Process technological parameter deviation risk quantitative analysis method

Abstract

The invention provides a process technological parameter deviation risk quantitative analysis method, and relates to the technical field of risk quantification.The process technological parameter deviation risk quantitative analysis method comprises the steps that firstly, a naive Bayesian model is trained through historical data, real-time data are input into the trained naive Bayesian model, and reasoning offset is obtained; adjusting the initial reference range of each process parameter based on the reasoning offset, obtaining the deviation through the real-time data and the adjusted reference range, calculating the deviation degree based on the deviation, and judging whether the risk exists or not according to the deviation degree. According to the method and the device, the process parameter deviation risk can be quantified by utilizing historical data.

Description

Technical Field

[0001] The present application relates to the technical field of risk quantification, and in particular to a method for quantitatively analyzing the risk of deviation of process parameters. Background Art

[0002] In modern industrial production, the stability and accuracy of process parameters are crucial for ensuring product quality, improving production efficiency, and ensuring production safety. However, in actual production, process parameters often deviate due to factors such as equipment aging, fluctuations in raw material quality, and operator errors. This can lead to reduced product quality, decreased production efficiency, and even safety accidents.

[0003] Traditional process parameter monitoring methods rely primarily on manual inspections and periodic testing. This approach is not only inefficient but also makes it difficult to promptly detect and address abnormal fluctuations in process parameters. Furthermore, with the continuous development of industrial automation and information technology, a large amount of historical data is being collected and stored. This historical data contains rich information on process parameter changes, but traditional monitoring methods often cannot fully utilize this historical data for in-depth analysis and prediction.

[0004] Therefore, how to effectively utilize these historical data, conduct quantitative analysis on the deviation risk of process parameters, and issue timely warnings when potential risks occur has become a technical problem that needs to be urgently solved in current industrial production. Summary of the Invention

[0005] In order to quantify the risk of process parameter deviation using historical data, the present application provides a process parameter deviation risk quantification analysis method.

[0006] This application provides a method for quantitative analysis of process parameter deviation risk, which adopts the following technical solutions: A method for quantitatively analyzing process parameter deviation risk includes the following steps: First collection: collect real-time data and historical data of various process parameters, as well as the historical offset corresponding to the historical data; Setting scope: including first setting, modeling and training, reasoning and first calculation; First setting: setting the initial reference range of various process parameters; Modeling and training: Establish a Naive Bayes model and train it using historical data and historical offsets to obtain a trained Naive Bayes model. Inference: Input real-time data into the trained Naive Bayes model to obtain the inference offset; First calculation: sum the initial reference range and the inference offset to obtain the final reference range; Second calculation: Calculate the deviation of various process parameters based on real-time data and the final reference range, standardize each deviation, and obtain the standardized deviation; The third calculation: Based on the normalized deviation, the deviation is calculated using the weighted average algorithm; First judgment: judging whether the deviation is greater than a preset deviation threshold, if so, executing the alarm step; if not, executing the first collection step; Alarm: Send an alarm signal.

[0007] This application establishes a naive Bayesian model and trains it using historical data and its corresponding historical offsets, so that it has the ability to predict the offset of real-time data. The real-time data is then input into the trained naive Bayesian model to obtain the inference offset, and the initial reference range is summed with the inference offset to obtain the final reference range. This application adjusts the reference range by inference offset, which can adapt to the dynamic changes of the process technology. The deviation of various process parameters is then calculated based on the real-time data and the final reference range, and each deviation is standardized. Based on the standardized deviation, the deviation is calculated using a weighted average algorithm. It is then determined whether the deviation is greater than the preset deviation. When the deviation is greater than the preset deviation, an alarm signal is issued. Otherwise, the first acquisition step is continued, thereby forming a closed-loop monitoring. Through the above scheme, this application can not only use historical data to quantify the risk of process parameter deviation, but also issue an alarm in a timely manner when potential risks occur.

[0008] Optionally, after performing the modeling and training steps and before performing the inference step, the method further includes: First acquisition: Based on the trained Naive Bayes model, obtain the conditional probability of historical data; Second acquisition: obtaining the prior probability of historical data; Fourth calculation: Calculate the correlation between historical data based on conditional probability and prior probability; Second judgment: determine whether each correlation degree is less than the preset correlation degree threshold. If so, execute the reasoning step; if not, execute the dimensionality reduction step; Dimensionality reduction: Use the principal component analysis algorithm to reduce the dimensionality of historical data to obtain the reduced-dimensional historical data. Use the reduced-dimensional historical data as new historical data and perform modeling and training steps.

[0009] This application is based on the trained Naive Bayes model to obtain the conditional probability of historical data. The conditional probability can help understand the dependency relationship between historical data, and then obtain the prior probability of historical data. By obtaining the conditional probability and prior probability, this application can more deeply understand the distribution and characteristics of historical data. The correlation between historical data is then calculated based on the conditional probability and the prior probability, which helps to identify which historical data have strong correlations. It is then judged whether each correlation is less than the preset correlation threshold. If so, it indicates that these historical data have a high degree of independence, and then the reasoning step is performed; otherwise, it indicates that there is a high correlation between each historical data. Highly correlated data often contains redundant information. Principal component analysis eliminates this redundant information by converting the original variables into principal components, reduces the dimension of the historical data, and retains the main information of the historical data as much as possible, thereby improving the training efficiency and prediction accuracy of the model. In addition, the data after dimensionality reduction is easier to process and analyze, which helps to reduce computing costs.

