Method for determining alarm priority of petrochemical enterprise

Abstract

The invention discloses a petrochemical enterprise alarm priority determination method, which comprises the steps of performing information extraction on an HAZOP report of a petrochemical enterprise by using a natural language processing technology to obtain consequence severity of a potential risk point; response time is determined through dynamic simulation software, and the urgency demand degree is obtained based on the response time; and based on the consequence severity and the urgency demand degree, determining the alarm priority of the petrochemical enterprise through a predefined matrix method. According to the method, HAZOP analysis which must be carried out by an enterprise is taken as a reference of a potential risk point consequence, and meanwhile, the response time is clearly defined by depending on ASPEN dynamic simulation. The invention aims to combine the existing knowledge and data, reduce the labor consumption, quickly determine the alarm priority, provide scientific and effective reference for operators, and play a role in auxiliary decision making in the rationalization stage of the alarm management life cycle.

Description

Technical Field

[0001] The invention belongs to the technical field of alarm level setting of enterprise production parameters, and in particular relates to a method for determining alarm priority of a petrochemical enterprise. Background Art

[0002] As a process hazard analysis technology, HAZOP not only considers the inherent risks of equipment, but also the dynamic risks of state parameter deviations in the process flow. It can comprehensively predict the dangers of chemical processes and plays a major role in improving plant safety and reliability. It is the most popular qualitative evaluation method in chemical safety assessment. HAZOP analysis takes the deviation of the node as the starting point, analyzes the causes, consequences, and protective measures that may cause the deviation to derive the risk level, and proposes countermeasures and suggestions for the scenarios with higher risks. The analysis results are retained in the form of report text. HAZOP reports are characterized by longer statements, more professional expressions, and unique logic and semantics.

[0003] HAZOP is widely used in petrochemical enterprises, and it is feasible to extract the required knowledge from HAZOP reports. However, the existing methods for determining alarm priorities in petrochemical enterprises still have the following problems: lack of a comprehensive method for qualitative and quantitative analysis, insufficient automated processing and knowledge extraction of HAZOP reports, high labor costs in determining alarm priorities, strong subjectivity in the process of determining alarm priorities, and lack of standardized processes. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes a method for determining alarm priority of a petrochemical enterprise to solve the problems existing in the above prior art.

[0005] To achieve the above object, the present invention provides a method for determining alarm priority of a petrochemical enterprise, comprising:

[0006] Use natural language processing technology to extract information from HAZOP reports of petrochemical enterprises and obtain the severity of the consequences of potential risk points;

[0007] Determine the response time by dynamic simulation software, and obtain the urgency of the demand based on the response time;

[0008] Based on the severity of the consequences and the urgency of the need, the alarm priority of the petrochemical enterprise is determined by a predefined matrix method.

[0009] Preferably, the process of extracting information from the HAZOP report of a petrochemical enterprise includes:

[0010] Based on the Chinese entity recognition model of the natural language processing technology, text recognition and entity extraction are performed on the text of the HAZOP report to obtain entity information;

[0011] The text similarity matching based on the natural language processing technology is used to match the deviation and consequence information in the HAZOP report and the alarm control point ledger, ensuring that the consequence level in the HAZOP report is filled in the consequence level of the alarm control point to obtain the consequence severity of the potential risk point.

[0012] Preferably, the process of performing text recognition and entity extraction on the text of the HAZOP report includes:

[0013] Adopting the LE-NEZHA layer in the Chinese entity recognition model, the word vectors in the entity dictionary are integrated into the neural context representation model for Chinese understanding through a dictionary adapter;

[0014] Extract long-range semantic information from text using neural context representation models;

[0015] Adopt iterative dilated convolutional neural network to learn more long-range semantic features;

[0016] The CRF layer is introduced to constrain the legitimacy of the label output, and the entity recognition result is obtained based on the output label probability.

[0017] Preferably, the process of matching the deviation and consequence information in the HAZOP report and the alarm control point ledger based on the text similarity matching of the natural language processing technology includes:

[0018] Performing text recognition on the deviation line and specific deviation line in the HAZOP report and the deviation line and control point in the alarm control point ledger to obtain a first concept set and a second concept set;

[0019] Select an alarm control point from the first concept set, calculate the text similarity between the alarm control point and the second concept set using the ESIM-Bert method, and obtain the highest matching result point;

[0020] The consequence row and consequence level in the highest point of the matching result are selected and filled into the corresponding row of the alarm control point ledger to obtain the consequence level of all control points.

