Method for evaluating safety state of chemical process based on multi-dimensional process parameters

Abstract

The application discloses a chemical process safety state evaluation method based on multi-dimensional process parameters and relates to the technical field of chemical process safety evaluation. The method comprises the following steps: acquiring multi-dimensional process parameter time series data of a plurality of monitoring nodes in a chemical process; constructing a dynamic recursive safety state evaluation model, which comprises a parameter coupling memory unit and a state transition probability deduction unit; performing time domain difference recursion and space domain cross coupling calculation on the data through the parameter coupling memory unit to generate a process parameter coupling state vector; inputting the coupling state vector into the state transition probability deduction unit, which iteratively calculates the probability distribution of the deviation of the current process state from the safety domain based on the Markov blanket dependence structure determined based on historical accident case statistics learning; and determining and outputting the safety state grade according to the cumulative probability value of the probability distribution exceeding the safety threshold.

Description

Technical Field

[0001] This invention relates to the field of chemical process safety assessment technology, specifically a method for assessing the safety status of chemical processes based on multi-dimensional process parameters. Background Technology

[0002] Safety status assessment of chemical processes relies on the continuous monitoring and analysis of multi-dimensional process parameters such as temperature, pressure, flow rate, and liquid level. Existing assessment methods often employ fixed threshold alarms or trend analysis of single parameters, triggering an alarm when the measured value exceeds a preset limit. These methods treat each process parameter as an independent variable, neglecting the dynamic coupling and interactive effects between parameters that are prevalent in chemical processes. Actual operational data shows that correlated fluctuations in multiple parameters are often precursors to accidents, but threshold discrimination mechanisms for single parameters struggle to capture these cross-parameter and cross-monitoring node coupling evolution characteristics, leading to delayed safety status assessments or missed detections.

[0003] Furthermore, existing probability-based evaluation schemes typically employ static Bayesian networks or fully connected dynamic Bayesian networks for state inference. Static networks cannot describe the state transition patterns over time, while fully connected dynamic networks, when faced with high-dimensional process parameters, exhibit overly complex inter-node dependencies, resulting in exponentially increasing conditional probability tables, enormous computational costs, and a susceptibility to overfitting, making them difficult to deploy in real-time monitoring scenarios. These schemes often neglect the selection of key parameter combinations that substantially contribute to the safe state when learning the network structure, leading to redundant model structures and low inference efficiency. Existing technologies have failed to effectively address two key issues: how to automatically extract spatiotemporal correlation features from multi-node, multi-parameter coupled time-series data, and how to construct a model with a concise structure that accurately characterizes the conditional independence relationships between parameters and enables dynamic probabilistic inference. Summary of the Invention

[0004] This invention aims to provide a method for assessing the safety status of chemical processes based on multi-dimensional process parameters. By constructing a dynamic recursive assessment model that includes parameter coupling memory units and state transition probability deduction units, it can extract time-space joint coupling features of time-series data from multiple monitoring nodes and multiple parameter types in chemical processes. Based on the Markov blanket dependency structure obtained from statistical learning of historical accident cases, it can perform iterative deduction of state transition probabilities, thereby accurately and in real time determining the process safety status level.

[0005] To achieve the above objectives, this invention provides the following technical solution: This invention relates to chemical process safety status assessment technology, specifically proposing a method for chemical process safety status assessment based on multi-dimensional process parameters. This method, through deep coupling analysis of multiple monitoring nodes and various types of process parameters in the chemical production process, constructs an assessment model with dynamic recursive capabilities, enabling real-time quantitative determination of the process safety status.

[0006] As a technical solution of the present invention, multidimensional process parameter time series data of multiple monitoring nodes in a chemical process are obtained. Preferably, temperature, pressure, flow, and liquid level parameters of each monitoring node are collected synchronously according to a preset sampling period. Missing value imputation is performed on the collected process parameters, and this imputation operation is based on linear interpolation of the historical trend of the same parameter type at the same monitoring node. The imputed multidimensional process parameters of each monitoring node are arranged in chronological order to form a multidimensional process parameter time series data matrix indexed by the sampling time. The rows of the matrix correspond to different sampling times, and the columns correspond to different combinations of monitoring nodes and parameter types. Preferably, the interpolation step size of the linear interpolation in the missing value imputation operation does not exceed twice the preset sampling period. When the number of consecutive missing data points exceeds a preset threshold, imputation is abandoned and the monitoring node is marked as a data failure state, thereby ensuring the integrity and reliability of the basic data.

[0007] As a technical solution of this invention, a dynamic recursive safety status assessment model is constructed. This model includes a parameter-coupled memory unit and a state transition probability inference unit. During model construction, multiple accident case segments and multiple normal case segments are extracted from historical operating data of the chemical process. Each case segment contains a multidimensional process parameter time series with multiple consecutive sampling times and corresponding safety status labels. The multidimensional process parameter time series in the case segments are grouped according to monitoring nodes, and the mutual information values ​​between parameter types within each monitoring node and the mutual information values ​​of the same parameter type between different monitoring nodes are calculated. Preferably, the mutual information value is calculated using a non-parametric method based on kernel density estimation, and the time window length is set to three times the main disturbance period of the chemical process to accurately capture the nonlinear correlation between parameters. Based on the mutual information values, key parameter combinations that contribute more than a preset contribution threshold to the safety status label are selected. These key parameter combinations are used to define the parent node set of each node in the Markov blanket dependency structure. This parent node set is used to limit the conditional probability table structure of the state transition probability inference unit. A Bayesian structure learning algorithm is used to perform directional edge search on the Markov blanket dependency structure to determine the directed dependencies between each node, completing the model initialization. In the process of directional edge search, candidate undirected edges are formed by calculating the mutual information values ​​of nodes. Two directional directions are tried respectively, and the optimal direction is selected based on the Bayesian information criterion. Gradually, a Markov blanket-dependent structure that accurately reflects the interaction mechanism of chemical process parameters is constructed.

