A method and system for predicting alarm events in refining and chemical plants.

Abstract

This invention discloses a method and system for predicting alarm events in refining and chemical plants, comprising: acquiring historical process alarm events of the unit to be evaluated from the DCS system, and establishing an original alarm data sequence based on alarm time arrangement; establishing a historical alarm flood data sequence library based on the original alarm data sequence, and predicting the alarm parameter similarity prediction sequence corresponding to the latest alarm segment by comparing the acquired latest alarm parameter flood data sequence with the historical alarm flood data sequence library; obtaining an alarm parameter model prediction sequence based on the latest alarm parameter flood data sequence using a fusion alarm prediction model; and fusing the alarm parameter similarity prediction sequence and the alarm parameter model prediction sequence to obtain an alarm parameter prediction result sequence.

Description

Technical Field

[0001] This invention relates to the field of process safety technology in refining and chemical plants, and in particular to a method and system for predicting alarm events in refining and chemical plants. Background Technology

[0002] With the continuous development of the oil refining and chemical industry, the safety and efficiency of the production process have become the focus of industry attention. In the oil refining and chemical production process, due to the complexity of the processes, the large number of equipment, and the variable operating conditions, factors such as high temperature, high pressure, and flammability and explosiveness are often involved, leading to certain risks during the operation of production facilities. Once an accident occurs, it can not only cause casualties and property damage but also potentially cause serious environmental pollution.

[0003] With the rapid development of industrial automation, Distributed Control Systems (DCS) are increasingly widely used in oil refining and chemical enterprises. Their high reliability, flexibility, and powerful data processing capabilities provide strong technical support for the production processes of these enterprises. Simultaneously, process alarm systems integrated into DCS systems have become an essential choice for monitoring and handling anomalies in the production processes of refining and chemical enterprises. In the oil refining and chemical production process, process alarms often indicate deviations, anomalies, or potential anomalies or malfunctions in the production process. Process alarm systems can monitor various key parameters in real time during production and issue timely alerts when these parameters show abnormalities, reminding operators to take appropriate measures to prevent accidents from occurring or escalating.

[0004] Although existing process alarm systems can ensure production safety to a certain extent, some problems still exist. This is because process alarm systems often only alert operators when deviations or abnormalities have already occurred in the equipment, using sound and light to inform and remind them. However, they are insufficient in predicting potential risks. This delayed alarm method may leave operators in a passive state when dealing with emergencies, increasing the risk of accidents.

[0005] Among existing process alarm prediction methods, some techniques often rely on human experience and statistical data, resulting in low prediction accuracy and slow response speed. They are only applicable to specific production environments and process conditions and are no longer suitable for current refining and chemical enterprises. Other techniques use manual or other methods to set alarm limits to determine whether an alarm has occurred or to assess equipment abnormalities. However, due to differences in researchers' experience levels and understanding of equipment operation, manually set alarm limits are often unreasonable. At the same time, because the equipment is affected by various factors during operation, the actual alarm situation often deviates from the manually set alarm limits. These problems will cause the process alarm prediction model to fail to accurately reflect the actual operating status of the equipment, thus potentially reducing the accuracy of the prediction. In addition, some techniques, although considering the correlation and dynamic characteristics between multiple parameters, often rely on complex mathematical models and a large amount of historical data. This may lead to problems such as high model complexity, large computational load, and poor real-time performance in practical applications.

[0006] Therefore, existing technologies need to provide a process alarm prediction scheme for refining and chemical plants that can deeply analyze and mine alarm events and discover the correlation and regularity between alarm events, so as to help more accurately predict the next possible alarm event, thereby improving the accuracy and reliability of the alarm system. Summary of the Invention

[0007] The purpose of this invention is to propose a process alarm prediction scheme for refining and chemical plants that can deeply analyze and mine alarm events and discover the correlation and regularity between alarm events. This scheme can help to more accurately predict the next possible alarm event, thereby improving the accuracy and reliability of the alarm system and effectively avoiding practical problems such as too many and too complicated process alarms, untimely alarm handling, and late abnormal intervention.

[0008] To address the aforementioned technical problems, embodiments of the present invention provide a method for predicting alarm events in refining and chemical plants, comprising: acquiring historical process alarm events of the plant to be evaluated from a DCS system, and establishing an original alarm data sequence arranged based on alarm time; establishing a historical alarm flood data sequence library based on the original alarm data sequence, and predicting an alarm parameter similarity prediction sequence corresponding to the latest alarm segment by comparing the acquired latest alarm parameter flood data sequence with the historical alarm flood data sequence library; obtaining an alarm parameter model prediction sequence based on the latest alarm parameter flood data sequence using a fusion alarm prediction model; and fusing the alarm parameter similarity prediction sequence and the alarm parameter model prediction sequence to obtain an alarm parameter prediction result sequence.

[0009] Preferably, the step of establishing a historical alarm flooding data sequence library based on the original alarm data sequence includes: identifying data sequence segments with alarm flooding from the original alarm data sequence and establishing an original alarm flooding data sequence; performing duplicate alarm removal processing on the original alarm flooding data sequence and establishing a historical alarm flooding data sequence library, so as to retain only the earliest alarm event description information for each parameter.

[0010] Preferably, the latest alarm parameter flooding data sequence is obtained through the following steps: obtaining the original alarm data sequence within the latest alarm collection time period, and inserting the original alarm data sequence within the latest alarm collection time period into the end of the historical alarm flooding data sequence library to form an alarm flooding data sequence to be adjusted; performing duplicate alarm removal processing on the alarm flooding data sequence to be adjusted to obtain an updated historical alarm flooding data sequence library; extracting data segments within the latest alarm collection time period from the updated historical alarm flooding data sequence library to obtain the latest alarm parameter flooding data sequence.

[0011] Preferably, the duplicate alarm removal process is performed through the following steps: Multiple truncated data sequence segments are sequentially extracted from the flooded alarm data sequence to be processed, according to a first sliding window with a first data point length and a first sliding step size; based on a preset duplicate alarm removal principle, flooded alarm data removal processing is performed on each truncated data sequence segment sequentially; and according to the alarm time, the truncated data sequence segments that satisfy the preset duplicate alarm removal principle are integrated to obtain a corresponding processing result sequence. Specifically, truncated data sequence segments that do not satisfy the preset duplicate alarm removal principle are deleted. The preset duplicate alarm removal principle includes: for each alarm parameter in the truncated data sequence segment, the data points between the earliest occurrence of the alarm event description information and the next non-self parameter's alarm event description information are deleted, and the number of remaining alarm parameters exceeds a preset number.