[0021] Preferably, the process of determining the response time by dynamic simulation software includes:

[0022] Aspen Plus software was used to perform steady-state simulation on the HAZOP report information to obtain a steady-state model;

[0023] The steady-state model is converted into a dynamic model and verified using Aspen Dynamics software;

[0024] Some parameters in the dynamic model are set to deviate from the normal state, and the propagation of the deviation to the critical deviation is simulated to obtain the response time.

[0025] Preferably, the predefined matrix method includes: a first level classification of consequence severity and a second level classification of urgency, and a plurality of alarm priorities are determined based on the first level classification and the second level classification.

[0026] Preferably, when determining the alarm priority, the influence of additional conditions such as the safety instrumented system and protection layer analysis is considered.

[0027] Compared with the prior art, the present invention has the following advantages and technical effects:

[0028] The present invention provides a method for determining the alarm priority of a petrochemical enterprise, comprising: extracting information from a HAZOP report of the petrochemical enterprise using natural language processing technology to obtain the consequence severity of a potential risk point; determining a response time using dynamic simulation software, and obtaining an urgency requirement based on the response time; and determining the alarm priority of the petrochemical enterprise using a predefined matrix method based on the consequence severity and the urgency requirement.

[0029] The present invention provides a consequence severity-response time matrix method, combining qualitative and quantitative methods, and proposes a more practical and feasible method for determining alarm priority to adapt to the actual situation of petrochemical enterprises. The present invention is based on the HAZOP analysis that enterprises must conduct, and uses it as a benchmark for the consequences of potential risk points, while relying on ASPEN dynamic simulation to clearly define the response time. The present invention aims to combine existing knowledge and data, reduce labor consumption, quickly determine the priority of alarms, provide scientific and effective references for operators, and play a role in assisting decision-making in the rationalization stage of the alarm management life cycle. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0031] Figure 1 A flow chart of a method for determining alarm priority according to an embodiment of the present invention;

[0032] Figure 2 A schematic diagram of an LNBIC model according to an embodiment of the present invention;

[0033] Figure 3 Schematic diagram of ordinary convolution and dilated convolution according to an embodiment of the present invention, (a) is ordinary convolution, (b) is dilated convolution;

[0034] Figure 4This is a flowchart of the ESIM-Bert similarity calculation steps of an embodiment of the present invention;

[0035] Figure 5 This is a schematic diagram of the alarm response time of an embodiment of the present invention;

[0036] Figure 6 A schematic diagram of the steps of the HDCT method according to an embodiment of the present invention;

[0037] Figure 7 A category confusion matrix diagram of an embodiment of the present invention;

[0038] Figure 8 is a simulation flow chart of an absorption stabilization system according to an embodiment of the present invention;

[0039] Fig. 9 It is a schematic diagram of the simulation result of the overpressure disturbance of the stabilizing tower according to an embodiment of the present invention;

[0040] Fig.10 It is a schematic diagram of the disturbance simulation result of the liquid level reduction in the reabsorption tower according to an embodiment of the present invention. DETAILED DESCRIPTION

[0041] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0042] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0043] Embodiment 1

[0044] like Figure 1 As shown, this embodiment provides a method for determining the alarm priority of a petrochemical enterprise, including:

[0045] S1. Use natural language processing technology to extract information from the HAZOP reports of petrochemical enterprises and obtain the severity of the consequences of potential risk points;

[0046] Specifically, Figure 1 The consequence determination module shown mainly uses NER technology to match the deviation and consequence information in the HAZOP report and the alarm point ledger to ensure that the consequence level in the HAZOP report is accurately filled into the consequence level of the alarm point. The consequence level and its description are shown in Table 1 below. Different companies may have different setting of the consequence level of HAZOP, so it is necessary to enter it according to the actual situation. This table is only used as an example for general situations.

[0047] Table 1

[0048]

[0049] (1) NER method adapted to Chinese HAZOP reports and alarm point records

[0050] In the Chinese NER task, word segmentation is a key preprocessing step because it can divide Chinese text into words, making it easier to label entities. However, incorrect entity boundaries are often generated due to word segmentation errors. Relatively speaking, the character level will have more advantages and obtain more delicate features and more sufficient semantic information. Therefore, the word fusion method has become a hot research object, and it has been found that introducing entity dictionaries into word fusion can improve the accuracy. In order to propose a NER method that is more suitable for Chinese HAZOP reports and alarm point ledgers, this chapter proposes the LE-NEZHA-Bi-LSTM-IDCNN-CRF Chinese entity recognition model, referred to as the LNBIC model, such as Figure 2 As shown. First, the LE-NEZHA (Lexicon-Natural Language Efficient Zero-shot Hierarchical Attentions) model is used to fuse the word vectors in the entity dictionary into the neural context representation model for Chinese understanding through a dictionary adapter. This fusion process specifically refers to the merging of dictionary information and text information between each Transformer layer. Next, Bi-LSTM is used to extract long-distance semantic information in the text to avoid losing semantic features in the local context. Furthermore, iterative dilated convolutional neural networks (IDCNN) are used to learn more long-distance semantic features. Finally, CRF is introduced to constrain the legitimacy of the label output, and the final entity recognition result is obtained based on the output label probability.

