Dynamic expansion fault detection method based on regression relationship

Abstract

The application discloses a dynamic expansion fault detection method based on a regression relation and relates to the technical field of fault detection engineering. The method comprises the following steps: obtaining a predictable dynamic information matrix according to an input data matrix and an output data matrix of a standard missile; obtaining a main score matrix and a number of principal components related to output according to the predictable dynamic information matrix; obtaining an input load matrix according to the main score matrix and a dynamic input information matrix; obtaining an input residual matrix according to the main score matrix and the input load matrix; obtaining a number of principal components irrelevant to output according to the input residual matrix; obtaining an input control limit and an output control limit according to the number of principal components; obtaining a statistic quantity related to output and a statistic quantity irrelevant to output according to an input data matrix and the input load matrix of a to-be-detected missile; and determining a fault detection result according to the statistic quantity related to output, the input control limit, the statistic quantity irrelevant to output and the output control limit. The application can reduce a false alarm rate and improve detection precision.

Description

Technical Field

[0001] This invention relates to the field of fault detection engineering technology, and in particular to a dynamic extended fault detection method based on regression relationships. Background Technology

[0002] With the rapid development of science and technology, the research and development and production processes of weaponry and equipment are gradually moving towards larger scale, greater complexity, integration, and intelligence. At the same time, the cross-linking of various components is increasing the complexity of the equipment. Furthermore, with the application of distributed control systems and the widespread use of data acquisition devices, equipment fault detection is becoming increasingly data-driven, making process monitoring and fault detection of complex equipment systems an inevitable trend.

[0003] In 1983, Wold et al. proposed the Partial Least Squares (PLS) algorithm, which extracts the information most relevant to quality from process data based on the principle of maximizing the covariance between process variables and quality variables. Subsequently, the PLS method was applied to disciplines such as statistics and chemistry, as well as process monitoring and fault detection. However, the standard PLS method still has shortcomings when monitoring complex equipment systems, especially given the correlations between components in modern equipment systems, which create dynamic relationships between data. This has led to the development of dynamic algorithms based on PLS.

[0004] In actual process monitoring, test data containing quality-independent faults are easily identified as quality-related faults, leading to false alarms. Therefore, the false alarm rate is a critical indicator when detecting quality-independent faults. A low false alarm rate ensures good system robustness, effectively preventing false alarms and facilitating the smooth operation of industrial processes. On the other hand, the basic idea of ​​the PLS algorithm is to use the correlation between process variables and quality variables to perform oblique decomposition of the variable space. Oblique decomposition increases the number of false alarms in the quality-related subspace, affecting fault detection effectiveness. Therefore, modifying the dynamic model to reduce the false alarm rate is of great significance for improving fault detection effectiveness. Summary of the Invention

[0005] The purpose of this invention is to provide a dynamic extended fault detection method based on regression relationships, which can reduce the false alarm rate and improve the detection accuracy.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A dynamic extended fault detection method based on regression relationships includes:

[0008] The predictable dynamic information matrix of the standard missile is obtained based on the input data matrix and output data matrix of the standard missile; the standard missile is a missile that has not experienced any malfunctions.

[0009] The PCA algorithm is used to orthogonally decompose the predictable dynamic information matrix of the standard missile to obtain the main score matrix and the number of principal components related to the output;

[0010] The input load matrix is ​​obtained based on the main score matrix and the dynamic input information matrix of the standard missile; the dynamic input information matrix of the standard missile is obtained by processing the standard missile using the DiPLS algorithm;

[0011] The input residual matrix is ​​obtained based on the main score matrix and the input load matrix;

[0012] The PCA algorithm is used to orthogonally decompose the input residual matrix to obtain the number of principal components that are independent of the output.

[0013] The input control limits and output control limits are obtained based on the number of principal components related to the output and the number of principal components unrelated to the output;

[0014] Obtain the input data matrix of the missile to be detected;

[0015] Based on the input data matrix of the missile to be detected and the input load matrix, statistical quantities related to the output and statistical quantities unrelated to the output are obtained;

[0016] If the output-related statistic is greater than the input control limit, the missile under test has a quality-related fault; if the output-independent statistic is greater than the output control limit, the missile under test has a quality-independent fault.

