Fault detection method and device based on collaborative neighbor feature analysis

Abstract

The invention discloses a fault detection method and device based on collaborative neighbor feature analysis, and aims to comprehensively and collaboratively analyze neighbor features of process data so as to implement fault detection by using anomaly of collaborative change of the neighbor features of the data. According to the fault detection method disclosed by the invention, collaborative neighbor feature analysis is carried out on a training data set to obtain a conversion matrix for calculating fault indexes; on the basis, corresponding improved technical schemes are disclosed for a collaborative neighbor feature analysis algorithm, process data processing and parameter determination, so that the detection sensitivity of the fault detection method is improved. Besides, the invention also discloses a corresponding fault detection device based on the same technical concept, and the device comprises a model training module, a data acquisition module, a data processing module, a detection display module and a fault alarm module. Wherein the model training module can implement collaborative neighbor feature analysis, and the data processing module can further implement self-correlation standardization processing.

Description

Technical Field

[0001] The present invention relates to the technical field of data-driven fault detection, and in particular to a fault detection method and device based on collaborative nearest neighbor feature analysis. Background Art

[0002] The continuous and stable operation of the production process in the expected working state is the focus of modern industrial production and management. The real-time monitoring of the production process status is inseparable from the fault detection during the operation process. With the widespread application of a large number of sensors and computer-aided production systems, the production process can collect, analyze and process a large amount of industrial process data in real time, which lays a solid data foundation for the implementation of data-driven industrial process monitoring. Compared with the traditional method of using mechanism models or expert experience to detect and diagnose faults in the production process, data-driven fault detection starts from the data and detects abnormal change characteristics in the data in real time according to the sampling time interval, thereby reflecting the faults in the production process. Its technical implementation is convenient and fast, and it is widely favored by technicians for its advantages in applicability and versatility.

[0003] A paper titled "Evaluation of the operating status of complex industrial processes based on collaborative perception of static-dynamic characteristics" published in the Journal of Automatica Sinica in 2023 disclosed an analysis algorithm for the changes in static and dynamic characteristics in collaborative perception process data, and on this basis established a process operating status evaluation system based on collaborative perception of static-dynamic characteristics, effectively realizing real-time detection of faults in industrial processes. The Chinese invention patent with application number / patent number 202311429386.3 discloses a chemical process fault detection method based on deep semi-non-negative matrix decomposition. This method decomposes the sampled data according to feature changes and implements chemical process fault detection from the perspective of fusion of multiple sub-block monitoring results. The Chinese invention patent with application number / patent number 202310866199.5 discloses an industrial boiler fault detection method based on improved ensemble learning. This method analyzes the differences between sampled data through multiple classification models, and also integrates the advantages of multiple models to solve the problem of industrial boiler operation fault detection.

[0004] It should be pointed out that the above three technical solutions all involve the use of the coordination or fusion of multi-faceted change characteristics of data to detect sampled data under fault conditions, which fully demonstrates the advantages of collaborative features in mining data change feature analysis. This is mainly due to the fact that collaborative feature analysis analyzes the change characteristics of data from multiple aspects, and more comprehensively represents the change characteristics of normal data. However, the sampling data of the industrial production process is not generated according to the expected normal distribution assumption. Different batches of the production process or different production environments will cause the corresponding sampling data to present an irregular and unclear multi-cluster distribution state. Considering that a comprehensive analysis of the change characteristics of normal working condition sampling data is the key to ensuring the sensitivity of subsequent fault detection, implementing feature analysis from the perspective of data distribution neighbors is a feasible solution. The Chinese invention patent with application number / patent number 202111639643.7 has disclosed a multi-condition process fault detection method and system based on local neighborhood standardization, which implements standardized processing through neighbor relationships to deal with the multi-cluster distribution problem of multi-condition sampling data.

[0005] However, data sampling has temporal continuity, and the distribution of data in space has both neighbor and angle relationships. The determination of the neighbor relationship between sampled data should not be limited to spatial distance. Data with similar sampling time also show similar change characteristics. In addition, from a geometric point of view, the angle between two data vectors also shows the close relationship between the two. Therefore, implementing fault detection from the perspective of abnormal changes in data neighbor features requires comprehensive collaborative analysis and processing of neighbor features, so as to provide detection sensitivity for data of different fault types. Summary of the invention

[0006] The main technical problem to be solved by the present invention is: how to comprehensively and collaboratively analyze the neighboring features of process data, so as to implement fault detection by utilizing the abnormalities of the collaborative changes of the neighboring features of the data.

