Quality-related process monitoring method for nonlinear mixed random distribution systems

Abstract

The present invention discloses a quality-related process monitoring method for a nonlinear mixed random distribution system, and relates to the technical field of fault detection. The method comprises the following steps: sequentially performing feature space mapping and zero-mean normalization processing on a Gaussian variable matrix and a non-Gaussian variable matrix of a device to be detected to obtain a Gaussian variable kernel matrix and a non-Gaussian variable kernel matrix of the device to be detected; calculating quality-related statistics of a Gaussian part and characteristic statistics of a non-Gaussian part of the device according to the Gaussian variable kernel matrix and the non-Gaussian variable kernel matrix; inputting the quality-related statistics of the Gaussian part and the characteristic statistics of the non-Gaussian part of the device to be detected into a trained RVM model to obtain the probability of a quality-related fault occurring in the device to be detected; and judging whether a quality-related fault occurs in the device to be detected according to the probability of the quality-related fault occurring in the device to be detected. The present invention improves the detection rate of quality-related faults and reduces the false alarm rate of fault detection.

Description

Technical Field

[0001] The present invention relates to the technical field of fault detection, and in particular to a quality-related process monitoring method for a nonlinear mixed random distribution system. Background Art

[0002] Existing methods for monitoring the health of large-scale equipment or complex industrial processes are mainly divided into two categories: single-variable process monitoring and multi-variable process monitoring. Single-variable process monitoring can only monitor the fluctuations of a single process variable, which is relatively inefficient in large-scale equipment or industrial process monitoring. Large-scale equipment health monitoring or complex industrial process monitoring is a complex process with multiple variables, multiple outputs, and a large amount of data. Multivariable process monitoring methods, on the other hand, have higher accuracy and faster speed in processing the complex characteristics of big data, and are more robust in the face of various interferences and noise. Therefore, multivariable process monitoring methods have gradually become a commonly used monitoring method for large-scale equipment health monitoring or complex industrial processes.

[0003] In actual equipment health monitoring, due to the influence of external disturbances and other factors, process variables often exhibit more than one distribution characteristic: Gaussian and non-Gaussian process variables coexist, and they also exhibit nonlinear characteristics. Quality-related information is masked by the fluctuations of these nonlinear process variables, making quality-related faults in nonlinear processes difficult to detect. This significantly interferes with the fault detection, location, and troubleshooting of key performance indicators, reducing the efficiency of equipment health monitoring. Therefore, a quality-related process monitoring method for nonlinear hybrid stochastic systems is needed to extract the quality-related information implicit in nonlinear hybrid stochastic process variables and thereby enable quality-related fault detection in nonlinear hybrid stochastic systems. This method first uses the Jarque-Bera test to separate Gaussian and non-Gaussian variables in the monitored process. The MKPLS model is then used to extract quality-related feature information from the Gaussian process variables, and the KICA model is used to extract non-Gaussian features from the non-Gaussian process variables. The extracted feature information is then used to construct an RVM classifier. The RVM classifier then monitors the fluctuations of the quality-related information in the nonlinear hybrid stochastic system, achieving quality-related fault detection in nonlinear non-Gaussian systems. This method achieves a high fault detection rate and a low false alarm rate. Summary of the Invention

[0004] The purpose of the present invention is to provide a quality-related process monitoring method for a nonlinear mixed random distribution system, which, to a certain extent, improves the detection rate of quality-related faults and reduces the false alarm rate of fault detection.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A quality-related process monitoring method for a nonlinear mixed random distribution system, comprising:

[0007] Obtaining a Gaussian variable matrix and a non-Gaussian variable matrix of the device to be detected; the Gaussian variable matrix includes sample values ​​of all process variables that meet Gaussian distribution characteristics; the non-Gaussian variable matrix includes sample values ​​of all process variables that meet non-Gaussian distribution characteristics;

[0008] Performing feature space mapping and zero-mean normalization processing on the Gaussian variable matrix and the non-Gaussian variable matrix of the device to be detected in sequence to obtain the Gaussian variable kernel matrix and the non-Gaussian variable kernel matrix of the device to be detected;

[0009] Calculate the quality-related statistics of the Gaussian part and the characteristic statistics of the non-Gaussian part of the device according to the Gaussian variable kernel matrix and the non-Gaussian variable kernel matrix of the device to be detected;

[0010] Inputting the quality-related statistics of the Gaussian part and the characteristic statistics of the non-Gaussian part of the device to be detected into the trained RVM model to obtain the probability of quality-related failure of the device to be detected;

[0011] Whether the device to be detected has a quality-related failure is determined according to the probability that the device to be detected has a quality-related failure.

