A chemical process dynamic risk assessment method and system and a storage medium

Abstract

The present application relates to the technical field of risk assessment, in particular to a chemical process dynamic risk assessment method and system and a storage medium, which first performs space-time double-layer adaptive filtering denoising and confidence estimation on a multi-source process state original sampling time sequence matrix to obtain a smooth feature evidence vector; then performs maximum posteriori expectation calibration learning on conditional probability parameters of a double-time slice dynamic Bayesian network to obtain a conditional probability binary structure configuration file; and finally performs online posterior probability particle filtering real-time inference tracking evaluation based on the conditional probability binary structure configuration file and the smooth feature evidence vector to obtain online posterior accident risk probability of a material serious crystal crack hidden state at each moment. The present application adopts space-time double-layer adaptive filtering denoising and confidence estimation, eliminates non-cooperative sharp pulse strong noise, avoids obvious phase angle rebound hysteresis during denoising, and greatly reduces the Internet of Things electrical clutter false alarm rate.

Description

Technical Field

[0001] This invention relates to the field of risk assessment technology, and in particular to a method, system and storage medium for dynamic risk assessment of chemical processes. Background Technology

[0002] Chemical pressure equipment such as reactors and high-pressure pipelines are highly susceptible to microcracks and hidden fatigue degradation under alternating loads and corrosive media. Since internal defects evolve in a latent state, traditional alarms based on fixed hard thresholds for pressure or external temperature exceedances have significant hysteresis and are difficult to accurately prevent before sudden material cracking and failure.

[0003] In recent years, several methods based on edge data acquisition and simple timing diagnostics for the Industrial Internet of Things (IIoT) have been proposed. However, these solutions face several intractable technical bottlenecks in practical industrial applications:

[0004] 1. On-site transient stray electrical signal interference and mottled industrial high-frequency electromagnetic noise: The transients of the start-up and shutdown of highly inductive equipment project extremely high electromotive force and ionization spike electromagnetic pulses onto the twisted-pair armored signal cables, which can easily lead to frequent unplanned emergency shutdowns of the equipment and serious economic losses due to material cold agglomeration and adhesion to the walls due to "top-level swing".

[0005] 2. Lack of a method to integrate the time-varying evolution mechanism of physical degradation into Bayesian hierarchical causal inference: The detailed description of simple isolated time series severs the causal evolution law of physical degradation accumulating over time, and lacks the ability to jointly estimate unmeasurable hidden states, resulting in the dual drawbacks of early excessive false alarms or late fatal missed alarms.

[0006] 3. Deadlock defect of zero frequency in major failure samples and high-dimensional gaps in offline maintenance data: It is explained that disaster samples are "almost extinct" zero-frequency samples, which leads to division by zero overflow deadlock when the traditional EM algorithm is used for offline likelihood calculation. In addition, long-term downtime and overhaul logs cause a bottleneck in high-dimensional time series data containing a large proportion of missing information.

[0007] 4. Collapse risks of twin 3D image surface rendering and end-point linkage under power failure and communication interruption: Analysis of the classic twin mesh is limited to surface shell rendering and cannot penetrate its deep material to show internal tensile principal direction crystal nucleus micro-anomalies; and the traditional safety interlock shut-off valve has the fatal flaw of complete failure and collapse of active reset under major disaster and complete paralysis (no power, communication interruption, pressure loss) scenarios. Summary of the Invention

[0008] The purpose of this invention is to provide a method, system, and storage medium for dynamic risk assessment of chemical processes to overcome the problems mentioned above.

[0009] To achieve the above objectives, the present invention provides a method for dynamic risk assessment of chemical processes, the method comprising:

[0010] The system continuously streams operating time-series signals of various physical sensors from industrial IoT edge network nodes in real time to obtain the original sampling time-series matrix of multi-source process states.

[0011] The original sampling time series matrix of the multi-source process state is subjected to spatiotemporal dual-layer adaptive filtering for denoising and confidence estimation to obtain a smooth feature evidence vector.

[0012] A dual-time-slice dynamic Bayesian network is constructed based on the smooth feature evidence vector; the dual-time-slice dynamic Bayesian network includes an observation layer and a hidden layer;

[0013] The conditional probability parameters of the dual-time-slice dynamic Bayesian network are subjected to maximum a posteriori expectation calibration learning to obtain a conditional probability binary structure configuration file.

[0014] Based on the conditional probability binary structure configuration file and the smoothed feature evidence vector, online posterior probability particle filtering real-time inference tracking and evaluation are performed to obtain the online posterior accident risk probability of the material's severe crystal crack hidden state at each time point.

[0015] Optionally, the step of performing spatiotemporal two-layer adaptive filtering for denoising and confidence estimation on the original sampling time series matrix of the multi-source process state to obtain a smooth feature evidence vector specifically includes:

[0016] Based on the original sampling time series matrix of the multi-source process state, the Pearson cross-correlation coefficients corresponding to each channel are solved on the time axis through a time sliding window.

[0017] The time-domain deviation confidence weights for each channel are generated based on the Pearson cross-correlation coefficients corresponding to each channel.

[0018] Based on the original sampling time series matrix of the multi-source process state, a spatial autocovariance matrix is ​​constructed on the spatial axis according to the spatiotemporal covariance physical correlation between different physical sensors;

[0019] The spatial autocovariance matrix is ​​used to extract the spatial deviation evaluation operator corresponding to each channel;

[0020] The temporal deviation confidence weight corresponding to each channel is multiplied with the spatial deviation evaluation operator to generate the spatiotemporal comprehensive confidence factor corresponding to each channel.

[0021] The smooth feature evidence vector is output based on the spatiotemporal integrated confidence factor of each channel.

[0022] Optionally, the smoothing feature evidence vector output based on the spatiotemporal integrated confidence factor of each channel specifically includes:

[0023] A diagonal weight matrix is ​​constructed based on the spatiotemporal integrated confidence factor of each channel;

[0024] The diagonal weight matrix is ​​used as the smoothing weights for local weighted sliding regression. The first derivative of the time series of each channel is calculated using the smoothing weights within the sliding window.

[0025] The first derivative is optimized using weighted least squares, and the coefficients of the constant term are selected. As a time-delay-free biased denoising and smoothing output, it forms a pure and smooth feature evidence vector.

