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

Abstract

This invention discloses a method to improve the early warning and accuracy of an assessment model. Based on on-site production anomaly records, a historical production anomaly time set D is selected for the modeling object j. The model captures the data change patterns within k time periods prior to the production anomaly. The model assesses the health index (HPI) variance and slope trends of the object within k time periods prior to the anomaly, reflecting the anomaly in advance. Thresholds are set for the health index (HPI) variance and slope; if these thresholds are exceeded, the model initiates an early warning. Based on the threshold parameter d... kv T, d ks T, the variance and slope overlap rate index pD obtained from historical production anomaly early warning. kvp pD ksp Establish the overlap rate, use a heuristic algorithm to solve the HPI trend early warning overlap rate loss function, and find the parameters k and d. kv T, d ks A suitable value for T maximizes the loss function p; during model runtime, the variance d of the health index (HPI) over k time intervals is calculated in real time. kv Does it exceed d? kv If T exceeds the limit, a variance warning will be issued; the health HPI slope d within the k-time period will be calculated in real time. ks Does it exceed d? ks If T is exceeded, a slope warning will be issued.

Description

Technical Field

[0001] This invention relates to a modeling method for intelligent assessment and diagnostic models, and in particular to a method for improving the early warning and accuracy of assessment models. Background Technology

[0002] Cement companies operate in a continuous production process, with hundreds or even thousands of sensors collecting data from key nodes in each production stage. Currently, this data is primarily used by users for monitoring threshold ranges at individual measuring points and for process debugging. Both of these methods are based on point monitoring or control loop (line) monitoring, and cannot provide a comprehensive (overall) status assessment.

[0003] Patent CN102270271B, "Method and System for Early Warning and Optimization of Equipment Failure Based on Similarity Curves," or simply the Evaluation System (IEM), utilizes historical normal operation data from factory equipment and process measurement points to establish equipment and process models for monitoring equipment failures, providing early warnings, or detecting drift in operating conditions. The Evaluation System (IEM) outputs a real-time Health Index (HPI) score for the modeled object, comparing it to a standard HPI value. An alarm is triggered when the real-time HPI remains below the standard HPI value for a period of time. The HPI value is a fixed standard value calculated based on historical modeling data.

[0004] The application of the aforementioned patented technology in the cement industry has encountered the following problems:

[0005] First, alarms obtained by comparing the model health index (HPI) benchmark value often indicate that the modeled object is already in an abnormal production state. Therefore, it fails to meet the need for advance prediction, forecasting, and early warning.

[0006] Second, the Evaluation and Assessment System (IEM) has a rich set of alarm cause rules, most of which describe data changes at key measurement points before production anomalies. However, in practical applications, based on the first point, it was found that when an HPI warning was generated, no cause rule matching the criteria was reported.

[0007] Analysis of the causes of the problems in the above applications:

[0008] First, the Health Index (HPI) value calculated by the Assessment System (IEM) model is the expected value at a single point in time, referencing health index values ​​corresponding to similar state spaces in historical data. Therefore, the health index obtained by the assessment model does not directly provide trend prediction functionality. Currently, general expert experience is based on matching the Health Index (HPI) or the recent trend of a specific measurement point. Essentially, the Assessment System (IEM) is a point-to-point classification model without time trends, while real-world rule matching involves multi-measurement point combinations with time trends. This is the fundamental difference, and it prevents the model from providing early warning functionality.

[0009] Second, the model health index (HPI) is influenced by the overall residuals of all measurement points, while alarm rules are triggered based on the residuals of specific measurement points and a combination of rules. Therefore, when the HPI generates an alarm, the alarm rules for the combined measurement points may not necessarily appear simultaneously.

[0010] Therefore, in order to achieve early warning of faults and report more corresponding fault causes, the early warning method of the IEM assessment system is improved. Summary of the Invention

[0011] To address the problems existing in the prior art, this invention provides a method to improve the early warning and accuracy of the evaluation model, which can reflect the early warning trend of faults, measure the overlap rate between the early warning trend and historical faults, and solve the problem of high cost of manual correction of early warning accuracy.

[0012] The objective of this invention is achieved through the following technical solutions.

