A fault diagnosis method for intermittent processes based on M-MForeCA
By using the M-MForeCA method during intermittent processes, the stability and transition phases are divided and dynamically modeled during the transition process, the problem of uncaptured multi-stage change characteristics in the existing technology is solved, and the accuracy of fault detection is improved and efficiency.
Patent Information
- Application Number
- CN202311151234.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-09-07
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2043-09-07
AI Technical Summary
The existing MForeCA method ignores the multi-stage production characteristics in intermittent process fault diagnosis and cannot effectively capture process correlation changes at different stages, resulting in low fault detection rate and high false alarm rate.
Using the intermittent process fault diagnosis method based on M-MForeCA, the stability stage and transition process are divided by similarity factor SI, and a multi-directional predictable component analysis model is established, and dynamic modeling is used for sliding window data during the transition process to improve fault detection. accuracy.
It improves the fault detection rate, reduces the false alarm rate, accurately extracts predictable characteristics of different operation stages, improves the process monitoring effect, and can detect fault occurrences earlier.
Smart Images

Figure CN117519063B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of industrial process fault diagnosis, and in particular to an intermittent process fault diagnosis method based on M-MForeCA. Background Art
[0002] In recent years, intermittent production has gradually become an important production method in modern industry. Fault diagnosis of intermittent processes is crucial to ensuring safe production operation, improving product quality, reducing cost losses and improving production efficiency.
[0003] Forecastable Component Analysis (ForeCA) is a new data dimension reduction and feature extraction algorithm proposed by Georg M. Goerg in 2013. It is mainly used to extract predictable characteristics from process data information. It has been applied to fund of funds forecasting in the financial field, data classification in macroeconomics, and fault diagnosis of continuous industrial processes. In order to process the three-dimensional data of intermittent processes, researchers further proposed Multiway Forecastable Component Analysis (MForeCA), which extracts predictable components from intermittent process production data information by transforming three-dimensional historical data into two-dimensional multi-directional expansion processing, and then captures the dynamic change characteristics of the system. However, MForeCA does not consider the multi-stage production characteristics of intermittent processes, and only establishes a single statistical model to perform online fault diagnosis on variable data at any stage of subsequent batches at any time, ignoring the large differences in process characteristics across stages. Therefore, how to explore the process correlation changes that often occur in different stages of intermittent production has become a challenging topic in intermittent process monitoring and fault diagnosis. Summary of the invention
[0004] The purpose of the present invention is to address the above-mentioned defects in the prior art and provide an intermittent process fault diagnosis method based on M-MForeCA. This method establishes a fault diagnosis model in stages for online monitoring, overcomes the shortcoming that a single model cannot reflect the stage change characteristics, improves the fault detection rate and reduces the false alarm rate.
[0005] The present invention provides an intermittent process fault diagnosis method based on M-MForeCA, and its technical solution comprises the following steps:
[0006] 1. Collect normal batch historical data of intermittent process , I is the batch number, J is the number of variables, K is the number of sampling times; Preprocessing: First, Get a two-dimensional matrix based on batch expansion , and then normalize each column to a dimensionless form, and then restore it to a three-dimensional form and regard it as a function of K Sampled 2D data slices composition;
[0007] 2. Calculate the similarity factor at each moment SI , used for batch processes K Subsampled data slices Conduct reproducible analysis;
[0008] Utilization SI Divide the intermittent process into a stable phase and a transition process;
[0009] 4. Assume that the batch process is separated N The multi-directional predictable component analysis (MForeCA) algorithm was used to establish N Fault diagnosis model for each stable stage;
[0010] 5. In view of the characteristics of the transition process, which has a short period, large fluctuations, and strong correlation between the dynamic time series before and after, the dynamic MForeCA method based on sliding window data is used to establish the fault diagnosis model of each transition process;
[0011] 6. Collect a new batch of k The sample time data is standardized according to the mean and standard deviation of the corresponding time when modeling. ;
[0012] VII. Confirmation k The stable sub-stage or transition process belongs to the corresponding model. W d calculate and ;
[0013] 8. Calculation and Statistics are compared with the control limits in the corresponding model. Once a statistic crosses the control limit, it is considered that a fault has occurred.
