State monitoring method and system based on data time-space domain change difference analysis

Abstract

The invention discloses a state monitoring method and system based on data time-space domain change difference analysis, which can analyze difference information of sampling data relative to historical normal data in time-space domain change in real time, so as to judge whether an operation state is normal or not according to the difference information. According to the state monitoring method, a time-space domain change difference analysis algorithm is used to analyze the change difference of data in a time-space domain, and the time sequence relationship of data at a plurality of continuous sampling moments and the spatial distribution neighbor relationship of data at each sampling moment are considered at the same time; and the abnormality is judged and identified through the difference evaluation index calculated in real time, so that the abnormality monitoring sensitivity is improved. In addition, the invention also discloses a state monitoring system for realizing the execution process of the state monitoring method, a program which is stored in an algorithm execution layer and is used for realizing a time-space domain change difference analysis algorithm is continuously called through a data operation module, and the operation state is also monitored based on data time-space domain difference analysis.

Description

Technical Field

[0001] The present invention relates to the technical field of data-driven fault detection, and in particular to a state monitoring method and system based on data temporal and spatial domain change difference analysis. Background Art

[0002] In the process industry, the continuous and stable operation of the production process is an important way to ensure production safety and corporate profitability. However, with the increasing complexity and large-scale production, it is becoming increasingly difficult to monitor the operation status of the process industry through mechanism modeling. In recent years, in the context of the intelligent and digital era, thanks to the various sensor data provided in real time by the data acquisition system in the process industry, the implementation of status monitoring from a data perspective to timely discover faults or anomalies in the production process has received more and more attention.

[0003] A paper titled "Just-in-time latent autoregressive residual generation for dynamic process monitoring" published in the Journal of Chemometrics in January 2024 discloses a technology for real-time generation of time series errors for process monitoring through an instant latent autoregressive error generation model. Although the errors generated in real time by this technology can reflect the abnormal changes of data in time series in real time, they cannot reflect the difference in spatial distribution changes between normal and faulty data. The corresponding fault detection efficiency needs to be further improved. The Chinese invention patent with application number / patent number 202210651302.X discloses a method for anomaly detection of the operation of a synthetic ammonia furnace using time series difference feature extraction. This method extracts time series difference features in real time through a real-time trained model. Although it analyzes the abnormality of time series changes in real time from another different angle, it still ignores the analysis of differences in spatial distribution of data.

[0004] In addition, the Chinese invention patent with application number / patent number 202210650166.2 discloses a batch production process monitoring method based on neighborhood difference feature analysis and extraction. This method timely discovers abnormalities in data neighborhood changes by implementing online neighborhood change difference analysis, thereby completing the fault detection task. However, this method only takes into account the neighborhood formed by the distance of the data in spatial distribution, and lacks further analysis of the difference changes in the time series relationship. However, this method also indirectly illustrates the advantages of implementing change difference analysis on the neighborhood of the data space in fault detection.

[0005] For process industry processes, continuous and stable operation will inevitably lead to continuous and stable changes in the relationship between sampled data in time series. The difference in time series domain changes will definitely reflect the abnormality or failure of the corresponding process object. However, due to the complex change characteristics of data caused by changes in production load, raw material batches, etc., it is also worth paying attention to the difference feature analysis of data in spatial distribution. However, time series and space are two different change characteristics, and the traditional difference analysis using all normal data as a reference set will have unnecessary interference. Therefore, feature analysis of the difference in data changes in the time and space domain requires the design of a suitable time and space domain construction method, so as to carry out targeted implementation of difference analysis, so as to ensure the sensitivity and effectiveness of subsequent fault detection. Summary of the invention

[0006] The main technical problem to be solved by the present invention is: how to analyze in real time the difference information of the sampling data used for status monitoring in terms of its temporal and spatial domain changes relative to the historical sampling data of the normal state, so as to judge whether the operating status of the process industry process is normal through the difference information.

[0007] The technical solution adopted by the present invention to solve the above-mentioned problem is: a state monitoring method and system based on the analysis of the difference between the temporal and spatial domain changes of data; wherein, a state monitoring method based on the analysis of the difference between the temporal and spatial domain changes of data includes steps 1 to 3 shown below.

[0008] Step 1: Obtain data of at least 500 consecutive sampling moments under normal working conditions from the historical database of the process industry, and then perform standardization on the data of each sampling moment to form a column vector corresponding to each sampling moment.