[0051] LE-NEZHA layer

[0052] NEZHA is based on the Transformer architecture and introduces a new hierarchical attention mechanism, which enables it to capture fine-grained and coarse-grained dependencies in sentences to better process Chinese text. The addition of a lexicon adapter (LE) can improve the performance of the pre-trained language model by incorporating vocabulary information for specific tasks. LE-NEZHA converts Chinese sentences into character and word sequence pairs through a lexicon tree structure. In the input part of the model, character-level features and entity dictionary features are used as input together. In addition, a lexicon adapter is introduced between the Transformer layers to effectively integrate the knowledge in the dictionary into the NEZHA model. The role of this adapter is to effectively combine the lexicon information with the intermediate representation of the model to enhance the performance of the model in entity recognition tasks.

[0053] a. Dictionary tree

[0054] A dictionary tree is a tree data structure that is specifically used to efficiently retrieve strings or words from a dictionary or word set. In order to make more full use of dictionary information, the character sequence in the Chinese sentence was expanded, and each character was paired with a word in the dictionary to obtain a sequence of character-word pairs. Next, a dictionary tree was built based on the entity dictionary. When processing the character sequence of the entire Chinese sentence, the system traverses the sequence and matches it with the dictionary tree to identify and obtain all potential words.

[0055] b. Dictionary adapter

[0056] In the dictionary adapter layer, the dictionary adapter is embedded into the Transformer layer to more fully integrate the dictionary information into the model. The specific operation is to input the character vector and word sequence into the dictionary adapter, and then process them through the attention layer and normalization layer. Finally, the weighted vocabulary information is merged with the character information.

[0057] The input to the dictionary adapter is represented as in is a character vector, which is the output of a Transformer layer in NEZHA. is a set of word embeddings. In order to effectively integrate character vectors and word vectors, a nonlinear transformation is also performed on the word vector information, as shown in Formula 3-1:

[0058]

[0059] In formula 1-1: is the word vector after nonlinear transformation; b1 and b2 both represent the bias of nonlinear transformation scaling; W1 is d c ×d w The matrix d cis the size of the hidden layer in BERT, d w is the dimension size of the word embedding of the entire dictionary; W2 is d c ×d c The matrix of .

[0060] When the word vector passes through the attention layer, the character-word pair attention mechanism is introduced to assign all the i-th character The vocabulary information of Where m is the total number of words assigned, and the similarity of each word (a i ) is calculated as:

[0061]

[0062] Where: W attention is the weight matrix of bilinear attention, is the weighted sum of all words.

[0063] Finally, the weighted word vector and character vector are fused to obtain the fused vector, as shown in Formula 3-4:

[0064]

[0065] The LE-NEZHA model fully integrates character information and dictionary information in this way, and uses the output fusion vector as the input of the IDCNN model. This structural design helps improve entity recognition performance.

[0066] IDCNN Layer

[0067] Traditional convolutional neural networks (CNNs) have problems such as small receptive fields of convolution kernels and difficulty covering long-distance texts. Usually, the network depth needs to be greatly increased to improve this problem, which often leads to computational difficulties and slow model convergence due to the surge in network parameters. IDCNN introduces holes in each layer of convolution kernels to increase the receptive field, and then stacks multiple layers of dilated convolutions together to form a dilated convolution block, which extracts features through iterative calculations. Figure 3 shown.

[0068] The specific implementation process of IDCNN is as follows:

[0069] First, for the input width L of the mth layer m The calculation of is shown in formula 1-5. As the number of layers m increases, the length of the covered sequence increases exponentially, thus effectively covering long sequence texts:

[0070] L m =2 m+1 -1(5)

[0071] c i As the joint input representation of LE-NEZHA, the output of the next layer is obtained by performing dilated convolution on the input of the previous layer. As the number of dilated convolution layers increases, the receptive field area of ​​the convolution kernel also increases. The convolution calculation process of each layer is shown in formula 1-6, where D represents the input width L m The mth layer of convolution, ReLU is the activation function, and the output of each layer is expressed as:

[0072]

[0073] Multiple dilated convolutional layers are stacked together to form a dilated convolutional block, where B() represents the convolutional block. Then the convolutional block is iterated k times to obtain As shown in formula (7):

[0074]