[0014] Preferably, in step 2, SI The definition formula is:
[0015] (1)
[0016] In the formula, for No. i According to the definition of formula (1), SI The value range is [0, 1).
[0017] Preferably, in step three, the stage division steps are:
[0018] I. Calculate the value of each sampling moment according to formula (1) SI Values and plot SI Graphs;
[0019] II. Determine the approximate starting time of each stable phase and the rough end moment , when the intermittent process tends to a new stable stage, SI The curve will drop significantly, and the time corresponding to the lowest point of the drop process is , if the intermittent process is initially a stable phase, then =1, when a stable phase tends to end, SI The curve will rise continuously or suddenly from a lowest point. The time corresponding to the lowest point of the rising process is taken as , if the intermittent process ends with a stable phase, then The value is the maximum sampling time K ;
[0020] III. Using the kernel density estimation method to estimate the stable stages in the initial division SI Determine each stage based on the value SI Threshold value;
[0021] IV. Determine the exact start time of each stable sub-phase and the exact end time ,exist SI On the curve graph, each threshold line has an intersection with the corresponding descending curve before and ascending curve after the stable stage. The first sampling time after the intersection with the descending curve is defined as the accurate starting time. The last sampling time before the intersection with the rising curve is the accurate end time , if the intermittent process is initially in a stable stage, then =1, if the intermittent process ends in a stable stage, then = K ;
[0022] V. The remaining part after removing the stable phase is identified as the transition process.
[0023] Preferably, in step 4, the steps of establishing the fault diagnosis model using the MForeCA algorithm in the stable stage are:
[0024] I. Assume that any n ( n =1,2,..., N ) three-dimensional data sets of stable sub-stages ,in K n is the sampling number of this stage, and the three-dimensional data of the sub-stage Expand into Normalize by column, then Rearrange into , and transpose to get ;
[0025] II. After the expansion The predictability analysis was performed through ForeCA transformation to obtain the predictable component matrix Y , the transformation formula is:
[0026] (2)
[0027] In the formula is the load transformation matrix;
[0028] III. Use the cumulative contribution rate of predictability to determine the number of main predictable components, as shown in formula (8):
[0029] (8)
[0030] Usually order Sure d value ,Pick ,according to Calculate the principal predictable component matrix ;
[0031] IV. Construction Fore 2 Statistics and SPE Statistics, as shown in formulas (9) and (10):
[0032] (9)
[0033] (10)
[0034] In the above two formulas, , Respectively represent matrices Y d , No. i Column vector, , yes , e ( i )yes Column vector of ;
[0035] V. Calculation using kernel density estimation Fore 2 and SPE Control limits for the statistic.
[0036] Preferably, in step II, determine W Here’s how:
[0037] i) Assumptions X The mean is , the covariance matrix is , the autocovariance function is , obtained by Fourier transform X The spectral function is:
[0038] (3)
[0039] In the formula i Represents an imaginary unit, k is an integer, is the angular frequency;
[0040] ii) Order , w yes W A column vector of , then the spectral function , y The spectral density function is defined as:
[0041] (4)
[0042] In the formula yes y The variance of
[0043] iii) Using information entropy, uncertainty is defined as:
[0044] (5)
[0045] In the formula, a >0 is the logarithmic base; for a white noise signal with a completely flat spectral density , cannot be predicted, so there is ,and ;
[0046] iv) Utilization y The uncertainty defines the predictability:
[0047] (6)
[0048] v) To achieve the extraction of predictable components, according to , the ForeCA method needs to find the optimal load vector w , thus establishing the following optimization problem:
[0049] (7)
[0050] Solve the above optimization problem and get the load vector matrix , the predictable component matrix is calculated by formula (2) ,and The corresponding predictability Arrange in order from high to low.
[0051] Preferably, in step 5, the steps of the transition process dynamic MForeCA modeling method are as follows:
[0052] I. Assume that a certain moment of the transition process is in the first k Sampling time, take the adjacent h 2D data slices constitute a 3D Total 2 h +1 continuous two-dimensional data slice is used as the training sample for establishing the fault diagnosis model at this moment in the transition process; k The training data window keeps sliding backwards; if the intermittent process starts or ends with a transition process, the number of adjacent samples in front or behind the training data set of the sliding window is less than h , there is a sliding window length < 2 h +1 situation;
[0053] II. Training samples at this moment of the transition process Transform to two dimensions Then ForeCA analysis is performed to obtain the W d and Y d , calculated by formula (9) and formula (10) respectively Fore 2 and SPE Statistics, and use the kernel density estimation method to obtain the control limits of the two statistics at that moment.
