A dynamic calculation system for major hazard source risks based on AI big model empowerment

Abstract

The present invention discloses a dynamic calculation system for major hazard source risks based on AI large-scale model empowerment, which relates to the field of risk calculation technology and includes: a multimodal data acquisition and preprocessing module for connecting sensors through industrial protocols to achieve synchronous acquisition of raw data; a heterogeneous data spatiotemporal fusion engine module for building a spatiotemporal graph model and analyzing the nonlinear coupling relationship of multi-source data; an online incremental learning and model fine-tuning module for triggering dynamic parameter adjustment based on real-time data streams; a risk transmission probability calculation module; a multimodal knowledge self-evolution module; and an adaptive threshold management and alarm module. The present invention adopts a sliding window dynamic confidence interval calculation combined with a time-varying confidence coefficient adjustment mechanism, and realizes adaptive fitting of the threshold to real environmental disturbances by embedding a periodic correction term, effectively resolving the two-way contradiction between detection sensitivity and false alarm suppression, and forming a complete technical closed loop from dynamic perception, cross-domain verification to rapid linkage.

Description

Technical Field

[0001] The present invention relates to the field of risk calculation technology, and in particular to a dynamic calculation system for major hazard source risks based on AI large model empowerment. Background Art

[0002] Risk assessment technology for major hazardous sources has gradually shifted from traditional manual experience to intelligent approaches. Typical technologies include: 1. Risk Analysis Models: The industry widely uses static models based on the Analytic Hierarchy Process (AHP), Fault Tree Analysis (FTA), and Event Tree Analysis (ETA). These models determine risk weights through expert scoring and combine historical accident statistics for risk assessment. These methods rely on a priori rule bases and have limited ability to model the dynamic evolution of complex systems. 2. Sensor Monitoring and Threshold Warning: Single-point sensors such as temperature, pressure, and gas concentration are deployed, integrated with PLC / DCS systems, to generate real-time alarms for out-of-limit parameters. Typical examples include tank pressure safety interlocks and ventilation systems triggered by toxic gas LEL concentration thresholds. These technologies focus solely on transient anomalies of a single parameter and fail to consider the potential risks of data coupling across multiple sources. 3. Early Application of Prediction Algorithms: Some improved solutions incorporate machine learning models such as time series analysis (such as ARIMA forecasting), support vector machines (SVMs), and random forests to predict equipment failure probabilities based on historical training data. However, these models are limited by the low dimensionality of their input features and their lack of generalization, making them difficult to adapt to the multimodal data fusion and rapid migration requirements of industrial scenarios. 4. Industry Standard Framework: Internationally accepted standard processes such as HAZOP (Hazard and Operability Analysis) and LOPA (Layer of Protection Analysis) are often used as the basis for risk identification, but they require regular manual review and cannot be dynamically corrected by linking the results to actual operating data in real time.

[0003] Existing technologies generally exhibit characteristics such as being dominated by static rules, focusing on local single-point analysis, and lacking cross-domain risk transmission mechanisms. This has led to significant bottlenecks in the timeliness and accuracy of dynamic risk assessments of complex hazardous sources. Therefore, there are technical problems in existing major hazardous source risk assessment methods, including the following:

[0004] 1. Traditional static risk assessment suffers from severe lags: Existing technologies rely heavily on regular manual inspections and fixed threshold judgments, with long data update cycles (usually measured in weeks or months). These technologies are unable to perceive dynamic changes in hazard sources (such as instantaneous fluctuations in tank pressure, cumulative trends in gas concentration, and accelerated equipment corrosion rates) in real time. This leads to delayed risk warnings and makes it difficult to respond promptly to rapidly evolving safety hazards.

[0005] 2. Insufficient ability to analyze the coupling of multi-source risk factors: Traditional methods typically use linear models or independent weight superposition methods, which make it difficult to effectively quantify the nonlinear interactions between multi-dimensional heterogeneous data such as environmental parameters (temperature, humidity, wind speed), equipment status (vibration, temperature), and human operation records. In particular, they lack the ability to model complex coupled risks such as the superposition of equipment aging and extreme weather, and the chain reaction of process anomalies and inspection omissions, which can easily lead to one-sided risk assessment results.

[0006] 3. Poor model generalization and adaptability: Current mainstream risk assessment models are often customized for specific scenarios (such as chemical parks and oil and gas storage), relying on expert experience to establish rule bases. When faced with process changes, new hazard types, or regional environmental differences, these models require extensive manual re-tuning, making them difficult to quickly adapt to new scenarios. Furthermore, traditional models are sensitive to the density and quality of historical data. Predictive reliability plummets when data is sparse (e.g., for newly built facilities) or when labels are missing (e.g., for unrecorded minor incidents).

