WO2024077983A1

WO2024077983A1 - Method for improving early-warning advance performance and accuracy of evaluation model

Info

Publication number: WO2024077983A1
Application number: PCT/CN2023/099140
Authority: WO
Inventors: 刘林; 崔保华; 李慧霞; 张成伟; 张爱民; 王磊; 李安平; 武艳文
Original assignee: 南京凯盛国际工程有限公司; 中国建材集团有限公司
Priority date: 2022-10-14
Filing date: 2023-06-08
Publication date: 2024-04-18
Also published as: CN115599646A

Abstract

Disclosed in the present invention is a method for improving the early-warning advance performance and accuracy of an evaluation model, comprising: screening for a historical production anomaly time set D of a modeling object j according to a field production anomaly record; a model capturing a data change rule within k time before a production anomaly; a health degree HPI variance and slope change trend of a model evaluation object within k time before the production anomaly reflecting an anomaly in advance; setting thresholds of a health degree HPI variance and a slope, and if the thresholds are exceeded, the model starting early-warning; obtaining variance and slope coincidence rate indexes pD_kvp and pD_ksp of historical production anomaly early-warning according to threshold parameters d_kvT and d_ksT; establishing a coincidence rate; solving for an HPI trend early-warning coincidence rate loss function by using a heuristic algorithm, and searching for suitable values of parameters k, d_kvT, and d_ksT to maximize the value of the loss function p; when the model runs, calculating in real time whether the variance d_kv of the health degree HPI in the k time period exceeds d_kvT, and if yes, prompting variance early-warning; and calculating in real time whether the slope d_ks of the health degree HPI in the k time period exceeds d_ksT, and if yes, prompting slope early-warning.

Description

A method to improve the early warning and accuracy of evaluation models

Technical Field

The present invention relates to a modeling method of an intelligent evaluation and diagnosis model, and in particular to a method for improving the early warning advanceness and accuracy of an evaluation model.

Background technique

Cement enterprises belong to the process industry of uninterrupted production. During the production process, hundreds of sensors collect key node data of each production link. At present, this part of data is mainly used by users for threshold range monitoring of a single measurement point and process production debugging. Both of them are based on point monitoring or control loop (line) monitoring, and cannot achieve overall (surface) status assessment.

Patent CN102270271B: "Method and system for early warning and optimization of equipment failure based on similarity curve", referred to as the evaluation system (IEM), uses the historical normal operation data of factory equipment measurement points and process measurement points to establish equipment models and process models to monitor equipment failures, early warnings or drift of operating conditions. The evaluation system (IEM) outputs the health (HPI) score of the modeled object in real time and compares it with the health HPI standard value. When the real-time health HPI continues to be lower than the standard HPI value for a period of time, an alarm is output. Among them, the health HPI value is a standard fixed value calculated based on the historical modeling data.

In the application of the above patented technology in the cement industry, the following problems have arisen:

First, the alarms obtained by comparing the HPI benchmark value of the model health often indicate that the modeled object is already in a state of abnormal production, so the needs of advance prediction, prejudgment, and early warning cannot be met.

Second, the evaluation system (IEM) has built-in rich alarm cause rules, most of which describe the data changes of key measurement points before production abnormalities. In actual applications, based on the first point, it was found that when the HPI warning was generated, no cause rules that met the conditions were reported.

Analysis of the causes of the above application problems:

First, the health HPI value calculated by the evaluation system (IEM) model is the expected value at a single time point, and refers to the health value corresponding to the similar state space in the historical data. Therefore, the health obtained by the evaluation model does not directly have the trend warning function. The general expert experience knowledge currently sorted out is matched based on the health HPI or the change trend of a certain measurement point in the recent period of time. So, in essence. The evaluation system (IEM) is a point-to-point classification model without time trend, while the actual rule matching is a multi-measurement point combination model with time trend. This is the most fundamental difference, which also makes the model unable to achieve early warning function.

Second, the model health index (HPI) is affected by the overall residuals of all measurement points, while the alarm rules are based on the residuals of specific measurement points, and the alarm is triggered after the combination of rules. Therefore, when the HPI generates an alarm, the combined measurement point alarm rules may not be triggered at the same time. Appear.

Therefore, in order to achieve early warning of faults and report more corresponding fault causes, the early warning method of the evaluation system IEM is improved.

Summary of the invention

In response to the problems existing in the prior art, the present invention provides a method for improving the advance warning and accuracy of the evaluation model, which can reflect the advance warning trend of faults, measure the advance warning trend and the overlap rate of historical faults, and solve the problem of high cost of manual correction of early warning accuracy.

