A dynamic detection method for early faults in hydraulic systems based on typical variable analysis

Abstract

The present application relates to a method for dynamic detection of early faults of hydraulic systems based on typical variable analysis. The method includes: collecting hydraulic system working parameters of the hydraulic system as a sample data set, performing standardized data preprocessing operations on the sample data set to obtain the preprocessed sample data set, constructing two sets of observation vectors adjacent at time k and the observation matrix corresponding to the two sets of observation vectors according to the preprocessed sample data set, analyzing the covariance and cross-covariance of the observation matrix, and performing singular value decomposition of the Hankel matrix, analyzing the singular value decomposition results and covariance, determining the projection matrix, dividing it into two orthogonal subspaces, constructing a state subspace with correlation, and the rest is a residual subspace, analyzing Hotelling's T at time point k 2 The value of and Q statistic are related to the preset threshold value T UCL 2 and Q UCL By comparison, it is determined whether the current hydraulic system is in a fault state, thereby improving the accuracy of early fault detection of the dynamic hydraulic system.

Description

Technical Field

[0001] The present application relates to the technical field of hydraulic system fault detection, and in particular to a method for dynamic detection of early hydraulic system faults based on typical variable analysis. Background Art

[0002] In modern industry, hydraulic systems have an increasingly profound impact on various engineering equipment. Due to their simple structure, convenient operation and strong output capacity, they have been widely used in manufacturing, transportation, construction, energy, agriculture and other fields, such as production platforms, vehicle-mounted drilling rigs and mining supports. In actual applications, hydraulic systems often work under a variety of complex working conditions, which greatly increases the possibility of various types of failures. Therefore, advanced condition monitoring technology plays a vital role in the reliable operation of hydraulic systems. It can promptly detect abnormal deviations caused by component failures and guide staff to take effective maintenance measures to keep the hydraulic system in a high-performance state.

[0003] According to the fault evolution characteristics of hydraulic systems, they can be divided into three types of faults: sudden, intermittent and initial. At present, most of the existing research results focus on monitoring sudden and intermittent faults in hydraulic systems, while there is still a great lack of research on early fault detection methods, which brings great hidden dangers to the normal operation of hydraulic systems. Since early faults in hydraulic systems usually manifest as slow degradation of system performance, if they are not discovered in time, the severity of early faults will further increase, eventually leading to catastrophic failures of the system.

[0004] After searching the literature on the prior art, it was found that the Chinese patent publication number is CN202110647378.0, the publication date is 2021.08.20, and the patent name is: A real-time early warning method for early failure of a mine hydraulic support system. The patent application describes itself as: "The present invention discloses a real-time early warning method for an early failure of a mine hydraulic support system, including two stages: offline training and online monitoring. The offline training determines the threshold value of the health index of the hydraulic support system. In the online monitoring stage, the working parameters of the hydraulic support system are collected in real time, and the Hotelling's and statistics of the working parameters are calculated and compared with the threshold value of the health index to determine whether the hydraulic support system has a fault; the abnormal change of the correlation between the working parameters of the hydraulic support system is used as the early failure feature of the hydraulic support system, and the principal component analysis method is used to project the standardized sample data of the working parameters of the hydraulic support system into the principal component subspace and the residual subspace, and the Hotelling's and statistics are constructed on these two subspaces respectively to quantitatively describe the degree of change of the correlation between the working parameters." Its shortcomings are: this diagnostic method can only detect weak early fault signals, and does not take into account the time correlation of the observed data and the non-Gaussianity of the working parameters, resulting in insufficient sensitivity and accuracy in early fault detection of dynamic hydraulic systems, and therefore cannot effectively guide staff to perform condition-based maintenance on the hydraulic system.

[0005] In practice, the hydraulic system is a time-varying system, and the state at the previous moment usually has a great influence on the working state at the next moment (that is, the parameters are strongly autocorrelated). In addition, there are uncertainties in load disturbances and system parameters, and the observation results of the system usually exhibit non-Gaussian characteristics, which makes it difficult to effectively obtain the dynamic characteristics of the hydraulic system by simply assuming time independence and Gaussianity. Therefore, principal component analysis and partial least squares techniques are not suitable for fault detection in nonlinear dynamic systems. Existing improved principal component analysis and partial least squares techniques have been successfully applied to dynamic systems, namely dynamic principal component analysis and dynamic partial least squares, but the extraction of its principal components is not necessarily the minimum dynamic representation. When lagging variables are involved, anomaly detection will be more complicated, greatly reducing the ability of early dynamic fault detection.

[0006] Therefore, the accuracy of early fault detection in current dynamic hydraulic systems is low. Summary of the invention

[0007] Based on this, it is necessary to provide a dynamic detection method for early faults of hydraulic systems based on typical variable analysis, which can improve the accuracy of early fault detection of dynamic hydraulic systems, in order to address the above technical problems.

