Calcium formate production process safety monitoring method and system

Abstract

The invention relates to the field of safety monitoring, and particularly discloses a calcium formate production process safety monitoring method and system, which utilize a distributed control system to collect original production data, and ensure time consistency of all monitoring variables through data synchronization processing, thereby laying a foundation for subsequent analysis. Thirdly, performing feature engineering processing on the synchronized data, extracting features such as short-term and medium-term change rates, moving average values and standard deviations of temperature, pressure and stirrer current, constructing state features of the reaction kettle, and capturing dynamic operation information of the system; in order to predict potential risks, a trend prediction model based on deep learning is adopted, a historical temperature sequence is converted into a future temperature trajectory, and prospective pre-judgment on the temperature trend is provided; meanwhile, the current system health condition is evaluated in combination with a state recognition model, the risk level is determined through a decision tree model according to the abnormal score and the predicted temperature maximum value, and graded early warning is achieved.

Description

Technical Field

[0001] The present application relates to the field of safety monitoring, and more specifically, to a method and system for safety monitoring of a calcium formate production process. Background Art

[0002] In chemical production processes, especially in the synthesis of calcium formate, the reactor, as a core piece of equipment, undertakes the critical task of carrying out complex chemical reactions requiring high temperatures and high pressures. Since the production of calcium formate involves the neutralization of strongly acidic substances (such as formic acid) with calcium salts, the intensity of the reaction is affected by multiple factors, including temperature, pressure, and stirring conditions. Therefore, extremely high requirements are placed on the safety monitoring of the reaction process. Problems such as temperature runaway, abnormal pressure, or agitator malfunction can easily lead to safety accidents, even resulting in serious consequences such as explosions or toxic gas leaks. Therefore, establishing an efficient, intelligent, and forward-looking safety monitoring system is crucial for ensuring production continuity, personnel safety, and environmental protection.

[0003] Currently, in the production process of calcium formate chemical products, traditional safety monitoring methods mainly rely on real-time data collected by distributed control systems and identify risks through simple threshold alarm mechanisms. However, this static threshold-based alarm method often has problems such as delayed response, high false alarm rate, and lack of trend prediction capabilities. For example, when the temperature rises slowly but has not yet reached the alarm threshold, the system cannot provide early warning of potential thermal runaway risks; or when multiple variables are coupled and changing, the anomaly of a single variable cannot accurately reflect the stability of the overall system. In addition, traditional methods usually only focus on the data snapshot at the current moment, ignoring the trend information and dynamic patterns contained in historical data, which limits their ability to identify abnormal behavior under complex working conditions.

[0004] Therefore, an optimized safety monitoring scheme for the calcium formate production process is desired. Summary of the Invention

[0005] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides a calcium formate production process safety monitoring method and system, which uses a distributed control system to collect original production data, and ensures the time consistency of all monitoring variables through data synchronization processing, thereby laying the foundation for subsequent analysis. Then, the synchronized data is subjected to feature engineering processing, and features such as the short-term and medium-term change rates, moving averages and standard deviations of temperature, pressure and agitator current are extracted to construct the reactor state features and capture the dynamic operation information of the system. In order to predict potential risks, a trend prediction model based on deep learning is adopted to convert historical temperature sequences into future temperature trajectories, providing a forward-looking prediction of temperature trends. At the same time, the current system health status is evaluated in combination with the state recognition model, and the risk level is determined according to the anomaly score and the predicted maximum temperature through a decision tree model to achieve accurate graded warning. The entire process effectively solves the problems of delayed response and high false alarm rate of traditional safety monitoring means through real-time data processing, feature extraction, trend prediction and intelligent decision support, thereby improving production safety.

[0006] According to one aspect of the present application, a method for safety monitoring of a calcium formate production process is provided, comprising: Obtain original DCS data collected by the distributed control system; performing data synchronization on the raw DCS data to obtain a time sequence of synchronized data frames, wherein the synchronized data frames include temperature, pressure, and agitator current; Performing feature engineering and state vector construction on the time series of the synchronized data frame to obtain a reactor state feature vector and a historical temperature series; Inputting the reactor state feature vector and the historical temperature sequence into a parallel monitoring component including a state recognition model and a trend prediction model to obtain an anomaly score and a predicted temperature trajectory; Inputting the anomaly score and the predicted temperature trajectory into a decision tree model to obtain a risk level; Based on the risk level, different levels of audible and visual alarms are triggered.

[0007] According to another aspect of the present application, a calcium formate production process safety monitoring system is provided, comprising: Original DCS data acquisition module, used to obtain original DCS data collected by the distributed control system; a data synchronization module, configured to synchronize the original DCS data to obtain a time sequence of synchronized data frames, wherein the synchronized data frames include temperature, pressure, and agitator current; A feature engineering module, configured to perform feature engineering and state vector construction on the time series of the synchronized data frame to obtain a reactor state feature vector and a historical temperature series; A parallel monitoring module, configured to input the reactor state feature vector and the historical temperature sequence into a parallel monitoring component comprising a state recognition model and a trend prediction model to obtain an anomaly score and a predicted temperature trajectory; a risk level decision module, configured to input the anomaly score and the predicted temperature trajectory into a decision tree model to obtain a risk level; The sound and light alarm triggering module is used to trigger different levels of sound and light alarms based on the risk level.

[0008] Beneficial effect: Compared with the prior art, the present application provides a method and system for monitoring the safety of the calcium formate production process, which uses a distributed control system to collect original production data and ensures the time consistency of all monitoring variables through data synchronization processing, thereby laying the foundation for subsequent analysis. Then, the synchronized data is subjected to feature engineering processing to extract features such as the short-term and medium-term change rates, moving averages and standard deviations of temperature, pressure and agitator current, construct the state characteristics of the reactor, and capture the dynamic operation information of the system. In order to predict potential risks, a trend prediction model based on deep learning is used to convert historical temperature sequences into future temperature trajectories, providing a forward-looking prediction of temperature trends. At the same time, the state recognition model is combined to evaluate the current system health status, and the decision tree model is used to determine the risk level based on the anomaly score and the predicted maximum temperature, thereby achieving accurate graded warning. The entire process effectively solves the problems of delayed response and high false alarm rate of traditional safety monitoring means through real-time data processing, feature extraction, trend prediction and intelligent decision support, thereby improving production safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] The above and other purposes, features, and advantages of the present application will become more apparent through a more detailed description of the embodiments of the present application in conjunction with the accompanying drawings. The accompanying drawings are intended to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 Flowchart of a method for safety monitoring of a calcium formate production process according to an embodiment of the present application; Figure 2 Schematic diagram of data flow of a method for safety monitoring of a calcium formate production process according to an embodiment of the present application; Figure 3 A flowchart of a method for safety monitoring of a calcium formate production process according to an embodiment of the present application for performing feature engineering and state vector construction on the time series of the synchronized data frame to obtain a reactor state feature vector and a historical temperature sequence; Figure 4A flowchart of the method for safety monitoring of a calcium formate production process according to an embodiment of the present application, wherein the historical temperature sequence is input into a pre-trained trend prediction model to obtain the predicted temperature trajectory; Figure 5 4 is a block diagram of a calcium formate production process safety monitoring system according to an embodiment of the present application. DETAILED DESCRIPTION

