Working condition monitoring method and system based on index slow characteristics, and storage medium

Through the working condition monitoring method based on the index slow feature, the problem of false alarms and difficulty in identification when the operating condition changes in traditional methods is solved, and the accurate identification of operation deviations and process abnormalities is achieved, which improves the accuracy of working condition monitoring and stability.

CN118939993APending Publication Date: 2024-11-12EAST CHINA UNIV OF SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410907594.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-08
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

Traditional dynamic multivariate statistical process monitoring methods are difficult to effectively extract dynamic information from process data, resulting in false positives easily when operating conditions change, and it is difficult to accurately identify operation deviations and process abnormalities.

Method used

Using the working condition monitoring method based on the exponential slow feature, the conditional driving variables affecting the energy consumption level are determined, and the clustering process divides high-energy, low-energy consumption and transitional working conditions are divided using the exponential slow feature method. Space, establish steady-state and dynamic statistics, determine their significance threshold and control range, and monitor the operating status of the current working conditions.

Benefits of technology

It effectively avoids false alarms caused by changes in operating conditions, accurately identify operating deviations and process abnormalities in new operating conditions, and improves the accuracy and stability of operating conditions monitoring.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118939993A_ABST
    Figure CN118939993A_ABST
Patent Text Reader

Abstract

The invention provides a working condition monitoring method and system based on index slow characteristics and a storage system. The working condition monitoring method based on the index slow characteristics comprises the following steps: determining a plurality of condition driving variables influencing the energy consumption level, and carrying out clustering processing on the condition driving variables; dividing the monitoring space of the online operation data of the atmospheric and vacuum distillation unit into a slowly-changing feature space and a quickly-changing feature space; respectively establishing a first steady-state statistic based on the slow change feature and a second steady-state statistic based on the quick change feature, and respectively determining a first threshold and a second threshold; respectively establishing a first dynamic statistical magnitude based on the first-order derivative of the slow change characteristic and a second dynamic statistical magnitude based on the first-order derivative of the quick change characteristic, and respectively determining a first control range and a second control range of the first dynamic statistical magnitude and the second dynamic statistical magnitude; and monitoring the running state of the current working condition according to the first steady-state statistic, the second steady-state statistic, the first dynamic statistic and the second dynamic statistic.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of petrochemical industry, and in particular to a working condition monitoring method based on exponential slow features, a working condition monitoring system based on exponential slow features, and a computer-readable storage medium. Background Art

[0002] In the distillation process of the atmospheric and vacuum unit in the refinery, the stable operation of the unit faces challenges due to factors such as the volatility of raw materials, the variability of operating conditions, equipment status, the technical level of operators, and changes in product demand. The complex energy and quality interactions and the instability of operating conditions caused by these factors require the actual production process to continuously adjust the operating parameters of the unit to adapt to different processing modes, which in turn causes the unit to exhibit dynamic unstable conditions.

[0003] When processing different types of crude oil, even under the same operating conditions, the unit will show different stable operating characteristics due to the different properties of the crude oil. Although multivariate statistical process control (MSPC) methods can effectively detect abnormal events in the process without basic model knowledge, these methods may fail and cause frequent false alarms when the unit operating conditions change.

[0004] Traditional dynamic multivariate statistical process monitoring methods, such as dynamic principal component analysis, dynamic partial least squares, and dynamic independent component analysis, are usually unable to effectively extract dynamic information from process data. These methods assume that latent variables are statistically independent when designed, and do not fully consider the slow adjustment effect of actual data, making it difficult to distinguish between changes in operating status and actual faults.

[0005] In order to overcome the above-mentioned defects of the prior art, the art urgently needs a condition monitoring technology based on exponential slow features to avoid false alarms due to changes in operating conditions and accurately identify operating deviations and process anomalies in new operating conditions. Summary of the invention

[0006] A brief summary of one or more aspects is given below to provide a basic understanding of these aspects. This summary is not an exhaustive overview of all conceived aspects, and is neither intended to identify the key or decisive elements of all aspects nor to define the scope of any or all aspects. Its only purpose is to give some concepts of one or more aspects in a simplified form as a prelude to a more detailed description that will be given later.

[0007] In order to overcome the above-mentioned defects of the prior art, the present invention provides a condition monitoring method based on exponential slow features, a condition monitoring system based on exponential slow features, and a computer-readable storage medium, which are used to avoid false alarms due to changes in operating conditions and accurately identify operating deviations and process anomalies in new operating conditions.

[0008] Specifically, the operating condition monitoring method based on exponential slow features provided according to the first aspect of the present invention includes the following steps: determining multiple conditional driving variables that affect the energy consumption level, and clustering the multiple conditional driving variables to divide and determine high energy consumption conditions, low energy consumption conditions and transition conditions; using the exponential slow feature method, dividing the monitoring space of the online operation data of the constant pressure reduction device into a slow-changing feature space and a fast-changing feature space to extract slow-changing features and fast-changing features respectively; establishing a first steady-state statistic based on the slow-changing feature and a second steady-state statistic based on the fast-changing feature, and determining a first threshold and a second threshold of their significance levels respectively; establishing a first dynamic statistic based on the first-order derivative of the slow-changing feature and a second dynamic statistic based on the first-order derivative of the fast-changing feature, and determining a first control range and a second control range respectively; and monitoring the operating status of the current operating condition according to the first steady-state statistic, the second steady-state statistic, the first dynamic statistic, and the second dynamic statistic.

[0009] Furthermore, in some embodiments of the present invention, the step of determining multiple conditional driving variables that affect the energy consumption level includes: obtaining multiple candidate conditional driving variables involved in the process flow to be monitored, and their corresponding sample data sets; and using the XGBoost model to perform a first variable screening based on regression prediction on each of the candidate conditional driving variables and their corresponding sample data in the sample data set, and / or based on plant mechanism experience, perform a second variable screening of key variables on multiple indicator models of the process flow, so as to determine the multiple conditional driving variables that affect the energy consumption level.