[0008] A method for dynamic detection of early faults in a hydraulic system based on typical variable analysis, characterized in that the method comprises:

[0009] Acquire the hydraulic system working parameters of the hydraulic system collected in a preset time period in real time as a sample data set, the hydraulic system working parameters including: lubricating oil pressure after the pump, lubricating oil pressure before the filter, lubricating oil pressure after the filter and lubricating oil temperature after the filter;

[0010] Performing a standardized data preprocessing operation on the sample data in the sample data set to obtain a preprocessed sample data set;

[0011] According to the preprocessed sample data set, construct the adjacent observation vector y at time k p,k and the observation vector y f,k , and the observation vector y p,k The corresponding observation matrix Y p and the observation vector y f,k The corresponding observation matrix Y f ;

[0012] According to the observation matrix Y p and the observation matrix Y f , analyze the observation matrix Y p The covariance ∑ pp , the observation matrix Y f The covariance ∑ ff And the observation matrix Y p and the observation matrix Y f The cross-covariance ∑ fp ;

[0013] According to the covariance ∑ pp , the covariance ∑ ff and the cross-covariance ∑ fp Perform singular value decomposition of the Hankel matrix H to obtain a singular value decomposition result;

[0014] According to the singular value decomposition result and the observation matrix Y p The covariance ∑ pp Analyze and determine the projection matrix J x and F;

[0015] According to the projection matrix J x and F, the observation vector y p,k Divide into two orthogonal subspaces and construct a state subspace z with correlation k , the rest is the residual subspace e k ;

[0016] According to the state subspace z k and the residual subspace e k , analyze Hotelling's T at time point k 2 Value and Q statistic Q k ;

[0017] The Hotelling's T 2 Value and Q statistic Q k With the preset threshold value T UCL 2 and Q UCL For comparison, or Q k >Q UCL , the current hydraulic system is determined to be in a fault state, otherwise, the current hydraulic system is in a normal state.

[0018] In one embodiment, the data standardization preprocessing operation expression is:

[0019]

[0020] Where μ is the mean vector of random variables, μ=[μ1,…,μ i …,μ n ],∈ i is the mean of the i-th random variable, i∈1,2,…,n, n is the number of sample data in the sample data; D σ is the variance matrix of the random variables, is the variance of the ith random variable, x is the sample data, and X is the preprocessed sample data.

[0021] In one embodiment, the adjacent observation vector y at time k is constructed based on the preprocessed sample data set. p,k and the observation vector y f,k , and the observation vector y p,k The corresponding observation matrix Y p and the observation vector y p,k The corresponding observation matrix Y f ,include:

[0022] According to the preprocessed sample data set, two adjacent sets of observation vectors y at time k p,k and f,k The analytical formulas are:

[0023]

[0024]

[0025] Among them, p and f are two sets of observation vectors y p,k and f,k The length of y kis the sample data at time k, q is the lag in the time window, is a real number, m is the length of the time window;

[0026] The two sets of observation vectors y p,k and f,k Defined as column matrices, we get two sets of observation matrices Y p and Y f , the observation matrix Y p and Y f The expressions are:

[0027]

[0028]

[0029] Wherein, M is the number of observation vectors in the observation matrix, M=n-2q+1, and n is the number of sample data in the sample data set.

[0030] In one embodiment, the observation matrix Y p The covariance ∑ pp , the observation matrix Y f The covariance ∑ ff And the observation matrix Y p and the observation matrix Y f The cross-covariance ∑ fp The expressions are:

[0031]

[0032]

[0033]

[0034] The superscript T indicates transposition.

[0035] In one embodiment, the covariance ∑ pp , covariance∑ ff and the cross-covariance∑ fp Perform singular value decomposition of the Hankel matrix H to obtain the singular value decomposition result, which is expressed as:

[0036]

[0037] Where U is the left singular column vector of H, V is the right singular column vector of H, and D is the diagonal matrix of ordered singular values.

[0038] In one embodiment, the projection matrix J x The expression of and F is:

[0039]

[0040]

[0041] Where I is the identity matrix, V x is a reduced matrix consisting of the first x columns of V.

[0042] In one embodiment, the state subspace z k and the residual subspace e k The expressions are:

[0043]

[0044]

[0045] Among them, V x is a reduced matrix consisting of the first x columns of V, and I is the identity matrix.

[0046] In one embodiment, the state subspace z k and the residual subspace e k , analyze Hotelling's T at time point k 2 The expressions of the and Q statistics are:

[0047]

[0048]

[0049] in, is Hotelling's T at time point k 2 Value, Q k is the Q statistic at time point k.

[0050] In one embodiment, the preset threshold value T UCL 2 and Q UCL It is obtained by offline training, and the training method is:

[0051] Collecting the hydraulic system working parameters of the hydraulic system as a training sample data set, the hydraulic system working parameters include: lubricating oil pressure after the pump, lubricating oil pressure before the filter, lubricating oil pressure after the filter and lubricating oil temperature after the filter;

[0052] Performing a standardized data preprocessing operation on the training sample data in the training sample data set to obtain a preprocessed training sample data set;

[0053] According to the preprocessed training sample data set, multiple time nodes t are defined, and for time node t, adjacent observation vectors y at time t are constructed p,t and the observation vector y f,t , and the observation vector y p,t The corresponding observation matrix and the observation vector y f,t The corresponding observation matrix