[0011] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described herein.

[0012] As used in this application and the claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not intended to refer to the singular but may include the plural. Generally speaking, the terms "comprises" and "include" only indicate the inclusion of the steps and elements specifically identified, and these steps and elements do not constitute an exclusive list. A method or apparatus may also include other steps or elements.

[0013] Although the present application makes various references to certain modules in the system according to embodiments of the present application, any number of different modules can be used and run on the user terminal and / or server. The modules are illustrative only, and different aspects of the system and method can use different modules.

[0014] Flowcharts are used in this application to illustrate the operations performed by the systems according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, the various steps may be processed in reverse order or simultaneously, as needed. Furthermore, other operations may be added to these processes, or one or more operations may be removed from these processes.

[0015] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described herein.

[0016] To address safety hazards in the calcium formate production process, particularly the potential risks arising from fluctuations in key parameters such as temperature, pressure, and agitator current, the technical solution of this application proposes a comprehensive safety monitoring method. This method begins with data acquisition, utilizing a distributed control system (DCS) to obtain raw production data. Data synchronization ensures temporal consistency across all monitored variables, addressing data analysis errors caused by inconsistent timestamps. Next, feature engineering is performed on the synchronized data frames to extract a variety of derivative and statistical features, including short-term and medium-term rates of change, moving averages, and standard deviations. This constructs a reactor state feature vector, capturing richer system operational information. To predict potential future anomalies, this solution introduces a deep learning-based trend prediction model. Taking historical temperature series as input, this model extracts local temporal features through one-dimensional convolutional coding, implements global pattern encoding using information transfer techniques, and ultimately decodes the future temperature trajectory. This method not only considers the state at a single point in time but also integrates dynamic trends over long time spans, improving the ability to predict potentially hazardous trends. In addition, the health status of the current system is evaluated through a pre-trained state recognition model, and the risk level is determined based on the anomaly score and the predicted maximum temperature using a decision tree model, thus achieving accurate graded warnings for different levels of risks.

[0017] In the technical solution of the present application, a method for safety monitoring of the calcium formate production process is proposed. Figure 1 Flowchart of a method for safety monitoring of a calcium formate production process according to an embodiment of the present application. Figure 2 Schematic diagram of data flow of the calcium formate production process safety monitoring method according to the embodiment of the present application. Figure 1 and Figure 2 As shown, according to an embodiment of the present application, a method for safety monitoring of a calcium formate production process includes the following steps: S100, acquiring original DCS data collected by a distributed control system; S200, performing data synchronization on the original DCS data to obtain a time series of synchronized data frames, wherein the synchronized data frames include temperature, pressure, and agitator current; S300, performing feature engineering and state vector construction on the time series of the synchronized data frames to obtain a reactor state feature vector and a historical temperature sequence; S400, inputting the reactor state feature vector and the historical temperature sequence into a parallel monitoring component including a state recognition model and a trend prediction model to obtain an anomaly score and a predicted temperature trajectory; S500, inputting the anomaly score and the predicted temperature trajectory into a decision tree model to obtain a risk level; and S600, triggering different levels of audible and visual alarms based on the risk level.

[0018] Specifically, in steps S100 and S200, raw DCS data collected by a distributed control system (DCS) is acquired and synchronized to produce a time series of synchronized data frames. The synchronized data frames include temperature, pressure, and agitator current. It should be understood that in the calcium formate production process, parameters such as temperature, pressure, and agitator current within the reactor are key indicators reflecting the system's operating status, and their changes are often closely related to the stability of the chemical reaction. Because these parameters are typically collected independently by a distributed control system (DCS), the data update frequency and communication timing of different variables vary, resulting in the raw DCS data being incompletely synchronized in the time dimension. This asynchrony directly impacts the accuracy of subsequent feature extraction and model analysis, thereby weakening the ability to determine system status. Therefore, in the technical solution of the present application, after acquiring the raw DCS data collected by the distributed control system, the raw DCS data is further synchronized to produce a time series of synchronized data frames, ensuring that all monitored variables form consistent data frames based on the same time reference. The synchronized data frames include temperature, pressure, and agitator current. This allows the construction of a structurally unified, time-aligned sequence of synchronized data frames, enabling key variables such as temperature, pressure, and agitator current to be analyzed collaboratively within the same time window, thus more realistically reflecting the real-time status of the reactor.

[0019] More specifically, in this embodiment of the present application, data synchronization of the raw DCS data to obtain a time sequence of synchronized data frames includes: In response to the operation of the master clock, performing a synchronization operation on the raw DCS data; specifically, the synchronization operation includes: looking back at the previous time window and finding the value with the latest timestamp for each variable to be monitored; and in response to the missing value of a variable, directly using the value of the variable at the previous time point to fill it in. In other words, each time the master clock is triggered, the system will look back at the timestamp of each variable in the previous time window and select the latest value closest to the current time; if a variable is missing at this time point, it will be filled in with the value at the previous time point, thereby maintaining the continuity and integrity of the data stream.