[0007] The technical solution adopted by the present invention to solve the above-mentioned problem is: a fault detection method and device based on collaborative neighbor feature analysis; wherein, the implementation process of a fault detection method based on collaborative neighbor feature analysis includes steps 1 to 3 shown below.

[0008] Step 1: Obtain process data of n consecutive sampling moments under normal working conditions from the historical database of the production process, and perform standardization on them to form a training data set.

[0009] Step 2: Perform collaborative neighbor feature analysis on the training data set to obtain the intermediate matrix M used to calculate the fault index.

[0010] Step 3: Obtain the process data at the latest sampling time in real time, perform standardization on it, calculate the fault index using the intermediate matrix M, and then determine whether a fault occurs based on whether the fault index exceeds the limit.

[0011] The collaborative neighbor feature analysis in the above step 2 is the core of the fault detection method, and the corresponding implementation process includes steps 2-1 to 2-3 shown below.

[0012] Step 2-1, after recording the training data set as a matrix X, set the distance neighbor order equal to K, the time series neighbor order equal to D, and the angle neighbor threshold equal to θ; wherein each column vector in X corresponds to the process data at each sampling time, K is an integer not greater than 10 and not less than 5, D is an integer not greater than 10 and not less than 2, and θ∈[0.95,0.99];

[0013] Step 2-2: Determine the distance neighbor matrix C1∈R according to formulas ① to ③ as shown below: n×n , the temporal neighbor matrix C2∈R n×n and the angle neighbor matrix C3∈R n×n :

[0014]

[0015]

[0016]

[0017] Step 2-3, where R n×n represents an n×n dimensional real matrix, x i With x j Respectively represent the i-th column vector and the j-th column vector in X, N(x i ) indicates that X and x i The set of the K nearest column vectors. The superscript T represents the transpose symbol of the matrix or vector. i || means calculating x i The length of C1(i,j) represents the element in the i-th row and j-th column of C1. |ij| means calculating the absolute value of ij. The numbers i∈{1,2,…,n} and j∈{1,2,…,n} are numbered.

[0018] Step 2-3, according to formula W α =C α U α After calculating the three weight matrices W1, W2, and W3 respectively, follow the formula and L=XX T Calculate G and L respectively; among them, C α represents any one of C1, C2, C3, the diagonal matrix Uα The elements on the diagonal U α (i,i)=1 / ∑ j C α (i,j),I n Represents the n×n dimensional unit matrix, numbered α=1,2,3.

[0019] Step 2-4: Solve the generalized eigenvalue problem Gp = λLp and find the eigenvectors p1, p2, ..., p corresponding to the smallest d eigenvalues d ; where λ and p represent the eigenvalue and its corresponding eigenvector respectively.

[0020] Step 2-5: According to the formula M = PΛ -1 P T / ε1+(I m -PB)(I m -PB) T / ε2 calculates the intermediate matrix M; wherein ε1 and ε2 represent two weight coefficients, and the matrix P = [p1, p2, ..., p d ], the matrix Λ=P T XX T P,I m Represents the m×m-dimensional identity matrix, matrix B = Λ -1 P T XX T .

[0021] The above steps 2-1 to 2-5 are the implementation process of the collaborative neighbor feature analysis algorithm involved in the present invention; the collaborative neighbor feature analysis is a new data feature analysis algorithm, which aims to minimize the distance between process data and its distance neighbor center, time sequence neighbor center, and angle neighbor center as much as possible through the transformation of a feature vector p, which can be quantified into a minimization problem through the data formula, namely: min p T (X-XW α )(X-XW α ) T p; wherein, numbers α=1, 2, 3 correspond to distance neighbors, time sequence neighbors, and angle neighbors, respectively.

[0022] It can be seen that the key to the minimization problem lies in how to determine the weight matrices W1, W2, and W3. First, the K column vectors with the smallest distance are determined. For this purpose, 0 or 1 is used in formula ① to indicate whether there is a neighbor relationship between the column vectors in X; secondly, the neighbor relationship in time continuity is determined, which is achieved through formula ②; finally, the angle neighbor relationship is determined by the cosine value of the angle between the column vectors in X, that is, formula ③.

[0023] Considering that the so-called nearest neighbor center is the mean of the K column vectors with the smallest distance, the elements in W1 are determined using the diagonal matrix U α For the distance neighbor matrix C α ∈R n×n Processing to ensure XW α What is calculated is the nearest center corresponding to each column vector in X.

