阅读说明：本技术 面向高端燃煤发电装备的变量协同动态报警阈值优化方法 (Variable collaborative dynamic alarm threshold optimization method for high-end coal-fired power generation equipment ) 是由 赵春晖 赵诣 李清毅 王文海 于 2021-09-07 设计创作，主要内容包括：本发明公开了一种面向高端燃煤发电装备的变量协同动态报警阈值优化方法。本发明针对高端燃煤发电过程中的磨煤机设备,考虑了非平稳变工况运行下的变量协同动态报警阈值优化。借助时序分割有序聚类方法将非平稳过程的历史正常数据划分至若干工况(包括稳态、过渡过程),每种工况用一个多元高斯分布模型进行数学表征。在线时,根据时序分割有序聚类方法将当前数据分类到所属工况,选取当前工况的多元高斯模型,从而根据条件概率进行阈值设计。该方法可满足全工况下的准确报警,对保证过程安全运行和人员安全有重大意义。(The invention discloses a variable collaborative dynamic alarm threshold optimization method for high-end coal-fired power generation equipment. The invention considers the variable collaborative dynamic alarm threshold optimization under the non-steady variable working condition operation aiming at the coal mill equipment in the high-end coal-fired power generation process. The historical normal data of the non-stationary process is divided into a plurality of working conditions (including a stationary state and a transition process) by means of a time sequence segmentation and ordered clustering method, and each working condition is mathematically represented by a multivariate Gaussian distribution model. And during online, classifying the current data to the working condition according to a sequential segmentation and sequential clustering method, and selecting a multivariate Gaussian model of the current working condition so as to design a threshold value according to the conditional probability. The method can meet the requirement of accurate alarm under the full working condition, and has great significance for ensuring the safe operation of the process and the safety of personnel.)

1. A variable collaborative dynamic alarm threshold optimization method for high-end coal-fired power generation equipment is characterized by comprising the following steps:

(1) obtaining non-stationary process data: acquiring T (T is 1, 2,.. T) continuous sampling data of J process variables during normal operation of high-end coal-fired power generation equipment, wherein a two-dimensional data matrix of the T (T × J) continuous sampling data is represented as X (T × J);

(2) clustering the non-stationary process data obtained in the step (1) by utilizing a time sequence segmentation and ordered clustering method to divide the working condition classes, and obtaining the mathematical representation of the multivariate Gaussian distribution of each working condition classWherein, mu_iExpressed as the mean, Σ, of the i-th operating condition class_iExpressed as covariance for the ith condition class.

(3) Acquiring new process variable data x of high-end coal-fired power generation equipment in real time during operation_new(1 XJ) for x_new(1 XJ) sequence expansion is carried out to obtain { x_new-ω+1，…，x_newω is the order of timing expansion; distributing the current new process variable data to a certain working condition category i clustered in the step (2) through a step E of a sequential segmentation and ordered clustering method;

(4) based on the multivariate Gaussian distribution corresponding to the working condition categories, variable collaborative threshold design is carried out on the process variable data after the time sequence expansion, and the method specifically comprises the following steps: representing the set of process variables in the J omega dimension as the variable x currently being thresholded¹And x is removed¹X of all process variables other than²Combination ofWherein x¹The dimension is 1; x is the number of²The dimension is J omega-1;

the mean vector is correspondingly split into:

the covariance matrix is correspondingly split into:

therein, sigma¹¹Represents the variable x¹And variable x¹Of (a) covariance matrix, ∑¹²Represents the variable x¹With other variables x²Of (a) covariance matrix, ∑²¹Representing other variables x²And variable x¹Of (a) covariance matrix, ∑¹¹Representing other variables x²With other variables x²The covariance matrix of (2).

Variable x¹In other variables x²When values are given, the condition Gaussian distribution is presented:

wherein the conditional mean is:

the conditional variance is:

determining a variable x from a conditional Gaussian distribution¹And (3) designing variable cooperative threshold values for each variable one by one to obtain the alarm threshold values of all the variables.

2. The variable collaborative dynamic alarm threshold optimization method for high-end coal-fired power generation equipment according to claim 1, wherein in the step 1, measuring the variable comprises: coal feeding amount, coal mill current, rotary separator current, outlet pressure, inlet primary air temperature, inlet primary air pressure, sealing air pressure and outlet air-powder mixture temperature.