[0012] Preferably, the step of predicting the alarm parameter similarity prediction sequence corresponding to the latest alarm segment by comparing the acquired latest alarm parameter flood data sequence with the historical alarm flood data sequence library includes: sequentially extracting multiple data sequence segments to be evaluated from the historical alarm flood data sequence library according to a second sliding window with a second data point length and a second sliding step size; calculating the similarity between the latest alarm parameter flood data sequence and each data sequence segment to be evaluated; selecting data sequence segments to be evaluated that meet the preset sequence similarity extraction conditions as target data sequences based on the similarity calculation results of each data sequence segment to be evaluated, wherein the preset sequence similarity extraction conditions include: the similarity calculation result exceeds a preset similarity coefficient threshold, and the starting parameter bit number in the data sequence segment to be evaluated is the latest alarm parameter bit number, and the data sequence segment to be evaluated with the highest similarity calculation result; locating the position of the latest alarm parameter in the target data sequence, and extracting the data segment arranged after the latest alarm parameter from the target data sequence as the alarm parameter prediction sequence.

[0013] Preferably, one or more of the following similarity calculation methods are used to calculate the similarity between the latest alarm parameter flood data sequence and each data sequence segment to be evaluated: Euclidean distance, Manhattan distance, cosine similarity, Pearson correlation coefficient, Jaccard similarity coefficient, and Hamming distance.

[0014] Preferably, the method further includes: if no data sequence fragment to be evaluated that meets the preset sequence similarity extraction conditions is selected, or if the latest alarm parameter is the last one in the historical alarm flood data sequence library, then the current alarm parameter similarity prediction sequence is set to empty.

[0015] Preferably, the fusion alarm prediction model includes multiple parallel initial prediction models and an integrated prediction model connected to the output of each initial prediction model. The step of obtaining an alarm parameter model prediction sequence using the fusion alarm prediction model based on the latest alarm parameter flooding data sequence includes: inputting the latest alarm parameter flooding data sequence to the input of each initial prediction model, outputting multiple alarm parameter initial prediction sequences; and inputting each set of alarm parameter initial prediction sequences into the integrated prediction model, outputting the alarm parameter model prediction sequence.

[0016] Preferably, the method further includes constructing the fusion alarm prediction model, including: performing duplicate alarm removal processing on the original alarm data sequence to establish an initial alarm prediction data sequence; sequentially extracting multiple sets of sequence segments for use as training input data from the initial alarm prediction data sequence according to a third sliding window with a third data point length and a third sliding step size; extracting the next alarm parameter from the last alarm parameter in each set of training data segments in the initial alarm prediction data sequence to be used as the training output data of the corresponding training data segment; establishing a first alarm prediction dataset based on all sets of training data segments and corresponding training output data; training and validating multiple pre-constructed initial learning models using the first alarm prediction dataset to obtain multiple initial prediction models; establishing a second alarm prediction dataset by using the training data segments as input and the initial model prediction sequences output by each initial prediction model, as well as the training output data; and training and validating a pre-constructed ensemble learning model using the second alarm prediction dataset to obtain an ensemble prediction model.

[0017] Preferably, before training and validating the initial prediction model / the integrated prediction model, the method further includes: uniquely encoding the alarm parameter tag numbers in the first alarm prediction dataset / the second alarm prediction dataset and performing standardization and normalization processing, wherein the standardization and normalization processing is max-min normalization processing.

[0018] Preferably, at least two initial learning models are selected from the following set of candidate models, and one ensemble learning model is selected from the following set of candidate models. The types of initial learning models selected do not overlap with the types of ensemble learning models selected. The set of candidate models includes: decision tree model, support vector machine model, K nearest neighbor model, logistic regression model, naive Bayes model, factorization machine model, gradient boosting tree model, and random forest model.

[0019] Preferably, the alarm parameter prediction sequence and the subsequent alarm parameter sequence are fused according to one or more of the following fusion rules: intersection principle between sequences, union principle between sequences, parameter priority principle, time interval priority principle, historical alarm count priority principle, and historical alarm duration priority principle.

[0020] Preferably, the method further includes: outputting a response to the alarm parameter prediction result sequence, including: sending a notification to the device monitoring and management personnel terminal in a preset format to the alarm parameters ranked in the top number of the alarm parameters in the alarm parameter prediction result sequence according to a preset number of pushes.

[0021] Preferably, the step of establishing the original alarm data sequence based on alarm time arrangement includes: identifying alarm events with the same alarm parameter according to the alarm time based on the process alarm history events of the device to be evaluated, so that adjacent alarm time points form alarm event description information with different alarm parameters, thereby forming the original alarm data sequence. The alarm event description information includes the name of the refining and chemical unit, alarm parameter tag number, alarm event description, start time and end time.

[0022] On the other hand, embodiments of the present invention also provide a system for predicting alarm events in refining and chemical plants. This system is implemented using the method described above, wherein the system includes: an alarm event acquisition module configured to acquire historical process alarm events of the unit to be evaluated from a DCS system and establish an original alarm data sequence based on alarm time arrangement; a first type of prediction sequence generation module configured to establish a historical alarm flood data sequence library based on the original alarm data sequence, and predict an alarm parameter similarity prediction sequence corresponding to the latest alarm segment by comparing the acquired latest alarm parameter flood data sequence with the historical alarm flood data sequence library; a second type of prediction sequence generation module configured to obtain an alarm parameter model prediction sequence based on the latest alarm parameter flood data sequence using a fusion alarm prediction model; and a prediction result fusion module configured to fuse the alarm parameter similarity prediction sequence and the alarm parameter model prediction sequence to obtain an alarm parameter prediction result sequence.

[0023] Preferably, the alarm event acquisition module has a data acquisition interface based on the OPC AE protocol, so that the DCS system can use a broadcast mechanism to actively push the historical process alarm events of the device to be evaluated to the outside through the AE interface.

[0024] Compared with the prior art, one or more embodiments of the above solutions may have the following advantages or beneficial effects:

[0025] This invention proposes a method and system for predicting alarm events in refining and chemical plants. The method and system, on the one hand, fully explore alarm flooding sequences in historical alarm events, accurately matching and identifying the latest alarm flooding problem through a preset similarity calculation method. On the other hand, it establishes an ensemble learning model based on the Stacking concept, selecting multiple learners to predict upcoming alarms. Finally, through a reasonable judgment mechanism and appropriate fusion strategy, the prediction results from the two channels are effectively integrated to improve the accuracy and reliability of the prediction model. This invention minimizes the occurrence of abnormal operating conditions in production units caused by alarm flooding, untimely alarm handling, delayed intervention, excessively late abnormal intervention, and inaccurate predictions, improving the emergency response capabilities of plant operators. By fully utilizing actual production process alarm events occurring in the DCS system for next alarm prediction, the accuracy and predictability of alarms can be improved, thereby enhancing the enterprise's safety production level.

[0026] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the description, claims, and drawings. Attached Figure Description

[0027] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0028] Figure 1 This is a schematic diagram of the overall steps of a method for predicting alarm events in a refining and chemical plant, according to an embodiment of this application.

[0029] Figure 2 This is a schematic flowchart illustrating a specific process in a method for predicting alarm events in a refining and chemical processing unit, as described in an embodiment of this application.

[0030] Figure 3 This is an example diagram illustrating the construction principle of the first alarm prediction dataset in the method for predicting alarm events in a refining and chemical plant according to an embodiment of this application.