[0013] A method for improving the early warning and accuracy of assessment models, comprising the following steps:

[0014] 1) On-site production anomaly records include the start and end times of production anomalies and the reasons for downtime. These records are used as the source of negative samples for the model. A sliding window method is used to extract the data pattern prior to the actual production anomalies. Based on these records, a set D of historical production anomaly times for the modeling object j is selected, where D = {d1, d2, ..., d...}. n}, where d n =t nbegin d n This is the nth production anomaly time, t nbegin It is the start time of the nth abnormal time period;

[0015] 2) The model captures the data change patterns within the k time intervals preceding the production anomaly: Based on the set of anomaly times D obtained in step 1), the model uses the start time t of each production anomaly... nbeginSubtracting k time periods forward represents the predictability of production anomalies in the data pattern. The initial value of k is determined based on the early warning requirements of the modeling object. This yields the set D of the k time periods preceding historical production anomalies. k D k ={d k1 d k2 , ..., d kn}, where d kn =(t n begin -k,t nbegin ), d kn It is the early warning time before each production anomaly, and also the time when data anomalies begin to appear;

[0016] 3) The model evaluates the health status (HPI) variance and slope trends of the object k time before the production anomaly, which reflect the anomaly in advance. Therefore, it is necessary to calculate the modeling object j in set D. k Health variance D within the HPI kv Health Index (HPI) Slope D ks D kv ={d kv1 d kv2 , ..., d kvn}, D ks ={d ks1 d ks2 , ..., d ksn};

[0017] 4) Set thresholds for the variance and slope of the Health Index (HPI). If these thresholds are exceeded, the model will issue an early warning.

[0018] 5) Calculate the threshold parameter d within the k time periods prior to the production anomaly. kv T, d ks T, the variance and slope overlap rate index pD obtained from historical production anomaly early warning. kvp pD ksp ,in

[0019]

[0020]

[0021] 6) Establish an HPI trend early warning overlap rate loss function to characterize the overlap rate of advance variance, abnormal slope changes, and historical production anomalies. The higher the overlap rate, the better the parameters k and d. kv T, d ks The more suitable T is;

[0022] p = pD kvp +pD ksp ;

[0023] 7) Use a heuristic algorithm to solve the HPI trend warning overlap rate loss function and find the parameters k and d. kv T, d ks A suitable value for T maximizes the value of the loss function p.

[0024] 8) Determine the optimal alarm lead time k for the modeling object and the most suitable health HPI variance threshold d for the modeling object. kv T, slope threshold d ks Substituting T into steps 1)-7), the solution is recalculated as the model is updated, thus exhibiting adaptability;

[0025] 9) During model runtime, the variance d of health index (HPI) within time period k is calculated in real time. kv Does it exceed d? kv If T exceeds the limit, a variance warning will be issued; the health HPI slope d within the k-time period will be calculated in real time. ks Does it exceed d? ks If T is exceeded, a slope warning will be issued.

[0026] Using historical data and statistical methods for calculating fault overlap rates based on threshold parameters, the fault overlap rate reflects whether the threshold settings are reasonable and can further determine whether the fault type has the potential for early warning.

[0027] First, initialize set D. kv D ks The element thresholds are d kv T, d ks T, and then based on the threshold parameter d kv T, d ks T obtains the fault overlap judgment set D. kvp D ksp ,

[0028] Initialize d kv T

[0029]

[0030] Initialize d ks T

[0031]

[0032] set D kvp ={d kvp1 d kvp2 , ..., d kvpn}

[0033] in

[0034]

[0035] set Dksp ={d ksp1 d ksp2 , ..., d kspn}

[0036] in

[0037]

[0038] Before obtaining historical production anomalies, a matrix of judgment results was obtained regarding whether the slope trend and variance trend of the Health Index (HPI) exceeded the threshold.

[0039] Compared to existing technologies, the advantages of this invention are: increased advance warning capability of the model, enabling predictive maintenance. Furthermore, for different evaluation objects, the hyperparameters k and d can be set... kv T, d ks T improves the model's adaptability. Based on this...

[0040] The assessment system's early warnings can be further categorized into the following types to prioritize fault warnings based on their severity and urgency:

[0041] The first category is a warning of a declining slope trend in the Health Index (HPI) or an oscillation in the Health Index (HPI) variance, with corresponding cause rules provided. This type of warning is a clear (knowing the specific problem) early warning of a general level. This type of warning is very valuable to users, as it allows for early detection and handling measures, which can usually prevent production anomalies.

[0042] The second category is: warnings of declining slope trend of Health Index (HPI) or oscillation of Health Index (HPI) variance, but without corresponding reason rules. These are vague (the specific problem is unknown) early warnings of a general level. These warnings are of slightly lower value to users because although they provide early warnings, they do not accurately tell the reason for the warning.