[0010] Optionally, after performing the first acquisition step and before performing the range setting step, the method further includes: Sorting: Sort the historical offsets of the i-th process parameter in descending order to obtain the first historical offset sequence, and select the j-th historical offset as the standard value; Data partitioning: historical offsets greater than the standard value are integrated into the first data set, and the remaining historical offsets are recorded as the second data set; Sixth calculation: calculating the difference between the first data set and the second data set; Difference judgment: judging whether the difference value is greater than a preset difference threshold, if so, executing the step of setting the range; if not, executing the step of updating the offset; Update the offset: Use the j+1th historical offset as the new standard value and perform the data partitioning step.

[0011] This application sorts the historical offsets of the i-th process parameter from largest to smallest and selects the j-th historical offset as the standard value, which helps identify the main distribution patterns and extreme values ​​of the historical offsets. Based on the selected standard value, the historical offsets are then divided into two data sets: the first data set contains historical offsets greater than the standard value, and the second data set contains the remaining historical offsets. The difference between the first and second data sets is then calculated, reflecting the distribution differences between the two data sets. The difference is then determined to be greater than a preset difference threshold, thereby assessing whether the difference between the two data sets is significant (thus distinguishing historical offsets within the normal range (historical offsets within the range of the second data set) from historical offsets within the abnormal range (historical offsets within the range of the first data set). If the difference does not meet the preset condition (i.e., is not greater than the preset difference threshold), the standard value is updated to the j+1-th historical offset, and the data partitioning step is repeated, helping to more accurately identify and process the characteristics of the historical offsets. By more deeply analyzing the distribution characteristics and differences of the historical offsets, the application can more accurately set the range of the process parameters, thereby improving the accuracy and reliability of subsequent steps.

[0012] Optionally, after performing the step of determining the difference and before performing the step of setting the range, the method further includes: First retention: retain the standard value corresponding to the difference value greater than the preset difference threshold, and record it as the first data; Get the maximum value: record the historical offset corresponding to the largest first data as the first offset; In the first setting step, the difference between the upper and lower limits of the initial reference range is no more than twice the first offset.

[0013] The present application retains the standard value corresponding to when the difference value is greater than the preset difference threshold, that is, it retains the standard value (that is, the first data) that can accurately distinguish the historical offsets within the normal range and the historical offsets within the abnormal range. Then, the largest first data and its corresponding historical offset (that is, the first offset) are selected among all the first data. Then, in the first setting step, the initial reference range is set according to the first offset, so that the difference between the upper and lower limits of the initial reference range is not greater than twice the first offset, that is, the initial reference range is set to a certain value plus or minus a first offset, so that the initial reference range can include the historical offsets within the normal range, and exclude the historical offsets within the abnormal range, which helps to identify those data points that are beyond the normal range, that is, outliers.

[0014] Optionally, after the step of performing the inference and before the step of performing the first calculation, the method further includes: Reasoning judgment: judging whether the reasoning offset is less than the first offset, if so, executing the first calculation step; if not, executing the feedback step; Feedback: Send feedback signals to operation and maintenance personnel, including real-time data and inference offsets of various process parameters.

[0015] After the reasoning step, the present application first determines whether the reasoning offset is less than the first offset, that is, less than the maximum historical offset within the normal allowable range. Through reasoning and judgment, the present application can screen out those situations where the reasoning results are reasonable or acceptable, which helps to improve the accuracy and reliability of subsequent steps. If the reasoning offset is greater than the first offset, it means that the reasoning result is abnormal or unreasonable (the current reasoning offset falls into the category of abnormal offsets, indicating that there is a problem with the current process parameters). At this time, the subsequent calculation steps will not be continued, but the feedback step will be turned to help the operation and maintenance personnel take corresponding measures to intervene or adjust based on the current situation. The feedback signal contains real-time data and reasoning offsets of various process parameters, so that the operation and maintenance personnel can have a more comprehensive understanding of the system status and reasoning process, which helps to improve transparency and traceability.

[0016] Optionally, after performing the third calculation step and before performing the first judgment step, the method further includes: Seventh calculation: Calculate the historical deviations of various process parameters based on historical data and the final benchmark range, standardize each historical deviation to obtain the standardized historical deviation, and calculate the historical deviation degree using a weighted average algorithm based on the standardized historical deviation; Calculating the standard deviation: Calculating the standard deviation of the historical deviation, and performing addition and subtraction operations on the preset deviation and the standard deviation to obtain the value range of the preset deviation; In the first judgment step, it is determined whether the deviation is greater than the maximum preset deviation in the value range. If so, the alarm step is executed; if not, the first collection step is executed.