[0031] Figure 4 This is a schematic diagram illustrating the principle of constructing a fusion alarm prediction model in the method for predicting alarm events in refining and chemical plants according to an embodiment of this application.

[0032] Figure 5 This is a schematic diagram of the structure of a system for predicting alarm events in a refining and chemical plant, according to an embodiment of this application. Detailed Implementation

[0033] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings and examples, so that the process of how the present invention uses technical means to solve technical problems and achieve technical effects can be fully understood and implemented accordingly. It should be noted that, as long as there is no conflict, the various embodiments and features in the various embodiments of the present invention can be combined with each other, and the resulting technical solutions are all within the protection scope of the present invention.

[0034] Furthermore, the steps illustrated in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Also, although a logical order is shown in the flowcharts, in some cases the steps shown or described may be performed in a different order than that shown here.

[0035] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments. Unless the context clearly indicates otherwise, the singular forms “a” and “an” as used herein are also intended to include the plural. It should also be understood that the terms “comprising” and / or “including” as used herein specify the presence of the stated features, integers, steps, operations, units, and / or components, without excluding the presence or addition of one or more other features, integers, steps, operations, units, components, and / or combinations thereof.

[0036] To address the aforementioned issues, this application provides a method and system for predicting alarm events in refining and chemical plants. This method and system combine an alarm flooding sequence matching prediction method based on similarity calculation with an alarm prediction method based on ensemble learning, forming an integrated framework with a dual-channel fusion alarm prediction approach. Through pre-defined mechanisms and rules, it solves the problems of excessive alarms and untimely alarm handling, and can effectively utilize alarm flooding sequences and machine learning methods to improve the accuracy of prediction results.

[0037] Example 1

[0038] Figure 1 This is a schematic diagram of the overall steps of a method for predicting alarm events in a refining and chemical plant, according to an embodiment of this application. Figure 2 This is a schematic flowchart illustrating a method for predicting alarm events in a refining and chemical processing unit, as described in an embodiment of this application. The following is a detailed explanation of the method. Figure 1 and Figure 2 The steps of the method for predicting alarm events in refining and chemical equipment (also known as the "alarm event prediction method") described in the embodiments of the present invention will be explained.

[0039] Step S110: Obtain the process alarm history events of the device to be evaluated from the DCS system and establish the original alarm data sequence based on alarm time.

[0040] Based on preset rules, the system first analyzes the historical alarm events of the device's DCS process to obtain key information such as timestamps, tag numbers, and descriptions, generates alarm history records based on time sequence, and establishes an original alarm data sequence database to provide data support for subsequent dual-channel alarm prediction.

[0041] In step S110, during the process of acquiring historical process alarm event data of the refining unit to be evaluated, the present invention first (e.g., via the OPC AE interface) collects historical process alarm events of the unit to be evaluated from the DCS system in real time. In this embodiment of the invention, the parameter tag number refers to the location number in the refining unit to be evaluated to which the parameter acquisition point belongs.

[0042] Then, based on the collected historical process alarm events of the device to be evaluated, alarm events with the same alarm parameter are identified according to the alarm time, so that adjacent alarm time points form alarm event description information with different alarm parameters, thus forming an original alarm data sequence database.

[0043] In one embodiment, data preprocessing is performed based on the collected historical process alarm events of the device under evaluation to establish an original alarm data sequence arranged by alarm time. Specifically, during the data preprocessing of the historical process alarm events of the device under evaluation, alarm events with the same alarm parameter need to be identified according to the alarm time. This ensures that not only does each alarm time point generate alarm event description information for a specific alarm parameter, but adjacent alarm time points also generate alarm event description information with different alarm parameters, thereby constructing the original alarm data sequence.

[0044] In this embodiment of the invention, the alarm event description information includes, but is not limited to: name of the refining unit, alarm parameter tag number, alarm event description, alarm start time, and alarm end time.

[0045] For parameter alarm events, according to the principle of two-in-one parsing of alarm events, alarm events of the same parameter are accurately identified in the order of their occurrence, the start alarm event and the end alarm event of the same parameter are identified, and they are merged to form an alarm record (the fields of an alarm record include: device, tag number, description, start time, and end time), and stored in the historical alarm record database to construct the original alarm data sequence.

[0046] After the original alarm data sequence is constructed, proceed to step S120.

[0047] Step S120: Based on the original alarm data sequence constructed in step S110, establish a historical alarm flood data sequence library, and predict the alarm parameter similarity prediction sequence corresponding to the latest alarm segment by comparing the latest alarm parameter flood data sequence with the historical alarm flood data sequence library.

[0048] First, step S120 will establish a historical alarm flood data sequence library based on the original alarm data sequence.

[0049] Specifically, in step S1, data sequence segments exhibiting alarm flooding are identified from the constructed original alarm data sequence, and an original alarm flooding data sequence is established.

[0050] Based on the original alarm data sequence database, an alarm flooding identification mechanism is implemented. Alarm flooding sequence identification rules are set, and using an initial sliding window with an initial time length (e.g., 10 minutes) and an initial sliding step size, multiple alarm data sequences are extracted along the alarm time point sequence of the original alarm data sequence database. If the number of alarms in a certain alarm data sequence exceeds a preset alarm count threshold, it is determined that the alarm data sequence for the current specified time period exhibits alarm flooding. Thus, by extracting alarm data sequences from different locations within a specified time period in the original alarm data sequence database, the original alarm flooding data sequence is formed. For example, if more than 10 alarms occur within 10 minutes, alarm flooding is considered to exist. Using 10 minutes as the initial time length, each alarm data sequence is sequentially checked for alarm flooding, thereby restoring all alarm data sequences exhibiting alarm flooding to the original alarm flooding data sequence.

[0051] After completing the construction of the original alarm flood data sequence, step S120 will also perform duplicate alarm removal processing on the original alarm flood data sequence in step S2, and establish a historical alarm flood data sequence library so that only the earliest alarm event description information is retained for each parameter.

[0052] In one embodiment, the number of duplicate alarms and alarm parameters in the original alarm flooding data sequence is determined based on preset rules, and a historical alarm flooding data sequence library is established.

[0053] Specifically, the embodiments of the present invention perform duplicate alarm rejection processing through the following steps:

[0054] The first step involves extracting multiple data sequence segments sequentially from the alarm flood data sequence to be processed (either the original alarm flood data sequence or the alarm flood data sequence to be adjusted described below) using a first sliding window with a length equal to the number of data points and a first sliding step. Starting with the first data point of the alarm flood data sequence to be processed, the first data sequence segment is extracted using the first sliding window. Then, the first sliding window is moved by the first sliding step to extract the second data sequence segment, and so on, until the last data point in the last extracted data sequence segment is the last data point in the alarm flood data sequence to be processed. The window sliding then ends (the data sequence segment extraction action ends), thus obtaining multiple data sequence segments.

