AI-based germane purification process anomaly early warning method and system

Abstract

The application provides an AI-based germane purification process anomaly early warning method and system. First, the key operation data set and process state data set in the germane purification process are obtained, then a purification process data correlation model is constructed to obtain a correlation feature set. The model can represent the correlation between operation parameters and equipment operating state, operation parameters and germane product characteristics, and equipment operating state and germane product characteristics. Then, a pre-trained process anomaly analysis AI model is called to identify the abnormal tendency of the correlation feature set, and a preliminary identification result of the process anomaly is obtained. Then, the preliminary identification result is used to determine the type and scope of the abnormal event, and detailed information of the abnormal event is obtained. Finally, based on the detailed information of the abnormal event, a germane purification process anomaly early warning instruction containing an abnormal positioning identifier and a response measure suggestion is generated and sent to the purification process control terminal. The method can improve the stability of the germane purification process and the product quality.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to an AI-based method and system for early warning of abnormalities in germane purification processes. Background Technology

[0002] In the field of germane purification technology, germane is an important semiconductor material precursor, and its purity has a crucial impact on the performance and quality of semiconductor devices. However, the germane purification process is complex, involving multiple steps and numerous parameters, and the operating status of the purification equipment and the characteristics of the germane product change continuously as the process progresses.

[0003] Currently, monitoring and early warning systems for germane purification processes mainly rely on periodic manual inspections and simple threshold alarm systems. Periodic manual inspections are not only time-consuming and labor-intensive, but also struggle to capture subtle changes in the process in real time, failing to promptly detect potential anomalies. Simple threshold alarm systems can only judge a single parameter based on a preset fixed threshold, issuing an alarm when the parameter exceeds the threshold. However, these systems cannot consider the complex relationships between parameters, easily leading to false alarms and missed alarms. They also struggle to accurately determine the type of abnormal event in the process and its impact range, thus failing to provide operators with effective countermeasures and severely impacting the stability of the germane purification process and product quality. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide an AI-based method for early warning of abnormalities in a germane purification process, the method comprising:

[0005] Acquire a set of key operation data and a set of process status data in the germane purification process. The set of key operation data includes records of operation parameters for each step of the purification process, and the set of process status data includes records of the operating status of the purification equipment and records of the characteristics of the germane product.

[0006] Based on the key operation data set and the process status data set, a purification process data association model is constructed to obtain the association relationship feature set output by the purification process data association model. The purification process data association model is used to characterize the association relationship between operation parameters and equipment operating status, operation parameters and germane product characteristics, and equipment operating status and germane product characteristics.

[0007] The pre-trained process anomaly analysis AI model is invoked to perform anomaly tendency identification processing on the correlation feature set to obtain preliminary process anomaly identification results;

[0008] Based on the preliminary identification results of the process anomalies, the types of abnormal events in the germane purification process and the scope of impact of the corresponding purification steps are determined, and detailed information on the abnormal events is obtained.

[0009] Based on the detailed information of the abnormal event, an abnormal warning instruction for the germane purification process is generated, which includes an abnormal location identifier and suggested countermeasures, and the abnormal warning instruction for the germane purification process is sent to the purification process control terminal.

[0010] Furthermore, embodiments of the present invention also provide an AI-based early warning system for abnormalities in the germane purification process, characterized in that it includes:

[0011] A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the aforementioned AI-based anomaly warning method for germane purification process by executing the machine-executable instructions.

[0012] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions, the machine-executable instructions being stored in a computer-readable storage medium, a processor of a computer device reading the machine-executable instructions from the computer-readable storage medium, the processor executing the machine-executable instructions, causing the computer device to execute the above-mentioned AI-based method for early warning of abnormalities in the germane purification process.

[0013] Based on the above, by comprehensively acquiring key operational data sets and process status data sets in the germane purification process, covering multi-dimensional information such as operational parameter records, purification equipment operating status records, and germane product characteristic records for each purification stage, a data association model for the purification process is constructed based on this data. This model can deeply explore the complex correlations between operational parameters and equipment operating status, operational parameters and germane product characteristics, and equipment operating status and germane product characteristics, obtaining a set of correlation feature sets, thus more comprehensively and accurately reflecting the inherent laws of the process. A pre-trained process anomaly analysis AI model is then used to identify anomaly tendencies in the correlation feature set. Leveraging the powerful learning and analytical capabilities of the AI ​​model, anomaly tendencies in the process can be accurately identified, improving the accuracy and reliability of anomaly identification. Based on the preliminary identification results of process anomalies, the type of anomaly event and the corresponding impact range of the purification stage are determined, obtaining detailed information about the anomaly event, providing a clear direction for subsequent countermeasures. Finally, based on the detailed information of the abnormal event, an abnormal warning instruction for the germane purification process is generated, which includes the abnormal location identifier and the corresponding countermeasure suggestions. This instruction is then sent to the purification process control terminal. This allows for timely and accurate feedback of abnormal situations to the operators, guiding them to take effective countermeasures. This effectively improves the stability of the germane purification process and product quality, reduces production costs, and increases production efficiency. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the execution flow of the AI-based early warning method for germane purification process provided in this embodiment of the invention.

[0015] Figure 2 This is a schematic diagram of exemplary hardware and software components of the AI-based germanane purification process anomaly early warning system provided in an embodiment of the present invention. Detailed Implementation

[0016] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating an AI-based early warning method for germane purification process anomalies, provided in one embodiment of the present invention. The following is a detailed description of this AI-based early warning method for germane purification process anomalies.

[0017] Step S110: Obtain the key operation data set and process status data set in the germane purification process. The key operation data set includes the operation parameter records of each purification stage, and the process status data set includes the purification equipment operation status records and germane product characteristic records.

[0018] This embodiment uses electronic-grade germane purification technology as an application scenario. This process is used to produce high-purity germane required for semiconductor manufacturing. The key operational data set covers the operational parameters of four core stages: pretreatment, adsorption, distillation, and post-purification treatment. These parameters include feed rate and filtration pressure in the pretreatment stage, adsorption temperature and adsorbent dosage in the adsorption stage, column bottom temperature and reflux ratio in the distillation stage, and drying temperature and vacuum level in the post-purification stage. The process status data set includes equipment operation status records such as the inlet and outlet pressure difference of the filter, the wall temperature of the adsorption column, the column pressure of the distillation column, and the speed of the vacuum pump. Germane product characteristic records include the purity, impurity content (such as silane and methane content), and dew point of the products at each stage. All data is collected in real time by sensors, flow meters, and analyzers deployed on the purification equipment and transmitted to the data processing terminal via industrial Ethernet.

[0019] Step S120: Construct a purification process data association model based on the key operation data set and the process status data set, and obtain the association relationship feature set output by the purification process data association model. The purification process data association model is used to characterize the association relationship between operation parameters and equipment operating status, operation parameters and germane product characteristics, and equipment operating status and germane product characteristics.

[0020] By conducting multi-dimensional correlation analysis of key operational data and process status data, a data correlation model for the purification process is constructed. This data correlation model needs to quantify the mutual influence between different parameters, such as the correlation between the feed rate in the pretreatment stage and the pressure difference of the filter, and the correlation between the reflux ratio of the distillation column and the purity of germane. Finally, the model outputs a set of features that reflect these correlation patterns.

[0021] Step S121: Classify the operation parameter records in the key operation data set. According to the pretreatment, adsorption, distillation and post-purification processes of germanane purification, divide the operation parameter records into a pretreatment operation parameter subset, an adsorption operation parameter subset, a distillation operation parameter subset and a post-purification operation parameter subset.

[0022] The key operational data sets are categorized at the data processing terminal: the pretreatment operational parameter subset includes records of parameters such as feed rate, feed preheating temperature, filtration pressure, and filtration time; the adsorption operational parameter subset includes records of parameters such as adsorption tower temperature, adsorption pressure, adsorbent loading, and feed flow rate; the distillation operational parameter subset includes records of parameters such as distillation column bottom temperature, column top temperature, column pressure, reflux ratio, and product rate; and the post-treatment operational parameter subset includes records of parameters such as drying temperature, drying time, vacuum degree, and inert gas purging flow rate. Each subset is sorted by timestamp to ensure data temporal consistency.

[0023] Step S122: Classify the equipment operation status records and germane product characteristic records in the process status data set. Divide the equipment operation status records into pretreatment equipment status subsets, adsorption equipment status subsets, distillation equipment status subsets, and post-treatment equipment status subsets corresponding to each purification stage. Divide the germane product characteristic records into pretreatment product characteristic subsets, adsorption product characteristic subsets, distillation product characteristic subsets, and post-treatment product characteristic subsets corresponding to each purification stage.

[0024] The equipment operation status records are categorized as follows: the pretreatment equipment status subset includes records such as the pressure difference between the inlet and outlet of the filter, the wall temperature of the raw material preheater, and the operating current of the raw material pump; the adsorption equipment status subset includes records such as the wall temperature distribution of the adsorption tower, the pressure difference between the inlet and outlet of the adsorption tower, and the valve opening and closing status; the distillation equipment status subset includes records such as the pressure fluctuation of the distillation tower, the temperature distribution inside the tower, and the speed of the reflux pump; and the post-treatment equipment status subset includes records such as the pressure of the drying tank, the vacuum degree of the vacuum pump, and the reading of the inert gas flow meter.