[0009] Step 2: Perform time-space domain variation difference analysis for each column vector in step 1, obtain the corresponding difference evaluation index, and then determine the threshold of the difference evaluation index.

[0010] Step 3: Repeat steps 3-1 and 3-2 according to the sampling time interval.

[0011] Step 3-1: Obtain the data at the latest sampling time in real time, and perform standardization on it to form a corresponding column vector.

[0012] Step 3-2: Perform spatiotemporal domain difference analysis on the column vector in step 3-1, and after obtaining the corresponding difference evaluation index, determine whether the difference evaluation index is greater than a threshold; if not, the process industry process is running normally; if so, the process industry process is running abnormally.

[0013] The key to the state monitoring method involved in the present invention lies in how to implement the time-space domain change difference analysis, and the specific implementation includes steps A-1 to A-4 shown below.

[0014] Step A-1, after setting the time series order to d and the neighbor order to k, the column vector for implementing the time-space domain change difference analysis is recorded as x0, and the column vectors corresponding to the d consecutive sampling moments before x0 are recorded as x1, x2, ..., x d ; wherein d represents an integer not greater than 15 and not less than 3, and k represents an integer not greater than 10 and not less than 5.

[0015] Step A-2: From the column vectors in step 1, they are x0, x1, x2, …, x d Determine the corresponding k nearest neighbor vectors to form the corresponding domain matrix X0,X1,X2,…,X d Then, x0,x1,x2,…,x d Form the d+1th domain matrix X d+1 =[x0,x1,x2,…,x d ].

[0016] Step A-3: Set i to 0, 1, 2, …, d, d+1 in sequence and solve the eigenvalue problem at the same time Get the eigenvector u corresponding to the maximum eigenvalue λ, and use the formula The corresponding domain coefficient vector β is calculated i ; where the superscript T represents the transpose of a matrix or vector, X i represents the i-th domain matrix.

[0017] Step A-4, after solving the generalized eigenvalue problem Gv=ηLv to obtain the maximum eigenvalue η and its corresponding eigenvector v, set the difference evaluation index equal to η; wherein,

[0018] It can be seen from the above steps A-1 to A-4 that the spatiotemporal domain involved in the present invention is specifically composed of the column vectors x1, x2, ..., x0 and the d consecutive sampling moments before it. d The domain matrix composed of the k nearest neighbor vectors and the column vector x0,x1,x2,…,x d It is composed of a domain matrix; it can be seen that the spatiotemporal domain takes into account both the neighbor relationship of the data in spatial distribution and the temporal relationship of the data in sampling time.

[0019] Determining the domain coefficient vector for each domain matrix in the above step A-3 is achieved by solving the eigenvalue problem. For those skilled in the art, the eigenvalue problem is to make X i The combined variance between the column vectors in is maximized, that is, The field coefficient vector β iThe calculation formula of is just to normalize the length of u, so that the relationship between the column vectors in the domain matrix can be quantitatively described on the same scale.

[0020] On this basis, the solution of the generalized eigenvalue problem in step A-4 is actually to further obtain X by transforming v d+1 The difference between the changes of the remaining d+1 domain matrices; the quantitative analysis of the difference is calculated in order to make X d+1 Each column vector in d+1 While maximizing the weighted variance, minimize the variance of the remaining d+1 domain matrices after weighting the corresponding domain coefficient vectors; mathematical expressions can be described as the maximization and minimization objectives as follows:

[0021]

[0022] To solve the above maximization and minimization problems, we can set v without loss of generality T When Lv=1, the equivalent maximization problem with constraints is obtained as follows:

[0023]

[0024] in, The solution of the above equation ② can be obtained by constructing the function J = v using the Lagrange multiplier method T Gv+η(1-v T Lv), and then let the partial derivatives of J with respect to v and η equal to 0, that is: 2Gv-2ηLv=0 and 1-v T Lv=0, thus we get the generalized eigenvalue problem Gv=ηLv in step A-4; if we multiply v on both sides of the equal sign of Gv=ηLv T , and combined with the constraint condition v T Lv=1, we can infer that v T Gv=ηv T Lv = η. Therefore, the largest eigenvalue η can be used to evaluate the difference in changes in the spatiotemporal domain.

