Chemical process abnormal type identification method based on multi-dimensional causal correlation index

Abstract

The application discloses a chemical process abnormal type recognition method based on multi-dimensional causal correlation indexes, and belongs to the technical field of chemical process monitoring and fault diagnosis. The method firstly carries out offline modeling: collecting normal working condition data, constructing a monitoring model by using a trend slow characteristic analysis algorithm, and determining a statistical limit. In online application, real-time data are calculated by the model to obtain a statistical value, and if the statistical value exceeds the limit, an alarm is given and a type recognition process is triggered. Whether a quality variable deviates is analyzed, and a root variable causing the abnormality is further analyzed from four dimensions, i.e., a position (CPI) in a process, an attribute (CAI) in a causal network and an influence propagation capability (CPD), to construct a multi-dimensional causal correlation index (MCAI). The application solves the problem that a traditional abnormal detection method cannot distinguish between process faults and new working conditions, and can effectively avoid false alarms, thereby providing a decision basis for accurate maintenance.

Description

Technical Field

[0001] This invention belongs to the field of chemical process monitoring and fault diagnosis technology, specifically relating to a method for identifying abnormal types of chemical processes based on multi-dimensional causal correlation indicators. Background Technology

[0002] With the continuous expansion of modern chemical production scale and the increasing complexity of processes, the operating conditions, reaction mechanisms, and control strategies in production processes exhibit characteristics such as high coupling, nonlinearity, and dynamic variability. Against this backdrop, process monitoring and anomaly identification technologies play a crucial role in ensuring production safety, improving product quality, and reducing economic losses. However, in actual chemical process operation, abnormal changes generally fall into two categories. One category is anomalies caused by process failures, such as instrument measurement malfunctions or severe changes in process parameters. These anomalies can reduce product quality or worsen process operation. The other category is anomalies caused by new operating conditions. Due to factors such as changes in external temperature and fluctuations in upstream raw materials, adjustments to operating conditions frequently occur. These new operating conditions differ from the original conditions and are considered normal fluctuations, distinct from process failures.

[0003] In recent years, data-driven machine learning methods have been widely used in the field of process anomaly detection. Trend-based Slow Feature Analysis (TSFA), as an effective feature extraction method, can capture the dynamic and static characteristics of process variables, establish normal operating condition models, and determine whether anomalies have occurred by monitoring whether statistics exceed limits. However, most anomaly detection methods, represented by TSFA, can only provide anomaly alarms and cannot identify the type of detected anomalies, that is, they cannot distinguish whether the anomaly is a hazardous "process failure" or an acceptable "new operating condition." This limitation easily leads to false alarms, causing operators to take unnecessary interventions for harmless new operating condition fluctuations or respond slowly to potential faults.

[0004] Therefore, based on existing anomaly detection technologies, developing a method that can automatically distinguish between different types of anomalies in chemical processes (especially process failures and new operating conditions) is of great practical significance for achieving accurate alarms, optimizing production and maintenance strategies, and reducing operating costs. Summary of the Invention

[0005] This invention aims to solve the technical problem that existing chemical process anomaly detection technologies (such as TSFA) can only issue alarms but cannot distinguish the type of anomaly (especially the process failure and new operating conditions). It provides a method that can identify anomaly types online and automatically, thereby reducing false alarms and providing operators with more accurate decision support.

[0006] To achieve the above objectives, the present invention adopts the following technical solution;

[0007] A method for identifying anomaly types in chemical processes based on multi-dimensional causal correlation indicators is applied to chemical production processes involving continuous stirred reactors. This method comprises two stages: offline modeling and online identification.

[0008] The offline modeling phase includes the following steps;

[0009] Step 1: Normal operating condition data acquisition;

[0010] Collection of chemical processes includes Sample Normal training dataset with 1 variable ,in For normal training datasets Perform preprocessing to obtain the preprocessed training dataset. ;

[0011] Step 2, anomaly detection model construction;

[0012] The online anomaly type identification phase includes the following steps;

[0013] Step 3, online anomaly detection;

[0014] Collecting online data from chemical processes After preprocessing, the data is input into the anomaly detection model to calculate the monitoring statistics corresponding to the online data. and If the monitored statistics exceed the control limit, an anomaly is determined to have occurred.

