Multi-agent response auxiliary decision-making method for process alarm under abnormal working condition

Abstract

The invention discloses a multi-agent response aided decision-making method for process alarm under an abnormal working condition, and relates to the technical field of industrial automation and artificial intelligence, and the method comprises the steps: collecting historical abnormal process data, an abnormal process type label and an actual processing decision-making scheme in a chemical production process, and based on the historical abnormal process data, the abnormal process type label and the actual processing decision-making scheme, obtaining an abnormal process type label; the method comprises the following steps: pre-training a first agent and a second agent by using a multi-strategy information acquisition optimization algorithm, then identifying real-time process data and determining process alarm information through the first agent, extracting real-time features by using the second agent and matching the real-time features with historical sample features when an alarm occurs, finding similar historical samples, and finally obtaining process alarm information; through cooperative work of the double intelligent agents and combination of a specific multi-strategy information acquisition optimization algorithm, the accuracy of anomaly recognition and the reliability of decision recommendation are improved, and the problems that response lags behind under the abnormal working condition of the chemical process, and decision depends on artificial experience are effectively solved.

Description

Technical Field

[0001] This application relates to the fields of industrial automation and artificial intelligence technology, and more specifically, to a multi-agent response auxiliary decision-making method for process alarms under abnormal operating conditions. Background Technology

[0002] In modern chemical production, processes are complex, involving numerous reaction units, heat exchangers, separation towers, and control loops. Due to raw material fluctuations, equipment aging, operational errors, or external environmental interference, abnormal states deviating from normal operating conditions frequently occur. If these abnormal conditions are not detected and addressed promptly, they can escalate into major safety accidents, causing significant economic losses and casualties. Traditional chemical process alarm management primarily relies on setting fixed thresholds (such as temperature and pressure limits), triggering an alarm once these limits are exceeded. However, this approach often suffers from delays and alarm proliferation, making it difficult for operators to promptly identify critical alarms. During the anomaly handling phase, existing solutions heavily depend on the operator's personal experience, lacking standardized, data-driven intelligent decision-making support. Because different operators possess varying skill levels, this can easily lead to misoperation or improper handling.

[0003] In recent years, fault diagnosis and decision support systems based on machine learning have gradually emerged. However, existing technologies often suffer from the following problems: 1) A single model struggles to simultaneously achieve high-precision anomaly identification and deep feature extraction; 2) Hyperparameter optimization of agent models often relies on human experience or traditional grid search, which is inefficient and makes it difficult to find the global optimum, resulting in poor generalization ability of the model on complex chemical engineering data; 3) When recommending decision-making schemes, there is a lack of accurate matching and clustering analysis of historically effective schemes. Summary of the Invention

[0004] This application aims to provide a multi-agent response auxiliary decision-making method for process alarms under abnormal operating conditions, in order to solve the technical problems existing in the prior art.

[0005] This application provides a multi-agent response auxiliary decision-making method for process alarms under abnormal operating conditions, including: Collect historical abnormal process data, abnormal process type labels, and actual abnormal handling decision-making schemes under various abnormal operating conditions; the abnormal process type labels include specific abnormal process types or normal types. Based on the historical abnormal process data and abnormal process type labels, a first intelligent agent is pre-trained using a multi-strategy information acquisition optimization algorithm to obtain the trained first intelligent agent. Based on historical abnormal process data, abnormal process type labels, and actual abnormal handling decision schemes, a second intelligent agent is pre-trained using a multi-strategy information acquisition optimization algorithm to obtain the trained second intelligent agent. The trained second intelligent agent is then used to extract the sample feature data corresponding to the historical abnormal process data and abnormal process type labels. Real-time process data is collected during chemical production, and the trained first intelligent agent is invoked to identify the real-time process data and determine process alarm information. When the process alarm information is a specific abnormal process type, a trained second agent is used to extract features from the real-time process data and the process alarm information to obtain real-time feature data. The real-time feature data is matched with the sample feature data to match similar historical abnormal process data and similar abnormal process anomaly type labels to obtain similar historical sample data. Based on the actual anomaly handling decision schemes corresponding to the similar historical sample data, a target auxiliary decision scheme for process alarms under abnormal operating conditions is generated.

[0006] Furthermore, based on the aforementioned historical abnormal process data and abnormal process type labels, a first intelligent agent is pre-trained using a multi-strategy information acquisition optimization algorithm to obtain the trained first intelligent agent, including: The hyperparameters of the first agent are randomly generated and encoded to obtain hyperparameter codes. Multiple different hyperparameter codes are obtained repeatedly. For any hyperparameter code, the fitness corresponding to the hyperparameter code is obtained based on the historical abnormal process data and the abnormal process type label. The hyperparameter encoding with the highest fitness is determined as the first optimal encoding; A multi-point guidance information acquisition mechanism is used to acquire multi-point guidance information for the hyperparameter encoding, resulting in the hyperparameter encoding after multi-point guidance information acquisition. Based on the first optimal encoding, an uncertainty-driven information acquisition mechanism is used to perform uncertainty-driven information acquisition on the hyperparameter encoding after multi-point guidance information acquisition, so as to obtain the hyperparameter encoding of uncertainty-driven information acquisition. A diversity information enhancement acquisition mechanism is adopted to perform diversity information enhancement acquisition on the hyperparameter encoding of uncertainty-driven information acquisition, and the hyperparameter encoding after diversity information enhancement acquisition is obtained. Determine whether the total number of optimization attempts is greater than or equal to the preset maximum number of optimization attempts. If so, enhance the hyperparameter encoding after collection based on diversity information and determine the second optimal encoding. Otherwise, return to the fitness acquisition step. Based on the second optimal encoding, obtain the first agent after training.