[0010] Furthermore, in some embodiments of the present invention, the step of clustering the multiple conditional driving variables to divide and determine high-energy consumption conditions, low-energy consumption conditions and transition conditions includes: using the DBSCAN algorithm to perform cluster analysis on multiple modes under each of the conditional driving variables to divide the multiple sample data in the sample data set into multiple clusters; and according to the operating status corresponding to each of the clusters, defining them as the high-energy consumption condition, the low-energy consumption condition and the transition condition, respectively.

[0011] Further, in some embodiments of the present invention, before dividing the monitoring space of the online operation data of the atmospheric and vacuum distillation device into a slowly changing feature space and a quickly changing feature space, the operating condition monitoring method further includes the following steps: calculating the mean and / or variance of each of the sample data in the sample data set; and performing data standardization on the online operation data of the atmospheric and vacuum distillation device according to the mean and / or the variance:

[0012]

[0013] Wherein, mean(X) represents the mean of each sample data in the sample data set, std(X) represents the variance of each sample data in the sample data set, and x new represents the online operation data before normalization, x′ new Represents standardized online operation data.

[0014] Further, in some embodiments of the present invention, the step of using the exponential slow feature method to divide the monitoring space of the online operation data of the atmospheric and vacuum distillation device into a slow-changing feature space and a fast-changing feature space to extract the slow-changing features and the fast-changing features respectively includes: constructing a covariance matrix B and a first-order derivative covariance matrix A of the input features composed of the multiple conditional driving variables:

[0015]

[0016] in, represents the variance of the nth eigenvector in the input feature, σ n,1 represents the covariance of two columns of different eigenvectors, represents the variance of the derivative of the nth eigenvector in the input feature, Representing the covariance of two columns of different eigenvector derivatives; and mapping all elements of the covariance matrix B and the first-order derivative covariance matrix A to the range of [-1,1], and performing SFA operation in the corresponding reduced space to divide the monitoring space of the online operation data of the constant pressure reduction device into a slow-varying feature space and a fast-varying feature space, and extracting the slow-varying features and the fast-varying features respectively.

[0017] Further, in some embodiments of the present invention, the step of performing SFA operation in the corresponding reduced space includes: for an m-dimensional input signal x(t)=[x1(t),x2(t),x3(t),…,x m (t)] T , through the linear mapping matrix g(x) = [g1(x),…,g j (x)] TMap the m-dimensional input signal to the corresponding output signal s(t)=[s1(t),…,s j (t)] T , and minimizes the square of the time derivative of the output signal

[0018]

[0019] Where t∈[t0,t1] represents time, s j (t) = g j (x(t)), 〈·〉 t represents the average value along the time axis, <s> t express ω i is the i-th mapping function g i The weight of (x), Apply singular value decomposition to perform singular value decomposition on the covariance matrix B:

[0020] B=X T X=UΛU T

[0021] Wherein, the m-dimensional input signal x(t) is whitened to obtain z=Λ -1 / 2 U T X=QX,Q=Λ -1 / 2 U T is the whitening matrix, According to the result of the singular value decomposition, the projection matrix W is determined:

[0022] W=Q T P=UΛ -1 / 2 P;

[0023] And according to the projection matrix W, the monitoring space of the online operation data of the atmospheric and vacuum distillation device is divided into the slowly changing feature space and the quickly changing feature space:

[0024] s d =W d X T

[0025] s e =W e X T

[0026] Among them, s d is the slow-varying feature, s e is the fast-changing feature.

[0027] Further, in some embodiments of the present invention, the first steady-state statistic is expressed as The second steady-state statistic is expressed as

[0028] Furthermore, in some embodiments of the present invention, the first-order derivative of the slowly varying feature is represented by The first dynamic statistic is expressed as The first-order derivative of the fast-varying characteristic is expressed as The second dynamic statistic is expressed as

[0029] Further, in some embodiments of the present invention, the step of respectively determining the first control range and the second control range includes: determining a probability density function of a kernel density estimation:

[0030]

[0031] Wherein, x represents the first dynamic statistic or the second dynamic statistic, x i represents the sample data point, n is the number of samples, h is the window bandwidth parameter, and K(·) is the Gaussian kernel function:

[0032]

[0033] Traverse each y data point of the probability density function to determine the first point y[i] that exceeds the preset confidence level α; based on the point y[i], calculate the control range of the corresponding dynamic statistic:

[0034] kzx=x[i-1]+(x[i]-x[i-1])*(y[i]-α) / (y[i]-y[i-1])

[0035] Among them, kzx represents the control limit.

[0036] Further, in some embodiments of the present invention, the step of monitoring the operating state of the current operating condition according to the first steady-state statistic, the second steady-state statistic, the first dynamic statistic, and the second dynamic statistic includes: in response to the first steady-state statistic being less than or equal to the first threshold, the second steady-state statistic being less than or equal to the second threshold, the first dynamic statistic being located in the first control range, and the second dynamic statistic being located in the second control range, determining that the current operating condition is operating normally; in response to the first steady-state statistic being less than or equal to the first threshold, the second steady-state statistic being less than or equal to the second threshold, and at least one of the first dynamic statistic and the second dynamic statistic being outside the corresponding control range, determining that the current operating condition has a condition drift phenomenon; in response to at least one of the first steady-state statistic and the second steady-state statistic being greater than the corresponding threshold, the first dynamic statistic being located in the first control range, and the second dynamic statistic being located in the second control range, determining the current operating condition; and in response to at least one of the first steady-state statistic and the second steady-state statistic being greater than the corresponding threshold, and at least one of the first dynamic statistic and the second dynamic statistic being outside the corresponding control range, determining that the current operating condition cannot be maintained.