[0008] As a technical solution of this invention, a parameter coupling memory unit performs time-domain difference recursive and spatial-domain cross-coupling calculations on multidimensional process parameter time series data to generate a time-stamped process parameter coupling state vector. For any current sampling time, the parameter coupling memory unit reads the historical coupling state vector output at the previous sampling time and the multidimensional process parameter time series data at the current sampling time. For each monitoring node, a first-order time difference value is calculated according to each parameter type within it. This value is the difference between the parameter value at the current sampling time and the parameter value at the previous sampling time. The first-order time difference values ​​of all parameter types within the node are weighted and summed to obtain the local time-domain variation characteristics of the monitoring node. The weight of each parameter type in the weighted summation is determined by the frequency of occurrence of that parameter type in the Markov blanket dependency structure. Simultaneously, the first-order spatial difference value of any two different monitoring nodes on the same parameter type is calculated to reflect the process parameter gradient between different spatial locations. The local temporal variation features of all monitoring nodes and the first-order spatial difference values ​​of all different monitoring node pairs are concatenated into vectors. This concatenation is then recursively superimposed with the historical coupling state vector output at the previous sampling time to generate the process parameter coupling state vector at the current sampling time. Preferably, the weights of the weighted summation are normalized before vector concatenation, ensuring that the sum of all weights within the same monitoring node is 1, thus guaranteeing the scale consistency of feature extraction. By introducing a decay mechanism that applies a forgetting coefficient to the historical coupling state vector, the model maintains sufficient sensitivity to new data changes while preserving long-term memory, achieving a continuous characterization of the dynamic changes in the process state.

[0009] As a technical solution of this invention, the process parameter coupled state vector is input into a state transition probability inference unit. This unit, based on the Markov blanket dependency structure determined by statistical learning results of historical accident cases in chemical processes, iteratively calculates the probability distribution of the current process state deviating from a preset safety domain. Specifically, the process parameter coupled state vector at the current moment is used as the observation evidence input. The state transition probability inference unit determines the set of parent nodes and child nodes directly related to the current process state node according to the Markov blanket dependency structure. The posterior probability of each hidden state node under the current observation evidence is queried from a pre-stored conditional probability table. This conditional probability table is obtained by expectation-maximization learning of the process parameter coupled state vectors at all sampling moments in the historical operation data of the chemical process. Preferably, during the query process, each component of the coupled state vector is discretized into multiple level intervals according to its numerical range, corresponding to discrete observation symbols, and the corresponding probability distribution is located and extracted from the conditional probability table. If there is no perfectly matching parent node combination, the posterior probability is generated by interpolation from the probability distribution of adjacent parent node combinations using the Laplace smoothing method, which enhances the robustness of the model. By using a Bayesian recursive formula to fuse the hidden state probability distribution at the previous sampling time with the observation likelihood probability at the current time, the hidden state probability distribution at the current sampling time is iteratively updated. Each component of this distribution corresponds to the probability value of a discrete safety sub-interval of the process. The probability values ​​of the corresponding hazardous safety sub-intervals are accumulated to obtain the probability distribution of the process state deviating from the preset safety domain at the current time. This process integrates time-varying, coupled, and uncertain reasoning, realizing dynamic perception of the safety situation.

[0010] As a technical solution of this invention, the safety status level of a chemical process is determined and output based on the cumulative probability value exceeding the safety threshold in the probability distribution. All probability values ​​corresponding to the hazardous safety sub-intervals are extracted from the probability distribution output by the state transition probability deduction unit and accumulated to obtain a cumulative hazard probability value. This cumulative hazard probability value is compared with pre-set first-level threshold, second-level threshold, and third-level threshold, with each level threshold increasing sequentially. When the cumulative hazard probability value is lower than the first-level threshold, it is determined to be a low-risk level; between the first and second-level thresholds, it is determined to be a medium-risk level; between the second and third-level thresholds, it is determined to be a high-risk level; and higher than the third-level threshold, it is determined to be a critical level. The determined safety status level is output as a discrete level identifier to the human-machine interface of the chemical process safety control system, providing operators with intuitive and quantitative decision support.

[0011] As a preferred embodiment of the present invention, after the dynamic recursive process of the safety status assessment model is completed at each sampling time, the coupled state vector of the process parameters output at the current time and the hidden state probability distribution are simultaneously stored in the circular buffer as input data for the parameter coupling memory unit at the next sampling time. When the hidden state probability distributions of multiple consecutive sampling times stored in the circular buffer all indicate that the safety status level is lower than the low-risk level, it indicates that the device is in a steady-state operation phase, triggering the online parameter update process of the safety status assessment model. This process uses the data of the most recent consecutive sampling times stored in the circular buffer to re-execute expectation-maximization learning and update the conditional probability table in the state transition probability inference unit, thereby enabling the model to adapt to the slow drift of device characteristics and operating conditions and continuously maintain the accuracy of the assessment.

[0012] The technical effects and advantages provided by the present invention in the above technical solution are as follows: By setting up parameter coupling memory units, the first-order time difference values ​​of various process parameters within each monitoring node at the current moment are calculated. The difference values ​​of all parameters within the same node are then weighted and summed according to their importance in the Markov blanket dependency structure corresponding to that unit, generating the local temporal variation characteristics of that node. Simultaneously, the first-order spatial difference values ​​of the same parameter type between different monitoring nodes are calculated to capture the coupling deviation of parameters in the spatial dimension. The local temporal variation characteristics of the monitoring nodes and the spatial difference characteristics of node pairs are concatenated and recursively superimposed with the historical coupling state vector attenuated by the forgetting coefficient from the previous moment to generate the process parameter coupling state vector at the current moment. This approach integrates the independent temporal variations of parameters, the spatial correlation variations across nodes, and the dynamic memory of states, enabling the extracted process parameter coupling state vector to dynamically and three-dimensionally represent the instantaneous evolution trend and spatial coupling relationship of the chemical process. This overcomes the shortcomings of traditional methods that isolate parameter analysis or only consider temporal variations while ignoring spatial correlations, providing a more discriminative state representation for subsequent probabilistic inferences and improving the sensitivity of assessments to complex process fluctuations.