[0025] The characteristics of germane products are categorized as follows: Pretreatment product characteristics subset includes records of the initial removal rate of impurities and water content of the pretreated raw material; Adsorption product characteristics subset includes records of the purity of the adsorbed product and the amount of target impurities adsorbed; Distillation product characteristics subset includes records of the purity, impurity content, and boiling point range of the distillate from the distillation column; Posttreatment product characteristics subset includes records of the purity (e.g., above 99.9999%), impurity content (e.g., silane content below a certain limit), and dew point of the final germane product.

[0026] Step S123: Determine the correlation dimensions between the subset of operating parameters and the corresponding subset of equipment status, the subset of operating parameters and the corresponding subset of product characteristics, and the subset of equipment status and the corresponding subset of product characteristics in each purification step. The correlation dimensions include time synchronization dimension, parameter influence dimension, and status feedback dimension.

[0027] For each purification step, three correlation dimensions are identified: the time synchronization dimension requires that the timestamps of operating parameters, equipment status, and product characteristic data be perfectly matched to ensure that the analysis focuses on parameter correlations at the same moment; the parameter impact dimension describes the degree of influence of changes in operating parameters on equipment status or product characteristics, such as the effect of changes in adsorption temperature on the wall temperature of the adsorption column; and the status feedback dimension describes the feedback effect of changes in equipment status on product characteristics, such as the feedback effect of distillation column pressure fluctuations on germane purity. For example, in the distillation step, the correlation between a subset of distillation operating parameters (such as reflux ratio) and a subset of distillation equipment status (such as column pressure) needs to cover all three dimensions. Adjustments to the reflux ratio (operating parameter) affect column pressure (equipment status) under the premise of time synchronization, while column pressure fluctuations will in turn affect germane purity (product characteristics).

[0028] Step S124: Construct a data association rule base for each purification step based on the association dimension. The data association rule base includes rules on the impact of changes in operating parameters on equipment status, rules on the impact of changes in operating parameters on product characteristics, and rules on the impact of changes in equipment status on product characteristics.

[0029] Step S1241: Collect the historical key operation data set and historical process status data set during the normal operation of the germane purification process, and clean the historical key operation data set and historical process status data set to obtain the effective historical data set.

[0030] Historical data from the past two years of normal process operation was collected and cleaned using the following methods: outliers caused by sensor malfunctions (such as pressure sudden drops to zero or temperature exceeding physical limits) were removed; missing data were filled in, with linear interpolation used for continuously missing parameters and the average of adjacent time points used for discretely missing state data; duplicate records were removed, and only data with unique timestamps were retained. This cleaned data set resulted in a valid historical dataset, ensuring the accuracy and completeness of the data.

[0031] Step S1242: Extract the historical operation parameter subset, historical equipment status subset, and historical product characteristic subset of each purification step from the effective historical data set, and align the historical operation parameter subset, historical equipment status subset, and historical product characteristic subset data of each purification step according to the time series.

[0032] Historical data subsets are extracted for each purification stage, and alignment is achieved through timestamp matching. For example, in the adsorption stage, the adsorption temperature (operating parameter), adsorption tower wall temperature (equipment status), and post-adsorption product purity (product characteristics) at a certain moment are correlated to form a time-aligned associated data record. Data from all stages are processed in this way to ensure that different types of parameters at the same time can be analyzed accordingly.

[0033] Step S1243: For each purification stage, analyze the change correspondence between the historical operating parameter subset of each purification stage and the historical equipment status subset of each purification stage. When the operating parameters change, record the change response of the equipment status, extract the influence law of the operating parameters on the equipment status, and form the influence rule of the change of operating parameters on the equipment status.

[0034] Taking the adsorption stage as an example, the relationship between the change in adsorption temperature and the wall temperature of the adsorption column is analyzed: when the adsorption temperature increases, the rise in wall temperature and the response time are recorded. It is found that for every certain range increase in adsorption temperature, the wall temperature of the adsorption column rises by a corresponding range within a certain time, and the rise is positively correlated with the temperature rise. Based on this pattern, an influence rule is formed, such as "adsorption temperature increases → adsorption column wall temperature rises within a set time, the rise is positively correlated with the temperature rise, and does not exceed the preset upper limit of wall temperature". Similarly, other stages are analyzed, such as "increased reflux ratio in distillation column → increased reflux pump speed, slight increase in column pressure".

[0035] Step S1244: Analyze the correspondence between the historical operating parameter subsets of each purification step and the historical product characteristic subsets of each purification step. When the operating parameters change, record the response of the product characteristics, extract the influence law of the operating parameters on the product characteristics, and form the influence rule of the operating parameter changes on the product characteristics.

[0036] Taking the distillation process as an example, the relationship between the reflux ratio and the purity of the top product is analyzed: when the reflux ratio increases within a reasonable range, the change in product purity is recorded. It is found that as the reflux ratio increases, the purity increases accordingly until it reaches a certain stable value; if the reflux ratio exceeds the reasonable range, the purity decreases. Based on this, rules are formed, such as "increasing the reflux ratio within a reasonable range in distillation → increasing the purity of the top product; exceeding the reasonable range in reflux ratio → decreasing the purity." Another example is the pretreatment process: "reducing the feed rate → increasing the impurity removal rate of the filtered feed."

[0037] Step S1245: Analyze the change correspondence between the historical equipment state subsets of each purification step and the historical product characteristic subsets of each purification step. When the equipment state changes, record the change response of product characteristics, extract the influence law of equipment state on product characteristics, and form the influence rule of equipment state change on product characteristics.

[0038] Taking the adsorption process as an example, the relationship between the pressure difference between the inlet and outlet of the adsorption tower and the product purity is analyzed: when the pressure difference is within the normal range, the product purity is stable; when the pressure difference increases (indicating adsorbent saturation), the product purity decreases. This forms a rule such as "pressure difference between the inlet and outlet of the adsorption tower exceeds the normal range → product purity decreases after adsorption." In the distillation process, "increased pressure fluctuation in the distillation tower → increased fluctuation in the purity of the top product."

[0039] Step S1246: Verify the effectiveness of the rules on the impact of changes in operating parameters on equipment status, the impact of changes in operating parameters on product characteristics, and the impact of changes in equipment status on product characteristics. By comparing historical data from different time periods, verify the applicability of these rules under different operating conditions. Eliminate rules whose applicable scope does not meet the set requirements and combine them to form a data association rule library for each purification stage.

[0040] Historical normal data from different seasons and raw material batches are selected as the validation set. For example, to validate the rule "adsorption temperature increases → adsorption tower wall temperature increases": under high-temperature conditions in summer and low-temperature conditions in winter, the response of the wall temperature to changes in adsorption temperature is checked to see if it conforms to the rule description. If the rule applies under most conditions, it is retained; if it only applies to specific raw material batches and the applicability rate is lower than a set threshold, it is discarded. The validated rules are categorized and stored according to pretreatment, adsorption, distillation, and post-treatment stages to form a data association rule library.

[0041] Step S125: Input the categorized subsets of operation parameters, device statuses, and product characteristics into the data association rule base, and calculate the association strength parameters under each association dimension through rule matching to form an association relationship matrix.

[0042] Input a subset of pretreatment operation parameters (such as feed rate) and a subset of pretreatment equipment status (such as filter pressure difference) of the current purification process into the data association rule library, match the influence rules of "feed rate and filter pressure difference", and calculate the association strength parameter. This association strength parameter is determined by the ratio of the change in operation parameters to the change in equipment status. For example, the number of units that the filter pressure difference changes for every unit change in feed rate is the association strength.

[0043] Calculate the correlation strength parameters of the three types of correlations (operation-equipment, operation-product, equipment-product) in each stage under the dimensions of time synchronization, parameter influence, and status feedback in this way. Arrange the above parameters by row (correlation type) and column (dimension) to form a correlation matrix. For example, the correlation matrix of the pretreatment stage includes correlation strength parameters such as feed rate-filter pressure difference and feed rate-impurity removal rate.

[0044] Step S126: Perform feature extraction processing on the association matrix, extract the distribution features, trend features and abnormal deviation features of the association strength parameters in the association matrix, and combine the distribution features, trend features and abnormal deviation features of the association strength parameters in the association matrix to form the association feature set output by the data association model of the purification process.

[0045] Extract the distribution characteristics of the correlation strength parameters: the range, mean, and variance of each correlation strength parameter in the statistical matrix, such as the mean and variance of the correlation strength between reflux ratio and purity in the distillation process, which reflect the overall stability of the correlation.

[0046] Extracting trend characteristics: Tracking the changes in correlation strength parameters by time series, such as the rising or falling trend of the adsorption temperature-wall temperature correlation strength in the continuous production process, reflecting the dynamic changes in the correlation relationship.

[0047] Extracting abnormal deviation features: Compare the current correlation strength parameters with the historical normal range to identify parameters that are out of range. For example, if the correlation strength between the feed rate in the pretreatment stage and the pressure difference of the filter suddenly exceeds three times the historical normal variance, it is marked as an abnormal deviation feature.

[0048] These three types of features are categorized and combined according to their stages to form a set of correlation features. This set of correlation features contains the distribution, trend, and deviation information of each correlation. For example, "the correlation between distillation reflux ratio and purity is: the distribution mean is a certain value, the trend is stable, and there is no abnormal deviation; the correlation between adsorption temperature and wall temperature is: the distribution variance increases, the trend is upward, and there is abnormal deviation."