[0024] Therefore, the optimization goal of the collaborative neighbor feature analysis algorithm involved in the present invention can be defined as follows:

[0025]

[0026] Among them, the constraint condition p T XX T p=1 is intended to constrain the variance change of X projected onto p; the constrained minimization problem can be easily solved by the Lagrange multiplier method, thereby converting it into a generalized eigenvalue problem: Gp=λLp, which is the generalized eigenvalue problem in steps 2-4.

[0027] After getting the matrix P = [p1, p2, ..., p d ], two fault indicators are usually used to detect faults in the process data at the latest sampling time; if the standardized process data is recorded as a column vector x, then the two fault indicators are and ψ = x T (I m -PB)(I m -PB) T x, the former detects the change of collaborative neighbor feature components, and the latter detects the change of model error.

[0028] As you can imagine, using two fault indicators is not conducive to judging whether a fault has occurred or not, because if only one of the fault indicators exceeds the limit, how can the fault be judged? To this end, the present invention performs weighted fusion of the two fault indicators in step 2-5, that is, through two weight coefficients ε1 and ε2, The weighted sum of ψ and ψ is used to obtain a corresponding fault index Since M is in the middle position, Therefore, in step 2, an intermediate matrix M is involved for calculating the fault indicator.

[0029] From the design reasoning process of the above algorithm, it can be seen that the collaborative neighbor feature analysis algorithm directly adds the weight matrices of distance neighbors, time sequence neighbors, and angle neighbors, that is: Although the three neighbor relationships are treated equally, the differences between the three neighbor relationships cannot be fully distinguished. To this end, the present invention discloses an improved solution, which replaces the formulas for calculating G and L in step 2-3 with and Among them, the weight coefficient φ α The condition φ1+φ2+φ3=1 is satisfied and φ1, φ2, φ3 are all greater than 0, the matrix Diagonal Matrix Elements on the diagonal

[0030] The improved scheme measures the differences of the three nearest neighbor relationships by differentiating the weighted coefficients, which can ensure a more comprehensive collaborative analysis of the nearest neighbor features of the process data; in addition, the calculation of the matrix L is to merge the change features between the data that do not have nearest neighbor relationships by weighted coefficients. Therefore, the improved scheme not only considers the collaborative nearest neighbor features, but also the non-nearest neighbor change features of the system, and its feature extraction is more comprehensive.

[0031] However, although the selection of the three weight coefficients gives corresponding conditions, it still depends on the experience of the technicians to determine. To this end, the present invention provides a further improvement scheme to specifically solve the problem of determining the weight coefficients, namely: first calculate the symmetric matrix H α =(I n -W α )(I n -W α ) T The maximum eigenvalue η α , and then through φ α =1-η α / (η1+η2+η3) determines φ1, φ2, φ3; this scheme does not rely on experience, but only objectively determines the corresponding weight coefficients from the perspective of data.

[0032] In the technical field, the maximum eigenvalue of a matrix can also be called the spectral radius of the matrix, which reflects the possible change capability contained in the matrix; taking the distance neighbor as an example, if the spectral radius of the corresponding Hankel matrix H1 is relatively small, more attention needs to be paid to the collaborative neighbor feature analysis to avoid ignoring the change of the distance neighbor feature. This is why the improved scheme uses 1-η1 / (η1+η2+η3) to determine the corresponding weights.

[0033] As another improvement of the present invention, after the process data at any sampling moment is standardized, it is necessary to further perform local neighborhood standardization, including steps A-1 to A-3 shown below.

[0034] Step A-1: ​​record the process data at any sampling time after the standardization process as a column vector x, select k column vectors with the smallest distance from x from the matrix X, and record them as x (1) ,x (2) ,…,x (k) .

[0035] Step A-2: Calculate k column vectors x (1) ,x (2) ,…,x (k) The mean μ and standard deviation δ.

[0036] Step A-3: According to the formula Perform local neighborhood normalization on x, and get the corresponding column vector in, Represents division of elements at the same position.

[0037] In this technical field, the purpose of standardization is to eliminate the size differences in the dimensional changes of each measured variable, so that the range of variation of each measured variable is a mean of 0 and a standard deviation of 1; taking into account the complex multi-cluster distribution characteristics that may exist in industrial process data, the local neighborhood standardization processing technology uses the mean and standard deviation of the k column vectors with the smallest distance to further process the corresponding column vectors, which can further eliminate the negative interference of the complex multi-cluster distribution problem of the data on the subsequent collaborative neighbor feature analysis.