[0055] The second step involves sequentially removing excessive alarm data from each extracted data sequence segment according to a preset duplicate alarm removal principle. The data is then integrated based on alarm time, resulting in a processed sequence of data sequence segments that meet the preset duplicate alarm removal principle. Specifically, data sequence segments that do not meet the preset duplicate alarm removal principle are deleted.

[0056] In this embodiment of the invention, the above-mentioned preset duplicate alarm elimination principle includes: deleting the data points between the earliest alarm event description information of each alarm parameter in the truncated data sequence segment and the alarm event description information between the next non-self parameter, and the number of remaining alarm parameters exceeds a preset number (e.g., 5).

[0057] After all the captured data sequence segments have been processed to remove flood alarm data, the remaining captured data sequence segments that meet the preset duplicate alarm removal criteria are the captured data sequence segments that have not been deleted. The remaining captured data sequence segments are the captured data sequence segments after the duplicate alarm removal process.

[0058] After establishing the original alarm flooding data sequence, a duplicate alarm removal mechanism is implemented. Duplicate alarm removal rules are set: if the same parameter continuously triggers an alarm, all other alarm records for that parameter are deleted, retaining only the earliest alarm record for that parameter. The number of deleted entries is equal to the number of times between the earliest alarm for that parameter and the next alarm triggered by that parameter within the original alarm flooding data sequence. For example, if a data sequence segment is [PT-863, PIC-511, PIC-511, PIC-511, PIC-511, FIC-611], after applying the above removal rules and processing the duplicate alarms, the new data sequence segment obtained is [PT-863, PIC-511, FIC-611], with 3 entries deleted. Then, after the duplicate alarms are removed, the number of remaining alarm parameters in the new extracted data sequence segment is determined. If the number is ≥5 (if there are too few alarm parameters, there will be fewer effective internal relationships and information between the parameters), the sequence is added to the historical alarm flood data sequence library; otherwise, the sequence segment is deleted.

[0059] Thus, after removing duplicate alarms from the original alarm flood data sequence, a historical alarm flood data sequence library for alarm parameter prediction is obtained.

[0060] Next, step S120 will also predict the alarm parameter similarity prediction sequence corresponding to the latest alarm segment based on the constructed historical alarm flood data sequence library and the obtained latest alarm parameter flood data sequence.

[0061] In one embodiment, the latest alarm parameter flooding data sequence is first obtained, and then the alarm parameter similarity prediction sequence corresponding to the latest alarm segment is predicted by comparing the obtained latest alarm parameter flooding data sequence with the historical alarm flooding data sequence library.

[0062] In one embodiment, step S120 first obtains the latest alarm parameter flood data sequence through the following steps: Step T1, obtain the original alarm data sequence within the latest alarm collection time period (e.g., the most recent 1 minute), and insert the original alarm data sequence within the latest alarm collection time period into the end of the historical alarm flood data sequence library to form the alarm flood data sequence to be adjusted; Step T2, use the above-mentioned duplicate alarm removal rules to perform duplicate alarm removal processing on the alarm flood data sequence to be adjusted to obtain an updated historical alarm flood data sequence library; Step T3, extract the data segment within the latest alarm collection time period from the updated historical alarm flood data sequence library to obtain the latest alarm parameter flood data sequence.

[0063] Obtain the latest alarm record information of the refining unit to be evaluated, and implement the above-mentioned duplicate alarm rejection rules, implement the duplicate alarm forward rejection mechanism and the alarm flood sequence forward identification mechanism, and remove duplicate alarms in reverse order of alarm occurrence time to obtain the flood sequence with the latest alarm as the end of the sequence, thereby obtaining the latest alarm parameter flood data sequence.

[0064] Then, step S120 will predict the alarm parameter prediction sequence corresponding to the latest alarm segment by comparing the latest alarm parameter flood data sequence with the historical alarm flood data sequence library.

[0065] Specifically, firstly, multiple data sequence segments to be evaluated are sequentially extracted from the updated historical alarm flood data sequence library according to a second sliding window with a second number of data points and a second sliding step size. In one embodiment, the second sliding step size is preferably 1.

[0066] Starting with the first data point of the updated historical alarm flood data sequence to be processed, the first data sequence segment to be evaluated is captured using a second sliding window. Then, the second sliding window is moved by a second sliding step to capture the second data sequence segment to be evaluated, and so on, until the last data point in the last captured data sequence segment to be evaluated is the last data point in the updated historical alarm flood data sequence to be processed. The window sliding ends (the data sequence segment capture action ends), thus obtaining multiple data sequence segments to be evaluated.

[0067] Then, the similarity between the latest alarm parameter flood data sequence and each data sequence segment to be evaluated is calculated.

[0068] One or more of the following similarity calculation methods are used to calculate the similarity between the latest alarm parameter flood data sequence and each data sequence segment to be evaluated: Euclidean distance, Manhattan distance, cosine similarity, Pearson correlation coefficient, Jaccard similarity coefficient, and Hamming distance.

[0069] Next, based on the similarity calculation results of each data sequence segment to be evaluated, the data sequence segments that meet the preset sequence similarity extraction conditions are selected as the target data sequences.

[0070] In this embodiment of the invention, the preset sequence similarity extraction conditions include: the similarity calculation result exceeds the preset similarity coefficient threshold (e.g., 80%), and the starting parameter number in the current data sequence segment to be evaluated is the latest alarm parameter number, and the remaining data sequence segments to be evaluated with the highest similarity calculation result satisfying the above two conditions are also included.

[0071] In addition, if no data sequence fragments that meet the above-mentioned preset sequence similarity extraction conditions are selected, or if the latest alarm parameter is the last one in the (updated) historical alarm flood data sequence library, then the current alarm parameter prediction sequence is set to empty.

[0072] Finally, the position of the latest alarm parameter (i.e., the last parameter in the current flood of alarm parameters) in the selected target data sequence is located, and the data segments arranged after the latest alarm parameter are extracted from the target data sequence as alarm parameter similarity prediction sequences.

[0073] The similarity between the latest alarm parameter flooding data sequence and the historical alarm flooding data sequence library is calculated. Similarity calculation methods include Euclidean distance, Manhattan distance, cosine similarity, Pearson correlation coefficient, Jaccard similarity coefficient, Hamming distance, etc., and one or more similarity methods can be selected. Based on preset similarity diagnosis rules, the most similar historical alarm flooding sequence is found from the flooding sequence library. In this case, a higher similarity indicates that the two sequences share more alarm events, have a stronger positive correlation, and have more similar trends and fluctuation patterns. Finally, the position of the latest alarm parameter in the selected target data sequence is located, and the next one or several parameters that may trigger an alarm are selected to obtain the candidate alarm parameter prediction sequence A, denoted as the alarm parameter similarity prediction sequence.

[0074] It should be noted that if no corresponding target data sequence is extracted based on the latest alarm, the candidate alarm parameter prediction sequence A is empty; if the target data sequence of the latest alarm is extracted, but the similarity does not meet the preset rules, the candidate alarm parameter prediction sequence A is empty; if the target data sequence of the latest alarm parameters is extracted, but after location, the latest alarm is the last one in the historical alarm flood data sequence library, then the candidate alarm parameter prediction sequence A is also empty.

