Safety critical system risk dynamic assessment system based on multi-modal fusion and uncertainty quantification
The dynamic risk assessment system for safety critical systems, which integrates multimodal fusion and uncertainty quantification, solves the problems of single risk assessment, insufficient fusion, and neglect of propagation mechanisms in existing technologies. It achieves real-time, accurate, and personalized risk assessment and provides systematic risk assessment and decision support.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ANHUI DIGITAL INTELLIGENCE PREDICTION TECHNOLOGY CO LTD
- Filing Date
- 2026-03-02
- Publication Date
- 2026-05-15
AI Technical Summary
Existing risk assessment systems suffer from problems such as the lack of a single and uncertain predictive model, insufficient fusion of multimodal information, neglect of risk propagation mechanisms, and inadequate personalization and adaptability, resulting in insufficient accuracy and reliability of the assessment.
A dynamic risk assessment system for safety critical systems employing multimodal fusion and uncertainty quantification includes multi-source heterogeneous data acquisition, hierarchical uncertainty quantification, adaptive multimodal fusion mechanism, risk propagation topology network, and personalized uncertainty calibration, providing interpretable risk assessment output.
It enables real-time, accurate, and personalized risk assessment of safety-critical systems, provides confidence indicators and risk propagation trend displays, and supports hierarchical decision-making.
Smart Images
Figure FT_1 
Figure FT_2 
Figure FT_3
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of industrial safety technology and artificial intelligence, specifically relating to a dynamic risk assessment system for safety-critical systems. More specifically, this invention relates to a real-time safety risk assessment system that integrates multimodal data and uncertainty quantification functions, applicable to fields with extremely high safety requirements such as industrial process control, critical infrastructure management, and high-reliability manufacturing systems. Background Technology
[0002] In safety-critical fields such as industrial control and infrastructure management, even minor system failures can lead to serious consequences, making real-time and accurate risk assessment crucial. Traditional safety assessment methods are mainly divided into two categories: physical model-based analytical methods and statistical methods based on historical data. Physical model methods assess risk by establishing fault propagation models of system components (such as fault trees and event trees), but they struggle to handle the nonlinear interactions of complex systems and unknown failure modes. Statistical methods rely on large amounts of historical incident data and predict risks through machine learning models, but they perform poorly in scenarios where data is scarce or its distribution varies.
[0003] In recent years, with the development of IoT and sensor technology, safety-critical systems can collect multi-source heterogeneous data, including sensor time-series data, device status logs, video surveillance streams, and operation log text. This multimodal data provides richer information dimensions for risk assessment, but also brings new technical challenges. Existing risk assessment systems suffer from the following technical problems:
[0004] 1. Lack of single prediction model and uncertainty: Most systems use a single prediction model, outputting deterministic risk assessment results without providing confidence intervals or uncertainty measures, making it impossible for decision-makers to judge the reliability of the prediction.
[0005] 2. Superficial multimodal information fusion: Existing multimodal fusion methods mostly use simple weighted averages or attention-based weighting, without fully considering the differences in quality, timeliness, and reliability of data from different modalities, and lack quality control mechanisms for the fusion process itself.
[0006] 3. Ignoring risk propagation mechanisms: Most systems treat risks as isolated events, failing to model the propagation paths and cascading effects of risks among system components, making it difficult to assess the impact of local failures on the overall system.
[0007] 4. Insufficient personalization adaptability: Risk assessment models are usually trained based on overall data and are not adjusted for different system states, operating stages and working conditions, which leads to a decrease in assessment accuracy in certain specific scenarios.