[0054] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0055] The present invention divides the three-dimensional data of the intermittent process into stages using the similarity factor SI, establishes a unified process monitoring model for each stable stage, and establishes a fault diagnosis model based on the sliding window data for the transition process at each moment, thereby forming the M-MForeCA modeling method of the present invention. Finally, the predictable components and new statistics of the new batch of test data sets are calculated according to the model, and compared with the statistical control limits in the model to determine whether there is a fault. The present invention mines the predictable characteristics in the process data information of different operating stages, accurately extracts the main predictable components that are helpful for fault diagnosis, objectively reveals the feature diversity of each stage, improves the detection rate of fault diagnosis, and effectively improves the process monitoring effect. BRIEF DESCRIPTION OF THE DRAWINGS
[0056] Figure 1 It is a flow chart of the intermittent process fault diagnosis method based on M-MForeCA of the present invention;
[0057] Figure 2 A schematic diagram of a process flow of a preprocessing method for a three-dimensional training data set of the present invention;
[0058] Figure 3 It is a diagram showing the stage division results of the penicillin fermentation process of the present invention;
[0059] Figure 4 The fault diagnosis results of M-MForeCA for batch 1 of faults are shown in Figure 2. Fore 2 Statistical graphs;
[0060] Figure 5 The fault diagnosis results of M-MForeCA for batch 1 of faults are shown in Figure 2. SPE Statistical graphs;
[0061] Figure 6 The fault diagnosis results of MForeCA for batch 1 are shown in Figure 2. Fore 2 Statistical graphs;
[0062] Figure 7 The fault diagnosis results of MForeCA for batch 1 are shown in Figure 2. SPE Statistical graphs;
[0063] Figure 8 The fault diagnosis results of M-MForeCA of the present invention for the fault 2 batches are shown in FIG. Fore 2 Statistical graphs;
[0064] Fig. 9 The fault diagnosis results of M-MForeCA of the present invention for the fault 2 batches are shown in FIG. SPE Statistical graphs;
[0065] Fig.10 The fault diagnosis results of MForeCA for fault batch 2 are shown in Figure 2. Fore 2 Statistical graphs;
[0066] Fig.11 The fault diagnosis results of MForeCA for fault batch 2 are shown in Figure 2. SPE Statistical graphs;
[0067] Fig.12 The fault diagnosis results of M-MForeCA for three batches of faults are shown in Figure 2. Fore 2 Statistical graphs;
[0068] Fig.13 The fault diagnosis results of M-MForeCA for three batches of faults are shown in Figure 2. SPE Statistical graphs;
[0069] Fig.14 The fault diagnosis results of MForeCA for batch 3 faults are shown in Figure 2. Fore 2 Statistical graphs;
[0070] Fig.15 The fault diagnosis results of MForeCA for batch 3 faults are shown in Figure 2. SPE Statistical chart. DETAILED DESCRIPTION
[0071] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.