[0015] Step 4: Calculate the deviation indicators of the quality variables;

[0016] When the statistical value exceeds the control limit and an alarm is triggered, the data acquisition window width is [value missing]. Online key quality variable data Calculate the deviation of quality variables This leads to the construction of the first dimension of a multi-dimensional causal correlation indicator, namely the deviation indicator of the quality variable. ;

[0017] Step 5: Isolate outliers and determine their causal location;

[0018] Collection time Online data Calculate online data The contribution of each variable is analyzed to identify outliers; based on the process mechanism, the causal position of each outlier is determined.

[0019] Step 6: Based on the transitive entropy method, determine the causal attributes of the outlier variables and the causal propagation indicators of the outlier variables;

[0020] Calculate the standard phase transfer entropy between each outlier variable to obtain the standard phase transfer entropy matrix. This allows us to determine the causal attributes of each outlier variable. ; Calculate the causal propagation markers for each outlier variable using the standard phase transfer entropy matrix. ;

[0021] Step 7: Constructing multi-dimensional causal relationship indicators;

[0022] Comprehensive utilization of quality variable deviation indicators Causal position of outlier variables causal attributes of abnormal variables and the causal propagation markers of anomalous variables Construct multi-dimensional causal relationship indicators.

[0023] Preferably, in step 1, the normal training dataset is processed using formula (1). After preprocessing, the expression for formula (1) is:

[0024] (1);

[0025] In the formula, For normal training dataset The mean, For normal training dataset The standard deviation.

[0026] Preferably, step 2 specifically includes the following steps:

[0027] Step 2.1: Calculate the linear transformation matrix using the slow trend feature analysis algorithm. The training dataset is obtained through formula (2). For each sample, perform slow trend feature analysis to obtain the slow trend feature matrix. ;

[0028] (2);

[0029] in, Represents a linear transformation matrix. Represents the first element in the linear transformation matrix. One vector; The generalized eigenvalues ​​are obtained by solving formula (3);

[0030] (3);

[0031] in, For the first The feature values ​​of a slow-trend feature, express The covariance matrix of the average first derivative after extracting the trend component. express The covariance matrix; It is a linear transformation matrix The first in List;

[0032] Step 2.2: Divide the slow trend feature matrix into main slow trend feature vectors. With secondary trend slow feature vector ;

[0033] The first The trend slow feature vector of each sample is denoted as... , slow trend feature vector Divided into main trend slow feature vectors With secondary trend slow feature vector ;in, The first d main trend slow feature vectors represent the core information of the system; The remaining md minor trend slow feature vectors are denoted by d, where d is the number of principal components and the value of d is determined by the cumulative contribution rate method.

[0034] Step 2.3: Construct the main monitoring statistics respectively. and secondary monitoring statistics ;

[0035] The following formulas (4) and (5) are used to indicate the changes in the system using two statistics;

[0036] (4);

[0037] (5);

[0038] This represents the main monitoring statistics constructed based on key features. This represents a secondary monitoring statistic constructed based on secondary features. and These represent the main trend slow feature vectors respectively. and secondary trend slow feature vectors The transpose of the statistic; Step 2.4, use the kernel density estimation method to determine the control limits of the primary monitoring statistic and the secondary monitoring statistic.

[0039] and .

[0040] Preferably, step 3 specifically includes the following steps:

[0041] Step 3.1, data collection Real-time online data The online data obtained after preprocessing the online data using formula (6) ;

[0042] (6);

[0043] in, express Real-time online data samples, express Online data obtained after preprocessing online data in real time;

[0044] Step 3.2, calculate the online data using formula (2). slow trend feature vector ;

[0045] Step 3.3, convert the slow trend feature vector Divided into two parts and ; Anomaly detection is performed using a trend slow feature analysis detection model, and the main monitoring statistics and secondary monitoring statistics corresponding to the online data are calculated according to formulas (7) and (8);

[0046] (7);

[0047] (8);

[0048] in, , They represent Real-time primary and secondary monitoring statistics calculated from constantly online data;

[0049] , express Slow feature vector of time trend The main and secondary parts are divided into two parts;

[0050] , They represent , transpose;

[0051] Step 3.4: Match the primary monitoring statistics and secondary monitoring statistics corresponding to the online data with their respective control limits. and If any value exceeds the control limit, it is considered abnormal.