[0013] By constructing a state transition probability inference unit based on a Markov blanket dependency structure, and utilizing accident case fragments from historical chemical process operation data, the safety state contribution of parameter types within and between each monitoring node is calculated. Key parameter combinations are selected to define the parent node set for each node, thus limiting the size of the conditional probability table. This structural learning strategy identifies simplified dependencies directly related to the safety state, discarding irrelevant or redundant parameter connections. After receiving the process parameter coupled state vector as observational evidence, the state transition probability inference unit queries and calculates the posterior probability of nodes only within the local dependency structure constrained by the Markov blanket. It then iteratively updates the hidden state probability distribution at the current moment by fusing the hidden state probability distribution from the previous moment according to the Bayesian recursive formula. By strictly limiting the probability inference to the key causal structures statistically derived from accident cases, computational complexity is significantly reduced. Simultaneously, it focuses on the process coupling patterns that truly lead to state transitions, avoiding the dimensionality curse and computational resource waste problems of fully connected networks with high-dimensional parameters. This ensures that dynamic safety level assessment achieves both computational efficiency and inference accuracy. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0015] Figure 1 This is a flowchart of a chemical process safety status assessment method based on multi-dimensional process parameters; Figure 2 This is a flowchart illustrating the process parameter coupling state vector generation process parameter coupling in the parameter coupling memory unit. Figure 3 This is a flowchart of the process state deviation probability iterative deduction based on the Markov blanket-dependent structure; Figure 4 This is a flowchart for determining the safety status level based on the cumulative hazard probability value. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] See Figure 1This invention provides a method for assessing the safety status of a chemical process based on multidimensional process parameters, comprising: acquiring time-series data of multidimensional process parameters from multiple monitoring nodes in the chemical process; constructing a dynamic recursive safety status assessment model, which includes a parameter coupling memory unit and a state transition probability deduction unit; performing time-domain difference recursion and spatial-domain cross-coupling calculations on the multidimensional process parameter time-series data through the parameter coupling memory unit to generate a time-stamped process parameter coupled state vector; inputting the process parameter coupled state vector into the state transition probability deduction unit, which iteratively calculates the probability distribution of the process state deviating from a preset safety domain at the current moment based on the Markov blanket dependency structure determined by statistical learning results of historical accident cases in the chemical process; and determining and outputting the safety status level of the chemical process based on the cumulative probability value exceeding the safety threshold in the probability distribution.

[0018] Example 1: In practice, the process of obtaining multidimensional process parameter time series data of multiple monitoring nodes in the chemical process is as follows.

[0019] Temperature, pressure, flow rate, and liquid level parameters of each monitoring node are collected synchronously according to a preset sampling period. The preset sampling period is a fixed time interval determined by the data sampling frequency of the chemical process control system, ensuring that the process parameters of all monitoring nodes are recorded synchronously under a unified time reference. Each monitoring node generates a set of process parameter measurements at each sampling moment, and each set of process parameter measurements includes temperature, pressure, flow rate, and liquid level measurements.

[0020] For each type of process parameter collected from each monitoring node, a missing value interpolation operation is performed. For any monitoring node, when the value of a certain type of process parameter collected at any sampling time is identified as missing, linear interpolation is performed using the values ​​of the preceding and following valid sampling times of the same process parameter type in that monitoring node that are temporally adjacent to the missing sampling time. The linear interpolation calculation process uses the following formula: ; in, This represents the process parameter value obtained from the interpolation of the missing sampling time. This refers to the value of the same process parameter type at the same monitoring node, which is the closest valid sampling time to the missing sampling time, prior to the missing sampling time. This represents the value of the same process parameter type at the same monitoring node, which is the closest valid sampling time to the missing sampling time after the missing sampling time. Indicates the sampling time corresponding to the missing value; express The corresponding sampling time; express The corresponding sampling time. The prerequisite for linear interpolation is... ,in Indicates the preset sampling period. When Greater than When the missing sampling time is not specified, linear interpolation is abandoned, and the process parameter values ​​at the missing sampling time are retained as missing values.

[0021] During the missing value imputation operation, the number of consecutive missing data points for each process parameter type at each monitoring node is monitored. When the number of consecutive missing data points for the same process parameter type at the same monitoring node exceeds a preset threshold, the missing value imputation operation for all process parameter types at that monitoring node is abandoned, and the monitoring node is marked as a data failure state. The preset threshold is an integer value predetermined based on the data integrity requirements of chemical processes. Monitoring nodes marked as data failure states do not participate in subsequent time-domain differential recursive calculations and spatial cross-coupling calculations. The temperature, pressure, flow, and level parameter values ​​corresponding to monitoring nodes marked as data failure states are filled with zero values ​​and an invalid flag is attached. The invalid flag is used to indicate that the corresponding value does not represent the actual process measurement value. The invalid flag adopts a Boolean type; when the invalid flag is true, it indicates that the corresponding value is invalid.

[0022] After completing the missing value interpolation and failure status marking for all monitoring nodes, the interpolated multidimensional process parameters of each monitoring node are arranged in chronological order of sampling time, forming a multidimensional process parameter time series data matrix indexed by sampling time. The rows of the multidimensional process parameter time series data matrix correspond to different sampling times, and the columns correspond to different combinations of monitoring nodes and process parameter types. If the chemical process includes... There are 10 monitoring nodes, each collecting four types of process parameters: temperature, pressure, flow rate, and liquid level. Therefore, the number of columns in the 3D process parameter time series data matrix is ​​equal to... Each column is uniquely identified by both the monitoring node identifier and the process parameter type name. The number of rows in the multidimensional process parameter time series data matrix is ​​determined by the total number of sampling times. Each element in the multidimensional process parameter time series data matrix stores the value or zero value obtained after interpolation of the corresponding sampling time, the corresponding monitoring node, and the corresponding process parameter type, and carries a corresponding invalidity flag for identification and elimination of invalid data in subsequent calculations.

[0023] Example 2: In practical implementation, the process of constructing a dynamic recursive security state assessment model is as follows. The security state assessment model includes a parameter-coupled memory unit and a state transition probability deduction unit. The core of the construction process is to determine the Markov blanket dependency structure on which the state transition probability deduction unit depends.

[0024] Multiple accident case segments and multiple normal case segments are extracted from historical operating data of chemical processes. Historical operating data is stored in a historical chemical process database. Each record includes the sampling time, monitoring node identifier, process parameter type, process parameter value, and a safety status label for the corresponding sampling time. An accident case segment refers to all historical operating data records within a sampling time interval where a continuous hazardous or accidental state occurs. A normal case segment refers to all historical operating data records within a sampling time interval where a continuous normal or safe state occurs. Each accident case segment and each normal case segment contains a multi-dimensional process parameter time series from multiple consecutive sampling times and a safety status label corresponding to each sampling time. The safety status labels are represented in discrete form and include at least two categories: normal state and hazardous state.

