Equipment dynamic risk assessment method based on neural differential equation and dynamic event tree

Abstract

The application discloses a kind of equipment dynamic risk assessment methods based on neural differential equation and dynamic event tree, comprising the following: neural event differential equation is used to generate system simulation path and node, then according to node merging simulation path generates DDET and carries out post-pruning to it;Black box event function is learned, and new DDET is generated using the black box event function;The simulation result of system dynamics is realized using the generated DDET;Above-mentioned link in the whole process, further acceleration is carried out using GPU parallel computing technology to it.In many PSA methods, the significant advantage of using DDET is that it can break the limitations of static analysis method in describing the dynamic behavior of complex equipment system, effectively describe the dynamic characteristics and accident evolution scene of the system, especially in the processing of the mutual influence between process variables and component state.

Description

Technical Field

[0001] This invention relates to the field of risk assessment technology, specifically to a method for dynamic risk assessment of equipment based on neural differential equations and dynamic event trees. Background Technology

[0002] Dynamic Event Tree (DET) analysis, as a probabilistic safety assessment (PSA) method, has been widely applied in aerospace, nuclear power, electronic communications, medical, energy, and military fields. DET modeling and analysis can integrate reliability-influencing factors such as external environment, human error, and equipment failure to obtain information on the evolution path, evolution time, and evolution probability of potential system accidents. This allows for further analysis and prediction of system safety risk levels and potential environmental impacts, guiding personnel to take appropriate measures to avoid or mitigate losses caused by accidents. Therefore, it provides important reference and decision support for equipment design, manufacturing, operation, and maintenance management. Dynamic Event Trees include Continuous Event Trees (CET) and Discrete Dynamic Event Trees (DDET), but DDET is more commonly used in practical scenarios. Applying the DDET model for risk / reliability analysis mainly involves the following four steps: 1) System definition and key factor / event identification; 2) Determining various rules in the tree-building process; 3) Simulation implementation; 4) Statistical analysis of simulation results.

[0003] However, when using DDET to quantitatively assess the operational reliability of complex equipment systems, the following two major problems are often encountered: 1) How to quickly construct DDET; 2) How to solve the problem of massive data and branch explosion caused by the randomness of branch path evolution direction due to uncertainty in the DDET generation process. Summary of the Invention

[0004] In view of the technical problems mentioned in the background, the purpose of this invention is to provide a method for dynamic risk assessment of equipment based on neural differential equations and dynamic event trees.

[0005] To achieve the objectives of this invention, the technical solution provided by this invention is as follows:

[0006] A method for dynamic risk assessment of equipment based on neural differential equations and dynamic event trees includes the following:

[0007] The system simulation path and nodes are generated using neural event differential equations. Then, the simulation path is merged based on the nodes to generate DDET and post-pruning is performed to remove unnecessary branches.

[0008] By learning black-box event functions from pruned DDET and system data, and then using the black-box event functions to generate new DDET, the system simulation efficiency can be improved.

[0009] The generated DDET simulation results are used to achieve dynamic system simulation, and further risk assessment based on statistics is completed to quantitatively evaluate the mission reliability of complex equipment systems. GPU parallel computing technology is used to further accelerate all the above-mentioned steps in the process.

[0010] Specifically, the method of generating system simulation paths and nodes using neural event differential equations includes:

[0011] After preprocessing the data, data that meets the conditions is extracted as the initial time and state for the system simulation, and input into the neural event differential equation solver along with the drift function and event function. The drift function reflects the differential relationship between the system state and time; the event function is the embodiment of the DDET tree building rules, which controls the triggering of events and directly determines the simulation results. The appropriate event function is selected according to the simulation requirements or the data conditions.

[0012] The neural event differential equation solver solves the problem based on the above inputs and obtains the event trigger time and system state. These values ​​are then input into the state update function. The system is then used to determine whether an accident has occurred. If an accident has occurred, the system simulation ends. Otherwise, the data is used as the initial time and state for the next simulation segment, and the system continues to be simulated until an accident occurs or the maximum simulation time is reached, at which point the simulation ends. Following this approach, countless system evolution paths can be generated.

[0013] Specifically, the process of generating a DDET based on the node merging simulation path and then pruning it includes:

[0014] The previously generated system evolution paths are merged. The merging logic is as follows: at each time step, the nodes and their states corresponding to the current time step of each evolution path are judged. If they are the same, the previous paths are merged, and the above judgment process is continued for the next time step. Otherwise, the paths are not merged. This process is repeated until a complete discrete dynamic event tree is generated. Then, a suitable pruning algorithm is called to remove redundant paths.