[0026] Optionally, the step of performing maximum a posteriori expectation calibration learning on the conditional probability parameters of the dual-time-slice dynamic Bayesian network to obtain a conditional probability binary structure profile specifically includes:

[0027] The joint posterior distribution expectation of the hidden node states is calculated using the exact connection tree belief propagation algorithm;

[0028] A multidimensional Dirichlet prior probability distribution smoothing constraint term is injected into the likelihood objective function of the maximization step, and a Lagrange multiplier constraint term is introduced to calculate the partial derivative. The conditional probability parameters of the dual-time-slice dynamic Bayesian network are updated offline using the joint posterior distribution expectation to obtain the updated conditional probability parameter values.

[0029] Based on the updated conditional probability parameter values, a self-correcting estimation is performed on the conditional probability table in the historical feature running log of the missing breakpoint, generating a conditional probability binary structure configuration file.

[0030] Optionally, the step of performing online posterior probability particle filtering real-time inference tracking and evaluation based on the conditional probability binary structure configuration file and the smoothed feature evidence vector to obtain the online posterior accident risk probability of the material's severe crystal crack latent state at each time point specifically includes:

[0031] Based on the conditional probability binary structure configuration file and the smooth feature evidence vector, the device material degradation hidden state trajectory is discretized and tracked in the state space by parallel particle swarm, and the activity index is calculated in real time.

[0032] Based on the activity index, it is determined whether the tracking of each particle in the parallel particle swarm is normal; if the tracking is normal, the tracking continues; if the tracking is abnormal, after resampling, a Gaussian stretching random walk evolution operator covariant with singular value diagonal stretching is adaptively introduced along the main dimension direction of material stretching failure on the particle state space trajectory to perform evolution update, and the non-homogeneous derived particle vector after evolution update is obtained.

[0033] When recursively tracking the particle state of multi-source process sensor input features in a multi-dimensional state space, the conditional likelihood function of the material's hidden state under the constraint of a multi-time-slice dynamic Bayesian network is analyzed in real time based on the evolved and updated non-homogeneous derived particle vector, and then the online posterior accident risk probability of the material's severe crystal crack hidden state at each time point is recursively solved.

[0034] Optionally, the method further includes: performing three-dimensional twin color temperature perspective display based on the online posterior accident risk probability and active explosion protection through physical spring interlocking of normally closed valves.

[0035] Optionally, the active explosion-proof method based on the online posterior accident risk probability for three-dimensional twin color temperature perspective display and physical spring interlocking of normally closed valves specifically includes:

[0036] The online posterior accident risk probability is sent to the shader through a dual-frequency transmission channel for RGBA diffuse dynamic color temperature thermodynamic rendering display. Simultaneously, it is monitored on the mechanical feedback side through time-sequential sliding absolute integral decision-making. When the safety protection threshold is exceeded, the safety electromagnetic interlock valve is triggered to close, thereby achieving active explosion protection.

[0037] This invention also provides a dynamic risk assessment system for chemical processes, the system comprising:

[0038] The acquisition module is used to continuously stream the operating time-series signals of each physical sensor from the edge network node of the Industrial Internet of Things in real time, and obtain the original sampling time-series matrix of the multi-source process state.

[0039] The smooth feature evidence vector determination module is used to perform spatiotemporal dual-layer adaptive filtering denoising and confidence estimation on the original sampling time series matrix of the multi-source process state to obtain the smooth feature evidence vector.

[0040] The network construction module is used to construct a dual-time-slice dynamic Bayesian network based on the smoothed feature evidence vector; the dual-time-slice dynamic Bayesian network includes an observation layer and a hidden layer;

[0041] The maximum a posteriori expectation calibration learning module is used to perform maximum a posteriori expectation calibration learning on the conditional probability parameters of the dual time slice dynamic Bayesian network to obtain a conditional probability binary structure configuration file.

[0042] The inference tracing evaluation module is used to perform online posterior probability particle filtering real-time inference tracing evaluation based on the conditional probability binary structure configuration file and the smoothed feature evidence vector, so as to obtain the online posterior accident risk probability of the material's severe crystal crack hidden state at each time.

[0043] The present invention also provides a computer-storable medium, characterized in that it stores a computer program, which, when executed by a processor, implements the steps of the above-described dynamic risk assessment method for chemical processes.

[0044] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method for dynamic risk assessment of a chemical process.

[0045] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0046] This invention provides a method, system, and storage medium for dynamic risk assessment of chemical processes. First, a spatiotemporal dual-layer adaptive filtering denoising and confidence estimation are performed on the original sampling time-series matrix of multi-source process states to obtain a smoothed feature evidence vector. Then, the conditional probability parameters of a dual-time-slice dynamic Bayesian network are calibrated using maximum a posteriori expectation to obtain a binary structure configuration file for conditional probabilities. Finally, based on the binary structure configuration file and the smoothed feature evidence vector, online a posteriori probability particle filtering real-time inference tracking and evaluation are performed to obtain the online a posteriori accident risk probability of the material's severe crystal crack latent state at each time point. This invention employs spatiotemporal dual-layer adaptive filtering denoising and confidence estimation to eliminate non-cooperative sharp impulse noise, avoiding significant phase angle recoil hysteresis during denoising, and significantly reducing the false alarm rate of electrical clutter in the Internet of Things (IoT). Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a flowchart of the dynamic risk assessment method for chemical processes according to an embodiment of the present invention;

[0049] Figure 2 This is a comparison curve showing the performance and alarm time window of the present invention and the traditional static hard threshold early warning method in cluttered environments and sudden disasters. Detailed Implementation

[0050] 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, and 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.

[0051] The purpose of this invention is to provide a method, system, and storage medium for dynamic risk assessment of chemical processes to overcome the problems mentioned above.

[0052] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0053] like Figure 1 As shown, this invention discloses a method for dynamic risk assessment of chemical processes, the method comprising:

[0054] Step S1: Retrieve the operating time-series signals of each physical sensor from the industrial IoT edge network node in real time and continuously to obtain the original sampling time-series matrix of the multi-source process state.

[0055] Step S2: Perform spatiotemporal dual-layer adaptive filtering denoising and confidence estimation on the original sampling time series matrix of the multi-source process state to obtain a smooth feature evidence vector.