[0072] Embodiment 1, a batch process fault diagnosis method based on M-MForeCA mentioned in the present invention, specifically a batch process fault diagnosis method based on multi-phase multi-way predictable component analysis, Multiphase MultiwayForecastable Component Analysis, referred to as: M-MForeCA;
[0073] See also Figure 1 , including the following steps:
[0074] 1. Collect normal batch historical data of intermittent process , I is the batch number, J is the number of variables, K is the number of sampling times; Preprocessing: First, Get a two-dimensional matrix based on batch expansion , and then normalize each column to a dimensionless form, and then restore it to a three-dimensional form and regard it as a function of K Sampled 2D data slices composition;
[0075] 2. Calculate the similarity factor at each moment SI , used for batch processes K Subsampled data slices Conduct reproducible analysis;
[0076] SI The definition formula is:
[0077] (1)
[0078] In the formula, for No. i According to the definition of formula (1), SI The value range is [0, 1);
[0079] Utilization SI Divide the intermittent process into a stable phase and a transition process;
[0080] The steps of phase division are as follows:
[0081] I. Calculate the value of each sampling moment according to formula (1) SI Values and plot SI Graphs;
[0082] II. Determine the approximate starting time of each stable phase and the rough end moment , when the intermittent process tends to a new stable stage, SI The curve will drop significantly, and the time corresponding to the lowest point of the drop process is , if the intermittent process is initially a stable phase, then =1, when a stable phase tends to end, SI The curve will rise continuously or suddenly from a lowest point. The time corresponding to the lowest point of the rising process is taken as , if the intermittent process ends with a stable phase, then The value is the maximum sampling time K ;
[0083] III. Using the kernel density estimation method to estimate the stable stages in the initial division SI Determine each stage based on the value SI Threshold value;
[0084] IV. Determine the exact start time of each stable sub-phase and the exact end time ,exist SI On the curve graph, each threshold line has an intersection with the corresponding descending curve before and ascending curve after the stable stage. The first sampling time after the intersection with the descending curve is defined as the accurate starting time. The last sampling time before the intersection with the rising curve is the accurate end time , if the intermittent process is initially in a stable stage, then =1, if the intermittent process ends in a stable stage, then = K ;
[0085] V. Determine the remaining portion after removing the stable phase as a transition process;
[0086] 4. Assume that the batch process is separated N The multi-directional predictable component analysis (MForeCA) algorithm was used to establish N Fault diagnosis model for each stable stage;
[0087] The steps of establishing the fault diagnosis model using the MForeCA algorithm in the stable stage are:
[0088] I. Assume that any n ( n =1,2,..., N ) three-dimensional data sets of stable sub-stages ,in K n is the number of samples in this stage, see Figure 2 3D data expansion processing method, the sub-stage 3D data Expand into Normalize by column, then Rearrange into , and transpose to get ;
[0089] II. After the expansion The predictability analysis was performed through ForeCA transformation to obtain the predictable component matrix Y , the transformation formula is:
[0090] (2)
[0091] In the formula is the load transformation matrix; determine W Here’s how:
[0092] i) Assumptions X The mean is , the covariance matrix is , the autocovariance function is , obtained by Fourier transform X The spectral function is:
[0093] (3)
[0094] In the formula i Represents an imaginary unit, k is an integer, is the angular frequency;
[0095] ii) Order , w yes W A column vector of , then the spectral function , y The spectral density function is defined as:
[0096] (4)
[0097] In the formula yes y The variance of
[0098] iii) Using information entropy, uncertainty is defined as:
[0099] (5)
[0100] In the formula, a >0 is the logarithmic base; for a white noise signal with a completely flat spectral density , cannot be predicted, so there is ,and ;
[0101] iv) Utilization y The uncertainty defines the predictability:
[0102] (6)
[0103] v) To achieve the extraction of predictable components, according to , the ForeCA method needs to find the optimal load vector w , thus establishing the following optimization problem:
[0104] (7)
[0105] Solve the above optimization problem and get the load vector matrix , the predictable component matrix is calculated by formula (2) ,and The corresponding predictability Arrange in order from high to low;
[0106] III. Use the cumulative contribution rate of predictability to determine the number of main predictable components, as shown in formula (8):
[0107] (8)
[0108] Usually order Sure d value ,Pick ,according to Calculate the principal predictable component matrix ;
[0109] IV. Construction Fore 2 Statistics and SPE Statistics, as shown in formulas (9) and (10):
[0110] (9)
[0111] (10)
[0112] In the above two formulas, , Respectively represent matrices Y d , No. i Column vector, , yes , e ( i )yes Column vector of ;
[0113] V. Calculation using kernel density estimation Fore 2 and SPE Control limits for the statistic.