[0038] Directly using the average value of k column vectors in the local area for local neighborhood normalization actually treats each column vector in the local neighborhood equally; however, there are differences in distance between these column vectors in the local area and x. In order to further take the distance difference into account, the present invention provides an improved solution. The improved solution uses the following formula ④ for the calculation of the average value and standard deviation of step A-2:

[0039]

[0040] Among them, x (c) Represents x (1) ,x (2) ,…,x (k) The cth column vector in, weight coefficient β c =exp(-||xx (c) || 2 / m), numbered c=1,2,…,k.

[0041] The calculation of the mean and standard deviation in formula ④ uses the weighted mean and weighted standard deviation, and the determination of the corresponding weighting coefficient involves the cth column vector x closest to x. (c)The square of the distance between them, that is, ||xx (c) || 2 Therefore, the improved scheme involving the local neighborhood normalization processing technology can further take the distance differences in the local area into consideration.

[0042] To further process the process data using the local neighborhood standardization technique, it is necessary to calculate the distance between x and each column vector in X, and then determine the k column vectors with the smallest distance from x. The subsequent mean and standard deviation only use the k column vectors with the smallest distance, and do not involve the remaining column vectors in X. It can be found that the selection of parameter k is very critical to the local neighborhood standardization technique; if k is too large, the local area standardization technique will not play a role in eliminating the negative impact of the complex multi-cluster distribution problem of the data; vice versa.

[0043] As another improvement of the present invention, after the process data at any sampling moment is subjected to standardization processing, it is necessary to further perform self-correlation standardization processing, including steps B-1 to B-4 shown below.

[0044] Step B-1: record the process data at any sampling time after the standardization process as a column vector x, and calculate the distance between each column vector in the matrix X and x, which are recorded as g1, g2, ..., g n .

[0045] Step B-2: According to the formula Calculate the weighting coefficients γ1,γ2,…,γ n ; Among them, g i With g j Respectively represent the distance between the i-th column vector and the j-th column vector in the matrix X and x.

[0046] Step B-3, calculate the mean value μ and standard deviation δ according to the following formula:

[0047]

[0048] Step B-4: According to the formula Implement autocorrelation normalization on x, and get the corresponding column vector in, Represents division of elements at the same position.

[0049] The above improved scheme also uses the weighted mean and standard deviation, but the difference is that the distance between all column vectors in X and x is used to determine the corresponding weight coefficient; this improved scheme avoids the selection of parameter k, and uses the exponential function exp() with natural constants as the base to give smaller weight values ​​to column vectors with larger distances, which can also play a role in eliminating the negative impact of complex multi-cluster distribution problems of data. Therefore, this improved scheme is more applicable and universal.

[0050] As another improvement of the present invention, the two weight coefficients ε1 and ε2 in step 2-5 are set according to steps C-1 to C-3 shown below.

[0051] Step C-1, according to the formula and Calculate n feature change indicators respectively and n error change indicators ψ1, ψ2, …, ψ n .

[0052] Step C-2: Calculation The average values ​​u1 and σ1 are used to calculate ψ1, ψ2, …, ψ n The average values ​​u2 and σ2.

[0053] Step C-3, set ε1=u1+3σ1 and ε2=u2+3σ2.

[0054] Generally speaking, technicians in this field can directly set ε1=ε2=1, that is, give the same weight to the feature change index and the error change index; however, this setting method has not taken into account the difference in the scale of change of these two indicators; and the above-mentioned improvement scheme objectively gives a method for determining the weight value, which takes the two indicators into consideration in a balanced manner, and does not rely on the technician's prior knowledge or engineering experience, and is more universal.

[0055] As another improvement of the present invention, when solving the generalized eigenvalue problem Gp=λLp in step 2-4, it is preferred to first convert it into an eigenvalue problem S -0.5 A T GAS -0.5 q=λq, and then solve the eigenvalue problem to find the eigenvalue vectors q1,q2,…,q corresponding to the smallest d eigenvalues. d , and finally through p τ =AS -0.5 q τ Calculate p1, p2, ..., p d ; Among them, the singular value decomposition of L can obtain the unitary matrix A and the diagonal matrix S composed of singular values, that is: L = A T SA. Since S is a diagonal matrix, all elements on the off-diagonal line are equal to 0, S 0.5It is to calculate the 0.5 power of the elements on the diagonal of S, S -0.5 That is, calculate the -0.5 power of the elements on the diagonal of S, numbered τ=1,2,…,d.