[0012] Optionally, the process of determining the trained RVM model is as follows:

[0013] Obtaining a process variable vector and a quality variable vector corresponding to each sampling moment when the sample device is in normal operating conditions, and obtaining an input data matrix and an output data matrix; the process variable vector includes the value of each process variable at the same sampling moment; the quality variable vector includes the value of each quality variable at the same sampling moment; the normal operating condition is a state in which the device is free of faults;

[0014] Using the Jarque-Bera test to process the input data matrix to obtain the distribution characteristics of each process variable; the distribution characteristics include satisfying Gaussian distribution characteristics and satisfying non-Gaussian distribution characteristics;

[0015] Obtaining a Gaussian variable matrix and a non-Gaussian variable matrix of the sample device according to the distribution characteristics of each process variable;

[0016] Performing feature space mapping and zero-mean normalization processing on the Gaussian variable matrix and the non-Gaussian variable matrix of the sample device in sequence to obtain the Gaussian variable kernel matrix and the non-Gaussian variable kernel matrix of the sample device;

[0017] Calculating quality-related statistics of the Gaussian part and feature statistics of the non-Gaussian part of the sample device according to the Gaussian variable kernel matrix and the non-Gaussian variable kernel matrix of the sample device;

[0018] The RVM model is trained with the quality-related statistics of the Gaussian part of the sample device and the characteristic statistics of the non-Gaussian part as input and the output data matrix of the sample device as output to obtain a trained RVM model.

[0019] Optionally, before using the Jarque-Bera test to process the input data matrix to obtain the distribution characteristics of each process variable, the method further includes:

[0020] The input data matrix and the output data matrix are standardized to obtain a standard input data matrix and a standard output data matrix respectively.

[0021] Optionally, calculating the quality-related statistics of the Gaussian part and the feature statistics of the non-Gaussian part of the sample device according to the Gaussian variable kernel matrix and the non-Gaussian variable kernel matrix of the sample device specifically includes:

[0022] For any sampling moment, the quality-related statistics of the Gaussian part and the characteristic statistics of the non-Gaussian part of the sample device at the sampling moment are calculated based on the elements corresponding to the sampling moment in the Gaussian variable kernel matrix of the sample device and the elements corresponding to the sampling moment in the non-Gaussian variable kernel matrix of the sample device.

[0023] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0024] The present invention classifies the process variables of the equipment to be tested into those that conform to Gaussian distribution and those that conform to non-Gaussian distribution, calculates the quality-related statistics of the Gaussian part and the characteristic statistics of the non-Gaussian part, and judges whether the equipment to be tested has a quality-related fault based on the two statistics. To a certain extent, the detection rate of quality-related faults is improved and the false alarm rate of fault detection is reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0026] Figure 1 A flow chart of a quality-related process monitoring method for a nonlinear mixed random distribution system provided by an embodiment of the present invention;

[0027] Figure 2 It is the fluctuation diagram of y before and after the fault IDV(7) occurs;

[0028] Figure 3 A diagram showing the detection results of a fault IDV (7) by a quality-related process monitoring method for a nonlinear mixed random distribution system provided by an embodiment of the present invention;

[0029] Figure 4 It is the fluctuation diagram of y before and after the fault IDV (14) occurs;

[0030] Figure 5 A diagram showing the detection results of a fault IDV (14) by a quality-related process monitoring method for a nonlinear mixed random distribution system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0031] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0032] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0033] like Figure 1 As shown, the present invention provides a quality-related process monitoring method for a nonlinear mixed random distribution system, comprising:

[0034] Step 101: Obtain a Gaussian variable matrix and a non-Gaussian variable matrix for the device to be inspected. The Gaussian variable matrix includes sample values ​​of all process variables that satisfy Gaussian distribution characteristics; the non-Gaussian variable matrix includes sample values ​​of all process variables that satisfy non-Gaussian distribution characteristics. Process variables are measurable factors that may affect quality variables throughout the entire process, such as iron ore content, oxygen content, furnace temperature, and other indicators in the steelmaking industry. Quality variables are the ultimate indicators of interest and are measurable, such as steel thickness and other indicators in the steelmaking industry.

[0035] Step 102: performing feature space mapping and zero-mean normalization processing on the Gaussian variable matrix and the non-Gaussian variable matrix of the device to be detected in sequence to obtain the Gaussian variable kernel matrix and the non-Gaussian variable kernel matrix of the device to be detected.

[0036] Step 103: Calculate the quality-related statistics of the Gaussian part and the characteristic statistics of the non-Gaussian part of the device according to the Gaussian variable kernel matrix and the non-Gaussian variable kernel matrix of the device to be detected.

[0037] Step 104: Inputting the quality-related statistics of the Gaussian part and the characteristic statistics of the non-Gaussian part of the device to be detected into the trained RVM model to obtain the probability of the quality-related failure of the device to be detected.

[0038] Step 105: Determine whether the device to be detected has a quality-related fault based on the probability of the device to be detected having a quality-related fault.

[0039] In practical applications, the process of determining the trained RVM model is as follows:

[0040] The process variable vectors and quality variable vectors corresponding to each sampling moment when the sample device is in normal operating conditions are obtained to obtain an input data matrix and an output data matrix. The process variable vector includes the values ​​of each process variable at the same sampling moment; the quality variable vector includes the values ​​of each quality variable at the same sampling moment. The normal operating condition is the state of the device being fault-free. The sample device and the device to be tested are of the same type. The specific physical meaning of the input data matrix and the output data matrix is ​​determined by different process industries. For example, in the steelmaking process, the input data can be samples of process variables that affect quality indicators, such as temperature; the output data can be samples of quality indicators, such as steel thickness.

[0041] The input data matrix is ​​processed using a Jarque-Bera test to obtain the distribution characteristics of each process variable; the distribution characteristics include those that satisfy Gaussian distribution characteristics and those that satisfy non-Gaussian distribution characteristics.

[0042] A Gaussian variable matrix and a non-Gaussian variable matrix of the sample device are obtained according to the distribution characteristics of each process variable.

[0043] The Gaussian variable matrix and the non-Gaussian variable matrix of the sample device are sequentially subjected to feature space mapping and zero-mean normalization processing to obtain the Gaussian variable kernel matrix and the non-Gaussian variable kernel matrix of the sample device.

[0044] The quality-related statistics of the Gaussian part and the characteristic statistics of the non-Gaussian part of the sample device are calculated according to the Gaussian variable kernel matrix and the non-Gaussian variable kernel matrix of the sample device.

[0045] The quality-related statistics of the Gaussian part of the sample device and the characteristic statistics of the non-Gaussian part are used as inputs, and the output data matrix of the sample device is used as output to train the RVM model to obtain a trained RVM model.

[0046] In practical applications, before using the Jarque-Bera test to process the input data matrix to obtain the distribution characteristics of each process variable, the method further includes:

[0047] The input data matrix and the output data matrix are standardized to obtain a standard input data matrix and a standard output data matrix respectively.

[0048] In practical applications, the quality-related statistics of the Gaussian part and the characteristic statistics of the non-Gaussian part of the sample device are calculated based on the Gaussian variable kernel matrix and the non-Gaussian variable kernel matrix of the sample device, specifically including:

[0049] For any sampling moment, the quality-related statistics of the Gaussian part and the characteristic statistics of the non-Gaussian part of the sample device at the sampling moment are calculated based on the elements corresponding to the sampling moment in the Gaussian variable kernel matrix of the sample device and the elements corresponding to the sampling moment in the non-Gaussian variable kernel matrix of the sample device.