[0037] In addition, the operating condition monitoring system based on exponential slow features provided according to the second aspect of the present invention includes: a memory on which computer instructions are stored; and a processor, connected to the memory and configured to execute the computer instructions stored on the memory to implement the operating condition monitoring method as described in any one of the first aspects of the present invention.

[0038] In addition, the computer-readable storage medium provided according to the third aspect of the present invention stores computer instructions, which, when executed by a processor, implement the working condition monitoring method as described in any one of the first aspect of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The above features and advantages of the present invention can be better understood after reading the detailed description of the embodiments of the present disclosure in conjunction with the following drawings. In the drawings, the components are not necessarily drawn to scale, and components with similar related properties or features may have the same or similar reference numerals.

[0040] Figure 1 A schematic flow chart of a method for monitoring operating conditions based on exponential slow features according to some embodiments of the present invention is shown.

[0041] Figure 2A A variable schematic diagram of the fuel gas flow rate and energy consumption of an atmospheric pressure furnace provided according to some embodiments of the present invention is shown.

[0042] Figure 2B A variable schematic diagram of the fuel gas flow rate and energy consumption of a pressure reducing furnace provided according to some embodiments of the present invention is shown.

[0043] Figure 2C A variable schematic diagram of fuel gas relative density and energy consumption provided according to some embodiments of the present invention is shown.

[0044] Figure 2D A variable schematic diagram of system high-pressure gas pressure and energy consumption provided according to some embodiments of the present invention is shown.

[0045] Figure 3 A schematic diagram of operating condition division based on DBSCAN provided according to some embodiments of the present invention is shown.

[0046] Figure 4 A schematic diagram of an online fine-scale monitoring solution based on the ESFA method provided according to some embodiments of the present invention is shown.

[0047] Figure 5A A schematic diagram showing monitoring results of the ESFA method provided according to some embodiments of the present invention is shown.

[0048] Figure 5B A schematic diagram of monitoring results of the SFA method provided according to some embodiments of the present invention is shown.

[0049] Figure 5C A schematic diagram of monitoring results of the PCA method provided according to some embodiments of the present invention is shown.

[0050] Figure 5D A schematic diagram of monitoring results of the ICA method provided according to some embodiments of the present invention is shown. DETAILED DESCRIPTION

[0051] The following specific embodiments illustrate the implementation of the present invention, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. Although the description of the present invention will be introduced in conjunction with the preferred embodiment, this does not mean that the features of this invention are limited to this implementation. On the contrary, the purpose of introducing the invention in conjunction with the implementation is to cover other options or modifications that may be extended based on the claims of the present invention. In order to provide a deep understanding of the present invention, the following description will include many specific details. The present invention can also be implemented without using these details. In addition, in order to avoid confusion or blurring the focus of the present invention, some specific details will be omitted in the description.

[0052] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0053] In addition, the terms "upper", "lower", "left", "right", "top", "bottom", "horizontal", and "vertical" used in the following description should be understood as the directions shown in the paragraph and the related drawings. Such relative terms are only used for the convenience of description and do not mean that the device described therein must be manufactured or operated in a specific direction, and therefore should not be understood as limiting the present invention.

[0054] It is understood that although the terms "first", "second", "third", etc. may be used herein to describe various components, regions, layers and / or parts, these components, regions, layers and / or parts should not be limited by these terms, and these terms are only used to distinguish different components, regions, layers and / or parts. Therefore, the first component, region, layer and / or part discussed below may be referred to as a second component, region, layer and / or part without departing from some embodiments of the present invention.

[0055] As mentioned above, traditional dynamic multivariate statistical process monitoring methods, such as dynamic principal component analysis, dynamic partial least squares, and dynamic independent component analysis, are usually unable to effectively extract dynamic information from process data. These methods assume that latent variables are statistically independent when designed, and do not fully consider the slow adjustment effect of actual data, making it difficult to distinguish between changes in operating status and actual faults.

[0056] In order to overcome the above-mentioned defects of the prior art, the present invention provides a condition monitoring method based on exponential slow features, a condition monitoring system based on exponential slow features, and a computer-readable storage medium, which are used to avoid false alarms due to changes in operating conditions and accurately identify operating deviations and process anomalies in new operating conditions.

[0057] In some non-limiting embodiments, the operating condition monitoring method based on the exponential slow feature provided in the first aspect of the present invention can be implemented based on the operating condition monitoring system based on the exponential slow feature provided in the second aspect of the present invention. Specifically, the operating condition monitoring system based on the exponential slow feature is configured with a memory and a processor. The memory includes but is not limited to the above-mentioned computer-readable storage medium provided in the third aspect of the present invention, on which computer instructions are stored. The processor is connected to the memory and is configured to execute the computer instructions stored on the memory to implement the above-mentioned operating condition monitoring method based on the exponential slow feature provided in the first aspect of the present invention.

[0058] The following will describe the working principle of the above-mentioned working condition monitoring system based on the exponential slow feature in combination with some embodiments of the working condition monitoring method based on the exponential slow feature. Those skilled in the art can understand that these embodiments of the working condition monitoring method based on the exponential slow feature are only some non-restrictive implementation methods provided by the present invention, which are intended to clearly demonstrate the main concept of the present invention and provide some specific solutions that are convenient for the public to implement, rather than to limit all functions or all working modes of the working condition monitoring system based on the exponential slow feature. Similarly, the working condition monitoring system based on the exponential slow feature is also only some restrictive implementation methods provided by the present invention, and does not constitute a limitation on the execution subject or execution order of each step in these working condition monitoring methods based on the exponential slow feature.