[0049] Step S130: Call the pre-trained process anomaly analysis AI model to perform anomaly tendency identification processing on the correlation feature set to obtain preliminary process anomaly identification results.

[0050] The set of correlation features is input into a pre-trained AI model. The AI ​​model needs to identify the abnormal patterns hidden in the features, such as sudden changes in correlation strength and trend reversals, and make a preliminary judgment on the type and possible location of the anomalies.

[0051] Step S131: Input the set of correlation features into the correlation feature comparison module of the process anomaly analysis AI model. The correlation feature comparison module of the process anomaly analysis AI model has a pre-stored standard correlation feature template for normal purification process. The difference parameter between the set of correlation features and the standard correlation feature template is calculated by the feature comparison algorithm.

[0052] The pre-stored standard association feature templates in the association feature comparison module are constructed based on a large amount of historical normal data and include standard distribution features (such as standard mean and variance) and standard change trends (such as stable or slow change) for each association. The current set of association features is compared feature by feature with the standard template: the differences in distribution features (such as the difference between the current mean and the standard mean) and the differences in trend features (such as the difference between the current trend slope and the standard slope) are calculated. The difference parameter is obtained by weighted summation. The larger the difference, the more significant the deviation between the current feature and the normal template.

[0053] Step S132: Based on the difference parameter, filter out the abnormal association feature subset in the association feature set whose difference exceeds the preset comparison threshold. The abnormal association feature subset includes features with abnormal association between operating parameters and equipment status, features with abnormal association between operating parameters and product characteristics, and features with abnormal association between equipment status and product characteristics.

[0054] Set a preset comparison threshold (determined based on historical anomaly data statistics), and filter features whose differences exceed this threshold: for example, if the difference in the correlation between adsorption temperature and wall temperature in the adsorption process exceeds the threshold, it is classified as an operation-equipment correlation anomaly; if the difference in the correlation between reflux ratio and purity in the distillation process exceeds the threshold, it is classified as an operation-product correlation anomaly; if the difference in the correlation between pressure difference and purity in the adsorption tower exceeds the threshold, it is classified as an equipment-product correlation anomaly. Summarize the above features to form an anomaly correlation feature subset.

[0055] Step S133: Input the subset of abnormal correlation features into the abnormal pattern matching module of the process abnormality analysis AI model. The abnormal pattern matching module of the process abnormality analysis AI model has a pre-stored library of correlation feature patterns corresponding to various process abnormalities. The library of correlation feature patterns includes abnormal patterns of pretreatment, adsorption, distillation and posttreatment.

[0056] The abnormal pattern matching module stores the associated feature pattern library, which stores the feature patterns corresponding to historical abnormal events: abnormal patterns in the pretreatment stage, such as "sudden increase in the correlation strength between feed rate and filter pressure difference + decrease in impurity removal rate"; abnormal patterns in the adsorption stage, such as "reversal of the correlation trend between adsorption temperature and wall temperature + decrease in purity"; abnormal patterns in the distillation stage, such as "increased fluctuation in the correlation between reflux ratio and tower pressure + fluctuation in purity"; and abnormal patterns in the post-treatment stage, such as "abnormal correlation between drying temperature and vacuum degree + increase in dew point".

[0057] Step S134: Match the subset of abnormal related features with each abnormal pattern in the related feature pattern library using a pattern matching algorithm, calculate the matching similarity parameter, and select the target abnormal pattern with the highest matching similarity parameter.

[0058] Step S1341: Perform feature standardization processing on the abnormal association feature subset, and convert the association strength parameter distribution features, change trend features, and abnormal deviation features in the abnormal association feature subset into feature vectors of a unified format, wherein each dimension of the feature vector corresponds to a feature type.

[0059] The distribution features (mean, variance), trend features (slope, rate of change), and deviation features (deviation magnitude, duration) in the subset of abnormal correlation features are converted into numerical variables and arranged in a fixed order to form a feature vector. For example, the vector of a certain abnormal feature is [mean difference, variance difference, trend slope, deviation magnitude, duration], ensuring that the dimensional position of each feature is fixed.

[0060] Step S1342: Perform feature vector conversion on each abnormal pattern in the associated feature pattern library, and convert the associated features corresponding to each abnormal pattern in the associated feature pattern library into a pattern feature vector that is consistent with the feature vector format of the abnormal associated feature subset.

[0061] Using the same conversion logic as step S1341, the features of the "adsorption temperature-wall temperature correlation trend reversal + purity decrease" pattern in the pattern library are converted into pattern feature vectors. The dimensions of these vectors are completely consistent with the feature vectors of the abnormal correlation feature subset, such as [standard mean difference, standard variance difference, standard trend slope, standard deviation magnitude, standard duration].

[0062] Step S1343: The cosine similarity algorithm is used to calculate the cosine similarity value between the feature vector of the abnormal associated feature subset and the pattern feature vector of each abnormal pattern in the associated feature pattern library. The cosine similarity value is used to characterize the directional consistency between the feature vector of the abnormal associated feature subset and the pattern feature vector of each abnormal pattern in the associated feature pattern library in the feature space.

[0063] The cosine similarity algorithm is used to calculate the cosine of the angle between two vectors. If the abnormal feature vector and the pattern feature vector are in the same direction (e.g., both show an increasing difference in mean and a negative trend slope), the cosine similarity value is close to 1, indicating high directional consistency. If the directions are opposite, the cosine similarity value is close to -1, indicating low directional consistency.

[0064] Step S1344: The Euclidean distance algorithm is used to calculate the Euclidean distance between the feature vector of the abnormal correlation feature subset and the pattern feature vector of each abnormal pattern in the correlation feature pattern library. The Euclidean distance value is used to characterize the degree of positional difference between the feature vector of the abnormal correlation feature subset and the pattern feature vector of each abnormal pattern in the correlation feature pattern library in the feature space.

[0065] The Euclidean distance algorithm is used to calculate the straight-line distance between two vectors in the feature space. The smaller the distance, the closer the feature values ​​of the two vectors are and the smaller the positional difference; the larger the distance, the greater the positional difference. For example, if the Euclidean distance between an anomaly feature vector and a certain pattern vector is small, it means that the mean difference, deviation magnitude, and other values ​​of the two are similar.

[0066] Step S1345: Based on the cosine similarity value and the Euclidean distance value, a matching similarity parameter is obtained by weighted calculation. The weights of the cosine similarity value and the Euclidean distance value are dynamically adjusted according to historical matching accuracy data so that the matching similarity parameter can accurately reflect the feature matching degree between the feature vector of the abnormal association feature subset and the pattern feature vector of each abnormal pattern in the association feature pattern library.

[0067] Weights are assigned to the cosine similarity value and the Euclidean distance value (e.g., initial weights of 0.6 and 0.4 respectively). The matching similarity parameter is calculated as (cosine similarity value × cosine weight) + (1 - normalized Euclidean distance value × distance weight). The normalized Euclidean distance value transforms the distance value to the same range as the cosine similarity value, ensuring that the two can be weighted together. For example, if the original range of the Euclidean distance value is different from the range of the cosine similarity value, a linear transformation is used to map the Euclidean distance value to the same range, resulting in the normalized Euclidean distance value.

[0068] When dynamically adjusting weights, the system periodically collects the correspondence between matching results and actual anomaly types from historical matching data, and calculates the matching accuracy under different weight combinations. If the matching accuracy of a certain anomaly pattern is lower than a preset threshold when using the current weight combination, the weight of the cosine similarity value is increased and the weight of the Euclidean distance value is decreased, or vice versa, until the matching accuracy reaches the preset requirement. Through this dynamic adjustment mechanism, the matching similarity parameter can more accurately reflect the actual matching degree between feature vectors, avoiding matching errors caused by the bias of a single indicator.

[0069] Step S135: Based on the anomaly determination rules corresponding to the target anomaly pattern, and combined with the specific feature content in the anomaly association feature subset, generate a preliminary process anomaly identification result containing anomaly association type, anomaly feature intensity, and anomaly occurrence time clues.

[0070] The system extracts the anomaly judgment rules corresponding to the target anomaly pattern. These rules clarify the judgment conditions for different anomaly association types, the evaluation criteria for the intensity of anomaly characteristics, and the extraction method for anomaly occurrence time clues. For example, if the target anomaly pattern is "anomaly in the correlation between column pressure and reflux ratio in the distillation process," the judgment rule might be "when the correlation strength parameter between the column pressure fluctuation amplitude and the reflux ratio adjustment amplitude is lower than the standard threshold and the duration exceeds the preset duration, it is judged as an anomaly in the correlation between operating parameters and equipment status."

[0071] Based on the specific features in the abnormal association feature subset, the abnormal association type is identified: if the abnormal association feature subset shows that "the correlation strength parameter between the distillation column pressure parameter and the reflux ratio parameter is lower than the standard threshold", then the abnormal association type is determined as "abnormal association between operating parameters and equipment status" according to the judgment rules.

[0072] Assess the strength of anomalous features: Based on the deviation of the correlation strength parameter in the anomalous feature subset from the standard threshold, and in conjunction with the evaluation criteria in the judgment rules, determine the anomalous feature strength level. If the deviation of the correlation strength parameter from the standard threshold is in the slight deviation range, the anomalous feature strength is "slight"; if it is in the moderate deviation range, it is "moderate"; if it is in the severe deviation range, it is "severe".