[0056] Step S3: Construct a dual-time-slice dynamic Bayesian network based on smooth feature evidence vectors; the dual-time-slice dynamic Bayesian network includes an observation layer and a hidden layer.

[0057] Step S4: Perform maximum a posteriori expectation calibration learning on the conditional probability parameters of the dual-time-slice dynamic Bayesian network to obtain the conditional probability binary structure configuration file.

[0058] Step S5: Based on the conditional probability binary structure configuration file and the smoothed feature evidence vector, perform online posterior probability particle filtering real-time inference tracking and evaluation to obtain the online posterior accident risk probability of the material's severe crystal crack hidden state at each time point.

[0059] The following steps are all executed primarily by the edge computing module. Each step will be discussed in detail below:

[0060] Step S1: Retrieve the operating time-series signals of each physical sensor from the industrial IoT edge network node in real time to obtain the original sampling time-series matrix of the multi-source process state; Specifically, deploy the physical sensor array on the main process pipe gallery and pressure vessel of the chemical plant, use the physical sensor array to collect the operating time-series signals of each chemical plant in real time, and send them to the industrial IoT edge network node for storage, so that the operating time-series signals can be retrieved from the industrial IoT edge network node through a continuous streaming parsing mechanism and combined online to construct the original sampling time-series matrix of the multi-source process state.

[0061] Step S2: Perform spatiotemporal two-layer adaptive filtering for denoising and confidence estimation on the original sampling time series matrix of the multi-source process state to obtain a smooth feature evidence vector, specifically including:

[0062] Step S21: Based on the original sampling time series matrix of the multi-source process state, solve the Pearson cross-correlation coefficient for each channel on the time axis using a time sliding window; the Pearson cross-correlation coefficient for each channel is actually the Pearson cross-correlation coefficient between the measured sequence of the current time period for each channel and the benchmark stationary operating condition process time series. The time sliding window disclosed in this embodiment is actually an adaptively movable length of [missing information]. Time window.

[0063] Specifically, based on the original sampling time series matrix of the multi-source process state, signals from each channel are sampled on the time axis through a time sliding window to obtain the measured sequence of the current time period corresponding to each channel; then, the Pearson cross-correlation coefficient between the measured sequence of the current time period corresponding to each channel and the benchmark stable operating condition process time series is calculated, and the specific formula is as follows:

[0064]

[0065] in, This represents the sampled value at the corresponding moment in the measured sequence of the current time period in channel c. The process timing is based on the baseline stable operating condition. To slide in time window Data sample index parameters within, As a rigidity factor, This is the arithmetic mean of the sequence measured in the current time period. This is the arithmetic mean of the process timings under baseline steady-state conditions. Stiffness factor. Usually set to

[0066] Step S22: Generate the time-domain deviation confidence weights for each channel based on the Pearson cross-correlation coefficients for each channel. The specific formula is as follows:

[0067] ;

[0068] ;

[0069] in, This represents the self-similarity difference deviation between the measured sequence of the current time period corresponding to channel c and the benchmark stationary operating condition process time series. Let Pearson cross-correlation coefficient be the number of channels c. The dimensionless system response stiffness factor is used to determine the weighted hard boundary characteristics. This is for occasional deviations from the safety hard boundary threshold. In this embodiment, the dimensionless system response rigidity factor is typically used. Set to 12.5; This will prevent occasional deviations from the safety hard threshold. Set it to 0.40. Additionally, the Pearson cross-correlation coefficient for channel c is... It was calculated within a time sliding window of size W.

[0070] Step S23: Based on the original sampling time series matrix of the multi-source process state, construct a spatial autocovariance matrix on the spatial axis according to the spatiotemporal covariance physical correlation between different physical sensors; the spatial autocovariance matrix is ​​also called the spatial adaptive correlation autocovariance matrix; the spatial autocovariance matrix has a dimension of 30*30.

[0071] The specific formula for constructing the spatial autocovariance matrix is ​​as follows:

[0072] ;

[0073] in, Let be the spatial autocovariance matrix. To track slice size for autocovariance, This refers to the index of independent variables for local discrete time-series samples within a slice of a spatial covariance sliding analysis. For the first The sampling time is a 30-dimensional multi-dimensional measurement and control feature column vector composed of observations from 30 physical multi-source sensors. In order to be in The expected mean vector of multidimensional sensor feature outputs within the sample analysis segment is represented by T, which is a transpose matrix operation. In this embodiment, the autocovariance tracks the slice size. The default value is 50. The range of values ​​for is: [t- +1,t].

[0074] Step S24: Extract the spatial bias evaluation operator corresponding to each channel using the spatial autocovariance matrix. The specific formula is as follows:

[0075] ;

[0076] ;

[0077] in, The residual between the measured and theoretical expected solutions for the c-th channel is... It is the log-inverse normalized matrix of the spatial autocovariance matrix. for The minimal regular constant term, where I is a 30-dimensional unit diagonal matrix. Let be the spatial autocovariance matrix. Let be the measured value collected by the sensor at time t for the c-th channel. This represents the theoretical value predicted by the model for the c-th channel at time t.

[0078] Step S25: Multiply the temporal deviation confidence weights corresponding to each channel with the spatial deviation evaluation operator to generate the spatiotemporal comprehensive confidence factor for each channel. The specific formula is as follows:

[0079] ;

[0080] in, For the spatial bias evaluation operator of the c-th channel, is the time-domain deviation confidence weight for the c-th channel.

[0081] This invention multiplies the temporal deviation confidence weight with the spatial deviation assessment operator to generate a spatiotemporal comprehensive confidence factor. This factor can filter out electromagnetic pulse noise and random process interference, eliminate non-cooperative spike noise, and apply zero-weighted damping to the first-order derivative of the LOESS local sliding weighted algorithm to smooth out spikes, thus eliminating unplanned false alarm shutdowns caused by "top-level pendulum jumps." This overcomes the first problem mentioned in the background technology.