[0059] Specifically, the operating condition monitoring method based on the exponential slow feature first determines multiple conditional driving variables that affect the energy consumption level, and clusters the multiple conditional driving variables to divide and determine high energy consumption conditions, low energy consumption conditions and transition conditions. Then, using the exponential slow feature method, the monitoring space of the online operation data of the constant pressure reduction device is divided into a slow-changing feature space and a fast-changing feature space to extract the slow-changing features and the fast-changing features respectively. After that, a first steady-state statistic based on the slow-changing feature and a second steady-state statistic based on the fast-changing feature are respectively established, and the first threshold and the second threshold (for example: 95%) of their significance levels are respectively determined. After that, a first dynamic statistic based on the first-order derivative of the slow-changing feature and a second dynamic statistic based on the first-order derivative of the fast-changing feature are respectively established, and their first control range and second control range are respectively determined. Finally, according to the first steady-state statistic, the second steady-state statistic, the first dynamic statistic, and the second dynamic statistic, the operating state of the current operating condition is monitored.

[0060] Please refer to Figure 1 , Figure 1 A schematic flow chart of a method for monitoring operating conditions based on exponential slow features according to some embodiments of the present invention is shown.

[0061] like Figure 1 As shown, the operating condition monitoring system may first execute step S1: using a clustering algorithm to divide different operating states under the drive of conditional variables.

[0062] Industrial processes in actual scenarios will switch between different operating modes, resulting in different working conditions, such as high energy consumption conditions, low energy consumption conditions, and unstable transition conditions. Clustering algorithms under the joint indication of multiple conditional driving variables can effectively identify and classify different working conditions: first, combine feature screening strategies and plant mechanism experience to find conditional driving variables that affect energy consumption levels, and cluster them under the joint indication of multiple conditional driving variables to divide working conditions. The working condition monitoring system can combine feature screening strategies and plant mechanism experience to find conditional driving variables that affect energy consumption levels, and adjust the operating state of the device by changing the four slow-changing input variables that determine the processing mode, namely, the fuel gas flow rate of the atmospheric pressure furnace and the reduced pressure furnace, the high-pressure gas pressure, and the relative density of the fuel gas.

[0063] Specifically, the condition monitoring system can use XGBoost for feature screening and summarize the key variables of each indicator model, namely, condition-driven variables, based on the plant mechanism experience. Here, XGBoost is an integrated machine learning algorithm based on decision trees, which can be used for data classification and variable regression prediction. In this paper, the feature importance of different performance indicators is evaluated and calculated by using its regression prediction characteristics.

[0064] In the process of using XGBoost for regression prediction, the algorithm processes variables based on a tree model. For a specific data set, during the tree model construction process, the system greedily selects the optimal feature split point at each layer to create leaf nodes, with the goal of maximizing the overall gain of the entire tree. In addition, when a feature is split, its average gain increases, indicating that the importance of the feature has increased. In this splitting mechanism, the weight of each leaf node is w(g i ,h i ), where g i and h i As shown in formula (1) and formula (2) respectively.

[0065]

[0066] Among them, the training error Represents the target value y i and predicted values The difference. The gain Gain after traversing each feature as the segmentation point makes the cost of the tree model after segmentation minimum. The gain formula is as shown in formula (3).

[0067]

[0068] For each feature, XGBoost will evaluate its importance in the feature set based on the importance metric. Therefore, the tree model in XGBoost sets up the feature splitting times FScore, feature average gain AverageGain, and feature average coverage AverageCover, as shown in formula (4), formula (5), and formula (6), respectively.

[0069] FScore=|X| (4)

[0070]

[0071]

[0072] Where X is the set of leaf nodes classified by the desired features; Gain is obtained by formula (3); Cover is the number of samples of data X contained in each node.

[0073] After that, a clustering algorithm is used to divide different working conditions. The original density-based clustering algorithm for analyzing and dividing conditional driving variables is changed to the DBSCAN algorithm, which analyzes and divides multiple modes under conditional driving variables. In order to identify clusters in the data set, the DBSCAN algorithm randomly selects a core point at the beginning and merges all data points that are directly or indirectly density-reachable from this point into one cluster. Then, the algorithm randomly selects another core point from the ungrouped data points, and similarly merges all data points reachable from this point into another cluster. This process continues until there are no core points to choose from. In addition, data points that cannot be connected to any core point through density reachability are regarded as abnormal points or outliers by DBSCAN. After being processed by the DBSCAN clustering algorithm, the operating state can be divided into a steady state and different transition states.

[0074] Please refer to Figure 2A to Figure 2D and Figure 3 , Figure 2A A variable schematic diagram of the fuel gas flow rate and energy consumption of an atmospheric pressure furnace provided according to some embodiments of the present invention is shown. Figure 2B A variable schematic diagram of the fuel gas flow rate and energy consumption of a pressure reducing furnace provided according to some embodiments of the present invention is shown. Figure 2C A variable schematic diagram of fuel gas relative density and energy consumption provided according to some embodiments of the present invention is shown. Figure 2D A variable schematic diagram of system high-pressure gas pressure and energy consumption provided according to some embodiments of the present invention is shown. Figure 3 A schematic diagram of operating condition division based on DBSCAN provided according to some embodiments of the present invention is shown.

[0075] like Figure 2A to Figure 2D and Figure 3 As shown in Figure 2, the non-stationary state of conditional driving variables and energy consumption over a period of time. In addition, the density-based clustering algorithm is used to cluster the working conditions. The clustering based on the joint indication of multiple conditional driving variables can effectively divide the transition state and different steady states, achieving the following Figure 3 The operation mode of the process data shown in the figure is divided, and the silhouette coefficient is introduced as an indicator to evaluate the quality of clustering. The calculation formula is as follows:

[0076]

[0077] Among them, i is the sample vector after dimensionality reduction, a(i) is the distance from i to all other points in the cluster to which it belongs, and b(i) is the average distance from i to all points in a cluster that does not contain it.

[0078] The average of all sample silhouette coefficients is the total silhouette coefficient of the clustering result. Its value range is [-1,1]. The closer it is to 1, the better the clustering effect.