[0017] This application calculates the historical deviations of various process parameters based on historical data and the final benchmark range, standardizes these historical deviations to obtain standardized historical deviations, and uses a weighted average algorithm to calculate the historical deviation based on these standardized historical deviations. The historical deviation calculated by the weighted average algorithm in this application can be used as an indicator for comprehensively evaluating the degree of deviation of historical data, which helps to more comprehensively understand the fluctuations of historical data. This application has higher adaptability and flexibility by calculating the standard deviation and setting the preset deviation value range accordingly. Compared with a fixed threshold, the use of a dynamically set preset deviation value range can more effectively reduce false alarms and omissions, and improve stability and reliability.

[0018] Optionally, after performing the first judgment step and before performing the alarm step, the method further includes: Third judgment: judging whether the deviation is less than the minimum value in the preset deviation range, if so, executing the first acquisition step; if not, executing the early warning step; Early warning: Send an early warning signal to the operation and maintenance personnel, the early warning signal including: the current deviation and the process parameters corresponding to the current deviation.

[0019] When the deviation is less than the minimum value in the preset deviation range, it means that the current deviation is completely allowed to exist. On the contrary, it means that although the current deviation is within the normal range, there is a certain risk of developing into an abnormal range. In this case, the early warning step is executed to notify the operation and maintenance personnel in advance before the problem becomes serious, so that the operation and maintenance personnel have enough time to take measures to prevent the problem from further deteriorating. Optionally, after performing the seventh calculation step and before performing the standard deviation step, the method further includes: First construction: constructing a deviation matrix based on the historical deviations, wherein the element in the i-th row and j-th column of the deviation matrix represents the j-th historical deviation of the i-th process parameter; The eighth calculation is to calculate the covariance matrix of the deviation matrix, solve the eigenvalue of the covariance matrix, and obtain the eigenvector when the eigenvalue is the largest; Product operation: Calculate the product of the deviation matrix and the eigenvector to obtain the matrix after operation, and use each element in the matrix as the new historical deviation.

[0020] This application constructs a deviation matrix based on historical deviations. The elements in the i-th row and j-th column of the matrix represent the j-th historical deviation of the i-th process parameter. The construction of the deviation matrix enables the historical deviation data to be organized and stored in an orderly manner, and provides a basis for multivariate analysis, so that the subsequent steps can simultaneously consider the historical deviations of multiple process parameters. Then, by solving the eigenvalues ​​and eigenvectors of the covariance matrix, the main direction of change in the deviation is extracted. The eigenvector with the largest eigenvalue represents the maximum direction of data change, that is, the most representative deviation change pattern. Then, the product of the deviation matrix and the eigenvector is calculated to obtain the matrix after operation, and each element in the matrix is ​​used as the new historical deviation. Through the product operation, the original deviation is converted to a new space defined by the eigenvector. The data in the new space (i.e., the new historical deviation) can better reflect the main change pattern of the data, thereby obtaining more representative and accurate results.

[0021] Optionally, after performing the third calculation step and before performing the seventh calculation step, the method further includes: Fourth judgment: judging whether the quantity of historical data of various process parameters is the same, if so, executing the first construction step; if not, executing the data generation step; Generate data: Build a generative adversarial network, train it with current historical data, obtain the trained generative adversarial network, and use the trained generative adversarial network to generate new historical data; Data completion: Use the new historical data in the data generation step to complete the historical data of various process parameters.

[0022] This application determines whether the quantity of historical data for various process parameters is the same. If so, the first construction step is executed. If not, the data generation step is executed. This ensures that the input data for subsequent steps (i.e., the historical data for various process parameters) is consistent in quantity, helping to reduce potential issues caused by inconsistent data quantity. If the historical data quantity is inconsistent, this application constructs and trains a generative adversarial network (GAN). The trained GAN is then used to generate new historical data. This GAN-generated data fills in missing data for various process parameters, ensuring consistency in quantity. This facilitates the smooth execution of subsequent steps (such as the first construction and the seventh calculation). The data generated by the GAN has a certain degree of diversity, which helps enrich the dataset and improve the accuracy and reliability of data analysis. The new historical data generated in the data generation step is then used to complete the historical data for various process parameters, ensuring that they are completely consistent in quantity. This data completion step improves the completeness of the historical data for various process parameters and provides reliable input for subsequent steps.

[0023] Optionally, after executing the step of generating data and before executing the step of completing data, the method further includes: Ninth calculation: using the Pearson correlation coefficient algorithm to calculate the correlation between the kth new historical data in the step of generating data and each historical data collected in the first collecting step; Fifth judgment: judging whether the correlation is greater than a preset correlation threshold, if so, executing the iterative step; if not, executing the second reserved step; Iteration: Delete the k-th new historical data, use the k+1-th new historical data as the new k-th historical data, and execute the ninth calculation step; Second retention: retain the kth new historical data.