[0020] Specifically, in step S300, feature engineering and state vector construction are performed on the time series of the synchronized data frame to obtain the reactor state feature vector and the historical temperature series. It should be understood that in the production process of calcium formate, changes in parameters such as temperature, pressure and agitator current inside the reactor often indicate the stability of the system operation state. In order to more accurately identify potential risks, it is not enough to rely solely on raw data. The time series must be deeply processed through feature engineering to extract more representative information for model analysis. Therefore, it is necessary to perform feature engineering and state vector construction on the synchronized data frame, which is conducive to improving the interpretability and predictive ability of the data, so that the monitoring system can capture the dynamic trends and abnormal patterns hidden behind the raw data.

[0021] Figure 3 This is a flow chart of performing feature engineering and state vector construction on the time series of the synchronous data frame to obtain the reactor state feature vector and historical temperature sequence according to an embodiment of the present application. Figure 3 As shown, according to the calcium formate production process safety monitoring method of an embodiment of the present application, step S300 includes: S310, extracting the historical temperature sequence from the time series of the synchronous data frames; S320, extracting the latest synchronous data frame from the time series of the synchronous data frames; S330, performing derivative feature calculation and statistical feature calculation based on the time series of the synchronous data frames to obtain derivative features and statistical features; S340, adding the derived features and statistical features to the latest synchronous data frame to obtain the reactor state feature vector. The derived features include the short-term temperature change rate, the medium-term temperature change rate, the short-term pressure change rate, the medium-term pressure change rate, the short-term agitator current change rate, and the medium-term agitator current change rate. The statistical features include the temperature moving average, the temperature moving standard deviation, the pressure moving average, the pressure moving standard deviation, the agitator current moving average, and the agitator current moving standard deviation. In other words, key features reflecting system behavior are extracted from the time series of the synchronous data frames, including trend sensitivity reflected by the short-term and medium-term change rates, and state stability and volatility reflected by statistical features such as the moving average and standard deviation. These derived and statistical features are integrated into the latest synchronized data frame to form a feature vector that comprehensively describes the reactor's current state, while retaining the historical temperature series for subsequent trend prediction modeling. This multi-dimensional feature fusion approach not only enhances the model's perception of complex operating conditions but also improves the accuracy of identifying potential anomalies.

[0022] In step S330, derived features and statistical features are calculated based on the time series of the synchronized data frames to obtain derived features and statistical features. Specifically, in a specific example of this application, the short-term and medium-term rates of change of the three key variables, temperature, pressure, and agitator current, are first calculated as derived features. Taking temperature as an example, the temperature difference between the current and previous moments is calculated using a sliding window difference method within a time window (e.g., the past 5 minutes) and divided by the time interval to obtain the short-term rate of change of temperature. The medium-term rate of change is calculated by fitting a local trend line within a longer time window (e.g., the past 30 minutes), and its slope is used as a representation of the trend. A similar method is applied to the rate of change calculation of pressure and agitator current. Secondly, statistical features are constructed based on aggregate calculations within a moving window, using a moving average and moving standard deviation to reflect the central tendency and volatility of the variable over a period of time. For example, using a 10-minute sliding window, the mean and standard deviation of temperature within the window are calculated sequentially to form smooth features on the time series, thereby revealing the stability of system operation. Similarly, this method is also applicable to extracting statistical features for pressure and agitator current. Finally, all these derived and statistical features are concatenated into the latest synchronized data frame to form a complete reactor state feature vector for subsequent modeling. This process can also be implemented using lightweight neural network models, such as one-dimensional convolutional networks (1D-CNN) or recurrent neural networks (such as LSTMs) to automatically learn and encode temporal features, thereby replacing some manual feature engineering and improving feature extraction efficiency and generalization capabilities.

[0023] Specifically, in step S400, the reactor state feature vector and the historical temperature sequence are input into a parallel monitoring component comprising a state recognition model and a trend prediction model to obtain an abnormality score and a predicted temperature trajectory. It should be understood that in the calcium formate production process, it is crucial to ensure the safe and stable operation of the reactor. However, traditional monitoring methods often rely on a fixed threshold alarm mechanism, which makes it difficult to provide early warning of potential risks, especially when faced with complex multivariable coupling changes. Therefore, by inputting the reactor state feature vector and the historical temperature sequence into a parallel monitoring component comprising a state recognition model and a trend prediction model, it is intended to achieve real-time assessment of the current system health status and make accurate predictions of abnormal situations that may occur in the future.

[0024] More specifically, in an embodiment of the present application, the reactor state feature vector and historical temperature sequence are input into a parallel monitoring component comprising a state recognition model and a trend prediction model to obtain an anomaly score and a predicted temperature trajectory, including: inputting the reactor state feature vector into a pre-trained state recognition model to obtain the anomaly score; and inputting the historical temperature sequence into a pre-trained trend prediction model to obtain the predicted temperature trajectory. In other words, since single state monitoring cannot fully reflect the actual operating status of the system, trend analysis combined with historical data can provide more forward-looking risk insights. Specifically, the reactor state feature vector contains the latest dynamic information of key parameters such as temperature, pressure, and agitator current, as well as their derived and statistical features. This information is input into the pre-trained state recognition model to calculate an anomaly score that reflects the degree to which the current system deviates from the normal operating range. At the same time, the historical temperature sequence is input into the trend prediction model as input. This model, based on a deep learning algorithm, can capture the complex patterns of temperature evolution over time and, based on this, infer the temperature trajectory over a period of time in the future.

[0025] Specifically, the reactor state feature vector is input into a pre-trained state recognition model to obtain the anomaly score. It should be understood that reactor state monitoring is crucial in the calcium formate production process, as any deviation from the normal operating range can potentially lead to serious safety issues. Traditional alarm mechanisms based on fixed thresholds struggle to adapt to complex industrial environments and cannot accurately capture all potential risk factors. To effectively identify these abnormal states, achieve real-time assessment of the reactor's operating status, and determine whether there are safety hazards requiring immediate action, the technical solution of this application inputs the reactor state feature vector into a pre-trained state recognition model to obtain an anomaly score. By introducing a machine learning or deep learning model, it can be trained using a large amount of historical data, enabling more accurate identification of whether the system is currently in an abnormal state. Specifically, the model analyzes variables such as temperature, pressure, and agitator current, as well as their derivatives and statistical characteristics, to calculate an anomaly score that reflects the extent to which the system deviates from its normal operating range. This score not only considers changes in individual variables but also integrates the interactions between multiple variables, providing a more comprehensive risk assessment.