[0025] From the perspective of the construction of the above-mentioned domain matrix, domain coefficient vector, optimization problem with constraints, and difference evaluation index, the spatiotemporal domain change difference analysis involved in the present invention is a new multivariate data feature analysis algorithm; the algorithm fully analyzes the change differences of the sampled data in its spatiotemporal domain, while considering the temporal relationship of the data at multiple consecutive sampling moments and the spatial distribution neighbor relationship of the data at each sampling moment, and then can judge whether the recognition is normal through the real-time calculated difference evaluation index, thereby improving the recognition efficiency and performance.

[0026] As an improvement of the present invention, the process of implementing the time-space domain change difference analysis also includes the following steps B-1 to B-2.

[0027] Step B-1: Set i to 0, 1, 2, …, d, d+1 in sequence and solve the eigenvalue problem at the same time Get the second largest eigenvalue The corresponding eigenvector And through the formula Recalculate the corresponding field coefficient vector β i .

[0028] Step B-2: According to the formula and After recalculating G and L, solve the generalized eigenvalue problem Get the maximum eigenvalue and its corresponding eigenvector After that, reset the difference evaluation index to

[0029] Step B-1 differs from step A-3 in that the second largest eigenvalue is used. The corresponding eigenvector Since different eigenvalues ​​correspond to different eigenvectors, the corresponding domain coefficient vectors weight each column vector in the domain matrix differently; in addition, since step B-2 also solves the generalized eigenvalue problem, the corresponding eigenvectors The objective function defined in the above formula ① or ② can also be satisfied, and a new feature vector is added from a different perspective to calculate the difference evaluation index. Therefore, this improved technical solution further uses more difference information on the changes in the time and space domain to determine whether the process industry process is faulty.

[0030] As another improvement of the present invention, the process of implementing the time-space domain change difference analysis includes steps C-1 to C-3 shown below in addition to steps A-1 to A-4.

[0031] Step C-1, according to the formula Y = [X0β0,X1β1,X2β2,…,X d β d ,X d+1 β d+1 ] After calculating the weighted matrix Y, we then use the formula s=v T Y and p = Ys T / (ss T ) calculates the score vector s and the loading vector p.

[0032] Step C-2: By formula X i =X i -pvT X i β i Update the domain matrices X0, X1, X2, …, X respectively d ,X d+1 Then, execute step A-3 to recalculate the corresponding domain coefficient vector.

[0033] Step C-3: After recalculating G and L, solve the generalized eigenvalue problem Get the maximum eigenvalue and its corresponding eigenvector After that, reset the difference evaluation index to

[0034] The improved technical solution recorded in the above steps C-1 to C-3 is to remove the analyzed score vector s from Y, and then execute the process of steps A-3 to A-4 again to obtain another eigenvector corresponding to the maximum eigenvalue; it can be seen that the purpose of this improvement scheme is consistent with that of the previous improvement scheme, both of which are to add a different eigenvector to satisfy the objective function defined in the above formula ① or ②, but the implementation methods are inconsistent.

[0035] Similarly, in order to obtain different eigenvectors satisfying the objective function defined in the above formula ① or ② from different angles, the present invention also provides another improvement, namely: the implementation process of the spatiotemporal domain change difference analysis includes step D shown below in addition to steps A-1 to A-4.

[0036] Step D: Solve the eigenvalue problem The largest eigenvalue and its corresponding eigenvector w, and then reset the difference evaluation index to Where I represents the identity matrix.

[0037] Different from the previous two improvement schemes, the above step D is to determine an additional eigenvector, while satisfying the objective function defined in equation ② above, and further make the newly solved eigenvector w and v orthogonal, that is, w T v = 0. Therefore, the feature vector w obtained by step D is orthogonal to the feature vector v obtained in the previous step A-4 in direction, and also realizes different analysis of difference change information, which can provide more valuable difference information for subsequent determination of whether the process industry process is running normally, so as to further improve the recognition efficiency and performance of abnormal or fault data.

[0038] In addition, considering that in step A-2 of the present invention, the construction method of the spatiotemporal domain is determined by forming a corresponding domain matrix with k nearest neighbor vectors, the present invention provides an improved scheme for determining the nearest neighbor vectors, namely: first, the implementation process of setting the nearest neighbor order equal to k in step A-1 is replaced by setting the nearest neighbor threshold equal to α, and then when determining the nearest neighbor vector in step A-2, according to steps E-1 to E-3 shown below.