[0007] Furthermore, a multi-point guidance information acquisition mechanism is used to acquire multi-point guidance information for the hyperparameter encoding, resulting in hyperparameter encoding after multi-point guidance information acquisition, including: Based on the total number of optimization iterations, the multi-point guidance speed control factor is obtained as follows: ; In the formula, For multi-point guidance speed control factor, It is an exponential function with the natural constant e as its base. It is a constant term, and is set to 5; This is the maximum value of the multi-point guidance speed control factor, and it is set to 0.95; This is the minimum value of the multi-point guidance speed control factor, and it is set to 0.0005. The total number of optimizations is T, where T is the preset maximum number of optimizations. The guidance speed is controlled using the multi-point guidance speed control factor. All hyperparameter codes are used as guidance parameters, and control is performed by combining the boundary information of the hyperparameter codes. The hyperparameter codes obtained after the multi-point guidance information acquisition are as follows: ; ; In the formula, For the first t During the second optimization process, the first k The first hyperparameter encoding d dimensional hyperparameters, for Corresponding multi-point guidance information, For the first k The first hyperparameter encoding after multi-point guidance information acquisition d dimensional hyperparameters, k =1,2,…,N, where N is the total number of hyperparameter codes. d =1,2,…,Din, where Din is the total dimension of hyperparameters in hyperparameter encoding. For multi-point guidance speed control factor, For the first d Upper limit of hyperparameters, For the first d Lower bound of hyperparameters This is the intensity coefficient of the group information parameter, and it is set to 0.5; It is a natural constant. For the first t During the second optimization process, the first j The first hyperparameter encoding d dimensional hyperparameters, For the first k The hyperparameter encoding and the first j The Euclidean distance between the hyperparameter codes, where c represents the population information parameter direction coefficient and is set to 1.5. Represents the first optimal code's... d Dimensional hyperparameters.

[0008] Furthermore, based on the first optimal encoding, an uncertainty-driven information acquisition mechanism is used to perform uncertainty-driven information acquisition on the hyperparameter encoding after multi-point guidance information acquisition, resulting in the hyperparameter encoding for uncertainty-driven information acquisition, including: Based on the fitness of the hyperparameter encoding after the acquisition of the multi-point guidance information, the uncertainty factor is obtained as follows: ; ; ; In the formula, As an uncertainty factor, It is a logarithmic function. The total number of hyperparameter codes. To optimize the state factor, For the first i The fitness percentage of hyperparameter encoding after multi-point guidance information collection. For the first i The fitness of hyperparameter coding after multi-point guidance information collection; Based on the total number of optimization attempts, the information collection control factor is as follows:

[0009] in, To collect control factors, This is the minimum value of the information collection control factor, and it is set to 0.05. This is the dynamic balance factor, and it is set to 0.5; The minimum value among the fitness values ​​of the hyperparameter encoding after multi-point guidance information acquisition. The maximum value in the fitness of the hyperparameter encoding after multi-point guidance information acquisition; Based on the first optimal encoding, the uncertainty factor, and the information acquisition control factor, the hyperparameter encoding for uncertainty-driven information acquisition is obtained as follows:

[0010] In the formula, For the first i Hyperparameter encoding after multi-point guidance information collection For the first i Hyperparameter encoding for information acquisition driven by uncertainty The first optimal code, This is the first learning coefficient, and it is set to 1 or 2; This is the second learning coefficient, and it is set to 0.75 or 1.5; The first random number between (0,1) The second random number between (0,1) For fitness greater than The random encoding, and in When fitness is at its maximum, Set a random code.

[0011] Furthermore, a diversity information enhancement acquisition mechanism is employed to perform diversity information enhancement acquisition on the hyperparameter encoding of uncertainty-driven information acquisition, resulting in hyperparameter encoding after diversity information enhancement acquisition, including: The diversity information factor for obtaining hyperparameter encoding of uncertainty-driven information acquisition is: ; ; In the formula, For the first t During the second optimization process, the first m The first hyperparameter encoding of the uncertainty-driven information acquisition d dimensional hyperparameters, m =1,2,…,N, where N is the total number of hyperparameter codes. The first hyperparameter encoding corresponding to the average hyperparameter encoding of all uncertainty-driven information acquisition. d dimensional hyperparameters, The first optimal code d dimensional hyperparameters, For the first m The diversity coding corresponding to the hyperparameter coding of uncertainty-driven information acquisition. d dimensional hyperparameters, for Corresponding diversity information factors; Based on the aforementioned diversity information factor, the hyperparameter encoding after enhanced diversity information acquisition is obtained as follows: ; In the formula, For the first m The first hyperparameter encoding after the acquisition of enhanced diversity information. d dimensional hyperparameters, The search coefficients are randomly distributed, and , The third random number between (0,1) The fourth random number between (0,1) Pi is the mathematical constant of a circle.

[0012] Furthermore, based on historical abnormal process data, abnormal process type labels, and actual abnormal handling decision schemes, a second intelligent agent is pre-trained using a multi-strategy information acquisition optimization algorithm to obtain the trained second intelligent agent. The trained second intelligent agent is then used to extract sample feature data corresponding to the historical abnormal process data and abnormal process type labels, including: K-means clustering was performed on the historical abnormal process data and the historical sample data composed of abnormal process type labels to obtain K cluster centers. Treat the actual anomaly handling decision scheme corresponding to any cluster center as the same scheme, and assign a unique numerical label to each cluster center. Based on the historical sample data in the cluster centers and the numerical labels corresponding to the cluster centers, a second intelligent agent is pre-trained using a multi-strategy information acquisition optimization algorithm to obtain the trained second intelligent agent. The historical sample data is used as input to the trained second agent, and the output of the feature layer of the second agent is obtained to obtain the sample feature data.

[0013] Furthermore, real-time process data is collected during the chemical production process, and the trained first intelligent agent is invoked to identify the real-time process data and determine process alarm information, including: Real-time process data is collected during chemical production, and the real-time process data is used as the input of the first intelligent agent after training to obtain the probability distribution of the output of the first intelligent agent after training. Based on the probability distribution output by the first intelligent agent after training, the abnormal process type with the highest probability is determined as the target abnormal process type. If the target abnormal process type is any specific abnormal process type, then the process alarm information is determined to be triggered; otherwise, the process alarm information is determined not to be triggered.