[0015] The process involves learning black-box event functions from pruned DDETs and system data, and then using these black-box event functions to generate new DDETs. Specifically, after generating a complete DDET and pruning it, the black-box event functions are learned from the data of this DDET, enabling it to trigger events sequentially given a termination time. Then, the black-box event functions are called, and neural event differential equations are used again to generate system simulation paths and nodes. Finally, the evolution paths are merged to obtain a new DDET.

[0016] The generated DDET is used to realize the dynamic simulation results of the system, and further complete the risk assessment based on statistics. Specifically, it includes: performing dynamic risk assessment of the system based on the paths in the new DDET; on the basis of constructing DDET, randomly simulating the paths of a large number of trees, then merging the identical paths and recording the number to obtain the simulation tree and the simplified simulation tree. As the set total number of simulations gradually increases, the probability of the accident corresponding to each path is then counted.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0018] Among the many PSA methods, the significant advantage of using DDET is that it can overcome the limitations of static analysis methods in characterizing the dynamic behavior of complex equipment systems, effectively describe the dynamic characteristics of the system and accident evolution scenarios, and has a strong advantage in dealing with the interaction between process variables and component states.

[0019] Building upon this foundation, since the differential dynamic system in this scheme is defined by a neural network, the present invention can perform system simulation in a GPU-accelerated environment, thereby significantly improving simulation efficiency. The present invention also refines the currently coarse DDET tree-building rules, reducing the uncertainty in the DDET construction process and solving the problems of massive data and evolutionary branch explosion. Simultaneously, the event function proposed in this invention can, to some extent, decouple the definition of system tasks from the traditional human reliability quantification evaluation framework.

[0020] Furthermore, the advantages of the neural differential equations used in this invention are as follows:

[0021] 1) In this context, the solver is treated as a black box, and the gradient is computed using the proximity sensitivity method. It solves an enhanced differential equation in time that includes the original state and the loss's sensitivity to the state. It can compute the gradient of the scalar loss for all inputs of any solver without backpropagation through computation. This enables the model to be trained at a constant memory cost, improves memory efficiency, and overcomes a major bottleneck in training deep models.

[0022] 2) Efficient and accurate modern ODE solvers can approximately ensure that the output value is within a given tolerance of the true value. By changing the tolerance, a trade-off can be made between accuracy and computational cost, and the model evaluation strategy can be adjusted so that the cost of evaluating the model is adapted to the complexity of the problem.

[0023] 3) Reduction in the number of parameters required for supervised learning tasks. When the dynamics of a hidden unit are parameterized as a continuous time function, the parameters of neighboring "layers" are automatically bundled together, resulting in higher parameter efficiency when using differential equations instead of residual networks for supervised learning.

[0024] 4) Continuous transformation makes the formula for variable change easier to calculate. Using it to build a new invertible density model can avoid the single bottleneck of normalized flow and can be trained directly through maximum likelihood.

[0025] 5) It enables prediction of continuous timelines. Unlike recurrent neural networks that require discretized observations, continuously defined models can uniquely define a latent trajectory given any latent state, thereby enabling predictions that can be made arbitrarily forward or backward in time. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the method provided in the embodiments of this application;

[0027] Figure 2 A schematic diagram of the simulation path and node flow of the system generated using neural event differential equations provided in this application embodiment;

[0028] Figure 3 This is a schematic diagram illustrating the process of generating DDET based on the node merging simulation path and then performing post-pruning, as provided in an embodiment of this application.

[0029] Figure 4 A schematic diagram illustrating the process of learning black-box event functions, generating new DDETs, and performing dynamic risk assessments of the system, as provided in this embodiment of the application.

[0030] Figure 5 A diagram illustrating the tree construction rules for DDET. Detailed Implementation

[0031] 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.

[0032] This invention defines a differential dynamic system defined by a neural network under GPU acceleration. By incorporating event functions that integrate the termination and branching rules of dynamic event trees, it refines the previously coarse DDET tree-building rules, thereby reducing the uncertainty in the DDET construction process and addressing the challenges of massive data volume and exponential growth in evolutionary branches. Simultaneously, the event functions presented in this technology can, to some extent, decouple the definition of dynamic event tree tasks from the traditional framework of human reliability quantification evaluation, thus overcoming the limitations of task definition.

[0033] like Figure 1 As shown, the technical solution provided by the present invention includes the following:

[0034] First, the system simulation path and nodes are generated using neural event differential equations. Then, the simulation paths are merged at the same nodes to generate DDET, and then post-pruning is performed to remove unnecessary branches.

[0035] By learning black-box event functions and then using them to generate new DDETs, the system simulation efficiency can be improved.

[0036] The generated DDET simulation results are used to achieve dynamic system performance, and a risk assessment based on statistics is further completed to quantitatively evaluate the mission reliability of complex equipment systems. For necessary steps in the entire process, GPU parallel computing technology is used to further accelerate the process.