[0056] Compared with solving generalized eigenvalues, solving eigenvalue problems is more convenient and faster, and has higher computational efficiency. Therefore, the above-mentioned improved solution for solving generalized eigenvalue problems in steps 2-4 has higher model training efficiency.

[0057] Based on the same inventive concept, the present invention also discloses a fault detection device based on collaborative neighbor feature analysis, including a model training module, a data acquisition module, a data processing module, a detection and display module and a fault alarm module; wherein, the model training module stores a computer program and an intermediate matrix M for calculating the fault index; the data acquisition module acquires the process data of the latest sampling time in real time and sends it to the data processing module; the data processing module, after receiving the process data sent by the data acquisition module, executes and implements step 3 of the above-disclosed fault detection method, and then sends the fault index and the judgment result of whether the fault occurs to the detection and display module and the fault alarm module respectively; the detection and display module, after receiving the fault index sent by the data processing module, displays its specific value and the current sampling time in real time.

[0058] The core of the device is that when the computer program stored in the model training module is executed, steps 1 and 2-1 to 2-5 in the above-mentioned fault detection method and its improved solution are implemented, and then the obtained intermediate matrix M is stored.

[0059] In addition, the present invention also provides an improved fault detection device, the data processing module of which performs standardization processing on the process data at any sampling moment, and then further performs self-correlation standardization processing on it, executing and realizing the above steps B-1 to B-4. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 The present invention is an implementation flow chart of the fault detection method disclosed in the present invention.

[0061] Figure 2 It is a schematic diagram of a specific application object of the present invention.

[0062] Figure 3 This is a fault detection details diagram of the fault detection method disclosed in the present invention.

[0063] Figure 4 This is a fault detection details diagram of the improved fault detection method disclosed in the present invention.

[0064] Figure 5 The flowchart of the local neighborhood normalization process implemented in the present invention.

[0065] Figure 6 Flow chart of the self-correlation normalization process implemented in the present invention.

[0066] Figure 7 The present invention is a schematic diagram of the structure of the fault detection device disclosed in the present invention. DETAILED DESCRIPTION

[0067] The technical solution of the embodiment of the present invention is described in detail below with reference to the accompanying drawings.

[0068] like Figure 1 As shown, the present invention discloses a fault detection method based on collaborative neighbor feature analysis; the method can be used to Figure 2 The chemical production process shown in the figure is used for fault detection; the chemical production process includes five main production links, namely a reactor, a condenser, a separator, a compressor and a stripping tower, and can transform four raw materials marked as A, B, C, and D to obtain corresponding chemical products; the process data that can be collected at each sampling moment include four categories of 22 process data, namely temperature (TI), liquid level (LI), pressure (PI), and flow (FI); the fault detection method disclosed in the present invention is applied, including steps 1 to 3 shown below.

[0069] Step 1: Obtain process data of normal working conditions for n=960 consecutive sampling moments from the historical database of the production process, and perform standardization processing on them to form a training data set.

[0070] Step 2: Perform collaborative neighbor feature analysis on the training data set to obtain an intermediate matrix M for calculating fault indicators. The corresponding implementation process includes steps 2-1 to 2-5 shown below.

[0071] Step 2-1, after recording the training data set as matrix X, set the distance neighbor order equal to 10, the time sequence neighbor order equal to 3, and the angle neighbor threshold equal to 0.95.

[0072] Step 2-2: Determine the distance neighbor matrix C1∈R according to formulas ① to ③ n×n , the temporal neighbor matrix C2∈R n×n and the angle neighbor matrix C3∈R n×n .

[0073] Step 2-3, according to formula W α =C α U α After calculating the three weight matrices W1, W2, and W3 respectively, follow the formula and L=XX T Calculate G and L separately.

[0074] Step 2-4: Solve the generalized eigenvalue problem Gp = λLp and find the eigenvectors p1, p2, ..., p corresponding to the smallest d = 10 eigenvalues d ; where λ and p represent the eigenvalue and its corresponding eigenvector respectively.

[0075] Step 2-5: According to the formula M = PΛ -1 P T / ε1+(I m -PB)(I m -PB) T / ε2 calculates the intermediate matrix M; where the two weight coefficients ε1 and ε2 are both equal to 1, and the matrix Λ=P T XX T P, matrix B = Λ -1 P T XX T .

[0076] Step 3: Obtain the process data at the latest sampling time in real time, perform standardization on it, calculate the fault index using the intermediate matrix M, and then determine whether a fault occurs based on whether the fault index exceeds the limit.