[0114] 5. In view of the characteristics of the transition process, which has a short period, large fluctuations, and strong correlation between the dynamic time series before and after, the dynamic MForeCA method based on sliding window data is used to establish the fault diagnosis model of each transition process;
[0115] The steps of the transition process dynamic MForeCA modeling method are as follows:
[0116] I. Assume that a certain moment of the transition process is in the first k Sampling time, take the adjacent h 2D data slices constitute a 3D Total 2 h +1 continuous two-dimensional data slice is used as the training sample for establishing the fault diagnosis model at this moment in the transition process;k The training data window keeps sliding backwards; if the intermittent process starts or ends with a transition process, the number of adjacent samples in front or behind the training data set of the sliding window is less than h , there is a sliding window length < 2 h +1 situation;
[0117] II. Training samples at this moment of the transition process Transform to two dimensions Then ForeCA analysis is performed to obtain the W d and Y d , calculated by formula (9) and formula (10) respectively Fore 2 and SPE Statistics, and use the kernel density estimation method to obtain the control limits of the two statistics at that moment;
[0118] 6. Collect a new batch of k The data at the sampling time is standardized according to the mean and standard deviation of the corresponding time when modeling. ;
[0119] VII. Confirmation k The stable sub-stage or transition process belongs to the corresponding model. W d calculate and ;
[0120] 8. Calculation and Statistics are compared with the control limits in the corresponding model. Once a statistic crosses the control limit, it is considered that a fault has occurred.
[0121] In the above method, steps 2 to 3 are stage division, steps 4 to 5 are offline modeling, and steps 6 to 8 are online fault diagnosis.
[0122] In addition, for example, penicillin fermentation is a typical intermittent production process with multi-operation stage characteristics. PenSim V2.0 is a penicillin production process simulation platform developed by Birol et al. of Illinois Institute of Technology. By changing the initial conditions and control parameters, different fermentation results can be simulated, and normal and faulty batches can be simulated. It is widely used in the verification of algorithms such as intermittent process monitoring, fault detection and diagnosis. The present invention uses PenSim software to generate 30 batches of penicillin fermentation processes, each batch reaction time is 200h, and 11 variables (as shown in Table 1) are sampled once every hour, and the historical training data set is obtained as follows: .
[0123] Table 1 Variable description
[0124]
[0125] In the experiment, three fault batch data generated by PenSim are used, as shown in Table 2, to verify the fault diagnosis effect of the algorithm.
[0126] Table 2 Fault description
[0127]
[0128] The multiphase-MForeCA (M-MForeCA) method proposed in this paper requires three-dimensional data Perform preprocessing, such as Figure 2 As shown, first expand it into two dimensions The data is then normalized by column, so that the mean of each column is 0 and the standard deviation is 1. The normalized data is then restored to a three-dimensional form and is considered to be composed of 200 The similarity factor is calculated by composing two-dimensional sampled data slices of different sizes. SI The penicillin fermentation process was divided into stages. Figure 3 As shown, it is obvious that the fermentation process has two stable stages. According to the above method, the rough starting time of each stable stage is first located. and the rough end moment , and then calculate the threshold line of each stage, and finally determine the exact starting time of each stage and end time , and the rest is the transition process. Figure 3 As a result, the various stages of penicillin production were identified in the experiment in the following time sequence: 0-7h is transition process 1, 8h-41h is stable stage 1, 42h-52h is transition process 2, and 43h-200h is stable stage 2, which is consistent with the physiological stages of penicillin production: bacterial adaptation transition, mycelial exponential growth stage, mycelial growth to penicillin synthesis transition, penicillin synthesis stage. The MForeCA projection algorithm was implemented in stable stage 1 and stable stage 2 respectively to obtain their respective fault diagnosis models, that is, their respective projection matrices W d and Fore 2 Statistical control limits, SPE Statistical control limits; In transition process 1 and transition process 2, dynamically changing fault diagnosis models are established based on sliding window data, that is, the projection matrix and statistical control limits change with time. Finally, for the three fault batch data in Table 2, the fault diagnosis results based on the M-MForeCA method of the present invention are as follows: Figure 4-Figure 5 , Figure 8-Figure 9, Figure 12-13 As shown in the above figure, in the transition period 0-7h and 42h-52h, Fore 2 and SPE The statistical control limit is a changing curve, while in the stable stage 8h-41h and 43h-200h, the statistical control limit is a fixed straight line.
[0129] In addition, the MForeCA method is used to compare the fault diagnosis performance with the M-MForeCA method of the present invention. After pretreatment, no stage division is performed. A unified monitoring model is established in the batch reaction period 0-200h to obtain a single W d and Fore 2 Control limits, SPE Control limits, the monitoring results of the new faulty batch are as follows Figure 6-Figure 7 , Figure 10-11 , Figure 14-15 As shown, Fore 2 and SPE The control limits of a statistical chart are all straight lines.