[0025] The time series of multidimensional process parameters in each case segment are grouped according to monitoring nodes. Specifically, for chemical processes, the time series of multidimensional process parameters are grouped according to monitoring nodes. Each monitoring node divides the multidimensional process parameter time series into several monitoring nodes. The monitoring nodes are grouped into groups, and each group contains time series data for temperature, pressure, flow rate, and level parameters at multiple consecutive sampling times. Based on this, the mutual information values ​​between different process parameter types within each monitoring node, as well as the mutual information values ​​for the same process parameter type between different monitoring nodes, are calculated. The mutual information values ​​are calculated using a nonparametric method based on kernel density estimation. The bandwidth parameter for kernel density estimation is determined through cross-validation.

[0026] The time window length used for calculating the mutual information value was set to three times the main disturbance period of the chemical process. The main disturbance period was determined through spectral analysis of historical operating data. The spectral analysis method was as follows: a representative temperature parameter time series of a monitoring node was selected from the historical operating data as the analysis object. A fast Fourier transform was performed on the temperature parameter time series to obtain the power spectral density curve. The frequency component with the largest amplitude was identified on the power spectral density curve and denoted as... The main disturbance period of the chemical process Through The calculated time window length for mutual information value calculation. Set as .

[0027] Based on all calculated mutual information values, key parameter combinations whose contribution to the safety status label exceeds a preset contribution threshold are selected. The safety status label contribution refers to the mutual information value between a parameter combination and the safety status label. The preset contribution threshold is a pre-defined positive real number, determined based on the statistical correlation strength distribution between abnormal parameter fluctuations and safety status changes in historical accident cases. All parameter combinations with mutual information values ​​greater than the preset contribution threshold are marked as key parameter combinations. Each parameter in a key parameter combination corresponds to a process parameter node in the Markov blanket dependency structure.

[0028] The set of parent nodes for each node in the Markov blanket dependency structure is defined using key parameter combinations. For any node in the Markov blanket dependency structure, if the process parameter represented by the node appears in a certain key parameter combination, and the other parameters in that key parameter combination are considered direct cause parameters of the node, then the nodes corresponding to the other parameters in the key parameter combination are added to the set of parent nodes for that node. The set of parent nodes is used to define the conditional probability table structure of the state transition probability deduction unit. Each row in the conditional probability table corresponds to a combination of values ​​for all nodes in the set of parent nodes.

[0029] A Bayesian structure learning algorithm is used to perform directional edge search on the Markov blanket dependency structure to determine the directed dependencies between nodes in the state transition probability inference unit.

[0030] The specific steps of the directed edge search are as follows: Initialize an empty set of edges, treating each process parameter node and safety status label node as an independent node in the Markov blanket dependency structure. The process parameter node corresponds to each process parameter type for each monitoring node, and the safety status label node represents the safety status.

[0031] Iterate through all possible node pairs and calculate the mutual information value between each pair. The calculation method for mutual information value is the same as described above, using a nonparametric method based on kernel density estimation, and the time window length is also set to three times the main disturbance period of the chemical process. Node pairs with mutual information values ​​higher than a preset mutual information threshold are added to the candidate edge set as candidate undirected edges. The preset mutual information threshold is based on the significance level. The chi-square test critical value for mutual information is determined.

[0032] For each candidate undirected edge in the candidate edge set, try two possible orientation directions and calculate the Bayesian information criterion score for the entire dependency structure under each orientation direction. The formula for calculating the Bayesian information criterion score is: ; in, This represents the current attempt to construct a Markov blanket-dependent structured directed graph with oriented edges; This represents a dataset composed of all extracted accident case fragments and normal case fragments; Indicates that in a given directed graph Dataset under the condition The likelihood probability, Through directed graphs The conditional probability table is obtained by maximum likelihood estimation under the constraint of the parent node; Represents a directed graph The total number of independent parameters in the middle. The value of is equal to that of the directed graph. The sum of the number of free parameters contained in the conditional probability distribution of each node in its parent node set; Represents the dataset The total number of sampling times included.

[0033] For a candidate undirected edge, compare the Bayesian information criterion scores under the two orientations. Select the orientation that maximizes the increase in the Bayesian information criterion score as the final orientation of this candidate undirected edge, and add this directed edge to the edge set of the Markov blanket dependency structure. If neither orientation increases the Bayesian information criterion score compared to not adding this directed edge, then this candidate undirected edge remains undirected and is not added to the edge set.

[0034] Repeat the directed edge search process described above. In each iteration, select one undirected candidate edge from the candidate edge set whose direction has not yet been determined, process it, and remove the processed undirected candidate edge from the candidate edge set. Repeat the iteration until the candidate edge set is empty, or until adding directed edges to any remaining undirected candidate edge in the candidate edge set fails to increase the Bayesian information criterion score. The final output is the directed Markov blanket dependency structure, which is a directed acyclic graph.

[0035] After completing the above process, the directed dependencies between nodes in the state transition probability deduction unit are determined by the final Markov blanket dependency structure, and the conditional probability table structure of the state transition probability deduction unit is determined accordingly, thus completing the initialization of the dynamic recursive security state assessment model.

[0036] Example 3: In specific implementation, please refer to Figure 2 The following is the implementation method for generating a time-stamped coupled state vector of process parameters by performing time-domain difference recursion and spatial cross-coupling calculations on multi-dimensional process parameter time series data through parameter coupling memory units.

[0037] For any given sampling moment, the parameter coupling memory unit reads the historical coupling state vector output from the previous sampling moment from the circular buffer, and reads all process parameter values ​​for the corresponding row of the current sampling moment from the multidimensional process parameter time series data matrix. The multidimensional process parameter time series data for the current sampling moment includes the temperature, pressure, flow rate, and liquid level parameters for each monitoring node at that sampling moment. The values ​​corresponding to monitoring nodes marked as having failed data are read as zero values ​​and marked as invalid.