[0014] Furthermore, when the process alarm information is a specific abnormal process type, a trained second agent is used to extract features from the real-time process data and the process alarm information to obtain real-time feature data, including: When the process alarm information is a specific abnormal process type, the real-time process data and the process alarm information are combined into a vector to obtain real-time sample data; The real-time sample data is used as input to the trained second agent to obtain the real-time feature data corresponding to the real-time sample data output by the trained second agent.

[0015] Further, the real-time feature data is matched with the sample feature data to match similar historical abnormal process data and similar abnormal process anomaly type labels, resulting in similar historical sample data, including: Obtain the cosine similarity between the real-time feature data and the sample feature data, and take the sample feature data with the largest cosine similarity as the target sample feature data; The similar historical abnormal process data corresponding to the target similar sample feature data and the similar abnormal process abnormality type label are used together as similar historical sample data.

[0016] Furthermore, based on the actual anomaly handling decision schemes corresponding to the similar historical sample data, a target auxiliary decision scheme for process alarms under abnormal operating conditions is generated, including: The cluster centers corresponding to the similar historical sample data are determined as the target cluster centers; The actual anomaly handling decision scheme corresponding to the similar historical sample data is determined as the first candidate auxiliary decision scheme. The M actual anomaly handling decision schemes that appear most frequently in the target cluster centers are identified as the second candidate auxiliary decision schemes; M represents the pre-set number of normal numbers. The first candidate auxiliary decision scheme and the second candidate auxiliary decision scheme are deduplicated to obtain at least one target auxiliary decision scheme for process alarm under abnormal operating conditions.

[0017] Beneficial effects: This application provides a multi-agent response-assisted decision-making method for process alarms under abnormal operating conditions. First, it collects historical abnormal process data, abnormal process type labels, and actual handling decision schemes from the chemical production process. Based on this, a first agent and a second agent are pre-trained using a multi-strategy information acquisition optimization algorithm. Then, the first agent identifies real-time process data and determines process alarm information. When an alarm occurs, the second agent extracts real-time features and matches them with historical sample features to find similar historical samples, thereby generating a target-assisted decision-making scheme. Through the collaborative work of the two agents, combined with a specific multi-strategy information acquisition optimization algorithm, the accuracy of anomaly identification and the reliability of decision recommendations are improved, effectively solving the problems of delayed response and reliance on human experience in abnormal chemical process conditions. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of a multi-agent response auxiliary decision-making method for process alarms under abnormal operating conditions, proposed in an embodiment of this application. Figure 2 This is a flowchart of obtaining the first intelligent agent after training, as proposed in one embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] To facilitate a better understanding of the technical solutions described in the embodiments of this application by those skilled in the art, illustrative examples of the application scenarios and key data of the technical solutions described in the embodiments of this application are provided. However, it is worth noting that these are merely examples and can also be applied to other chemical and manufacturing processes.

[0022] Historical abnormal process data: Historical time-series data segments of the above process parameters stored in the production process over the past 3 years, with each data segment covering data from 10 minutes before the occurrence of the abnormality to 30 minutes after the occurrence.

[0023] Real-time process data (input to the first intelligent agent): including reactor temperature (T_reactor, unit °C), reactor pressure (P_reactor, unit MPa), stirring motor current (I_agitator, unit A), catalyst feed flow rate (F_cat, unit kg / h), propylene monomer feed flow rate (F_mono, unit kg / h), and jacket cooling water outlet temperature (T_cw_out, unit °C). The data acquisition frequency is once per minute.

[0024] Abnormal process type tags: These correspond to the manually labeled tags for the aforementioned historical data segments. Specific categories include: Type 0: Normal state; Type 1: Overheating of the reaction (precursor to explosive polymerization); Type 2: Catalyst poisoning reduces activity; Type 3: Agitator malfunction / scaling; Type 4: Clogged feed line; Actual anomaly handling decision-making scheme: The specific operational measures taken by operators in response to the above-mentioned anomalies in the historical records are formatted as text or instruction codes.

[0025] Process alarm information: The judgment result output by the first intelligent agent, such as "Type 1: Reaction over-temperature".

[0026] Target-assisted decision-making scheme: Emergency response suggestions recommended to operators, such as "Emergency opening of the cooling water bypass valve, setting the opening degree to 50%; suspension of catalyst injection pump".

[0027] like Figure 1 As shown in the figure, this application provides a multi-agent response auxiliary decision-making method for process alarms under abnormal operating conditions, including: S101. Collect historical abnormal process data, abnormal process type labels, and actual abnormal handling decision schemes under various abnormal operating conditions; the abnormal process type labels include specific abnormal process types or normal types.

[0028] For example, historical abnormal process data, abnormal process type labels, and actual abnormal handling decision-making schemes can be collected. In this embodiment, 1000 sets of historical samples are exported from the DCS (Distributed Control System) database. Among them, there are 200 sets of reaction overheating (Type 1) samples, 150 sets of catalyst poisoning (Type 2) samples, 100 sets of stirring failure (Type 3) samples, 100 sets of feed blockage (Type 4) samples, and 450 sets of normal state (Type 0) samples. The collected data is normalized to map data of different dimensions to the [0,1] interval to eliminate the influence of units.

[0029] S102. Based on the historical abnormal process data and abnormal process type labels, a first intelligent agent is pre-trained using a multi-strategy information acquisition optimization algorithm to obtain the trained first intelligent agent. A multi-strategy information acquisition optimization algorithm is used to pre-train the first agent (such as a BP neural network, a long short-term memory neural network, or a support vector machine). The input of the first agent is the normalized process parameters (6-dimensional), and the output is the probability distribution of each anomaly type.

[0030] Existing technologies often suffer from problems such as getting stuck in local optima, slow training speed, and poor training results during the training of intelligent agents. Therefore, this application provides a multi-strategy information acquisition optimization algorithm to solve the technical problems existing in the prior art and improve the accuracy of process alarms.