[0073] Extracting time clues of anomaly occurrence: From the time series data in the subset of anomaly-related features, locate the time point when the correlation strength parameter first falls below the standard threshold, and the time interval during which the parameter continues to deviate. Use the above time information as time clues of anomaly occurrence.

[0074] By integrating the types of abnormal associations, the intensity of abnormal features, and the time clues of abnormal occurrence, a preliminary identification result of process abnormalities is formed. This preliminary identification result of process abnormalities is presented in a structured form, which clarifies the abnormal associations and basic characteristics that occur in the current process.

[0075] Step S140: Based on the preliminary identification results of the process anomalies, determine the types of abnormal events in the germane purification process and the scope of impact of the corresponding purification steps, and obtain detailed information on the abnormal events.

[0076] By analyzing the core information in the preliminary identification results of process anomalies, and combining the process logic and parameter relationships of each step in the germane purification process, the specific type of the anomaly event is identified, and the scope of its impact on the current and subsequent purification steps is analyzed. Finally, detailed information containing key details of the anomaly event is generated.

[0077] Step S141: Analyze the abnormal association types in the preliminary identification results of the process abnormality, and determine the abnormal event type by combining the pre-constructed abnormality type classification system. The abnormality type classification system includes abnormalities of operating parameters deviation, abnormalities of equipment status failure, abnormalities of product characteristics not meeting standards, and abnormalities of multi-factor coupling.

[0078] The pre-constructed anomaly classification system clearly defines the types of anomalies and their judgment criteria: Operational parameter deviation anomalies refer to deviations between the actual and standard values ​​of operational parameters exceeding the allowable range, leading to abnormal correlations; Equipment status failure anomalies refer to abnormal equipment operating status parameters, causing abnormal correlations with operational parameters or product characteristics; Product characteristic non-compliance anomalies refer to substandard purity, concentration, and other characteristic parameters of germane products, resulting in abnormal correlations with operational parameters or equipment status; Multi-factor coupling anomalies refer to abnormal correlations caused by the combined effects of multiple factors such as operational parameter deviations and equipment status failures.

[0079] Analyze the anomaly correlation types in the preliminary identification results of process anomalies. If the anomaly correlation type is "abnormal correlation between operating parameters and equipment status", and further analysis shows that it is caused by the actual value of the reflux ratio operating parameter deviating too much from the standard value, then the anomaly event type is determined to be "operating parameter deviation anomaly" based on the classification system. If the anomaly correlation type is "abnormal correlation between equipment status and product characteristics", and the abnormal equipment status parameters are caused by the temperature sensor failure in the distillation column, then it is determined to be "equipment status failure anomaly". If the anomaly correlation type is "abnormal correlation between operating parameters and product characteristics", and the product purity parameter does not meet the standard, then it is determined to be "product characteristic non-compliance anomaly". If the anomaly is caused by both operating parameter deviation and equipment status failure, then it is determined to be "multi-factor coupling anomaly".

[0080] Step S142: Extract the time clues of the abnormality occurrence from the preliminary identification results of the process abnormality, and combine them with the time series records of each step of the germane purification process to determine the target purification step where the abnormal event first occurred. The target purification step is any one or more of the pretreatment step, adsorption step, distillation step, and posttreatment step.

[0081] The initial identification results of process anomalies were used to identify the time of occurrence of the anomalies, specifically the time when the correlation strength parameter first showed an anomaly. Time series records of each step in the germane purification process were retrieved, detailing the start and end times of each step and the time information of key operational nodes.

[0082] By comparing the time point corresponding to the anomaly occurrence with the time series records of each stage, if the time point falls between the start and end times of the distillation stage, the target purification stage is identified as the "distillation stage." If the time point simultaneously falls within the transition period between the end time of the adsorption stage and the start time of the distillation stage, and the anomaly correlation features involve parameters from both stages, the target purification stage is identified as the "adsorption and distillation stage." Through this time comparison method, the purification stage where the anomaly first occurred can be accurately located.

[0083] Step S143: Call the purification process impact analysis module to analyze the impact of the abnormal event on each sub-step within the target purification process based on the abnormal event type and the target purification process. The sub-steps include the raw material filtration sub-step and raw material preheating sub-step in the pretreatment process, the adsorbent loading sub-step and adsorption temperature control sub-step in the adsorption process, the tower pressure adjustment sub-step and reflux ratio control sub-step in the distillation process, and the drying sub-step and purity detection sub-step in the post-treatment process.

[0084] Step S1431: Input the abnormal event type and the target purification step into the purification step impact analysis module. The purification step impact analysis module pre-stores the correlation impact matrix between sub-steps and abnormal types within each purification step. The correlation impact matrix between sub-steps and abnormal types within each purification step contains the potential impact coefficients of different abnormal types on each sub-step.

[0085] After receiving the abnormal event type and the target purification step, the purification process impact analysis module retrieves a pre-stored correlation impact matrix. This matrix is ​​organized with sub-steps as rows and abnormality types as columns. Each cell corresponds to a potential impact coefficient, which reflects the degree of potential impact of a certain type of abnormality on a specific sub-step. For example, in the correlation impact matrix for the "distillation process," the potential impact coefficient for the "operating parameter deviation type abnormality" corresponding to the "tower pressure regulation sub-step" is higher than the coefficient for the corresponding "reflux ratio control sub-step," indicating that this type of abnormality has a greater potential impact on the tower pressure regulation sub-step.

[0086] Step S1432: Extract the target influence coefficient subset corresponding to the abnormal event type and the target purification step from the correlation influence matrix of sub-steps and abnormal types within each purification step. The target influence coefficient subset contains the potential influence coefficient of the abnormal event type on each sub-step within the target purification step.

[0087] If the target purification step is the "distillation step" and the abnormal event type is "operating parameter deviation anomaly", then all potential influence coefficients corresponding to the "distillation step" row and the "operating parameter deviation anomaly" column are extracted from the correlation influence matrix to form a subset of target influence coefficients. This subset of target influence coefficients includes the potential influence coefficients of this anomaly type on each sub-step within the distillation step, such as the column pressure regulation sub-step and the reflux ratio control sub-step.

[0088] Step S1433: Collect the current operating parameter data of each sub-step in the target purification process. The current operating parameter data of each sub-step in the target purification process includes the actual values ​​of the sub-step's operating parameters, the actual values ​​of the equipment's operating status, and the actual values ​​of the product characteristics.

[0089] The current operating parameter data of each sub-step in the target purification process is collected through the data acquisition interface: for the column pressure regulation sub-step in the distillation process, the actual monitored value of the column pressure, the opening status of the column pressure regulating valve, and the concentration value of germane product in the corresponding area are collected; for the reflux ratio control sub-step, the actual set value of the reflux ratio, the operating speed of the reflux pump, and the temperature value of the condenser are collected. The above data covers three dimensions: operating parameters, equipment status, and product characteristics, to ensure that the current operating status of the sub-step can be fully reflected.

[0090] Step S1434: Analyze the degree of deviation between the current operating parameter data of each sub-step in the target purification process and the standard value of the normal operating parameters, and calculate the deviation rate parameter. The deviation rate parameter is used to characterize the deviation between the current operating state of the sub-step and the normal state.

[0091] The actual values ​​of the current operating parameters for each sub-step are compared with the corresponding standard values ​​of normal operating parameters. The deviation rate parameter is obtained by calculating the ratio of (actual value - standard value) to the standard value. If the actual value is higher than the standard value, the deviation rate parameter is positive; if it is lower than the standard value, the deviation rate parameter is negative. The larger the absolute value of the deviation rate parameter, the greater the deviation between the current operating state and the normal state of the sub-step. For example, in the tower pressure regulation sub-step, if there is a difference between the actual tower pressure value and the standard value, the deviation rate parameter of the tower pressure parameter is obtained through the above calculation; similarly, the deviation rate parameter of the reflux ratio parameter is obtained.

[0092] Step S1435: Standardize the deviation rate parameters of each sub-step in the target purification process and the potential influence coefficients corresponding to the subset of target influence coefficients into dimensionless form. Perform a weighted product operation on the standardized deviation rate parameters and the potential influence coefficients to obtain the quantitative value of the influence degree of each sub-step in the target purification process. The quantitative value of the influence degree is used to compare the severity of the impact of abnormal events on each sub-step.

[0093] Using the same standardization method, the deviation rate parameter and the potential influence coefficient are transformed to the same dimensionless range, eliminating calculation errors caused by dimensional differences. For example, a linear transformation can be used to map both to a specific range, allowing the standardized values ​​to be directly used in mathematical operations.

[0094] Multiplying the standardized deviation rate parameter by the corresponding potential impact coefficient yields the quantified impact value of each sub-step. If the absolute value of the standardized deviation rate parameter of a sub-step is large, and the corresponding potential impact coefficient is also large, then its quantified impact value is large, indicating that the abnormal event has a more serious impact on the sub-step; conversely, the quantified impact value is small, indicating a less severe impact.

[0095] Step S1436: Based on the relative magnitude of the quantitative values ​​of the impact of each sub-step within the target purification process, sort the sub-steps within the target purification process from largest to smallest impact, and generate a sub-step impact ranking list, which serves as the result of the relative severity analysis of the impact of abnormal events on each sub-step within the target purification process.