[0008] To address these issues, this invention provides a dynamic risk assessment system for safety critical systems that integrates multimodal fusion and uncertainty quantification. Summary of the Invention
[0009] The purpose of this invention is to provide a dynamic risk assessment system for safety-critical systems that integrates multimodal fusion and uncertainty quantification, thereby overcoming the aforementioned deficiencies of existing technologies. Specific objectives include:
[0010] 1. Establish a hierarchical uncertainty quantification framework to simultaneously measure cognitive uncertainty and accidental uncertainty, providing confidence indicators for risk assessment;
[0011] 2. Design an adaptive multimodal fusion mechanism to monitor the data quality of each modality in real time, dynamically adjust the fusion strategy, and avoid low-quality data from polluting the evaluation results;
[0012] 3. Construct a risk propagation topology network, model the propagation path of risk among system components, and realize systemic risk assessment;
[0013] 4. Develop personalized uncertainty calibration methods to provide conditional confidence intervals for different system states and operating conditions, thereby improving the accuracy of evaluation;
[0014] 5. Provide interpretable risk assessment outputs, visually displaying the sources of risk, confidence levels, and propagation trends, supporting tiered decision-making.
[0015] This invention provides a dynamic risk assessment system for safety-critical systems that integrates multimodal fusion and uncertainty quantification, comprising the following core modules:
[0016] 1. Multi-source heterogeneous data acquisition module:
[0017] Responsible for collecting four types of modal data from the safety-critical system:
[0018] Sensor timing data: sensor readings of physical quantities such as temperature, pressure, vibration, and current;
[0019] Equipment status log: Equipment self-test report, fault codes, maintenance records;
[0020] Video surveillance stream: Real-time monitoring video of key areas;
[0021] Operation log texts include operator instructions, work orders, shift handover records, etc.
[0022] The acquisition module supports adaptive protocol parsing and can handle a variety of industrial communication protocols.
[0023] 2. Data Preprocessing and Time Series Alignment Module
[0024] Cleaning, normalizing, and synchronizing multimodal data:
[0025] Missing value handling: A multidimensional imputation algorithm based on spatiotemporal correlation is used;
[0026] Outlier detection: Combining statistical tests with the Isolation Forest algorithm;
[0027] Timestamp alignment: Based on high-precision network time protocols, all modal data are unified to millisecond-level time precision;
[0028] Feature engineering: Extracting time-domain features, frequency-domain features, and time-frequency-domain features.
[0029] 3. Multimodal feature fusion module
[0030] A deep fusion of multimodal features is achieved using an improved low-rank attention mechanism:
[0031] Modal feature encoding: For each modal of data, a specialized encoding network is used to extract feature vectors;
[0032] Low-rank attention computation: Calculate cross-modal attention weights through matrix factorization with dynamically adjusted rank;
[0033] Feature fusion: Based on attention weights, features from each modality are weighted and aggregated to generate a cross-modal fusion feature vector.
[0034] 4. Uncertainty Quantification and Risk Assessment Module
[0035] The core innovative module of this invention adopts a layered architecture:
[0036] Basic risk assessment network: a hybrid architecture based on deep residual networks and gated recurrent units;
[0037] Cognitive uncertainty quantification: Monte Carlo Dropout is introduced into the basic network to measure model uncertainty through the output variance of multiple forward propagations;
[0038] Quantification of random uncertainty: Deep evidence learning is used to replace the last layer of the network with a Dirichlet distribution parameter estimation layer;
[0039] Personalized uncertainty calibration: Based on the conditional conformal prediction framework, personalized confidence intervals are provided for different system operating conditions.
[0040] 5. Risk Propagation Topology Network Construction Module
[0041] Abstracting the safety-critical system as a directed weighted graph:
[0042] Node: System component;
[0043] Edge: The physical connection or functional dependency between components;
[0044] Weight: Intensity of risk transmission.
[0045] Graph attention networks are used to calculate the probability of risk propagation.
[0046] 6. Risk Visualization and Decision Support Module
[0047] Provides a multi-dimensional visualization interface:
[0048] Risk heatmap: On the system topology diagram, the risk level of each component is represented by the intensity of the color.