[0038] For each monitoring node at the current sampling time, the first-order time difference value is calculated according to the various process parameter types within the monitoring node. The first-order time difference value is the difference between the parameter value at the current sampling time and the parameter value of the same monitoring node and the same process parameter type at the previous sampling time. When the corresponding parameter value at the previous sampling time exists and is valid, the first-order time difference value is calculated directly; when the corresponding parameter value at the previous sampling time is zero (marked as a data failure status), the first-order time difference value is set to zero.

[0039] The first-order time difference values ​​of temperature, pressure, flow rate, and level parameters within the same monitoring node are weighted and summed to obtain the local time-domain variation characteristics of that monitoring node. The weight of each process parameter type in the weighted summation is determined by the frequency of occurrence of that process parameter type in the Markov blanket dependency structure. The frequency of occurrence of a process parameter type in the Markov blanket dependency structure refers to the number of times that process parameter type is represented by its starting or ending node in the directed edge set of the Markov blanket dependency structure. For the same monitoring node, initial weights for the four process parameter types are generated based on the ratio of their frequencies in the Markov blanket dependency structure. These initial weights are then normalized so that the sum of all weights within the same monitoring node equals 1. The normalization process is as follows: divide the initial weight of each type of process parameter by the sum of the initial weights of the four types of process parameters of the monitoring node, and the result is the final weight used for weighted summation.

[0040] Calculate the first-order spatial difference between any two different monitoring nodes. For any pair of monitoring nodes, including node A and node B, calculate the differences between node A and node B in terms of temperature, pressure, flow rate, and liquid level parameters. These differences are all first-order spatial difference values. The first-order spatial difference value for each pair of monitoring nodes is taken as the arithmetic mean of the first-order spatial difference values ​​for all four types of process parameters.

[0041] The local time-domain variation characteristics of each monitoring node are arranged in ascending order according to the monitoring node number to form the first feature sub-vector. The monitoring node numbers are pre-assigned from 1 to... The integer sequence number. Let the monitoring node number be... Monitoring nodes The local time-domain variation characteristics are denoted as Then the first feature vector is .

[0042] The average of the first-order spatial differences of all different monitoring node pairs is arranged in ascending order of node pair numbers to form the second feature vector. The node pair numbers are arranged using unique binary tuples, first fixing the smaller monitoring node number and then fixing the larger monitoring node number and then increasing the order. Let the node pairs... The average value of the corresponding first-order spatial difference is Then the second feature vector is all A sequence arranged in order.

[0043] The first and second feature vectors are concatenated end-to-end to form a combined feature vector. The dimension of the combined feature vector is equal to the total number of monitoring nodes. Total number of monitoring nodes sum.

[0044] A forgetting coefficient is applied to the historical coupled state vector output at the previous sampling time to obtain the decayed historical coupled state vector. The forgetting coefficient is a pre-defined real number greater than zero and less than one. The setting of the forgetting coefficient is based on the inertial variation characteristics of the process parameters of the chemical process: the closer the forgetting coefficient is to 1, the higher the retention of the historical coupled state vector in the recursive superposition; the closer the forgetting coefficient is to 0, the faster the historical coupled state vector decays. In this implementation, the forgetting coefficient is set to 0.85. This setting value is determined by analyzing the autocorrelation coefficient of the coupled state vectors at adjacent sampling times under normal operating conditions of the chemical process, so that the forgetting coefficient is equal to the square of the average first-order autocorrelation coefficient.

[0045] The process parameter coupling state vector at the current sampling moment is obtained by element-wise addition of the combined feature vector and the attenuated historical coupling state vector. The recursive superposition formula for the process parameter coupling state vector is expressed as: ; in, Indicates the current sampling time The process parameter coupling state vector; This represents the forgetting coefficient, with a value range of [value missing]. In this implementation, the value is taken as 0.85; Indicates the previous sampling time Output historical coupling state vector; Indicates the current sampling time The generated combined feature vector. At the initial running time of the system, there is no historical coupling state vector output from the previous sampling time. Take a vector containing all zeros.

[0046] After completing the above calculations, the coupled state vector of the process parameters at the current sampling time is... It is assigned a corresponding timestamp, which is the precise time value of the current sampling moment.

[0047] Example 4: In specific implementation, please refer to Figure 3 The process of inputting the process parameters coupled with the state vector into the state transition probability deduction unit and iteratively calculating the probability distribution of the process state deviating from the preset safety domain at the current moment based on the Markov blanket dependency structure is as follows.

[0048] After receiving the coupled state vector of process parameters at the current sampling moment, the state transition probability inference unit inputs this coupled state vector as observation evidence into the inference calculation process. Observation evidence is the real-time information upon which the state transition probability inference unit bases its hidden state probability updates.

[0049] The state transition probability inference unit identifies the set of parent nodes and the set of child nodes directly related to the current process state node based on the established Markov blanket dependency structure. The current process state node is a hidden state node representing the process safety state in the Markov blanket dependency structure. The set of parent nodes consists of nodes connected by directed edges pointing to the current process state node, and the set of child nodes consists of nodes connected by directed edges originating from the current process state node. Each node in the set of parent nodes corresponds to a specific component in the process parameter coupled state vector or historical hidden state information.

[0050] The posterior probability of each hidden state node under the current observational evidence is retrieved from a pre-stored conditional probability table. This pre-stored conditional probability table is obtained by expectation-maximization learning of the process parameter coupled-state vectors at all sampling times in the historical operating data of the chemical process. The expectation-maximization learning process is executed during model initialization and subsequent online updates. Based on a large number of historical process parameter coupled-state vector samples, using the hidden states as latent variables, the probability values ​​in the conditional probability table are iteratively optimized until the log-likelihood function converges.

[0051] The specific method for querying the posterior probability from the conditional probability table is as follows: Each component in the coupled state vector of the process parameters is discretized into multiple level intervals according to its numerical range, with each level interval corresponding to a discrete observation symbol. The division of the numerical range is based on the statistical distribution of the historical values ​​of each component in the coupled state vector of the process parameters. For each component, a minimum and maximum value are determined, and the interval between the minimum and maximum values ​​is divided into equal intervals. Each level range A value of 5 is chosen to ensure that each level interval has a certain historical sample coverage. Each level interval is assigned a discrete observation symbol, and when a component value falls into a certain level interval, that component is represented by the corresponding discrete observation symbol.