[0050] In practical applications, the input data matrix and the output data matrix are respectively standardized to obtain a standard input data matrix and a standard output data matrix, specifically including:

[0051] Assume that the input data matrix is ​​under normal working conditions and the output data matrix Contains n training samples, m process variables and p quality variables, where the process variable vector Quality variable vector i=1,2,...,n,x i is a vector of 1 row and m columns, each column represents the value of a process variable, x i represents the i-th sampling of all process variables, y i A vector with 1 row and p columns, each column represents the value of a quality variable, y i represents the i-th sampling of all quality variables, and the following standardization is performed first:

[0052]

[0053] Where: and s x are the mean and standard deviation of the matrix X respectively; and s y are the mean and standard deviation of matrix Y respectively.

[0054] In practical applications, the Jarque-Bera test is used to process the input data matrix X to obtain the distribution characteristics of each process variable. According to the distribution characteristics of each process variable, the Gaussian variable matrix and the non-Gaussian variable matrix of the sample device are obtained, specifically including:

[0055] First, calculate the JB statistic for each process variable:

[0056]

[0057] Where S and K are the skewness and kurtosis of the training samples, respectively. Each process variable has its own skewness and kurtosis.

[0058] Then, the JB statistic for each process variable is compared with For comparison, is the critical value of the chi-square distribution with 2 degrees of freedom and 0.05 confidence level. If If , the variable belongs to Gaussian distribution; otherwise, it belongs to non-Gaussian distribution.

[0059] According to the distribution characteristics of the variables, X is divided into two parts of Gaussian variable matrix and non-Gaussian variable matrices where X g Satisfies Gaussian distribution, X n Satisfies non-Gaussian distribution, m1+m2=m.

[0060] In practical applications, considering the nonlinearity of process variables, the following Gaussian kernel function is used to transform the input vector into the feature space F:

[0061]

[0062] Where: i∈[1,n],j∈[1,n], c is the width of the Gaussian kernel function, and c is selected based on the principle that the model's false alarm rate for faults under normal operating conditions is less than 10%. The Gaussian kernel function is also the basis function of the Bayesian classifier constructed in this embodiment of the present invention, the RVM model.

[0063] X g After mapping from the original space to the feature space F, it becomes:

[0064]

[0065] For Φ raw Perform zero-mean normalization:

[0066]

[0067]

[0068] in: is Φ raw The mean of 1 n Represents an n-dimensional column vector whose elements are all 1. Similarly, X n After mapping from the original space to the feature space F, the process data Φ can be obtained after zero-mean preprocessing. n .

[0069] In order to avoid Φ gExplicit use of, defining the Gaussian kernel matrix K g :

[0070]

[0071] Therefore, calculate K g :

[0072]

[0073] in: K raw The element K in row i and column j of raw (i,j)=K(x i ,x j ).

[0074] Through the KPLS algorithm, under the condition that the number of principal components is equal to a, construct Φ g The nonlinear regression relationship between y and y is as follows:

[0075]

[0076]

[0077] Among them, the number of principal components a is a parameter in the KPLS algorithm, which is determined by the cross-validation method, and C is Φ g and Y, To model the error, T and U are the latent variable matrices of X and Y respectively.

[0078] Since Φ g It cannot be obtained explicitly and Φ cannot be directly established g and Y regression model. We can establish K g and Y regression model, indirectly establish Φ g and the regression model of Y. Substituting formula (10) into formula (9), we get:

[0079]

[0080] make:

[0081] M=U(T Τ K g U) -1 T Τ Y(12)

[0082] so:

[0083]

[0084] Obviously, equation (13) establishes a linear relationship between the kernel matrix and the output matrix. Τ Perform SVD decomposition, that is:

[0085]

[0086] in:

[0087] Assumptions is the Gaussian part of the sample being tested. Then, x g After zero-mean preprocessing, it is mapped into the feature space F and recorded as φ(x g ). k g (x g ) is φ(x g ) corresponds to the kernel vector.