[0049] Confidence curve: Shows how the confidence interval of the risk assessment results changes over time;
[0050] Risk propagation animation: Dynamically demonstrates the path and speed of risk propagation within the system;
[0051] Early warning dashboard: Displays the current risk level, major risk sources, and recommended measures. Attached Figure Description
[0052] Figure 1 : Schematic diagram of the overall system architecture of this invention
[0053] Figure 2 Detailed flowchart of the multimodal feature fusion module
[0054] Figure 3 Layered Architecture Diagram of Uncertainty Quantification and Risk Assessment Module
[0055] Figure 4 Schematic diagram of risk propagation topology network construction
[0056] Figure 5 Personalized Uncertainty Calibration Flowchart
[0057] Figure 6 Example image of a risk visualization interface Detailed Implementation
[0058] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The following embodiments use industrial process control systems as an application scenario, but the present invention is also applicable to other safety critical systems.
[0059] Example 1: Risk Assessment of Industrial Process Control Systems
[0060] Basic configuration:
[0061] A chemical process control system includes key components such as reactors, pipelines, heat exchangers, and control units. This embodiment deploys the following data acquisition points:
[0062] Sensors: Temperature sensors (15), pressure sensors (12), flow meters (8), level sensors (6), sampling frequency 1Hz-10Hz;
[0063] Video surveillance: 5 cameras in total, covering key equipment areas and control room, with a resolution of 1080p and a frame rate of 25fps;
[0064] Equipment logs: pump status log, valve action record, control system self-test report;
[0065] Operation logs: operator instructions, process parameter adjustment records, shift handover logs.
[0066] 1. Data preprocessing implementation:
[0067] Timestamp alignment: All sensor data is synchronized via a high-precision clock synchronization protocol, with millisecond-level accuracy. Each frame of the video stream is embedded with a timestamp, and text logs are aligned by parsing the time field.
[0068] Outlier detection: Robust principal component analysis (PCA) is used to separate normal patterns from outliers. For process variable data, the model is represented as follows:
[0069]
[0070] in This is a low-rank normal mode. This is a sparse outlier. It is solved using the augmented Lagrange multiplier method:
[0071]
[0072] Detected outliers are marked as "suspicious" for reference by subsequent fusion modules.
[0073] 2. Implementation of Multimodal Feature Fusion
[0074] Modal feature encoding:
[0075] Sensor time-series data: Temporal features were extracted using bidirectional LSTM encoding with a hidden layer dimension of 128.
[0076] Video data: Encoded using 3D CNN, based on the I3D architecture, to extract spatial-temporal features;
[0077] Device logs: Structured encoding is used to convert log entries into feature vectors;
[0078] Manipulating text: Encodes text using a pre-trained language model and extracts semantic features.
[0079] Low-rank attention computation: setting the initial rank The rank is dynamically adjusted based on the correlation of modal features. The cross-modal attention weight matrix is calculated. ,in Representing modes For modes The level of attention. Introducing the modal confidence imbalance coefficient:
[0080]
[0081] in This represents the confidence weight vector for each modality. When... When this happens, the fusion strategy is adjusted to reduce the attention weight of low-confidence modalities.
[0082] Modal missing detection latency coefficient: Monitor the heartbeat interval of each modal data stream and calculate the average latency.
[0083]
[0084] when At a certain time, it is determined that there is a delay problem in this mode, and the historical data interpolation strategy is activated.
[0085] 3. Implementation of Uncertainty Quantification
[0086] Basic risk assessment network: A hybrid architecture of ResNet-34 and GRU is adopted. ResNet extracts spatial features, and GRU handles temporal dependencies. The output layer is Softmax, predicting 5 risk levels: normal, attentive, alert, high risk, and emergency.
[0087] Cognitive uncertainty quantification: Dropout is added to the last three convolutional layers and the GRU layer of ResNet, and the Dropout rate during training is... During reasoning Sub-Monte Carlo sampling. Cognitive uncertainty is calculated as the variance of 50 predictions.