3. The variable collaborative dynamic alarm threshold optimization method for high-end coal-fired power generation equipment according to claim 1, wherein the step 2 is implemented by the following substeps:

(2.1) parameter selection: determining the category number I, the time sequence expansion order omega, a norm regularization parameter lambda and a time sequence consistency constraint coefficient beta in advance by combining Bayesian Information Criterion (BIC) indexes and mechanisms;

(2.2) clustering: expanding the order omega according to the selected time sequence, each sample x in the non-stationary process data_tIs time-sequence expanded to { x_t-ω+1，…，x_tForm (E) }; the two-dimensional data matrix of non-stationary process data is denoted as X (T × J ω); clustering is carried out according to a time sequence segmentation and ordered clustering method, and a multivariate Gaussian distribution representation corresponding to I (I is 1, 2

4. Variable collaborative dynamic reporting for high-end coal-fired power generation equipment according to claim 1The alarm threshold optimization method is characterized in that in the step 4, the variable x is determined according to the conditional Gaussian distribution¹The alarm threshold value specifically is as follows:

if the confidence is 95%, x¹High alarm threshold of_1|2+2∑_1|2Low alarm threshold is mu_1|2-2∑_1|2。

Technical Field

The invention belongs to the field of alarm management in a non-stable high-end coal-fired power generation process, and particularly relates to a variable collaborative dynamic alarm threshold optimization method.

Background

The alarm system plays an important role in safe and efficient operation of modern industrial processes. Such as Distributed Control Systems (DCS) and supervisory control and data acquisition Systems (SCADA), compare some important process variables in real time with their configured high (low) alarm thresholds and trigger an alarm to notify operators if the variables exceed the thresholds.

Generally, the alarm threshold is determined according to factory protection fixed values or carefully during debugging, and if the alarm threshold is not properly set, false alarm and false alarm can occur, so that the performance of the whole alarm system is concerned. However, after an industrial process is put into production operation, regular, large-scale maintenance and modification of alarm thresholds is lacking. At present, the common problem of the existing alarm system is that the alarm is inundated, and the generated alarm is divided into an interference alarm and a correct alarm. First, a large number of alarms fall into the category of nuisance alarms, which do not provide any useful information and are merely distracting to the operator. Secondly, even if all alarms are correct, the operator cannot process a large amount of alarm information in a short time, and can only ignore many alarms that have occurred. When a serious disaster really occurs, the fault source is probably buried in a large number of alarms and cannot be positioned and responded in time.

Aiming at the problem of alarm flooding of the alarm system, the academic world provides a plurality of threshold optimization design methods based on mathematical statistics, such as the most classical three-time standard deviation rule. Another mainstream threshold design idea is to balance the false alarm rate and the false alarm rate, and the threshold design at this time is actually an optimization problem. However, the threshold design methods are all from a univariate perspective, and the inherent coupling of the industrial process is ignored. The characteristic is not ignored when the alarm threshold value is designed, otherwise, the false alarm rate and the missing alarm rate are easily caused. Therefore, from the perspective of variable cooperation, a scholars respectively calculate the correlation between two or more process variables and alarm data, and set a threshold value to make the difference as small as possible, which indicates that the alarm condition relatively reflects the process characteristic, but cannot dynamically optimize the threshold value according to different working conditions.

Disclosure of Invention

The invention aims to provide a variable collaborative dynamic alarm threshold value optimization method for high-end coal-fired power generation equipment aiming at unreasonable design of the existing alarm threshold value. According to the method, the fact that a fixed threshold cannot adapt to different working conditions is considered, historical normal data of a non-stationary process are divided into a plurality of working conditions (including a steady state process and a transition process) by means of a time sequence segmentation and ordered clustering method, each working condition is represented by a multivariate Gaussian model, and each process variable under the working conditions estimates a dynamic threshold under the condition of taking values of other current variables by means of conditional probability distribution. And during online, classifying the current data to the working condition according to a sequential segmentation and sequential clustering method, and selecting a multivariate Gaussian model of the current working condition so as to design a threshold value according to the conditional probability. The invention provides a new analysis idea for threshold optimization work in a non-steady process, and the method can meet accurate alarm under all working conditions and has great significance for ensuring safe operation of the process and personnel safety.

The purpose of the invention is realized by the following technical scheme:

a variable collaborative dynamic alarm threshold optimization method for high-end coal-fired power generation equipment comprises the following steps:

(1) obtaining non-stationary process data: acquiring T (T is 1, 2,.. T) continuous sampling data of J process variables during normal operation of high-end coal-fired power generation equipment, wherein a two-dimensional data matrix of the T (T × J) continuous sampling data is represented as X (T × J);

(2) clustering the non-stationary process data obtained in the step (1) by utilizing a time sequence segmentation and ordered clustering method to divide the working condition classes, and obtaining the mathematical representation of the multivariate Gaussian distribution of each working condition classWherein, mu_iExpressed as the mean, Σ, of the i-th operating condition class_iExpressed as covariance for the ith condition class.