[0031] S103. Based on historical abnormal process data, abnormal process type labels and actual abnormal handling decision schemes, a second intelligent agent is pre-trained using a multi-strategy information acquisition optimization algorithm to obtain the trained second intelligent agent, and the trained second intelligent agent is used to extract the sample feature data corresponding to the historical abnormal process data and abnormal process type labels. The second intelligent agent can be configured as a convolutional neural network. It can first perform K-means clustering (K=5) on samples composed of historical abnormal process data and labels. Samples with similar clustering results (meaning similar process states) are grouped into one class; for example, cluster center 1 corresponds to "slight over-temperature, normal pressure," and cluster center 2 corresponds to "severe over-temperature, increased pressure." A unique numerical label is assigned to each cluster center, and the "actual anomaly handling decision scheme" that appears most frequently in that cluster is taken as the representative scheme for that cluster.

[0032] After training, the output of the feature layer of the second agent (e.g., a fully connected layer; if other neural networks are used as the second agent, the output of a network layer with fixed dimensions can be extracted) is used as the sample feature data for that historical sample. This feature data is a low-dimensional dense expression of high-dimensional process parameters, containing the essential characteristics of the operating conditions.

[0033] S104. Collect real-time process data during chemical production, and call the trained first intelligent agent to identify the real-time process data and determine process alarm information. To ensure accurate data identification, the data structures of real-time process data and historical abnormal process data should be identical, and both need to be normalized. For example, real-time process data such as the reactor temperature and pressure can be acquired. This real-time process data is then input into a pre-trained agent, which outputs a probability distribution: [0.02, 0.88, 0.05, 0.03, 0.02]. The type with the highest probability is determined as Type 1 (reaction over-temperature), with a probability of 88%. Finally, if it is Type 0, no alarm is triggered; if it is Type 1-4, a process alarm is triggered, and the alarm information should include the specific abnormal process type.

[0034] It is worth noting that both real-time process data and historical abnormal process data can be set as time-series data. For example, when collecting temperature data, temperature data can be collected continuously for a period of time, thereby improving data accuracy.

[0035] S105. If the process alarm information is a specific abnormal process type, the trained second agent is used to extract features from the real-time process data and the process alarm information to obtain real-time feature data.

[0036] S106. Match the real-time feature data with the sample feature data, and match similar historical abnormal process data and similar abnormal process anomaly type labels to obtain similar historical sample data. For example, real-time process data and the identified Type 1 data can be combined into a vector, which is then input into a trained second agent to extract real-time feature data. The cosine similarity between the real-time feature data and the feature data of all historical samples in the database is calculated. The historical sample with the highest similarity is found; for example, the operating condition corresponding to this sample is "the overheating event at 14:00 on May 12, 2022".

[0037] S107. Based on the actual anomaly handling decision scheme corresponding to the similar historical sample data, generate a target auxiliary decision scheme for process alarm under abnormal operating conditions.

[0038] For example, the cluster center to which the similar sample belongs (e.g., cluster center A) can be determined, the actual treatment plan for the similar sample can be obtained as the first candidate plan, the top 3 plans that appear most frequently in cluster center A can be obtained as the second candidate plans, the candidate plans can be deduplicated and sorted, and the target auxiliary decision plan can be output: emergency injection of terminating agent (level: high); increase jacket cooling water flow rate to 120%; reduce propylene feed flow rate to 80%.

[0039] This application provides a multi-agent response-assisted decision-making method for process alarms under abnormal operating conditions. First, it collects historical abnormal process data, abnormal process type labels, and actual handling decision schemes from the chemical production process. Based on this, a first agent and a second agent are pre-trained using a multi-strategy information acquisition optimization algorithm. Then, the first agent identifies real-time process data and determines process alarm information. When an alarm occurs, the second agent extracts real-time features and matches them with historical sample features to find similar historical samples, thereby generating a target-assisted decision-making scheme. Through the collaborative work of the two agents, combined with a specific multi-strategy information acquisition optimization algorithm, the accuracy of anomaly identification and the reliability of decision recommendations are improved, effectively solving the problems of delayed response and reliance on human experience in abnormal chemical process conditions.

[0040] The first agent focuses on rapid classification, while the second agent focuses on deep feature extraction. When an anomaly occurs, there is no need to rerun complex optimization algorithms; features can be obtained simply through forward derivation, meeting the millisecond-level response speed requirements of chemical production.

[0041] This method directly uses existing process data as input and outputs specific control commands or operation suggestions, eliminating the need for operators to perform complex secondary analysis and making it extremely easy to integrate and deploy in existing industrial control systems.

[0042] like Figure 2As shown, based on the historical abnormal process data and abnormal process type labels, a first intelligent agent is pre-trained using a multi-strategy information acquisition optimization algorithm to obtain the trained first intelligent agent, which includes: S201. Randomly generate hyperparameters of the first agent and encode them to obtain hyperparameter codes. Repeatedly obtain multiple different hyperparameter codes. For example, hyperparameters can be randomly initialized between upper and lower bounds and encoded as vectors, thus obtaining a vector-based hyperparameter encoding that can be used for continuous optimization.

[0043] S202. For any hyperparameter code, based on the historical abnormal process data and the abnormal process type label, obtain the fitness corresponding to the hyperparameter code. For example, historical abnormal process data can be used as the actual input, and the abnormal process type label can be used as the expected output. The cross-entropy loss function value or root mean square loss function value can be obtained. Then, the cross-entropy loss function value or root mean square loss function value can be added to a constant term (such as 0.001), and the reciprocal can be taken to obtain the fitness corresponding to the hyperparameter encoding.

[0044] S203. The hyperparameter encoding with the highest fitness is determined as the first optimal encoding; S204. A multi-point guidance information acquisition mechanism is used to acquire multi-point guidance information for the hyperparameter encoding to obtain the hyperparameter encoding after multi-point guidance information acquisition. S205. Based on the first optimal encoding, an uncertainty-driven information acquisition mechanism is used to perform uncertainty-driven information acquisition on the hyperparameter encoding after multi-point guidance information acquisition, so as to obtain the hyperparameter encoding of uncertainty-driven information acquisition. S206. Employ a diversity information enhancement acquisition mechanism to perform diversity information enhancement acquisition on the hyperparameter encoding of uncertainty-driven information acquisition, and obtain the hyperparameter encoding after diversity information enhancement acquisition. S207. Determine whether the total number of optimization attempts is greater than or equal to the preset maximum number of optimization attempts. If so, determine the second optimal code based on the hyperparameter encoding after the diversity information is collected; otherwise, return to the fitness acquisition step. S208. Obtain the first intelligent agent after training according to the second optimal encoding.