The purpose of the present invention is achieved through the following technical solutions.

A method for improving the early warning and accuracy of an evaluation model, comprising:

1) The on-site production abnormality records include the start and end time of the production abnormality and the reason for the shutdown. The on-site production abnormality records are used as the source of the model negative samples. The data pattern before the real production abnormality is extracted by the sliding window length method. According to the on-site production abnormality records, the historical production abnormality time set D of the modeling object j is screened out, D = {d ₁ , d ₂ , ..., d _n }, where d _n = t _nbegin , d _n is the nth production abnormality time, and t _nbegin is the start time of the nth abnormal time period;

2) The model captures the data change rules within k time periods before the production anomaly: According to the abnormal time set D obtained in step 1), use the start time of each production anomaly _tnbegin minus k time periods to represent the advance of the production anomaly in the data pattern. The initial k value is determined according to the early warning advance requirements of the modeling object. Get the set of k time periods before the historical production anomaly _Dk , _Dk = { _dk1 , _dk2 , ..., _dkn }, where _dkn = ( _tnbegin -k, _tnbegin ), _dkn is the early warning advance time before each production anomaly, and is also the time when the data anomaly begins to exist;

3) The health HPI variance and slope change trends of the model evaluation object within k time before the production abnormality reflect the abnormality in advance. Therefore, it is necessary to calculate the health HPI variance D _kv and health HPI slope D _ks of the modeling object j in the set D _k , where D _kv = {d _kv1 , d _kv2 , …, d _kvn }, D _ks = {d _ks1 , d _ks2 , …, d _ksn };

4) Set the thresholds for the variance and slope of the health HPI. If the thresholds are exceeded, the model will start to warn.

5) Calculate the variance and slope overlap rate indicators pD _kvp and pD _ksp of the historical production abnormality warning obtained within k time periods before the production abnormality according to the threshold parameters d _kv T and d _ks T, where

6) Establish the HPI trend warning coincidence rate loss function to characterize the coincidence rate of advance variance, slope abnormal change and historical production abnormality. The higher the coincidence rate, the more appropriate the parameters k, d _kv T and d _ks T are;

p＝pD _kvp +pD _ksp ;

7) Use the heuristic algorithm to solve the HPI trend warning coincidence rate loss function and find the appropriate values of parameters k, d _kv T, and d _ks T to maximize the loss function p value;

8) Substitute the optimal alarm advance time k of the modeling object, the most appropriate health HPI variance threshold d _kv T and the slope threshold d _ks T of the modeling object into steps 1)-7), and re-solve as the model is updated, so it is adaptive;

9) When the model is running, the health HPI variance d _kv within the k time period is calculated in real time to see whether it exceeds d _kv T. If so, a variance warning is prompted; the health HPI slope d _ks within the k time period is calculated in real time to see whether it exceeds d _ks T. If so, a slope warning is prompted.

Using historical data, according to the statistical method of threshold parameters for fault reclosing rate, the fault reclosing rate reflects whether the above parameter threshold settings are reasonable, and can further determine whether the fault type has the possibility of early warning.

First, the element thresholds of the sets D _kv and D _ks are initialized to d _kv T and d _ks T respectively. Then, the fault coincidence judgment sets D _kvp and D _ksp are obtained according to the threshold parameters d _kv T and d _ks T.

Initialize d _kv T

Initialize _dksT

Set D _kvp ={d _kvp1 , d _kvp2 , ..., d _kvpn }

in

Set D _ksp ={d _ksp1 , d _ksp2 , ..., d _kspn }

in

Obtain the judgment result matrix of whether the slope trend and variance trend of the health HPI exceed the threshold before the historical production abnormality.

Compared with the prior art, the advantages of the present invention are: increasing the advance of model warning and realizing predictive maintenance. At the same time, for different evaluation objects, the adaptability of the model is improved by setting hyperparameters k, d _kv T, and d _ks T.

The evaluation system warnings can be divided into the following categories to achieve the priority classification of fault warnings:

Category 1: Health HPI slope trend decline warning, or health HPI variance oscillation warning, and there are corresponding reason rules for reporting, which are clear (know the specific problem) early general level warnings. This type of warning is valuable to users. If it is very large, it can be discovered in advance and treated, which can usually avoid production abnormalities.

The second category: Health HPI slope trend decline warning, or Health HPI variance oscillation warning, but no corresponding reason rules are reported. This is a fuzzy (unknown specific problem) general level early warning. This type of warning is less valuable to users because although it warns in advance, it does not accurately tell the reason for the warning.