[0088] Quantification of random uncertainty: Replace the Softmax layer with an evidence collection layer, and output the Dirichlet distribution parameters. The network output has been modified as follows:
[0089]
[0090] in This is the $k$-th output of the evidence gathering layer. The random uncertainty is calculated as follows:
[0091]
[0092] Personalized uncertainty calibration: Dividing the operating state of the process control system into 4 operating conditions:
[0093] Startup phase: The system gradually transitions from a shutdown state to normal operating conditions;
[0094] Steady-state operation: The system operates stably within the normal process parameter range;
[0095] Load adjustment: Changes in processing volume or product specifications;
[0096] Planned shutdown: The system will gradually cease operation according to a plan.
[0097] Maintain an independent calibration set for each operating condition, size Conditional conformal prediction is used, and the non-conformal scoring function is:
[0098]
[0099] in The standard deviation of the model estimate. Prevent division by zero. For new samples, select the appropriate confidence threshold based on their operating conditions. Generate a prediction set.
[0100] 4. Implementation of the Risk Propagation Topology Network
[0101] System Graph Construction: The process control system is abstracted as a directed graph containing 45 nodes and 62 edges:
[0102] Nodes: Reactor (2), Heat Exchanger (3), Pump (6), Piping Section (24), Valve (8), Controller (6);
[0103] Edges: physical connections (pipe connections) and functional dependencies (control signals).
[0104] Propagation probability calculation: A graph attention network is used, with node features including device type, years of operation, maintenance records, and historical failure rate. After three layers of GAT (Graph Attention Network), with each layer having a head count of $h = 4$, the node representation $h_i'$ is obtained. The propagation probability of edge $(i, j)$ is:
[0105]
[0106] in Here is the weight matrix, and $b_p$ is the bias. This is the Sigmoid function.
[0107] Critical path identification: An improved A* algorithm is used, with the evaluation function as follows:
[0108]
[0109] in To account for the actual propagation risk from the starting point to $n$, For heuristic estimation, This is the sum of the uncertainties of each node on the path. Uncertainty weights are used. The algorithm identifies high-risk propagation paths, such as "pump failure → abnormal pipeline pressure → decreased heat exchanger efficiency → runaway reaction temperature".
[0110] 5. Implementation of Visualization and Decision Support
[0111] Risk heatmap: On the process system diagram, components are colored according to their risk level.
[0112] Green: Normal ( );
[0113] Yellow: Attention ( );
[0114] Orange: Alert ( );
[0115] Red: High risk ( );
[0116] Crimson: Emergency ( ).
[0117] Confidence curve: Shows the change in the confidence interval of the risk assessment of critical equipment over the past hour, including the 95% confidence interval and the 68% confidence interval.
[0118] Early warning decision: When the system detects "abnormal pump vibration" (risk level: alert, confidence level: 0.85) and identifies that the fault may propagate to the reactor along the critical path, the system issues a level three early warning:
[0119] Level 1: Notify the control room operator;
[0120] Level 2: It is recommended to increase the frequency of vibration monitoring;
[0121] Level 3: Prepare a backup pump start-up plan.
[0122] Example 2: Health Management of High-Reliability Manufacturing Systems
[0123] 1. System Specificity:
[0124] High-reliability manufacturing systems operate in complex environments, with sensor data exhibiting high noise, multiple modes, and significant uncertainties. This embodiment focuses on demonstrating the robustness of the present invention in high-noise environments.
[0125] Data characteristics:
[0126] High-noise vibration signals: low signal-to-noise ratio;
[0127] Multiple sampling rate data: vibration signal (high frequency), temperature signal (low frequency), pressure signal (medium frequency);
[0128] Frequent data loss: Some data packets are lost due to transmission interruptions.
[0129] 2. Key technology adjustments
[0130] Robust Feature Extraction: For high-noise vibration signals, wavelet packet transform is used instead of Fourier transform to extract frequency band energy features. Wavelet packet decomposition formula:
[0131]
[0132] in For wavelet packet basis functions, the decomposition depth Energy characteristics of 64 frequency bands were obtained.
[0133] Missing data handling: Generative adversarial networks (GANs) are used to fill in large-scale missing data. The generator is trained. and discriminator Objective function:
[0134]
[0135] in For complete data, The noise is random. The generator learns the data distribution and generates appropriate imputations for missing parts.