[0039] Step E-1, calculate the distance between each column vector obtained in step 1 and the given column vector x, and then record the number of column vectors whose distance is less than α as Where, given a column vector x, it represents x0,x1,x2,…,x d Any column vector in .

[0040] Step E-2: Perform numerical correction, specifically: If less than 5, set like If it is greater than 10, set Other situations remain constant.

[0041] Step E-3: The vector with the smallest distance to the given column vector x column vectors are determined as the neighbor vectors, so that The neighbor vectors are merged into the corresponding domain matrix.

[0042] It can be seen from the numerical correction process in step E-2 that this improved scheme can use different numbers of neighbor vectors to form the corresponding domain matrix for different given column vectors x, and can flexibly and specifically determine the corresponding domain matrix, which is a more suitable method for determining the spatiotemporal domain.

[0043] Since the above-mentioned neighbor threshold is a key parameter of the improved technical solution, the preferred process of setting α is: first calculate the distance between each column vector obtained in step 1, and then set the neighbor threshold equal to the average value of all calculated distances; this method of determining the neighbor threshold α is more objective than manually determining α through experience, and there will be no problem of setting α too small or too large, resulting in the use of the same number of neighbor vectors to form the domain matrix for different given column vectors.

[0044] Finally, since steps 3-1 to 3-2 are the implementation process of activating the spatiotemporal domain change difference analysis in real time based on the data at the latest sampling time, it is necessary to solve the generalized eigenvalue problem in step A-4 in real time. In order to improve the calculation and solution efficiency, the present invention preferably solves according to the following steps:

[0045] Step G-1: Multiply L on both sides of the equality sign of the generalized eigenvalue problem Gv=ηLv -1 Get the eigenvalue problem L - 1 After Gv=ηv, v is initialized to an arbitrary real vector.

[0046] Step G-2, sequentially through v = L -1 Gv and Keep updating v until v converges.

[0047] Step G-3, by the formula η = v T L -1 Gv calculates the maximum eigenvalue η.

[0048] The above solution to the generalized eigenvalue problem makes full use of the technical advantages of computers in numerical iterative updating, and can quickly obtain the eigenvector v and its corresponding maximum eigenvalue η. Therefore, the generalized eigenvalue problems in step B-2 and step C-3 can be solved in the same way.

[0049] In addition, since step A-4 directly determines the maximum eigenvalue as the difference index, and whether the process industry process is running normally is determined by comparing the difference index with the threshold, that is, step 3-2. Therefore, in order to further improve the computational efficiency of implementing the time-space domain change difference analysis in step 3-1, when solving the generalized eigenvalue problem Gv=ηLv according to the above-mentioned improved scheme, it is preferred to replace the implementation process of step G-2 with: sequentially passing v=L -1 Gv and Constantly update v and at the same time according to the formula σ=v T L -1 Gv calculates the corresponding eigenvalue σ. If σ is greater than the threshold determined in step 3, the update of v is stopped and step G-3 is executed again.

[0050] Finally, it should be noted that in step 2 of the above-mentioned various technical solutions, there are many ways to determine the threshold. The kernel density estimation method can be used to implement probability density estimation on all the difference evaluation indicators in step 2, thereby setting the value under the given confidence limit condition as the threshold. Alternatively, the threshold can be simply and directly determined to be equal to the maximum value of all the difference evaluation indicators in step 2; wherein the confidence limit can usually be 99% or 99.75%.

[0051] Regarding the determination of the threshold in step 2, the state monitoring method and improved technical solution thereof according to the present invention can preferably determine the threshold of the difference evaluation index according to steps H-1 to H-2 shown below.

[0052] Step H-1, calculating the average value μ and standard deviation δ of all the difference evaluation indicators in step 2.

[0053] Step H-2: Set the threshold value equal to μ+3×δ.

[0054] Compared with the kernel density estimation method, the above method of determining the threshold does not require artificial setting of hyperparameters such as kernel parameters and confidence limits; at the same time, it can also avoid the problem of imprecise threshold setting caused by extreme measures such as the maximum value; the above method of determining the threshold comprehensively considers the possible range of changes in the difference evaluation index from an objective level, and the threshold it sets is relatively more universal.

[0055] Based on the same inventive concept, the present invention also provides a state monitoring system based on data temporal and spatial domain change difference analysis; the state monitoring system includes a software subsystem consisting of a model training layer and an algorithm execution layer, and a hardware subsystem consisting of a data acquisition module, a data calculation module and a state display module.