[0024] This application uses the Pearson correlation coefficient algorithm to calculate the correlation between the kth new historical data in the data generation step and each historical data collected in the first collection step, thereby quantifying the degree of linear correlation between the kth new historical data and the original historical data, which helps to understand the quality of the newly generated data, that is, their similarity or consistency with the original data. It is then determined whether the correlation calculated in the ninth calculation step is greater than a preset correlation threshold. By setting the preset correlation threshold, newly generated data with low correlation with the original historical data can be screened out, which helps to improve the data quality in the data completion step.

[0025] In summary, this application includes at least one of the following beneficial technical effects: 1. This application establishes a naive Bayesian model and trains it using historical data and its corresponding historical offsets, enabling it to predict the offset of real-time data. The real-time data is then input into the trained naive Bayesian model to obtain the inference offset. The initial reference range and the inference offset are summed to obtain the final reference range. This application adjusts the reference range by the inference offset, which can adapt to the dynamic changes of the process technology. The deviations of various process parameters are then calculated based on the real-time data and the final reference range, and each deviation is standardized. Based on the standardized deviations, a weighted average algorithm is used to calculate the deviation. It is then determined whether the deviation is greater than a preset deviation. If the deviation is greater than the preset deviation, an alarm signal is issued. Otherwise, the first acquisition step is continued, thereby forming a closed-loop monitoring system. Through the above scheme, this application can not only use historical data to quantify the risk of process parameter deviation, but also issue an alarm in a timely manner when potential risks occur.

[0026] 2. This application obtains the conditional probability of historical data based on the trained Naive Bayes model. The conditional probability can help understand the dependency relationship between historical data, and then obtains the prior probability of historical data. By obtaining the conditional probability and prior probability, this application can more deeply understand the distribution and characteristics of historical data. The correlation between historical data is then calculated based on the conditional probability and prior probability, which helps to identify which historical data have strong correlations. It is then judged whether each correlation is less than the preset correlation threshold. If so, it indicates that these historical data have a high degree of independence, and then the reasoning step is performed; otherwise, it indicates that there is a high correlation between the various historical data. Highly correlated data often contains redundant information. Principal component analysis eliminates this redundant information by converting the original variables into principal components, reducing the dimension of the historical data while retaining the main information of the historical data as much as possible, thereby improving the training efficiency and prediction accuracy of the model. In addition, the data after dimensionality reduction is easier to process and analyze, which helps to reduce computing costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 is a flow chart of Example 1 of the present application; Figure 2 This is a flow chart of Example 2 of the present application; Figure 3 is a flow chart of Example 3 of the present application; Figure 4 This is a flowchart of Example 4 of the present application. DETAILED DESCRIPTION

[0028] The following combination Figures 1 to 4 This application is described in further detail.

[0029] Example 1: This example discloses a method for quantitative analysis of process parameter deviation risk, referring to Figure 1 The method includes: S1 first acquisition, S2 range setting, S3 deviation calculation, S4 first judgment and S5 alarm. First, real-time data and historical data of various process parameters and historical offsets corresponding to the historical data are collected. Then, initial reference ranges of various process parameters are set. Then, the initial reference ranges are optimized using the offsets to obtain final reference ranges. Then, deviations are calculated using the real-time data and the final reference ranges, and the deviation is calculated based on the deviation. If the deviation is greater than a preset deviation threshold, an alarm signal is issued. The process of this embodiment is as follows: S1 is the first collection, which collects the real-time data and historical data of various process parameters, as well as the historical offset corresponding to the historical data.

[0030] The various process parameters refer to any of temperature, pressure, flow rate and other parameters.

[0031] Historical offset refers to the offset data of the historical data of each process parameter relative to the reference value (or reference range), that is, the value that needs to be corrected or compensated, which can be positive or negative.

[0032] S2 setting scope, including S21 first setting, S22 modeling and training, S23 reasoning and S24 first calculation.

[0033] S21 is the first setting, which sets the initial reference ranges of various process parameters.

[0034] S22 modeling and training: establish a naive Bayes model and use historical data and historical offsets to train the naive Bayes model. Historical data is used as training data, and historical offsets are used as training labels (i.e., an offset is a classification label), so that it can learn the relationship between the data and offsets of process parameters and obtain the trained naive Bayes model.

[0035] S23 inference inputs the real-time data into the trained Naive Bayes model. The trained Naive Bayes model infers the classification label corresponding to the real-time data based on the input data, that is, the inference offset.

[0036] In step S24 , a first calculation is performed to add the upper and lower limit values ​​of the initial reference range to the inferred offset value, and the calculation result is used as the final reference range.

[0037] S3 calculates the deviation, including S31 second calculation and S32 third calculation.

[0038] S31 is the second calculation, which calculates the deviations of various process parameters based on the real-time data and the final reference range, and standardizes each deviation to obtain the standardized deviation.

[0039] The deviation of various process parameters is calculated based on the real-time data and the final benchmark range: determine whether the real-time data is within the final benchmark range. If the real-time data is within the final benchmark range, it indicates that the deviation is 0. Otherwise, calculate the difference between the real-time data and the upper limit or lower limit of the final benchmark range respectively, and take the difference with the smaller absolute value as the deviation of the process parameter.