[0052] Preferably, step 4 specifically includes the following steps:

[0053] Step 4.1, calculate the time using formula (9). At the time Medium quality variable data Deviation of quality variables ;

[0054] (9);

[0055] in, For reference values ​​of quality variables, for With a window width of Mean within the window Tolerance;

[0056] Step 4.2, use formula (10) to represent the deviation indicator of the quality variable. The judgment logic;

[0057] (10);

[0058] when This indicates that the value of the quality variable fluctuates within an acceptable range; an anomaly at this point may be due to a new operating condition or sensor malfunction. If the fault is not found, it is directly identified as a process failure, and no further analysis is required.

[0059] Preferably, step 5 specifically includes the following steps:

[0060] Step 5.1: Using contribution analysis, calculate the contribution value of each variable to the main monitoring statistic;

[0061] When the statistic exceeds the control limit alarm and the quality variable deviation flag equals 1, the anomaly type is further identified: contribution analysis is used to calculate... The contribution value of each variable is used to locate outlier variables by analyzing the magnitude of the contribution value. , , , The number of outlier variables; at time [time]. ,variable Main monitoring statistics Contribution value Calculated using formula (11);

[0062] (11);

[0063] in, express At that moment, the The measured values ​​of each process variable, the variable Main monitoring statistics The contribution value,

[0064] Furthermore, the contribution of the variable is expressed by formula (12);

[0065] (12);

[0066] in, It is a slow trend feature analysis projection matrix The Row vectors; Represents the first in the projection matrix One vector;

[0067] Calculate using formula (13) Each variable affects the main monitoring statistics. The mean of the sum of contributions ;

[0068] (13);

[0069] Variables whose contribution value exceeds the average contribution value of all variables are identified as outliers.

[0070] Step 5.2: Identify variables whose contribution values ​​exceed the average contribution value of all variables as outliers;

[0071] Will More than The variable was identified as an anomalous variable. , ;

[0072] Step 5.3, determine the first step based on the process mechanism relationship. Causal location of an outlier variable , If the variable If the input or stimulus is external to the system and is not directly controlled by internal state feedback, then define the variable. Let be the input variable. If the variable is an internal state or output variable of the process, then let The judgment logic is determined by formula (14);

[0073] (14).

[0074] Preferably, step 6 specifically includes the following steps:

[0075] Step 6.1, based on the preset window Using online data within the dataset, calculate the standard phase transfer entropy between each abnormal variable and construct the standard phase transfer entropy matrix;

[0076] The width of the acquisition window is Online data Calculate each abnormal variable Between moments At the time Standard phase transfer entropy within the window As shown in formula (15);

[0077] (15);

[0078] in, and abnormal variables Any two variables in abnormal variables to abnormal variables Phase transfer entropy, abnormal variables to abnormal variables The phase transfer entropy, ignoring values ​​less than 0.5. The value is used to obtain the standard phase transfer entropy matrix. This indicates that abnormal information is passed from row variables to column variables, using the standard phase transfer entropy matrix. As shown in formula (16);

[0079] (16);

[0080] Step 6.2: Based on the standard phase transit entropy matrix, calculate the causal propagation degree of each anomalous variable. ;

[0081] At any moment At the time In the window, the standard phase transfer entropy matrix is ​​used to calculate the first phase using formula (17). Causal propagation of an outlier variable ;

[0082] (17);

[0083] Step 6.3, Formula (18) represents the first... A causal propagation marker for an anomalous variable;

[0084] (18);

[0085] Step 6.4: If the anomalous variable satisfies the condition of passing information to other variables without being affected by the information passing of other variables, then the anomalous variable has a causal attribute. ,otherwise .

[0086] Preferably, in step 7, deviation indicators of quality variables are comprehensively utilized. Causal position of outlier variables causal attributes of abnormal variables = Causal propagation degree of outliers Four dimensions to construct multi-dimensional causal relationship indicators As shown in formula (20);

[0087] (20);

[0088] in, The number of abnormal variables. For sign functions; when When the quality has not deviated and the source variable causing the anomaly is an input variable with causal propagation, it meets the characteristics of a new operating condition, and the anomaly type is determined to be a new operating condition; when When this occurs, the exception type is determined to be a process failure.