[0026] More specifically, in one example of this application, a deep learning model suitable for processing time series data, such as a long short-term memory (LSTM) network or a one-dimensional convolutional neural network (1D-CNN), is first constructed. These models automatically extract key patterns from input feature vectors that help distinguish normal from abnormal conditions. Taking the LSTM as an example, during the training phase, a large amount of historical data, including known normal and abnormal samples, is used as input. The model learns how to predict the system health status based on the feature vectors. Once the model is trained, it can be used for online monitoring, receiving the latest state feature vectors from the reactor in real time and calculating an anomaly score through forward propagation. This deep learning-based state recognition method not only detects obvious faults promptly, but also identifies subtle deviations that are less noticeable but can still lead to serious consequences. Furthermore, because the model is pre-trained, it can respond quickly in real-world applications without adding additional time delays, ensuring immediate warning of emergencies. This provides an efficient and reliable solution for risk management in chemical production, significantly improving production safety.

[0027] Specifically, the historical temperature sequence is input into a pre-trained trend prediction model to obtain the predicted temperature trajectory. It should be understood that traditional monitoring methods can generally only provide status information at the current moment and lack the ability to effectively predict future trends. In order to effectively identify and prevent possible abnormal situations, especially those that may lead to temperature runaway, the historical temperature sequence is input into a pre-trained trend prediction model to obtain a predicted temperature trajectory. By utilizing historical data for trend analysis, potential risks can be warned in advance, thereby buying time to take preventive measures. This not only enables real-time monitoring of current temperature conditions, but also predicts future temperature change trends based on past data patterns, which is crucial for developing scientific and reasonable control strategies. For example, if the temperature is found to be gradually rising but has not yet reached the warning value, the system can identify this slow but continuous increase through the trend prediction model and issue a warning signal, reminding the operator to adjust process parameters or check equipment status to prevent accidents.

[0028] Figure 4 The flowchart of the method for monitoring the safety of the calcium formate production process according to the embodiment of the present application is to input the historical temperature sequence into the pre-trained trend prediction model to obtain the predicted temperature trajectory. Figure 4As shown, according to the calcium formate production process safety monitoring method of the embodiment of the present application, the historical temperature sequence is input into the pre-trained trend prediction model to obtain the predicted temperature trajectory, including: S410, performing temperature local time series feature extraction based on one-dimensional convolutional coding on the historical temperature sequence to obtain a sequence of temperature local time series pattern feature vectors; S420, performing temperature pattern feature time series transfer on the sequence of temperature local time series pattern feature vectors to obtain a temperature full time domain pattern feature coding vector; S430, performing feature decoding on the temperature full time domain pattern feature coding vector to obtain the predicted temperature trajectory.

[0029] More specifically, step S410 performs temperature local time series feature extraction based on one-dimensional convolution coding on the historical temperature sequence to obtain a sequence of temperature local time series pattern feature vectors. It should be understood that in the calcium formate production process, the temperature changes inside the reactor are directly related to the safety and efficiency of the chemical reaction. In order to accurately predict potential risks and take timely measures, it is necessary to deeply analyze the historical temperature sequence to capture the complex patterns and trends contained therein. Traditional statistical methods are often difficult to fully reveal the subtle but important dynamic characteristics of temperature data. Therefore, in the technical solution of the present application, the historical temperature sequence is further subjected to temperature local time series feature extraction based on one-dimensional convolution coding to obtain a sequence of temperature local time series pattern feature vectors, which can effectively extract local time dependencies and patterns from the original time series, which is crucial for understanding the physical mechanism behind temperature changes. Local time series features in the temperature sequence are mined through deep learning technology. These features include not only short-term fluctuations and long-term trends, but also cover the complex interactions between temperatures in different time periods. Specifically, by scanning the entire temperature time series through a one-dimensional convolutional layer, representative local patterns at different time scales are automatically detected and converted into feature vectors, forming a continuous sequence of feature vectors of local temperature time series patterns. This enables the model to more meticulously understand and describe the temporal variations of temperature, providing strong support for subsequent trend predictions.

[0030] More specifically, step S420 involves performing a temperature pattern feature time-series transfer on the sequence of local temperature time-series pattern feature vectors to obtain a temperature full-time-domain pattern feature encoding vector. It should be understood that during the calcium formate production process, the changing trend of the reactor's internal temperature is a key factor affecting process safety and product quality. To accurately model the temperature evolution process and provide proactive early warning, relying solely on feature extraction within a local time window is often insufficient to fully reflect the dynamic evolution of the system state. That is, while local, representative temperature pattern time-series features are extracted from historical temperature data via one-dimensional convolution, these local features are limited by their perception range and are unable to reveal long-range dependencies between nodes in the entire time series and global variation patterns. In complex chemical processes, temperature changes often exhibit phased, nonlinear, or even sudden characteristics, making it easy to miss key trend signals through local observation alone. Therefore, the sequence of local temperature time-series pattern feature vectors is further subjected to a temperature pattern feature time-series transfer to obtain a temperature full-time-domain pattern feature encoding vector.

[0031] Specifically, at the micro level, an intelligent perception window is set for each local temperature temporal pattern feature vector. This window is not simply fixed in length, but adaptively adjusted based on contextual relevance to capture the most explanatory local contextual information of the temperature temporal pattern. Subsequently, deep embedding enhancement technology is used to fuse and optimize the local temperature temporal pattern features within the window, generating enhanced node encodings rich in local semantic information and forming a preliminary refined feature representation. This information is then integrated at the macro level, where the enhanced local encodings are fed into a global information aggregation network based on a Transformer architecture. This network uses a self-attention mechanism to model global information interactions and associations across all time nodes in the entire time series. This ensures that the features at each moment are not only influenced by neighboring time periods but also dynamically incorporate feedback from more distant time points, effectively capturing long-range dependencies and overall evolutionary trends in temperature variations. The final output, the global temperature temporal pattern feature encoding vector, retains the detailed characterization capabilities of the original local features while incorporating global structural information across time periods, forming a high-level representation that is both semantically rich and structurally coherent. This encoding method improves the subsequent prediction model's ability to predict the future trajectory of temperature, enabling the system to identify potential risks and respond at an earlier time point, enhancing the intelligence and robustness of the safety monitoring system, and providing strong technical support for the safe control of the calcium formate production process.