[0082] Step S26: Output a smooth feature evidence vector based on the spatiotemporal comprehensive confidence factor of each channel; specifically, construct a diagonal weight matrix based on the spatiotemporal comprehensive confidence factor of each channel. The specific formula is as follows:

[0083] ;

[0084] The diagonal weight matrix is ​​used as the smoothing weights for the locally weighted regression (LOESS) moving regression. Within the moving window, the first derivative of each channel's time series is calculated using the smoothing weights, and the local minimization of the sum of the second-order self-partial residual matrices is obtained. The specific formula is as follows:

[0085]

[0086] in, To minimize the fitting parameters of the subsequent algebraic expression and The solution set; , These are all the undetermined fitting coefficients of the local first-order dynamic linear regression equation, representing the constant intercept and first-order slope of the current local regression estimation line, respectively. The constructed spatiotemporal two-level comprehensive confidence factor is used as the smoothing weight factor in the LOESS active window local weighted regression. These are the sampled values ​​of the physical measurement channel at each corresponding moment. In the sliding window Data sample index parameters within ( ).

[0087] The solution is obtained by optimizing the first derivative using weighted least squares, and a smooth feature evidence vector is output. Specifically, select the coefficients of the constant term. As a time-delay-free biased denoising and smoothing output, it forms a pure, smooth feature evidence vector. Because this regression algorithm has a local whitening function under the control of the spatial and temporal weighted partial mode parameters, it can completely zero-weight damp and smooth out abrupt electromagnetic ionization spikes, thereby outputting a smooth true feature sequence without phase recoil hysteresis.

[0088] Step S3: Construct a dual-time-slice dynamic Bayesian network based on smoothed feature evidence vectors, specifically including:

[0089] Step S31: Smooth the feature evidence vector As the observation layer of a dual-time-slice dynamic Bayesian network.

[0090] Step S32: Set discrete fault hidden nodes corresponding to the internal damage and deterioration state of the pressure-bearing material in the hidden layer of the dual-time-slice dynamic Bayesian network. The discrete fault hidden nodes have 3 damage and deterioration state values ​​(a value of 0 represents the undamaged state; a value of 1 represents the fatigue damage state; and a value of 2 represents the critical failure state of severe cracking).

[0091] Step S33: Based on the equipment's cumulative service cycle hours Using the time-domain independent variable, the degradation path integral and Weibull fatigue damage probability distribution are applied to dynamically calculate the time-varying transition probability of the conditional state corresponding to the damage degradation state of the discrete fault hidden node in the adjacent time slices of the dual-time-slice dynamic Bayesian network. The specific formula is as follows:

[0092]

[0093] in, Let t represent the damage and degradation state of the discrete faulty hidden node at time t. The value represents the damage and degradation state of the discrete fault hidden node at the previous adjacent time point, and t represents the cumulative service cycle hours of the currently evaluated equipment. The conversion split ratio factor between moderate fatigue and severe damage. They represent the adjacent previous time slices respectively. With the current time slice The corresponding damage and degradation state index of the discrete fault hidden node. The degradation state is set as follows: 0 is completely healthy and normal, 1 is moderate fatigue concentration, and 2 is severe damage and cracks.

[0094] Due to the principle of irreversible material mechanical damage, the probability of reverse transition is set to a fixed constant of 0. The first row of the matrix corresponds to the calculation of the transition and evolution probability in a fully healthy state, setting the probability of mutation and decay in the healthy state. It conforms to the Weibull cumulative failure time physical and mechanical continuous distribution density function law, therefore the probability of abrupt decay in the healthy state can be calculated. The specific formula is as follows:

[0095]

[0096] in, Let t represent the probability of a healthy state undergoing a sudden decay, and t represent the cumulative service cycle hours of the currently evaluated equipment. To adjust the geometric parameters of the degraded curve shape, The constant parameter for fatigue damage limit life. The hyperparameter for the geometric adjustment of the degradation curve shape. Limited by strong constraints Within the real number space.

[0097] The above-mentioned degradation curve shape geometric adjustment scale hyperparameter The optimal fixed calibration design is as follows Fatigue damage clearance life scale constant parameter The limiting constant is set as The conversion splitting ratio factor between moderate fatigue and severe damage is set as follows: The probability of the transition from moderate fatigue state 1 to severe material cracking 2 is dynamically calculated using an exponential crack propagation model targeting the sudden burst of the fission limit.

[0098] ;

[0099] in, This represents the probability of unhealthy state mutation and decay. For shape hyperparameters, To accelerate the characteristic boundary scale parameter, a steeper crack propagation abrupt change factor is often set as a shape hyperparameter. And accelerate the feature bounding scale parameter Shrink to h.

[0100] This invention discretizes damage into discrete hidden fault nodes containing moderate cracks and macroscopic yielding catastrophes. Based on the Weibull cumulative failure physical law of time-varying service time, it dynamically calculates the transition matrix of adjacent time slices, overcoming static limitations and completely eliminating early false alarms and late missed alarms. That is, it overcomes problem two in the background technology.

[0101] Step S32: Based on the time-varying transition probability of the conditional state, establish a dual-time-slice dynamic Bayesian network that incorporates the physical constraints of material fatigue life decay, also known as a directed acyclic causal probabilistic evolution network.

[0102] Step S4: Perform maximum a posteriori expectation calibration learning on the conditional probability parameters of the dual-time-slice dynamic Bayesian network to obtain the conditional probability binary structure configuration file, specifically including:

[0103] Step S41: Calculate the joint posterior distribution expectation of the hidden node states using the exact connection tree belief propagation algorithm.

[0104] Step S42: Inject a multidimensional Dirichlet prior probability distribution smoothing constraint term into the likelihood objective function of the maximization step, and introduce a Lagrange multiplier constraint term to calculate the partial derivative. Use the joint posterior distribution expectation to perform offline iterative updates on the conditional probability parameters of the dual-time-slice dynamic Bayesian network to obtain the updated conditional probability parameter values. The specific formula is as follows:

[0105] ;

[0106] in, The updated conditional probability parameter value, i.e., during the [number]th [period]. During the second offline parameter fitting iteration, the node Its parent node is configured with the following state arrangement: The discrete state of its own value is The conditional probability parameter at that time; D is the historical execution log containing missing breakpoints. For the joint posterior distribution expectation, that is, in the expectation E-step, the parameters of the current and previous steps are configured. As a basis for transfer, from historical operation logs The set of hidden layer variables obtained from the solution in the unmeasurable network is The joint posterior distribution expectation of each failure event under the given conditions. Let k' be the joint expected count value, and k' be the cumulative state index variable used to sum all possible fault state values ​​for this node in the summation denominator. Let be the expected number of transitions from state i to state k under condition j, and be the expected number of transitions to state k. To give the expected number of transitions to all possible states k′ from the same parent node combination (i, j), Self-indexing of nodes in the network. Configure a composite index for the corresponding parent node state. The index for the current node's state value (i.e., when the fault state is divided into 3 levels). ), n is the number of self-iteration rounds ( ).