[0089] The beneficial technical effects of this invention are as follows:

[0090] This invention provides a method for identifying anomaly types in chemical processes based on multi-dimensional causal correlation indicators. It constructs multi-dimensional causal correlation indicators from four dimensions: deviation indicators of quality variables, causal location of anomaly variables, causal attributes of anomaly variables, and causal propagation indicators of anomaly variables. This effectively distinguishes the specific types of abnormal operating conditions, avoiding misjudging new operating conditions as faults and facilitating subsequent targeted system maintenance. This method provides a valuable reference for online anomaly type identification in chemical processes, solving the technical problem of traditional anomaly detection methods misjudging new operating conditions as faults. Attached Figure Description

[0091] Figure 1 This is an overall flowchart of the method described in this invention. Detailed Implementation

[0092] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0093] A method for identifying anomaly types in chemical processes based on multi-dimensional causal correlation indicators is proposed. This method includes two stages: offline modeling and online identification.

[0094] The offline modeling phase includes the following steps;

[0095] Step 1: Normal operating condition data acquisition;

[0096] Collection of chemical processes includes Sample Normal training dataset with 1 variable ,in For normal training datasets Perform preprocessing to obtain the preprocessed training dataset. ;

[0097] Using formula (1) on the normal training dataset After preprocessing, the expression for formula (1) is:

[0098] (1);

[0099] In the formula, For normal training dataset The mean, For normal training dataset The standard deviation.

[0100] Step 2, anomaly detection model construction;

[0101] The online anomaly type identification phase includes the following steps;

[0102] Step 2.1: Calculate the linear transformation matrix using the slow trend feature analysis algorithm. The training dataset is obtained through formula (2). For each sample, perform slow trend feature analysis to obtain the slow trend feature matrix. ;

[0103] (2);

[0104] in, Represents a linear transformation matrix. Represents the first element in the linear transformation matrix. One vector; The generalized eigenvalues ​​are obtained by solving formula (3);

[0105] (3);

[0106] in, For the first The feature values ​​of a slow-trend feature, express The covariance matrix of the average first derivative after extracting the trend component. express The covariance matrix; It is a linear transformation matrix The first in List;

[0107] Step 2.2: Divide the slow trend feature matrix into main slow trend feature vectors. With secondary trend slow feature vector ;

[0108] The first The trend slow feature vector of each sample is denoted as... , slow trend feature vector Divided into main trend slow feature vectors With secondary trend slow feature vector ;in, The first d main trend slow feature vectors represent the core information of the system; The remaining md minor trend slow feature vectors are denoted by d, where d is the number of principal components and the value of d is determined by the cumulative contribution rate method.

[0109] Step 2.3: Construct the main monitoring statistics respectively. and secondary monitoring statistics ;

[0110] The following formulas (4) and (5) are used to indicate the changes in the system using two statistics;

[0111] (4);

[0112] (5);

[0113] This represents the main monitoring statistics constructed based on key features. This represents a secondary monitoring statistic constructed based on secondary features. and These represent the main trend slow feature vectors respectively. and secondary trend slow feature vectors The transpose of the statistic; Step 2.4, use the kernel density estimation method to determine the control limits of the primary monitoring statistic and the secondary monitoring statistic.

[0114] and .

[0115] Step 3, online anomaly detection;

[0116] Collecting online data from chemical processes After preprocessing, the data is input into the anomaly detection model to calculate the monitoring statistics corresponding to the online data. and If the monitored statistics exceed the control limit, an anomaly is determined to have occurred.

[0117] Specifically, the steps include the following:

[0118] Step 3.1, data collection Real-time online data The online data obtained after preprocessing the online data using formula (6) ;

[0119] (6);

[0120] in, express Real-time online data samples, express Online data obtained after preprocessing online data in real time;

[0121] Step 3.2, calculate the online data using formula (2). slow trend feature vector ;

[0122] Step 3.3, convert the slow trend feature vector Divided into two parts and ; Anomaly detection is performed using a trend slow feature analysis detection model, and the main monitoring statistics and secondary monitoring statistics corresponding to the online data are calculated according to formulas (7) and (8);

[0123] (7);

[0124] (8);

[0125] in, , They represent Real-time primary and secondary monitoring statistics calculated from constantly online data;

[0126] , express Slow feature vector of time trend The main and secondary parts are divided into two parts;

[0127] , They represent , transpose;

[0128] Step 3.4: Match the primary monitoring statistics and secondary monitoring statistics corresponding to the online data with their respective control limits. and If any value exceeds the control limit, it is considered abnormal.