[0096] The impact values ​​of each sub-step are compared and sorted in descending order of value. For example, if the impact value of the column pressure regulation sub-step in the distillation process is greater than that of the reflux ratio control sub-step, the sorting is "column pressure regulation sub-step > reflux ratio control sub-step", generating a corresponding sub-step impact ranking list to clearly show the priority of the impact of abnormal events on each sub-step in the target purification process.

[0097] Step S144: Analyze the transmission impact of the abnormal event on the subsequent related purification steps of the target purification step, wherein the subsequent related purification steps are the purification steps that are connected to the target purification step in the process flow, and determine the degree and scope of the transmission impact.

[0098] Clearly define the position of the target purification step within the germane purification process and determine its subsequent related purification steps. For example, if the target purification step is the "distillation step," its subsequent related purification step is the "post-treatment step"; if the target purification step is the "adsorption step," its subsequent related purification step is the "distillation step."

[0099] Based on the abnormal event type and the process function of the target purification step, the transmission path is analyzed: If the target purification step is the "distillation step" and the abnormal event type is "operating parameter deviation type abnormality" which leads to a decrease in product purity, then the transmission path is "distillation step → post-processing step". That is, the unqualified product from the distillation step enters the post-processing step, affecting the drying effect of the post-processing step and the final purity test result.

[0100] Assess the degree of transmission impact: Based on the degree of deviation of the product characteristic parameters in the target purification stage, and in conjunction with the requirements of subsequent related purification stages on the raw material characteristics, determine the degree of transmission impact. If the deviation of the product purity in the target purification stage is slight, and the subsequent related purification stages have a certain correction capability, then the degree of transmission impact is "slight"; if the deviation is severe, exceeding the correction range of the subsequent related purification stages, then the degree of transmission impact is "severe".

[0101] Determine the scope of the impact: Analyze the flow path of abnormal products in subsequent related purification steps to identify the affected sub-steps. For example, if the substandard product from the distillation step enters the post-processing step, it will affect the drying sub-step (requiring extended drying time) and the purity detection sub-step (resulting in substandard test results). The scope of the impact is "the drying sub-step and the purity detection sub-step in the post-processing step".

[0102] Step S145: Based on the comprehensive abnormal event type, target purification stage, the degree of impact of sub-steps within the target purification stage, and the scope of impact of subsequent related purification stages, generate detailed abnormal event information including abnormal event classification identifier, initial occurrence stage identifier, sub-step impact list, and subsequent propagation scope description.

[0103] Assign corresponding classification labels to abnormal event types. For example, “Operating parameter deviation type abnormality” is labeled as “OPD”, and “Equipment status failure type abnormality” is labeled as “ESF”, thus forming abnormal event classification labels.

[0104] Assign initial occurrence stage identifiers to the target purification stage, such as "PP" for "pretreatment stage", "AD" for "adsorption stage", "DS" for "distillation stage", and "PT" for "posttreatment stage", thus forming initial occurrence stage identifiers.

[0105] Convert the sub-step impact ranking list into a sub-step impact list, which includes the name of each sub-step, the quantified value of its impact, and the ranking result.

[0106] The degree and scope of the transmission impact of subsequent purification steps are integrated into a description of the subsequent transmission scope, clarifying the affected subsequent steps and sub-steps and the degree of impact.

[0107] The above information is integrated according to a preset format to generate detailed information on the abnormal event. This detailed information comprehensively and clearly presents the core characteristics and scope of impact of the abnormal event.

[0108] For example, if the abnormal event type is "Operating Parameter Deviation Anomaly," then the abnormal event classification identifier is "OPD"; the target purification step is "Distillation Step," and the initial occurrence step is identified as "DS"; the sub-step impact list is "1. Tower Pressure Adjustment Sub-step - Impact Level Quantification Value - Ranking 1; 2. Reflux Ratio Control Sub-step - Impact Level Quantification Value - Ranking 2"; the subsequent propagation range description is "The subsequent associated purification step is the post-processing step, the affected sub-steps are the drying sub-step and the purity detection sub-step, the propagation impact level is moderate, specifically, the drying sub-step requires extended processing time, and the purity detection sub-step may have a risk of substandard test results." This information is integrated according to the preset format of "Abnormal Event Classification Identifier - Initial Occurrence Step Identifier - Sub-step Impact List - Subsequent Propagation Range Description" to ultimately form a structured abnormal event detail, which can be directly used to generate subsequent early warning instructions.

[0109] Step S150: Generate an anomaly warning instruction for the germane purification process based on the detailed information of the abnormal event, which includes an anomaly location identifier and suggested countermeasures, and send the anomaly warning instruction for the germane purification process to the purification process control terminal.

[0110] Based on detailed information about abnormal events, a precise abnormality location identifier is constructed, and targeted response measures are matched. After generating early warning instructions in a standardized format, the instructions are transmitted to the purification process control terminal through a communication interface to achieve timely early warning and handling guidance for abnormalities.

[0111] Step S151: Parse the abnormal event classification identifier and the initial occurrence stage identifier in the abnormal event details, and combine them with the stage coding system and sub-step coding system of the germane purification process to generate the initial location code of the abnormal event. The initial location code of the abnormal event is used to uniquely identify the purification stage and sub-step where the abnormal event first occurs.

[0112] The germane purification process uses a process coding system that assigns a unique two-digit code to each purification process. For example, the pretreatment process is "01", the adsorption process is "02", the distillation process is "03", and the post-treatment process is "04". The sub-step coding system assigns a two-digit code to each sub-step within each process. For example, the pressure adjustment sub-step within the distillation process is "01" and the reflux ratio control sub-step is "02".

[0113] Analyze the detailed information of the abnormal event. If the abnormal event classification identifier is "OPD", the initial occurrence stage identifier is "DS" (corresponding stage code "03"), and the first sub-step in the sub-step impact list is the tower pressure regulation sub-step (corresponding sub-step code "01"), then the initial location code consists of "abnormal event classification identifier - stage code - sub-step code", i.e. "OPD-03-01". This code uniquely corresponds to the initial occurrence location of "operating parameter deviation anomaly - distillation stage - tower pressure regulation sub-step".

[0114] Step S152: Extract the sub-step impact list and subsequent propagation range description from the detailed information of the abnormal event, determine all purification links and sub-steps currently affected by the abnormal event, and generate an extended location code set for the abnormal event, which contains the codes of all affected links and sub-steps.

[0115] Extract all affected sub-steps from the sub-step impact list and generate corresponding codes using the process coding system. For example, the reflux ratio control sub-step in the distillation process is coded as "03-02". Extract affected subsequent processes and sub-steps from the subsequent transmission range description and generate corresponding codes. For example, the drying sub-step in the post-processing process is coded as "04-01" and the purity detection sub-step is coded as "04-02". Integrate the above codes to form an extended location code set, such as "{03-02, 04-01, 04-02}". This extended location code set completely covers all processes and sub-steps currently affected by the abnormal event.

[0116] Step S153: Combine the initial location code of the abnormal event with the extended location code set of the abnormal event to form an abnormal location identifier, wherein the abnormal location identifier includes an initial occurrence location code and a current influence range code.

[0117] The initial location code “OPD-03-01” generated in step S151 is combined with the extended location code set “{03-02, 04-01, 04-02}” generated in step S152, connected by a separator, to form the anomaly location identifier “OPD-03-01|{03-02, 04-01, 04-02}”. This anomaly location identifier clearly identifies the initial location of the anomaly and clearly presents the current scope of impact, facilitating maintenance personnel to quickly locate the core of the problem and the affected area.

[0118] Step S154: Call the pre-built anomaly response measures library. The pre-built anomaly response measures library contains response strategies corresponding to different anomaly event types and different impact ranges. The response strategies corresponding to different anomaly event types and different impact ranges include operating parameter adjustment schemes, equipment maintenance schemes, and temporary process flow adjustment schemes.

[0119] The pre-built anomaly response library adopts a hierarchical storage structure, with the top level categorized by anomaly event type and the lower levels subdivided by impact scope. For example, the "Operating Parameter Deviation Anomaly" category includes subcategories such as "Single-Stage Impact" and "Multi-Stage Transmission Impact," with each subcategory corresponding to a specific response strategy. The operating parameter adjustment plan clearly defines the names of the parameters to be adjusted, the direction of adjustment, and the operating procedure; the equipment maintenance plan includes the inspection procedure for faulty equipment, component replacement guidelines, and maintenance precautions; and the temporary process flow adjustment plan covers the conditions for pausing, jumping, and activating alternative processes at process nodes.

[0120] Step S155: Based on the description of the abnormal event type and scope of impact in the abnormal event details, select matching target response strategies from the pre-built abnormal response measures library.

[0121] Step S1551: Extract keywords from the abnormal event types in the abnormal event details to obtain abnormal type keywords, which are used to accurately characterize the core type features of the abnormal event.

[0122] For the abnormal event type "operation parameter deviation anomaly", the keywords "operation parameter" and "deviation" are extracted. These keywords directly reflect the core type characteristics of the anomaly and are used for preliminary retrieval in the response measures database.

[0123] Step S1552: Perform semantic analysis on the impact range description in the detailed information of the abnormal event, and extract impact range keywords. The impact range keywords include the name of the affected purification process, the name of the sub-step, and words describing the degree of impact.

[0124] Semantic analysis was performed on the description of the scope of impact: "The subsequent purification step is the post-processing step, the affected sub-steps are the drying sub-step and the purity detection sub-step, and the degree of influence is moderate." The keywords for the scope of impact were extracted as "post-processing step," "drying sub-step," "purity detection sub-step," and "moderate influence." These keywords clearly define the specific scope and degree of the abnormal impact.