[0056] Among them, the model training layer provides at least 500 standardized column vectors for the data operation module, and also provides a threshold for judging whether it is abnormal for the status display module; the data acquisition module obtains the data of the latest sampling time in real time according to the sampling time interval of the process industry process, and sends it to the data operation module immediately; the data operation module performs standardized processing on the received data, and then calls the stored program in the algorithm execution layer to perform further calculation processing on the data, and after obtaining the corresponding difference evaluation index, the difference evaluation index is sent to the status display module immediately; the data operation module performs standardized processing on the received data, and then calls the stored program in the algorithm execution layer to perform further calculation processing on the data, and after obtaining the corresponding difference evaluation index, the difference evaluation index is sent to the status display module immediately.

[0057] The key to this state monitoring system lies in that the algorithm execution layer stores programs that can implement the time-space domain change difference analysis in the above-mentioned technical solutions, and the model training layer can execute steps 1 to 2 in the above-mentioned technical solutions according to the initialization instructions to obtain the column vector required for implementing the time-space domain change difference analysis, as well as the threshold for judging whether it is abnormal. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 The present invention is a schematic diagram of the implementation process of the state monitoring method.

[0059] Figure 2 Schematic diagram of the three-phase flow mixing process.

[0060] Figure 3The present invention relates to a condition monitoring method and two conventional condition monitoring diagrams.

[0061] Figure 4 It is a schematic diagram of implementing three improved solutions for the time-space domain change difference analysis involved in the present invention.

[0062] Figure 5 for Figure 4 The status monitoring diagrams corresponding to the three improvement schemes.

[0063] Figure 6 The present invention relates to a schematic diagram of an improved implementation flow of a condition monitoring method.

[0064] Figure 7 The present invention is a schematic diagram of the structure of the state monitoring system. DETAILED DESCRIPTION

[0065] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0066] according to Figure 1 The implementation process shown in the figure, the specific implementation mode of the present invention will be Figure 2 The three-phase flow mixing process shown implements condition monitoring; Figure 2 The three-phase flow mixing process shown in the figure can mix air (AIR), oil (OIL) and water (WATER) in various proportions. The industrial process can measure the data of 24 sensors per second, including Figure 2 The four types of sensors shown in the figure, PT, FT, LI, VC, and PO, measure pressure, flow, liquid level, valve opening, and current respectively;

[0067] according to Figure 1 The implementation process shown implements state monitoring of the three-phase flow mixing process, including steps 1 to 3 shown below.

[0068] Step 1: From the historical database of the three-phase flow mixing process, obtain the data of 10371 continuous sampling moments under normal working conditions, and then perform standardization on the data of each sampling moment to form a column vector corresponding to each sampling moment.

[0069] Step 2: Perform time-space domain variation difference analysis for each column vector in step 1, obtain the corresponding difference evaluation index, and then determine the threshold of the difference evaluation index.

[0070] Step 3: Repeat steps 3-1 and 3-2 according to the sampling time interval.

[0071] Step 3-1: Obtain the data at the latest sampling time in real time, and perform standardization on it to form a corresponding column vector.

[0072] Step 3-2, perform time-space domain variation difference analysis on the column vector in step 3-1, and after obtaining the corresponding difference evaluation index, determine whether the difference evaluation index is greater than a threshold; if not, the three-phase flow mixing process operates normally; if so, the three-phase flow mixing process operates abnormally.

[0073] The implementation process of the spatiotemporal domain change difference analysis includes steps A-1 to A-4 shown below.

[0074] Step A-1, after setting the time series order to 3 and the neighbor order to 5, the column vector for implementing the time-space domain change difference analysis is recorded as x0, and the column vectors corresponding to the three consecutive sampling moments before x0 are recorded as x1, x2, and x3 respectively.

[0075] Step A-2: Determine the five nearest neighbor vectors for x0, x1, x2, and x3 from the column vectors in step 1, thereby forming the corresponding domain matrices X0, X1, X2, and X3, and then combine x0, x1, x2, and x3 to form the fourth domain matrix X d+1 =[x0,x1,x2,x3].

[0076] Step A-3: Set i to 0, 1, 2, 3, 4 in sequence and solve the eigenvalue problem at the same time Get the eigenvector u corresponding to the maximum eigenvalue λ, and use the formula The corresponding domain coefficient vector β is calculated i .