[0054] According to the observation matrix and the observation matrix Analyze the observation matrix The covariance of Observation Matrix The covariance of And the observation matrix and the observation matrix The cross-covariance

[0055] According to the covariance Covariance and mutual covariance Perform singular value decomposition of the Hankel matrix H to obtain a singular value decomposition result;

[0056] According to the singular value decomposition result and the observation matrix The covariance of Perform analysis to determine the projection matrix and

[0057] According to the projection matrix and The observation vector y p,t Divide into two orthogonal subspaces and construct a state subspace with correlation The rest is the residual subspace

[0058] According to the state subspace and the residual subspace Analyze Hotelling's T at time point t 2 and Q statistics to obtain Hotelling's T at multiple time nodes t 2 and Q statistics;

[0059] According to Hotelling's T at multiple time nodes t 2 And Q statistics, using adaptive kernel density estimation algorithm to analyze and obtain the preset threshold value T UCL 2 and Q UCL .

[0060] The above-mentioned hydraulic system early fault dynamic detection method based on typical variable analysis obtains the hydraulic system working parameters collected in a preset time period in real time as a sample data set, performs standardized data preprocessing operations on the sample data in the sample data set to obtain the preprocessed sample data set, and constructs the adjacent observation vector y at time k according to the preprocessed sample data set. p,k and the observation vector y f,k , and the corresponding observation matrix Y p and the observation matrix Y f , according to the observation matrix Y p and the observation matrix Y f , analyze the observation matrix Y p The covariance ∑ pp , the observation matrix Y f The covariance ∑ ff And the observation matrix Y p and the observation matrix Y f The cross-covariance ∑ fp , according to the covariance ∑ pp , the covariance ∑ ff and the cross-covariance ∑ fp Perform singular value decomposition of the Hankel matrix H to obtain a singular value decomposition result, and then use the singular value decomposition result and the observation matrix Y to obtain the singular value decomposition result. p The covariance ∑ pp Analyze and determine the projection matrix J x and F, according to the projection matrix J x and F, the observation vector y p,k Divide into two orthogonal subspaces and construct a state subspace z with correlation k , the rest is the residual subspace e k , analyze Hotelling's T at time point k 2 Value and Q statistic Q k , the Hotelling's T 2 Value and Q statistic Q k With the preset threshold value T UCL 2 and Q UCL For comparison, when T 2 >T UCL 2 or Q>Q UCL When the current hydraulic system is in a fault state, it is determined that the current hydraulic system is in a normal state, thereby improving the accuracy of early fault detection of the dynamic hydraulic system. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 A schematic diagram of a flow chart of a method for dynamic early fault detection of a hydraulic system based on typical variable analysis in one embodiment;

[0062] Figure 2 This is a diagram showing the detection effect of the existing univariate threshold method (Pauta criterion) on the hydraulic system based on the lubricating oil pressure after the pump;

[0063] Figure 3 This is a diagram showing the detection effect of the existing single variable threshold method (Pauta criterion) on the hydraulic system based on the oil pressure before the filter;

[0064] Figure 4 This is a diagram showing the detection effect of the existing single variable threshold method (Pauta criterion) on the hydraulic system based on the oil pressure after the filter;

[0065] Figure 5 This is a diagram showing the effect of the existing univariate threshold method (Pauta criterion) on the detection of the oil temperature after the filter in the hydraulic system;

[0066] Figure 6 The existing multivariate statistical method (principal component analysis) is used in hydraulic systems to analyze the T 2 Implementation effect diagram of threshold value;

[0067] Figure 7 It is the effect diagram of the implementation of the existing multivariate statistical method (principal component analysis) on the Q threshold value in the hydraulic system;

[0068] Figure 8 The hydraulic system early fault dynamic detection method based on typical variable analysis in this application is about T 2 Implementation effect diagram of threshold value;

[0069] Fig. 9 This is a diagram showing the implementation effect of the hydraulic system early fault dynamic detection method based on typical variable analysis in the hydraulic system regarding the Q threshold value of the present application. DETAILED DESCRIPTION

[0070] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0071] In one embodiment, Figure 1 As shown, a method for dynamic detection of early faults of hydraulic systems based on typical variable analysis is provided. The method is applied to online monitoring of hydraulic systems by terminals as an example, and includes the following steps:

[0072] Step 1-1, real-time acquisition of hydraulic system working parameters of the hydraulic system collected within a preset time period as a sample data set, the hydraulic system working parameters including: pump after lubricating oil pressure, filter before lubricating oil pressure, filter after lubricating oil pressure and filter after lubricating oil temperature.

[0073] Among them, the preset time period can be set according to actual conditions, such as the first 10 minutes, the first 20 minutes or the first 30 minutes and so on.

[0074] In one example, the preset time period is set to the first 10 minutes, that is, the hydraulic system operating parameters of the hydraulic system collected within the first 10 minutes are obtained as the sample data set.

[0075] Among them, common faults of hydraulic system include pump failure, pipeline leakage and filter blockage; the monitored working parameters of hydraulic system include four types: oil pressure after pump, oil pressure before filter, oil pressure after filter and oil temperature after filter.