[0052] For each hidden state node in the Markov blanket dependency structure, the conditional probability value of its parent node set under specific combinations of discrete state values ​​is read from the conditional probability table. The set of discrete state values ​​of the hidden state node corresponds to different discrete safety sub-intervals in the process. These discrete safety sub-intervals include normal safety sub-intervals, low deviation safety sub-intervals, medium deviation safety sub-intervals, high deviation safety sub-intervals, and dangerous safety sub-intervals, each corresponding to a discrete state value. The conditional probability table is a multidimensional table in its storage structure. The dimension is determined by the number of discrete values ​​of the hidden state node and its parent node set. Each cell in the table stores the probability of the hidden state taking a specific value under a given combination of parent node values.

[0053] Based on the discrete observation symbols of each parent node corresponding to the current observation evidence, locate the row in the conditional probability table corresponding to the parent node combination that matches the current observation evidence. Extract the probability distribution of all possible discrete state values ​​for that hidden state node in the conditional probability table row as the query result. When there is a parent node combination in the conditional probability table that completely matches the current observation evidence, directly extract the probability distribution under this combination.

[0054] When no parent node combination in the conditional probability table perfectly matches the current observation evidence, the Laplace smoothing method is used to interpolate the probability distributions of adjacent parent node combinations to generate the posterior probability under the current observation evidence. An adjacent parent node combination is one whose discrete observation symbol differs from the current parent node combination by only one adjacent level. The interpolation process involves collecting the hidden state probability distributions corresponding to all adjacent parent node combinations in the conditional probability table. For each hidden state discrete state value, the average probability of that discrete state value under adjacent parent node combinations is calculated, and a Laplace smoothing constant is added to the average value. , The value is set to 0.1, and then the probability values ​​after adjusting all discrete state values ​​are normalized to obtain the posterior probability distribution under the current observation evidence. The value of 0.1 is chosen to avoid numerical problems caused by zero probability in chemical process safety assessments, while maintaining a reasonable bias towards historical data statistics.

[0055] The hidden state probability distribution at the previous sampling time is fused with the observation likelihood probability at the current time using a Bayesian recursive formula, and the hidden state probability distribution at the current sampling time is iteratively updated. The Bayesian recursive formula is expressed as: ; in, Indicates the current sampling time Hidden state nodes. Indicates the previous sampling time Hidden state nodes; and These represent the specific discrete state values ​​of the hidden state nodes, and all possible discrete safe subintervals are traversed. Indicates the current sampling time The observational evidence consists of a combination of discrete observation symbols after discretization of each component of the coupled state vector of process parameters; It is the probability distribution of the hidden state at the previous sampling time, stored in the circular buffer; It is the hidden state transition probability, which is obtained by querying the hidden state historical dependency information provided by the set of parent nodes of the current process state node in the Markov blanket dependency structure from the conditional probability table. It is the observation likelihood probability, representing the probability given the current hidden state as... Observed under the conditions The probability can be obtained by looking up the conditional probability table; It is the updated probability distribution of the hidden state at the current sampling time. Each component in the hidden state probability distribution corresponds to the probability value of a discrete safety sub-interval of the process.

[0056] At the initial sampling time of the system, if there is no hidden state probability distribution from the previous sampling time, a non-information prior distribution is used instead, and the probability values ​​of all discrete safe sub-intervals are equal.

[0057] The probability values ​​corresponding to the dangerous safety sub-intervals in the hidden state probability distribution are accumulated to obtain the probability distribution of the process state deviating from the preset safety domain at the current moment. The dangerous safety sub-intervals correspond to two categories in the discrete state values ​​of the hidden state, representing high deviation from the safety sub-interval and dangerous safety sub-interval. The probability values ​​corresponding to these two categories of discrete state values ​​are extracted from the hidden state probability distribution and summed.

[0058] Example 5:

[0059] In specific implementation, please refer to Figure 4 The process of determining and outputting the safety status level of a chemical process based on the cumulative probability value exceeding the safety threshold in the probability distribution is as follows.

[0060] The hidden state probability distribution output by the state transition probability deduction unit extracts all probability values ​​corresponding to the hazardous safety sub-intervals. The hidden state probability distribution output by the state transition probability deduction unit is a probability vector with a length equal to the total number of discrete safety sub-intervals. Each component in the hidden state probability distribution corresponds to a discrete safety sub-interval in the process. Discrete safety sub-intervals include normal safety sub-intervals, low deviation safety sub-intervals, medium deviation safety sub-intervals, high deviation safety sub-intervals, and hazardous safety sub-intervals. Hazardous safety sub-intervals specifically include two categories: high deviation safety sub-intervals and hazardous safety sub-intervals. The probability values ​​corresponding to the high deviation safety sub-intervals and the hazardous safety sub-intervals are extracted from the hidden state probability distribution. These two probability values ​​are then summed to obtain a single value, called the cumulative hazardous probability value.

[0061] The cumulative hazard probability value is compared with multiple pre-set safety status level thresholds. These thresholds include a first-level threshold, a second-level threshold, and a third-level threshold, where the first-level threshold is lower than the second-level threshold, and the second-level threshold is lower than the third-level threshold. The first-level threshold is set to 0.1, the second-level threshold to 0.3, and the third-level threshold to 0.6. The first-level threshold of 0.1 is based on the basic control requirements for acceptable risk probability in chemical industry safety management, meaning that the cumulative hazard probability value should be maintained below 0.1 under normal operating conditions. The second-level threshold of 0.3 is based on the fact that when the cumulative hazard probability value exceeds 0.3, the trend of the process deviating from the safety domain has become statistically significant, requiring proactive monitoring measures. The third-level threshold of 0.6 is based on the fact that when the cumulative hazard probability value exceeds 0.6, the probability of the chemical process entering a hazardous state exceeds the probability of a safe state, requiring immediate intervention.

[0062] When the cumulative hazard probability value is below the first-level threshold, the safety status of the chemical process is determined to be low-risk. A low-risk level indicates that the chemical process is operating within the normal range of a preset safety domain, and fluctuations in process parameters are at an acceptable level.