[0032] Accordingly, according to an embodiment of the present application, step S420, performing temperature pattern feature temporal transfer on the sequence of the temperature local temporal pattern feature vectors to obtain a temperature full-time domain pattern feature encoding vector, including: determining the window size of the temperature local temporal semantic enhancement perception window of each temperature local temporal pattern feature vector based on the feature distribution of each temperature local temporal pattern feature vector in the sequence of the temperature local temporal pattern feature vectors; performing local temporal pattern enhancement on each temperature local temporal pattern feature vector based on all the temperature local temporal pattern feature vectors in the temperature local temporal semantic enhancement perception window of each temperature local temporal pattern feature vector to obtain a sequence of temperature local temporal pattern feature enhancement encoding vectors; inputting the sequence of temperature local temporal pattern feature enhancement encoding vectors into the global information aggregation network of the Transformer architecture to obtain a temperature full-time domain pattern feature encoding vector.

[0033] More specifically, in an embodiment of the present application, based on the characteristic distribution of each temperature local time series pattern feature vector in the sequence of the temperature local time series pattern feature vector, the window size of the temperature local time series semantic enhancement perception window of each temperature local time series pattern feature vector is determined, including: based on the characteristic distribution of each temperature local time series pattern feature vector, calculating the neighbor distribution correlation factor of each temperature local time series pattern feature vector to obtain a sequence of neighbor distribution correlation factors; based on the neighbor distribution correlation factor to obtain a sequence of neighbor distribution correlation factors, calculating the window size of the temperature local time series semantic enhancement perception window of each temperature local time series pattern feature vector.

[0034] More specifically, based on the characteristic distribution of each temperature local time series pattern feature vector, the neighbor distribution correlation factor of each temperature local time series pattern feature vector is calculated to obtain a sequence of neighbor distribution correlation factors, which is expressed as follows:

[0035] in, and are respectively the temperature local time series pattern feature vectors in the sequence of the temperature local time series pattern feature vectors, is the logarithmic function value with base 2, is the first in the sequence of the temperature local time series pattern feature vector The neighbor distribution correlation factor of the temperature local time series pattern feature vector, is the one-norm of the vector, is an exponential function with the natural constant e as the base, for and The characteristic metric coefficient of the temperature local time series pattern between is the activity parameter, for and The characteristic metric coefficient of the temperature local time series pattern between is the preset neighborhood, is the second smoothing term coefficient, for and Neighborhood representation factor of the local temporal pattern of temperature between .

[0036] It should be understood that traditional analysis methods often ignore the subtle differences and complex correlations in temperature patterns between different time points, which is important for fully understanding the laws of temperature evolution. To accurately predict temperature trends, it is necessary to deeply explore the local time series features and their interrelationships in historical temperature data. Therefore, in the technical solution of this application, the nearest neighbor distribution correlation factor is calculated based on the characteristic distribution of each temperature local time series pattern eigenvector. Specifically, by quantifying the similarity or difference between each temperature local time series pattern eigenvector and its neighboring eigenvectors, the hidden structural information in the data can be revealed. This characteristic distribution-based method can identify which time periods have relatively stable and consistent temperature change patterns and which ones show large fluctuations or anomalies. This is crucial for distinguishing normal operating conditions from potential risk areas. For example, the speed and amplitude of temperature changes may vary at different stages of a chemical reaction. By calculating the nearest neighbor distribution correlation factor, the details of these changes can be more accurately captured, providing a basis for subsequent risk assessment. In this way, not only the temperature value at a single time point is considered, but also its interactions and dependencies with other time points. This approach allows for a more detailed description of temperature variations over time at the microscopic level, while also supporting macroscopic trend analysis. The implementation involves first determining the characteristic distribution of each local temperature time series pattern eigenvector. Based on this, the degree of difference between the eigenvector and its adjacent eigenvectors is then calculated, generating a series of neighbor distribution correlation factor values. These entropy value sequences reflect the dynamic changes in the temperature pattern over time.

[0037] More specifically, based on the neighbor distribution correlation factor to obtain a sequence of neighbor distribution correlation factors, the window size of the temperature local time series semantic enhancement perception window of each temperature local time series pattern feature vector is calculated. , expressed as:

[0038] in, is the first in the sequence of the temperature local time series pattern feature vector The neighbor distribution correlation factor of the temperature local time series pattern feature vector, is the first smoothing term coefficient, To preset the maximum window size, To round down.

[0039] It should be understood that the traditional fixed window method has difficulty adapting to the complex local dynamic characteristics of the data, while the adaptive window size can be flexibly adjusted according to the characteristic distribution of the data itself, thereby better capturing the inherent laws of temperature changes. Therefore, in order to more finely capture the subtle patterns and potential risks in these temperature changes, in the technical solution of this application, based on the neighbor distribution correlation factor to obtain a sequence of neighbor distribution correlation factors, the window size of the temperature local time series semantic enhancement perception window of each temperature local time series pattern feature vector is calculated.

[0040] Specifically, by calculating the neighbor distribution correlation factor of each temperature local time series pattern feature vector, a sequence is generated that quantifies the similarity or difference between the temperature pattern at each time point and its neighboring patterns. These neighbor distribution correlation factor values ​​are then used to guide the selection of window size, ensuring that each window contains the most representative contextual information while excluding information that may introduce noise or misleading information. For example, during certain time periods, when temperature changes are relatively stable, a smaller window size can capture sufficient detail; however, during other time periods, when there are large fluctuations or trend shifts, a larger window size is required to cover the entire pattern evolution. The window size is no longer fixed, but is adjusted based on the characteristics of the actual data, ensuring that each window maximizes the preservation of local semantic information while minimizing the impact of redundant information. This adaptive approach is particularly suitable for complex and changing industrial environments because it can automatically optimize the window size under different operating conditions, thereby improving the flexibility and adaptability of the model.

[0041] More specifically, based on the temperature local temporal semantic enhancement of each temperature local temporal pattern feature vector in the perception window, local temporal pattern enhancement is performed on each temperature local temporal pattern feature vector to obtain a sequence of temperature local temporal pattern feature enhancement encoding vectors, which can be expressed as follows:

[0042] in, and are the trainable query weight matrix and the trainable value weight matrix, for The scale, for function, for function, for and The temperature local temporal context semantic fusion representation vector, is the trainable key weight matrix, For the Temperature local temporal pattern feature enhanced encoding vector, Enhance the sequence of encoding vectors for the local temporal pattern features of temperature, are the first, second and third sequences of the temperature local temporal pattern feature enhanced encoding vector. The temperature local temporal pattern feature enhancement encoding vector.