[0017] Optionally, obtaining the predictable dynamic information matrix of the standard missile based on its input and output data matrices specifically includes:

[0018] The input data matrix and output data matrix of the standard missile are normalized respectively to obtain the normalized input matrix and normalized output matrix of the standard missile;

[0019] The normalized input matrix and normalized output matrix of the standard missile are respectively augmented to obtain the dynamic input matrix and dynamic output matrix of the standard missile;

[0020] The predictable dynamic information matrix of the standard missile is obtained based on the dynamic input matrix and dynamic output matrix of the standard missile.

[0021] Optionally, obtaining the input residual matrix based on the main score matrix and the input load matrix specifically includes:

[0022] Based on the main score matrix and the input load matrix, obtain the main information related to the predictable dynamic information matrix;

[0023] The input residual matrix is ​​obtained based on the key information related to the predictable dynamic information matrix.

[0024] Optionally, obtaining output-related statistics and output-independent statistics based on the input data matrix of the missile to be detected and the input load matrix specifically includes:

[0025] The input data matrix and the output data matrix of the missile to be detected are normalized respectively to obtain the normalized input matrix and the normalized output matrix of the missile to be detected.

[0026] The normalized input matrix and normalized output matrix of the missile to be detected are respectively augmented to obtain the dynamic input matrix and dynamic output matrix of the missile to be detected;

[0027] Based on the dynamic input matrix of the missile to be detected and the input load matrix, a score vector related to the output and a score vector unrelated to the output are obtained;

[0028] Based on the output-related score vector and the output-independent score vector, we obtain the output-related statistics and the output-independent statistics.

[0029] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: The present invention obtains a predictable dynamic information matrix based on the input data matrix and output data matrix of a standard missile; obtains a main score matrix and the number of output-related principal components based on the predictable dynamic information matrix; obtains an input load matrix based on the main score matrix and dynamic input information matrix; obtains an input residual matrix based on the main score matrix and input load matrix; obtains the number of output-independent principal components based on the input residual matrix; obtains input control limits and output control limits based on the number of principal components; obtains output-related statistics and output-independent statistics based on the input data matrix and input load matrix of the missile to be detected; and determines the fault detection result based on the output-related statistics, input control limits, output-independent statistics, and output control limits, thereby reducing the false alarm rate and improving detection accuracy. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 A flowchart of a dynamic extended fault detection method based on regression relationships provided in an embodiment of the present invention;

[0032] Figure 2 A graph showing the monitoring effect of quality-related fault IDV(13);

[0033] Figure 3 The monitoring effect diagram of quality-independent fault IDV(4);

[0034] Figure 4 This is a diagram illustrating the effectiveness of quality-related fault monitoring. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0037] like Figure 1 As shown, this invention provides a dynamic extended fault detection method based on regression relationships, comprising:

[0038] The predictable dynamic information matrix of the standard missile is obtained based on its input and output data matrices; the standard missile is a missile without malfunctions; the input data matrix includes sensor data from all key components that can affect the final nozzle sway angle of the missile within a set time period, such as voltage, current, frequency, and environmental changes; the output data matrix includes the nozzle sway angle within the set time period. Data can be acquired by deploying sensors at key locations and is considered as input data; the final output of the missile system is the amplitude of the nozzle sway; the ultimate goal of the entire system is to ensure the normal operation of the launch mission by adjusting the nozzle.

[0039] The PCA algorithm is used to orthogonally decompose the predictable dynamic information matrix of the standard missile to obtain the main score matrix and the number of principal components A related to the output. y Number of principal components A y Determined by cross-validation.

[0040] The input load matrix is ​​obtained based on the main score matrix and the dynamic input information matrix of the standard missile; the dynamic input information matrix of the standard missile is obtained by processing the standard missile using the DiPLS algorithm.

[0041] The input residual matrix is ​​obtained based on the main score matrix and the input load matrix.

[0042] The input residual matrix is ​​orthogonally decomposed using the PCA algorithm to obtain the number A of principal components that are independent of the output. r Number of principal components A r Determined by cross-validation.

[0043] The input control limits and output control limits are obtained based on the number of principal components related to the output and the number of principal components unrelated to the output.

[0044] Based on the input data matrix of the missile to be detected and the input load matrix, statistics related to the output and statistics unrelated to the output are obtained.