[0077] Step A-4, solve the generalized eigenvalue problem Gv = ηLv to obtain the eigenvector v corresponding to the maximum eigenvalue η, and then set the difference evaluation index equal to η; where the matrix matrix

[0078] According to the above steps 3-1 to 3-2, the data at the 5811 latest sampling moments are monitored, where the abnormal state is introduced at the 1566 latest sampling moments and eliminated at the 5181 sampling moments; the difference evaluation index at each sampling moment can be plotted as follows Figure 3 The monitoring diagram shown in Figure (a) is as follows; the horizontal line represents the threshold value, and ln(Q) represents the logarithmic value of the abnormal evaluation index Q.

[0079] In order to demonstrate the technical advantages of the condition monitoring method according to the present invention, Figure 3 Figures (b) and (c) of the paper show the monitoring diagrams of two traditional methods. Figure 3From the comparison of the three figures, it can be clearly found that the state monitoring method involved in the present invention can detect abnormal conditions earlier and has a great advantage in the sensitivity of abnormality detection. This is mainly due to the fact that the method simultaneously implements the time-space domain change difference analysis for both the temporal relationship and the spatial neighbor relationship.

[0080] The second implementation case of the present invention adopts an improved time-space domain change difference analysis based on the above implementation case. The corresponding implementation process is as follows: Figure 4 As shown in Figure (a), specifically, steps B-1 to B-2 are added on the basis of steps A-1 to A-4.

[0081] The third implementation example of the present invention uses another improved time-space domain change difference analysis, and the corresponding implementation process is as follows: Figure 4 As shown in Figure (b), specifically, steps C-1 to C-3 are added on the basis of steps A-1 to A-4.

[0082] The fourth implementation case of the present invention also uses another improved time-space domain change difference analysis, and the corresponding implementation process is as follows: Figure 4 As shown in Figure (c), step D is added on the basis of steps A-1 to A-4.

[0083] Figure 5 Given Figure 4 The state monitoring diagrams corresponding to the three improvement schemes; compared with Figure 3 As shown in Figure (a), the monitoring diagrams corresponding to these three improvement schemes can capture anomalies earlier, further improving the reliability of condition monitoring.

[0084] The fifth implementation case of the present invention is based on the fourth implementation case, and the implementation process of setting the neighbor order equal to k in step A-1 is replaced by setting the neighbor threshold equal to α, and when determining the neighbor vector in step A-2, according to Figure 6 Steps E-1 to E-3 shown in ; in addition, the implementation case also sets the neighbor threshold according to the following process: first calculate the distance between each column vector obtained in step 1, and then set the neighbor threshold equal to the average of all calculated distances.

[0085] In addition, in the above five implementation cases, the generalized eigenvalue problem can also be solved by steps G-1 to G-3 as shown below.

[0086] Step G-1: Multiply L on both sides of the equality sign of the generalized eigenvalue problem Gv=ηLv -1 Get the eigenvalue problem L - 1 After Gv=ηv, v is initialized to an arbitrary real vector.

[0087] Step G-2, sequentially through v = L -1 Gv and Keep updating v until v converges.

[0088] Step G-3, by the formula η = v T L -1 Gv calculates the maximum eigenvalue η.

[0089] In order to improve the efficiency of online computing, when implementing the time-space domain change difference analysis in step 3-2, if the above steps G-1 to G-3 are used to solve the generalized eigenvalue problem, the implementation process of step G-2 can be replaced by: sequentially passing v=L - 1 Gv and Constantly update v and at the same time according to the formula σ=v T L -1 Gv calculates the corresponding eigenvalue σ. If σ is greater than the threshold value determined in step 3, the update of v is stopped and step G-3 is executed again.

[0090] Based on the same inventive concept, the present invention is also as follows Figure 7 The structural schematic diagram shown implements a state monitoring system based on data temporal and spatial domain change difference analysis, including a software subsystem (10) composed of a model training layer (101) and an algorithm execution layer (102), and a hardware subsystem (20) composed of a data acquisition module (201), a data calculation module (202) and a state display module (203).