[0077] Fault detection is performed on the process data of the 2000 latest sampling moments of the production process. The corresponding fault detection details are shown in the figure below. Figure 3 As shown in the figure, according to the feedback from the on-site technicians, the first 400 sampling moments were in normal working condition, and from the 1600 sampling moments onwards, a sensor had a measurement offset failure. Figure 3 The horizontal line in is the upper limit of the fault index. If the fault index exceeds the limit, it can be determined that a fault has occurred.

[0078] from Figure 3 It can be found that the fault detection method disclosed in the present invention has exceeded the limit of fault indicators at most sampling moments starting from the 401st sampling moment. This is because a fault did occur in the production process starting from the 401st moment. It can be seen that the fault detection method disclosed in the present invention can effectively detect faults.

[0079] The second embodiment of the present invention is an improvement on step 2-3, where the formulas for calculating G and L are replaced by and Among them, the weight coefficients φ1, φ2, φ3 that satisfy the condition φ1+φ2+φ3=1 are determined as follows: First calculate the symmetric matrix H α =(I n -W α )(I n -W α ) T The maximum eigenvalue η α, and then through φ α =η α / (η1+η2+η3), and the final weight coefficients are φ1=0.36, φ2=0.42, and φ3=0.22 respectively.

[0080] According to the second embodiment, fault detection is performed on the same 2000 latest sampling time process data, and the corresponding fault detection details are as follows: Figure 4 shown; through Figure 3 and Figure 4 By comparison, it can be found that after the failure occurs, Figure 4 The underreporting of the fault indicators shown in the figure is significantly reduced, which is mainly due to the fact that the improved scheme takes both the neighbor features and the non-neighbor features into consideration when constructing the matrices G and L.

[0081] Furthermore, after the fault detection method disclosed in the present invention performs standardization processing on the process data at any sampling time, it is necessary to further perform local neighborhood standardization processing or self-correlation standardization processing. The corresponding processing flows are as follows: Figure 5 and Figure 6 As shown; among them, Figure 5 The local neighborhood normalization process shown includes steps A-1 to A-3 shown below.

[0082] Step A-1: ​​record the process data at any sampling time after the standardization process as a column vector x, select k column vectors with the smallest distance from x from the matrix X, and record them as x (1) ,x (2) ,…,x (k) .

[0083] Step A-2: Calculate k column vectors x (1) ,x (2) ,…,x (k) The mean μ and standard deviation δ.

[0084] Step A-3: According to the formula Perform local neighborhood normalization on x, and obtain the corresponding column vector x; where Represents division of elements at the same position.

[0085] exist Figure 5 In the implementation process of the local neighborhood normalization process shown in FIG. 1 , the mean value μ and the standard deviation δ may also be calculated in a weighted manner, with specific reference to the aforementioned formula ④.

[0086] like Figure 6 As shown, the self-correlation normalization process is implemented, including steps B-1 to B-4 shown below.

[0087] Step B-1: record the process data at any sampling time after the standardization process as a column vector x, and calculate the distance between each column vector in the matrix X and x, which are recorded as g1, g2, ..., g n .

[0088] Step B-2: According to the formula Calculate the weighting coefficients γ1,γ2,…,γ n ; Among them, g i With g j Respectively represent the distance between the i-th column vector and the j-th column vector in the matrix X and x.

[0089] Step B-3, calculate the mean value μ and standard deviation δ according to the above formula ⑤.

[0090] Step B-4: According to the formula Implement autocorrelation normalization on x, and get the corresponding column vector

[0091] Another embodiment of the present invention may be to set the two weight coefficients ε1 and ε2 in step 2-5 in any of the aforementioned implementation schemes according to steps C-1 to C-3 shown below.

[0092] Step C-1, according to the formula and Calculate n feature change indicators separately and n error change indicators ψ1, ψ2, …, ψ n .

[0093] Step C-2: Calculation The average values ​​u1 and σ1 are used to calculate ψ1, ψ2, …, ψ n The average values ​​u2 and σ2.

[0094] Step C-3, set ε1=u1+3σ1 and ε2=u2+3σ2.

[0095] In addition, when solving the generalized eigenvalue problem Gp=λLp in step 2-4, it can be first converted into an eigenvalue problem S -0.5 A T GAS -0.5 q=λq, and then solve the eigenvalue problem to find the eigenvalue vectors q1,q2,…,q corresponding to the smallest d eigenvalues. d , and finally through p τ =AS -0.5 q τ Calculate p1, p2, ..., p d , which can speed up the computational efficiency of solving generalized eigenvalue problems.