[0088] The quality-related characteristic statistics of nonlinear Gaussian distributed process variables are calculated as follows:

[0089]

[0090] In the feature space, the covariance matrix S of the non-Gaussian process variables is:

[0091]

[0092] Compute and scale the non-Gaussian kernel matrix K:

[0093]

[0094]

[0095] K n Perform eigenvalue decomposition:

[0096] λα=K n α(19)

[0097] Where λ represents the eigenvalue and α represents the orthogonal eigenvector. K is selected by the relative eigenvalue criterion n The u largest eigenvalues ​​of :

[0098]

[0099] where λ i (i=1,2,...n) represents K y The characteristic value of α1, α2, … α is obtained by formula (19) u Represents the orthogonal eigenvectors corresponding to the u largest eigenvalues. On this basis, the orthogonal eigenvectors corresponding to the u largest eigenvalues ​​of S are as follows:

[0100]

[0101] The whitening matrix V is:

[0102] V=[β1,β2,…β u ]=Φ n HΛ -1 / 2 (twenty two)

[0103] Among them: Λ=diag(λ1,λ2,…λ u ),

[0104] After normalizing V, it becomes:

[0105]

[0106] In order to whiten the mapped data in the feature space, the i-th (i=1, 2, ..., n) whitening score vector is calculated as follows:

[0107]

[0108] Where: φ n (x i ) is Φ n The i-th column of Z, z(i) is the i-th column of Z, k n (i) is K n The i-th row of .

[0109] Then, by maximizing the non-Gaussianity of the independent element s, we make E{ss Τ}=D=diag{λ1,λ2,...λ e ) holds true, and obtain e (e≤u) principal independent components (ICs) from Z.

[0110] s=C Τ z(i) (25)

[0111] in: C Τ C=D. The obtained ICs are sorted according to the size of the variance. Define the standardized ICs (s n )for:

[0112]

[0113] in: and

[0114] Let initialize matrix for:

[0115]

[0116] in is the normalized form of C, I erepresents the e-dimensional identity matrix, and 0 represents the e×(ue)-dimensional zero matrix. Obtained by FastICA algorithm, so s n It has the greatest non-Gaussianity.

[0117] Get C n After that, the core ICs can be obtained as follows:

[0118]

[0119] Where s(i) is the i-th row of s.

[0120] Assumptions is the non-Gaussian part of the sample being tested. Then, x n After zero-mean preprocessing, it is mapped into the feature space F and recorded as φ(x n ). k n (x n ) is φ(x n ) corresponds to the kernel vector.

[0121] The characteristic statistics of nonlinear non-Gaussian distributed process variables are calculated as follows:

[0122]

[0123]

[0124] Therefore, the Gaussian variable matrix and the non-Gaussian variable matrix of the sample device are sequentially subjected to feature space mapping and zero-mean normalization processing to obtain the Gaussian variable kernel matrix and the non-Gaussian variable kernel matrix of the sample device, specifically including:

[0125] According to the distribution characteristics of the process variables after standardization, an appropriate amount of training samples Divided into Gaussian parts and the non-Gaussian part

[0126] Will and Mapped into the feature space F, and after zero-mean preprocessing, they are and The corresponding kernel matrices are and

[0127] In practical applications, for any sampling moment, the quality-related statistics of the Gaussian part and the characteristic statistics of the non-Gaussian part of the sample device at the sampling moment are calculated based on the elements corresponding to the sampling moment in the Gaussian variable kernel matrix of the sample device and the elements corresponding to the sampling moment in the non-Gaussian variable kernel matrix of the sample device, specifically including:

[0128] Extracting quality-related statistics of Gaussian parts using MKPLS model

[0129]

[0130] in: represents the i-th training sample Statistics, Representation matrix The i-th column, i∈[1,t], the i-th training sample is collected at the i-th sampling time, is calculated using the training samples collected at the i-th moment.

[0131] The KICA model is used to extract the characteristic statistics of the non-Gaussian part:

[0132]

[0133]

[0134] in: and SPE train (i) represents the I of the i-th training sample 2 and the SPE statistic, Representative Matrix The i-th column of .