[0007] 4. Lack of tracking of dynamic risk transmission paths: Existing technologies focus on independent assessments of local hazard sources and lack computable modeling of cross-regional and cross-level risk transmission mechanisms (such as the spatiotemporal diffusion path of pipeline leaks triggering chain explosions, and the probabilistic transmission chain of equipment failures triggering safety system malfunctions). This makes it difficult to predict the cascading amplification effects of major risks, affecting the precise scheduling of prevention and control resources.

[0008] 5. Rigid alarm rules and high false alarm rates: Traditional methods trigger alarms based on fixed thresholds or simple statistical models (such as the 3σ principle). These methods are unable to dynamically adapt to operating condition fluctuations (such as normal deviations in equipment deformation caused by diurnal temperature differences) and data quality disturbances (such as temporary sensor drift). This leads to frequent false alarms (such as normal maintenance operations being misidentified as leaks) or missed alarms (such as slow-changing corrosion not being identified), reducing system credibility. Summary of the Invention

[0009] Based on this, it is necessary to provide a dynamic calculation system for major hazard source risks based on AI big models to address the above technical issues.

[0010] The present invention provides a major hazard source risk dynamic calculation system based on AI large model empowerment, including:

[0011] Multimodal data acquisition and preprocessing module, used to connect sensors through industrial protocols to achieve synchronous acquisition of raw data;

[0012] Heterogeneous data spatiotemporal fusion engine module, used to build spatiotemporal graph models and analyze nonlinear coupling relationships between multi-source data;

[0013] Online incremental learning and model fine-tuning module, used to trigger dynamic parameter adjustments based on real-time data streams;

[0014] Risk transmission probability calculation module, used to combine Bayesian networks and CFD simulation to generate risk diffusion paths;

[0015] Multimodal knowledge self-evolution module, used to continuously update the risk knowledge base using large model technology;

[0016] Adaptive threshold management and alarm module, used to dynamically optimize alarm rules and output emergency decisions.

[0017] Furthermore, the multimodal data acquisition and preprocessing module includes:

[0018] The sensor data acquisition module is used to collect analog and digital signals from the physical device itself, distinguish the input type based on the signal type judgment circuit, and then divert and forward the signals;

[0019] Sliding window segmentation module, which is used to cut data blocks by timestamp based on the ring buffer of edge computing terminals;

[0020] Denoising processing module, used to perform wavelet packet decomposition on data signals to achieve high-frequency noise filtering;

[0021] The time alignment module is used to calibrate the clocks of each sensor. If the clock deviation exceeds the tolerance, interpolation compensation is triggered.

[0022] The missing value LSTM interpolation module is used to obtain time series data collected in the past N sampling periods, predict the current missing values ​​and replace the placeholders based on the long short-term memory network;

[0023] The modal normalization module is used to normalize the interpolated time series data.

[0024] Furthermore, the heterogeneous data spatiotemporal fusion engine module includes:

[0025] Input module, used to receive preprocessed time series data and extract multimodal feature vectors;

[0026] The spatiotemporal graph construction module is used to construct a spatiotemporal graph consisting of nodes and edges, and use a temporal convolutional network to extract the temporal dependencies of each node;

[0027] Attention network module, which is used to dynamically adjust the weights between nodes based on the attention adaptation mechanism;

[0028] The risk level prediction module is used to output the risk level prediction results and trigger the adaptive adjustment of the local neuron learning rate based on the latest accumulated data.

[0029] Furthermore, we construct a spatiotemporal graph consisting of nodes and edges, and use a temporal convolutional network to extract the temporal dependencies of each node, including:

[0030] Set each sensor as a node with device attribute labels;

[0031] Based on the historical accident database, the coupling strength of risk factors is statistically analyzed and the edge weights of the two nodes are calculated;

[0032] A temporal convolutional network is constructed based on causal dilated convolution, and the network structure parameters are set. The temporal dependency of each node is extracted using the temporal convolutional network. The network structure parameters include input dimension, convolution kernel diameter, dilation factor, number of output channels and activation function.

[0033] Furthermore, based on the attention adaptation mechanism, the weights between nodes are dynamically adjusted, including:

[0034] Based on the node feature vector, learnable parameter matrix and attention vector, the attention coefficient is calculated and updated in real time through back propagation of the learnable parameter matrix and attention vector, so that the connection weights between sensor nodes can be adaptively adjusted as the distribution of sensor data changes.

[0035] Furthermore, the attention coefficient calculation formula is:

[0036] ;

[0037] Where, α ij represents the connection weight between node i and node j; a represents the attention vector; W represents the learnable matrix; h i and h j represents the node feature vector; T represents the transposed operation matrix.