[0043] It's understandable that while local temperature temporal pattern feature vectors provide important information about temperature variations within a specific time period, they often lack context and fail to fully reflect the dynamic patterns and long-range dependencies within the entire time series. By introducing local semantic enhancement perception windows and deeply processing and fusing the local temperature temporal pattern feature vectors within these windows, the semantic expressiveness and contextual awareness of each local temperature temporal pattern feature vector can be enhanced. For example, during different stages of a chemical reaction, temperature changes are not merely numerical fluctuations but also encompass complex physical and chemical processes. Through local semantic enhancement, the system can better capture these subtle changes and their underlying physical meaning, providing a richer information foundation for subsequent trend prediction. Specifically, the size of each perception window is first determined based on the distribution of the local temperature temporal pattern feature vectors, ensuring that the window not only covers structurally adjacent time points but also captures the most semantically relevant local context. Then, using a specific local semantic embedding enhancement mechanism, all local temperature temporal pattern feature vectors within the window are deeply processed to generate a sequence of enhanced local temperature temporal pattern feature encoding vectors. This method goes beyond traditional feature extraction methods and improves the richness and accuracy of temperature time series feature representation by comprehensively considering the interactions and correlations between local temperature time series features.

[0044] More specifically, the sequence of the temperature local temporal pattern feature enhancement encoding vector is input into the global information aggregation network of the Transformer architecture to obtain the temperature full temporal pattern feature encoding vector, which is expressed as follows:

[0045] in, It is a global information aggregation network based on the Transformer architecture. It is the temperature full time domain pattern feature encoding vector.

[0046] Understandably, relying solely on enhanced local temporal pattern features of individual temperatures to accurately model temperature trends and provide proactive early warnings is insufficient to meet monitoring requirements under complex operating conditions. Traditional feature extraction methods often overlook the contextual dependencies between nodes in the time series, which are crucial for understanding the dynamic mechanisms underlying temperature changes. For example, the rate of temperature change, fluctuation amplitude, and coupling relationship with other variables (such as pressure and agitator current) can vary significantly at different stages of a reaction. Modeling data solely based on a single time point can easily miss this semantically valuable evolutionary information. Therefore, by constructing a local semantically enhanced perception window and integrating information from multiple time points within it, we can more comprehensively characterize the feature representation at each moment, reflecting not only the current state but also the influence of preceding and following trends. This approach enhances the model's understanding of temperature trends and its generalization capabilities. The enhanced full-temporal pattern feature encoding vector not only retains the local details of the original features but also incorporates contextual information, enabling the model to more accurately identify potential risk signals, even those not readily apparent at a single time point. In addition, because this method enhances the model's ability to model long-range dependencies, it demonstrates strong robustness and adaptability in the face of sudden temperature anomalies or slowly evolving trend deviations. For example, when the temperature is detected to be gradually rising over a certain period of time but has not yet reached the alarm threshold, the system can use the enhanced features to identify possible thermal runaway risks in advance and issue an early warning based on this, providing operators with an intervention window. This semantically enhanced feature processing method essentially improves the intelligence level and response efficiency of the entire safety monitoring system, providing solid technical support for the safe control of the calcium formate production process.

[0047] More specifically, step S430 performs feature decoding on the temperature full-time-domain pattern feature encoding vector to obtain the predicted temperature trajectory. It should be understood that although high-quality temperature full-time-domain pattern feature encoding vectors have been obtained through the previous steps, the temperature full-time-domain pattern feature encoding vectors themselves cannot be directly used to guide actual operations or early warning decisions. Therefore, these abstract encoding vectors must be converted into concrete and intuitive temperature prediction results to truly realize their value. Therefore, the temperature full-time-domain pattern feature encoding vectors are further feature decoded to obtain the predicted temperature trajectory. Leveraging the powerful representation capabilities of deep learning models, the complex temperature full-time-domain pattern feature encoding vectors are mapped back to the original time series space, thereby generating an accurate prediction of future temperature changes. This requires not only that the model accurately capture key patterns and trends in the input data, but also that it possesses good generalization performance to cope with new, unseen operating conditions. For example, if a temperature trend of gradually rising within a certain time period is detected, the system can not only issue a timely warning but also pre-plan corresponding control strategies based on the predicted temperature trajectory to ensure the safety and stability of the production process.

[0048] Specifically, in one specific example of this application, a backpropagation mechanism or a specially designed decoder network is used to perform the feature decoding task. For example, a decoder structure based on a long short-term memory (LSTM) or gated recurrent unit (GRU) is used. First, the full-time temperature pattern feature encoding vector is fed into the decoder network as input. This network gradually unfolds information along the time dimension through a series of hidden layers, outputting a temperature prediction corresponding to a future moment at each step. During this process, the decoder continuously adjusts its internal state using a forward feedback mechanism and incorporates contextual information to optimize prediction accuracy at each time point. Furthermore, an attention mechanism can be introduced to enhance the decoder's performance, enabling it to dynamically focus on historical information most important for the current prediction. Ultimately, through this feature decoding process, the system outputs a continuous predicted temperature trajectory that not only reflects the expected direction of future temperature changes but also includes the potential fluctuation range and uncertainty. This approach enhances the understanding of temperature trends and prediction accuracy, enabling the monitoring system to respond to potential problems before they manifest, significantly improving production continuity and safety. It also provides reliable data support for the development of more scientific and reasonable emergency response plans, further strengthening enterprises' risk management capabilities.

[0049] Specifically, in step S500, the anomaly score and the predicted temperature trajectory are input into the decision tree model to obtain the risk level. It should be understood that a single risk indicator is often difficult to fully reflect the complex state of the system. For example, relying solely on the anomaly score may ignore the evolution path of future trends, and relying solely on the maximum temperature prediction may miss information on whether the current system is already in an unstable state. Therefore, in the technical solution of the present application, the anomaly score and the predicted temperature trajectory are input into the decision tree model to obtain the risk level. This constructs a decision logic based on two key variables, integrates the degree of anomaly of the current state and the evolution direction of future trends, and forms a multi-dimensional risk judgment mechanism with explanatory and operational properties. It can more accurately portray the risk stage of the system and improve the scientificity and pertinence of the early warning.