The third category: The health HPI drops below the baseline. At this time, the residuals of the associated measurement points are generally large, and most of them are lagging (problems may have been discovered on site) and are serious warnings.

BRIEF DESCRIPTION OF THE DRAWINGS

Figure 1 is a trend warning diagram of the preheater health index (HPI).

Figure 2 Preheater health index (HPI) trend warning configuration interface.

Figure 3 Preheater Health Index (HPI) trend warning list.

Detailed ways

The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments.

Step 1: The on-site production anomaly records contain the start and end time of the production anomaly, the reason for the shutdown, and other information. The above information can be used as the source of negative samples for the model. Through algorithm extraction, the data pattern before the real production anomaly can be obtained, thereby achieving early warning. According to the on-site production anomaly records, the historical production anomaly time set D of the modeling object j is screened out, D = {d ₁ , d ₂ , ..., d _n }, where d _n = t _nbegin , d _n is the nth production anomaly time, and t _nbegin is the start time of the nth anomaly time period.

Step 2: The model needs to capture the data change pattern within k time before the production anomaly. According to the abnormal time set D obtained in step 1, use the start time of each production anomaly _tnbegin minus k time to represent the advance of the production anomaly in the data pattern. The initial k value is determined according to the advance warning requirements of the modeling object. The K value is the K period of time before the modeling object failure occurs. The physical meaning is to characterize the advance of the monitoring object failure in the data pattern.

The initial K value is determined based on the equipment operating characteristics and the type of fault. For example, temperature faults generally occur more early (K is 2-3 hours), while process faults generally occur less early (K is 5-15 minutes).

The optimization of K value is shown in step 7): Use a heuristic algorithm to find a suitable K value based on the loss function.

The set of k time periods before the historical production anomaly is obtained _{, D k} ₌ {d _k1 , d _k2 , ..., d _kn }, where d _kn = (t _{n begin} - k, t _{n begin} ). d _kn is the warning lead time before each production anomaly, and is also the time when the data anomaly begins to exist.

Step 3: The health HPI variance and slope change trend of the model evaluation object within k time before production abnormality can reflect the abnormality in advance. Therefore, it is necessary to calculate the health HPI variance D _kv and health HPI slope of the modeling object j in the set D _k D _ks , where D _kv ={d _kv1 , d _kv2 , ... , d _kvn }, D _ks ={d _ks1 , d _ks2 , ... , d _ksn };

Step 4: Set the thresholds for the variance and slope of the health HPI. If the thresholds are exceeded, the model starts to warn. The following is a statistical method for the fault overlap rate based on the threshold parameters using historical data. The fault overlap rate reflects whether the above parameter threshold settings are reasonable, and can further determine whether the fault type has the possibility of early warning.

Firstly, the element thresholds of the sets D _kv and D _ks are initialized to d _kv T and d _ks T respectively. Secondly, the fault coincidence judgment sets D _kvp and D _ksp are obtained according to the threshold parameters d _kv T and d _ks T.

Initialize d _kv T

Initialize _dksT

Set D _kvp ={d _kvp1 , d _kvp2 , ..., d _kvpn }

in

Set D _ksp ={d _ksp1 , d _ksp2 , ..., d _kspn }

in

After the above steps, we can get the judgment result matrix of whether the slope trend and variance trend of health HPI exceed the threshold before the historical production abnormality.

Step 5: Different evaluation models have different value ranges for the related parameters k, d _kv T, and d _ks T. In order to improve the accuracy of the early warning, it is necessary to calculate the variance and slope overlap rate indicators pD _kvp and pD _ksp of the historical production abnormality early warning based on the threshold parameters d _kv T and d _ks T within k hours before the production abnormality, where

Step 6: Establish the HPI trend warning coincidence rate loss function, which can characterize the coincidence rate of advance variance, slope abnormal change and historical production anomaly. The higher the coincidence rate, the more appropriate the parameters k, d _kv T and d _ks T are.

p＝pD _kvp +pD _ksp

Step 7: Use a heuristic algorithm to solve the loss function of step 6 and find the appropriate values of parameters k, d _kv T, and d _ks T to maximize the loss function p value.

Step 8: After steps 1 to 7, the optimal alarm advance time k of the modeling object, the most appropriate health HPI variance threshold d _kv T, and the slope threshold d _ks T of the modeling object can be determined. The above steps 1 to 7 are re-solved as the model is updated, so they are adaptive.