[0040] In other embodiments, calculating the deviation of various process parameters based on the real-time data and the final reference range may also include determining whether the real-time data is within the final reference range. If the real-time data is within the final reference range, calculating its deviation from the midpoint of the final reference range. If the real-time data exceeds the final reference range, calculating its deviation from the upper or lower limit of the final reference range. The deviation may also be a relative value from the final reference range (indicating a percentage deviation).

[0041] For example, if the final reference range of temperature is 200-250°C and the real-time temperature is 260°C, the deviation can be +10°C (absolute value) or +20% (relative value, calculated relative to the range of 50°C).

[0042] S32 is a third calculation, which uses a weighted average algorithm to calculate the deviation based on the standardized deviation.

[0043] S4 is a first judgment, which determines whether the deviation is greater than a preset deviation threshold. If so, S5 is executed to give an alarm; if not, S1 is executed to collect the first data.

[0044] S5 alarm, sends an alarm signal.

[0045] The following further explains this embodiment by taking the example of a chemical plant production line requiring quantification of process parameter deviation risks for three process parameters, namely temperature, pressure and flow.

[0046] S1 first collection: Sensor deployment: Temperature sensors, pressure sensors, and flow sensors are installed on the reactor to collect real-time data on three process parameters: temperature, pressure, and flow. The results of simultaneous measurements by various sensors are: real-time temperature of 270°C, real-time pressure of 1.5 MPa, and real-time flow of 5 m³ / h.

[0047] Historical data acquisition: The historical data of temperature, pressure and flow rate for the past year and the deviation of these data from their respective baseline values ​​(historical offset) are acquired from the factory database.

[0048] S21 First Setting: Set the initial reference ranges for temperature, pressure, and flow rate based on production process requirements and historical data analysis. For example, the initial reference range for temperature is set to 200-250°C, the initial reference range for pressure is set to 1-2 MPa, and the initial reference range for flow rate is set to 10-20 m³ / h.

[0049] S22 Modeling and Training: Establish a naive Bayes model and train the naive Bayes model using historical data and historical offsets to obtain a trained naive Bayes model.

[0050] S23 Inference: The real-time collected temperature, pressure, and flow data are input into the trained Naive Bayes model. The trained Naive Bayes model infers the temperature, pressure, and flow offsets based on the input data. The inferred temperature offset is +5°C, the pressure offset is 0.1 MPa, and the flow offset is 1 m³ / h.

[0051] S24 First Calculation: Add the initial reference range to the inferred offset to obtain the final reference ranges for temperature, pressure, and flow. The final reference range for temperature is 205-255°C, the final reference range for pressure is 1.1-2.1 MPa, and the final reference range for flow is 11-21 m³ / h.

[0052] S31 Second calculation: Taking the absolute value calculation method as an example, the temperature deviation is +15°C, the pressure deviation is 0, and the flow deviation is -6 m³ / h.

[0053] 15, 0, and -6 are standardized. Taking the Z-score standardization method as an example, the standardized deviations are 1.36, -0.34, and -1.02, respectively.

[0054] S32 Third calculation: According to the importance of temperature, pressure and flow in the production process, corresponding weights are assigned to them (such as temperature 0.5, pressure 0.3, flow 0.2), and the overall deviation is calculated using the weighted average algorithm: 0.374.

[0055] S4 first judgment: compare the deviation with the preset deviation threshold. If the deviation is greater than the preset deviation threshold, execute S5 alarm. Otherwise, execute S1 first collection again.

[0056] This embodiment collects real-time and historical data of process parameters such as temperature, pressure, and flow, as well as their historical offsets, to first establish an initial baseline range. A naive Bayesian model is then used to train and reason with historical data to adjust the final baseline range. Next, the standardized deviation between the real-time data and the final baseline range is calculated, and the overall deviation is derived through weighted averaging. When this deviation exceeds a preset threshold, an alarm signal is triggered, effectively monitoring the process parameters and quantifying the risk of process parameter deviation.

[0057] Example 2: Reference Figure 2 The difference between this embodiment and embodiment 1 is that after executing S22 modeling and training and before executing S23 reasoning, the following is further included: S221 first obtains: the trained naive Bayes model will form a conditional probability table, through which the conditional probability of historical data can be directly obtained.

[0058] S222 second acquisition, using a statistical analysis algorithm to calculate the prior probability of each historical data based on the historical data.

[0059] S223 The fourth calculation is to calculate the correlation between historical data based on conditional probability and prior probability. In probability theory and machine learning, using conditional probability and prior probability to calculate the correlation between samples is a prior art. This embodiment is explained using mutual information as an example.

[0060] Calculate the joint probability distribution: Calculate the joint probability distribution P(x,y) of historical data A and historical data B.

[0061] ; in, is the probability that historical data y exists under the condition that historical data x exists, is the marginal probability distribution of historical data y. is the probability that historical data x exists under the condition that historical data y exists, is the marginal probability distribution of historical data x.

[0062] Calculate marginal probabilities: ; .

[0063] Calculate the correlation: .