[0043] The third category is when the HPI (Health Index) simply drops below the baseline. In this case, the residuals of the associated measurement points are generally large, and most of them are lagging warnings (the problem may have already been found on-site). Attached Figure Description

[0044] Figure 1 This is a trend warning chart for the health index (HPI) of the preheater.

[0045] Figure 2 Preheater Health Index (HPI) Trend Warning Configuration Interface.

[0046] Figure 3 Preheater Health Index (HPI) Trend Warning List. Detailed Implementation

[0047] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0048] A method for improving the early warning and accuracy of assessment models, comprising the following steps:

[0049] Step 1: The on-site production anomaly records contain information such as the start and end times of the anomaly, and the reason for the downtime. This information can serve as a source of negative samples for the model. Through algorithmic extraction, the data pattern before the actual production anomaly can be obtained, thereby achieving early warning. Based on the on-site production anomaly records, a set D of historical production anomaly times for the modeling object j is selected, where D = {d1, d2, ..., d...}. n}, where d n =t nbegin d n This is the nth production anomaly time, t nbegin It is the start time of the nth abnormal time period.

[0050] Step 2: The model needs to capture the data change patterns within the k time periods preceding the production anomaly. Based on the anomaly time set D obtained in Step 1, the start time t of each production anomaly is used... nbegin Subtracting k time periods forward represents the advance nature of production anomalies in the data pattern. The initial value of k is determined based on the advance warning requirements of the modeled object. The value of k represents the k time periods before the model needs to monitor the failure of the modeled object. Physically, it characterizes the advance nature of the failure of the monitored object in the data pattern.

[0051] The initial K value is determined based on the equipment's operating characteristics and the type of fault. For example, temperature-related faults generally occur with a relatively large lead time (K is set to 2-3 hours), while process-related faults generally occur with a smaller lead time (K is set to 5-15 minutes).

[0052] See step 7 for K value optimization: Use a heuristic algorithm to find a suitable K value based on the loss function.

[0053] Obtain the set D of the k time periods preceding the historical production anomalies. k D k ={d k1 d k2 , ..., d kn}, where d kn =(t n begin -k,t nbegin ). d kn It is the early warning time before each production anomaly, and also the time when data anomalies begin to appear.

[0054] Step 3: The model evaluates the health status (HPI) variance and slope trends of the object within k time intervals prior to production anomalies. These trends can predict anomalies in advance. Therefore, it is necessary to calculate the modeling object j within set D.k Health variance D within the HPI kv Health Index (HPI) Slope D ks D kv ={d kv1 d kv2 , ..., d kvn}, D ks ={d ks1 d ks2 , ..., d ksn};

[0055] Step 4: Set thresholds for the HPI variance and slope. If these thresholds are exceeded, the model will issue an early warning. The following is a statistical method for calculating the fault overlap rate based on historical data and threshold parameters. The fault overlap rate reflects the reasonableness of the above threshold settings and can further determine whether this fault type has the potential for early warning.

[0056] First, initialize set D. kv D ks The element thresholds are d kv T, d ks T, and then based on the threshold parameter d kv T, d ks T obtains the fault overlap judgment set D. kvp D ksp .

[0057] Initialize d kv T

[0058]

[0059] Initialize d ks T

[0060]

[0061] set D kvp ={d kvp1 d kvp2 , ..., d kvpn}

[0062] in

[0063]

[0064] set D ksp ={d ksp1 d ksp2 , ..., d kspn}

[0065] in

[0066]

[0067] After the above steps, a matrix of judgment results can be obtained to determine whether the slope trend and variance trend of the Health Index (HPI) exceeded the threshold before the historical production anomaly.

[0068] Step 5: Different evaluation models, their relevant parameters k, d kv T, d ks T has different value ranges. To improve the accuracy of early warning, it is necessary to calculate the threshold parameter d within k time periods before the production anomaly. kv T, d ks T, the variance and slope overlap rate index pD obtained from historical production anomaly early warning. kvp pD ksp ,in

[0069]

[0070]

[0071] Step 6: Establish the HPI trend early warning overlap rate loss function. This function can characterize the overlap rate of advance variance, abnormal slope changes, and historical production anomalies. The higher the overlap rate, the better the parameters k and d are. kv T, d ks The more suitable T is.