[0075] For example, the cosine similarity method and the Jaccard similarity coefficient method are selected, and the final sequence similarity is obtained by averaging the results of the two methods. This allows the target data sequence with the highest similarity to be found from the historical alarm flood data sequence database. If the similarity is less than 80% (determined based on expert knowledge), the two sequences are considered to have a weak similarity relationship, and no similarity relationship is established.

[0076] After generating the predicted alarm parameter sequence, step S130 is used to predict the alarm parameter sequence based on real-time running process data.

[0077] Next, in step S130, based on the latest alarm parameter flood data sequence obtained in step S120, the alarm parameter model prediction sequence is obtained using the fusion alarm prediction model.

[0078] In this embodiment of the invention, the fusion alarm prediction model is an integrated learning model built using the Stacking concept.

[0079] In one embodiment, the fusion alarm prediction model includes: multiple parallel initial prediction models and an integrated prediction model, wherein the integrated prediction model is connected to the output of each initial prediction model.

[0080] refer to Figure 4 In step S130, firstly, the latest alarm parameter flood data sequence obtained above is input to the input end of each initial prediction model, and multiple sets of alarm parameter initial prediction sequences are output. Then, each set of alarm parameter initial prediction sequences is input to the integrated prediction model, and the alarm parameter model prediction sequence (i.e., candidate abnormal parameter prediction sequence B) is output to fuse multiple prediction results.

[0081] In addition, before step S130 is implemented, the alarm event prediction method described in this embodiment of the invention further includes the step of constructing a fusion alarm prediction model.

[0082] In one embodiment, a fusion alarm prediction model is constructed through the following steps.

[0083] Step 1: According to the duplicate alarm rejection rules set in step S120, perform duplicate alarm rejection processing on the original alarm data sequence constructed in step S110 to establish an initial alarm prediction data sequence.

[0084] Step 2: According to the third sliding window with the third number of data points and the third sliding step size, extract multiple sets of sequence segments as training input data from the initial alarm prediction data sequence constructed in Step 1 to obtain multiple sets of training data segments.

[0085] like Figure 3 As shown, starting with the first data point of the initial alarm prediction data sequence to be processed, the first training data segment is extracted using the third sliding window. Then, the third sliding window is moved by the third sliding step to extract the second training data segment, and so on, until the last data point in the last extracted training data segment is the last data point in the initial alarm prediction data sequence to be processed. The window sliding ends (the data sequence segment extraction action ends), thus obtaining multiple training data segments.

[0086] Step 3: Extract the next alarm parameter from the last alarm parameter in each training data segment from the initial alarm prediction data sequence, and use it as the training output data for the corresponding training data segment. Then, in step 4, establish the first alarm prediction dataset based on all training data segments and the corresponding training output data.

[0087] The method in step S120 is used to remove duplicate alarms. Based on preset sampling rules, the alarm history is reconstructed, transforming the original alarm history into an initial alarm prediction dataset that the model can understand and process. The sampling rule adopts a sliding window segmentation method based on the alarm parameter tag number (i.e., sequence data points). This aims to extract record segments of a specific length from continuous alarm history. A fixed window length is set for each segmentation, determining the range of records contained in each segment. To fully utilize the information in the alarm history, a sliding method is used to segment the records. This means that after each segmentation, the window slides backward by a certain step size to generate the next sub-record segment. The choice of step size needs to consider both the continuity of the records and the training requirements of the model. A smaller step size can generate more data segments but may lead to higher redundancy between segments; while a larger step size may result in information loss.

[0088] like Figure 3 As shown, taking an alarm history record with 12 alarms as an example (only the alarm parameter tag number field is retained), according to the above sliding window segmentation method (sliding window length is 10, sliding step size is 1), a continuous alarm record sequence is finally discretized to obtain two fixed-length alarm prediction learning samples that can be used for supervised learning.

[0089] In this way, a series of sub-record fragments can be generated, which will serve as inputs to the model. Simultaneously, to build a supervised ensemble learning model, we also need to find the corresponding output for each input fragment, using data from the next time step of the same length as the output. This process is repeated sequentially, transforming the original alarm history records into a complete first alarm prediction dataset. This first alarm prediction dataset not only contains a large number of samples needed for model training but also retains the temporal order and correlation information from the original data, helping the model learn accurate prediction patterns subsequently.

[0090] After completing the construction of the first alarm prediction dataset, this embodiment of the invention will also use the currently constructed first alarm prediction dataset to construct a fusion alarm prediction model.

[0091] Further, in step five, using the first alarm prediction dataset, the pre-built multiple initial learning models are trained and validated respectively to obtain multiple initial prediction models. In step six, using the training data segments as input and the initial model prediction sequences output by each initial prediction model, along with the aforementioned training output data, a second alarm prediction dataset is established. Finally, in step seven, using this second alarm prediction dataset, the pre-built ensemble learning model is further trained and validated to obtain an ensemble prediction model.

[0092] Furthermore, before training and validating the initial prediction model, the alarm event prediction method described in this embodiment of the invention further includes: uniquely encoding the alarm parameter tag numbers in the first alarm prediction dataset and performing standardization and normalization processing. Moreover, before training and validating the integrated prediction model, the alarm event prediction method described in this embodiment of the invention further includes: uniquely encoding the alarm parameter tag numbers in the second alarm prediction dataset and performing standardization and normalization processing. The standardization and normalization processing is a max-min normalization process.

[0093] In this embodiment of the invention, a fusion alarm prediction model is established using the Stacking concept. In the initial learning phase, multiple initial learning models are constructed and combined to complete the initial alarm prediction task. A first alarm prediction dataset is used to train and validate each initial learning model, obtaining initial alarm prediction results. In the ensemble learning phase, the alarm prediction results of the initial learning models are used as new input features and fused with the original prediction features in the first alarm prediction dataset to establish a second alarm prediction dataset. This second alarm prediction dataset is used to train and validate the ensemble learning model, achieving the final prediction of the next or several possible alarms. The trained fusion alarm prediction model is then used to generate a candidate alarm parameter prediction sequence B online.

[0094] In this embodiment of the invention, at least two initial learning models are selected from the following set of candidate models, and one ensemble learning model is selected from the following set of candidate models. The types of initial learning models and selected ensemble learning models do not overlap. The set of candidate models includes, but is not limited to: decision tree models, support vector machine models, K-nearest neighbor models, logistic regression models, Naive Bayes models, factorization machine models, gradient boosting tree models, and random forest models.

[0095] like Figure 4As shown, a fusion alarm prediction model is established using the Stacking concept. In the initial learning phase, four initial learning models (selected as Support Vector Machine, K-Nearest Neighbors, Factorization Machine, and Gradient Boosting Tree) are constructed to complete the initial alarm prediction task. Each initial learning model is trained using the first alarm prediction dataset to obtain the initial alarm prediction results. In the ensemble learning phase, the alarm prediction results of the initial learning models are fused and used as new model input features to train the ensemble learning model (selected as Logistic Regression), achieving the final prediction of the next or several possible alarms, in order to generate a candidate alarm parameter prediction sequence B.