[0096] like Figure 7 As shown, based on the same inventive concept, the present invention designs a fault detection device based on collaborative neighbor feature analysis, and its component modules and the functions of each module are introduced as follows.

[0097] The model training module stores a computer program for implementing step 1 and step 2-1 to step 2-5, is used to calculate the intermediate matrix M of the fault index, and can store the obtained intermediate matrix M.

[0098] The data acquisition module acquires the process data of the latest sampling moment in real time and sends it to the data processing module; after receiving the process data sent by the data acquisition module, the data processing module performs data processing according to step 3 of the fault detection method disclosed in the present invention, and then sends the fault indicator and the judgment result of whether the fault occurs to the detection display module and the fault alarm module respectively.

[0099] The detection and display module, after receiving the fault indicators sent by the data processing module, displays the specific values ​​and current sampling time in real time; the fault alarm module triggers the fault alarm according to the judgment result of whether the fault occurs or not. The function of these two modules is to output the fault detection results in different forms.

[0100] The fault detection device can be further improved by further performing self-correlation normalization processing on the process data at any sampling moment after the data processing module performs normalization processing on the process data.

[0101] The above-mentioned embodiments of the present invention are described in detail in conjunction with the accompanying drawings, and the corresponding embodiments can make the purpose, technical solutions and advantages of the present invention more clearly described. However, the specific embodiments of the present invention are only used to explain the present invention, and are not intended to limit the technical content of the present invention. All equivalent substitutions, obvious improvements, etc. made within the method ideas provided by the present invention should fall within the protection scope of the present invention.

Claims

1. A fault detection method based on collaborative neighbor feature analysis, comprising steps 1 to 3 as shown below: Step 1: Obtain process data of n consecutive sampling moments under normal working conditions from the historical database of the production process, and perform standardization on them respectively to form a training data set; Step 2: Perform collaborative neighbor feature analysis on the training data set to obtain an intermediate matrix M for calculating fault indicators; Step 3: Obtain the process data at the latest sampling time in real time, perform standardization on it, calculate the fault index using the intermediate matrix M, and then determine whether a fault occurs based on whether the fault index exceeds the limit; Features: The implementation process of step 2 includes steps 2-1 to 2-5 shown below: Step 2-1, after recording the training data set as a matrix X, set the distance neighbor order equal to K, the time series neighbor order equal to D, and the angle neighbor threshold equal to θ; wherein each column vector in X corresponds to the process data at each sampling time; K is an integer not greater than 10 and not less than 5, D is an integer not greater than 10 and not less than 2, and θ∈[0.95,0.99]; Step 2-2: Determine the distance neighbor matrix C1∈R according to formulas ① to ③ as shown below: n×n , the temporal neighbor matrix C2∈R n×n and the angle neighbor matrix C3∈R n×n : Among them, R n×n represents an n×n dimensional real matrix, x i With x j Respectively represent the i-th column vector and the j-th column vector in X, N(x i ) indicates that X and x i The set of the K nearest column vectors. The superscript T represents the transpose symbol of the matrix or vector. i || means calculating x i The length of C1(i,j) is the element in the i-th row and j-th column of C1. |ij| means calculating the absolute value of ij. The numbers i∈{1,2,…,n} and j∈{1,2,…,n} are numbered. Step 2-3, according to formula W α =C α U α After calculating the three weight matrices W1, W2, and W3 respectively, follow the formula and L=XX T Calculate G and L respectively; among them, C α represents any one of C1, C2, C3, the diagonal matrix U α The elements on the diagonal U α (i,i)=1 / Σ j C α (i,j),I n Represents the n×n-dimensional identity matrix, numbered α=1,2,3; Step 2-4: Solve the generalized eigenvalue problem Gp = λLp and find the eigenvectors p1, p2, ..., p corresponding to the smallest d eigenvalues d ; Among them, λ and p represent the eigenvalue and its corresponding eigenvector respectively; Step 2-5: According to the formula M = PΛ -1 P T / ε1+(I m -PB)(I m -PB) T / ε2 calculates the intermediate matrix M; wherein ε1 and ε2 represent two weight coefficients, and the matrix P = [p1, p2, ..., p d ], the matrix Λ=P T XX T P,I m represents the m×m-dimensional identity matrix, where m is equal to the total number of row vectors in X, and the matrix B = Λ -1 P T XX T .