[0038] Furthermore, the risk transmission probability calculation module includes:

[0039] The Bayesian network inference engine module is used to build a Bayesian network using physical devices and environmental states. It uses the acquired real-time sensor data to regularly revise the joint probability distribution in the conditional probability table and calculate the risk probability corresponding to each physical device node.

[0040] The CFD physical simulation engine module is used to start the CFD simulation engine when the risk probability of any physical device node exceeds the preset threshold, and simulate and output the spatiotemporal distribution of the combustible gas diffusion concentration cloud map;

[0041] The risk transmission path integration module is used to mark the risk path of combustible gas leakage based on the concentration field data and equipment topology diagram output by the CFD simulation engine.

[0042] Furthermore, the multimodal knowledge self-evolution module includes:

[0043] Multimodal input parsing module, used to parse structured data and unstructured data separately;

[0044] A knowledge triple generation module is used to establish an entity relationship grammar template, generate a knowledge triple consisting of equipment-failure mode-root cause, and calculate the confidence of the newly generated knowledge triple;

[0045] The knowledge graph fusion module is used to detect conflicts between new triples and old knowledge. If there is a conflict, it triggers expert review. After the review is passed, the knowledge graph node attributes are updated.

[0046] Furthermore, the confidence calculation formula of the newly generated knowledge triple is:

[0047] ;

[0048] Where, Confidence represents confidence; W i represents the weight factor of data source i; C i represents the content consistency score of the triple in data source i; η consistency Represents the consistency factor among multiple sources;

[0049] Detection of conflicts between new triples and old knowledge includes: attribute value conflicts, statistical significance conflicts, logical reasoning conflicts, and spatiotemporal context conflicts.

[0050] Furthermore, the adaptive threshold management and alarm module includes:

[0051] A benchmark curve generation module is used to fit parameter change trends based on the equipment's service life;

[0052] Time fluctuation correction module, used to calculate instantaneous variance through exponentially weighted moving average;

[0053] The dynamic threshold calculation module is used to trigger the linkage control mechanism when the hazard source parameters exceed the preset threshold for three consecutive sampling cycles and the output risk level is greater than or equal to medium.

[0054] The beneficial effects of the present invention are:

[0055] 1. To address the false alarm problem caused by traditional static thresholds being susceptible to periodic interference in dynamic industrial scenarios (such as natural pressure fluctuations caused by day-night temperature differences in storage tanks), the present invention adopts a sliding window dynamic confidence interval calculation combined with a time-varying confidence coefficient adjustment mechanism. By embedding periodic correction terms (such as under pressure conditions), the threshold can be adaptively fitted to real-world disturbances. This effectively resolves the two-way contradiction between detection sensitivity and false alarm suppression, forming a complete technical closed loop from dynamic perception, cross-domain verification to rapid linkage, and systematically overcoming the long-standing key problem of achieving both accuracy and real-time performance in the field of major hazardous source risk monitoring.

[0056] 2. In view of the fact that the reliability of the original monitoring system knowledge graph is not strong enough, the present invention constructs a cognitive closed loop and introduces a physical verification mechanism based on CFD simulation to conduct a secondary verification of knowledge conflicts. When semantic reasoning determines the risk of leakage of the tank flange, the simulation engine reconstructs the three-dimensional fluid field distribution in real time, verifies the physical consistency of the pressure gradient and the leakage diffusion pattern, and corrects the confidence level of the mislabeled "suspected leakage" node in the knowledge base from 0.72 to below 0.18, significantly improving the reliability of cross-domain knowledge coupling.

[0057] 3. Based on the response delay of the original monitoring system, the present invention further combines the resource collaboration strategy of edge-side signal preprocessing and cloud-side deep analysis, deploys a lightweight AI model to achieve millisecond-level anomaly perception, and only uploads high-confidence events to the cloud to trigger the dynamic evolution of the knowledge graph. The overall data transmission volume is reduced by 73%, and the end-to-end response time of the system is compressed from 5 to 8 minutes in the traditional architecture to within 45 seconds. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0059] Figure 1 This is a system principle block diagram of a major hazard source risk dynamic calculation system based on AI large model empowerment according to an embodiment of the present invention;

[0060] Figure 2 is a data preprocessing flow chart according to an embodiment of the present invention;

[0061] Figure 3 2. It is a structural diagram of a spatiotemporal graph neural network (STGNN) according to an embodiment of the present invention;

[0062] Figure 4 2. A schematic diagram of a Bayesian network structure according to an embodiment of the present invention (including CFD module linkage);

[0063] Figure 5 is a flowchart of knowledge self-evolution according to an embodiment of the present invention;

[0064] Figure 6 is a flow chart of dynamic threshold generation according to an embodiment of the present invention;

[0065] Figure 7 4 is a timing diagram of system operation according to an embodiment of the present invention. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0067] See also Figure 1 The present invention provides a major hazard source risk dynamic calculation system based on AI large model empowerment, including: multimodal data acquisition and preprocessing module, heterogeneous data spatiotemporal fusion engine module, online incremental learning and model fine-tuning module, risk transmission probability calculation module, multimodal knowledge self-evolution module and adaptive threshold management and alarm module.