[0037] The technical solution will be broken down and explained below.

[0038] (1) Generating system simulation paths and nodes using neural event differential equations

[0039] This invention uses neural event differential equations to generate system simulation paths and nodes given initial system values, termination time, drift function, event function, and instantaneous update function. For specific usage of each parameter and function, please refer to [reference needed]. Figure 2 .

[0040] After preprocessing the data in this invention, data meeting certain conditions are extracted as the initial time and state for system simulation, and input into the neural event differential equation solver along with the drift function and event function. Here, the drift function reflects the differential relationship between the system state and time; therefore, its determination requires a clear understanding of the physical laws satisfied by the system. The event function embodies the DDET tree-building rules, controlling the triggering of events and directly determining the simulation results. This invention can select appropriate event functions based on simulation requirements or data conditions.

[0041] The neural event differential equation solver solves the equations based on the above inputs, obtaining the event trigger time and system state. These values ​​are then input into the state update function. This invention uses this data to determine if an accident has occurred. If an accident occurs, the system simulation ends; otherwise, this data is used as the initial time and state for the next simulation segment, continuing the simulation until an accident occurs or the maximum simulation time is reached, at which point the simulation ends. Following this approach, this invention can generate countless system evolution paths.

[0042] (2) Generate DDET based on the node merging simulation path and perform post-pruning on it.

[0043] This invention merges previously generated system evolution paths. The merging logic is relatively simple and can be found above. Figure 3 On the left. This invention, at each moment, assesses the nodes and their states corresponding to each evolution path. If they are the same, the previous paths are merged, and the assessment process continues for the next moment; otherwise, the paths are not merged. This process is repeated until a complete discrete dynamic event tree is generated. Then, this invention calls a suitable pruning algorithm to remove redundant paths.

[0044] (3) Learn black-box event functions, generate new DDETs and conduct dynamic risk assessment of the system.

[0045] After generating and pruning a complete DDET tree, this invention uses the data from this DDET tree to learn black-box event functions, enabling it to trigger events sequentially given a termination time. For example... Figure 4 As shown.

[0046] Subsequently, this invention calls the black-box event function and uses neural event differential equations again to generate system simulation paths and nodes. Then, this invention merges the evolution paths to obtain a new DDET.

[0047] This invention performs dynamic risk assessment of the system based on paths in a new DDET (Directed Path Estimation and Path Calculation). Building upon the DDET, this invention randomly simulates paths in a large number of trees, then merges identical paths and records the number of occurrences, resulting in a simulated tree and a simplified simulated tree. As the total number of simulations gradually increases, this invention then calculates the probability of accidents occurring for each path.

[0048] The theory upon which this invention is based is as follows:

[0049] (1) System simulation process based on discrete dynamic event tree

[0050] In simple terms, the system simulation process can be described as follows: Starting from the initial node, the system state changes according to branching rules and evolution direction triggered by randomly generated events, thus generating new nodes. These new nodes are stored, and the state transition operation described above is repeated for each of them until the evolution path of the current node is about to enter the absorption state or reaches the termination time. At this point, the simulation stops, thus generating a discrete dynamic event tree reflecting the evolution process of the simulated system.

[0051] (2) Discrete dynamic event tree based on neural event differential equation

[0052] The core innovation of this scheme lies in proposing an event-driven system simulation method based on neural differential equations, with a focus on explaining the implementation of this step. The simulation process of an event-driven system based on neural differential equations is as follows: Figure 1 As shown.

[0053] With the help of neural event differential equations, this invention only requires initial system values ​​(t0, z(t0)), termination time T, drift function f, event function g, and instantaneous update function h to model the system by repeatedly calling the solver, obtaining a series of event occurrence times and system evolution paths z. Therefore, this invention can complete the construction of DDET by integrating the evolution paths and further achieve the corresponding probabilistic security assessment of the system.

[0054] The initial value, termination time, and instantaneous update function can be directly given. For other variables, this invention assumes that the system state is described by n discrete variables, denoted as k1, k2, ..., k. n , where the kth i There are N discrete variables. i If there are n states (i = 1, 2, ..., n), then the system has There are several combinations of discrete variable state values; the system also has m continuous parameter variables, denoted as x1, x2, ..., xm. m The system state at time t can then be represented as a row vector, i.e.

[0055] z(X,K,t)=[x1,x2,…,x m ,k1,k2,…,k n ,t].

[0056] Differentiating the node state z(t) with respect to time t, we obtain the drift function as follows:

[0057]

[0058] Where x' i (i = 1, 2, ..., n) represents the derivative of a continuous variable with respect to time, k' j(j=1,2,…,m) represents the time derivative of a discrete variable in a continuous-time system.