(3) Acquiring new process variable data x of high-end coal-fired power generation equipment in real time during operation_new(1 XJ) for x_new(1 XJ) sequence expansion is carried out to obtain { x_new-ω+1，…，x_newω is the order of timing expansion; and distributing the current new process variable data through the step E of the sequential segmentation and sequential clustering methodEntering a certain working condition category i clustered in the step (2);

(4) based on the multivariate Gaussian distribution corresponding to the working condition categories, variable collaborative threshold design is carried out on the process variable data after the time sequence expansion, and the method specifically comprises the following steps: representing the set of process variables in the J omega dimension as the variable x currently being thresholded¹And x is removed¹X of all process variables other than²Combination ofWherein x¹The dimension is 1; x is the number of²The dimension is J omega-1;

the mean vector is correspondingly split into:

the covariance matrix is correspondingly split into:

therein, sigma¹¹Represents the variable x¹And variable x¹Of (a) covariance matrix, ∑¹²Represents the variable x¹With other variables x²Of (a) covariance matrix, ∑²¹Representing other variables x²And variable x¹Of (a) covariance matrix, ∑¹¹Representing other variables x²With other variables x²The covariance matrix of (2).

Variable x¹In other variables x²When values are given, the condition Gaussian distribution is presented:

wherein the conditional mean is:

the conditional variance is:

determining a variable x from a conditional Gaussian distribution¹And (3) designing variable cooperative threshold values for each variable one by one to obtain the alarm threshold values of all the variables.

Further, in step 1, measuring the variable includes: coal feeding amount, coal mill current, rotary separator current, outlet pressure, inlet primary air temperature, inlet primary air pressure, sealing air pressure and outlet air-powder mixture temperature.

Further, the step 2 is realized by the following substeps:

(2.1) parameter selection: determining the category number I, the time sequence expansion order omega, a norm regularization parameter lambda and a time sequence consistency constraint coefficient beta in advance by combining Bayesian Information Criterion (BIC) indexes and mechanisms;

(2.2) clustering: expanding the order omega according to the selected time sequence, each sample x in the non-stationary process data_tIs time-sequence expanded to { x_t-ω+1，…，x_tForm (E) }; the two-dimensional data matrix of non-stationary process data is denoted as X (T × J ω); clustering is carried out according to a time sequence segmentation and ordered clustering method, and a multivariate Gaussian distribution representation corresponding to I (I is 1, 2

Further, in the step 4, the variable x is determined according to the conditional gaussian distribution¹The alarm threshold value specifically is as follows:

if the confidence is 95%, x¹High alarm threshold of_1|2+2∑_1|2Low alarm threshold is mu_1|2-2∑_1|2。

Compared with the prior art, the invention has the beneficial effects that: the invention provides a new research idea for the alarm variable threshold optimization technology of the high-end coal-fired power generation non-steady operation process data. The original non-stationary operation data is reconstructed into a plurality of stationary working condition sections based on the working condition indicating variables, probability statistics is carried out in each section, and the threshold value under the working condition is obtained. When the method is applied on line, the corresponding threshold value is matched according to the value of the working condition indicating variable, dynamic threshold value setting is realized according to the current working condition, and the method is favorable for reducing misinformation. The method is fully experimental research aiming at actual industrial data and is successfully applied.

Drawings

FIG. 1 is a flow chart of variable collaborative dynamic alarm threshold optimization according to the present invention.

Fig. 2 is a schematic diagram of the working condition modal division result and the corresponding threshold value according to the present invention.

FIG. 3 is a graph of the results of the application of fault data with threshold optimization in accordance with the present invention.

FIG. 4 is a graph of results of a conventional non-threshold optimized fault data application.

Detailed Description

The invention deeply considers the characteristics of multi-working-condition operation and variable coupling of high-end coal-fired power generation equipment, adopts an ordered segmentation and clustering method to divide working conditions, further realizes variable collaborative alarm threshold design through multivariate Gaussian distribution representation, and provides a variable collaborative alarm threshold optimization method facing the high-end coal-fired power generation equipment. The invention is further described with reference to the following drawings and specific examples.

The invention provides a single variable alarm threshold value optimization method for non-steady operation of high-end coal-fired power generation equipment, which comprises the following steps:

(1) obtaining non-stationary process data: in this example, about 24480 samples were collected for modeling, with 8 measured variables: coal feeding amount, coal mill current, rotary separator current, outlet pressure, inlet primary air temperature, inlet primary air pressure, sealing air pressure and outlet air-powder mixture temperature. 2000 samples are collected in the fault set, wherein the fault starting time is 674 th sample, and the fault set is used for checking the alarm performance of the threshold optimization method, wherein the fault is that the outlet temperature of the coal mill is low.