[0075] Finally, the output results of the iterative dilated convolution block are randomly discarded at a certain ratio to improve the generalization ability of the model. Dropout is the discard function, is the final model probability output, as shown in formula (8):

[0076]

[0077] (2) Similarity calculation based on ESIM-Bert

[0078] After entity recognition of the text, text similarity matching needs to be performed as needed so that the form can be filled in correctly. Text similarity matching is mainly divided into traditional text matching technology, such as word frequency-text frequency mining method, and deep learning-based text matching technology, such as deep structure speech model. Traditional text matching technology performs statistics based on word frequency but ignores its own semantic information, while deep learning-based text matching technology combines the word vector obtained by unsupervised learning of text data with neural network to calculate the similarity between texts. Deep learning-based text matching technology is further divided into representation-based deep matching method and interactivity-based deep matching method. The representation method converts text information into tensor representation through deep neural network, and obtains the similarity score of text by combining tensor information. The interactive method interacts the text information to be matched and judges the semantic similarity of two texts from the interactive features. Compared with the representation-based text matching method, the interactive method interacts the two texts to be matched, and in order to better obtain the interactive features between text information, an interaction matrix is ​​usually constructed to further extract the semantic information of the text.

[0079] Therefore, the ESIM-Bert method was selected for text similarity matching. The Enhanced Sequential Inference Model (ESIM) uses enhanced Bi-LSTM to implement information reasoning and encoding, uses the Attention mechanism to interact with the encoded information between sentences, and finally splices the vectors to obtain the degree of semantic matching between texts. BERT uses Transformer Decoder (including Masked Multi-HeadAttention) as an extractor and uses the corresponding mask training method to utilize all contextual information of each word in the encoding process of the input text. This section combines and improves these two algorithms, changing the method of obtaining text features through Bi-LSTM in ESIM to using BERT to obtain text features, so as to improve the accuracy of similarity calculation.

[0080] The required original document is input to obtain the text entity recognition result, and the ESIM-Bert method is used to calculate the similarity. The specific algorithm steps are as follows Figure 4 As shown:

[0081] Input: HAZOP report of the device and alarm control point record

[0082] Step 1: Use the LNBIC algorithm introduced in Section 1.1.1 to perform text recognition on the deviation lines and specific deviation lines in the HAZOP reports of the two documents and the deviation lines and control points in the alarm control point ledger, and store them in the concept sets {a,b} and {c,d} respectively, and store the consequence lines and consequence levels in the HAZOP reports in the concept set {e,f;

[0083] Step 2: Select one of the concept pairs {c i ,d j};

[0084] Step 3: Replace {c i ,d j} and {a,b} are used to calculate the text similarity using the ESIM-Bert method;

[0085] Step 4: Get the point with the highest matching result {a i ,b j};

[0086] Step 5: Select {a i ,b j} in the line {e i ,f j} Fill in the corresponding row of the alarm control point ledger;

[0087] Step 6: Repeat steps 3 to 5 until the consequences of all control points are filled in;

[0088] Output: consequence description and consequence level of all control points in the alarm control point ledger.

[0089] S2. Determine the response time by dynamic simulation software, and obtain the urgency of the demand based on the response time; specifically, a response time determination module based on the HDCT method

[0090] In the process of an event possibly evolving into an accident, when the process deviation reaches the alarm set point, human intervention is required to prevent the terrible consequences. Assuming that the operator does not take any action and allows the consequences to continue to spread until they reach the critical point, this time period is the response time. Figure 5 As shown. In order to quantitatively estimate the response time, this embodiment uses Aspen Plus + Dynamics as a tool for dynamic simulation of the process simulation method. Compared with steady-state simulation, dynamic simulation can better reflect the changes in chemical processes under abnormal scenarios, and its results are more credible. This section proposes a HDCT (HAZOP-Dynamics-consequence time) method, which estimates the consequence response time by dynamically simulating the deviation propagation process of HAZOP. Minor consequences have been screened in the consequence level, and combined with actual investigations of several companies, it was found that the response time of minor consequences is usually longer. Therefore, this method only simulates the time for the deviation propagation with more serious consequence levels to reach the proposed consequences. This method only targets abnormal scenarios that can be represented by abnormal model structures and parameters, and does not consider non-technical scenarios involving operational adjustments, external interference, mechanical damage, etc.