[0079] Generally speaking, the average of all sample silhouette coefficients is the total silhouette coefficient of the clustering result, and its value range is [-1,1]. Here, a silhouette coefficient greater than 0.5 can be considered a good clustering effect, and the closer it is to 1, the better the clustering effect. Conversely, a silhouette coefficient less than 0.5 means that the clustering effect is poor. At this time, you can adjust the parameters, use PCA dimensionality reduction technology to reduce the dimension of the data, or change other clustering algorithms.

[0080] By using the DBSCAN algorithm to analyze and divide different working conditions, the non-convex and irregular characteristics of the joint probability density distribution of conditional driving variables are effectively solved, and the interference factors are also shielded. In addition, due to the lack of prior knowledge of the working condition changes contained in the data set, the commonly used center-based clustering algorithm cannot effectively solve this problem. Instead, using density-based clustering can effectively identify effective working conditions for irregularly distributed data sets.

[0081] Afterwards, the operating condition monitoring system may execute step S2: performing data standardization on the online data using the mean and variance of the offline modeling training data.

[0082] Specifically, the working condition monitoring system uses the mean and / or variance of the training data of the offline modeling to normalize the online data, and the formula is as follows:

[0083]

[0084] Among them, mean(X) represents the mean of offline modeling training data, std(X) represents the variance of offline modeling training data, and x new represents online data, x' new Represents normalized online data.

[0085] Afterwards, the operating condition monitoring system may execute step S3: using an exponential slow feature analysis method to divide the monitoring space into two different monitoring subspaces with slower changes and faster changes, and extracting slow features and fast features.

[0086] Specifically, the data format of the covariance matrix B and the first-order derivative covariance matrix A is as follows:

[0087]

[0088]

[0089] in, represents the variance of the nth eigenvector in the input feature, σ n,1 represents the covariance of two columns of different eigenvectors, represents the variance of the derivative of the nth eigenvector in the input feature, represents the covariance of the derivatives of two different eigenvectors.

[0090] For the exponential slow feature analysis method, the covariance matrix B of the input feature x(t) and the first-order derivative covariance matrix A need to be mapped to a new space:

[0091]

[0092] A→Φ(A)=exp(A)=∑ A (12)

[0093] A→Φ(B)=exp(B)=∑ B (13)

[0094] Specifically, taking the first-order derivative covariance matrix A as an example, its exponential form can be calculated as:

[0095]

[0096] In addition, if A or B involves large numbers, the calculation accuracy may be affected. Therefore, a scaling parameter θ should be set to map all elements in A and B into a small range. In the present invention, by setting θ to 1 / max{max|A|,max|B|}, all elements are mapped to [-1,1], where max|A| represents the element with the largest absolute value in A.

[0097] ESFA is equivalent to transforming the original data into a new space through exponential mapping, and then applying the SFA criterion in the new space. The principle is as follows:

[0098] Given an m-dimensional input signal x(t) = [x1(t), x2(t), x3(t), …, x m (t)] T , where t∈[t0,t1] represents time, the basis of SFA operation is a transformation g(x)=[g1(x),…,g j (x)] T , which maps the input data x to the output signal s(t)=[s1(t),…,s j (t)] T , and minimize the square of the time derivative of the output signal, that is, where s j (t) = g j (x(t)). The objective function is as follows:

[0099]

[0100] At the same time, the constraint function must be satisfied:

[0101] j > t =0 (16)

[0102]

[0103] Among them, <·> t Represents the average along the time axis. <s> t It can be expressed as

[0104]

[0105] This requires that the output signal has zero mean, unit variance, and that the different output components are orthogonal to each other. These conditions ensure the uniqueness of the transformation and the validity of the solution. Through this transformation, SFA is able to simplify the complex dynamics of the input data into a few key features that represent the long-term behavior of the system. Here, s j (t) is the slow feature of the input data x(t).

[0106] Assume that the mapping function is linear, that is, it satisfies ω i is the weight of the ith function. From formula (16), we can get

[0107]

[0108] The objective function of SFA can be expressed as follows, corresponding to equation (15):

[0109]

[0110] in is the first derivative of x.

[0111] Combining (17) and (18) we get

[0112]

[0113] in

[0114]

[0115] In this way, the optimization problem of SFA can be expressed as:

[0116]

[0117] To calculate these slow features, SFA usually applies singular value decomposition (SVD) to find the optimal transformation matrix W. This matrix W is a mapping from the input space to the slow feature space. SFs can be expressed as

[0118] s=Wx (25)

[0119] Where W = [ω1,ω2,…,ω m ] T is the weight matrix. The process of solving SFs is to solve the matrix W.

[0120] Perform singular value decomposition on the covariance matrix B of x(t).

[0121] B=X T X=UΛU T (26)

[0122] After whitening, x(t) can be obtained

[0123] z=Λ -1 / 2 U T X=QX (27)

[0124] Where Q = Λ -1 / 2 U T is the whitening matrix. To find the mapping direction that minimizes the slowness, Perform SVD decomposition:

[0125]

[0126] The projection matrix is ​​derived:

[0127] W=Q T P=UΛ -1 / 2 P (29)

[0128] From the obtained W matrix, the slow feature s(t) can be further obtained:

[0129] s=P T z=P T Λ -1 / 2 U T x=W T x (30)

[0130] In addition, the slowness of s(t) is the singular value after SVD decomposition:

[0131]

[0132] ESFA is to replace A and B in formula (24) to make it exponential form Σ A ,Σ B , as follows:

[0133]

[0134] Equation (32) and Equation (33) are an optimization problem, which can be solved by the Lagrange multiplier method.

[0135] (Σ B ) -1 ∑ A w i =λ i w i (34)

[0136] The order of slowness can be assigned according to the difference value Δ(·) of the slowness feature obtained. Δ(·) is the difference of the slowness feature, which can be obtained using the formula Specifically, among the extracted features, the one with the smallest Δ(·) is indexed as the slowest feature, and among the extracted features, the one with the largest Δ(·) is indexed as the fastest feature. The monitoring space is divided into two different monitoring subspaces with slower and faster changes:

[0137] s d =W d X T (35)

[0138] s e =W e X T (36)

[0139] s d For the derived slower feature, s e For the faster feature derivation, W is the corresponding projection matrix solved.