[0107] In this invention, It is a conjugate symmetry-adjusted smoothing base hyperparameter of the multidimensional Dirichlet, constantly set and calibrated as a dimensionless parameter, confined to the interval [1.05, 1.15], with a preferred calibration value of 1.08, forcibly restricting the bias provided by the numerator and denominator bases. When When there is a lack of historical faults and the value is zero, both the numerator and denominator of the calculation are locked to a value based on... Under the robustness constant basis, the surplus / deficit rate of historical operation log D is in the range of 20% to 25%. When fault disaster data is extremely scarce, the likelihood estimation expectation value is... At that time, the molecule is symmetrically adjusted and a priori forced to remain constant at a non-zero base. .

[0108] This invention injects a symmetric multidimensional conjugate Dirichlet prior into the log-likelihood maximum a posteriori (MAP-EM) objective estimation, setting the base of its bias parameter as a limiting protection constant. This ensures that even under cold-state zero-history counting, the numerator and denominator can obtain a symmetric base difference protection of no less than a certain value, thus shielding register-triggered division-by-zero overflow deadlock at the mathematical algorithm level. This overcomes problem three in the background technology.

[0109] Step S43: Perform self-correcting estimation of the conditional probability table in the historical feature running log of the missing breakpoint based on the updated conditional probability parameter values, and generate a binary structure configuration file for conditional probability.

[0110] Step S5: Based on the conditional probability binary structure configuration file and smoothed feature evidence vector, perform online posterior probability particle filtering real-time inference tracking and evaluation to obtain the online posterior accident risk probability of the material's severe crystal crack latent state at each time point, specifically including:

[0111] Step S51: Based on the conditional probability binary structure configuration file and smooth feature evidence vector, the hidden state trajectory of the device material degradation is discretized and tracked in the state space using parallel particle swarm optimization, and the activity index is calculated in real time. The specific formula is as follows:

[0112] ;

[0113] in, For discrete parallel particle swarm, The normalized importance weight factor corresponding to the likelihood weight allocation of the i-th particle at tracking time t (satisfying the normalization condition). And the retrieval index ).

[0114] Specifically, this invention first deploys a conditional probability binary structure configuration file in an edge computing module as a priori physical knowledge base. Then, it streams smooth feature evidence vectors into the edge computing module at the microsecond level. Finally, using the streamed edge computing module, it performs discrete approximation tracking of the device material degradation hidden state trajectory in the state space through parallel particle swarm optimization. Furthermore, this invention recursively normalizes the importance weight of each particle based on the time-varying conditional transition matrix A(t). The system adaptively matches the observed smooth feature evidence vector Y(t) with the state points of each particle after advancement through transition probabilities. The higher the matching degree between the particle and the actual physical occurrence, the higher the importance weight w obtained by the particle based on the transition probability. To reflect the tracking state diversity and trajectory quality of the parallel particle swarm in each time-series particle flow iteration, the system is based on... Conduct activity indicators for the current tracking health and diversity trajectory quality of particle swarm optimization. Real-time calculation.

[0115] Step S52: Determine whether the tracking of each particle in the parallel particle swarm is normal based on the activity index; if the tracking is normal, the particle continues to survive by passing down the state level by level, preserving the continuity of the atomic history physical trajectory; if the tracking is abnormal, after resampling, an evolutionary update is performed on the particle state space trajectory along the main dimension of the material tensile failure, adaptively introducing a Gaussian stretching random walk evolutionary operator that is covariant with the singular value diagonal stretching, to obtain the non-homogeneous derived particle vector after the evolutionary update, the specific formula of which is:

[0116] ;

[0117] in, This represents the evolved and updated non-homogeneous derived particle vector. The stretching coefficient of the particles is randomly adjusted by Gaussian. The main diagonal basis matrix is ​​an orthogonal eigentransformation matrix. The square root of the diagonal matrix stretched to the angular feature scale. For standard high-dimensional Gaussian noise, This is the state position vector of the i-th particle obtained through resampling.

[0118] The method for determining whether the tracking of each particle in a parallel particle swarm is normal is based on the activity index. Specifically, it includes determining whether the activity index has fallen below the preset limit protection line. If it falls below the preset limit protection line, it indicates that the tracking is abnormal. If it is above the preset limit protection line, it indicates that the tracking is normal.

[0119] In this embodiment, the preset limit protection line is: , The number of particles in a discrete parallel particle swarm optimization is usually set to 1. Standard high-dimensional Gaussian noise This is a dimensionless constant, typically set to 0.035. The particle Gaussian random fine-tuning stretching coefficient... It is a dimensionless constant, usually set to 0.035.

[0120] If the activity index of this invention is at a high level above the limit protection line, the tracking algorithm is judged to have normal activity quality. The particle swarm state is then issued along the original approximation trajectory, forming a smooth long-term time integral sensitivity, while avoiding particle diversity degradation and computational power loss caused by frequent resampling. If the activity index is judged to fall below the limit protection line, it indicates that due to the homogeneous overlap of the swarm orbits in multiple rounds, particle depletion occurs, and the tracking algorithm sensitivity drops significantly. Resampling will be triggered immediately, and the resampled particle position vector will be obtained during the resampling allocation. An adaptive Gaussian stretching random walk evolution operator, covariant with singular value diagonal stretching, is introduced along the principal dimension of material tensile failure on the particle state space trajectory for evolutionary update. This "injects new blood of diversity" into the particle swarm to maintain ultra-sensitive tracking of micro-pulse-level crack propagation on the welded microcrystalline cross section, and obtains the non-homogeneous derived particle vector after evolutionary update.