[0135] In practical applications, the RVM model is trained using the quality-related statistics of the Gaussian part of the sample device and the feature statistics of the non-Gaussian part as input and the output data matrix of the sample device as output to obtain a trained RVM model, specifically including:

[0136] Let the matrix As the input of the RVM model; its corresponding target response class label is where y t ∈{0,1}. When a quality-related failure occurs, y t =1, otherwise under normal working conditions y t =0.

[0137] Typically, predictions are made using the following function:

[0138] y(x;θ)=φθ (34)

[0139] Among them: θ=[θ0,θ1,θ2...θ t ] Τ is the weight, φ(x t )=[1,K(x t ,x1),K(x t,x2)...K(x t ,x t )] Τ is the basis function. The sparse restriction is implemented by introducing a hyperparameter vector α on θ to avoid overfitting:

[0140]

[0141] Among them, the hyperparameters Indicates the weight strength.

[0142] In classification problems, the posterior probabilities of different classes given an input x are crucial and are determined by the sigmoid function as shown below:

[0143]

[0144] According to the Bernoulli distribution, we can get P(y|θ):

[0145]

[0146] When α is fixed, P(θ|y,α)∝P(y|θ)P(θ|α), and the maximum value of θ can be obtained using the following formula:

[0147]

[0148] Each weight is associated with a hyperparameter, A = diag{α0,α1...α t Then, the most likely θ MP It can be obtained by iterative weighted least squares algorithm. Differentiating (38) twice yields the following equation:

[0149]

[0150] Where B=diag{β0,β1...β t}, β i =σ{y(x i ;θ)}[1-σ{y(x i ;θ)}]. Because Under the Gaussian approximation, the following two statistics can be obtained:

[0151] Σ=(φ Τ Bφ+A) -1 (40)

[0152] θ MP =Σφ Τ By(41)

[0153] The hyperparameter α is updated using the following formula:

[0154]

[0155] γ i =1-α i Σ ii (43)

[0156] where Σ ii is the i-th diagonal element in formula (40). i When converged, the corresponding correlation weight θ can be obtained i and the corresponding correlation vector φ(x i ). The remaining hyperparameters tend to infinity, and the corresponding weights are 0, thus ensuring the sparsity of RVM. When all relevant vectors and weights are obtained, the RVM model training ends.

[0157] In practical applications, the quality-related statistics of the Gaussian part and the feature statistics of the non-Gaussian part of the device to be detected are input into a trained RVM model to obtain the probability of the device to be detected having a quality-related fault. Determining whether the device to be detected has a quality-related fault based on the probability of the device to be detected having a quality-related fault specifically includes:

[0158] Calculate the Gaussian quality-related statistics and non-Gaussian feature statistics of the device to be tested, and use the Bayesian classifier to determine whether it has a quality-related fault. Assume that the statistics of the device to be tested are The corresponding class is:

[0159] y(x new θ r )=φ new θ r (44)

[0160] where θ r is the relevant weight, φ new =[K(x new ,φ(x1)),K(x new ,φ(x2))...K(x new ,φ(x r ))] is the basis function, φ(x r ) is the correlation vector, y(x new θ r ) represents the relevant weight θ r Under the constraint of x new A mapping relationship between and y.

[0161] The probability of quality-related failures occurring is calculated as follows:

[0162]

[0163] When P(y=1|θr )≥0.5, it is considered that y=1, indicating that the current sample has a quality-related fault; otherwise, it is considered that the current sample is in normal operation.

[0164] The present invention provides an embodiment. Considering the characteristics of large and complex equipment systems, such as multiple indicators, high dimensionality, large sample sizes, and the presence of process variables and key performance indicators, similar to the Tennessee Eastman Process (TEP), the TEP is used to verify the effectiveness of the proposed method. The method proposed in the present invention is verified using data collected during a TEP experiment.