[0059] Event functions are represented in a time-dependent form. Where t is time, and z(t) is the system state. These are a series of parameters. Since system evolution depends on the occurrence of events, this invention also requires specifying triggering conditions for the event function so that the event is successfully triggered when the conditions are met.

[0060] (3) Tree building rules and their optimization

[0061] This invention uses tree-building rules in DDET to solve problems such as branch explosion, massive data, and redundant data generated during evolution, such as... Figure 5 As shown, the branching rule determines when DDET triggers a branch, the absorption rule limits the number of paths and the expansion of the tree, and the pruning rule controls the size of the tree by merging similar paths. In this context, the present invention utilizes event functions to describe and apply various tree-building rules.

[0062] However, as system complexity increases, this invention may be unable to determine all events that occur and their triggering conditions. To address this issue, this invention proposes a black-box event function based on machine learning. After generating DDET using traditional tree-building rules, this invention inputs relevant data as a training set into a suitable machine learning model for iterative learning, generating a black-box event function model. Based on this, this invention defines a prediction function that can predict and output results reflecting events and their occurrence times given a simulation termination time. In this way, this invention not only effectively triggers events but also provides a more comprehensive understanding of events in complex systems by combining system states, which is beneficial for in-depth system modeling.

[0063] Finally, it should be noted that the above embodiments are merely illustrative and explanatory of the present invention, and are not intended to limit the present invention to the scope of the described embodiments. Furthermore, those skilled in the art will understand that the present invention is not limited to the above embodiments, and many more variations and modifications can be made based on the teachings of the present invention, all of which fall within the scope of protection claimed by the present invention.

Claims

1. A method for equipment dynamic risk assessment based on neural differential equations and dynamic event trees, characterized in that, Including the following: The system simulation path and nodes are generated using neural event differential equations. Then, the simulation path is merged based on the nodes to generate DDET and post-pruning is performed to remove unnecessary branches. By learning black-box event functions from pruned DDET and system data, and then using the black-box event functions to generate new DDET, the system simulation efficiency can be improved. The generated DDET simulation results are used to achieve dynamic system simulation, and further risk assessment based on statistics is completed to quantitatively evaluate the mission reliability of complex equipment systems. GPU parallel computing technology is used to further accelerate all the above-mentioned steps in the process. The method of generating system simulation paths and nodes using neural event differential equations specifically includes: After preprocessing the data, data that meets the conditions is extracted as the initial time and state for the system simulation, and input into the neural event differential equation solver along with the drift function and event function. The drift function reflects the differential relationship between the system state and time; the event function is the embodiment of the DDET tree building rules, which controls the triggering of events and directly determines the simulation results. The appropriate event function is selected according to the simulation requirements or the data conditions. The neural event differential equation solver solves the problem based on the above inputs, obtains the event trigger time and system state, and then inputs these values ​​into the state update function. Use these data to determine whether an accident has occurred in the system. If an accident has occurred, the system simulation ends. Otherwise, use these data as the initial time and state for the next simulation and continue to simulate the system until an accident occurs or the maximum simulation time is reached, at which point the simulation ends. Following this approach, countless system evolution paths can be generated.

2. The equipment dynamic risk assessment method based on neural differential equations and dynamic event trees according to claim 1, characterized in that, DDET is generated based on the node merging simulation path and then post-pruned, specifically including: The previously generated system evolution paths are merged. The merging logic is as follows: at each time step, the nodes and their states corresponding to the current time step of each evolution path are judged. If they are the same, the previous paths are merged, and the above judgment process is continued for the next time step. Otherwise, the paths are not merged. This process is repeated until a complete discrete dynamic event tree is generated. Then, a suitable pruning algorithm is called to remove redundant paths.

3. The equipment dynamic risk assessment method based on neural differential equations and dynamic event trees according to claim 1, characterized in that, By learning black-box event functions from pruned DDETs and system data, and then using these black-box event functions to generate new DDETs, the process involves: after generating a complete DDET and pruning it, using the data from this DDET to learn black-box event functions, enabling it to trigger events sequentially given a termination time; then, calling the black-box event functions and again using neural event differential equations to generate system simulation paths and nodes, and finally merging the evolution paths to obtain a new DDET.

4. The equipment dynamic risk assessment method based on neural differential equations and dynamic event trees according to claim 1, characterized in that, The generated DDET is used to realize the dynamic simulation results of the system, and further complete the risk assessment based on statistics. Specifically, it includes: performing dynamic risk assessment of the system based on the paths in the new DDET; based on the construction of DDET, randomly simulating the paths of a large number of trees, then merging the identical paths and recording the number, to obtain the simulation tree and the simplified simulation tree. As the set total number of simulations gradually increases, the probability of the accident corresponding to each path is then counted.

Images