[0125] Step S1553: Establish the first mapping relationship between exception type keywords and response strategies in the pre-built exception response measures library, and establish the second mapping relationship between impact scope keywords and response strategies in the pre-built exception response measures library.

[0126] The first mapping relationship is established through a keyword matching algorithm, which associates abnormal type keywords such as "operational parameters" and "deviation" with strategies of the "operational parameter adjustment plan" category in the response measures library; the second mapping relationship is also established through keyword matching, which associates impact range keywords such as "post-processing stage" and "moderate impact" with response strategies for moderate impact on the post-processing stage.

[0127] Step S1554: Based on the first mapping relationship, filter out a set of candidate response strategies that match the abnormality type keywords. The set of candidate response strategies includes all response strategies applicable to the abnormal event type.

[0128] Based on the first mapping relationship, all strategies containing "operating parameter adjustment" are selected from the response measures library to form a candidate response strategy set. For example, this set includes strategies such as "tower pressure parameter adjustment scheme" and "reflux ratio parameter adjustment scheme" that are applicable to abnormal deviations in operating parameters.

[0129] Step S1555: Based on the second mapping relationship, perform a second screening on the candidate response strategy set, eliminate response strategies that are not suitable for the current scope of influence, and retain response strategies that match the keywords of the scope of influence to form a candidate subset of target response strategies.

[0130] Based on the second mapping relationship, strategies that simultaneously include the relevant conditions of "post-processing stage" and "moderate impact" are selected from the candidate response strategy set. Strategies that only apply to a single stage with a slight impact are eliminated, forming a candidate subset of target response strategies. For example, this subset includes "adjustment of distillation column pressure parameters + extension of drying time in post-processing stage".

[0131] Step S1556: Prioritize the candidate subset of target response strategies and determine the response strategy with the highest priority as the final target response strategy. The ranking criteria include the historical evaluation of the implementation effect of the response strategy, the difficulty of implementation, the time required for implementation, and the implementation cost.

[0132] Collect historical data for each strategy in the candidate subset of target response strategies: evaluate the historical effectiveness by assessing the anomaly resolution rate in historical applications; classify the implementation difficulty according to the complexity of the operation steps; determine the implementation time based on the estimated time to complete the operation; and calculate the implementation cost based on the required consumables and manpower costs.

[0133] Based on these indicators, a comprehensive score is calculated, and the strategy with the highest score has the highest priority. For example, the "adjustment of distillation column pressure parameters + extension of drying time in post-processing" solution has a high historical resolution rate, low implementation difficulty, short time requirement, and low cost, and thus has the highest comprehensive score and has been identified as the final target response strategy.

[0134] Step S1557: If there are multiple response strategies with the same priority in the candidate subset of the target response strategy, then the response strategy that best matches the current severity is selected as the target response strategy by further combining the severity identifier of the abnormal event extracted from the abnormal event details.

[0135] If two strategies in the candidate subset have the same overall score and the same priority, the severity indicator, such as "moderately severe," is extracted from the anomaly details. The adaptability of the two strategies to moderately severe anomalies is compared, and the strategy explicitly labeled as suitable for moderately severe scenarios is selected as the target response strategy to ensure that the strategy accurately matches the severity of the anomaly.

[0136] Step S156: Refine the target response strategy by adding specific operational steps, implementation time requirements, and key monitoring points during implementation to form detailed response measures recommendations.

[0137] Taking the target response strategy of "adjusting the column pressure parameters in the distillation stage + extending the drying time in the post-treatment stage" as an example, the specific operation steps are detailed as follows: In the distillation stage, first turn off the automatic control mode of the column pressure regulating valve and switch to manual mode, and gradually adjust the valve opening by rotating the adjustment knob; in the post-treatment stage, modify the set time on the control panel of the drying equipment.

[0138] Supplementary implementation time requirements: Adjustments to tower pressure parameters must be implemented immediately after an abnormality warning is issued, and extensions to drying time must be completed before non-conforming products enter the post-processing stage.

[0139] The key monitoring points during implementation are clearly defined: In the distillation stage, real-time monitoring of changes in column pressure parameters and the corresponding changes in the reflux ratio is required; in the post-processing stage, monitoring of the temperature and humidity parameters of the drying equipment is necessary to ensure that the adjustment process meets expectations. These detailed points form a comprehensive set of recommended countermeasures.

[0140] Step S157: Integrate the anomaly location identifier and the suggested countermeasures according to the preset warning instruction format, add basic information such as the warning instruction generation time, instruction number, and sending object identifier, and generate an anomaly warning instruction for the germane purification process containing the anomaly location identifier and the suggested countermeasures.

[0141] The preset warning instruction format is "Instruction Number | Generation Time | Sending Target | Anomaly Location Identifier | Suggested Countermeasures". Information is integrated according to this format: the instruction number is generated as "YW + Date + Serial Number"; the generation time is the current system time; the sending target identifier is the equipment number of the purification process control terminal; the anomaly location identifier is "OPD-03-01|{03-02, 04-01, 04-02}"; and the suggested countermeasures include detailed operating steps, time requirements, and monitoring points. After integration, a complete anomaly warning instruction for the germane purification process is generated.

[0142] Step S158: Send the abnormal warning command of the germane purification process to the purification process control terminal.

[0143] The generated early warning command is sent to the purification process control terminal via the industrial Ethernet communication protocol. Upon receiving the command, the control terminal displays an early warning prompt window on its screen, showing the anomaly location and suggested countermeasures, and issues an audible and visual alarm to remind maintenance personnel to handle the anomaly promptly. Simultaneously, the control terminal feeds back the command reception status to the sending end to ensure successful command transmission.

[0144] Furthermore, the pre-training process of the process anomaly analysis AI model includes:

[0145] For example, step S211: collect the set of historical correlation features and historical abnormal event tags corresponding to historical abnormal events in the germane purification process. The historical abnormal event tags include the type identifier, scope of impact identifier, and severity identifier of the historical abnormal event.

[0146] Historical anomaly events occurring in the germane purification process over the past five years were collected. Each event includes a corresponding set of historical correlation features (such as the distribution and trends of correlation parameters between historical operating parameters and equipment status) and a historical anomaly event label. The label includes type identifiers such as "OPD" and "ESF," impact range identifiers such as "03-01|{03-02}," and severity identifiers such as "mild," "moderate," and "severe," ensuring data completeness and labeling accuracy.

[0147] Step S212: Divide the collected historical correlation feature set into training feature subset, validation feature subset and test feature subset, and allocate the data volume according to a preset ratio, wherein the preset ratio is determined based on the total amount of data and the model training requirements.

[0148] Based on the total amount of data and the need for sample diversity in model training, a random stratified sampling method is used to divide the dataset, ensuring that the distribution of anomaly types and impact ranges in each subset is consistent with the original data. For example, if the total amount of data is large, a preset ratio is used to allocate the dataset in the order of training feature subset, validation feature subset, and test feature subset, to ensure sufficient training samples and enough samples for validating and testing model performance.

[0149] Step S213: Construct the initial network structure of the process anomaly analysis AI model. The initial network structure of the process anomaly analysis AI model includes an input layer, a correlation feature comparison layer, an anomaly pattern matching layer, and an output layer. The input layer of the initial network structure of the process anomaly analysis AI model is used to receive the correlation feature set. The correlation feature comparison layer of the initial network structure of the process anomaly analysis AI model is used to calculate the difference degree with the standard feature template. The anomaly pattern matching layer of the initial network structure of the process anomaly analysis AI model is used to perform pattern matching. The output layer of the initial network structure of the process anomaly analysis AI model is used to output the anomaly identification result.

[0150] The number of neurons in the input layer matches the feature dimensions of the associated feature set, ensuring complete feature data input. The associated feature comparison layer uses a fully connected network structure, calculating the difference parameter between the input features and the standard feature template through an activation function. The anomaly pattern matching layer introduces an attention mechanism, focusing on the matching process between key feature dimensions and anomaly patterns. The output layer uses a softmax activation function, outputting the recognition probability of different anomaly types and the corresponding anomaly feature information. All layers are connected through weight parameters to form a complete initial network structure.

[0151] Step S214: Input the training feature subset and the corresponding historical abnormal event labels into the initial network structure of the process anomaly analysis AI model, and use the gradient descent optimization algorithm to iteratively update the model parameters of the initial network structure of the process anomaly analysis AI model. In each iteration, calculate the loss value between the model prediction result of the initial network structure of the process anomaly analysis AI model and the historical abnormal event labels, and adjust the weight parameters and bias parameters of the initial network structure of the process anomaly analysis AI model according to the loss value.

[0152] The training feature subset and historical anomaly event labels are input into the initial network structure in batches. Each batch of data is processed through the input layer, the associated feature comparison layer, and the anomaly pattern matching layer, and the output layer outputs the prediction result. The cross-entropy loss function is used to calculate the loss value between the prediction result and the historical anomaly event labels. The gradient descent optimization algorithm is used to adjust the weight parameters and bias parameters of each layer in the direction of decreasing loss value. This process is repeated until the model's prediction accuracy on the training data reaches the preset requirement.