[0096] It should be noted that, in order to facilitate model training and eliminate the influence of inconsistent units, before training the initial prediction model and the ensemble prediction model, this embodiment of the invention also adopts a unique coded numerical representation (e.g., PT-863 is represented by the numerical value 1, PIC-511 is represented by the numerical value 2, etc.) and maximum and minimum value normalization processing (to map the coded numerical values ​​to the range [0,1]) for the alarm parameter tag numbers in the first alarm prediction dataset and the second alarm prediction dataset, and adopts inverse normalization and inverse encoding representation for the prediction results.

[0097] Therefore, after completing the construction of the fusion alarm prediction model and obtaining the alarm parameter model prediction sequence, the process proceeds to step S140, where the prediction results of the two types of data sequences are fused.

[0098] Step S140: The alarm parameter similarity prediction sequence obtained in step S120 and the alarm parameter model prediction sequence obtained in step S130 are fused to obtain the alarm parameter prediction result sequence.

[0099] In this embodiment of the invention, the alarm parameter prediction sequence and the subsequent alarm parameter sequence can be fused according to one or more of the following fusion rules: intersection principle between sequences, union principle between sequences, parameter priority principle, time interval priority principle, historical alarm count priority principle, and historical alarm duration priority principle.

[0100] When performing sequence fusion processing according to the principle of taking the intersection between sequences, the new sequence obtained by taking the intersection of the alarm parameter prediction sequence and the subsequent alarm parameter sequence is taken as the alarm parameter prediction result sequence.

[0101] When performing sequence fusion processing according to the principle of taking the union of sequences, the new sequence obtained by taking the union of the alarm parameter prediction sequence and the subsequent alarm parameter sequence is used as the alarm parameter prediction result sequence.

[0102] When performing sequence fusion processing according to the parameter priority principle, the different alarm parameters are ranked according to their importance in the entire refining unit to be evaluated. Each alarm parameter appearing in the alarm parameter prediction sequence and subsequent alarm parameter sequences is sorted, and the resulting new sequence is used as the alarm parameter prediction result sequence.

[0103] When performing sequence fusion processing according to the time interval priority principle, the alarm start time of each alarm parameter appearing in the alarm parameter prediction sequence and the subsequent alarm parameter sequence is first counted. Then, all alarm parameters appearing in these two data sequences are sorted according to the order of their alarm start times. The resulting new sequence is used as the abnormal process parameter prediction result sequence.

[0104] When performing sequence fusion processing based on the principle of prioritizing historical alarm counts, the total number of alarms for each alarm parameter in the alarm parameter prediction sequence and subsequent alarm parameter sequences is first counted. Then, all alarm parameters appearing in these two data sequences are sorted according to the total number of alarms, and the resulting new sequence is used as the alarm parameter prediction result sequence.

[0105] When performing sequence fusion processing according to the principle of prioritizing historical alarm duration, the total alarm duration of each alarm parameter appearing in the alarm parameter prediction sequence and subsequent alarm parameter sequences is first counted. Then, all alarm parameters appearing in these two data sequences are sorted according to the total alarm duration, and the resulting new sequence is used as the alarm parameter prediction result sequence.

[0106] In addition, the alarm event prediction method described in this embodiment of the invention will output a response to the alarm parameter prediction result sequence after obtaining the final required alarm parameter prediction result sequence.

[0107] In one embodiment, according to a preset number of pushes, the alarm parameters ranked in the top number of pushes in the alarm parameter prediction result sequence obtained in step S140 are sent to the device monitoring and management personnel terminal in a preset form.

[0108] The final alarm parameter prediction sequence is determined, and the alarm parameter push sequence is notified to the device technicians through DCS screen pop-ups, SMS, email, etc., according to the required order and number of pushes, so as to remind them to take relevant technical measures to achieve the purpose of real-time alarm warning and alarm flood elimination.

[0109] Therefore, the core objective of this invention is to conduct in-depth analysis and utilization of historical alarm events generated during the production process of refining and chemical plants. On one hand, through sophisticated data processing techniques, a series of alarm flood sequences are extracted and generated. During this process, duplicate and invalid alarm information is filtered and removed to ensure the accuracy and validity of the data. Simultaneously, the similarity relationships between sequences are calculated to discover hidden patterns and correlations in the data, completing preliminary alarm prediction. On the other hand, to achieve more accurate alarm prediction, a training set is specifically constructed based on the extracted and optimized alarm history records, and an alarm prediction model based on ensemble learning is trained, also completing preliminary alarm prediction. Finally, a parallel fusion strategy is constructed to fuse prediction results from both channels. This dual-channel design increases the accuracy and diversity of predictions. The ability to predict the next impending alarm through this entire technical process framework is extremely important for plant technicians, as this solution effectively assists them in promptly handling abnormal situations. In this way, technicians can react quickly and take corresponding measures to improve the safety and stability of the plant, thereby ensuring the smooth operation of the production process and minimizing potential risks and losses.

[0110] Example 2

[0111] Based on the alarm event prediction method described in Embodiment 1 above, a specific example of an alarm event prediction method is given below.

[0112] Taking a coking unit of a refining and chemical enterprise as an example, the effects of the present invention will be further illustrated.

[0113] Alarm events collected from the DCS system OPC server AE interface from August 30, 2022 to March 31, 2024 were processed to obtain a total of 191,312 DCS alarm records, 6,849 historical alarm flood sequences that met the requirements, and 10,023 alarm prediction datasets.

[0114] Select any alarm at any time in 2024 as the latest alarm to verify the accuracy of the prediction. For example, take the high alarm with parameter LIC-12702 at 23:37:01 on 2024 / 3 / 18 as the latest alarm. The next three alarms are the high alarm with parameter PI-11501 at 23:39:29 on 2024 / 3 / 18, the high alarm with parameter AI-11001-NO at 23:40:52 on 2024 / 3 / 18, and the high alarm with parameter AI-00059 at 23:43:18 on 2024 / 3 / 18.

[0115] Channel 1: The latest alarm LIC-12702's flooded sequence is [LIC-50102, LI-12075, PI-11062, LI-12071, AI-00059, PI-11051, FIC-21301, LIC-12702]. The sequence with the highest similarity matched in the historical alarm flooded sequence library (using data before 2024 / 3 / 18, a total of 6831 sequences) is [PI-11062, PI-11501, LIC-50102, LI-12075, LIC-12702, PI-11501, AI-11001-NO, LI-12071]. The resulting candidate alarm parameter prediction sequence a is [PI-11501, AI-11001-NO, LI-12071].

[0116] Channel 2: The initial learning model and the ensemble learning model are trained using the first 9987 data points (data before 2024 / 3 / 18) as the training set. The ratio of the training set to the validation set is 9:1. Five-fold cross-validation is used to optimize the relevant parameters. The alarm prediction results of the ensemble learning model are output and sorted according to the probability. Finally, the candidate alarm parameter prediction sequence b is obtained as [PI-11501, AI-00059, PI-11602].