2. A fault detection method according to claim 1, characterized in that: The formulas for calculating G and L in step 2-3 are replaced by and Among them, the weight coefficient φ α The condition φ1+φ2+φ3=1 is satisfied and φ1, φ2, φ3 are all greater than 0, the matrix Diagonal Matrix Elements on the diagonal 3. A fault detection method according to claim 2, characterized in that: The weight coefficient φ α The way to determine is: first calculate the symmetric matrix H α =(I n -W α )(I n -W α ) T The maximum eigenvalue η α , and then through φ α =η α / (eta1+eta2+eta3) determines φ1, φ2, φ3.

4. A fault detection method according to any one of claims 1 to 3, characterized in that: After the process data at any sampling time is standardized, it is necessary to further implement local neighborhood standardization, including steps A-1 to A-3 shown below: Step A-1: ​​record the process data at any sampling time after the standardization process as a column vector x, select k column vectors with the smallest distance from x from the matrix X, and record them as x (1) ,x (2) ,…,x (k) ; Step A-2: Calculate k column vectors x (1) ,x (2) ,…,x (k) The mean μ and standard deviation δ of ; Step A-3: According to the formula Perform local neighborhood normalization on x, and obtain the corresponding column vector x; in, Represents division of elements at the same position.

5. A fault detection method according to claim 4, characterized in that: The step A-2 is to calculate the mean value μ and the standard deviation δ according to the following formula: Among them, x (c) Represents x (1) ,x (2) ,…,x (k) The cth column vector in, weight coefficient β c =exp(-||xx (c) || 2 / m), numbered c=1,2,…,k.

6. A fault detection method according to any one of claims 1 to 3, characterized in that: After the process data at any sampling time is standardized, it is necessary to further implement self-correlation standardization, including steps B-1 to B-4 shown below: Step B-1: record the process data at any sampling time after the standardization process as a column vector x, and calculate the distance between each column vector in the matrix X and x, which are recorded as g1, g2, ..., g n ; Step B-2: According to the formula Calculate the weighting coefficients γ1,γ2,…,γ n ; Among them, g i With g j Respectively represent the distance between the i-th column vector and the j-th column vector in the matrix X and x; Step B-3, calculate the mean value μ and standard deviation δ according to the following formula: Step B-4: According to the formula Implement autocorrelation normalization on x, and get the corresponding column vector in, Represents division of elements at the same position.

7. A fault detection method according to any one of claims 1 to 3, characterized in that: The two weight coefficients ε1 and ε2 in step 2-5 are set according to steps C-1 to C-3 as shown below: Step C-1, according to the formula and Calculate n feature change indicators separately and n error change indicators ψ1, ψ2, …, ψ n ; Step C-2: Calculation The average values ​​u1 and σ1 are used to calculate ψ1, ψ2, …, ψ n The average values ​​u2 and σ2; Step C-3, set ε1=u1+3σ1 and ε2=u2+3σ2.

8. A fault detection method according to any one of claims 1 to 3, characterized in that: When solving the generalized eigenvalue problem Gp=λLp in step 2-4, it is first converted into an eigenvalue problem S -0.5 A T GAS -0.5 q=λq, then solve the eigenvalue problem and find the eigenvectors q1,q2,…,q corresponding to the smallest d eigenvalues. d , and finally through p τ =AS -0.5 q τ Calculate p1, p2, ..., p d ; Among them, performing singular value decomposition on L can obtain a unitary matrix A and a diagonal matrix S composed of singular values, numbered τ=1,2,…,d.

9. A fault detection device based on collaborative neighbor feature analysis, comprising: A model training module, on which a computer program and an intermediate matrix M for calculating fault indicators are stored; The data acquisition module acquires the process data at the latest sampling moment in real time and sends it to the data processing module; The data processing module, after receiving the process data sent by the data acquisition module, implements step 3 of the fault detection method according to any one of claims 1 to 3, and then sends the fault indicator and the judgment result of whether the fault occurs to the detection display module and the fault alarm module respectively; The detection and display module, after receiving the fault indicators sent by the data processing module, displays the specific values ​​and current sampling time in real time; A fault alarm module triggers a fault alarm according to the received judgment result of whether the fault occurs or not; It is characterized in that when the computer program stored in the model training module is executed, step 1 and step 2-1 to step 2-5 of the fault detection method according to any one of claims 1 to 3 are implemented, and then the obtained intermediate matrix M is stored.

10. A fault detection device according to claim 9, characterized in that: After the data processing module performs standardization processing on the process data at any sampling time, it further performs self-correlation standardization processing according to a fault detection method according to claim 6.

Images