[0091] The HDCT method is based on the HAZOP analysis results and the piping and instrumentation diagram (PID diagram). It uses process simulation to convert steady-state analysis into dynamic analysis. Its main processes are as follows: Figure 6 As shown, it includes: (1) preparing accurate and effective information such as process flow diagram (PFD), process instrument diagram (PID), accident manual, etc.; (2) dividing the node units, using Aspen Plus to perform steady-state simulation under normal operating conditions based on data such as mass and energy balance, and providing all parameters for dynamic simulation on this basis and verifying it; (3) converting the steady-state model into a dynamic model using Aspen Dynamics and verifying it; (4) setting some parameters in the dynamic model to deviate from the normal state, simulating the propagation of the deviation to the critical deviation to obtain the response time; (5) iterating each deviation that needs to be simulated; (6) recording the results in a table, obtaining the level of response time and organizing and improving it.

[0092] API-RP1167 provides three recommended levels for the factors affecting the timeliness of operator actions on alarm priority:

[0093] (1) "Less than 5 minutes" or "Immediately". Abandon all actions except the response to this alarm. A value less than 5 (usually 3) can be used.

[0094] (2) "5 to 15 minutes" or "Quickly". Complete any ongoing tasks quickly, but no new tasks will start until this alarm is processed. You can use a value greater than 15 (usually 10).

[0095] (3) "15 to 30 minutes" or "as soon as possible". Some short-term tasks can be completed before handling this alarm. You can use a value higher than 30 (sometimes 60).

[0096] In this method, emergency response time is defined as the degree of urgency, that is, the degree of urgency of taking measures for the deviation to spread to the critical point, which is divided into only two levels:

[0097] (1) Urgent: Once the safety limit is exceeded, adverse consequences will immediately occur and a quick response is required.

[0098] (2) Non-urgent: There is still some time before the safety limit is exceeded and adverse consequences occur, or there are other independent protection layers to prevent the consequences from occurring, so the response action requirement is not too urgent.

[0099] Table 2

[0100]

[0101] S3. Based on the severity of the consequences and the urgency of the need, the alarm priority of the petrochemical enterprise is determined by a predefined matrix method.

[0102] Specifically, after running the consequence module and response time module, they need to be filled in the alarm priority confirmation matrix. The specific steps are as follows:

[0103] (1) Integrate the results obtained by the consequence extraction module and the response time module.

[0104] (2) Assign priorities to alarms based on the severity of the consequences and simulated response time.

[0105] (3) Considering additional conditions, the alarm priority can be modified. This step mainly searches and matches the Safety Instrumentation System (SIS) ledger and the Layer of Protection Analysis (LOPA) results. The main principles are as follows:

[0106] ① For interlock protection of parameter settings, the critical level can be reduced to important level;

[0107] ② If the fluctuation of different parameters may cause the same consequence, if the main influencing parameter has been set to alarm, the other parameters can be set to normal or not set;

[0108] ③ When the alarm is used as an independent protection layer, the alarm level is directly judged as important.

[0109] (4) Use the priority matrix to determine the alarm priority. This matrix uses the consequence level classification in Table 1 and combines it with the urgency level in Table 2 to define four alarm priorities: critical alarm, important alarm, general alarm, and no alarm needs to be set, as shown in Table 3.

[0110] Table 3

[0111]

[0112] Verification and demonstration of this example

[0113] As one of the most important processes in an oil refinery, the catalytic cracking unit has the characteristics of complex processes, a large number of variables, and high risks, and has received continuous attention from researchers. This section selects the absorption stabilization system of the catalytic cracking unit of a petrochemical enterprise for example application demonstration, obtains real documents and data from the factory for modeling and verification, and uses the method proposed above in this embodiment to carry out detailed research steps. In order to ensure the scientific nature of the example, all parameters (such as equipment size, operating data, etc.) and documents (operation manual, PFD, PID, DCS screenshots, etc.) involved in this section are provided by real petrochemical enterprises.

[0114] 1. Introduction to the example process

[0115] The absorption and stabilization system in the catalytic cracking unit is used to further process the intermediate products from the fractionation system. The process includes the use of an absorption tower (C301), a desorption tower (C302), a reabsorption tower (C303), a stabilization tower (C304) and corresponding cold exchange equipment to produce products such as liquefied gas and stabilized gasoline. The fractionation system sends the compressed rich gas from the catalytic cracking unit to the absorption tower after gas-liquid separation, where the rich gas is absorbed under the action of the absorbent, and the lean gas further enters the reabsorption tower for further absorption. In order to control the heat release generated during the absorption process and prevent the operating temperature from being too high, the absorption tower is equipped with four mid-section refluxes to ensure the absorption effect. In the absorption tower, the rich gas is sent from the top of the tower to the bottom of the reabsorption tower after being absorbed. After the gasoline component is recovered through light diesel in the absorption tower, it becomes dry gas. This dry gas is divided into two paths: one path enters the riser reactor and is used as pre-lift dry gas; the other path enters the product refining part for desulfurization and is used as factory fuel gas.