[0140] Thereafter, the operating condition monitoring system may execute step S4: calculating steady-state online monitoring statistics based on the slower characteristics and the faster characteristics.

[0141] Specifically, according to the slower feature and the faster feature obtained in step S3, in order to measure the steady-state change of the feature, d and e Two T 2 Statistics, the formula is as follows:

[0142]

[0143] The above two statistics describe the steady-state changes during the operation of the atmospheric and vacuum distillation unit. and Exceeding a threshold determined by a certain significance level α. For example, the significance level is set to 95%. If the result of the statistical test exceeds this critical value, it means that the steady state of the operating state has changed.

[0144] Here, the present invention takes the energy consumption of the atmospheric and vacuum distillation device as the performance index research object. The atmospheric and vacuum distillation device, whose full name is atmospheric and vacuum distillation device, is an important equipment in the oil refining process, mainly used for the preliminary processing of crude oil, separating crude oil into different components, and providing raw materials for subsequent deep processing. The principle of atmospheric distillation is that crude oil is heated to a certain temperature under normal pressure, so that the components with lower boiling points (such as gasoline and kerosene) evaporate, and then cool and separate through a fractionating tower. The principle of vacuum distillation is that for heavy oil components with higher boiling points, it is necessary to heat them under reduced pressure to reduce their boiling points, evaporate them and separate them through a fractionating tower. The present invention mainly relates to the fractionation processing link of the atmospheric and vacuum distillation device.

[0145] Thereafter, the operating condition monitoring system may execute step S5: calculating dynamic online monitoring statistics based on the slower feature and the faster feature.

[0146] According to the slower and faster features obtained in step S3, the first-order derivative of the slower feature in the feature space is calculated. In this way, a dynamic monitoring index is established to monitor its dynamic process. The formula is as follows:

[0147]

[0148] in, and The first-order derivative representing the slow feature in the feature space can be calculated by the following formula:

[0149]

[0150] Based on the first-order slow feature in the above formula, the dynamic changes of the operating data can be evaluated. In order to measure the dynamic abnormality of the process, and Based on this, two statistics are defined as follows:

[0151]

[0152] The above two statistics describe the distribution of dynamic changes in operating data. The control range of the two statistics can be calculated using kernel density estimation.

[0153] First, the probability density function is calculated.

[0154]

[0155] Among them, x is the input sample value, that is, each statistic, x i represents the sample data points, which are the points in the data set used to calculate the density, n is the number of samples, h is the window bandwidth parameter, which is set to 0.5, and K(·) is the Gaussian kernel function.

[0156]

[0157] Iterate over the y data points of the probability density function and find the first point y[i]>α that exceeds the confidence level.

[0159] The control limit (i.e. threshold) is calculated by the following formula:

[0160] kzx=x[i-1]+(x[i]-x[i-1])*(y[i]-α) / (y[i]-y[i-1]) (47)

[0161] Among them, kzx represents the control limit.

[0162] Here, the dynamic monitoring statistics are used to evaluate the trend of changes in the process. When these statistics remain within the calculated control limits, it indicates that the production process of the device is in a normal control state and the process performance is stable. Once any of them or Exceeding the threshold means that potential abnormalities in the dynamic process are monitored, indicating that the process may be affected by external faults or changes in environmental conditions. At this time, it is necessary to focus on the operating conditions in this stage and its small range.

[0163] Please refer to Figure 4 , Figure 4 A schematic diagram of an online fine-scale monitoring solution based on the ESFA method provided according to some embodiments of the present invention is shown.

[0164] like Figure 4 As shown, "yes" means that both statistics are within the control range, and "no" means that both indices exceed the corresponding confidence range.

[0165] Case 1: If the four statistics Specifically, in response to the first steady-state statistic being less than or equal to the first threshold, the second steady-state statistic being less than or equal to the second threshold, the first dynamic statistic being within the first control range, and the second dynamic statistic being within the second control range, it is determined that the current operating condition is operating normally.

[0166] Case 2: When dynamic monitoring statistics and All are within the limit, steady-state monitoring statistics or Specifically, in response to the first steady-state statistic being less than or equal to the first threshold, the second steady-state statistic being less than or equal to the second threshold, and at least one of the first dynamic statistic and the second dynamic statistic exceeding the corresponding control range, it is determined that the current operating condition has an operating condition drift phenomenon.

[0167] Case 3: When dynamic monitoring statistics or If the steady-state monitoring statistics and If both are within the limit, it means that the current operating condition of the atmospheric and vacuum system can adapt to the changes in the current environment, but considering that dynamic changes may cause abnormalities in the steady-state operation, it is necessary to keep a continuous eye on the subsequent operating state of the atmospheric and vacuum system. Specifically, in response to at least one of the first steady-state statistic and the second steady-state statistic being greater than the corresponding threshold, the first dynamic statistic being within the first control range, and the second dynamic statistic being within the second control range, the current operating condition is determined.

[0168] Case 4: When dynamic monitoring statistics or Over-limit, steady-state monitoring statistics or When the limit is exceeded, it indicates that the current operating condition of the atmospheric and vacuum pump cannot adapt to the changes in the current environment, and the working parameters should be adjusted. Specifically, in response to at least one of the first steady-state statistic and the second steady-state statistic being greater than the corresponding threshold, and at least one of the first dynamic statistic and the second dynamic statistic exceeding the corresponding control range, it is determined that the current operating condition cannot be maintained.

[0169] In addition, those skilled in the art can use the data obtained from the atmospheric and vacuum production process based on the above-mentioned working condition monitoring method and system based on the exponential slow feature to demonstrate the monitoring performance of the proposed method in steady-state and dynamic processes, and compare it with SEF, ICA and PCA.