[0129] Step 4: Calculate the deviation indicators of the quality variables;

[0130] When the statistical value exceeds the control limit and an alarm is triggered, the data acquisition window width is [value missing]. Online key quality variable data Calculate the deviation of quality variables This leads to the construction of the first dimension of a multi-dimensional causal correlation indicator, namely the deviation indicator of the quality variable. ;

[0131] Specifically, the steps include the following:

[0132] Step 4.1, calculate the time using formula (9). At the time Medium quality variable data Deviation of quality variables ;

[0133] (9);

[0134] in, For reference values ​​of quality variables, for With a window width of Mean within the window Tolerance;

[0135] Step 4.2, use formula (10) to represent the deviation indicator of the quality variable. The judgment logic;

[0136] (10);

[0137] when This indicates that the value of the quality variable fluctuates within an acceptable range; an anomaly at this point may be due to a new operating condition or sensor malfunction. If the fault is not found, it is directly identified as a process failure, and no further analysis is required.

[0138] Step 5: Isolate outliers and determine their causal location;

[0139] Collection time Online data Calculate online data The contribution of each variable is analyzed to identify outliers; based on the process mechanism, the causal position of each outlier is determined.

[0140] Specifically, the steps include the following:

[0141] Step 5.1: Using contribution analysis, calculate the contribution value of each variable to the main monitoring statistic;

[0142] When the statistic exceeds the control limit alarm and the quality variable deviation flag equals 1, the anomaly type is further identified: contribution analysis is used to calculate... The contribution value of each variable is used to locate outlier variables by analyzing the magnitude of the contribution value. , , , The number of outlier variables; at time [time]. ,variable Main monitoring statistics Contribution value Calculated using formula (11);

[0143] (11);

[0144] in, express At that moment, the The measured values ​​of each process variable, the variable Main monitoring statistics The contribution value,

[0145] Furthermore, the contribution of the variable is expressed by formula (12);

[0146] (12);

[0147] in, It is a slow trend feature analysis projection matrix The Row vectors; Represents the first in the projection matrix One vector;

[0148] Calculate using formula (13) Each variable affects the main monitoring statistics. The mean of the sum of contributions ;

[0149] (13);

[0150] Variables whose contribution value exceeds the average contribution value of all variables are identified as outliers.

[0151] Step 5.2: Identify variables whose contribution values ​​exceed the average contribution value of all variables as outliers;

[0152] Will More than The variable was identified as an anomalous variable. , ;

[0153] Step 5.3, determine the first step based on the process mechanism relationship. Causal location of an outlier variable , If the variable If the input or stimulus is external to the system and is not directly controlled by internal state feedback, then define the variable. Let be the input variable. If the variable is an internal state or output variable of the process, then let The judgment logic is determined by formula (14);

[0154] (14).

[0155] Step 6: Based on the transitive entropy method, determine the causal attributes of the outlier variables and the causal propagation indicators of the outlier variables;

[0156] Calculate the standard phase transfer entropy between each outlier variable to obtain the standard phase transfer entropy matrix. This allows us to determine the causal attributes of each outlier variable. ; Calculate the causal propagation markers for each outlier variable using the standard phase transfer entropy matrix. ;

[0157] Specifically, the steps include the following:

[0158] Step 6.1, based on the preset window Using online data within the dataset, calculate the standard phase transfer entropy between each abnormal variable and construct the standard phase transfer entropy matrix;

[0159] The width of the acquisition window is Online data Calculate each abnormal variable Between moments At the time Standard phase transfer entropy within the window As shown in formula (15);

[0160] (15);

[0161] in, and abnormal variables Any two variables in abnormal variables to abnormal variables Phase transfer entropy, abnormal variables to abnormal variables The phase transfer entropy, ignoring values ​​less than 0.5. The value is used to obtain the standard phase transfer entropy matrix. This indicates that abnormal information is passed from row variables to column variables, using the standard phase transfer entropy matrix. As shown in formula (16);

[0162] (16);

[0163] Step 6.2: Based on the standard phase transit entropy matrix, calculate the causal propagation degree of each anomalous variable. ;

[0164] At any moment At the time In the window, the standard phase transfer entropy matrix is ​​used to calculate the first phase using formula (17). Causal propagation of an outlier variable ;

[0165] (17);

[0166] Step 6.3, Formula (18) represents the first... A causal propagation marker for an anomalous variable;

[0167] (18);

[0168] Step 6.4: If the anomalous variable satisfies the condition of passing information to other variables without being affected by the information passing of other variables, then the anomalous variable has a causal attribute. ,otherwise .