[0063] When the cumulative hazard probability value is greater than or equal to the first-level threshold but less than the second-level threshold, the safety status of the chemical process is determined to be at a medium-risk level. A medium-risk level indicates that the process parameters of the chemical process have deviated to a certain extent, requiring increased monitoring frequency and attention to further development of the deviation trend.

[0064] When the cumulative hazard probability value is greater than or equal to the second-level threshold but less than the third-level threshold, the safety status of the chemical process is determined to be high-risk. A high-risk level indicates that the process parameters have deviated significantly, and the cumulative probability value of the deviation from the preset safety domain has reached a significant level. An early warning signal needs to be issued to prompt operators to take control measures to curb the continued development of the deviation trend.

[0065] When the cumulative hazard probability value is greater than or equal to the third-level threshold, the safety status of the chemical process is determined to be critical. A critical level indicates that the chemical process is in or about to enter a dangerous state, and the cumulative probability value deviating from the preset safety domain has exceeded the upper limit tolerable for safe operation, requiring immediate triggering of interlock protection actions or emergency shutdown procedures. In some implementations, the upper limit of the cumulative hazard probability value corresponding to the critical level is 1. When the cumulative hazard probability value equals 1, it means that the sum of the probability values ​​corresponding to the dangerous safety sub-intervals is 1, definitively indicating that the chemical process is in a dangerous state.

[0066] The determined safety status level is output to the human-machine interface (HMI) of the chemical process safety control system in the form of discrete level identifiers. The discrete level identifiers use numerical codes: low risk level corresponds to identifier 0, medium risk level to identifier 1, high risk level to identifier 2, and critical level to identifier 3. Each identifier includes a corresponding Chinese level name string and a timestamp, which are transmitted to the display buffer of the HMI. The HMI switches the corresponding display color and alarm sound based on the identifier: low risk level displays green and no alarm sound is triggered; medium risk level displays yellow and triggers an intermittent alarm sound; high risk level displays orange and triggers a continuous warning sound; and critical level displays red and triggers an emergency alarm sound.

[0067] The dynamic recursive process of the safety state assessment model, after completing at each sampling time, simultaneously stores the coupled state vector of the process parameters output at the current sampling time and the hidden state probability distribution into a circular buffer. The circular buffer is a fixed-capacity first-in-first-out (FIFO) storage queue, with the fixed capacity set to... Data records at each sampling time point The value is 120. The value of 120 is chosen because, under a preset sampling period of 1 second, the circulating buffer can hold 120 seconds (2 minutes) of continuous process status data, which can cover the complete time window from the inception to the development of most abnormal conditions in the chemical process. When the number of data records stored in the circulating buffer reaches a fixed capacity... When a new data record is written, it will overwrite the oldest data record. The circular buffer always retains the most recent data record. The process parameters coupled with the state vector and the probability distribution of the hidden state at each sampling time.

[0068] When the latent state probability distributions of multiple consecutive sampling times stored in the circular buffer all indicate a security status level lower than the low-risk level, the online parameter update process of the security status assessment model is triggered. A security status level lower than the low-risk level means that the extracted and calculated cumulative hazard probability value is lower than the first-level threshold. The specific number of consecutive sampling times is set to... indivual, The value is 10. The value of 10 is chosen because 10 consecutive sampling times at low risk levels strongly confirm that the chemical process is operating stably. Data from stable operation is representative enough to serve as learning samples for expectation maximization. The online parameter update process utilizes the most recent consecutive sampling time data stored in the circular buffer. Specifically, the most recent consecutive sampling time data refers to data that traces backward from the current sampling time. The system generates a sequence of sampling times, including all coupled state vectors of process parameters and the probability distribution of hidden states within a continuous sampling time window. Expectation-maximization (EM) learning is then re-executed using this data. The input data for EEM consists of the coupled state vectors of process parameters from the currently stored continuous sampling times, and the initial hidden state distribution uses the currently stored hidden state probability distribution as its initial value. During the iteration process, EEM updates the conditional probability table in the state transition probability deduction unit. When the log-likelihood function of the EEM process converges, the updated conditional probability table is output, replacing the original one, and applied to the state transition probability deduction calculation at subsequent sampling times.

[0069] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for assessing the safety status of chemical processes based on multi-dimensional process parameters, characterized in that, include: Acquire multidimensional process parameter time series data from multiple monitoring nodes in a chemical process; A dynamic recursive security state assessment model is constructed, which includes a parameter-coupled memory unit and a state transition probability deduction unit. The time-domain difference recursion and spatial cross-coupling calculations are performed on the multidimensional process parameter time series data through the parameter coupling memory unit to generate a process parameter coupling state vector with timestamps. The process parameters are coupled with the state vector and input into the state transition probability inference unit. The state transition probability inference unit determines the Markov blanket dependency structure based on the statistical learning results of historical accident cases of chemical processes, and iteratively calculates the probability distribution of the process state deviating from the preset safety domain at the current moment. Based on the cumulative probability value exceeding the safety threshold in the probability distribution, the safety status level of the chemical process is determined and output.

2. The method for assessing the safety status of chemical processes based on multi-dimensional process parameters according to claim 1, characterized in that, The acquisition of multidimensional process parameter time series data from multiple monitoring nodes in the chemical process specifically includes: Temperature, pressure, flow rate, and liquid level parameters of each monitoring node are collected synchronously according to the preset sampling cycle. For each process parameter collected from each monitoring node, a missing value interpolation operation is performed. The missing value interpolation operation is based on the historical trend of the same parameter type of the same monitoring node and performs linear interpolation. The interpolated multidimensional process parameters of each monitoring node are arranged in chronological order to form a multidimensional process parameter time series data matrix indexed by the sampling time. The rows of the multidimensional process parameter time series data matrix correspond to different sampling times, and the columns correspond to different combinations of monitoring nodes and parameter types.

3. The method for assessing the safety status of chemical processes based on multi-dimensional process parameters according to claim 2, characterized in that, In the missing value interpolation operation, the interpolation step size of linear interpolation shall not exceed twice the preset sampling period, and when the number of consecutive missing data points exceeds the preset number threshold, the interpolation shall be abandoned and the monitoring node shall be marked as a data failure state.