[0045] For example, the hyperparameters in the second optimal encoding can be used as the final hyperparameters of the first agent to obtain the trained first agent. It's worth noting that after each optimization, the hyperparameter encoding can be subjected to limit-breaking processing to ensure that all optimizations are effective.

[0046] Compared with the prior art, the multi-strategy information acquisition and optimization algorithm provided in this application has the characteristics of fast convergence speed, strong global optimization capability and good optimization effect. It can effectively improve the robustness of the agent and enable the trained agent to effectively identify data, thereby realizing assisted response decision-making.

[0047] In this embodiment of the application, a multi-point guidance information acquisition mechanism is used to acquire multi-point guidance information for the hyperparameter encoding, resulting in hyperparameter encoding after multi-point guidance information acquisition, including: Based on the total number of optimization iterations, the multi-point guidance speed control factor is obtained as follows: ; In the formula, For multi-point guidance speed control factor, It is an exponential function with the natural constant e as its base. It is a constant term, and is set to 5; This is the maximum value of the multi-point guidance speed control factor, and it is set to 0.95; This is the minimum value of the multi-point guidance speed control factor, and it is set to 0.0005. The total number of optimizations is T, where T is the preset maximum number of optimizations. The guidance speed is controlled using the multi-point guidance speed control factor. All hyperparameter codes are used as guidance parameters, and control is performed by combining the boundary information of the hyperparameter codes. The hyperparameter codes obtained after the multi-point guidance information acquisition are as follows: ; ; In the formula, For the first t During the second optimization process, the first k The first hyperparameter encoding d dimensional hyperparameters, for Corresponding multi-point guidance information, For the first k The first hyperparameter encoding after multi-point guidance information acquisition d dimensional hyperparameters, k =1,2,…,N, where N is the total number of hyperparameter codes. d =1,2,…,Din, where Din is the total dimension of hyperparameters in hyperparameter encoding. For multi-point guidance speed control factor, For the first d Upper limit of hyperparameters, For the first d Lower bound of hyperparameters This is the intensity coefficient of the group information parameter, and it is set to 0.5; It is a natural constant. For the first t During the second optimization process, the first j The first hyperparameter encoding d dimensional hyperparameters, For the first k The hyperparameter encoding and the first j The Euclidean distance between the hyperparameter codes, where c represents the population information parameter direction coefficient and is set to 1.5. Represents the first optimal code's... d Dimensional hyperparameters.

[0048] In neural network hyperparameter optimization, the population tends to prematurely cluster around a suboptimal solution, leading to premature convergence. This application addresses this issue through a multi-point guided velocity control factor and a multi-point guided information acquisition mechanism, which enhances global search capability and maintains diversity in the early and mid-stages of the algorithm. Furthermore, in the later stages of the algorithm, as the encodings continuously cluster within the solution space, the local space can be searched more effectively, thereby improving convergence accuracy.

[0049] In this embodiment of the application, based on the first optimal encoding, an uncertainty-driven information acquisition mechanism is used to perform uncertainty-driven information acquisition on the hyperparameter encoding after multi-point guidance information acquisition, to obtain the hyperparameter encoding for uncertainty-driven information acquisition, including: Based on the fitness of the hyperparameter encoding after the acquisition of the multi-point guidance information, the uncertainty factor is obtained as follows: ; ; ; In the formula, As an uncertainty factor, It is a logarithmic function. The total number of hyperparameter codes. To optimize the state factor, For the first i The fitness percentage of hyperparameter encoding after multi-point guidance information collection. For the first i The fitness of hyperparameter coding after multi-point guidance information collection; Based on the total number of optimization attempts, the information collection control factor is as follows:

[0050] in, To collect control factors, This is the minimum value of the information collection control factor, and it is set to 0.05. This is the dynamic balance factor, and it is set to 0.5; The minimum value among the fitness values ​​of the hyperparameter encoding after multi-point guidance information acquisition. The maximum value in the fitness of the hyperparameter encoding after multi-point guidance information acquisition; Based on the first optimal encoding, the uncertainty factor, and the information acquisition control factor, the hyperparameter encoding for uncertainty-driven information acquisition is obtained as follows:

[0051] In the formula, For the first i Hyperparameter encoding after multi-point guidance information collection For the first i Hyperparameter encoding for information acquisition driven by uncertainty The first optimal code, This is the first learning coefficient, and it is set to 1 or 2; This is the second learning coefficient, and it is set to 0.75 or 1.5; The first random number between (0,1) The second random number between (0,1) For fitness greater than The random encoding, and in When fitness is at its maximum, Set a random code.

[0052] In the embodiments of this application, the algorithm does not simply perform a blind search, but evaluates the current search state through uncertainty factors and adjusts the search step size and strategy accordingly. By utilizing uncertainty factors and combining the first optimal code and the random excellent code, the algorithm can dynamically adjust the step size towards the optimal solution based on the quality of the current solution, thereby ensuring the accuracy of the final hyperparameters and improving the convergence speed of the algorithm.

[0053] In this embodiment of the application, a diversity information enhancement acquisition mechanism is used to perform diversity information enhancement acquisition on the hyperparameter encoding of uncertainty-driven information acquisition, resulting in hyperparameter encoding after diversity information enhancement acquisition, including: The diversity information factor for obtaining hyperparameter encoding of uncertainty-driven information acquisition is: ; ; In the formula, For the first t During the second optimization process, the first m The first hyperparameter encoding of the uncertainty-driven information acquisition d dimensional hyperparameters, m=1,2,…,N, where N is the total number of hyperparameter codes. The first hyperparameter encoding corresponding to the average hyperparameter encoding of all uncertainty-driven information acquisition. d dimensional hyperparameters, The first optimal code d dimensional hyperparameters, For the first m The diversity coding corresponding to the hyperparameter coding of uncertainty-driven information acquisition. d dimensional hyperparameters, for Corresponding diversity information factors; Based on the aforementioned diversity information factor, the hyperparameter encoding after enhanced diversity information acquisition is obtained as follows: ; In the formula, For the first m The first hyperparameter encoding after the acquisition of enhanced diversity information. d dimensional hyperparameters, The search coefficients are randomly distributed, and , The third random number between (0,1) The fourth random number between (0,1) Pi is the mathematical constant of a circle.