[0165] The TE process consists of 11 manipulated variables XMV(1-11) and 41 measured variables XMEAS(1-41), of which 22 are process variables and 19 are analytical variables. The 11 manipulated variables and 22 process variables are selected to form the input matrix X, and the product G(XMEAS(35)) is selected as the output y. The training set consists of 500 normal operating condition samples; the test set selects 11 different types of faults, namely IDV(1,3-7,9-12,14), each type consisting of 960 samples, of which the first 160 are normal samples and the last 800 are single-type fault samples. Faults IDV(1,3-7,9-12) are weak quality-related faults, and IDV(14) is a quality-independent fault.

[0166] Considering that the data collected during the TE process has a time delay, the current output is caused by the input at a certain time in the past, y(t) = f(t-Δt). The process monitoring performance indicator uses the fault detection absolute deviation (FDAD), which is defined as follows:

[0167]

[0168] Where: N nea N represents the total number of fault samples detected by the process monitoring algorithm after the fault occurs; tfs It represents the total number of samples where the actual key performance indicators exceed the standard after the fault occurs. After determining the quality variable y, the maximum value y in the training sample is selected. max and the minimum value y min The normal fluctuation range of the quality variable y[y min ,y max ], if the key performance indicators of the sample after the failure exceed [y min ,y max ], it is determined that the current sample has a quality-related failure. N tsIt represents the total number of samples collected after the fault occurs. In process monitoring, the lower the absolute deviation of the fault detection algorithm, the better.

[0169] Taking the weak quality-related fault IDV (7) as an example, Figure 2 The figure shows the fluctuation of the quality variable y before and after the occurrence of the weak quality-related fault IDV (7). The two dotted lines represent the maximum and minimum values ​​of the quality variable y under normal working conditions, and the solid line represents the quality variable y. Figure 3 The figure shows the detection results of the fault IDV (7) by the method proposed in the present invention. The solid line represents the fault detection results of the method proposed in the present invention for the current sample. Class 0 indicates normal, and class 1 indicates a quality-related fault. Figure 3 It can be seen that when fault IDV (7) occurs, QRBC can detect the fault in time and issue an alarm. Table 1 lists the absolute deviations of fault detection by QRBC for TE process fault IDV (1, 3-7, 9-12, 14). As shown in Table 1, QRBC maintains a high fault detection rate and a low fault false alarm rate when detecting quality fault IDV (7), which proves the effectiveness and accuracy of the proposed method.

[0170] Table 1 FDADs (%) of IDV (1, 3-7, 9-12, 14) detected by QRBC

[0171]

[0172]

[0173] As shown in Table 1, QRBC still has good process monitoring performance when detecting other weak quality-related faults, maintaining a high fault detection rate and a low fault false alarm rate. In summary, when process data contains quality-related faults, the proposed method has a good detection effect on quality-related faults.

[0174] Taking the quality-independent fault IDV (14) as an example, Figure 4 The fluctuation of the quality variable y before and after the occurrence of fault IDV (14) is shown. Figure 5 The results of QRBC's detection of quality-independent fault IDV (14) are shown. Figure 5 It can be seen that QRBC can accurately identify the fault nature of IDV (14) and there is no false alarm. As shown in Table 1, the absolute deviation of fault detection by QRBC when detecting quality-independent fault IDV (14) is 0%, and there is no false alarm. This proves that when a quality-independent fault occurs, the method proposed in the present invention can accurately identify the fault type and reduce the interference of quality-independent faults on process monitoring.

[0175] Combined with TE process experimental data, the following results were obtained: QRBC can effectively detect quality-related faults in nonlinear hybrid random systems, providing reliable process monitoring performance and reducing safety and product quality issues caused by faults in industry. Furthermore, this method can accurately identify quality-independent faults, significantly reducing false alarms in process monitoring. These experimental results verify the effectiveness of the method provided by the embodiments of the present invention.

[0176] The present invention has the following technical effects:

[0177] The quality-related process monitoring method for a nonlinear mixed random distribution system provided by the present invention is a quality-related process monitoring method for a nonlinear mixed random distribution system. It can realize real-time monitoring of large-scale equipment, complex industrial processes and equipment, alarm for quality-related faults, improve the reliability and safety of complex systems, reduce the occurrence of major accidents, and has strong robustness. It is an early quality-related process monitoring method for a nonlinear mixed random distribution system, which can detect early faults in time and avoid unnecessary economic losses and casualties.