[0116] 2. Instance consequence extraction

[0117] (1) Experimental data and parameter settings

[0118] The experimental data set comes from the HAZOP analysis report (441 items in total) and the alarm control point ledger (482 items) of a catalytic cracking unit. After the original text is processed by the data preprocessing method, the LNBIC method is used for entity recognition to extract the key information in the text. The focus is on the deviations, specific deviations, consequences, and consequence levels in the HAZOP analysis report, and the control points and deviations in the alarm control point ledger. In entity recognition, the equipment (such as distillation tower, separator) and materials (such as crude oil, recycled oil, etc.) in the original text are regarded as named entities and marked with abbreviations, as shown in Table 4. In this model, the BMES data annotation method is used. Specifically, it includes the following labels: "B" represents the first character of the entity, indicating the beginning of the entity; "M" represents the middle character of the entity, indicating the middle character of the entity; "E" represents the last character of the entity, indicating the end of the entity; "S" represents an entity composed of a single character, that is, it constitutes an entity alone; "O" represents other non-entity characters, indicating characters that do not belong to any entity. The annotation examples are shown in Table 5. These labeled data sets are divided into training set, test set and validation set in the ratio of 8:1:1. The experimental environment of the model is shown in Table 6, and the parameter settings are shown in Table 7.

[0119] Table 4

[0120]

[0121]

[0122] Table 5

[0123]

[0124] Table 6

[0125]

[0126] Table 7

[0127]

[0128]

[0129] (2) Evaluation indicators and comparative experiments

[0130] aEvaluation index

[0131] The selection of evaluation indicators is very important for the evaluation of the model. Since this model is mainly aimed at entity recognition tasks, its main goal is to evaluate the correct recognition of entity boundaries and the correctness of entity type labeling. Therefore, during the experiment, precision (P), recall (R) and F1 value (F1) are still used to evaluate the model.

[0132] b. Comparative experiment

[0133] In order to verify the advancedness of the proposed method, this example compares the entity recognition performance of the LNBIC model in Chinese HAZOP reports and alarm records with other mainstream models. The model is introduced as follows:

[0134] Bi-LSTM-CRF model: It uses a word vector with a dimension of 300 as input, then extracts text features through Bi-LSTM, and finally inputs these features into the CRF layer for sequence labeling.

[0135] Bi-LSTM-Attention-CRF model: uses word vectors as input, but in the feature extraction stage, not only Bi-LSTM is used, but also the self-attention mechanism is introduced, which helps to enhance the dependencies between entities. Finally, like the Bi-LSTM-CRF model, it also inputs the feature sequence into the CRF layer for entity sequence labeling.

[0136] BERT-Bi-LSTM-CRF model: word vectors are used as Bert outputs. After Bert encoding, Bi-LSTM is used for feature extraction and the attention mechanism is combined to strengthen the dependencies between entities. Finally, CRF is used to annotate the entity sequence.

[0137] The comparison of experimental results of mainstream models and the proposed method is shown in Table 8.

[0138] Table 8

[0139]

[0140] According to the data in Table 8, it can be clearly seen that the LNBIC model performs significantly better than other models in evaluation indicators such as P, R and F1. For example, the F1 score of the LNBIC model is 97.8%, which is 19.4% higher than the basic model Bi-LSTM-CRF. This shows that the proposed LNBIC model performs significantly better in entity recognition tasks and has significant superiority.

[0141] c Example training results

[0142] In the case analysis, this section takes the HAZOP analysis report of a catalytic cracking unit of a petrochemical enterprise as an example and lists the entities that need to be extracted and identified. The original HAZOP report is relatively rich in content. In order to show it more clearly, some columns are excerpted in Table 9. In Table 10, only the columns required by the consequence extraction module are shown to show the results of entity recognition. For example, in the specific deviation, the report pointed out that "the pressure of the stabilization tower C304 is high". After entity recognition through LNBIC, the following entities were successfully identified: "EQU: stabilization tower; EQUN: C304; PAR: pressure; SAT: high".

[0143] Table 9

[0144]

[0145]

[0146] Table 10

[0147]

[0148] dSimilarity matching

[0149] Table 11 shows the experimental results of the ESIM-BERT method proposed in this embodiment and other mainstream text matching methods. It can be seen that in terms of the F1 score that comprehensively considers precision and recall, the ESIM-BERT method reaches 0.938, which is significantly better than other methods, which proves the effectiveness and performance advantages of this method in text matching tasks.

[0150] Table 11

[0151]

[0152] In order to study the adaptability of text matching in HAZOP reports and alarm point records, 12 corresponding text matches were randomly selected as examples, and their confusion matrices were visualized, as shown in Figure 2. Figure 7 shown.