(2) The working condition is divided by using a time sequence segmentation and order Clustering method (Toeplitz Inverse Covariance-Based Clustering, TICC), and the step is realized by the following sub-steps:

(2.1) parameter selection: determining the category number 6, the time sequence expansion order omega of 3, the one-norm regularization parameter lambda of 0.11 and the time sequence consistency constraint coefficient beta of 15 in advance by combining Bayesian Information Criterion (BIC) indexes and mechanisms;

(2.2) clustering: expanding the order omega according to the selected time sequence, each sample x in the non-stationary process data_tAre all time-sequentially expanded into x_t-2，x_t-1，x_tForm (E) }; the two-dimensional data matrix of non-stationary process data is denoted as X (T × J ω); clustering is carried out according to a time sequence segmentation and ordered clustering method, and a multivariate Gaussian distribution representation corresponding to I (I is 1, 2And a category label i corresponding to each sample time t; here mu_iExpressed as the mean, Σ of the ith class_iCovariance expressed as ith class; therefore, different categories can be regarded as different working conditions of a non-stationary process, and the mathematical representation of the multivariate Gaussian distribution of the non-stationary process is obtained;

therefore, each process variable under the working condition can estimate the dynamic threshold under the current value of other variables by means of conditional probability distribution. The variable collaborative threshold design method specifically comprises the following steps: taking a certain sample time t as an example, selecting a corresponding multivariate Gaussian distribution model according to the class label iThen the process variable set in the J omega dimensionWherein x¹The dimension of the variable designed for the current threshold value is 1; x is the number of²Is divided by x¹All process variables except for the process variable, the dimension is J omega-1;

the mean vector is correspondingly split into:

the covariance matrix is correspondingly split into:

variable x¹In other variables x²When values are given, the condition Gaussian distribution is presented:

wherein the conditional mean is:

the conditional variance is:

further determining variable x according to conditional Gaussian distribution¹The alarm threshold value of (2) is, in this embodiment, 95% confidence level, x¹High alarm threshold of_1|2+2∑_1|2Low alarm threshold is mu_1|2-2∑_1|2(ii) a Performing variable collaborative threshold design on each variable one by one, namely taking the variable to be subjected to threshold design currently as x¹Remove x¹All process variables except for x²Determining a high alarm threshold value and a low alarm threshold value according to the current step;

(3) when the method is used online, the method specifically comprises the following steps:

(3.1) acquiring new sample data of the process variable and identifying the working condition: collecting new process variable data x_new(1 xj), wherein the subscript new represents a new sample; reference step(2.2) for x_new(1 XJ) sequence expansion is carried out to obtain { x_new-ω+1，…，x_newAllocating the current sample to a certain working condition type i through a step E of a sequential segmentation and ordered clustering method algorithm;

(3.2) alarm threshold selection: design method of x according to variable collaborative threshold_new-ω+1，…，x_newAndsubstituting the formulas (2) to (5), and calculating high alarm threshold values and low alarm threshold values of all process variables at the current moment one by one;

(3.3) judging the process running state on line: comparing the current values of the high and low alarm thresholds and the corresponding variables in real time; when any variable value exceeds the threshold range, the alarm information of the variable is triggered, and the field check is needed.

The following is the result of online monitoring of fault samples using the threshold optimization method of the present invention and the conventional method to further illustrate the effect of the present invention:

a section of normal data is selected and modeled by a sequential segmentation and ordered clustering method, the clustering result is shown as the attached figure 2, 6 types of working conditions and respective covered sample ranges are obtained, the horizontal coordinate in the figure is the sample time, and the vertical coordinate is the corresponding type of each time.

The results are shown in fig. 3, following the online application of a 2000 sample number of fault samples. As can be seen from fig. 3, the threshold is dynamically adjusted as the operating conditions switch, and the inherent coupling between variables is taken into account.

In contrast, the same section of fault sample is applied online with a non-optimized threshold, and the results are shown in fig. 4. According to calculation, during the fault period, the alarm rate of threshold optimization is 86.7%, and the alarm rate of non-threshold optimization is 92.9%; during normal operation, the false alarm rate with threshold optimization is 3.7%, and the alarm rate without threshold optimization is 37.5%. The method can flexibly and dynamically adjust the threshold value along with the switching of the working conditions, effectively reduce the false alarm rate and also have good fault detection performance.

Generally speaking, the time sequence segmentation-based ordered clustering and condition threshold calculation method provided by the invention considers the characteristic that different working conditions of a non-stationary process have great influence on data distribution, can dynamically select a proper working condition model, and calculates the condition threshold based on the multivariate Gaussian distribution representation of each working condition, thereby being beneficial to reducing false alarm and improving the performance of an alarm system.

It should be understood that the present invention is not limited to the coal pulverizer process in the high end coal-fired power plant of the above-described embodiment, and that equivalent modifications or substitutions may be made by those skilled in the art without departing from the spirit of the present invention, and such equivalent modifications or substitutions are intended to be included within the scope of the present invention as defined by the appended claims.