[0076] Step 1-2: performing a standardized data preprocessing operation on the sample data in the sample data set to obtain a preprocessed sample data set.

[0077] It should be understood that due to the different dimensions of system operating parameters, standardizing the preprocessing operations on the collected data sample sets can facilitate typical variable analysis.

[0078] In one embodiment, the data standardization preprocessing operation expression is:

[0079]

[0080] Where μ is the mean vector of random variables, μ=[μ1,…,μ i …,μ n ],μ i is the mean of the i-th random variable, i∈1,2,…,n, n is the number of sample data in the sample data set; D σ is the variance matrix of the random variables, is the variance of the ith random variable, x is the sample data, and X is the preprocessed sample data.

[0081] Step 1-3, construct the adjacent observation vector y at time k according to the preprocessed sample data set p,k and the observation vector y f,k , and the observation vector y p,k The corresponding observation matrix Y p and the observation vector y f,k The corresponding observation matrix Y f .

[0082] In one embodiment, the adjacent observation vector y at time k is constructed based on the preprocessed sample data set. p,k and the observation vector y f,k , and the observation vector y p,k The corresponding observation matrix Y p and the observation vector y p,k The corresponding observation matrix Y f ,include:

[0083] According to the preprocessed sample data set, two adjacent sets of observation vectors y at time k p,k and f, The analytical formulas for k are:

[0084]

[0085]

[0086] Among them, p and f are two sets of observation vectors y p,k and f,k The length of y k is the sample data at time k, q is the lag in the time window, is a real number, m is the length of the time window;

[0087] The two sets of observation vectors y p,k and f,k Defined as column matrices, we get two sets of observation matrices Y p and Y f , the observation matrix Y p and Y f The expressions are:

[0088]

[0089]

[0090] Wherein, M is the number of observation vectors in the observation matrix, M=n-2q+1, and n is the number of sample data in the sample data set.

[0091] It should be understood that for a sample data set with n sample data, a total of M=n-2q+1 observation vectors can be obtained through the movement of a time window. p and Y f The observation vectors are obtained by moving a time window respectively.

[0092] Steps 1-4, according to the observation matrix Y p and the observation matrix Y f , analyze the observation matrix Yp The covariance ∑ pp , the observation matrix Y f The covariance ∑ ff And the observation matrix Y p and the observation matrix Y f The cross-covariance ∑ fp .

[0093] In one embodiment, the observation matrix Y p The covariance ∑ pp , the observation matrix Y f The covariance ∑ ff And the observation matrix Y p and the observation matrix Y f The cross-covariance ∑ fp The expressions are:

[0094]

[0095]

[0096]

[0097] The superscript T indicates transposition.

[0098] It should be understood that since canonical variable analysis constructs the state space of the system by finding the linear combination of two sets of variables that are adjacent at time k and have the greatest correlation, it can be achieved by setting a T (y f,k ) and b T (y p,k ) is the coefficient vector of the linear combination, and the correlation between these two sets of variables can be expressed as:

[0099]

[0100] Among them, ∑ pp ,∑ ff and∑ fp Represents two sets of variables (Y p and Y f ) are used to estimate the changing trends of the two groups of variables.

[0101] So we can get the observation matrix Y p The covariance ∑ pp , the observation matrix Y f The covariance ∑ ff And the observation matrix Y p and the observation matrix Y f The cross-covariance ∑ fp The expressions can be:

[0102]

[0103]

[0104]

[0105] The superscript T indicates transposition.

[0106] Step 1-5, according to the covariance ∑ pp , the covariance ∑ ff and the cross-covariance ∑ fp Perform singular value decomposition of the Hankel matrix H to obtain a singular value decomposition result.

[0107] In one embodiment, the covariance ∑ pp , covariance∑ ff and the cross-covariance∑ fp Perform singular value decomposition of the Hankel matrix H to obtain the singular value decomposition result, which is expressed as:

[0108]

[0109] Where U is the left singular column vector of H, V is the right singular column vector of H, and D is the diagonal matrix of ordered singular values.

[0110] in:

[0111]

[0112]

[0113]

[0114] Where r is the rank of the Hankel matrix H, the i-th singular vector u i and v i corresponds to the coefficient of the linear combination of system variables, σ i is the correlation of the i-th combined variable, i∈1,2…r.

[0115] Step 1-6, based on the singular value decomposition result and the observation matrix Y p The covariance ∑ pp Analyze and determine the projection matrix J x and F.

[0116] In one embodiment, the projection matrix J x The expression of and F is:

[0117]

[0118]

[0119] Where I is the identity matrix, V x is a reduced matrix consisting of the first x columns of V.

[0120] Step 1-7, according to the projection matrix J x and F, the observation vector y p,k Divide into two orthogonal subspaces and construct a state subspace z with correlation k , the rest is the residual subspace e k .

[0121] In one embodiment, the state subspace z k and the residual subspace e k The expressions are:

[0122]

[0123]

[0124] Among them, V x is a reduced matrix consisting of the first x columns of V, and I is the identity matrix.

[0125] Among them, the state subspace is represented by the first x typical variables, which represents the main characteristics of the system measurement; the residual subspace is represented by the remaining typical variables, which represents the remaining information of the system measurement.