[0169] Step 7: Constructing multi-dimensional causal relationship indicators;

[0170] Comprehensive utilization of quality variable deviation indicators Causal position of outlier variables causal attributes of abnormal variables = Causal propagation degree of outliers Four dimensions to construct multi-dimensional causal relationship indicators As shown in formula (20);

[0171] (20);

[0172] in, The number of abnormal variables. For sign functions; when When the quality has not deviated and the source variable causing the anomaly is an input variable with causal propagation, it meets the characteristics of a new operating condition, and the anomaly type is determined to be a new operating condition; when When this occurs, the exception type is determined to be a process failure.

[0173] The following is in conjunction with the appendix Figure 1 The present invention will be described in detail with reference to specific embodiments. This embodiment uses a typical continuous stirred reactor (CSTR) process as the test object. This process records 10 variables: product concentration... (Key quality variables), reactor temperature Cooling jacket temperature Liquid level Heat release Feed flow rate Feed concentration Feed temperature Coolant flow rate Coolant inlet temperature .

[0174] 1. Offline modeling;

[0175] The training set consists of 2000 sample data points collected during the stable operation of the CSTR process under normal working conditions. .right Standardization (subtract the mean, divide by the standard deviation) preprocessing is performed to obtain The TSFA algorithm is applied to determine the number of principal components based on a cumulative contribution rate > 85%. The characteristic transformation matrix is ​​obtained. And calculate the normal data. and Statistic. The control limit at 99% confidence level is determined using kernel density estimation. and .

[0176] 2. Online anomaly type identification and verification;

[0177] During the online testing phase, test datasets containing 2000 samples were generated for each of the five typical anomalies listed in Table 1. A pre-defined anomaly was introduced at the 261st sample point. Key parameter settings are as follows: Quality variable ( Reference value Tolerance Window width for quality variable analysis , transfer entropy calculation window width .

[0178] Taking Case 5 (New operating condition: feed flow rate increases slightly by 4 L / h) as an example, the online identification process is explained in detail:

[0179] Anomaly detection; online data flows through an established TSFA model, and real-time statistics are calculated. At time t=261, Statistics exceed control limits The system alarmed, indicating an anomaly in the judgment process.

[0180] Quality deviation analysis; take the period from t=261 to t=263 after the alarm ( ) quality variables Data, calculate its mean Because the step amplitude is small, Still within the target range Therefore, the calculation yields... ,and then This indicates that the product quality has not deviated.

[0181] Outlier identification and CPI assessment; contribution analysis of the sample at time t=261 to identify outliers. The contribution of (feed flow rate) is significantly higher than that of other variables, therefore the variable is... It was identified as the only outlier variable. According to the CSTR process mechanism, the feed flow rate... It is a typical external input variable, not directly controlled by the internal state of the reactor (such as temperature and concentration), therefore the variable The causal position is determined as the input variable, i.e. .

[0182] Causal attributes and propagation analysis; taking window data from t=261 to t=310 ( ), calculate abnormal variables Standardized phase transfer entropy between the variable and other process variables. Analyze the transfer entropy matrix. Discovery, Existence In this situation, The values ​​of all of them are close to 0. Therefore, the calculated variables are... causal propagation Therefore Meanwhile, variables The source variable condition of "only outputting without affecting" is satisfied, therefore... .

[0183] Comprehensive judgment; substitute the indicators of each dimension into formula (20) to calculate the multi-dimensional causal relationship indicators. Therefore, the system ultimately determined that this anomaly was a new operating condition.

[0184] Table 1: Description of 5 Abnormal Situations;

[0185] .

[0186] Table 2: Results of Differentiation of 5 Abnormal Situations;

[0187] .

[0188] Table 2 shows the complete identification results for the five anomalies. The results indicate that:

[0189] Situations 1 and 2 (process failures) cause quality variables. A significant and persistent deviation occurred. The system directly identifies the fault as a process failure, consistent with the preset type.

[0190] Case 3 (External instrument failure), quality variable No deviation ( The isolated abnormal variables are both input variables and source variables. However, an external instrument malfunction prevented the sensor measurements from truly driving process changes, thus hindering the causal propagation of the isolated abnormal variable. ,final The fault was correctly identified as an external instrument malfunction.

[0191] Case 4 (Instrument failure within the loop), quality variable No deviation ( This identified several outlier variables. Further analysis revealed that the source variables were internal process state variables rather than input variables. Therefore The fault was identified as an instrument malfunction within the circuit.