[0130] In order to directly compare the fault diagnosis effects of the two methods, Table 3 summarizes the M-MForeCA method and the MForeCA method. Fore 2 Statistical charts and SPE The statistical graph shows the start time of the slowly changing slope fault. From Table 3 and the above result graph, it can be clearly seen that the M-MForeCA method of the present invention is superior to MForeCA. Fore 2 The statistical chart can detect these three types of faults earlier. SPE The detection of the first two faults is also better than that of the MForeCA method. In summary, the fault diagnosis method based on M-MForeCA can detect the occurrence of ramp faults more quickly and improve the fault detection rate.
[0131] Table 3 The start time of slope fault detection by M-MForeCA and MForeCA methods
[0132]
[0133] Note: In the table Indicates that a fault cannot be detected.
[0134] The above are only some preferred embodiments of the present invention. Any person skilled in the art may modify the above technical solutions or modify them into equivalent technical solutions. Therefore, the corresponding simple modifications or equivalent transformations made according to the technical solutions of the present invention shall fall within the scope of protection claimed by the present invention.
Claims
1. A fault diagnosis method for intermittent processes based on M-MForeCA, Its characteristics are The following steps are involved:
1. Collect normal batch historical data of intermittent process X (I×J×K), I is the batch number, J is the number of variables, and K is the number of sampling times; X (I×J×K) is preprocessed: First, X (I×J×K) is expanded based on the batch to obtain a two-dimensional matrix X(I×KJ), and each column is dimensionlessly normalized, and then restored to a three-dimensional form and regarded as a matrix consisting of K sampled two-dimensional data slices {X k ∈R I×J , k = 1, 2, ..., K}; Second, calculate the similarity factor SI at each moment, which is used for the K-times sampling data slice {X k ∈R I×J , k = 1, 2, ..., K} for repeatability analysis; 3. Use SI to divide the intermittent process into stable phase and transition process; Fourth, assuming that the intermittent process is separated into N stable stages, the multi-directional predictable component analysis (MForeCA) algorithm is used to establish the fault diagnosis model of each of the N stable stages; In step 4, the steps of establishing the fault diagnosis model using the MForeCA algorithm in the stable stage are: I. Assume that any nth (n=1, 2, ..., N) stable sub-stage 3D data set X (I×J×K n ), where K n is the sampling number of this stage, and the three-dimensional data of the sub-stage X (I×J×K n ) is expanded into X(I×K n J) is standardized by column, and then X(I×K n J) is rearranged into X(K n I×J), and transpose to get X(J×K n I); II. For the expanded X(J×K n I) The predictability analysis is performed through ForeCA transformation to obtain the predictable component matrix Y, and the transformation formula is: Y=W T X (2) Where W = [w 1 , w 2 , ..., w J ]∈R J×J is the load transformation matrix; III. Use the cumulative contribution rate of predictability to determine the number of main predictable components, as shown in formula (8): Usually, p(Ω)≥80% is used to determine the d value (d<J), and W is taken d =[w 1 , w 2 , ..., w d ], according to Y d =W d T X calculates the main predictable component matrix IV. Constructing Fore 2 Statistics and SPE statistics, as shown in formulas (9) and (10): Fore 2 (i)=y d (i) T y d (i)=x(i) T W d W d T x(i) (9) In the above two equations, y d (i) and x(i) represent the matrix Y d 、X(J×K n I) the i-th column vector, i = 1, 2, ..., K n , yes , e(i) is Column vector of ; V. Calculate Fore using kernel density estimation 2 and control limits of the SPE statistic; 5. In view of the characteristics of the transition process, which has a short period, large fluctuations, and strong correlation between the dynamic time series before and after, the dynamic MForeCA method based on sliding window data is used to establish the fault diagnosis model of each transition process; 6. Collect the k-th sampling time data of a new batch and standardize it according to the mean and standard deviation of the corresponding time when modeling to obtain x new (k); 7. Determine the stable sub-stage or transition process to which k belongs, and use W in the corresponding model d Calculate y new (k) = W d T x new (k) and e new (k) = x new (k)-W d W d T x new (k); 8. Calculation and SPE new (k) = e new (k) T e new (k) Statistics are compared with the control limits in the corresponding models. Once a statistic crosses the control limit, it is considered that a fault has occurred.