[0126] Steps 1-8, according to the state subspace z k and the residual subspace e k , analyze Hotelling's T at time point k 2 Value and Q statistic Q k .

[0127] In one embodiment, the state subspace z k and the residual subspace e k , analyze Hotelling's T at time point k 2 Value and Q statistic Q k The expression is:

[0128]

[0129]

[0130] in, is Hotelling's T at time point k 2 Value, Qk is the Q statistic at time point k.

[0131] Step 1-9, the Hotelling's T 2 Value and Q statistic Q k With the preset threshold value T UCL 2 and Q UCL For comparison, or Q k >Q UCL , the current hydraulic system is determined to be in a fault state, otherwise, the current hydraulic system is in a normal state.

[0132] Among them, according to the hydraulic system fault judgment rules, That is, when the multivariate statistics or Q k When the data fluctuation exceeds its upper control limit, an abnormal alarm will be issued.

[0133] The above hydraulic system early fault dynamic detection method based on typical variable analysis obtains the hydraulic system working parameters collected in a preset time period as a sample data set, performs standardized data preprocessing operations on the sample data in the sample data set, obtains the preprocessed sample data set, and constructs the adjacent observation vector y at time k according to the preprocessed sample data set. p,k and the observation vector y f,k , and the observation vector y p,k The corresponding observation matrix Y p and the observation vector y f,k The corresponding observation matrix Y f , according to the observation matrix Y p and the observation matrix Y f , analyze the observation matrix Y p The covariance ∑ pp , the observation matrix Y f The covariance ∑ ff And the observation matrix Y p and the observation matrix Y f The cross-covariance ∑ fp , according to the covariance ∑ pp , the covariance ∑ ff and the cross-covariance ∑ fp Perform singular value decomposition of the Hankel matrix H to obtain a singular value decomposition result, and then use the singular value decomposition result and the observation matrix Y to obtain the singular value decomposition result. p The covariance ∑ pp Analyze and determine the projection matrix J x and F, according to the projection matrix J xand F, the observation vector y p,k Divide into two orthogonal subspaces and construct a state subspace z with correlation k , the rest is the residual subspace e k , analyze Hotelling's T at time point k 2 and Q statistics, Hotelling's T 2 and Q statistics and the preset threshold value T UCL 2 and Q UCL For comparison, when T 2 >T UCL 2 or Q>Q UCL When the current hydraulic system is in a fault state, it is determined that the current hydraulic system is in a normal state, thereby improving the accuracy of early fault detection of the dynamic hydraulic system.

[0134] In one embodiment, the preset threshold value T UCL 2 and Q UCL It is obtained by offline training, and the training method is:

[0135] Step 2-1, collecting hydraulic system working parameters of the hydraulic system as a training sample data set, the hydraulic system working parameters include: pump after the lubricating oil pressure, filter before the lubricating oil pressure, filter after the lubricating oil pressure and filter after the lubricating oil temperature.

[0136] Step 2-2: performing a standardized data preprocessing operation on the training sample data in the training sample data set to obtain a preprocessed training sample data set.

[0137] The data standardization preprocessing operation expression is:

[0138]

[0139] Where μ is the mean vector of random variables, μ=[μ1,…,μ i ,…,μ N ],μ i is the mean of the i-th random variable, i∈1,2,…,N, N is the number of sample data in the training sample data set; D σ is the variance matrix of the random variables, is the variance of the ith random variable, is the training sample data, is the preprocessed training sample data.

[0140] Step 2-3, based on the preprocessed training sample data set, define multiple time nodes t, and for time node t, construct the adjacent observation vector y at time t p,t and the observation vector y f,t , and the observation vector y p,t The corresponding observation matrix and the observation vector y f,t The corresponding observation matrix

[0141] In one embodiment, the construction of the adjacent observation vector y at time t p,t and the observation vector y f,t , and the observation vector y p,t The corresponding observation matrix and the observation vector y f,t The corresponding observation matrix include:

[0142] According to the preprocessed training sample data set, two adjacent sets of observation vectors y at time t p,t and f,t The analytical formulas are:

[0143]

[0144]

[0145] Among them, p and f are two sets of observation vectors y p,t and f,t The length of y t is the training sample data at time t, q is the lag in the time window, is a real number, m is the length of the time window;

[0146] The two sets of observation vectors y p,t and f,t They are defined as column matrices, and two sets of observation matrices are obtained. and Observation Matrix and The expressions are:

[0147]

[0148]

[0149] Wherein, M is the number of observation vectors in the observation matrix, M=N-2q+1, and N is the number of training sample data in the training sample data set.

[0150] It should be understood that for a training sample data set with N training sample data, a total of M=N-2q+1 observation vectors can be obtained through the movement of a time window. and The observation vectors are obtained by moving a time window respectively.

[0151] Step 2-4, according to the observation matrix and the observation matrix Analyze the observation matrix The covariance of Observation Matrix The covariance of And the observation matrix and the observation matrix The cross-covariance

[0152] In one embodiment, the observation matrix The covariance of The observation matrix The covariance of And the observation matrix and the observation matrix The cross-covariance The expressions are:

[0153]

[0154]

[0155]

[0156] The superscript T indicates transposition.