[0045] If the output-related statistic is greater than the input control limit, the missile under test has a quality-related fault; if the output-independent statistic is greater than the output control limit, the missile under test has a quality-independent fault.

[0046] In practical applications, obtaining the predictable dynamic information matrix of the standard missile based on its input and output data matrices specifically includes:

[0047] The input data matrix and output data matrix of the standard missile are normalized respectively to obtain the normalized input matrix and normalized output matrix of the standard missile;

[0048] The normalized input matrix and normalized output matrix of the standard missile are respectively augmented to obtain the dynamic input matrix and dynamic output matrix of the standard missile;

[0049] The predictable dynamic information matrix of the standard missile is obtained based on the dynamic input matrix and dynamic output matrix of the standard missile.

[0050] In practical applications, obtaining the input residual matrix based on the main score matrix and the input load matrix specifically includes:

[0051] Based on the main score matrix and the input load matrix, obtain the main information related to the predictable dynamic information matrix;

[0052] The input residual matrix is ​​obtained based on the key information related to the predictable dynamic information matrix.

[0053] In practical applications, obtaining output-related and output-independent statistics based on the input data matrix and input load matrix of the missile to be detected specifically includes:

[0054] The input data matrix and the output data matrix of the missile to be detected are normalized respectively to obtain the normalized input matrix and the normalized output matrix of the missile to be detected.

[0055] The normalized input matrix and normalized output matrix of the missile to be detected are respectively augmented to obtain the dynamic input matrix and dynamic output matrix of the missile to be detected;

[0056] Based on the dynamic input matrix of the missile to be detected and the input load matrix, a score vector related to the output and a score vector unrelated to the output are obtained;

[0057] Based on the output-related score vector and the output-independent score vector, we obtain the output-related statistics and the output-independent statistics.

[0058] In practical applications, the input and output data matrices of the standard missile are normalized to obtain the normalized input and output matrices, respectively. Then, the normalized input and output matrices are augmented to obtain the dynamic input and output matrices. Specifically, to improve the model's convergence speed and accuracy, and to eliminate the influence of unit differences, it is necessary to first obtain the original input X of the large and complex equipment. xun Output Y xun Normalized preprocessing results in a data matrix X = [x0, x1, ..., x2] with a mean of 0 and a variance of 1. m-1 ] T ∈R n×m Y = [y0, y1, ..., y p-1 ] T ∈R n×p Where n represents the number of samplings, m and p represent the number of variables, and the dynamic input matrix Z is obtained by data augmentation. s and dynamic output matrix Y s .

[0059] X i =[x i ,x i+1 ,…,x i+N ] T ∈R (n-s+1)×m (1)

[0060] Z s =[X s ,X s-1 [,…,X0]∈R (n-s+1)×sm (2)

[0061] Y s =[y s ,y s+1,…,y s+N ] T ∈R (n-s+1)×p (3)

[0062] In the formula, s is the order of the dynamic system, which is determined by the Akaike information criterion (AIC).

[0063] In practical applications, the predictable dynamic information matrix of the standard missile is obtained based on its dynamic input matrix and dynamic output matrix. Specifically, this includes: to better obtain dynamic output data related to dynamic input data, using the regression relationship matrix M to obtain a predictable dynamic information matrix of the output. The regression matrix M is calculated as shown in equation (4):

[0064]

[0065] The predictable dynamic output information matrix is ​​shown in equation (5):

[0066]

[0067] In practical applications, the PCA algorithm is used to orthogonally decompose the predictable dynamic information matrix of the standard missile to obtain the main score matrix T. y Specifically, this includes: using the PCA algorithm to... Orthogonal decomposition into the main parts useful for predicting the output and residual part The decomposition formula is shown in equation (6):

[0068]

[0069] In the formula T y For the main score matrix, For the secondary score matrix, Q y , T respectively y , The load matrix.

[0070] In practical applications, the input load matrix is ​​obtained based on the main score matrix and the dynamic input information matrix of the standard missile. Specifically, this includes: [the process of obtaining the predictable output...] The main score matrix T y As predictable input data The main score matrix can predict the input load matrix P. x Depend on and T y The regression yields the result shown in equation (7):

[0071]

[0072] In the formula Predicted by the DiPLS algorithm, the steps of which are shown below:

[0073] Step (1): Initialization u s Take Y s any column;

[0074] Step (2): Calculate w and normalize it.