2. The intermittent process fault diagnosis method based on M-MForeCA according to claim 1, Its characteristics are: In step 2, the definition formula of SI is: In the formula, For X k In the i-th row of (I×J), according to the definition of formula (1), the SI value range is [0,1).
3. The intermittent process fault diagnosis method based on M-MForeCA according to claim 2, Its characteristics are: In step 3, the stages are divided into: I. Calculate the SI value at each sampling moment according to formula (1) and draw an SI curve graph; II. Determine the rough starting time k of each stable stage rough_start and the rough end time k rough_end When the intermittent process tends to a new stable stage, the SI curve will drop significantly, and the time corresponding to the lowest point of the drop process is taken as k rough_start , if the intermittent process is initially a stable phase, then k rough_start =1, when a stable stage tends to end, the SI curve will rise continuously or suddenly from a lowest point, and the time corresponding to the lowest point of the rising process is taken as k rough_end , if the intermittent process ends in a stable stage, then k rough_end The value is the maximum sampling time K; III. Determine the SI threshold of each stage based on the SI value of each stable stage preliminarily divided by using the kernel density estimation method; IV. Determine the exact starting time k of each stable sub-phase exact_start and the exact end time k exact_end On the SI curve, each threshold line has an intersection with the corresponding descending curve before and after the stable stage. The first sampling time after the intersection with the descending curve is defined as the accurate starting time k exact_start , the last sampling time before the intersection with the rising curve is the accurate end time k exact_enf , if the intermittent process is initially in a stable stage, then k exact_start =1, if the intermittent process ends in a stable stage, then k exact_end =K; V. The remaining part after removing the stable phase is identified as the transition process.
4. The intermittent process fault diagnosis method based on M-MForeCA according to claim 3, Its characteristics are: In step II, the method for determining W is as follows: i) Assume that the mean of X is μ and the covariance matrix is Σ X , the autocovariance function is Γ X (k), the X spectrum function is obtained by Fourier transform: Where i represents the imaginary unit, k is an integer, and λ is the angular frequency; ii) Let y = w T X, w is a column vector of W, then the spectral function S y (λ)=w T S X The spectral density function of (λ)w,y is defined as: In the formula is the variance of y; iii) Using information entropy, uncertainty is defined as: Where a>0 is the logarithmic base; for a white noise signal ε with a completely flat spectral density, it cannot be predicted, so there exists Hs ,a( y)≤H s,a (ε), and iv) Define predictability using the uncertainty of y: v) To achieve the extraction of predictable components, according to y=w T X, ForeCA method needs to find the optimal load vector w, thus establishing the following optimization problem: Solving the above optimization problem, we get the load vector matrix W = [w 1 , w 2 , ..., w J ], and the predictable component matrix Y = [y 1 ,y 2 , ..., y J ] T , and y 1 ,y 2 , ..., y J The corresponding predictability Ω is arranged in order from high to low.
5. The intermittent process fault diagnosis method based on M-MForeCA according to claim 4, Its characteristics are: In step 5, the steps of the transition process dynamic MForeCA modeling method are as follows: I. Assume that a certain moment of the transition process is at the kth sampling moment of the entire intermittent process, and take the h adjacent two-dimensional data slices before and after it to form a three-dimensional X k (I×J×(2h+1))={X k-h , ..., X k-1 , X k , X k+1 , ..., X k+h A total of 2h+1 continuous two-dimensional data slices are used as training samples for establishing the fault diagnosis model at this moment of the transition process; as k increases, the training data window continues to slide backward; if the intermittent process is the transition process at the beginning or the end, the number of adjacent samples in front or behind taken by some sliding window training data sets is less than h, and there is a situation where the sliding window length is less than 2h+1; II. Training samples at this moment in the transition process X k (I×J×(2h+1)) is transformed into two-dimensional X k (J×(2h+1)I) form, and then perform ForeCA analysis to obtain W at that moment d and Y d , Fore is calculated by formula (9) and formula (10) respectively 2 and SPE statistics, and use the kernel density estimation method to obtain the control limits of the two statistics at that moment.