[0136] Uncertainty weight adjustment: In high-noise environments, the weight of random uncertainty is increased. The overall uncertainty calculation is adjusted as follows:
[0137]
[0138] Data noise is the main source of uncertainty.
[0139] 3. Real-time optimization
[0140] Manufacturing system health management requires rapid response; this embodiment incorporates the following optimizations:
[0141] Lightweight Model: Training a Small Risk Assessment Network Using Knowledge Distillation
[0142] Teacher network: ResNet-50 + BiLSTM, with a large number of parameters;
[0143] Student network: Lightweight network + GRU, small number of parameters;
[0144] Distillation loss: Combining prediction loss and distribution matching loss.
[0145] Streaming processing: Employs a sliding window mechanism to achieve near real-time evaluation.
[0146] Example 3: Infrastructure Park Security Monitoring System
[0147] 1. Large-scale distributed deployment
[0148] The infrastructure park comprises multiple subsystems and thousands of monitoring points. This example demonstrates the system's scalability and distributed computing capabilities.
[0149] Architecture Design:
[0150] Edge nodes: Each subsystem deploys an edge computing unit to perform local data collection and preliminary risk assessment;
[0151] Regional gateway: Each 5-10 subsystems are connected to a regional gateway for multi-system risk correlation analysis;
[0152] Central platform: Comprehensive risk assessment and visualization of the entire park.
[0153] Data flow: Edge nodes periodically send risk assessment results and uncertainty indicators to the regional gateway; the regional gateway sends a regional risk summary to the central platform.
[0154] 2. Cross-system risk propagation modeling
[0155] Infrastructure parks are characterized by strong cascading effects; localized failures can trigger park-wide incidents. This embodiment extends the risk propagation network to the cross-system level:
[0156] Hypergraph construction: In addition to intra-system propagation, add cross-system propagation edges, based on:
[0157] Material flow connections, energy flow connections, spatial proximity.
[0158] Multi-scale propagation model: employing a hierarchical graphical neural network.
[0159] Micro level: Risk propagation within the subsystem;
[0160] Meso-level: Risk propagation between subsystems;
[0161] Macro level: Risk transmission at the park level.
[0162] Each layer uses graph representations of different granularities, connected through graph pooling and graph upsampling.
[0163] 3. Collaborative Decision Support
[0164] When multiple systems experience risks simultaneously, the system provides collaborative decision-making suggestions:
[0165] Resource optimization allocation: Assume there are $M$ high-risk systems, but only One emergency team ( The system prioritizes data based on the following factors:
[0166] Risk level and confidence level;
[0167] Risk transmission potential;
[0168] Accessibility of emergency resources;
[0169] Historical handling results.
[0170] Establish a multi-objective optimization model to solve for the optimal resource allocation scheme.
Claims
1. A dynamic risk assessment system for safety-critical systems based on multimodal fusion and uncertainty quantification, characterized in that, include: Multi-source heterogeneous data acquisition module: used to acquire raw data in four modalities from safety-critical systems: sensor time-series data, equipment status logs, video surveillance streams, and operation log text. Data preprocessing and time series alignment module: performs missing value imputation, outlier detection and timestamp alignment on multimodal data to generate feature sequences with a unified time base; Multimodal feature fusion module: It adopts an improved low-rank attention mechanism to dynamically calculate the contribution weight of each modality data, and extracts cross-modal correlation features through orthogonal matrix factorization and bidirectional long short-term memory network; Uncertainty Quantification and Risk Assessment Module: Based on a deep evidence learning framework and an improved conformal prediction algorithm, it generates assessment results that include risk probability distribution and uncertainty intervals. Risk visualization and decision support module: Visualizes risk assessment results in the form of heatmaps, confidence curves and risk propagation paths, and provides tiered early warning suggestions.