[0153] Finally, the results of similarity matching are filled into the consequence level column in the alarm classification table, as shown in Table 12.

[0154] Table 12

[0155]

[0156] 3. Determine the instance response time

[0157] This section selects the absorption stabilization system of the catalytic cracking unit to demonstrate the application of the urgent demand analysis example.

[0158] (1) Aspen steady-state simulation

[0159] First, the steady-state simulation of the absorption stabilization system of the catalytic cracking unit was constructed using Aspen Plus software. During the simulation, the virtual component method was used to describe the complex petroleum fractions and the RadFrac model was selected as the tower model. Figure 8 shown.

[0160] The RK-SOAVE method was used in the simulation of the absorption stabilization system. This method is suitable for the physical properties of non-polar or weakly polar mixtures. In order to control the temperature during the absorption process to ensure the absorption effect, four mid-section refluxes are set in the absorption tower. These mid-section refluxes come from their respective oil collection tanks and return to the corresponding tray layers of the absorption tower. During the simulation, the four mid-section refluxes were set separately. In addition, the tray efficiency was also set accordingly, as shown in Table 13.

[0161] Table 13

[0162]

[0163]

[0164] The simulation results of product distribution in the simulation results of the absorption stabilization system are shown in Table 14. It can be seen that the overall yield simulation value of the simulation results is basically consistent with the calibration value, among which the dry gas yield is higher and the error is more obvious. The analysis of the Engler distillation data and liquefied gas composition of the stabilized gasoline of the simulation results are shown in Tables 14 and 15 respectively.

[0165] Table 14

[0166]

[0167] Table 15

[0168]

[0169]

[0170] According to the data in the table, the Engler distillation data of the stabilized gasoline is very close to the calibration data, with only a slight deviation in the initial distillation point temperature. In addition, according to the table data, the liquefied gas composition in the simulation results is very similar to the calibration results, with a small error. Based on the comparison of the above simulation results, it can be concluded that the steady-state simulation results of the absorption stabilization system are reliable and can meet the needs of subsequent research.

[0171] (2) Aspen dynamic simulation

[0172] The focus of steady-state simulation is to solve the balance convergence in process design and obtain the steady-state value of process parameters, while dynamic simulation is closer to the characteristics of parameters changing over time in actual working conditions. Dynamic simulation introduces time variables based on the three major balances, providing a more accurate reference for the process. On the basis of steady-state simulation, necessary container size calculations and pressure calibration were carried out, and dynamic simulation was performed using the Aspen Dynamics module to obtain the time required for the operator to reach the extreme value without taking action after the fault occurs, so as to judge the urgency of the demand.

[0173] a Stabilizing tower overpressure disturbance

[0174] In the HAZOP report, the consequence of overpressure in the stabilizing tower is C, which is a serious level, so a dynamic simulation is performed. The E1310 cooling effect, one of the causes of overpressure disturbance in the stabilizing tower (C304), is input as a disturbance, and the response is recorded from the beginning of the change in the variable. Fig. 9 According to the dynamic simulation results, the pressure of the stabilization tower C304 increased rapidly within a few seconds without any operator action. In a short period of time, the liquid level at the bottom of the stabilization tower and the liquid level of the analytical tower C302 also showed a downward trend (due to the increase in pressure, the process flow rate increased, thus showing a downward trend of oscillation). The time was within 3 minutes, so the urgency of the specific deviation was determined to be urgent.

[0175] b. Reabsorption tower liquid level reduction disturbance

[0176] In the HAZOP report, the decrease in the liquid level of the reabsorption tower caused the stable tower bottom liquid level to drop. The consequence of the liquefied gas entering the product refining unit was more serious, so a dynamic simulation was performed. The LICA1311 control fault, which was one of the reasons for the decrease in the liquid level of the reabsorption tower (C303), was input as a disturbance, and the reaction was recorded from the beginning of the change in the variable. Fig.10 The dynamic simulation results show that without any operator action, the change in the liquid level of the reabsorption tower (C303) shows a linear characteristic and takes more than 3 minutes to reach the extreme value. Fig.10 As shown in the figure, the time exceeds 3 minutes, so the urgency of this specific deviation is determined to be non-urgent.

[0177] 4. Determination of instance alarm priority

[0178] Fill the obtained results into the alarm priority record form example, focusing on the consequence level and urgency column, as shown in Table 16. Among them, the control point and instrument number are obtained from the alarm ledger; the deviation, consequence description and consequence level are obtained from the HAZOP report; the urgency is determined according to the rules; the alarm level is determined by the consequence level, urgency and constraints.