[0064] S224 is the second judgment, which determines whether each correlation is less than the preset correlation threshold. If so, it indicates that the correlation between the historical data is not high, and then S23 reasoning is executed; if not, it indicates that the correlation between the historical data is high, and S225 dimensionality reduction needs to be executed to reduce the dimensionality of these historical data.

[0065] S225 dimensionality reduction uses the principal component analysis algorithm to reduce the dimensionality of historical data to obtain the historical data after dimensionality reduction. The historical data after dimensionality reduction is used as the new historical data, and S22 modeling and training are performed.

[0066] This embodiment first counts the frequency or distribution of each historical data under different historical offsets to obtain the estimated value of the conditional probability, and stores these estimated values ​​in the conditional probability table for use in the prediction stage. Secondly, the statistical analysis algorithm is used to calculate the prior probability of each historical data based on the historical data. Next, the correlation between the historical data is calculated. Then, it is determined whether each correlation is less than the preset correlation threshold. If it is less than, it indicates that the correlation between the historical data is not high, and the subsequent reasoning steps are continued; if it is greater than or equal to, it indicates that the correlation between the historical data is high, and it is necessary to use the principal component analysis algorithm to reduce the dimension of the historical data to obtain the historical data after dimensionality reduction, and use these data as new historical data to re-execute the modeling and training process to speed up the training speed of the naive Bayes model.

[0067] Example 3: Reference Figure 3 The difference between this embodiment and embodiment 1 is that after executing S1 first acquisition and before executing S2 setting range, it further includes: S11 sorts the historical offsets of the i-th process parameter in descending order to obtain a first historical offset sequence, and selects the j-th historical offset as a standard value.

[0068] S12 data partitioning divides all historical offsets into two groups: historical offsets greater than the standard value are integrated into the first data set, and the remaining historical offsets are integrated into the second data set.

[0069] S13: Sixth calculation: calculating the difference between the first data set and the second data set.

[0070] The calculation model of the difference value between the first data set and the second data set can be: ; in, is the difference between the first and second datasets; A is the number of historical offsets in the first dataset; B is the number of historical offsets in the second dataset; is the average value of historical offsets in the first dataset; is the average of the historical offsets in the second dataset.

[0071] In other embodiments, the calculation model of the first data set and the second data set may also be: ; in, is the difference between the first and second datasets; A is the number of historical offsets in the first dataset; B is the number of historical offsets in the second dataset; is the average value of historical offsets in the first dataset; is the average value of historical offsets in the second dataset; is the average of all historical offsets.

[0072] S14 performs a difference determination, determining whether the difference value in the sixth calculation in S13 is greater than a preset difference threshold. If so, it indicates that the current difference value can distinguish between historical offsets within a normal range and historical offsets within an abnormal range, and then S16 performs the first retention. If not, it indicates that the current difference value is insufficient to distinguish between historical offsets within a normal range and historical offsets within an abnormal range, and S15 performs the offset update.

[0073] S15 updates the offset, takes the j+1th historical offset as the new standard value, and executes S12 data division until a preset stop condition is met. The preset stop condition can be reaching a preset number of iterations or all historical offsets have been processed.

[0074] S16 first retention, retaining the standard value corresponding to when the difference value in S14 difference judgment is greater than the preset difference threshold, and recording it as the first data.

[0075] S17 obtains the maximum value, and records the historical offset corresponding to the maximum value of the first data as the first offset, so as to obtain a threshold value that can distinguish the historical offset within the normal range from the historical offset within the abnormal range to the greatest extent, that is, the first offset.

[0076] When setting the initial reference range in the first setting S21, the difference between the upper and lower limits of the initial reference range is not greater than twice the first offset.

[0077] Then, S22 modeling and training and S23 reasoning are performed, and after S23 reasoning is performed and before S24 first calculation is performed, the following steps are further included: S18 inference judgment, judges whether the inference offset is less than the first offset. If so, it indicates that the inference offset is a difference within the normal range, and then executes S24 first calculation; if not, it indicates that the inference offset is a difference within the abnormal range, and then executes S19 feedback.

[0078] S19 feedback: sending a feedback signal to the operation and maintenance personnel, wherein the feedback signal includes: real-time data of various process parameters and inference offset.

[0079] By dividing historical offsets into two sets and calculating the difference between them, this embodiment can more effectively distinguish between historical offsets within the normal range and those within the abnormal range, facilitating more accurate identification and processing of abnormal data. By implementing an inference judgment process, this embodiment monitors system status in real time and determines whether the offset is within the normal range based on the size of the inferred offset. This helps operations and maintenance personnel promptly detect abnormal conditions and take appropriate intervention measures, thereby avoiding potential risks and losses.

[0080] Example 4: Reference Figure 4 The difference between this embodiment and embodiment 1 is that after executing the third calculation in step S32 and before executing the first judgment in step S4, the following is further included: S61 is the fourth judgment, which uses statistical analysis methods to count the number of historical data of various process parameters, and then determines whether the number of historical data of various process parameters is the same. If so, execute S69 the first construction to construct a deviation matrix based on the consistent data volume; if not, execute S62 to generate data.