[0050] More specifically, in a specific example of the present application, the anomaly score and the predicted temperature trajectory are input into a decision tree model to obtain a risk level, including: extracting the predicted maximum temperature from the predicted temperature trajectory; inputting the anomaly score and the predicted maximum temperature into the decision tree model to obtain the risk level. In other words, the system extracts the predicted maximum temperature from the predicted temperature trajectory as a quantitative indicator of future thermal runaway risk; and at the same time combines the anomaly score at the current moment to measure the degree to which the system deviates from normal operating conditions. These two parameters are input together into a preset decision tree model, and judgment is made according to the set threshold combination rules, thereby outputting four distinct levels of risk: normal, trend warning, abnormal status warning, and double-confirmed high-risk signal.

[0051] More specifically, the anomaly score and the predicted maximum temperature are input into the decision tree model to determine the risk level. The following steps are performed: if the anomaly score is less than a first preset threshold and the predicted maximum temperature is less than a second preset threshold, the risk level is determined to be normal; if the anomaly score is greater than the first preset threshold and the predicted maximum temperature is less than the second preset threshold, the risk level is determined to be a trend warning; if the anomaly score is less than the first preset threshold and the predicted maximum temperature is greater than the second preset threshold, the risk level is determined to be a state abnormality warning; and if the anomaly score exceeds the first preset threshold and the predicted maximum temperature is greater than the second preset threshold, the risk level is determined to be a double-confirmed high-risk signal. This dual-factor linkage judgment approach significantly improves the accuracy of risk identification and response efficiency. On the one hand, it avoids the problem of misjudgment or omission of a single indicator, making the warning results more credible. On the other hand, by clearly defining different risk levels, it also provides a basis for subsequent alarm levels, control actions, and manual intervention. For example, when the anomaly score is low but the maximum temperature prediction value is high, the system can determine it as a status abnormality warning, prompting operators to pay attention to future trend changes without having to take drastic measures such as emergency shutdown immediately; when both the anomaly score and the temperature prediction value exceed the threshold, the highest level of double confirmation high-risk signal is triggered, prompting the system to activate multiple protection mechanisms to ensure that the safety boundary is not breached.

[0052] Specifically, in step S600, different levels of sound and light alarms are triggered based on the risk level. In this way, a multi-level, dynamically adjusted alarm system can be established so that each level of risk can receive appropriate attention and treatment. Specifically, once the risk level output by the decision tree model is determined, the system will automatically select the appropriate sound and light alarm method based on preset rules. For lower-level "normal" or "trend warning", only low-frequency prompt sounds and flashing green indicator lights are required to remind operators to pay attention and observe; for higher-level "abnormal status warning" or "double-confirmed high-risk signal", high-frequency sound alarms (such as continuous ringing) and flashing red warning lights should be immediately activated to attract the attention of on-site personnel and trigger protective measures such as emergency shutdown.

[0053] More specifically, in one specific example of this application, this task is accomplished through an automated monitoring system integrating multiple sensors and control modules. First, the system receives risk level information from a decision tree model and converts it into specific control instructions. These instructions are then sent to audio and visual alarm devices distributed throughout the production workshop. For example, a PLC (programmable logic controller) serves as the central processing unit. Based on the received risk level signal, it drives corresponding relay switches according to a predefined logic program, thereby controlling alarm lights and sound devices of different colors and frequencies. Furthermore, SCADA (supervisory control and data acquisition) systems can be incorporated to further enhance the system's intelligence, enabling remote monitoring and real-time adjustment of alarm strategies to ensure flexible and accurate responses. Ultimately, this risk-level-based audio and visual alarm triggering method significantly improves the response speed and reliability of the safety monitoring system. It not only provides differentiated warning methods for different risk levels, but also effectively coordinates human-computer interaction processes, improving emergency response efficiency. For example, in the face of a sudden high temperature risk, the system can rapidly escalate the alarm level, using strong visual and auditory stimulation to ensure that all relevant personnel are immediately notified of the dangerous situation and take necessary countermeasures to minimize losses and ensure the safety of personnel and equipment. In short, by integrating advanced data analysis technology and mature hardware facilities, the entire safety monitoring system has achieved full-process optimization from risk identification to immediate response, providing strong support for the sustained and stable chemical production.

[0054] In summary, the calcium formate production process safety monitoring method according to the embodiment of the present application is explained, which uses a distributed control system to collect original production data, and ensures the time consistency of all monitoring variables through data synchronization processing, thereby laying the foundation for subsequent analysis. Then, the synchronized data is subjected to feature engineering processing, and the features such as the short-term and medium-term change rates, moving averages and standard deviations of temperature, pressure and agitator current are extracted to construct the reactor state features and capture the dynamic operation information of the system. In order to predict potential risks, a trend prediction model based on deep learning is adopted to convert historical temperature sequences into future temperature trajectories, providing a forward-looking prediction of temperature trends. At the same time, the current system health status is evaluated in combination with the state recognition model, and the risk level is determined according to the anomaly score and the predicted temperature maximum value through the decision tree model to achieve accurate graded warning. The entire process effectively solves the problems of delayed response and high false alarm rate of traditional safety monitoring means through real-time data processing, feature extraction, trend prediction and intelligent decision support, thereby improving production safety.

[0055] Furthermore, a calcium formate production process safety monitoring system is also provided.

[0056] Figure 5FIG. 1 is a block diagram of a safety monitoring system for the calcium formate production process according to an embodiment of the present application. Figure 5 As shown, the calcium formate production process safety monitoring system 500 according to the embodiment of the present application includes: an engineering real-time data acquisition module 510, which is used to obtain the original DCS data collected by the distributed control system; a BIM component information acquisition module 520, which is used to synchronize the original DCS data to obtain a time series of synchronized data frames, wherein the synchronized data frames include temperature, pressure and agitator current; a casting state dynamic digital twin construction module 530, which is used to perform feature engineering and state vector construction on the time series of the synchronized data frames to obtain a reactor state feature vector and a historical temperature sequence; a temperature curve anomaly detection module 540, which is used to input the reactor state feature vector and the historical temperature sequence into a parallel monitoring component including a state recognition model and a trend prediction model to obtain an anomaly score and a predicted temperature trajectory; a surface defect detection module 550, which is used to input the anomaly score and the predicted temperature trajectory into a decision tree model to obtain a risk level; and a result display module 560, which is used to trigger different levels of sound and light alarms based on the risk level.

