Represents the weakly stationary and non-stationary components of a high-dimensional degradation index vector group in the frequency domain. Step 4.2: According to the generated high-dimensional degradation index vector group in the time domain and frequency domain, calculate the required difference step size through spectrum analysis, and perform a difference operation to generate a periodic degradation index to input into the next degradation trend prediction model. 6 . The method for predicting the degradation trend of rotating machinery by autoregressive external source vector in stationary subspace according to claim 5 , characterized in that: in the step 4.2 , the difference step size is the reciprocal of the frequency corresponding to the main peak of the spectrum analysis. 7 . 7. a kind of stationary subspace exogenous vector autoregressive autoregressive rotary machinery degradation trend prediction method according to claim 1, is characterized in that: in described step 5, the construction of exogenous vector autoregressive degradation trend prediction model comprises: Source, endogenous variables, input degradation index stationarity test, model order determination, impulse response analysis, parameter estimation and degradation trend prediction five parts, the basic steps are as follows: Step 5.1: Define the weak stationary component of the high-dimensional degradation index vector group without significant fluctuations in each degradation stage in the life cycle as endogenous variables, and the rest are exogenous variables to assist in the prediction of degradation trend; Step 5.2: Test for stationarity by augmenting Dickey Fuller's method to detect whether the selected input quantity has a unit root: H ₀ is assumed to be as follows, namely: H ₀ : y _t =c+y _t-1 +β ₁ Δy _t-1 +β ₂ Δy _t-2 +...+β _p Δy _p-1 +ε _t H ₁ is assumed to be as follows, namely: H ₀ : y _t =c+dt+θy _t-1 +β ₁ Δy _t-1 +β ₂ Δy _t-2 +…+β _p Δy _p-1 +ε _t In the formula, θ<1, [β ₁ ,…,β _p ] and d are the coefficients of the regression term and the trend term, respectively, ε _t represents the random error, and c is the constant term. Then, the order interval range with stationary characteristics is preliminarily determined by this method; Step 5.3: Calculate the value of the Akaike information criterion in the order interval with stationary characteristics, namely:

Search for the order corresponding to the minimum value of the above formula, which is defined as the model order; Step 5.4: Use impulse response analysis to further analyze the endogenous variables, the impact of the disturbance of the time domain and frequency domain degradation indicators through the vector autoregressive model lag structure, and check the impact of the shock on all endogenous variables from a short-term or long-term perspective; Step 5.5: Use maximum likelihood estimation to estimate the parameters of the exogenous vector autoregressive model, and then use the model to predict the degradation trend under different starting points to test the generalization ability of the model. 8. The method for predicting the degradation trend of rotating machinery by autoregression of a stationary subspace exogenous vector according to claim 7, wherein the method for determining the order interval with stationary characteristics in the step 5.2 is as follows: Including the unit root, that is, rejecting all suitable intervals of H ₀ , and defining li and u _i as the upper and lower bounds of the _ith interval, respectively; and then perform the intersection operation on all intervals, namely: [l ₁ ,...,u ₁ ]∪,…,∪[l _i ,…,u _i ]∪,…,∪[l _n ,…,u _n ], so as to determine the order interval range with stationary characteristics.

Images