[0157] Step 2-5, according to the covariance Covariance and mutual covariance Perform singular value decomposition of the Hankel matrix H to obtain a singular value decomposition result.

[0158] In one embodiment, the covariance Covariance and mutual covariance Perform singular value decomposition of the Hankel matrix H to obtain the singular value decomposition result, which is expressed as:

[0159]

[0160] Where U is the left singular column vector of H, V is the right singular column vector of H, and D is the diagonal matrix of ordered singular values.

[0161] Step 2-6, based on the singular value decomposition result and the observation matrix The covariance of Perform analysis to determine the projection matrix and

[0162] In one embodiment, the projection matrix and The expression is:

[0163]

[0164]

[0165] Where I is the identity matrix, V x is a reduced matrix consisting of the first x columns of V.

[0166] Step 2-7, according to the projection matrix and The observation vector y p,t Divide into two orthogonal subspaces and construct a state subspace with correlation The rest is the residual subspace

[0167] In one embodiment, the state subspace and the residual subspace The expressions are:

[0168]

[0169]

[0170] Among them, V x is a reduced matrix consisting of the first x columns of V, and I is the identity matrix.

[0171] Step 2-8, according to the state subspace and the residual subspace Analyze Hotelling's T at time point t 2 and Q statistics to obtain Hotelling's T at multiple time nodes t 2 and the Q statistic.

[0172] In one embodiment, the state subspace and the residual subspace Analyze Hotelling's T at time point t 2 The value of T t 2 and Q statistic Q t The expression is:

[0173]

[0174]

[0175] Among them, T t 2 is Hotelling's T at time t 2 Value, Q t is the Q statistic at time point t.

[0176] Step 2-9, based on Hotelling's T at multiple time nodes t 2 And Q statistics, using adaptive kernel density estimation algorithm to analyze and obtain the preset threshold value T UCL 2 and Q UCL .

[0177] It should be understood that the number of sample data in the training sample data set is N, and one sample data is used at each time node. Each time node is regarded as a time node t, and the corresponding Hotelling's T is calculated. 2 and Q statistics, we can obtain N Hotelling's T 2 and N Q statistics.

[0178] Among them, the analysis principle of the adaptive kernel density estimation algorithm is:

[0179] Set the initial bandwidth h0, for a set of random variables {u1,u2,...,u i ...,u N}, i∈1,2,…,N,, the local bandwidth factor τ is obtained by calculating each random variable u i , the formula is as follows:

[0180]

[0181] in, is the initial adaptive probability density estimation result, u is the current random variable, and h is the bandwidth determined during the kernel density estimation. K(·) represents the kernel function used to fit the probability distribution of the training sample data. This application selects the Gaussian kernel function, whose formula is:

[0182]

[0183] Among them, e is a natural constant and c is a variable in the Gaussian kernel function.

[0184] Step 2-9-2: By using the local bandwidth factor τ iModify the initial bandwidth h0 to obtain the final adaptive probability density estimation result

[0185] Calculate each random variable u to obtain the local bandwidth factor τ i , the formula is as follows:

[0186]

[0187] Among them, α is the sensitivity factor.

[0188] By the local bandwidth factor τ i Modify the initial bandwidth h0 to obtain the final adaptive probability density estimation result Final adaptive probability density estimation result The expression is:

[0189]

[0190] Among them, at the preset confidence level Next, according to the final adaptive probability density estimation result Derivation of Hotelling's T 2 And the threshold value of Q statistic:

[0191] At a preset confidence level Next, using the final adaptive probability density estimation result Expression, analyze each Hotelling's T 2 The value of and the probability density estimation results corresponding to each Q statistic value are used to estimate the probability density using Hotelling T 2 The analytical formula for the threshold value of the Q statistic is given by each Hotelling's T 2 The corresponding threshold value is determined from the value of and each Q statistic value, Hotelling T 2 The analytical formula for the threshold value of the Q statistic is:

[0192]

[0193]

[0194] It should be understood that for the probability density function Corresponding Hotelling T 2 The analysis of the threshold value of the Q statistic sets the variable u in the adaptive probability density estimation to be the Hotelling T at time t. 2 and Q statistics, among N Hotelling's T 2The Hotelling's T is determined by the analytical formula of its threshold value from the values ​​of and N Q statistics. 2 and the threshold value of the Q statistic.

[0195] It should be understood that the above hydraulic system early fault dynamic detection method based on typical variable analysis uses a combination of typical variable analysis and adaptive kernel density function algorithm to perform early fault dynamic detection, which can improve the dynamic detection capability of hydraulic system early faults; by constructing a state subspace z with strong correlation based on the observation vector through typical variable analysis k and the residual subspace e k , solve the time correlation problem of observed variables, and provide a basis for dynamic detection and index construction of hydraulic systems; according to the different density distribution of sample points and the non-Gaussianity of working parameters, the adaptive kernel density estimation algorithm is used to adaptively adjust the bandwidth h of the probability density function to ensure that the probability density estimate is closer to the true value, and the threshold value T for entering early faults can be accurately determined. UCL 2 and Q UCL , which can effectively improve the accuracy of dynamic detection of early faults in hydraulic systems.