[0170] Please refer to Figure 5A to Figure 5D , Figure 5A A schematic diagram showing monitoring results of the ESFA method provided according to some embodiments of the present invention is shown. Figure 5B A schematic diagram of monitoring results of the SFA method provided according to some embodiments of the present invention is shown. Figure 5C A schematic diagram of monitoring results of the PCA method provided according to some embodiments of the present invention is shown. Figure 5D A schematic diagram of monitoring results of the ICA method provided according to some embodiments of the present invention is shown.

[0171] like Figure 5A to Figure 5D As shown, the training data is sampled at intervals of 10 minutes. Some monitoring results are shown in Figure 5A to Figure 5D In the sampled data, the operating condition fluctuation began at the moment of the 106th sample point, the gas volume changed significantly, and both the dynamic and static indicators of the process were affected. After a period of transition adjustment, the processing process switched to a new processing mode.

[0172] according to Figure 5C and 5D The SPE statistical indicators calculated by the PCA and ICA algorithms exceeded the control limit after the gas volume changed, indicating that the current process deviated from the original operating state. After a period of time, it temporarily recovered to the control limit, indicating that the model temporarily adapted to this transition state. After switching to the new operating condition, the SPE indicator continued to exceed the control limit, and a large number of false alarms occurred. The model did not effectively identify the normal operating condition switching, which was inconsistent with the actual situation.

[0173] For the SFA algorithm, Figure 5B Medium static indicators and All of them exceeded the control limit at the 107th sampling time, and the process change was detected after a delay of 10 minutes. However, after the working condition was switched, the dynamic indicators and There was also a large-scale over-limit phenomenon, which caused a large number of false alarms.

[0174] Figure 5A The monitoring results of ESFA are shown. and The indicator exceeded the control limit after 106 sample points, which timely reflected the abnormality of the process. After the working conditions were switched, the slow characteristic characteristics were further amplified on the basis of SFA, which could better capture the overall trend of atmospheric and vacuum processing production. and It is continuously within the control limit, indicating that this method effectively characterizes the normal switching process of the working condition.

[0175] Please refer to Tables 1 and 2 for further information. Table 1 shows the fault detection rate and false alarm rate of each method in the experiment, and Table 2 shows the sampling time at which these methods detect the beginning of the fault in the experiment.

[0176] Table 1 Fault detection rate and false alarm rate of each method

[0177] Algorithm / operation status False Positive Rate This program method 5.45% SFA 20.46% ICA 33.25% PCA 32.14%

[0178] Table 2 Sampling time of fault onset detected by each method

[0179] algorithm Sampling time This program method 106 SFA 107 ICA 107 PCA 108

[0180] As shown in Tables 1 and 2, combining the information in the two tables, we can see that the false alarm rate and the time to detect faults of the proposed method are the lowest.

[0181] In summary, the operating condition monitoring method based on exponential slow features proposed in the present invention takes the energy consumption of the constant pressure reduction device as the performance indicator research object. By combining the feature screening strategy and the plant mechanism experience, the conditional driving variables that affect the energy consumption level are found, and based on the common indication of these multiple conditional driving variables, the process data is clustered, so that the data can be divided according to the operating mode. In addition, the present invention also combines an improved slow feature analysis method, namely the exponential slow feature analysis method, to monitor the dynamic and static characteristics of the process. This method can extract time-varying features more sensitively. Through the comparative verification of multiple algorithms, it is proved that this method can not only effectively avoid false alarms caused by changes in operating conditions, but also accurately identify operating deviations and process anomalies in new operating conditions.

[0182] Although the above methods are illustrated and described as a series of actions for simplicity of explanation, it should be understood and appreciated that these methods are not limited by the order of the actions, because according to one or more embodiments, some actions may occur in a different order and / or concurrently with other actions from those illustrated and described herein or not illustrated and described herein but understandable to those skilled in the art.

[0183] The previous description of the disclosure is provided to enable any person skilled in the art to make or use the disclosure. Various modifications to the disclosure will be apparent to those skilled in the art, and the general principles defined herein may be applied to other variations without departing from the spirit or scope of the disclosure. Thus, the disclosure is not intended to be limited to the examples and designs described herein, but should be granted the widest scope consistent with the principles and novel features disclosed herein.< / s> ​< / s>

Claims

1. A working condition monitoring method based on exponential slow features, characterized in that: The following steps are involved: Determine a plurality of conditional driving variables that affect energy consumption levels, and perform clustering processing on the plurality of conditional driving variables to divide and determine high energy consumption operating conditions, low energy consumption operating conditions, and transitional operating conditions; The exponential slow feature method is used to divide the monitoring space of the online operation data of the atmospheric and vacuum distillation unit into a slow-changing feature space and a fast-changing feature space, so as to extract the slow-changing features and the fast-changing features respectively. Establishing a first steady-state statistic based on the slowly varying feature and a second steady-state statistic based on the rapidly varying feature, respectively, and determining a first threshold and a second threshold of their significance levels, respectively; Establishing a first dynamic statistic based on the first-order derivative of the slowly varying feature and a second dynamic statistic based on the first-order derivative of the quickly varying feature, and determining a first control range and a second control range thereof respectively; and The operating state of the current operating condition is monitored according to the first steady-state statistic, the second steady-state statistic, the first dynamic statistic, and the second dynamic statistic.

2. The working condition monitoring method according to claim 1, characterized in that: The step of determining a plurality of condition-driven variables that affect the energy consumption level comprises: Acquire multiple candidate condition-driven variables involved in the process to be monitored, and their corresponding sample data sets; and Using the XGBoost model, each of the candidate condition-driven variables in the sample data set and its corresponding sample data is screened by a first variable based on regression prediction, and / or based on the plant mechanism experience, a second variable screening of key variables is performed on multiple indicator models of the process flow to determine the multiple condition-driven variables that affect the energy consumption level.