[0192] Case 5 (New Operating Condition), as analyzed above, satisfies... And the source variable that caused the exception was the input variable ( ), and produced effective causal propagation to other variables in the process ( Therefore It accurately identifies the new operating condition.

[0193] The above embodiments fully demonstrate that the method of the present invention can not only effectively detect abnormalities in chemical processes, but also accurately distinguish the specific types of abnormalities, especially accurately identify new operating conditions, thereby significantly reducing the false alarm rate and providing reliable technical support for intelligent decision-making in production operations.

[0194] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for identifying anomaly types in chemical processes based on multi-dimensional causal correlation indicators, applied to chemical production processes including continuous stirred reactors, characterized in that... This method comprises two stages: offline modeling and online recognition; among which, The offline modeling phase includes the following steps; Step 1: Normal operating condition data acquisition; Collection of chemical processes includes Sample Normal training dataset with 1 variable ,in For normal training datasets Perform preprocessing to obtain the preprocessed training dataset. ; Step 2, anomaly detection model construction; The online anomaly type identification phase includes the following steps; Step 2.1: Calculate the linear transformation matrix using the slow trend feature analysis algorithm. The training dataset is obtained through formula (2). For each sample, perform slow trend feature analysis to obtain the slow trend feature matrix. ; （2）； in, Represents a linear transformation matrix. Represents the first element in the linear transformation matrix. One vector; The generalized eigenvalues ​​are obtained by solving formula (3); （3）； in, For the first The feature values ​​of a slow-trend feature, express The covariance matrix of the average first derivative after extracting the trend component. express The covariance matrix; It is a linear transformation matrix The first in List; Step 2.2: Divide the slow trend feature matrix into main slow trend feature vectors. With secondary trend slow feature vector ; The first The trend slow feature vector of each sample is denoted as... , slow trend feature vector Divided into main trend slow feature vectors With secondary trend slow feature vector ;in, The first d main trend slow feature vectors represent the core information of the system; The remaining md minor trend slow feature vectors are denoted by d, where d is the number of principal components and the value of d is determined by the cumulative contribution rate method. Step 2.3: Construct the main monitoring statistics respectively. and secondary monitoring statistics ; The following formulas (4) and (5) are used to indicate the changes in the system using two statistics; （4）； （5）； This represents the main monitoring statistics constructed based on key features. This represents a secondary monitoring statistic constructed based on secondary features. and These represent the main trend slow feature vectors respectively. and secondary trend slow feature vectors The transpose of the statistic; Step 2.4, use the kernel density estimation method to determine the control limits of the primary monitoring statistic and the secondary monitoring statistic. and ; Step 3, online anomaly detection; Collecting online data from chemical processes After preprocessing, the data is input into the anomaly detection model to calculate the monitoring statistics corresponding to the online data. and If the monitored statistics exceed the control limit, an anomaly is determined to have occurred. Step 4: Calculate the deviation indicators of the quality variables; When the statistical value exceeds the control limit and an alarm is triggered, the data acquisition window width is [value missing]. Online key quality variable data Calculate the deviation of quality variables This leads to the construction of the first dimension of a multi-dimensional causal correlation indicator, namely the deviation indicator of the quality variable. ; Step 5: Isolate outliers and determine their causal location; Collection time Online data Calculate online data The contribution of each variable is analyzed to identify outliers; based on the process mechanism, the causal position of each outlier is determined. Step 6: Based on the transitive entropy method, determine the causal attributes of the outlier variables and the causal propagation indicators of the outlier variables; Calculate the standard phase transfer entropy between each outlier variable to obtain the standard phase transfer entropy matrix. This allows us to determine the causal attributes of each outlier variable. ; Calculate the causal propagation markers for each outlier variable using the standard phase transfer entropy matrix. ; Step 7: Constructing multi-dimensional causal relationship indicators; Comprehensive utilization of quality variable deviation indicators Causal position of outlier variables causal attributes of abnormal variables and the causal propagation markers of anomalous variables Construct multi-dimensional causal relationship indicators.

2. In step 1, the normal training dataset is processed using formula (1). After preprocessing, the expression for formula (1) is: （1）； In the formula, For normal training dataset The mean, For normal training dataset The standard deviation.