[0121] Step S53: When recursively tracking the input features of multi-source process sensors in a multi-dimensional state space, the conditional likelihood function of the material's hidden state under the constraints of a multi-time-slice dynamic Bayesian network is analyzed in real time based on the evolved and updated non-homogeneous derived particle vectors. Then, the online posterior accident risk probability of the material's severe crystal crack hidden state at each time point is recursively solved. .

[0122] Step S6: Based on online posterior accident risk probability The system employs 3D twin color temperature perspective display and active explosion-proof interlocking with normally closed valves via physical springs. Specifically, it uses a dual-frequency transmission channel to perform online post-accident risk probability assessment. The data is fed into the shader for RGBA diffuse dynamic color temperature thermodynamic rendering and display. Simultaneously, on the mechanical feedback side, monitoring is performed through time-sequential sliding absolute integral decision-making. When the safety protection threshold is exceeded, a safety electromagnetic interlock valve is triggered to close, thereby achieving active explosion protection. Specifically, the inherently safe physical shutdown operation is performed by relying on the mechanical molecular deformation restoring force of a high-load-bearing spiral reset tension steel spring. The shader can be a WebGLFragment shader.

[0123] The online post-hoc accident risk probability of the material's severe crystallization latent state at each moment is obtained through a dual-frequency transmission channel. The data is fed into the shader for RGBA diffuse dynamic color temperature thermodynamic rendering display. The specific formula is as follows:

[0124]

[0125] in, It is a multidimensional color vector containing red. ,green ,blue and opacity component (represent (the fourth dimension parameter in the color vector), superscript This represents the vector transpose operation; the reference primary color adjustment constants are set to... Semi-transparency and opacity Thus, high contrast indicates the space of hidden strain hotspots under tensile stress.

[0126] When the posterior probability determines that the risk is extremely low, the model displays a healthy green with high chromaticity at the corresponding location; when the posterior probability determines that the risk occurs during a concentrated instantaneous micro-crack, the RGBA color gradation automatically and adaptively changes to a bright warm golden-orange through pixel interpolation performed by the webpage front end; when the posterior probability determines that the risk occurs beyond a catastrophic event, i.e., when the posterior probability exceeds the catastrophic threshold, the red channel is restricted and forcibly locked to a maximum brightness of 255, presenting an overall self-illuminating deep red, accurately locating the hot spot space.

[0127] Specifically, on the mechanical feedback side, the edge computing module calculates the post-accident risk probability. Along the time axis The absolute integral of time is calculated using a sliding time window of seconds. The specific formula is as follows:

[0128]

[0129] in, The probability integral blocking factor; Let be a continuous independent variable for time-domain integration.

[0130] When the integral factor of 5 consecutive uninterrupted sampling windows Exceeding the danger threshold ( When the control causes the safety PLC controller to cut off the excitation circuit, the solenoid valve loses power and exhausts air. The main pneumatic shut-off gate valve loses its air pressure thrust. Then, the high-load spiral reset tension steel spring, which is sleeved on the outer side of the piston valve core and is in an energy storage compression deformation, releases at high speed instantly by relying on the elastic restoring force of its own molecular deformation. Under the extreme danger of no passive signal, the gate valve is forced to move downward and close within one second, blocking the feeding line.

[0131] On the display side (digital twin multidimensional spatial projection): adaptively maps and drives the color temperature RGBA diffuse reflection state of the WebGLFragment shader, translating the implicit damage risk of the material into a visualized spatial cloud map hotspot temperature, realizing the spatial deconstruction of risk level and dynamic rendering presentation.

[0132] On the mechanical feedback side (closed-loop active safety response): by making a decision based on the sliding absolute integral of the execution time, when it is calculated that the system is in an intolerable catastrophic crisis range (exceeding the safety protection threshold), the safety electromagnetic interlock valve is immediately triggered to close. The mechanical molecular deformation restoring force of the high-load helical reset tension steel spring is used to perform physical shutdown, thereby achieving the intrinsically safe closed-loop tripping of the physical interface and eliminating potential sources of safety hazards.

[0133] ① Tracking and Diagnosis Side: When the feature active particle ESS scale falls below the 550 feature depletion line, this invention adaptively combines singular value decomposition (SVD) covariance matrix stretching Gaussian random walk particle self-reconstruction to dehomogenize; the solution is used as the RGBA shading rendering interpolation independent variable input to the WebGL fragment shader, and the perspective penetrates the three-dimensional space independent unit to render deep internal hotspots.

[0134] ② Purely Physically Intrinsically Safe Cut-off Side: The edge computing module controls the excitation power loss during the 2.5-second over-limit sliding integral catastrophe critical moment. Utilizing the molecular deformation recovery force of a pure helical energy-storing large steel spring (possessing a burst recovery elastic force of 4500N), a high-fail-safe elastic one-shot normally closed emergency closed loop is executed within 1 second under the "no power, no signal, no air pressure" environment, locking in the lowest production line. Based on the above methods, problem four in the background technology is solved.

[0135] In other words, this invention uses the calculated real-time online posterior accident risk probability as a quantitative decision target for dynamic risk assessment, and sends it to the digital twin display terminal and the physical security feedback loop through dual-frequency transmission channels to achieve multi-level risk prevention and control response under physical-digital joint interaction conditions.

[0136] The beneficial effects of this invention include:

[0137] 1. Spatiotemporal dual-layer Pearson self-similarity and spatial autocovariance inverse normalization denoising is adopted to prevent the denoising from producing obvious phase angle recoil hysteresis, thereby significantly reducing the false alarm rate of IoT electrical noise.

[0138] 2. By introducing the Weibull material mechanics physical degradation prior mechanism to reconstruct the dynamic Bayesian adjacent time slice state transition probability table, the shortcomings of oversensitivity of new equipment and slow alarm failure of old equipment are effectively prevented, and the prediction window is moved forward.

[0139] 3. Inject Dirichlet conjugate virtual hyperparameters to eliminate the abnormal deadlock caused by singularities in the log-likelihood maximization iterative solution resulting from cold-state low-probability explosions and other zero-frequency events in the mathematical and algorithmic intrinsic mechanisms.

[0140] 4. By introducing a Gaussian random walk evolution operator with feature stretching along the principal feature axis of the particle swarm spatial autocovariance (SVD) during the resampling stage, particle homogeneity depletion is eliminated, and the millisecond-level sensitivity of edge sensing and tracking of extremely low perturbations of microscopic and fine microcrack anomalies is improved.