[0054] The diversity information enhancement acquisition mechanism provided in this application embodiment can effectively enhance the random search characteristics of the algorithm in the early and middle stages, ensure the comprehensiveness of the solution space exploration, and thus ensure that the algorithm does not get stuck in local optima. However, by utilizing the diversity characteristics of group information, it can avoid completely random search, which not only improves the efficiency of global search, but also avoids destroying the original search experience during the solution space search process. When the search stalls, the algorithm can be forced to jump out of the current local area through random perturbation and continue to explore the unknown solution space.

[0055] Although the algorithm itself is a hyperparameter optimization, it directly serves the training of the first agent (neural network). The second optimal encoding obtained through the above mechanism serves as the final hyperparameter, enabling the first agent to minimize cross-entropy loss or root mean square error during training. This results in a higher accuracy rate for the trained agent in identifying historical abnormal process data. Because the optimization process considers diversity (avoiding overfitting to a specific distribution), the agent exhibits stronger adaptability and anti-interference capabilities when faced with unseen and new abnormal process data. It can more effectively identify anomaly type labels, thereby providing accurate data support for subsequent automatic control or manual intervention, and enabling rapid response and handling of process faults.

[0056] In this embodiment, based on historical abnormal process data, abnormal process type labels, and actual abnormal handling decision schemes, a second intelligent agent is pre-trained using a multi-strategy information acquisition optimization algorithm to obtain the trained second intelligent agent. The trained second intelligent agent is then used to extract sample feature data corresponding to the historical abnormal process data and abnormal process type labels, including: K-means clustering was performed on the historical abnormal process data and the historical sample data composed of abnormal process type labels to obtain K cluster centers. Treat the actual anomaly handling decision scheme corresponding to any cluster center as the same scheme, and assign a unique numerical label to each cluster center. Based on the historical sample data in the cluster centers and the numerical labels corresponding to the cluster centers, a second intelligent agent is pre-trained using a multi-strategy information acquisition optimization algorithm to obtain the trained second intelligent agent. The historical sample data is used as input to the trained second agent, and the output of the feature layer of the second agent is obtained to obtain the sample feature data.

[0057] The training process of the second agent is similar to that of the first agent, except that the input data and the expected output data are different. This application will not elaborate further in its embodiments.

[0058] In this embodiment of the application, real-time process data is collected during the chemical production process, and the trained first intelligent agent is invoked to identify the real-time process data and determine process alarm information, including: Real-time process data is collected during chemical production, and the real-time process data is used as the input of the first intelligent agent after training to obtain the probability distribution of the output of the first intelligent agent after training. Based on the probability distribution output by the first intelligent agent after training, the abnormal process type with the highest probability is determined as the target abnormal process type. If the target abnormal process type is any specific abnormal process type, then the process alarm information is determined to be triggered; otherwise, the process alarm information is determined not to be triggered.

[0059] By using a pre-trained agent to perform online inference on real-time streaming data, subtle fluctuations and abnormal trends in the production process can be quickly detected without human intervention. An alarm is triggered only when the most probable anomaly type is identified as a specific type. This probability-based discrimination mechanism is more robust than simple threshold-based judgment, effectively filtering out ambiguous or uncertain interference signals, significantly reducing the probability of false alarms and missed alarms, and ensuring the accuracy of alarms.

[0060] In this embodiment of the application, when the process alarm information is a specific abnormal process type, a trained second agent is used to extract features from the real-time process data and the process alarm information to obtain real-time feature data, including: When the process alarm information is a specific abnormal process type, the real-time process data and the process alarm information are combined into a vector to obtain real-time sample data; The real-time sample data is used as input to the trained second agent to obtain the real-time feature data corresponding to the real-time sample data output by the trained second agent.

[0061] In this embodiment of the application, the real-time feature data is matched with the sample feature data to match similar historical abnormal process data and similar abnormal process anomaly type labels to obtain similar historical sample data, including: Obtain the cosine similarity between the real-time feature data and the sample feature data, and take the sample feature data with the largest cosine similarity as the target sample feature data; The similar historical abnormal process data corresponding to the target similar sample feature data and the similar abnormal process abnormality type label are used together as similar historical sample data.

[0062] Feature matching can quickly find the most similar historical sample data for the current process alarm, thereby providing historical operation data to assist staff in making quick decisions. This feedback learning mechanism will become more and more accurate as staff handle more events.

[0063] In this embodiment of the application, based on the actual anomaly handling decision scheme corresponding to the similar historical sample data, a target auxiliary decision scheme for process alarms under abnormal operating conditions is generated, including: The cluster centers corresponding to the similar historical sample data are determined as the target cluster centers; The actual anomaly handling decision scheme corresponding to the similar historical sample data is determined as the first candidate auxiliary decision scheme. The M actual anomaly handling decision schemes that appear most frequently in the target cluster centers are identified as the second candidate auxiliary decision schemes; M represents a pre-set number of normals, such as 5. The first candidate auxiliary decision scheme and the second candidate auxiliary decision scheme are deduplicated to obtain at least one target auxiliary decision scheme for process alarm under abnormal operating conditions.

[0064] In addition to the polypropylene reactor described above, the technical solutions described in this application can also be widely applied to the following chemical scenarios: Distillation column control: Input data includes column top temperature, column bottom temperature, reflux ratio, and feed composition. Anomaly types include flooding and leakage. Output decisions include adjusting the reflux ratio or feed position.