[0075] Step (3): Calculate the score vector t = Xw of X;

[0076] Step (4): Calculate the load vector q of Y and normalize it, i.e.

[0077] Step (5): Calculate the score vector u of Y s =Y s q;

[0078] Step (6): Calculate the weighting coefficient β and normalize it, i.e. m is the number of variables input to X, I m This represents an m-dimensional unit variable.

[0079] Step (7): Estimate the coefficients using the least squares method: b j =(T s T T s ) -1 T s T u s T s =[t s ,t s-1 [,…,t0]; Taking different columns of t yields t s , t s-1 , ..., t0.

[0080] Step (8): Estimate the score, i.e. Calculate Y s residual Y s : At this time, Y s Considered as the new Y s ;

[0081] Step (9): Return to step (2) until t i convergence.

[0082] In practical applications, the main information related to the predictable dynamic information matrix is ​​obtained based on the main score matrix and the input load matrix. Specifically, this includes: the space of the input X is ultimately decomposed by the DiPLS algorithm into a space related to the output. and residual space Where T = [t1, t2, ..., t A ]、P=[t1X,t2X,…,t A X]. At this time Projected onto The related low-dimensional subspace was obtained with Key related information

[0083] In practical applications, based on the aforementioned predictable dynamic information matrix The relevant key information is obtained from the input residual matrix E. x Specifically, this includes: [and] Unrelated input residual matrix E x The PCA method was used to separate the variation of large variance and noise variation in the residuals. r We construct the Q statistic to monitor noise with small variance.

[0084]

[0085] In summary, the constructed model of the dynamic extended fault detection method based on regression relationships is as follows:

[0086]

[0087] In the formula It represents predictable dynamic information related to quality variables in process variables; This represents dynamic information with large variance in predictable process variables that is unrelated to quality variables; E r E represents noise in a predictable process variable. p This represents unpredictable information in process variables; It represents dynamic information with a large variance that can be explained by process variables; E represents dynamic information with small variance that can be explained by process variables; y This represents unpredictable residual information that is independent of process variables.

[0088] For a single sample {x,y}, its score is constructed as follows:

[0089]

[0090]

[0091] The predictable input residual matrix is ​​constructed as follows:

[0092]

[0093] The unpredictable input residual matrix is ​​constructed as follows:

[0094]

[0095] The main part of the predictable output, the residual part, and the unpredictable residual are constructed as follows:

[0096]

[0097]

[0098]

[0099] The meanings of the different subspaces of this model are shown in Table 1.

[0100] Table 1. Meaning of Different Subspaces

[0101]

[0102] In practical applications, for the input data matrix {x} of the missile to be detected new ,y new}, using equations (10)-(11) to construct a new score t y_new t r_new That is, based on the dynamic input matrix Z of the missile to be detected s_new and the input load matrix P x Obtain the score vector t associated with the output. y_new and the score vector t that is independent of the output r_new Specifically, this includes: according to the formula calculate.

[0103] In practical applications, output-related statistics and output-independent statistics are obtained based on the output-related score vector and the output-independent score vector, specifically including:

[0104] Λ is obtained based on the score vector related to the output and the score vector unrelated to the output. y v y_new and Λ r v r_new ;

[0105] According to Λ y v y_new and Λ r v r_new Obtain statistics that are relevant to the output and statistics that are irrelevant to the output.

[0106] In practical applications, latent variables often contain significant dynamic characteristics within the data, requiring the establishment of dynamic models to reflect their internal autocorrelation properties. In this case, a time series model (i.e., a vector autoregressive model) can be used to describe the input latent variable t at the current time step. k As shown in the following formula:

[0107]

[0108] Where the model parameters Θ=[α1 α2…α s ] T , If v k This can be viewed as a zero-mean white noise sequence, and the parameters Θ can be estimated using the multivariate least squares (LS) algorithm:

[0109]

[0110] It can be seen that the dynamic score residual v k With dynamic score vector t k It is time-independent, therefore, when monitoring dynamic score statistics, it can be done by analyzing v. k The monitoring of t k Monitoring, namely:

[0111]

[0112] In the formula

[0113] Therefore, in practical applications, according to Λ y v y_new and Λ r v r_new Obtain relevant statistics for the output. and statistics unrelated to output Specifically, this includes: according to the formula

[0114]

[0115]

[0116] Calculate, where They are used to monitor faults that are related to the output and faults that are not.