[0091] exist Figure 7 In the invention, the model training layer (101) provides 10371 column vectors after standardization processing for the data operation module (202), and also provides a threshold value for judging whether it is abnormal for the state display module (203); the data acquisition module (201) obtains the data of the latest sampling time in real time according to the sampling time interval of the three-phase flow mixing process, and sends it to the data operation module (202) in real time; the data operation module (202) performs standardization processing on the received data, and then calls the program stored in the algorithm execution layer (102) that can realize the implementation process of time-space domain change difference analysis to perform further operation processing on the data, and after obtaining the corresponding difference evaluation index, the difference evaluation index is sent to the state display module (203) in real time; after receiving the difference evaluation index, the state display module (203) uses the latest sampling time as the horizontal axis data and the received difference evaluation index as the vertical axis data for real-time display, and at the same time displays the threshold value in the form of a horizontal line, and the displayed scatter plot is consistent with the state display module (203). Figure 3 and Figure 5The same as the scatter plot in , when the scatter points are above the horizontal line, it means that the process object has entered an abnormal state.

[0092] In addition, the model training layer (201) can execute steps 1 to 2 in the above-mentioned various technical solutions and their improved solutions according to the initialization instructions, thereby obtaining and providing 10371 standardized column vectors to the data operation module (202) and the threshold value used to judge whether there is an abnormality in the status display module (203).

[0093] The above explanation of the technical solution disclosed in the present invention in combination with the accompanying drawings is illustrative rather than restrictive. The technical solution disclosed in the present invention is not limited to the above-mentioned implementation cases. Any modifications made without departing from the purpose of the technical solution disclosed in the present invention and the scope of protection of the claims shall fall within the protection scope of the present invention.

Claims

1. A condition monitoring method based on data temporal and spatial domain change difference analysis, comprising steps 1 to 3 shown below: Step 1: Obtain data from the historical database of the process industry for at least 500 consecutive sampling moments under normal operating conditions. After the data at each sampling moment are standardized, the corresponding column vectors at each sampling moment are formed; Step 2: Perform time-space domain variation difference analysis for each column vector in step 1, obtain the corresponding difference evaluation index, and then determine the threshold of the difference evaluation index; Step 3, repeating steps 3-1 to 3-2 according to the sampling time interval; Step 3-1, obtain the data of the latest sampling time in real time, and perform standardization on it to form a corresponding column vector; Step 3-2, performing time-space domain variation difference analysis for the column vector in step 3-1, and after obtaining the corresponding difference evaluation index, determining whether the difference evaluation index is greater than a threshold; if not, the process industry process is running normally; if so, the process industry process is running abnormally; Features: The implementation process of the spatiotemporal domain change difference analysis includes the following steps A-1 to A-4; Step A-1, after setting the time series order to d and the neighbor order to k, the column vector for implementing the time-space domain change difference analysis is recorded as x0, and the column vectors corresponding to the d consecutive sampling moments before x0 are recorded as x1, x2, ..., x d ; wherein d represents an integer not greater than 15 and not less than 3, and k represents an integer not greater than 10 and not less than 5; Step A-2: From the column vectors in step 1, they are x0, x1, x2, …, x d Determine the corresponding k nearest neighbor vectors to form the corresponding domain matrix X0,X1,X2,…,X d Then, x0,x1,x2,…,x d Form the d+1th domain matrix X d+1 =[x0,x1,x2,…,x d ]; Step A-3: Set i to 0, 1, 2, …, d, d+1 in sequence and solve the eigenvalue problem at the same time Get the eigenvector u corresponding to the maximum eigenvalue λ, and use the formula The corresponding domain coefficient vector β is calculated i ; where the superscript T represents the transpose of a matrix or vector, X i represents the i-th domain matrix; Step A-4, solve the generalized eigenvalue problem Gv = ηLv to obtain the maximum eigenvalue η and its corresponding eigenvector v, and then set the difference evaluation index equal to η; where the matrix matrix 2. A condition monitoring method according to claim 1, characterized in that: The implementation process of the temporal and spatial domain change difference analysis also includes the following steps B-1 to B-2: Step B-1: Set i to 0, 1, 2, …, d, d+1 in sequence and solve the eigenvalue problem at the same time Get the second largest eigenvalue The corresponding eigenvector And through the formula Recalculate the corresponding field coefficient vector β i ; Step B-2: According to the formula and After recalculating G and L, solve the generalized eigenvalue problem Get the maximum eigenvalue and its corresponding eigenvector After that, reset the difference evaluation index to 3. A condition monitoring method according to claim 1, characterized in that: The implementation process of the temporal and spatial domain change difference analysis also includes the following steps C-1 to C-3: Step C-1, according to the formula Y = [X0β0,X1β1,X2β2,…,X d β d ,X d+1 β d+1 ] After calculating the weighted matrix Y, we then use the formula s=v T Y and p = Ys T / (ss T ) Calculate the score vector s and the load vector p; Step C-2: By formula X i =X i -pv T X i β i Update the domain matrices X0, X1, X2, …, X respectively d ,X d+1 Then, execute step A-3 to recalculate the corresponding domain coefficient vector; Step C-3: After recalculating G and L, solve the generalized eigenvalue problem Get the maximum eigenvalue and its corresponding eigenvector After that, reset the difference evaluation index to 4. A condition monitoring method according to claim 1, characterized in that: The implementation process of the temporal and spatial domain change difference analysis also includes the following step D: Step D: Solve the eigenvalue problem The largest eigenvalue and its corresponding eigenvector w, and then reset the difference evaluation index to Where I represents the identity matrix.