[0065] Boiler combustion system: Input data includes furnace temperature, flue gas oxygen content, and fuel pressure. Anomaly types include flameout and coking. Output decisions include adjusting the air-fuel ratio or performing soot blowing.

[0066] Continuous reforming unit: Monitor reactor bed temperature drop, identify abnormal catalyst carbon buildup, and recommend regeneration solutions.

[0067] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0068] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, apparatuses, electronic devices, and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0069] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0070] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0071] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0072] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0073] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A multi-agent response-assisted decision-making method for process alarms under abnormal operating conditions, characterized in that, include: Collect historical abnormal process data, abnormal process type labels, and actual abnormal handling decision-making schemes under various abnormal operating conditions; The abnormal process type label includes a specific abnormal process type or a normal process type; Based on the historical abnormal process data and abnormal process type labels, a first intelligent agent is pre-trained using a multi-strategy information acquisition optimization algorithm to obtain the trained first intelligent agent. Based on historical abnormal process data, abnormal process type labels, and actual abnormal handling decision schemes, a second intelligent agent is pre-trained using a multi-strategy information acquisition optimization algorithm to obtain the trained second intelligent agent. The trained second intelligent agent is then used to extract the sample feature data corresponding to the historical abnormal process data and abnormal process type labels. Real-time process data is collected during chemical production, and the trained first intelligent agent is invoked to identify the real-time process data and determine process alarm information. When the process alarm information is a specific abnormal process type, a trained second agent is used to extract features from the real-time process data and the process alarm information to obtain real-time feature data. The real-time feature data is matched with the sample feature data to match similar historical abnormal process data and similar abnormal process anomaly type labels to obtain similar historical sample data. Based on the actual anomaly handling decision schemes corresponding to the similar historical sample data, a target auxiliary decision scheme for process alarms under abnormal operating conditions is generated.

2. The multi-agent response auxiliary decision-making method for process alarms under abnormal operating conditions according to claim 1, characterized in that, Based on the aforementioned historical abnormal process data and the labels of abnormal process types, a first intelligent agent is pre-trained using a multi-strategy information acquisition optimization algorithm to obtain the trained first intelligent agent, which includes: The hyperparameters of the first agent are randomly generated and encoded to obtain hyperparameter codes. Multiple different hyperparameter codes are obtained repeatedly. For any hyperparameter code, the fitness corresponding to the hyperparameter code is obtained based on the historical abnormal process data and the abnormal process type label. The hyperparameter encoding with the highest fitness is determined as the first optimal encoding; A multi-point guidance information acquisition mechanism is used to acquire multi-point guidance information for the hyperparameter encoding, resulting in the hyperparameter encoding after multi-point guidance information acquisition. Based on the first optimal encoding, an uncertainty-driven information acquisition mechanism is used to perform uncertainty-driven information acquisition on the hyperparameter encoding after multi-point guidance information acquisition, so as to obtain the hyperparameter encoding of uncertainty-driven information acquisition. A diversity information enhancement acquisition mechanism is adopted to perform diversity information enhancement acquisition on the hyperparameter encoding of uncertainty-driven information acquisition, and the hyperparameter encoding after diversity information enhancement acquisition is obtained. Determine whether the total number of optimization attempts is greater than or equal to the preset maximum number of optimization attempts. If so, determine the second optimal code by enhancing the hyperparameter encoding after the collection of diversity information; otherwise, return to the fitness acquisition step. Based on the second optimal encoding, obtain the first agent after training.

3. The multi-agent response auxiliary decision-making method for process alarms under abnormal operating conditions according to claim 2, characterized in that, A multi-point guidance information acquisition mechanism is used to acquire multi-point guidance information for the hyperparameter encoding, resulting in hyperparameter encoding after multi-point guidance information acquisition, including: Based on the total number of optimization iterations, the multi-point guidance speed control factor is obtained as follows: ； In the formula, For multi-point guidance speed control factor, It is an exponential function with the natural constant e as its base. It is a constant term, and is set to 5; This is the maximum value of the multi-point guidance speed control factor, and it is set to 0.95; This is the minimum value of the multi-point guidance speed control factor, and it is set to 0.0005. The total number of optimizations is T, where T is the preset maximum number of optimizations. The guidance speed is controlled using the multi-point guidance speed control factor. All hyperparameter codes are used as guidance parameters, and control is performed by combining the boundary information of the hyperparameter codes. The hyperparameter codes obtained after the multi-point guidance information acquisition are as follows: ； ； In the formula, For the first t During the second optimization process, the first k The first hyperparameter encoding d dimensional hyperparameters, for Corresponding multi-point guidance information, For the first k The first hyperparameter encoding after multi-point guidance information acquisition d dimensional hyperparameters, k =1,2,…,N, where N is the total number of hyperparameter codes. d =1,2,…,Din, where Din is the total dimension of hyperparameters in hyperparameter encoding. For multi-point guidance speed control factor, For the first d Upper limit of hyperparameters, For the first d Lower bound of hyperparameters This is the intensity coefficient of the group information parameter, and it is set to 0.5; It is a natural constant. For the first t During the second optimization process, the first j The first hyperparameter encoding d dimensional hyperparameters, For the first k The hyperparameter encoding and the first j The Euclidean distance between the hyperparameter codes, where c represents the population information parameter direction coefficient and is set to 1.

5. The first optimal code represents the first... d Dimensional hyperparameters.

4. The multi-agent response auxiliary decision-making method for process alarms under abnormal operating conditions according to claim 3, characterized in that, Based on the first optimal encoding, an uncertainty-driven information acquisition mechanism is used to perform uncertainty-driven information acquisition on the hyperparameter encoding after multi-point guidance information acquisition, resulting in the hyperparameter encoding for uncertainty-driven information acquisition, including: Based on the fitness of the hyperparameter encoding after the acquisition of the multi-point guidance information, the uncertainty factor is obtained as follows: ； ； ； In the formula, As an uncertainty factor, It is a logarithmic function. The total number of hyperparameter codes. To optimize the state factor, For the first i The fitness percentage of hyperparameter encoding after multi-point guidance information collection. For the first i The fitness of hyperparameter coding after multi-point guidance information collection; Based on the total number of optimization attempts, the information collection control factor is as follows: in, To collect control factors, This is the minimum value of the information collection control factor, and it is set to 0.