3. The working condition monitoring method according to claim 2, characterized in that: The step of clustering the multiple conditional driving variables to divide and determine the high energy consumption working condition, the low energy consumption working condition and the transition working condition comprises: Using the DBSCAN algorithm, cluster analysis is performed on the multiple modes under each of the conditional driving variables to divide the multiple sample data in the sample data set into multiple clusters; and According to the operating states corresponding to the clusters, they are respectively defined as the high energy consumption operating condition, the low energy consumption operating condition and the transitional operating condition.

4. The working condition monitoring method according to claim 1, characterized in that: Before dividing the monitoring space of the online operation data of the atmospheric and vacuum distillation device into a slowly changing feature space and a quickly changing feature space, the operating condition monitoring method further includes the following steps: Calculating the mean and / or variance of each of the sample data in the sample data set; and According to the mean value and / or the variance, the online operation data of the atmospheric and vacuum distillation device is standardized: Wherein, mean(X) represents the mean of each sample data in the sample data set, std(X) represents the variance of each sample data in the sample data set, and x new represents the online operation data before normalization, x′ new Represents standardized online operation data.

5. The working condition monitoring method according to claim 1, characterized in that: The step of using the exponential slow feature method to divide the monitoring space of the online operation data of the atmospheric and vacuum distillation device into a slow-changing feature space and a fast-changing feature space to extract the slow-changing features and the fast-changing features respectively includes: Construct the covariance matrix B and the first-order derivative covariance matrix A of the input features composed of the multiple conditional driving variables: in, represents the variance of the nth eigenvector in the input feature, σ n,1 represents the covariance of two columns of different eigenvectors, represents the variance of the derivative of the nth eigenvector in the input feature, represents the covariance of the derivatives of two different eigenvectors; and All elements of the covariance matrix B and the first-order derivative covariance matrix A are mapped to the range of [-1,1], and an SFA operation is performed in the corresponding reduced space to divide the monitoring space of the online operation data of the constant pressure reduction device into a slow-varying feature space and a fast-varying feature space, and the slow-varying features and the fast-varying features are extracted respectively.

6. The working condition monitoring method according to claim 5, characterized in that: The step of performing the SFA operation in the corresponding reduced space includes: For a given m-dimensional input signal x(t) = [x1(t), x2(t), x3(t), …, x m (t)] T , through the linear mapping matrix g(x) = [g1(x),…,g j (x)] T Map the m-dimensional input signal to the corresponding output signal s(t)=[s1(t),…,s j (t)] T , and minimizes the square of the time derivative of the output signal Where t∈[t0,t1] represents time, s j (t) = g j (x(t)), 〈·〉 t represents the average value along the time axis, <s> t express ω i is the i-th mapping function g i The weight of (x), < / s> <s> Apply singular value decomposition to perform singular value decomposition on the covariance matrix B: B=X T X=UΛU T Wherein, the m-dimensional input signal x(t) is whitened to obtain z=Λ -1 / 2 U T X=QX,Q=Λ -1 / 2 U T is the whitening matrix, According to the result of the singular value decomposition, the projection matrix W is determined: W=Q T P=UΛ -1 / 2 P; and According to the projection matrix W, the monitoring space of the online operation data of the atmospheric and vacuum distillation device is divided into the slowly changing feature space and the quickly changing feature space: s d =W d X T s e =W e X T Among them, s d is the slow-varying feature, s e is the fast-changing feature.

7. The working condition monitoring method according to claim 6, characterized in that: The first steady-state statistic is expressed as The second steady-state statistic is expressed as 8. The working condition monitoring method according to claim 7, characterized in that: The first-order derivative of the slowly varying characteristic is expressed as The first dynamic statistic is expressed as The first-order derivative of the fast-varying characteristic is expressed as The second dynamic statistic is expressed as 9. The working condition monitoring method according to claim 8, characterized in that: The steps of respectively determining the first control range and the second control range include: Determine the probability density function of the kernel density estimate: Wherein, x represents the first dynamic statistic or the second dynamic statistic, x i represents the sample data point, n is the number of samples, h is the window bandwidth parameter, and K(·) is the Gaussian kernel function: Traversing each y data point of the probability density function to determine the first point y[i] that exceeds a preset confidence level α; According to the point y[i], the control range of the corresponding dynamic statistics is calculated: kzx=x[i-1]+(x[i]-x[i-1])*(y[i]-α) / (y[i]-y[i-1]) Among them, kzx represents the control limit.

10. The working condition monitoring method according to claim 1, characterized in that: The step of monitoring the operating state of the current working condition according to the first steady-state statistic, the second steady-state statistic, the first dynamic statistic, and the second dynamic statistic comprises: In response to the first steady-state statistic being less than or equal to the first threshold, the second steady-state statistic being less than or equal to the second threshold, the first dynamic statistic being within the first control range, and the second dynamic statistic being within the second control range, determining that the current operating condition is operating normally; In response to the first steady-state statistic being less than or equal to the first threshold, the second steady-state statistic being less than or equal to the second threshold, and at least one of the first dynamic statistic and the second dynamic statistic exceeding a corresponding control range, determining that an operating condition drift phenomenon occurs in a current operating condition; In response to at least one of the first steady-state statistic and the second steady-state statistic being greater than a corresponding threshold, the first dynamic statistic being within the first control range, and the second dynamic statistic being within the second control range, determining a current operating condition; and In response to at least one of the first steady-state statistic and the second steady-state statistic being greater than a corresponding threshold, and at least one of the first dynamic statistic and the second dynamic statistic exceeding a corresponding control range, it is determined that the current operating condition cannot be maintained.

11. A working condition monitoring system based on exponential slow features, characterized in that: include: a memory having computer instructions stored thereon; as well as A processor is connected to the memory and is configured to execute computer instructions stored in the memory to implement the working condition monitoring method according to any one of claims 1 to 10.

12. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the computer instructions are executed by a processor, the operating condition monitoring method according to any one of claims 1 to 10 is implemented. < / s>