2. The system according to claim 1, characterized in that, The multimodal feature fusion module further includes: Modal confidence imbalance coefficient calculation unit: Based on the signal-to-noise ratio, historical accuracy and real-time consistency index of each modal data, dynamically calculate the confidence distribution vector between modalities; Modal Missing Detection Delay Coefficient Calculation Unit: By monitoring the heartbeat signal and data packet arrival interval of each modal data link, the time delay from actual modal failure to system failure identification is quantified; Fusion Hysteresis Drift Failure Coupling Unit: The two coefficients mentioned above are nonlinearly coupled to generate a fusion hysteresis drift failure coupling index. When the index exceeds a preset threshold, the fusion strategy is dynamically adjusted.
3. The system according to claim 1, characterized in that, The uncertainty quantification and risk assessment module adopts a hierarchical uncertainty quantification architecture, including: Cognitive uncertainty quantification layer: A Bayesian neural network with Monte Carlo Dropout is used to measure the uncertainty of the model itself through the output variance of multiple forward propagations; Random uncertainty quantification layer: A deep evidence learning method is used to estimate the inherent noise uncertainty of the data through the Dirichlet distribution parameters; Personalized Uncertainty Calibration Layer: Based on the conditional conformal prediction framework, it provides personalized confidence intervals for different risk scenarios and system states, achieving individual condition coverage guarantee.
4. The system according to claim 3, characterized in that, The personalized uncertainty calibration layer implements the following steps: Constructing a nonconformal scoring function: A scoring function based on Mahalanobis distance is used to measure the degree of deviation between the predicted results and the true risk value; Dynamic partitioning of calibration sets: Based on the operating stage and operating conditions of the safety-critical system, historical data is divided into multiple conditional calibration subsets; Adaptive threshold determination: For each subset of conditional calibration, the corresponding confidence threshold is determined based on the quantile distribution of its nonconformal scores; Online continuous calibration: The calibration set and confidence threshold are updated in real time through a sliding window mechanism and exponentially weighted moving average to address data distribution drift.
5. The system according to claim 1, characterized in that, The system also includes a risk propagation topology network construction module, which: The components, subsystems, and connections of a safety-critical system are abstracted as nodes and edges in a directed weighted graph; A graph neural network is used to calculate the probability of risk propagation between nodes; An improved A* algorithm was applied to identify key risk propagation paths; Based on the risk propagation dynamics model, the spatiotemporal evolution trend of risk in the system is predicted.
6. A method for dynamic assessment of safety criticality system risk based on the system described in any one of claims 1-5, characterized in that, Includes the following steps: Step S1: Real-time acquisition of multimodal data from the safety critical system, followed by preprocessing and timing alignment; Step S2: Calculate the confidence imbalance coefficient and missing detection delay coefficient for each modality, and dynamically adjust the fusion strategy; Step S3: Extract cross-modal spatiotemporal correlation features through the multimodal feature fusion module; Step S4: Based on the hierarchical uncertainty quantification architecture, generate assessment results that include risk level, probability of occurrence, and confidence interval; Step S5: Construct a risk propagation topology network, identify key risk paths, and predict propagation trends; Step S6: Output the risk assessment results through a visual interface, and provide tiered early warning and intervention suggestions.
7. The method according to claim 6, characterized in that, The specific implementation of the dynamic adjustment of the fusion strategy in step S2 includes: When the modal confidence imbalance coefficient exceeds the threshold α, the weight allocation of low-confidence modes is reduced; When the modality missing detection delay coefficient exceeds the threshold β, the historical data interpolation or adjacent modality replacement strategy for that modality is enabled. When the coupling index of fusion hysteresis drift failure continues to rise, the online retraining mechanism of the fusion model is triggered.
8. The method according to claim 6, characterized in that, The mathematical expression for the hierarchical uncertainty quantification in step S4 is: Among them, cognitive uncertainty Calculated using Monte Carlo Dropout sampling: Random uncertainty Calculation of Dirichlet distribution parameters through deep evidence learning: in The first of the Dirichlet distributions One concentration parameter, This represents the number of risk categories.
9. A non-transitory computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 6-8.
10. An electronic device, characterized in that, It includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method of any one of claims 6-8.