05. This is the dynamic balance factor, and it is set to 0.5; The minimum value among the fitness values ​​of the hyperparameter encoding after multi-point guidance information acquisition. The maximum value T in the fitness of hyperparameter encoding after multi-point guidance information acquisition is , t is , ; Based on the first optimal encoding, the uncertainty factor, and the information acquisition control factor, the hyperparameter encoding for uncertainty-driven information acquisition is obtained as follows: In the formula, For the first i Hyperparameter encoding after multi-point guidance information collection For the first i Hyperparameter encoding for information acquisition driven by uncertainty The first optimal code, This is the first learning coefficient, and it is set to 1 or 2; This is the second learning coefficient, and it is set to 0.75 or 1.5; The first random number between (0,1) The second random number between (0,1) For fitness greater than The random encoding, and in When fitness is at its maximum, Set a random code.

5. The multi-agent response auxiliary decision-making method for process alarms under abnormal operating conditions according to claim 4, characterized in that, A diversity information enhancement acquisition mechanism is employed to perform diversity information enhancement acquisition on the hyperparameter encoding of uncertainty-driven information acquisition, resulting in hyperparameter encoding after diversity information enhancement acquisition, including: The diversity information factor for obtaining hyperparameter encoding of uncertainty-driven information acquisition is: ； ； In the formula, For the first t During the second optimization process, the first m The first uncertainty-driven hyperparameter encoding of information acquisition d dimensional hyperparameters, m =1,2,…,N, where N is the total number of hyperparameter codes. The first hyperparameter encoding corresponding to the average hyperparameter encoding of all uncertainty-driven information acquisition. d dimensional hyperparameters, The first optimal code d dimensional hyperparameters, For the first m The diversity coding corresponding to the hyperparameter coding of uncertainty-driven information acquisition. d dimensional hyperparameters, for Corresponding diversity information factors; Based on the aforementioned diversity information factor, the hyperparameter encoding after enhanced diversity information acquisition is obtained as follows: ； In the formula, For the first m The first hyperparameter encoding after the acquisition of enhanced diversity information. d dimensional hyperparameters, The search coefficients are randomly distributed, and , The third random number between (0,1) The fourth random number between (0,1) Pi is the mathematical constant of a circle.

6. The multi-agent response auxiliary decision-making method for process alarms under abnormal operating conditions according to claim 1, characterized in that, Based on historical abnormal process data, abnormal process anomaly type labels, and actual anomaly handling decision schemes, a second intelligent agent is pre-trained using a multi-strategy information acquisition optimization algorithm to obtain the trained second intelligent agent. The trained second intelligent agent is then used to extract sample feature data corresponding to the historical abnormal process data and abnormal process anomaly type labels, including: K-means clustering was performed on the historical abnormal process data and the historical sample data composed of abnormal process type labels to obtain K cluster centers. Treat the actual anomaly handling decision scheme corresponding to any cluster center as the same scheme, and assign a unique numerical label to each cluster center. Based on the historical sample data of the cluster centers and the numerical labels corresponding to the cluster centers, a second intelligent agent is pre-trained using a multi-strategy information acquisition optimization algorithm to obtain the trained second intelligent agent. The historical sample data is used as input to the trained second agent, and the output of the feature layer of the second agent is obtained to obtain the sample feature data.

7. The multi-agent response auxiliary decision-making method for process alarms under abnormal operating conditions according to claim 6, characterized in that, Real-time process data is collected during chemical production, and the trained first intelligent agent is invoked to identify the real-time process data and determine process alarm information, including: Real-time process data is collected during chemical production, and the real-time process data is used as the input of the first intelligent agent after training to obtain the probability distribution of the output of the first intelligent agent after training. Based on the probability distribution output by the first intelligent agent after training, the abnormal process type with the highest probability is determined as the target abnormal process type. If the target abnormal process type is any specific abnormal process type, then the process alarm information is determined to be triggered; otherwise, the process alarm information is determined not to be triggered.

8. The multi-agent response auxiliary decision-making method for process alarms under abnormal operating conditions according to claim 7, characterized in that, When the process alarm information is a specific abnormal process type, a trained second agent is used to extract features from the real-time process data and the process alarm information to obtain real-time feature data, including: When the process alarm information is a specific abnormal process type, the real-time process data and the process alarm information are combined into a vector to obtain real-time sample data; The real-time sample data is used as input to the trained second agent to obtain the real-time feature data corresponding to the real-time sample data output by the trained second agent.

9. The multi-agent response auxiliary decision-making method for process alarm under abnormal operating conditions according to claim 8, characterized in that, The real-time feature data is matched with the sample feature data to obtain similar historical abnormal process data and similar abnormal process anomaly type labels, including: Obtain the cosine similarity between the real-time feature data and the sample feature data, and take the sample feature data with the largest cosine similarity as the target sample feature data; The similar historical abnormal process data corresponding to the target similar sample feature data and the similar abnormal process abnormality type label are used together as similar historical sample data.

10. The multi-agent response auxiliary decision-making method for process alarms under abnormal operating conditions according to claim 9, characterized in that, Based on the actual anomaly handling decision schemes corresponding to the aforementioned similar historical sample data, a target auxiliary decision scheme for process alarms under abnormal operating conditions is generated, including: The cluster centers corresponding to the similar historical sample data are determined as the target cluster centers; The actual anomaly handling decision scheme corresponding to the similar historical sample data is determined as the first candidate auxiliary decision scheme. The M actual anomaly handling decision schemes that appear most frequently in the target cluster center are determined as the second candidate auxiliary decision schemes, where M represents the pre-set normal number; The first candidate auxiliary decision scheme and the second candidate auxiliary decision scheme are deduplicated to obtain at least one target auxiliary decision scheme for process alarm under abnormal operating conditions.

Images