[0081] S62 generates data, constructs a generative adversarial network, uses the current historical data to train the generative adversarial network, obtains the trained generative adversarial network, and uses the trained generative adversarial network to generate new historical data to complete or balance the data set.

[0082] S63 is the ninth calculation, which uses the Pearson correlation coefficient algorithm to calculate the correlation between the k-th new historical data in the data generated in S62 and each historical data collected in the first collection in S1.

[0083] S64 is the fifth judgment, judging whether the correlation is greater than the preset correlation threshold. If so, it indicates that the new historical data has a high correlation with the original historical data, and it is necessary to execute S65 iteration; if not, it indicates that the new historical data has a low correlation with the original historical data, and then execute S66 second retention.

[0084] S65 is an iteration, where the k-th new historical data is deleted, the k+1-th new historical data is used as the new k-th historical data, and the ninth calculation in S63 is performed until a preset stopping condition for iteration is met.

[0085] S66 second retention, retaining the kth new historical data with low correlation with the original historical data to increase the diversity of the historical data.

[0086] S67 data completion uses the new historical data in the data generated by S62 to complete the historical data of various process parameters, so that various process parameters have the same amount of historical data.

[0087] The seventh calculation in S68 is to calculate the historical deviations of various process parameters based on the historical data and the final reference range. Using the same method as in the second calculation in S31, each historical deviation is normalized to obtain a standardized historical deviation. Based on the standardized historical deviations, a weighted average algorithm is used to calculate the historical deviation degree.

[0088] S69 is a first construction, which is to construct a deviation matrix based on the historical deviations, wherein the element in the i-th row and j-th column of the deviation matrix represents the j-th historical deviation of the i-th process parameter.

[0089] S70 is an eighth calculation, which is to calculate the covariance matrix of the deviation matrix, and solve the eigenvalue of the covariance matrix to obtain the eigenvector when the eigenvalue is the largest.

[0090] S71 product operation calculates the product of the deviation matrix and the eigenvector to obtain the matrix after operation, and uses each element in the matrix as the new historical deviation.

[0091] S72 calculates the standard deviation. Based on the new historical deviations in the product operation of S71, the average value of these historical deviations is calculated. Then, the square of the difference between each historical deviation and the average value is calculated to obtain multiple square differences. Then, the square root of the average value of these square differences is calculated to obtain the standard deviation of the historical deviations. The preset deviations are added and subtracted to the standard deviation to obtain the value range of the preset deviations.

[0092] In the first judgment of S4, it is also judged whether the deviation is greater than the maximum preset deviation. If so, the third judgment of S81 is executed; if not, the first collection of S1 is executed.

[0093] S81 is the third judgment, which determines whether the deviation is less than the minimum value in the preset deviation range. If so, it indicates that the deviation of the current process parameters is within the safe range, and then the first acquisition of S1 is executed; if not, it indicates that the deviation of the current process parameters is within the normal range, but may develop into the abnormal range, and it is necessary to execute S82 early warning.

[0094] S82 Early warning: Send an early warning signal to the operation and maintenance personnel. The early warning signal includes: the current deviation and the process parameters corresponding to the current deviation.

[0095] In this embodiment, in the data generation step S62, a generative adversarial network (GAN) is used to train and generate new historical data to complete or balance the data set so that all process parameters have the same amount of historical data. The second retention step S66 focuses on retaining new data with low correlation with the original data to increase data diversity. The seventh calculation step S68 uses standardization to eliminate the influence of different process parameter dimensions and numerical ranges, making historical deviations comparable. The eighth calculation step S70 analyzes the characteristics of the deviation matrix in the multiplication operation of step S71, and further processes the new historical deviation to obtain the preset deviation value range, which is combined with the standard deviation calculated in step S72 to provide an accurate basis for the first judgment step S4.

[0096] The above are all preferred embodiments of the present application, and are not intended to limit the scope of protection of the present application. Therefore, any equivalent changes made based on the structure, shape, and principle of the present application should be included in the scope of protection of the present application.

Claims

1. A method for quantitative analysis of process parameter deviation risk, characterized in that: include: First collection: collect real-time data and historical data of various process parameters, as well as the historical offset corresponding to the historical data; Setting scope: including first setting, modeling and training, reasoning and first calculation; First setting: setting the initial reference range of various process parameters; Modeling and training: Establish a Naive Bayes model and train it using historical data and historical offsets to obtain a trained Naive Bayes model. Inference: Input real-time data into the trained Naive Bayes model to obtain the inference offset; First calculation: sum the initial reference range and the inference offset to obtain the final reference range; Second calculation: Calculate the deviation of various process parameters based on real-time data and the final reference range, standardize each deviation, and obtain the standardized deviation; The third calculation: Based on the normalized deviation, the deviation is calculated using the weighted average algorithm; First judgment: judging whether the deviation is greater than a preset deviation threshold, if so, executing the alarm step; if not, executing the first collection step; Alarm: Send an alarm signal.