[0117] Dual-channel fusion alarm prediction: The intersection of the sequences is PI-11501, and the historical time interval priority principle is AI-11001-NO, LI-12071, AI-00059, PI-11602. The final alarm prediction result is [PI-11501, AI-11001-NO, LI-12071, AI-00059, PI-11602]. It can be seen that, compared with the actual situation, the first two of the subsequent three alarms were accurately predicted. The third alarm in the actual situation is the fourth in the predicted alarm result. Although it is not a complete match, the proposed technology also predicted that the third parameter may have an abnormal situation, which verifies the effectiveness of the technology.

[0118] Example 3

[0119] Based on the alarm event prediction method described in Embodiments 1 and 2 above, this invention also provides a system for predicting alarm events in refining and chemical processing equipment (also called an "alarm event prediction system"), which is used to implement the alarm event prediction method described above.

[0120] Figure 5 This is a schematic diagram of the structure of a system for predicting alarm events in a refining and chemical plant, according to an embodiment of this application. Figure 5As shown, the alarm event prediction system includes: an alarm event acquisition module 501, a first type of prediction sequence generation module 502, a second type of prediction sequence generation module 503, and a prediction result fusion module 504.

[0121] The alarm event acquisition module 501 is implemented according to the method described in step S110 above, and is configured to acquire historical process alarm events of the device to be evaluated from the DCS system and establish an original alarm data sequence based on alarm time. The first type of prediction sequence generation module 502 is implemented according to the method described in step S120 above, and is configured to establish a historical alarm flood data sequence library based on the original alarm data sequence, and predict the alarm parameter similarity prediction sequence corresponding to the latest alarm segment by comparing the acquired latest alarm parameter flood data sequence with the historical alarm flood data sequence library. The second type of prediction sequence generation module 503 is implemented according to the method described in step S130 above, and is configured to obtain the alarm parameter model prediction sequence based on the latest alarm parameter flood data sequence using a fusion alarm prediction model. The prediction result fusion module 504 is implemented according to the method described in step S140 above, and is configured to fuse the alarm parameter similarity prediction sequence and the alarm parameter model prediction sequence to obtain the alarm parameter prediction result sequence.

[0122] Furthermore, the alarm event acquisition module 501 has a data acquisition interface based on the OPC AE protocol, which enables the DCS system to actively push the historical process alarm events of the device to be evaluated to the outside world through the AE interface using a broadcast mechanism.

[0123] The alarm event acquisition module is a data acquisition interface developed based on the OPC AE protocol, supporting the OPC AE protocol. This module can acquire alarm events from the DCS system and store them in the process alarm history event database. The OPC AE protocol uses a broadcast mechanism for alarm events, meaning that generated alarm events are actively pushed outward from the DCS system through the AE interface. Therefore, the alarm event acquisition module 501 described in this embodiment does not need to set a data acquisition frequency; it can receive alarm events by always keeping the module in a listening or subscribing state. Then, it stores the alarm events in chronological order in the process alarm history event database, enabling the first type of prediction sequence generation module 502 and the second type of prediction sequence generation module 503 to use the process alarm history events of the device to be evaluated in the database to predict the corresponding alarm parameter data sequences.

[0124] This invention discloses a method and system for predicting alarm events in refining and chemical plants. The method and system, on the one hand, fully explore alarm flooding sequences in historical alarm events, accurately matching and identifying the latest alarm flooding problem through a preset similarity calculation method. On the other hand, it establishes an ensemble learning model based on the Stacking concept, selecting multiple learners to predict upcoming alarms. Finally, through a reasonable judgment mechanism and appropriate fusion strategy, the prediction results from the two channels are effectively integrated to improve the accuracy and reliability of the prediction model. This invention minimizes the occurrence of abnormal operating conditions in production units caused by alarm flooding, untimely alarm handling, delayed intervention, late abnormal intervention, and inaccurate predictions, improving the emergency response capabilities of plant operators. By fully utilizing actual production process alarm events occurring in the DCS system for next alarm prediction, the accuracy and predictability of alarms can be improved, thereby enhancing the enterprise's safety production level.

[0125] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0126] In the description of this invention, unless otherwise stated, "a plurality of" means two or more; the terms "upper," "lower," "left," "right," "inner," "outer," "front end," "rear end," "head," "tail," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0127] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "connected" and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0128] It should be understood that the embodiments disclosed herein are not limited to the specific structures, processing steps, or materials disclosed herein, but should be extended to equivalent substitutions of these features as understood by those skilled in the art. It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting.

[0129] The phrase "an embodiment" or "an embodiment" used in this specification means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the invention. Therefore, the phrase "an embodiment" or "an embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment.

[0130] While the embodiments disclosed in this invention are as described above, the content is merely for the purpose of facilitating understanding of the invention and is not intended to limit the invention. Any person skilled in the art to which this invention pertains may make any modifications and changes in form and detail of the implementation without departing from the spirit and scope disclosed herein; however, the scope of patent protection of this invention shall still be determined by the scope defined in the appended claims.

Claims

1. A method for predicting refinery alarm events, the method comprising: The method comprises the following steps: obtaining process alarm historical events of a device to be evaluated from a DCS system, and establishing an original alarm data sequence based on alarm time arrangement; according to the original alarm data sequence, establishing a historical alarm flooding data sequence library, and predicting an alarm parameter similar prediction sequence corresponding to a current latest alarm segment by comparing a latest alarm parameter flooding data sequence obtained with the historical alarm flooding data sequence library; according to the latest alarm parameter flooding data sequence, obtaining an alarm parameter model prediction sequence by using a fusion alarm prediction model; fusing the alarm parameter similar prediction sequence and the alarm parameter model prediction sequence to obtain an alarm parameter prediction result sequence.

2. The method of claim 1, wherein, In the step of establishing a historical alarm flooding data sequence library according to the original alarm data sequence, the method comprises the following steps: identifying data sequence segments with alarm flooding phenomenon from the original alarm data sequence, and establishing an original alarm flooding data sequence; performing repeated alarm elimination processing on the original alarm flooding data sequence, and establishing a historical alarm flooding data sequence library to retain only the earliest occurring alarm event description information for each parameter.

3. The method of claim 2, wherein, The latest alarm parameter flooding data sequence is obtained by the following steps: obtaining an original alarm data sequence in a latest alarm collection time period, and inserting the original alarm data sequence in the latest alarm collection time period into the end of the historical alarm flooding data sequence library to form a to-be-adjusted alarm flooding data sequence; performing repeated alarm elimination processing on the to-be-adjusted alarm flooding data sequence to obtain an updated historical alarm flooding data sequence library; extracting a data segment in the latest alarm collection time period from the updated historical alarm flooding data sequence library to obtain the latest alarm parameter flooding data sequence.