[0141] 5. Digital twin color temperature diffuse reflection perspective diagnosis, combined with passive spiral energy storage multi-turn heavy-duty special steel large spring, returns to its normally closed position upon impact, ensuring highly reliable "Fail-Safe" active shutdown in severe and extreme situations such as power grid interruption due to lightning strikes, power outages, and loss of credit, achieving true intrinsic safety.

[0142] This invention discloses and compares the anti-exhaustion covariance stretched dual-particle filtering inference process for distributed edge computing:

[0143] After dozens of rounds of computation involving finite particles (N=1000 parallel discrete particles), their distribution easily collapses due to polarity overlap, converging onto a single high-weight particle. The effective particle count plummets, and particle orbits largely collapse to a very small number of homogeneous particles (particle depletion, causing depolarization lock-in). If, subsequently, extremely weak microscopic strain cracking occurs, causing transmission fluctuations, homogeneous particles will exhibit extremely severe hysteresis or even false negatives. This invention introduces particle estimation logic with a principal elastic covariance stretched Gaussian random walk into the RAM cache of the computation module:

[0144] 1. The edge computing module allocates a cache cluster to track 1000 particles in the state space: .

[0145] 2. Estimate its state transition changes and the nonlinear morphism of multidimensional IoT measurements:

[0146]

[0147]

[0148] in, and To introduce measured non-Gaussian clutter noise based on the material's temperature characteristics.

[0149] 3. Calculate the current effective evaluation dimension of the parallel particle swarm optimization, i.e., the feature activity index. :

[0150]

[0151] 4. When The drop exceeded the resampling safety threshold set by the first-generation system, i.e., below... Resampling is performed when there are fewer than 550 effective particles. Trajectories are resampled and assigned in the parallel particle space. The evolutionary differential term of the superimposed principal diagonal mechanical covariant and covariant stretching is given by the following formula:

[0152]

[0153] Singular value decomposition (SVD) was performed on the self-covariance matrix of 1000 sampled particles in the state space (multiple physical characteristic variables such as pressure, wall temperature difference, acoustic vibration, tensile force, etc.). Among them, the principal characteristic alternating basis matrix. It indicates the principal shear polar direction of the physical material most prone to fracture or systematic covariance under the combined alternating state of multiple physical fields under stress, and It is the square root matrix of the eigenvalues ​​of the principal feature stretched by the scale, for the walking operator. (where the dimensionless Gaussian stretching characteristic coefficient is set to) Elastic scaling and stretching adjustment under coupled diagonal multiphysics conformal deformation mechanism was carried out.

[0154] Through the above formula, each particle reconstruction, resampling, and update walk not only effectively resolves the depletion caused by the extreme replication of high-order particles, but also generates adaptive stretching based on the mechanical coupling covariation law, enabling particles to perfectly distribute and walk along the combined mechanical background, thereby improving the discrimination accuracy of weak deformations below megapascal.

[0155] like Figure 2 As shown, (a) is a sensor measurement voltage signal containing instantaneous noise and glitches, (b) is a schematic diagram of a traditional static pre-alarm, and (c) is a schematic diagram of the posterior probability of DBN risk introduced in this invention; Figure 2 As shown, in the 150-minute lead-up to the cataclysm, the welding defects, under stress, underwent microcrystalline cracking and skipping, resulting in only extremely weak and indistinguishable macroscopic effects. Weak wobbling. Traditional static safety warning systems cannot detect this early hazard before a serious explosion occurs; however, the method introduced in this invention… With adaptive principal covariant stretching, the edge computing module accurately captured weak fluctuations within a 1-millisecond calculation cycle. The probability of delayed danger began to rise significantly after 151 minutes, exceeding the judgment limit at 155.1 minutes. This enables a high degree of predictive and proactive blocking, establishing a forward-looking security interception window.

[0156] This invention also discloses a dynamic risk assessment system for chemical processes, the system comprising:

[0157] The acquisition module is used to continuously stream the operating time-series signals of each physical sensor from the edge network node of the Industrial Internet of Things in real time, and obtain the original sampling time-series matrix of the multi-source process state.

[0158] The smooth feature evidence vector determination module is used to perform spatiotemporal dual-layer adaptive filtering denoising and confidence estimation on the original sampling time series matrix of the multi-source process state to obtain the smooth feature evidence vector.

[0159] The network building module is used to build dynamic Bayesian networks with two time slices.

[0160] The maximum a posteriori expectation calibration learning module is used to perform maximum a posteriori expectation calibration learning on the conditional probability parameters of the dual time-slice dynamic Bayesian network to obtain a conditional probability binary structure configuration file.

[0161] The inference tracing evaluation module is used to perform online posterior probability particle filtering real-time inference tracing evaluation based on the conditional probability binary structure configuration file and the smoothed feature evidence vector, so as to obtain the online posterior accident risk probability of the material's severe crystal crack hidden state at each time.

[0162] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0163] The present invention also discloses a computer-storable medium storing a computer program, which, when executed by a processor, implements the steps of the above-described dynamic risk assessment method for chemical processes.

[0164] The present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described dynamic risk assessment method for chemical processes.

[0165] Embodiments of the present invention may be provided as methods, systems, or computer program products. The invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the invention may take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0166] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0167] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0168] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0169] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for dynamic risk assessment of a chemical process, characterized in that, The method includes: The system continuously streams operating time-series signals of various physical sensors from industrial IoT edge network nodes in real time to obtain the original sampling time-series matrix of multi-source process states. The original sampling time series matrix of the multi-source process state is subjected to spatiotemporal dual-layer adaptive filtering for denoising and confidence estimation to obtain a smooth feature evidence vector. A dual-time-slice dynamic Bayesian network is constructed based on the smooth feature evidence vector; the dual-time-slice dynamic Bayesian network includes an observation layer and a hidden layer; The conditional probability parameters of the dual-time-slice dynamic Bayesian network are subjected to maximum a posteriori expectation calibration learning to obtain a conditional probability binary structure configuration file. Based on the conditional probability binary structure configuration file and the smoothed feature evidence vector, online posterior probability particle filtering real-time inference tracking and evaluation are performed to obtain the online posterior accident risk probability of the material's severe crystal crack hidden state at each time point.