[0057] As described above, the calcium formate production process safety monitoring system 500 according to the embodiment of the present application can be implemented in various wireless terminals, such as a server equipped with a calcium formate production process safety monitoring algorithm. In one possible implementation, the calcium formate production process safety monitoring system 500 according to the embodiment of the present application can be integrated into the wireless terminal as a software module and / or hardware module. For example, the calcium formate production process safety monitoring system 500 can be a software module in the operating system of the wireless terminal, or can be an application developed for the wireless terminal; of course, the calcium formate production process safety monitoring system 500 can also be one of the many hardware modules of the wireless terminal.

[0058] While various embodiments of the present disclosure have been described above, the above descriptions are illustrative, non-exhaustive, and not intended to be limiting of the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is selected to best explain the principles of the embodiments, their practical applications, or improvements to existing technologies, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for safety monitoring of a calcium formate production process, characterized in that: include: Obtain original DCS data collected by the distributed control system; performing data synchronization on the raw DCS data to obtain a time sequence of synchronized data frames, wherein the synchronized data frames include temperature, pressure, and agitator current; Performing feature engineering and state vector construction on the time series of the synchronized data frame to obtain a reactor state feature vector and a historical temperature series; Inputting the reactor state feature vector and the historical temperature sequence into a parallel monitoring component including a state recognition model and a trend prediction model to obtain an anomaly score and a predicted temperature trajectory; Inputting the anomaly score and the predicted temperature trajectory into a decision tree model to obtain a risk level; Based on the risk level, different levels of audible and visual alarms are triggered.

2. The calcium formate production process safety monitoring method according to claim 1, wherein The method further comprises: performing a data synchronization on the original DCS data to obtain a time sequence of a synchronized data frame, comprising: performing a synchronization operation on the original DCS data in response to the operation of a master clock, wherein the synchronization operation comprises: Look back at the previous time window and find the value with the latest timestamp for each variable to be monitored; In response to the missing value of a variable, the value of the variable at the previous time point is directly used to fill it.

3. The calcium formate production process safety monitoring method according to claim 1, wherein Performing feature engineering and state vector construction on the time series of the synchronized data frame to obtain a reactor state feature vector and a historical temperature series includes: Extracting the historical temperature sequence from the time series of the synchronized data frame; Extracting the latest synchronization data frame from the time series of synchronization data frames; Performing derivative feature calculation and statistical feature calculation based on the time series of the synchronized data frame to obtain derivative features and statistical features; The derived features and the statistical features are added to the latest synchronous data frame to obtain the reactor state feature vector.

4. The calcium formate production process safety monitoring method according to claim 3, wherein The derived features include the short-term rate of change of temperature, the medium-term rate of change of temperature, the short-term rate of change of pressure, the medium-term rate of change of pressure, the short-term rate of change of agitator current and the medium-term rate of change of agitator current.

5. The calcium formate production process safety monitoring method according to claim 3, wherein: The statistical features include a temperature moving average, a temperature moving standard deviation, a pressure moving average, a pressure moving standard deviation, an agitator current moving average, and an agitator current moving standard deviation.

6. The calcium formate production process safety monitoring method according to claim 1, wherein The reactor state feature vector and the historical temperature sequence are input into a parallel monitoring component including a state recognition model and a trend prediction model to obtain an anomaly score and a predicted temperature trajectory, including: Inputting the reactor state feature vector into a pre-trained state recognition model to obtain the abnormality score; The historical temperature sequence is input into a pre-trained trend prediction model to obtain the predicted temperature trajectory.

7. The calcium formate production process safety monitoring method according to claim 6, wherein: Inputting the historical temperature sequence into a pre-trained trend prediction model to obtain the predicted temperature trajectory includes: Performing temperature local time series feature extraction based on one-dimensional convolution coding on the historical temperature sequence to obtain a sequence of temperature local time series pattern feature vectors; Performing temperature pattern feature time series transfer on the sequence of the temperature local time series pattern feature vectors to obtain a temperature full time domain pattern feature coding vector; Feature decoding is performed on the temperature full-time-domain pattern feature coding vector to obtain the predicted temperature trajectory.

8. The calcium formate production process safety monitoring method according to claim 1, wherein: The anomaly score and predicted temperature trajectory are input into a decision tree model to obtain a risk level, including: extracting a predicted temperature maximum value from the predicted temperature trajectory; The anomaly score and the temperature prediction maximum value are input into the decision tree model to obtain the risk level.

9. The calcium formate production process safety monitoring method according to claim 8, wherein: Inputting the anomaly score and the predicted maximum temperature into the decision tree model to obtain the risk level includes: If the abnormality score is less than a first preset threshold and the predicted maximum temperature is less than a second preset threshold, determining that the risk level is normal; If the anomaly score is greater than a first preset threshold and the predicted maximum temperature is less than a second preset threshold, determining the risk level as a trend warning; If the abnormality score is less than a first preset threshold and the temperature prediction maximum value is greater than a second preset threshold, the risk level is determined to be a state abnormality warning; If the abnormality score exceeds a first preset threshold and the temperature prediction maximum value exceeds a second preset threshold, the risk level is determined to be a double-confirmed high-risk signal.

10. A calcium formate production process safety monitoring system, characterized in that: include: Original DCS data acquisition module, used to obtain original DCS data collected by the distributed control system; a data synchronization module, configured to synchronize the original DCS data to obtain a time sequence of synchronized data frames, wherein the synchronized data frames include temperature, pressure, and agitator current; A feature engineering module, configured to perform feature engineering and state vector construction on the time series of the synchronized data frame to obtain a reactor state feature vector and a historical temperature series; A parallel monitoring module, configured to input the reactor state feature vector and the historical temperature sequence into a parallel monitoring component comprising a state recognition model and a trend prediction model to obtain an anomaly score and a predicted temperature trajectory; a risk level decision module, configured to input the anomaly score and the predicted temperature trajectory into a decision tree model to obtain a risk level; The sound and light alarm triggering module is used to trigger different levels of sound and light alarms based on the risk level.

Images