[0178] The present invention utilizes the extracted feature information to construct a Bayesian classifier to monitor quality-related information in a nonlinear mixed random distribution system, which not only avoids the threshold construction in the traditional method, but also makes up for the disadvantage of insufficient single statistical information.

[0179] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0180] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A quality-related process monitoring method for a nonlinear mixed random distribution system, characterized in that: include: Obtaining a Gaussian variable matrix and a non-Gaussian variable matrix of the device to be detected; the Gaussian variable matrix includes sample values ​​of all process variables that meet Gaussian distribution characteristics; the non-Gaussian variable matrix includes sample values ​​of all process variables that meet non-Gaussian distribution characteristics; Performing feature space mapping and zero-mean normalization processing on the Gaussian variable matrix and the non-Gaussian variable matrix of the device to be detected in sequence to obtain the Gaussian variable kernel matrix and the non-Gaussian variable kernel matrix of the device to be detected; Calculate the quality-related statistics of the Gaussian part and the characteristic statistics of the non-Gaussian part of the device according to the Gaussian variable kernel matrix and the non-Gaussian variable kernel matrix of the device to be detected; specifically, use the MKPLS model to obtain the quality-related statistics of the Gaussian part; use the KICA model to obtain the characteristic statistics of the non-Gaussian part; Inputting the quality-related statistics of the Gaussian part and the characteristic statistics of the non-Gaussian part of the device to be detected into the trained RVM model to obtain the probability of the quality-related failure of the device to be detected; Whether the device to be detected has a quality-related failure is determined according to the probability that the device to be detected has a quality-related failure.

2. A quality-related process monitoring method for a nonlinear mixed random distribution system according to claim 1, characterized in that: The determination process of the trained RVM model is as follows: Obtaining a process variable vector and a quality variable vector corresponding to each sampling moment when the sample device is in normal operating conditions, and obtaining an input data matrix and an output data matrix; the process variable vector includes the value of each process variable at the same sampling moment; the quality variable vector includes the value of each quality variable at the same sampling moment; the normal operating condition is a state in which the device is free of faults; Using the Jarque-Bera test to process the input data matrix to obtain the distribution characteristics of each process variable; the distribution characteristics include satisfying Gaussian distribution characteristics and satisfying non-Gaussian distribution characteristics; Obtaining a Gaussian variable matrix and a non-Gaussian variable matrix of the sample device according to the distribution characteristics of each process variable; Performing feature space mapping and zero-mean normalization processing on the Gaussian variable matrix and the non-Gaussian variable matrix of the sample device in sequence to obtain the Gaussian variable kernel matrix and the non-Gaussian variable kernel matrix of the sample device; Calculating quality-related statistics of the Gaussian part and feature statistics of the non-Gaussian part of the sample device according to the Gaussian variable kernel matrix and the non-Gaussian variable kernel matrix of the sample device; The RVM model is trained with the quality-related statistics of the Gaussian part of the sample device and the characteristic statistics of the non-Gaussian part as input and the output data matrix of the sample device as output to obtain a trained RVM model.

3. A quality-related process monitoring method for a nonlinear mixed random distribution system according to claim 2, characterized in that: Before using the Jarque-Bera test to process the input data matrix to obtain the distribution characteristics of each process variable, the method further includes: The input data matrix and the output data matrix are standardized to obtain a standard input data matrix and a standard output data matrix respectively.

4. A quality-related process monitoring method for a nonlinear mixed random distribution system according to claim 2, characterized in that: Calculating the quality-related statistics of the Gaussian part and the feature statistics of the non-Gaussian part of the sample device according to the Gaussian variable kernel matrix and the non-Gaussian variable kernel matrix of the sample device specifically includes: For any sampling moment, the quality-related statistics of the Gaussian part and the characteristic statistics of the non-Gaussian part of the sample device at the sampling moment are calculated based on the elements corresponding to the sampling moment in the Gaussian variable kernel matrix of the sample device and the elements corresponding to the sampling moment in the non-Gaussian variable kernel matrix of the sample device.

Images