3. Step 3 specifically includes the following steps: Step 3.1, data collection Real-time online data The online data obtained after preprocessing the online data using formula (6) ; （6）； in, express Real-time online data samples, express Online data obtained after preprocessing online data in real time; Step 3.2, calculate the online data using formula (2). slow trend feature vector ; Step 3.3, convert the slow trend feature vector Divided into two parts and ; Anomaly detection is performed using a trend slow feature analysis detection model. The main monitoring statistics and secondary monitoring statistics corresponding to the online data are calculated according to formulas (7) and (8). （7）； （8）； in, , They represent Real-time primary and secondary monitoring statistics calculated from constantly online data; , express Slow feature vector of time trend The main and secondary parts are divided into two parts; , They represent , transpose; Step 3.4: Match the primary monitoring statistics and secondary monitoring statistics corresponding to the online data with their respective control limits. and If any value exceeds the control limit, it is considered abnormal.

4. Step 4 specifically includes the following steps: Step 4.1, calculate the time using formula (9). At the time Deviation of quality variables in quality variable data ; （9）； in, For reference values ​​of quality variables, for With a window width of Mean within the window Tolerance; Step 4.2, use formula (10) to represent the deviation indicator of the quality variable. The judgment logic; （10）； when This indicates that the value of the quality variable fluctuates within an acceptable range; an anomaly at this point may be due to a new operating condition or sensor malfunction. If the fault is not found, it is directly identified as a process failure, and no further analysis is required.

5. Step 5 specifically includes the following steps: Step 5.1: Using contribution analysis, calculate the contribution value of each variable to the main monitoring statistic; When the statistic exceeds the control limit alarm and the quality variable deviation flag equals 1, the anomaly type is further identified: contribution analysis is used to calculate... The contribution value of each variable is used to locate outlier variables by analyzing the magnitude of the contribution value. , , , The number of outlier variables; at time [time]. ,variable Main monitoring statistics Contribution value Calculated using formula (11); (11)； in, express At that moment, the Measured values ​​of process variables; Furthermore, the contribution of the variable is expressed by formula (12); (12)； in, It is a slow trend feature analysis projection matrix The Row vectors; Represents the first in the projection matrix One vector; Calculate using formula (13) Each variable affects the main monitoring statistics. The mean of the sum of contributions ; (13)； Variables whose contribution value exceeds the average contribution value of all variables are identified as outliers. Step 5.2: Identify variables whose contribution values ​​exceed the average contribution value of all variables as outliers; Will More than The variable was identified as an anomalous variable. , ; Step 5.3, determine the first step based on the process mechanism relationship. Causal location of an outlier variable , If the variable If the input or stimulus is external to the system and is not directly controlled by internal state feedback, then define the variable. Let be the input variable. If the variable is an internal state or output variable of the process, then let The judgment logic is determined by formula (14); (14)。 6. Step 6 specifically includes the following steps: Step 6.1, based on the preset window Using online data within the dataset, calculate the standard phase transfer entropy between each abnormal variable and construct the standard phase transfer entropy matrix; The width of the acquisition window is Online data Calculate each abnormal variable Between moments At the time Standard phase transfer entropy within the window As shown in formula (15); （15）； in, and abnormal variables Any two variables in abnormal variables to abnormal variables Phase transfer entropy, abnormal variables to abnormal variables The phase transfer entropy, ignoring those less than 0.5 The value is used to obtain the standard phase transfer entropy matrix. This indicates that abnormal information is passed from row variables to column variables, using the standard phase transfer entropy matrix. As shown in formula (16); （16）； Step 6.2: Based on the standard phase transit entropy matrix, calculate the causal propagation degree of each anomalous variable. ; At any moment At the time In the window, the standard phase transfer entropy matrix is ​​used to calculate the first phase using formula (17). Causal propagation of an outlier variable ; （17）； Step 6.3, Formula (18) represents the first... A causal propagation marker for an anomalous variable; （18）； Step 6.4: If the anomalous variable satisfies the condition of passing information to other variables without being affected by the information passing of other variables, then the anomalous variable has a causal attribute. ,otherwise .

7. In step 7, the deviation indicators of quality variables are comprehensively utilized. Causal position of outlier variables causal attributes of abnormal variables = Causal propagation degree of outliers Four dimensions to construct multi-dimensional causal relationship indicators As shown in formula (20); (20)； in, The number of abnormal variables. For sign functions; when When the quality has not deviated and the source variable causing the anomaly is an input variable with causal propagation, it meets the characteristics of a new operating condition, and the anomaly type is determined to be a new operating condition; when When this occurs, the exception type is determined to be a process failure.

Images