[0117] Using equation (22), new predictable residual information is obtained, and a new Q statistic is constructed as follows:

[0118] Q r_new =||e r_new || 2 (twenty two)

[0119] In the formula Q r_new Used to monitor predictable residual faults that are independent of the output.

[0120] In practical applications, the input control limits are obtained based on the number of principal components related to the output and the number of principal components unrelated to the output. and output control limits Specifically, this includes: calculation based on formula (23). T 2 The control limits constructed using the Q statistic and the Q statistic are shown below:

[0121]

[0122]

[0123] In equation (23), A is the number of principal components, n is the number of samples, and α is the confidence level; in equation (24), the sample mean μ and sample variance S are determined under a normal distribution and follow the χ² distribution. 2 distributed.

[0124] This invention provides a Tennessee-Eastman simulation experiment to illustrate the above method:

[0125] In the Tennessee-Eastman simulation experiment, the selected input variables were process variables XMEAS (1-36) and manipulated variables XMV (1-11), and the selected output variable was XMEAS (37-41). The number of principal components obtained from two-dimensional cross-validation was A=3, and the system model order obtained from the Akaike information criterion was s=2. The training set selected in the data modeling phase contained 480 normal samples. In the model testing phase, the first 160 samples in the test set were normal data, and samples 161-960 were fault samples, including 15 known fault states.

[0126] Table 2 summarizes the quality-related fault detection results of the dynamic extended fault detection method based on regression relationships. The data in Table 2 show that, within the predictable quality-related dynamic subspace, the dynamic extended fault detection method based on regression relationships exhibits the best monitoring performance for minor changes in faults (5) and (10). For fault (8) reflecting changes in feed concentration, the dynamic extended fault detection method based on regression relationships can detect 83% of quality-related faults. When fault (12) reflecting changes in the condensate temperature inside the compressor occurs, the effective monitoring rate of the dynamic extended fault detection method based on regression relationships reaches 93.49%. For fault (13) reflecting the degree of material reaction, the fault detection rate reaches 91.97%. Therefore, the dynamic extended fault detection method based on regression relationships maintains good detection performance for quality-related faults.

[0127] Table 2. Quality-related fault detection rate (%) of the dynamic extended fault detection method based on regression relationship

[0128]

[0129]

[0130] Figure 2 To assess the monitoring effect of the dynamic extended fault detection method based on regression relationships on fault (13). In the quality-related dynamic subspace, i.e. Figure 2 In (a), the algorithm basically detected the occurrence of quality-related faults near 200 samples, but it could not detect the fault immediately at the moment of its occurrence. Figure 2 In (b), this space is mainly responsible for monitoring the part of the predictable information that does not contribute to the prediction output. It can be seen that this part can extract most of the information that does not contribute, and has a good monitoring effect. Figure 2 (c) Calculated according to formula (24), in Figure 2 In (c), this space is mainly responsible for residual information, from Figure 2 As can be seen, almost all noise and other interference can exceed the control limits. Therefore, it can be concluded that while the fault detection speed of this algorithm is not fast, it can maintain continuous alarms after a fault is detected.

[0131] The following is a summary of the experimental results of the dynamic extended fault detection method based on regression relationships in the detection of quality-independent faults, as shown in Table 3. From the data in Table 3, it is clear that in the quality-related dynamic subspace, the false alarm rate of the proposed method is low for fault (3), less than 5%. For fault (4), the false alarm rate is only 2.38%. For fault (9), the false alarm rate is low. For fault (11), the false alarm rate is only 1.63%. As can be seen from the data in Table 3, for faults 4 and 11, which reflect changes in the internal temperature of the cooling water, the proposed method can effectively detect the occurrence of quality-independent faults, especially achieving a 100% detection rate for fault 4.

[0132] Table 3. False alarm rate (%) of the dynamic extended fault detection method based on regression relationship.