2. The process parameter deviation risk quantitative analysis method according to claim 1 is characterized in that: After the modeling and training steps, and before the inference step, the following steps are also included: First acquisition: Based on the trained Naive Bayes model, obtain the conditional probability of historical data; Second acquisition: obtaining the prior probability of historical data; Fourth calculation: Calculate the correlation between historical data based on conditional probability and prior probability; Second judgment: determine whether each correlation degree is less than the preset correlation degree threshold. If so, execute the reasoning step; if not, execute the dimensionality reduction step; Dimensionality reduction: Use the principal component analysis algorithm to reduce the dimensionality of historical data to obtain the reduced-dimensional historical data. Use the reduced-dimensional historical data as new historical data and perform modeling and training steps.

3. The process parameter deviation risk quantitative analysis method according to claim 1 or 2, characterized in that: After executing the first acquisition step and before executing the range setting step, the method further includes: Sorting: Sort the historical offsets of the i-th process parameter in descending order to obtain the first historical offset sequence, and select the j-th historical offset as the standard value; Data partitioning: historical offsets greater than the standard value are integrated into the first data set, and the remaining historical offsets are recorded as the second data set; Sixth calculation: calculating the difference between the first data set and the second data set; Difference judgment: judging whether the difference value is greater than a preset difference threshold, if so, executing the step of setting the range; if not, executing the step of updating the offset; Update the offset: Use the j+1th historical offset as the new standard value and perform the data partitioning step.

4. The process parameter deviation risk quantitative analysis method according to claim 3 is characterized in that: After executing the step of determining the difference and before executing the step of setting the range, the method further includes: First retention: retain the standard value corresponding to the difference value greater than the preset difference threshold, and record it as the first data; Get the maximum value: record the historical offset corresponding to the largest first data as the first offset; In the first setting step, the difference between the upper and lower limits of the initial reference range is no more than twice the first offset.

5. The process parameter deviation risk quantitative analysis method according to claim 4 is characterized in that: After the step of performing the inference and before the step of performing the first calculation, the method further includes: Reasoning judgment: judging whether the reasoning offset is less than the first offset, if so, executing the first calculation step; if not, executing the feedback step; Feedback: Send feedback signals to operation and maintenance personnel, including real-time data and inference offsets of various process parameters.

6. The process parameter deviation risk quantitative analysis method according to claim 1 or 2, characterized in that: After performing the third calculation step and before performing the first judgment step, the method further includes: Seventh calculation: Calculate the historical deviations of various process parameters based on historical data and the final benchmark range, standardize each historical deviation to obtain the standardized historical deviation, and calculate the historical deviation degree using a weighted average algorithm based on the standardized historical deviation; Calculating the standard deviation: Calculating the standard deviation of the historical deviation, and performing addition and subtraction operations on the preset deviation and the standard deviation to obtain the value range of the preset deviation; In the first judgment step, it is determined whether the deviation is greater than the maximum preset deviation in the value range. If so, the alarm step is executed; if not, the first collection step is executed.

7. The process parameter deviation risk quantitative analysis method according to claim 6, characterized in that: After executing the first judgment step and before executing the alarm step, the method further includes: Third judgment: judging whether the deviation is less than the minimum value in the preset deviation range, if so, executing the first acquisition step; if not, executing the early warning step; Early warning: Send an early warning signal to the operation and maintenance personnel, the early warning signal including: the current deviation and the process parameters corresponding to the current deviation.

8. The process parameter deviation risk quantitative analysis method according to claim 7, characterized in that: After performing the seventh calculation step and before performing the step of calculating the standard deviation, the method further includes: First construction: constructing a deviation matrix based on the historical deviations, wherein the element in the i-th row and j-th column of the deviation matrix represents the j-th historical deviation of the i-th process parameter; The eighth calculation is to calculate the covariance matrix of the deviation matrix, solve the eigenvalue of the covariance matrix, and obtain the eigenvector when the eigenvalue is the largest; Product operation: Calculate the product of the deviation matrix and the eigenvector to obtain the matrix after operation, and use each element in the matrix as the new historical deviation.

9. The process parameter deviation risk quantitative analysis method according to claim 8, characterized in that: After performing the third calculation step and before performing the seventh calculation step, the method further includes: Fourth judgment: judging whether the quantity of historical data of various process parameters is the same, if so, executing the first construction step; if not, executing the data generation step; Generate data: Build a generative adversarial network, train it with current historical data, obtain the trained generative adversarial network, and use the trained generative adversarial network to generate new historical data; Data completion: Use the new historical data in the data generation step to complete the historical data of various process parameters.

10. The process parameter deviation risk quantitative analysis method according to claim 9, characterized in that: After executing the step of generating data and before executing the step of filling data, it also includes: Ninth calculation: using the Pearson correlation coefficient algorithm to calculate the correlation between the kth new historical data in the step of generating data and each historical data collected in the first collecting step; Fifth judgment: judging whether the correlation is greater than a preset correlation threshold, if so, executing the iterative step; if not, executing the second reserved step; Iteration: Delete the k-th new historical data, use the k+1-th new historical data as the new k-th historical data, and execute the ninth calculation step; Second retention: retain the kth new historical data.

Images