Step 9: When the model is running, the health HPI variance d _kv within the k time period is calculated in real time to see whether it exceeds d _kv T. If so, a variance warning is prompted; the health HPI slope d _ks within the k time period is calculated in real time to see whether it exceeds d _ks T. If so, a slope warning is prompted.

Example

The health value is defined as a decrease in two consecutive hours, and the decrease exceeds α. α is a loss function based on the fault coincidence rate and is solved by a heuristic algorithm. In the HPI trend warning case of the preheater health, α = -0.01 and k = 2.

Figure 1 is the trend analysis of the preheater health index (HPI). Within two hours, the health index of the monitoring model decreased by 2.3%.

Figure 2 Preheater health index (HPI) trend warning configuration interface. According to the offline model calculation results, configure the slope trend warning parameters: continuous decline time, decline amplitude threshold.

Figure 3 Preheater health index (HPI) trend warning list. A series of warning information is obtained through the health index decline trend.

Claims

A method for improving the early warning and accuracy of an evaluation model, characterized in that the steps include:

1) The on-site production abnormality records include the start and end time of the production abnormality and the reason for the shutdown. The on-site production abnormality records are used as the source of the model negative samples. The data pattern before the real production abnormality is extracted by the sliding window length method. According to the on-site production abnormality records, the historical production abnormality time set D of the modeling object j is screened out, D = {d 1 , d 2 , ..., d n }, where d n = t nbegin , d n is the nth production abnormality time, and t nbegin is the start time of the nth abnormal time period;

2) The model captures the data change rules within k time periods before the production anomaly: According to the abnormal time set D obtained in step 1), each production anomaly start time tnbegin minus k time periods is used to represent the advance time of the production anomaly in the data pattern. The initial k value is determined according to the requirements of the early warning advance of the modeling object, and the historical production anomaly k time period set Dk is obtained, Dk = { dk1 , dk2 , ..., dkn }, where dkn = ( tnbegin -k, tnbegin ), dkn is the early warning advance time before each production anomaly, and is also the time when the data anomaly begins to exist;

3) The health HPI variance and slope change trends of the model evaluation object within k time before the production abnormality reflect the abnormality in advance. Therefore, it is necessary to calculate the health HPI variance D kv and health HPI slope D ks of the modeling object j in the set D k , where D kv = {d kv1 , d kv2 , …, d kvn }, D ks = {d ks1 , d ks2 , …, d ksn };

4) Set the thresholds for the variance and slope of the health HPI. If the thresholds are exceeded, the model will start to warn.

5) Calculate the variance and slope overlap rate indicators pD kvp and pD ksp of the historical production abnormality warning obtained within k time periods before the production abnormality according to the threshold parameters d kv T and d ks T, where

6) Establish the HPI trend warning coincidence rate loss function to characterize the coincidence rate of advance variance, slope abnormal change and historical production abnormality. The higher the coincidence rate, the more appropriate the parameters k, d kv T and d ks T are;
p＝pD kvp +pD ksp ;

7) Use the heuristic algorithm to solve the HPI trend warning coincidence rate loss function and find the appropriate values of parameters k, d kv T, and d ks T to maximize the loss function p value;

8) Substitute the optimal alarm advance time k of the modeling object, the most appropriate health HPI variance threshold d kv T and the slope threshold d ks T of the modeling object into steps 1)-7), and re-solve as the model is updated, so it is adaptive;

9) When the model is running, the health HPI variance d kv within the k time period is calculated in real time to see whether it exceeds d kv T. If so, a variance warning is prompted; the health HPI slope d ks within the k time period is calculated in real time to see whether it exceeds d ks T. If so, a slope warning is prompted.
The method for improving the early warning and accuracy of the evaluation model according to claim 1 is characterized in that the historical Based on historical data, the fault reclosing rate is statistically analyzed according to the threshold parameters. The fault reclosing rate reflects whether the above parameter threshold settings are reasonable, and can further determine whether the fault type has the possibility of early warning.

First, the element thresholds of the sets D kv and D ks are initialized to d kv T and d ks T respectively. Then, the fault coincidence judgment sets D kvp and D ksp are obtained according to the threshold parameters d kv T and d ks T.

Initialize d kv T

Initialize dksT

Set D kvp ={d kvp1 , d kvp2 , ..., d kvpn }

in

Set D ksp ={d ksp1 , d ksp2 , ..., d kspn }

in

Obtain the judgment result matrix of whether the slope trend and variance trend of the health HPI exceed the threshold before the historical production abnormality.