4. The method according to claim 2 or 3, characterized in that, The repeated alarm elimination processing is performed by the following steps: extracting a plurality of intercepted data sequence segments from the alarm flooding data sequence to be processed according to a first sliding window with a first data point number length and a first sliding step; according to a preset repeated alarm elimination principle, performing flooding alarm data elimination processing on each intercepted data sequence segment, and integrating the intercepted data sequence segments that meet the preset repeated alarm elimination principle according to alarm time to obtain a corresponding processing result sequence, wherein the intercepted data sequence segments that do not meet the preset repeated alarm elimination principle are deleted, the preset repeated alarm elimination principle comprises: deleting, in each alarm parameter of the intercepted data sequence segment, data points between the earliest occurring alarm event description information of the alarm parameter and alarm event description information of the next non-own parameter, and the number of remaining alarm parameters exceeds a preset number.

5. The method according to any one of claims 1 to 4, characterized in that, In the step of predicting an alarm parameter similar prediction sequence corresponding to a current latest alarm segment by comparing a latest alarm parameter flooding data sequence obtained with the historical alarm flooding data sequence library, the method comprises the following steps: extracting a plurality of to-be-evaluated data sequence segments from the historical alarm flooding data sequence library according to a second sliding window with a second data point number length and a second sliding step; Calculate the similarity between the latest alarm parameter flooding data sequence and each to-be-evaluated data sequence segment, respectively; According to the similarity calculation result of each to-be-evaluated data sequence segment, screen a to-be-evaluated data sequence segment that meets a preset sequence similarity extraction condition as a target data sequence, wherein the preset sequence similarity extraction condition includes: the similarity calculation result exceeds a preset similarity coefficient threshold, the starting point parameter bit number in the to-be-evaluated data sequence segment is the latest alarm parameter bit number, and the to-be-evaluated data sequence segment has the highest similarity calculation result; Locate the position of the latest alarm parameter in the target data sequence, and extract a data segment arranged after the latest alarm parameter from the target data sequence as the alarm parameter prediction sequence.

6. The method of claim 5, wherein, The similarity between the latest alarm parameter flooding data sequence and each to-be-evaluated data sequence segment is calculated by using one or more of the following similarity calculation methods: Euclidean distance, Manhattan distance, cosine similarity, Pearson correlation coefficient, Jaccard similarity coefficient, and Hamming distance.

7. The method according to claim 5 or 6, characterized in that, The method further includes: if no to-be-evaluated data sequence segment that meets the preset sequence similarity extraction condition is screened out or the latest alarm parameter is the last one in the historical alarm flooding data sequence library, setting the current alarm parameter similarity prediction sequence as empty.

8. The method according to any one of claims 1 to 7, characterized in that, The fusion alarm prediction model includes multiple parallel initial prediction models and an integrated prediction model connected to the output end of each initial prediction model, wherein in the step of obtaining an alarm parameter model prediction sequence by using the fusion alarm prediction model according to the latest alarm parameter flooding data sequence, the following steps are included: Input the latest alarm parameter flooding data sequence into the input end of each initial prediction model, and output multiple alarm parameter initial prediction sequences; Input each group of alarm parameter initial prediction sequences into the integrated prediction model, and output the alarm parameter model prediction sequence.

9. The method of claim 8, wherein, The method further includes constructing the fusion alarm prediction model, which includes: Perform repeated alarm elimination processing on the original alarm data sequence to establish an initial alarm prediction data sequence; According to a third sliding window with a third data point number length and a third sliding step, sequentially extract multiple groups of sequence segments for training input data from the initial alarm prediction data sequence; Extract the next alarm parameter of the last alarm parameter in each group of training data segments from the initial alarm prediction data sequence as the training output data of the corresponding training data segment; Establish a first alarm prediction data set according to all groups of training data segments and corresponding training output data; Train and verify multiple initial learning models pre-constructed by using the first alarm prediction data set, and obtain multiple initial prediction models; Establish a second alarm prediction data set by using the training data segments as input and the initial model prediction sequences output by each initial prediction model, and the training output data; Train and verify an integrated learning model pre-constructed by using the second alarm prediction data set, and obtain an integrated prediction model.

10. The method of claim 9, wherein, Before training and verifying the initial prediction model / the integrated prediction model, the method further comprises: uniquely coding the alarm parameter bit number in the first alarm prediction data set / the second alarm prediction data set into a numerical value and performing standardization normalization processing, wherein the standardization normalization processing is max-min normalization processing.

11. The method according to claim 9 or 10, characterized in that, The plurality of initial learning models selects at least two models from the following alternative model set, and the integrated learning model selects one from the following alternative model set, the category of the selected initial learning model and the category of the selected integrated learning model are not repeated, wherein, The alternative model set includes: decision tree model, support vector machine model, K nearest neighbor model, logistic regression model, naive Bayes model, factor decomposition machine model, gradient boosting tree model and random forest model.

12. The method according to any one of claims 1 to 11, characterized in that, The alarm parameter prediction sequence and the subsequent alarm parameter sequence are fused according to one or a combination of several of the following fusion rules: sequence intersection principle, sequence union principle, parameter priority principle, time interval priority principle, historical alarm frequency priority principle and historical alarm duration priority principle.

13. The method according to any one of claims 1 to 12, characterized in that, The method further comprises: outputting a response to the alarm parameter prediction result sequence, which includes: sending a notification to the device monitoring and management personnel terminal in a preset form according to a preset push quantity, and the alarm parameters ranked in the top push quantity in the alarm parameter prediction result sequence.

14. The method according to any one of claims 1 to 13, characterized in that, In the step of establishing the original alarm data sequence based on alarm time arrangement, it comprises: According to the process alarm history event of the to-be-evaluated device, the alarm events of the same alarm parameter are identified according to the alarm time, so that the adjacent alarm time points form alarm event description information of different alarm parameters, so as to form the original alarm data sequence, and the alarm event description information includes the name of the refining device, the alarm parameter bit number, the alarm event description, the start time and the end time.

15. A system for predicting refinery alarm events, the system comprising: The system is realized by using the method of any one of claims 1-14, wherein the system comprises: An alarm event acquisition module configured to obtain process alarm history events of a to-be-evaluated device from a DCS system and establish an original alarm data sequence based on alarm time arrangement; A first type of prediction sequence generation module configured to establish a historical alarm flooding data sequence library according to the original alarm data sequence, and predict an alarm parameter similar prediction sequence corresponding to the current latest alarm segment by comparing the obtained latest alarm parameter flooding data sequence with the historical alarm flooding data sequence library; A second type of prediction sequence generation module configured to obtain an alarm parameter model prediction sequence by using a fusion alarm prediction model according to the latest alarm parameter flooding data sequence; A prediction result fusion module configured to fuse the alarm parameter similar prediction sequence and the alarm parameter model prediction sequence to obtain an alarm parameter prediction result sequence.

16. The system of claim 15, wherein The alarm event acquisition module has a data acquisition interface based on the OPC AE protocol, which enables the DCS system to actively push the historical process alarm events of the device to be evaluated to the outside world through the AE interface using a broadcast mechanism.

Images