[0179] Table 16

[0180]

[0181]

[0182] The comparison of the optimization results of the alarm priority of the catalytic cracking unit is shown in Table 17. It can be seen from the table that the original priority setting of some alarm parameters of the unit may not meet the standard recommended distribution ratio because it did not consider the severity of the consequences after the deviation of the parameters, the effectiveness of the protection measures and the timeliness of the personnel response. After using the alarm priority determination method proposed in this embodiment and adjusting and optimizing it, the alarm priority division of the unit meets the recommended ratio of EEMUA, and the proposed method has good scientificity and effectiveness.

[0183] Table 17

[0184]

[0185] In summary, this embodiment provides a comprehensive and objective method for determining alarm priorities of petrochemical enterprises, which mainly includes:

[0186] (1) The LNBIC model was developed for entity recognition from HAZOP reports and alarm point records. The F1 score of the model was 0.978 in the experimental results, which showed a high accuracy. The consequence level was filled into the corresponding column of the alarm point through similarity matching using the ESIM-Bert method (the experimental result F1 score was 0.938).

[0187] (2) The HDCT method is used to determine the response time. The response time of the deviation with serious consequences in the HAZOP is dynamically simulated using ASPEN and input into the proposed urgency matrix to determine the level of urgency.

[0188] (3) Alarm priorities are assigned according to the consequence level and the urgency level, and some restrictions on alarm priorities are proposed, such as the impact of SIS and LOPA.

[0189] (4) The method was applied in a catalytic cracking unit. The results showed that the priority ratio determined by the alarm priority determination method was more in line with the EEMUA recommended ratio than before, and the method was scientific.

[0190] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for determining alarm priority in a petrochemical enterprise, characterized in that: The following steps are involved: Use natural language processing technology to extract information from HAZOP reports of petrochemical enterprises and obtain the severity of the consequences of potential risk points; Determine the response time by dynamic simulation software, and obtain the urgency of the demand based on the response time; Based on the severity of the consequences and the urgency of the need, the alarm priority of the petrochemical enterprise is determined by a predefined matrix method.

2. The method for determining alarm priority of a petrochemical enterprise according to claim 1, characterized in that: The process of extracting information from the HAZOP report of a petrochemical enterprise includes: Based on the Chinese entity recognition model of the natural language processing technology, text recognition and entity extraction are performed on the text of the HAZOP report to obtain entity information; The text similarity matching based on the natural language processing technology is used to match the deviation and consequence information in the HAZOP report and the alarm control point ledger, ensuring that the consequence level in the HAZOP report is filled in the consequence level of the alarm control point to obtain the consequence severity of the potential risk point.

3. The method for determining the alarm priority of a petrochemical enterprise according to claim 2, characterized in that: The process of performing text recognition and entity extraction on the text of the HAZOP report includes: Adopting the LE-NEZHA layer in the Chinese entity recognition model, the word vectors in the entity dictionary are integrated into the neural context representation model for Chinese understanding through a dictionary adapter; Extract long-range semantic information from text using neural context representation models; Adopt iterative dilated convolutional neural network to learn more long-range semantic features; The CRF layer is introduced to constrain the legitimacy of the label output, and the entity recognition result is obtained based on the output label probability.

4. The method for determining alarm priority of a petrochemical enterprise according to claim 2, characterized in that: The process of matching the deviation and consequence information in the HAZOP report and the alarm control point ledger based on the text similarity matching of the natural language processing technology includes: Performing text recognition on the deviation line and specific deviation line in the HAZOP report and the deviation line and control point in the alarm control point ledger to obtain a first concept set and a second concept set; Select an alarm control point from the first concept set, calculate the text similarity between the alarm control point and the second concept set using the ESIM-Bert method, and obtain the highest matching result point; The consequence row and consequence level in the highest point of the matching result are selected and filled into the corresponding row of the alarm control point ledger to obtain the consequence level of all control points.

5. The method for determining alarm priority of a petrochemical enterprise according to claim 1, characterized in that: The process of determining response time through dynamic simulation software includes: Aspen Plus software was used to perform steady-state simulation on the HAZOP report information to obtain a steady-state model; The steady-state model is converted into a dynamic model and verified using Aspen Dynamics software; Some parameters in the dynamic model are set to deviate from the normal state, and the propagation of the deviation to the critical deviation is simulated to obtain the response time.

6. The method for determining alarm priority of a petrochemical enterprise according to claim 1, characterized in that: The predefined matrix method includes: a first level classification of consequence severity and a second level classification of urgency, and a plurality of alarm priorities are determined based on the first level classification and the second level classification.

7. The method for determining alarm priority of a petrochemical enterprise according to claim 6, characterized in that: When determining alarm priorities, consider the impact of additional conditions such as safety instrumented systems and protection layer analysis.

Images