[0153] Step S215: After each iteration, input the verification feature subset into the process anomaly analysis AI model in the current training stage, calculate the recognition accuracy and error rate of the process anomaly analysis AI model in the current training stage on the verification feature subset, and stop training the process anomaly analysis AI model when the recognition accuracy no longer improves or the error rate no longer decreases in multiple iterations.

[0154] After each iteration, a subset of validation features is input into the current model, and the model's recognition accuracy (the ratio of correctly identified anomalies to the total number of validation events) and error rate (the ratio of incorrectly identified anomalies to the total number of validation events) on the validation set are calculated. If the recognition accuracy remains stable for several consecutive iterations without improvement, or the error rate remains stable for several consecutive iterations without decrease, the model is considered to have reached convergence, training is stopped, and overfitting is avoided.

[0155] Step S216: Input the test feature subset into the trained process anomaly analysis AI model, evaluate the generalization ability of the trained process anomaly analysis AI model on the test data. If the generalization ability meets the preset standard, the training of the process anomaly analysis AI model is completed, and a pre-trained process anomaly analysis AI model is obtained. If the generalization ability does not meet the preset standard, adjust the initial network structure of the process anomaly analysis AI model or increase the amount of training data, and retrain the process anomaly analysis AI model until the generalization ability of the process anomaly analysis AI model meets the standard.

[0156] The tested feature subset is input into the trained model, and the generalization ability is evaluated by calculating the recognition accuracy, precision, recall, and F1 score on the test set. If these indicators all meet the preset standards (e.g., accuracy is higher than a preset threshold, and F1 score is higher than a preset threshold), the model training is complete. If the standards are not met, the network structure can be optimized by increasing the number of network layers or adjusting the number of neurons, or by supplementing more training data from different scenarios, and the training process can be restarted until the model's generalization ability meets the requirements, ultimately resulting in a pre-trained AI model for process anomaly analysis.

[0157] Based on the same inventive concept, please refer to Figure 2 The diagram shows a schematic block diagram of an AI-based germanane purification process anomaly early warning system 100 provided in this application embodiment for performing the above-described inspection video stream processing method. The AI-based germanane purification process anomaly early warning system 100 may include a communication unit 110, a machine-readable storage medium 120, and a processor 130.

[0158] In this embodiment, both the machine-readable storage medium 120 and the processor 130 are located within the AI-based germanane purification process anomaly early warning system 100 and are separately configured. However, it should be understood that the machine-readable storage medium 120 may also be independent of the AI-based germanane purification process anomaly early warning system 100 and may be accessed by the processor 130 via a bus interface. Alternatively, the machine-readable storage medium 120 may also be integrated into the processor 130 and may communicate with external systems via the communication unit 110.

[0159] The processor 130 is the control center of the AI-based germane purification process anomaly early warning system 100. It connects various parts of the system via various interfaces and lines, and performs overall monitoring by running or executing software programs and / or modules stored in the machine-readable storage medium 120, and by calling data stored in the machine-readable storage medium 120, thereby executing various functions and processing data of the AI-based germane purification process anomaly early warning system 100. Optionally, the processor 130 may include one or more processing cores; for example, the processor 130 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may also not be integrated into the processor. The machine-readable storage medium 120 is used to store machine-executable instructions for executing the scheme of this application, and the processor 130 is used to execute the machine-executable instructions stored in the machine-readable storage medium 120 to implement the inspection video stream processing method provided in the aforementioned method embodiments.

[0160] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. An AI-based method for early warning of process anomalies in germane purification, characterized in that, The method includes: Acquire a set of key operation data and a set of process status data in the germane purification process. The set of key operation data includes records of operation parameters for each step of the purification process, and the set of process status data includes records of the operating status of the purification equipment and records of the characteristics of the germane product. Based on the key operation data set and the process status data set, a purification process data association model is constructed to obtain the association relationship feature set output by the purification process data association model. The purification process data association model is used to characterize the association relationship between operation parameters and equipment operating status, operation parameters and germane product characteristics, and equipment operating status and germane product characteristics. The pre-trained process anomaly analysis AI model is invoked to perform anomaly tendency identification processing on the correlation feature set to obtain preliminary process anomaly identification results; Based on the preliminary identification results of the process anomalies, the types of abnormal events in the germane purification process and the scope of impact of the corresponding purification steps are determined, and detailed information on the abnormal events is obtained. Based on the detailed information of the abnormal event, an abnormal warning instruction for the germane purification process is generated, which includes an abnormal location identifier and a suggestion for countermeasures, and the abnormal warning instruction for the germane purification process is sent to the purification process control terminal. The process involves determining the type of abnormal event in the germane purification process and the scope of impact of the corresponding purification stage based on the preliminary identification results of the process anomaly, and obtaining detailed information about the abnormal event, including: The abnormal association types in the preliminary identification results of the process abnormalities are analyzed, and the abnormal event types are determined by combining them with the pre-constructed abnormality type classification system. The abnormality type classification system includes abnormalities caused by deviation of operating parameters, abnormalities caused by equipment status failure, abnormalities caused by non-compliance of product characteristics, and abnormalities caused by multiple factors. Extract the time clues of the occurrence of the anomaly from the preliminary identification results of the process anomaly, and combine them with the time series records of each step of the germane purification process to determine the target purification step where the anomaly event first occurred. The target purification step is any one or more of the pretreatment step, adsorption step, distillation step, and posttreatment step. The purification process impact analysis module is invoked to analyze the impact of the abnormal event on each sub-step within the target purification process based on the abnormal event type and the target purification process. The sub-steps include the raw material filtration sub-step and raw material preheating sub-step in the pretreatment process, the adsorbent loading sub-step and adsorption temperature control sub-step in the adsorption process, the tower pressure adjustment sub-step and reflux ratio control sub-step in the distillation process, and the drying sub-step and purity detection sub-step in the post-treatment process. The study analyzes the transmission impact of abnormal events on subsequent related purification steps of the target purification step, where the subsequent related purification steps are the purification steps that are connected to the target purification step in the process flow, and determines the degree and scope of the transmission impact. Based on the comprehensive abnormal event type, target purification stage, the degree of impact of sub-steps within the target purification stage, and the scope of impact of subsequent related purification stages, detailed abnormal event information is generated, including abnormal event classification identifier, initial occurrence stage identifier, sub-step impact list, and subsequent propagation scope description.

2. The AI-based method for early warning of abnormalities in germane purification process according to claim 1, characterized in that, The purification process data association model is constructed based on the key operation data set and the process status data set, resulting in a set of association relationship features output by the purification process data association model, including: The operation parameter records in the key operation data set are classified and processed according to the pretreatment, adsorption, distillation and post-purification processes of germane purification. The operation parameter records are divided into pretreatment operation parameter subset, adsorption operation parameter subset, distillation operation parameter subset and post-purification operation parameter subset. The equipment operation status records and germane product characteristic records in the process status data set are classified and processed. The equipment operation status records are divided into pretreatment equipment status subsets, adsorption equipment status subsets, distillation equipment status subsets, and post-treatment equipment status subsets corresponding to each purification step. The germane product characteristic records are divided into pretreatment product characteristic subsets, adsorption product characteristic subsets, distillation product characteristic subsets, and post-treatment product characteristic subsets corresponding to each purification step. Determine the correlation dimensions between the subset of operating parameters and the corresponding subset of equipment status, the subset of operating parameters and the corresponding subset of product characteristics, and the subset of equipment status and the corresponding subset of product characteristics in each purification step. The correlation dimensions include time synchronization dimension, parameter influence dimension, and status feedback dimension. Based on the aforementioned correlation dimensions, a data association rule base for each purification stage is constructed. The data association rule base includes rules on the impact of changes in operating parameters on equipment status, rules on the impact of changes in operating parameters on product characteristics, and rules on the impact of changes in equipment status on product characteristics. The categorized subsets of operation parameters, device status, and product characteristics are input into the data association rule base. The association strength parameters under each association dimension are obtained through rule matching calculation, forming an association relationship matrix. Feature extraction processing is performed on the association matrix to extract the distribution features, trend features, and abnormal deviation features of the association strength parameters in the association matrix. The distribution features, trend features, and abnormal deviation features of the association strength parameters in the association matrix are combined to form the association feature set output by the data association model of the purification process.

3. The AI-based early warning method for germane purification process abnormalities according to claim 2, characterized in that, The construction of a data association rule base for each purification stage based on the aforementioned association dimension includes: Collect historical key operation data sets and historical process status data sets during the normal operation of the germane purification process, and clean the historical key operation data sets and historical process status data sets to obtain effective historical data sets. Extract the historical operation parameter subset, historical equipment status subset, and historical product characteristic subset of each purification step from the effective historical data set, and align the historical operation parameter subset, historical equipment status subset, and historical product characteristic subset of each purification step according to the time series. For each purification stage, the relationship between the historical operating parameter subsets and the historical equipment status subsets of each purification stage is analyzed. When the operating parameters change, the response of the equipment status changes is recorded, the influence of the operating parameters on the equipment status is extracted, and the influence rules of the operating parameter changes on the equipment status are formed. Analyze the correspondence between the historical operating parameter subsets of each purification step and the historical product characteristic subsets of each purification step. When the operating parameters change, record the response of the product characteristics to the change, extract the influence law of the operating parameters on the product characteristics, and form the influence rule of the operating parameter changes on the product characteristics. Analyze the correspondence between the historical equipment status subsets and the historical product characteristic subsets of each purification step. When the equipment status changes, record the response of product characteristics to the change, extract the influence law of equipment status on product characteristics, and form the influence rule of equipment status change on product characteristics. The effectiveness of the rules on the impact of changes in operating parameters on equipment status, product characteristics, and product characteristics is verified. By comparing historical data from different time periods, the applicability of these rules under different operating conditions is verified. Rules whose applicable scope does not meet the set requirements are eliminated, and a data association rule library for each purification stage is formed.