2. The dynamic risk assessment method for chemical processes according to claim 1, characterized in that, The step of performing spatiotemporal dual-layer adaptive filtering for denoising and confidence estimation on the original sampling time-series matrix of the multi-source process state to obtain a smooth feature evidence vector specifically includes: Based on the original sampling time series matrix of the multi-source process state, the Pearson cross-correlation coefficients corresponding to each channel are solved on the time axis through a time sliding window. The time-domain deviation confidence weights for each channel are generated based on the Pearson cross-correlation coefficients corresponding to each channel. Based on the original sampling time series matrix of the multi-source process state, a spatial autocovariance matrix is ​​constructed on the spatial axis according to the spatiotemporal covariance physical correlation between different physical sensors; The spatial autocovariance matrix is ​​used to extract the spatial deviation evaluation operator corresponding to each channel; The temporal deviation confidence weight corresponding to each channel is multiplied with the spatial deviation evaluation operator to generate the spatiotemporal comprehensive confidence factor corresponding to each channel. The smooth feature evidence vector is output based on the spatiotemporal integrated confidence factor of each channel.

3. The dynamic risk assessment method for chemical processes according to claim 2, characterized in that, The smoothed feature evidence vector output based on the spatiotemporal integrated confidence factor of each channel specifically includes: A diagonal weight matrix is ​​constructed based on the spatiotemporal integrated confidence factor of each channel; The diagonal weight matrix is ​​used as the smoothing weights for local weighted sliding regression. The first derivative of the time series of each channel is calculated using the smoothing weights within the sliding window. The first derivative is optimized using weighted least squares, and the coefficients of the constant term are selected. As a time-delay-free biased denoising and smoothing output, it forms a pure and smooth feature evidence vector.

4. The dynamic risk assessment method for chemical processes according to claim 1, characterized in that, The step of performing maximum a posteriori expectation calibration learning on the conditional probability parameters of the dual-time-slice dynamic Bayesian network to obtain a conditional probability binary structure configuration file specifically includes: The joint posterior distribution expectation of the hidden node states is calculated using the exact connection tree belief propagation algorithm; A multidimensional Dirichlet prior probability distribution smoothing constraint term is injected into the likelihood objective function of the maximization step, and a Lagrange multiplier constraint term is introduced to calculate the partial derivative. The conditional probability parameters of the dual-time-slice dynamic Bayesian network are updated offline using the joint posterior distribution expectation to obtain the updated conditional probability parameter values. Based on the updated conditional probability parameter values, a self-correcting estimation is performed on the conditional probability table in the historical feature running log of the missing breakpoint, generating a conditional probability binary structure configuration file.

5. The dynamic risk assessment method for chemical processes according to claim 1, characterized in that, The online posterior probability particle filtering real-time inference tracking and evaluation, based on the conditional probability binary structure configuration file and the smoothed feature evidence vector, to obtain the online posterior accident risk probability of the material's severe crystal crack latent state at each time point, specifically includes: Based on the conditional probability binary structure configuration file and the smooth feature evidence vector, the device material degradation hidden state trajectory is discretized and tracked in the state space by parallel particle swarm, and the activity index is calculated in real time. Based on the activity index, it is determined whether the tracking of each particle in the parallel particle swarm is normal; if the tracking is normal, the tracking continues; if the tracking is abnormal, after resampling, a Gaussian stretching random walk evolution operator covariant with singular value diagonal stretching is adaptively introduced along the main dimension direction of material stretching failure on the particle state space trajectory to perform evolution update, and the non-homogeneous derived particle vector after evolution update is obtained. When recursively tracking the particle state of multi-source process sensor input features in a multi-dimensional state space, the conditional likelihood function of the material's hidden state under the constraint of a multi-time-slice dynamic Bayesian network is analyzed in real time based on the evolved and updated non-homogeneous derived particle vector, and then the online posterior accident risk probability of the material's severe crystal crack hidden state at each time point is recursively solved.

6. The dynamic risk assessment method for chemical processes according to claim 1, characterized in that, The method further includes: performing three-dimensional twin color temperature perspective display based on the online post-hoc accident risk probability and active explosion protection through physical spring interlocking of normally closed valves.

7. The dynamic risk assessment method for chemical processes according to claim 1, characterized in that, The active explosion-proof method based on the online posterior accident risk probability for three-dimensional twin color temperature perspective display and physical spring interlocking of normally closed valves specifically includes: The online posterior accident risk probability is sent to the shader through a dual-frequency transmission channel for RGBA diffuse dynamic color temperature thermodynamic rendering display. Simultaneously, it is monitored on the mechanical feedback side through time-sequential sliding absolute integral decision-making. When the safety protection threshold is exceeded, the safety electromagnetic interlock valve is triggered to close, thereby achieving active explosion protection.

8. A dynamic risk assessment system for chemical processes, characterized in that, The system includes: The acquisition module is used to continuously stream the operating time-series signals of each physical sensor from the edge network node of the Industrial Internet of Things in real time, and obtain the original sampling time-series matrix of the multi-source process state. The smooth feature evidence vector determination module is used to perform spatiotemporal dual-layer adaptive filtering denoising and confidence estimation on the original sampling time series matrix of the multi-source process state to obtain the smooth feature evidence vector. The network construction module is used to construct a dual-time-slice dynamic Bayesian network based on the smooth feature evidence vector; the dual-time-slice dynamic Bayesian network includes an observation layer and a hidden layer; The maximum a posteriori expectation calibration learning module is used to perform maximum a posteriori expectation calibration learning on the conditional probability parameters of the dual time slice dynamic Bayesian network to obtain a conditional probability binary structure configuration file. The inference tracing evaluation module is used to perform online posterior probability particle filtering real-time inference tracing evaluation based on the conditional probability binary structure configuration file and the smoothed feature evidence vector, so as to obtain the online posterior accident risk probability of the material's severe crystal crack hidden state at each time.

9. A computer-storable medium, characterized in that, The device contains a computer program that, when executed by a processor, implements the steps of a dynamic risk assessment method for a chemical process as described in any one of claims 1-7.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the dynamic risk assessment method for chemical processes according to any one of claims 1-7.

Images