[0133]

[0134] Figure 3 The monitoring effect of the dynamic extended fault detection method based on regression relationship on fault (4) is shown. Figure 3 It can be seen from this that: Figure 3 (a) In the quality-dependent dynamic subspace, the algorithm maintains a low false alarm rate, with only a small number of false alarms between 200 and 400 sampling points. Figure 3(b) In the quality-independent subspace, the dynamic extended fault detection method based on regression relationship can quickly detect the occurrence of quality-independent faults at the fault moment and has a high detection rate. Figure 3 (c) Calculated according to formula (24), in Figure 3 (c) In the residual subspace, the proposed method is relatively sensitive to noise, and the results are almost always above the control limits. The dynamic extended fault detection method based on regression relationships is also relatively sensitive to faults unrelated to quality, and can achieve a lower false alarm rate.

[0135] The embodiments of the present invention provide a simulation experiment of a three-phase flow system to illustrate the above method.

[0136] The following example, a three-phase flow system (TPFS) provided by Cranfield University, is used to verify the effectiveness of the dynamic extended fault detection method based on regression relationships. The variable operating conditions and the scale and complexity of the test bench make the three-phase flow system an ideal benchmark case for evaluating novel multivariate process monitoring technologies based on real experimental data. This benchmark case provides 24 different process variables and 6 specific faults. However, only the first 23 variables are applicable to all fault cases and algorithm training. The main purpose of this system is to provide controllable and measurable flow rates of water, oil, or air for pressurized systems.

[0137] Fault 1 simulates air blockage over time by gradually closing the air pipe valves before the mixing point of the three materials. At the start of data collection, the valves were fully open (normal condition); then, starting from the 1566th sample, the valves were gradually closed, introducing the simulated blockage fault state; then, at the 5181st sample, the fault completely disappeared; subsequently, samples under normal conditions were collected until the sampling was completed. The collected dataset contains a total of 5811 samples. To facilitate comparison between normal and fault data, this section only uses samples before the fault disappeared for experimental verification. Eight variables were selected as input variables, and one variable as output, as shown in Table 4.

[0138] Table 4 Variable Descriptions

[0139]

[0140] In the comparative experiment of quality-related faults, the parameter selected for the dynamic extended fault detection method based on regression relationship was A=4. The process monitoring diagram of this method is shown below. Figure 4 As shown. Calculations show that in the mass-related dynamic subspace, i.e. Figure 4 In (a), the effective detection rate of the dynamic extended fault detection method based on regression relationships is 87.41%, indicating good detection results. Figure 4In (b), it can be seen that before the 3000th sampling point, the detection effect of faults is relatively average, while after 3000, the fault statistics are significantly above the control limit. Figure 4 (c) Calculated according to formula (24), in Figure 4 In (c), it can be seen that between 1500 and 2000 samples (when the fault is slowly introduced), the proposed method does not show a significant effect on noise, but after 2000 samples, the proposed method can better detect noise and other interference information.

[0141] In summary, the dynamic extended fault detection method based on regression relationships directly decomposes predictable dynamic output information for detecting quality-related faults and uses a vector autoregressive model to separate dynamic and static information, thereby maintaining a better detection rate for quality-related faults and reducing the number of quality-related faults. However, when faults develop slowly, the monitoring effect is not as good as when faults are fully developed.

[0142] Dynamic inner partial least squares (DiPLS) is a dynamic partial least squares algorithm proposed in recent years. This algorithm provides explicit explanations of the dynamic internal and external models by modifying the PLS internal model. This invention, based on the Dynamic Inner Partial Least Squares (DiPLS) method, innovatively proposes a quality-related dynamic process modeling method and conducts research on its application. The research results can provide technical support for quality-related dynamic process modeling and fault detection technologies, and have broad application prospects for the health maintenance of dynamic processes in weaponry and large industrial systems.