[0196] In one embodiment, the present application takes the hydraulic lubrication system of a large ship power plant as an example. Considering that the working state of the lubrication system is crucial to the large ship power plant, the lubricating oil pressure and temperature are the key parameters for the control of the lubrication system. Considering the frequency and severity of the fault, the three common progressive faults in the hydraulic lubrication system of the ship power plant are "pump fault", "pipeline leakage" and "filter blockage". The performance of the hydraulic lubrication system is described by four key working parameters, namely "lubricating oil pressure after the pump", "lubricating oil pressure before the filter", "lubricating oil pressure after the filter" and "lubricating oil temperature after the filter". The working parameters are collected by the PCI-6225 data acquisition system, and the collected data are preprocessed by median filtering to reduce the measurement noise. The ship power plant works under three typical working conditions of 1800r / min full load 25%, 50% and 75%, and the collected system working parameters are used to construct the training data set and the test data set. The training data set consists of normal data and is used to determine the control limit of the system parameters, while the test data set consists of normal data and fault data, which is used to evaluate the performance of the fault detection method.

[0197] This application builds a hydraulic system experimental test platform, and tests four types of "pump failure", "pipeline leakage", "filter blockage" and normal under three working conditions of 25%, 50% and 75%. For the convenience of introduction and analysis, taking "pump failure" under 50% working condition as an example, the univariate threshold method (Pauta criterion), multivariate statistical method (principal component analysis) and multivariate statistical method (typical variable analysis) are introduced respectively.

[0198] like Figure 2-Figure 5 The effect diagram of the existing univariate threshold method (Pauta criterion) implemented on the hydraulic system is shown. The first 100 blue points are normal data, the last 400 red points are fault data, and the upper and lower red dotted lines correspond to the upper and lower limits of the detection threshold respectively. Figure 2-Figure 5 It can be seen that a small number of faults can be detected in the working parameters of "lubricating oil pressure after filter" and "lubricating oil temperature after filter", but the other two conditions are almost undetectable. The overall detection rate is 36%, and the detection effect is not ideal.

[0199] like Figure 6-Figure 7 The effect diagram of the existing multivariate statistical method (principal component analysis) implemented on the hydraulic system is shown. In the principal component analysis detection method based on Hotelling's statistics, the black solid line is the waveform of the statistic, and the red dotted line is the threshold value of the statistic obtained by training. Figure 6-Figure 7 It can be seen that most faults can be effectively detected using the four working parameters of this method, but there are still a large number of fault data points below the threshold value and cannot be effectively detected. The overall detection rate is 74.75%, and the detection result is relatively general.

[0200] like Figure 8-Figure 9 The effect diagram of the implementation of the multivariate statistical method (typical variable analysis) of the present application on the hydraulic system is shown in FIG. In the typical variable analysis detection method based on Hotelling's statistics, it can be seen that among the four working parameters, no matter for Hotelling's T 2 Still using Q statistics, almost all fault data points are above the threshold value, and only a few fault data points are below the threshold value, with a detection rate of 86.75%. All normal data points are below the threshold value, and the misdiagnosis rate is 0. Compared with the Pauta criterion and principal component analysis method, the detection rate of the canonical variable analysis method is significantly improved, and the comprehensive detection effect is ideal, meeting the expected goal.

[0201] In summary, in view of the shortcomings of existing methods in assuming the time independence of observed data and the Gaussianity of working parameters, a dynamic detection method for early faults of hydraulic systems based on typical variable analysis is proposed. Taking the hydraulic lubrication system of a ship power unit as an example, the feasibility of this application is illustrated. Assuming three typical workloads and speeds and three common hydraulic system faults, the Pauta criterion and the principal component analysis method based on multivariate statistical process monitoring are compared with the method combining typical variable analysis with adaptive kernel density estimation. The results show that the multivariate statistical process monitoring method can effectively detect early faults that are difficult to detect with the traditional Pauta criterion, and the multivariate statistical process monitoring technology combined with adaptive kernel density estimation and typical variable analysis can further improve the detection rate of the hydraulic system, which illustrates the advancement of this application. This application can significantly improve the detection rate of early faults in hydraulic systems and provide an effective technology for hydraulic system status monitoring.