5. A condition monitoring method according to any one of claims 1 to 4, characterized in that: The implementation process of setting the neighbor order equal to k in step A-1 is replaced by setting the neighbor threshold equal to α, and when determining the neighbor vector in step A-2, according to steps E-1 to E-3 shown below: Step E-1, calculate the distance between each column vector obtained in step 1 and the given column vector x, and then record the number of column vectors whose distance is less than α as Where, given a column vector x, it represents x0,x1,x2,…,x d Any column vector in ; Step E-2: Perform numerical correction if If less than 5, set like If it is greater than 10, set Other situations remain constant; Step E-3: The vector with the smallest distance to the given column vector x column vectors are determined as the neighbor vectors, so that The neighbor vectors are merged into the corresponding domain matrix.

6. A condition monitoring method according to claim 5, characterized in that: The process of setting the neighbor threshold is: firstly calculate the distances between the column vectors obtained in step 1, and then set the neighbor threshold to be equal to the average value of all calculated distances.

7. A condition monitoring method according to claim 1, characterized in that: The generalized eigenvalue problem in step A-4 is solved according to the following steps: Step G-1: Multiply L on both sides of the equality sign of the generalized eigenvalue problem Gv=ηLv -1 Get the eigenvalue problem L -1 After Gv=ηv, initialize v to any real vector; Step G-2, sequentially through v = L -1 Gv and Keep updating v until v converges; Step G-3, by formula η = v T L -1 Gv calculates the maximum eigenvalue η.

8. A condition monitoring method according to claim 7, characterized in that: The implementation process of step G-2 is replaced by: sequentially passing v=L -1 Gv and Constantly update v and at the same time according to the formula σ=v T L -1 Gv calculates the corresponding eigenvalue σ. If σ is greater than the threshold value determined in step 3, the update of v is stopped and step G-3 is executed again.

9. A condition monitoring method according to any one of claims 1 to 4, 6 and 7, characterized in that: The step 2 is to determine the threshold of the difference evaluation index according to the following steps: Step H-1, calculating the average value μ and standard deviation δ of all difference evaluation indicators in step 2; Step H-2: Set the threshold value equal to μ+3×δ.

10. A state monitoring system based on data spatiotemporal difference analysis, comprising a software subsystem consisting of a model training layer and an algorithm execution layer, and a hardware subsystem consisting of a data acquisition module, a data calculation module and a state display module; wherein, The model training layer provides at least 500 standardized column vectors for the data operation module, and also provides the threshold for judging abnormality for the status display module; The data acquisition module obtains the data of the latest sampling time in real time according to the sampling time interval of the process industry process, and sends it to the data calculation module immediately; The data operation module performs standardization processing on the received data, and then calls the stored program in the algorithm execution layer to perform further operation processing on the data, and after obtaining the corresponding difference evaluation index, the difference evaluation index is immediately sent to the status display module; After receiving the difference evaluation index, the status display module uses the latest sampling time as the horizontal axis data and the received difference evaluation index as the vertical axis data for real-time display, and at the same time displays the threshold value in the form of a horizontal line; Its characteristic is that the algorithm execution layer stores a program that can implement the time-space domain change difference analysis of a condition monitoring method described in any one of claims 1 to 7, and the model training layer can execute steps 1 to 2 of the condition monitoring method described in any one of claims 1 to 7 according to the initialization instructions.

Images