[0143] This invention relates to a dynamic extended fault detection method based on regression relationships. The main steps of this method are as follows: 1) Considering the regression relationship matrix, extracting dynamic information, and using the DiPLS algorithm to establish a dynamic model between input and output; 2) Establishing dynamic extended fault monitoring indicators based on regression relationships, calculating regression coefficients based on dynamic input and output data, and establishing a PCA model of the residuals; 3) Constructing appropriate statistical indicators for each subspace based on a vector autoregression model. Since this invention is based on multivariate statistical methods, it only requires analyzing relevant data obtained from large and complex equipment systems and establishing relevant mathematical models to achieve fault detection in complex equipment systems. The "regression relationship matrix" is mainly constructed by building a regression relationship between input X and output Y, forming an orthogonal projection space, reducing some highly variable information contained in the mass-independent subspace, and constructing a dynamic extended fault detection method model based on regression relationships. The "monitoring indicators" are designed using a vector autoregression model. Compared with traditional monitoring indicators based on quadratic design, monitoring indicators constructed based on a vector autoregression model can remove dynamic information contained in the score vector, thus improving fault detection accuracy.

[0144] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0145] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A dynamic extended fault detection method based on regression relationships, characterized in that, include: The predictable dynamic information matrix of the standard missile is obtained based on its input and output data matrices. Specifically, the input and output data matrices of the standard missile are normalized and expanded to obtain its dynamic input and output matrices. A regression relation matrix is ​​obtained based on the dynamic input and output matrices of the standard missile. The predictable dynamic information matrix of the standard missile is then obtained based on the regression relation matrix. The standard missile is a missile that has not experienced any malfunctions. The PCA algorithm is used to orthogonally decompose the predictable dynamic information matrix of the standard missile to obtain the main score matrix and the number of principal components related to the output; The input load matrix is ​​obtained based on the main score matrix and the dynamic input information matrix of the standard missile, specifically as follows: according to the formula... Calculation, where Indicates the input load matrix. Represents the main score matrix. Indicates to transpose, The dynamic input information matrix of the standard missile; the dynamic input information matrix of the standard missile is obtained by processing the standard missile using the DiPLS algorithm; The input residual matrix is ​​obtained based on the main score matrix and the input load matrix; The PCA algorithm is used to orthogonally decompose the input residual matrix to obtain the number of principal components that are independent of the output. The input control limits and output control limits are obtained based on the number of principal components related to the output and the number of principal components unrelated to the output; Obtain the input data matrix of the missile to be detected; Based on the input data matrix of the missile to be detected and the input load matrix, statistical quantities related to the output and statistical quantities unrelated to the output are obtained; If the output-related statistic is greater than the input control limit, the missile under test has a quality-related fault; if the output-independent statistic is greater than the output control limit, the missile under test has a quality-independent fault.

2. The dynamic extended fault detection method based on regression relationship according to claim 1, characterized in that, The process of obtaining the predictable dynamic information matrix of the standard missile based on its input and output data matrices specifically includes: The input data matrix and output data matrix of the standard missile are normalized respectively to obtain the normalized input matrix and normalized output matrix of the standard missile; The normalized input matrix and normalized output matrix of the standard missile are respectively augmented to obtain the dynamic input matrix and dynamic output matrix of the standard missile; The predictable dynamic information matrix of the standard missile is obtained based on the dynamic input matrix and dynamic output matrix of the standard missile.

3. The dynamic extended fault detection method based on regression relationship according to claim 1, characterized in that, The step of obtaining the input residual matrix based on the main score matrix and the input load matrix specifically includes: Based on the main score matrix and the input load matrix, the main information related to the predictable dynamic information matrix is ​​obtained; specifically, according to the formula... Calculate, where, This represents the key information related to the predictable dynamic information matrix; The input residual matrix is ​​obtained based on the key information related to the predictable dynamic information matrix.

4. The dynamic extended fault detection method based on regression relationship according to claim 1, characterized in that, The step of obtaining output-related statistics and output-independent statistics based on the input data matrix and input load matrix of the missile to be detected specifically includes: The input data matrix and the output data matrix of the missile to be detected are normalized respectively to obtain the normalized input matrix and the normalized output matrix of the missile to be detected. The normalized input matrix and normalized output matrix of the missile to be detected are respectively augmented to obtain the dynamic input matrix and dynamic output matrix of the missile to be detected; Based on the dynamic input matrix of the missile to be detected and the input load matrix, a score vector related to the output and a score vector unrelated to the output are obtained; Based on the output-related score vector and the output-independent score vector, we obtain the output-related statistics and the output-independent statistics.

Images