4. The method for assessing the safety status of chemical processes based on multi-dimensional process parameters according to claim 1, characterized in that, The construction of a dynamic recursive security state assessment model specifically includes: Multiple accident case segments and multiple normal case segments are extracted from the historical operating data of chemical processes. Each case segment contains a multi-dimensional process parameter time series at multiple consecutive sampling times and the corresponding safety status label. The time series of multidimensional process parameters in each case segment are grouped according to monitoring nodes, and the mutual information values ​​between parameter types within each monitoring node and the mutual information values ​​of the same parameter type between different monitoring nodes are calculated. Based on the mutual information value, key parameter combinations that contribute more than a preset contribution threshold to the security status label are selected. The key parameter combinations are used to define the set of parent nodes for each node in the Markov blanket dependency structure. The set of parent nodes is used to limit the conditional probability table structure of the state transition probability inference unit. A Bayesian structure learning algorithm is used to perform directional edge search on the Markov blanket dependency structure to determine the directed dependencies between nodes in the state transition probability inference unit, thereby completing the initialization of the dynamic recursive security state assessment model.

5. The method for assessing the safety status of chemical processes based on multi-dimensional process parameters according to claim 4, characterized in that, The mutual information value is calculated using a nonparametric method based on kernel density estimation, and the time window length is set to three times the main disturbance period of the chemical process.

6. The method for assessing the safety status of chemical processes based on multi-dimensional process parameters according to claim 1, characterized in that, The process of performing time-domain difference recursive and spatial-domain cross-coupling calculations on multidimensional process parameter time series data through parameter-coupled memory units specifically includes: For any current sampling time, the parameter coupling memory unit reads the historical coupling state vector output from the previous sampling time and the multi-dimensional process parameter time series data of the current sampling time; For each monitoring node at the current sampling time, calculate the first-order time difference value according to the parameter types within that monitoring node. The first-order time difference value is the difference between the parameter value at the current sampling time and the parameter value at the previous sampling time. The local time domain variation characteristics of the monitoring node are obtained by weighted summing of the first-order time difference values ​​of all parameter types within the same monitoring node. The weight of each parameter type in the weighted summation is determined by the frequency of occurrence of that parameter type in the Markov blanket-dependent structure. Calculate the first-order spatial difference between any two different monitoring nodes, where the first-order spatial difference is the difference between the parameter values ​​of the two monitoring nodes on the same parameter type; The local time-domain variation characteristics of all monitoring nodes and the first-order spatial difference values ​​of all different monitoring node pairs are vectorized and then recursively superimposed with the historical coupling state vector output at the previous sampling time to generate the process parameter coupling state vector at the current sampling time.

7. The method for assessing the safety status of chemical processes based on multi-dimensional process parameters according to claim 6, characterized in that, The weights of each parameter type in the weighted summation are normalized before vector concatenation, so that the sum of all weights within the same monitoring node is 1.

8. The method for assessing the safety status of chemical processes based on multi-dimensional process parameters according to claim 1, characterized in that, The process parameters coupled with the state vector are input into the state transition probability inference unit. Based on the Markov blanket dependency structure determined by statistical learning results of historical accident cases in chemical processes, the state transition probability inference unit iteratively calculates the probability distribution of the process state deviating from the preset safety domain at the current moment, specifically including: The process parameters coupled with the state vector at the current sampling moment are used as observation evidence and input into the state transition probability inference unit; The state transition probability deduction unit determines the set of parent nodes and the set of child nodes that are directly related to the current process state node based on the Markov blanket dependency structure. The posterior probability of each hidden state node under the current observation evidence is retrieved from a pre-stored conditional probability table, which is obtained by expectation maximization learning of the process parameters coupled with the state vector at all sampling times in the historical operation data of the chemical process. The hidden state probability distribution at the previous sampling time is fused with the observation likelihood probability at the current time using the Bayesian recursive formula, and the hidden state probability distribution at the current sampling time is iteratively updated. Each component of the hidden state probability distribution corresponds to the probability value of a discrete safety sub-interval of the process. The probability values ​​of the corresponding dangerous safety sub-intervals in the hidden state probability distribution are summed to obtain the probability distribution of the process state deviating from the preset safety domain at the current moment.

9. The method for assessing the safety status of chemical processes based on multi-dimensional process parameters according to claim 8, characterized in that, The step of querying the posterior probability of each hidden state node under the current observation evidence from a pre-stored conditional probability table specifically includes: Each component in the coupled state vector of the process parameters is discretized into multiple level intervals according to its numerical range, and each level interval corresponds to a discrete observation symbol. For each hidden state node in the Markov blanket dependency structure, read the conditional probability value of the set of parent nodes under specific combination conditions when the hidden state node takes each discrete state value from the conditional probability table. Based on the discrete observation symbols of each parent node corresponding to the current observation evidence, locate the corresponding conditional probability row from the conditional probability table, and extract the probability distribution of all hidden state values ​​in that row as the query result; When there is no parent node combination in the conditional probability table that perfectly matches the current observation evidence, the Laplace smoothing method is used to interpolate from the probability distribution of adjacent parent node combinations to generate the posterior probability under the current observation evidence.

10. The method for assessing the safety status of chemical processes based on multi-dimensional process parameters according to claim 1, characterized in that, Based on the cumulative probability value exceeding the safety threshold in the probability distribution, the safety status level of the chemical process is determined and output, specifically including: Extract all probability values ​​corresponding to the dangerous and safe sub-intervals from the probability distribution output by the state transition probability deduction unit, and sum all the probability values ​​to obtain the cumulative dangerous probability value; The cumulative risk probability value is compared with a plurality of pre-set safety status level thresholds, including a first level threshold, a second level threshold, and a third level threshold, wherein the first level threshold is less than the second level threshold, and the second level threshold is less than the third level threshold. When the cumulative hazard probability value is lower than the first-level threshold, the safety status level of the chemical process is determined to be low-risk. When the cumulative hazard probability value is between the first-level threshold and the second-level threshold, the safety status level of the chemical process is determined to be medium risk. When the cumulative hazard probability value is between the second-level threshold and the third-level threshold, the safety status level of the chemical process is determined to be high-risk. When the cumulative hazard probability value is higher than the third-level threshold, the safety status of the chemical process is determined to be critical. The determined safety status level is output to the human-machine interface of the chemical process safety control system in the form of discrete level identifiers.

Images