[0202] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0203] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A dynamic detection method for early faults of hydraulic system based on typical variable analysis, characterized in that: The method comprises: Acquire the hydraulic system working parameters of the hydraulic system collected in a preset time period in real time as a sample data set, the hydraulic system working parameters including: lubricating oil pressure after the pump, lubricating oil pressure before the filter, lubricating oil pressure after the filter and lubricating oil temperature after the filter; Performing a data standardization preprocessing operation on the sample data in the sample data set to obtain a preprocessed sample data set; According to the preprocessed sample data set, construct the adjacent observation vector y at time k p,k and the observation vector y f,k , and the observation vector y p,k The corresponding observation matrix Y p and the observation vector y f,k The corresponding observation matrix Y f ; According to the observation matrix Y p and the observation matrix Y f , analyze the observation matrix Y p The covariance ∑ pp , the observation matrix Y f The covariance ∑ ff And the observation matrix Y p and the observation matrix Y f The cross-covariance ∑ fp ; According to the covariance ∑ pp , the covariance ∑ ff and the cross-covariance ∑ fp Perform singular value decomposition of the Hankel matrix H to obtain a singular value decomposition result; According to the singular value decomposition result and the observation matrix Y p The covariance ∑ pp Analyze and determine the projection matrix J x and F; According to the projection matrix J x and F, the observation vector y p,k Divide into two orthogonal subspaces and construct a state subspace z with correlation k , the rest is the residual subspace e k ; According to the state subspace z k and the residual subspace e k , analyze Hotelling's T at time point k 2 The value of T k 2 and Q statistic Q k ; The Hotelling's T 2 Value and Q statistic Q k With the preset threshold value T UCL 2 and Q UCL For comparison, or Q k >Q UCL , the current hydraulic system is determined to be in a fault state, otherwise, the current hydraulic system is in a normal state.

2. The method according to claim 1, characterized in that The data standardization preprocessing operation expression is: Where μ is the mean vector of random variables, μ=[μ1,…,μ i ,…,μ n ],μ i is the mean of the i-th random variable, i∈1,2,…,n, n is the number of sample data in the sample data set; D σ is the variance matrix of the random variables, is the variance of the ith random variable, x is the sample data, and X is the preprocessed sample data.

3. The method according to claim 2, characterized in that The method constructs the adjacent observation vector y at time k according to the preprocessed sample data set. p,k and the observation vector y f,k , and the observation vector y p,k The corresponding observation matrix Y p and the observation vector y p,k The corresponding observation matrix Y f ,include: According to the preprocessed sample data set, two adjacent sets of observation vectors y at time k p,k and f,k The analytical formulas are: Among them, p and f are two sets of observation vectors y p,k and f,k The length of y k is the sample data at time k, q is the lag in the time window, is a real number, m is the length of the time window; The two sets of observation vectors y p,k and f,k Defined as column matrices, we get two sets of observation matrices Y p and Y f , the observation matrix Y p and Y f The expressions are: Wherein, M is the number of columns of the observation matrix, M=n-2q+1, and n is the number of sample data in the sample data set.

4. The method according to claim 3, characterized in that The observation matrix Y p The covariance ∑ pp , the observation matrix Y f The covariance ∑ ff And the observation matrix Y p and the observation matrix Y f The cross-covariance ∑ fp The expressions are: The superscript T indicates transpose.

5. The method according to claim 4, characterized in that The covariance ∑ pp , covariance∑ ff and the cross-covariance∑ fp Perform singular value decomposition of the Hankel matrix H to obtain the singular value decomposition result, which is expressed as: Where U is the left singular column vector of H, V is the right singular column vector of H, and D is the diagonal matrix of ordered singular values.

6. The method according to claim 5, characterized in that The projection matrix J x The expression of and F is: Where I is the identity matrix, V x is a reduced matrix consisting of the first x columns of V.

7. The method according to claim 6, characterized in that The state subspace z k and the residual subspace e k The expressions are: Among them, V x is a reduced matrix consisting of the first x columns of V, and I is the identity matrix.

8. The method according to claim 7, characterized in that According to the state subspace z k and the residual subspace e k , analyze Hotelling's T at time point k 2 Value and Q statistic Q k The expression is: in, is Hotelling's T at time point k 2 Value, Q k is the Q statistic at time point k.

9. The method according to claim 8, characterized in that The preset threshold value T UCL 2 and Q UCL It is obtained by offline training, and the training method is: Collecting the hydraulic system working parameters of the hydraulic system as a training sample data set, the hydraulic system working parameters include: lubricating oil pressure after the pump, lubricating oil pressure before the filter, lubricating oil pressure after the filter and lubricating oil temperature after the filter; Performing a standardized data preprocessing operation on the training sample data in the training sample data set to obtain a preprocessed training sample data set; According to the preprocessed training sample data set, multiple time nodes t are defined, and for time node t, adjacent observation vectors y at time t are constructed. p,t and the observation vector y f,t , and the observation vector y p,t The corresponding observation matrix and the observation vector y f,t The corresponding observation matrix According to the observation matrix and the observation matrix Analyze the observation matrix The covariance of Observation Matrix The covariance of And the observation matrix and the observation matrix The cross-covariance According to the covariance Covariance and mutual covariance Perform singular value decomposition of the Hankel matrix H to obtain a singular value decomposition result; According to the singular value decomposition result and the observation matrix The covariance of Perform analysis to determine the projection matrix and According to the projection matrix and The observation vector y p,t Divide into two orthogonal subspaces and construct a state subspace with correlation The rest is the residual subspace According to the state subspace and the residual subspace Analyze Hotelling's T at time point t 2 and Q statistics to obtain Hotelling's T at multiple time nodes t 2 and Q statistics; According to Hotelling's T at multiple time nodes t 2 And Q statistics, using adaptive kernel density estimation algorithm to analyze and obtain the preset threshold value T UCL 2 and Q UCL .

Images