[0072] p = pD kvp +pD ksp

[0073] Step 7: Use a heuristic algorithm to solve for the loss function in step 6, and find the parameters k and d. kv T, d ks A suitable value for T maximizes the value of the loss function p.

[0074] Step 8: After steps 1 to 7, the optimal alarm lead time k and the most suitable health HPI variance threshold d for the modeling object can be determined. kv T, slope threshold d ks T. Steps 1 to 7 above are solved again as the model is updated, thus exhibiting adaptability.

[0075] Step 9: During model runtime, calculate the variance d of health index (HPI) in real time over k time intervals. kv Does it exceed d? kv If T exceeds the limit, a variance warning will be issued; the health HPI slope d within the k-time period will be calculated in real time. ks Does it exceed d? ks If T is exceeded, a slope warning will be issued.

[0076] Example

[0077] The health status value is defined as decreasing for two consecutive hours, and the decrease exceeds α. α is a loss function based on the fault overlap rate, and the threshold for the slope of the preheater health status decrease is calculated using a heuristic algorithm. In the preheater health status HPI trend early warning case, α = -0.01, k = 2.

[0078] Figure 1 The preheater health index (HPI) trend analysis showed that the health index of the monitoring model decreased by 2.3% within two hours.

[0079] Figure 2 The preheater health index (HPI) trend warning configuration interface allows you to configure slope trend warning parameters based on the offline model calculation results: continuous decline time and decline magnitude threshold.

[0080] Figure 3 Preheater Health Index (HPI) Trend Warning List. A series of warning messages are generated based on the declining health index trend.

Claims

1. A method for improving the early warning capability and accuracy of assessment models, characterized by the following steps: include: 1) On-site production anomaly records include the start and end times of the anomaly and the reason for the downtime. These records are used as the source of negative samples for the model. A sliding window method is used to extract the data pattern before the actual production anomaly. Based on these records, the modeling objects are selected. Historical production anomaly time set , ,in , It is the first An abnormal production time, It is the first The start time of each abnormal time period; 2) Model captures production anomalies before Data variation patterns within a time period: Based on the abnormal time set obtained in step 1). Use the start time of each production exception Subtract forward This time period represents the advance nature of production anomalies in the data pattern, the initial... The value is determined based on the early warning requirements of the modeling object, and the historical production anomalies before... A set of time periods , ,in , It is the early warning time before each production anomaly, and also the time when data anomalies begin to appear; 3) The model evaluates the object before the production anomaly. The variation trend of HPI variance and slope over a period of time can reflect anomalies in advance, therefore it is necessary to calculate the modeling object. In the set Health variance within HPI Health Index (HPI) Slope ,in , ; 4) Set thresholds for the variance and slope of the Health Index (HPI). If these thresholds are exceeded, the model will issue an early warning. 5) Calculate before production anomalies Within a time period, based on the threshold parameter , The variance and slope overlap rate of historical production anomaly early warning indicators were obtained. , ,in ， ； Fault overlap judgment set , ,gather , in ， gather , in ; 6) Establish an HPI trend early warning overlap rate loss function to characterize the overlap rate of advance variance, abnormal slope changes, and historical production anomalies. The higher the overlap rate, the stronger the parameter. , , The more suitable; ； 7) Use a heuristic algorithm to solve the HPI trend warning overlap rate loss function and find the parameters. , , A suitable value makes the loss function Maximum value; 8) Determine the optimal alarm lead time for the modeling object. The most suitable health HPI variance threshold for the modeling object Slope threshold Substituting into steps 1)-7), the solution is recalculated as the model is updated, thus exhibiting adaptability; 9) Real-time calculation during model runtime Health Index (HPI) variance over the time period Does it exceed If the variance exceeds the limit, a variance warning will be issued; real-time calculation. Health Index (HPI) slope over a given time period Does it exceed If the slope is exceeded, a slope warning will be displayed.

2. The method for improving the early warning and accuracy of the assessment model according to claim 1, characterized in that... Using historical data and based on the statistical method of fault overlap rate according to threshold parameters, the fault overlap rate reflects whether the threshold parameter setting is reasonable, and can further determine whether the fault type has the possibility of early warning. First, initialize the collection. , The element thresholds are respectively , Secondly, based on the threshold parameter , Obtain the fault overlap judgment set , , initialization , ， initialization , ， Before obtaining historical production anomalies, a matrix of judgment results was obtained regarding whether the slope trend and variance trend of the Health Index (HPI) exceeded the threshold.

Images