4. The AI-based early warning method for anomalies in germane purification process according to claim 1, characterized in that, The process anomaly analysis AI model is invoked to perform anomaly tendency identification processing on the correlation feature set, resulting in preliminary process anomaly identification results, including: The set of correlation features is input into the correlation feature comparison module of the process anomaly analysis AI model. The correlation feature comparison module of the process anomaly analysis AI model has a pre-stored standard correlation feature template for the normal purification process. The difference parameter between the set of correlation features and the standard correlation feature template is calculated by the feature comparison algorithm. Based on the difference parameter, an abnormal association feature subset that exceeds a preset comparison threshold is selected from the association feature set. The abnormal association feature subset includes features that are abnormally associated with operating parameters and equipment status, features that are abnormally associated with operating parameters and product characteristics, and features that are abnormally associated with equipment status and product characteristics. The subset of abnormal correlation features is input into the abnormal pattern matching module of the process abnormality analysis AI model. The abnormal pattern matching module of the process abnormality analysis AI model has a pre-stored correlation feature pattern library corresponding to various process abnormalities. The correlation feature pattern library includes abnormal patterns of pretreatment, adsorption, distillation and posttreatment. The abnormal correlation feature subset is matched with each abnormal pattern in the correlation feature pattern library using a pattern matching algorithm. The matching similarity parameter is calculated, and the target abnormal pattern with the highest matching similarity parameter is selected. Based on the anomaly determination rules corresponding to the target anomaly pattern, and combined with the specific feature content in the anomaly association feature subset, a preliminary process anomaly identification result is generated, which includes the anomaly association type, anomaly feature intensity, and anomaly occurrence time clues.

5. The AI-based early warning method for germane purification process abnormalities according to claim 4, characterized in that, The step of matching the subset of abnormal related features with each abnormal pattern in the related feature pattern library using a pattern matching algorithm, and calculating the matching similarity parameter, includes: The abnormal association feature subset is subjected to feature standardization processing, and the association strength parameter distribution features, change trend features, and abnormal deviation features in the abnormal association feature subset are converted into feature vectors in a unified format. Each dimension of the feature vector corresponds to a feature type. The feature vectors of each abnormal pattern in the associated feature pattern library are transformed into pattern feature vectors that are consistent with the feature vector format of the abnormal associated feature subset. The cosine similarity algorithm is used to calculate the cosine similarity value between the feature vector of the abnormal correlation feature subset and the pattern feature vector of each abnormal pattern in the correlation feature pattern library. The cosine similarity value is used to characterize the directional consistency between the feature vector of the abnormal correlation feature subset and the pattern feature vector of each abnormal pattern in the correlation feature pattern library in the feature space. The Euclidean distance algorithm is used to calculate the Euclidean distance between the feature vector of the abnormal correlation feature subset and the pattern feature vector of each abnormal pattern in the correlation feature pattern library. The Euclidean distance value is used to characterize the degree of positional difference between the feature vector of the abnormal correlation feature subset and the pattern feature vector of each abnormal pattern in the correlation feature pattern library in the feature space. Based on the cosine similarity value and the Euclidean distance value, a matching similarity parameter is obtained through weighted calculation. The weights of the cosine similarity value and the Euclidean distance value are dynamically adjusted according to historical matching accuracy data so that the matching similarity parameter can accurately reflect the feature matching degree between the feature vector of the abnormal association feature subset and the pattern feature vector of each abnormal pattern in the association feature pattern library.

6. The AI-based early warning method for germane purification process abnormalities according to claim 1, characterized in that, The call to the purification process impact analysis module, based on the abnormal event type and the target purification process, analyzes the degree of impact of the abnormal event on each sub-step within the target purification process, including: The abnormal event type and the target purification step are input into the purification step impact analysis module. The purification step impact analysis module pre-stores the correlation impact matrix between sub-steps and abnormal types in each purification step. The correlation impact matrix between sub-steps and abnormal types in each purification step contains the potential impact coefficients of different abnormal types on each sub-step. Extract the target influence coefficient subset corresponding to the abnormal event type and the target purification step from the correlation influence matrix of sub-steps and abnormal types in each purification step. The target influence coefficient subset contains the potential influence coefficient of the abnormal event type on each sub-step in the target purification step. Collect the current operating parameter data of each sub-step in the target purification process. The current operating parameter data of each sub-step in the target purification process includes the actual values ​​of the operating parameters of the sub-step, the actual values ​​of the equipment operating status, and the actual values ​​of the product characteristics. The deviation between the current operating parameter data of each sub-step in the target purification process and the standard value of the normal operating parameters is analyzed, and the deviation rate parameter is calculated. The deviation rate parameter is used to characterize the deviation between the current operating state of the sub-step and the normal state. The deviation rate parameters of each sub-step in the target purification process and the potential influence coefficients corresponding to the subset of target influence coefficients are standardized and dimensionless. The standardized deviation rate parameters and the potential influence coefficients are then weighted and multiplied to obtain the quantitative value of the influence degree of each sub-step in the target purification process. The quantitative value of the influence degree is used to compare the severity of the impact of abnormal events on each sub-step. Based on the relative magnitude of the quantitative values ​​of the impact of each sub-step within the target purification process, the sub-steps within the target purification process are sorted from largest to smallest in terms of impact, generating a sub-step impact ranking list, which serves as the result of the relative severity analysis of the impact of abnormal events on each sub-step within the target purification process.

7. The AI-based early warning method for germane purification process abnormalities according to claim 1, characterized in that, The generation of an anomaly warning instruction for the germane purification process based on the detailed information of the anomaly event, including an anomaly location identifier and suggested countermeasures, includes: The abnormal event classification identifier and initial occurrence stage identifier in the abnormal event details are analyzed. Combined with the stage coding system and sub-step coding system of the germane purification process, an initial location code for the abnormal event is generated. The initial location code for the abnormal event is used to uniquely identify the purification stage and sub-step where the abnormal event first occurs. Extract the sub-step impact list and subsequent propagation range description from the detailed information of the abnormal event, determine all purification links and sub-steps currently affected by the abnormal event, and generate an extended location code set for the abnormal event, which contains the codes of all affected links and sub-steps; The initial location code of the abnormal event is combined with the extended location code set of the abnormal event to form an abnormal location identifier, which includes an initial occurrence location code and a current impact range code. The pre-built anomaly response measures library is invoked. The pre-built anomaly response measures library contains response strategies corresponding to different anomaly event types and different impact ranges. The response strategies corresponding to different anomaly event types and different impact ranges include operation parameter adjustment schemes, equipment maintenance schemes, and temporary process flow adjustment schemes. Based on the description of the abnormal event type and scope of impact in the abnormal event details, select matching target response strategies from the pre-built abnormal response measures library; The target response strategy is further refined by adding specific operational steps, implementation time requirements, and key monitoring points during implementation, thus forming detailed response recommendations. The abnormal location identifier and the suggested countermeasures are integrated according to a preset warning instruction format. Basic information such as the warning instruction generation time, instruction number, and target identifier are added to generate an abnormal warning instruction for the germane purification process that includes the abnormal location identifier and the suggested countermeasures.

8. The AI-based method for early warning of abnormalities in germane purification process according to claim 7, characterized in that, The step of calling a pre-built exception handling measures library, and filtering out matching target handling strategies from the pre-built exception handling measures library based on the exception event type and impact scope description in the exception event details, includes: Keyword extraction is performed on the abnormal event type in the abnormal event details to obtain abnormal type keywords, which are used to accurately characterize the core type features of the abnormal event. Semantic analysis is performed on the description of the scope of impact in the detailed information of the abnormal event to extract keywords of the scope of impact. The keywords of the scope of impact include the name of the affected purification process, the name of the sub-step, and words describing the degree of impact. Establish a first mapping relationship between anomaly type keywords and response strategies in the pre-built anomaly response measures library; establish a second mapping relationship between impact scope keywords and response strategies in the pre-built anomaly response measures library. Based on the first mapping relationship, a set of candidate response strategies matching the abnormality type keywords is selected, and the set of candidate response strategies includes all response strategies applicable to the abnormal event type. Based on the second mapping relationship, the candidate response strategy set is further filtered to remove response strategies that are not suitable for the current scope of influence, and to retain response strategies that match the keywords of the scope of influence, thus forming a subset of target response strategy candidates. The candidate subset of target response strategies is prioritized, and the response strategy with the highest priority is determined as the final target response strategy. The ranking criteria include the historical evaluation of the implementation effect of the response strategy, the difficulty of implementation, the time required for implementation, and the implementation cost. If there are multiple response strategies with the same priority in the candidate subset of the target response strategy, then the response strategy that best matches the current severity level is selected as the target response strategy by further combining the severity identifier of the abnormal event extracted from the abnormal event details.

9. An AI-based early warning system for abnormalities in germane purification processes, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the AI-based germanane purification process anomaly early warning method according to any one of claims 1 to 8 by executing the machine-executable instructions.

Images