[0068] like Figure 1 As shown in Figure 2, the output ports of the multimodal data acquisition and preprocessing module include:

[0069] Port A: transmits cleaned time series data (including 9 types of parameters such as temperature and pressure) to the heterogeneous data spatiotemporal fusion engine module;

[0070] Port B: Provides labeled samples (including abnormal event labels and sensor raw values) to the online incremental learning and model fine-tuning module.

[0071] The connection methods of the heterogeneous data spatiotemporal fusion engine module include:

[0072] Bidirectional connection: Shares weight parameters with the online incremental learning and model fine-tuning modules (arrows indicate "real-time synchronization of model parameters");

[0073] Lower-level port: transmits the fused feature vector to the risk transmission probability calculation module.

[0074] The output branches of the risk transmission probability calculation module include:

[0075] Branch 1: Risk diffusion path data (including probability values) are input into the multimodal knowledge self-evolution module;

[0076] Branch 2: triggers the real-time threshold adjustment signal of the adaptive threshold management and alarm module.

[0077] Feedback path of the multimodal knowledge self-evolution module: push updated entity relationship weights to the heterogeneous data spatiotemporal fusion engine module (arrow marked "knowledge graph reverse update").

[0078] The input dependencies of the adaptive threshold management and alarm module are: receiving the real-time evaluation results of the risk transmission probability calculation module and the latest threshold rules of the multimodal knowledge self-evolution module.

[0079] The multimodal data acquisition and preprocessing module is used to connect sensors through industrial protocols to achieve synchronous acquisition of raw data. The multimodal data acquisition and preprocessing module implements the data preprocessing process as follows: Figure 2 shown.

[0080] In the description of the present invention, the multimodal data acquisition and preprocessing module includes: a sensor data acquisition module (not shown in the figure), a sliding window segmentation module (not shown in the figure), a denoising processing module (not shown in the figure), a time alignment verification module (not shown in the figure), a missing value LSTM interpolation module (not shown in the figure) and a modal normalization module (not shown in the figure).

[0081] The sensor data acquisition module is used to collect analog and digital signals from the physical device body, distinguish the input type based on the signal type judgment circuit, and forward the signals.

[0082] Specifically, the sensor data acquisition module receives analog signals (current 4-20mA) and digital signals (RS485 messages) from the device itself. It uses a signal type judgment circuit (integrated circuit model: TI ADS1115) to distinguish input types and forward them accordingly.

[0083] Sliding window segmentation is used to cut data blocks by timestamp based on the ring buffer of edge computing terminals.

[0084] Specifically, the parameter settings for sliding window segmentation include: window length = 60 seconds, step length = 10 seconds, and data reuse in overlapping areas; during the implementation of the sliding window segmentation module, it is necessary to cut data blocks by timestamp based on the ring buffer (capacity 1MB) of the edge computing terminal.

[0085] The denoising processing module is used to perform wavelet packet decomposition on the data signal to achieve high-frequency noise filtering.

[0086] Specifically, the algorithm flow of the denoising processing module includes:

[0087] 1. Perform 5-layer wavelet packet decomposition on the vibration signal (mother wavelet: Daubechies 4);

[0088] 2. Calculate the energy proportion of each sub-band. If the high-frequency band energy is less than the 5% threshold, set it to zero.

[0089] 3. Reconstruct the signal and output it to the downstream.

[0090] The time alignment verification module is used to correct the clock of each sensor. If the clock deviation exceeds the tolerance, interpolation compensation is triggered.

[0091] Specifically, the synchronization mechanism of the time alignment verification module uses the NTP protocol (Network Time Protocol) to correct the clock of each sensor with an error tolerance of ±5ms. If the clock deviation exceeds the tolerance, it is identified as an abnormality and triggers interpolation compensation (linear interpolation method).

[0092] The missing value LSTM interpolation module is used to obtain time series data collected in the past N sampling periods, predict the current missing values ​​and replace the placeholders based on the long short-term memory network.

[0093] Specifically, the network structure of the long short-term memory network is a single-layer LSTM (long short-term memory network), with the number of hidden layer neurons = 64; the input format is the data sequence of the past 10 minutes (sampling interval is 1 second); the long short-term memory network output correction is used to predict the current missing value and replace the placeholder.

[0094] The modal normalization module is used to normalize the interpolated time series data.

[0095] Specifically, the calculation formula for modal normalization is:

[0096] ;

[0097] Where μ windows represents the mean value of this window; σ base represents the baseline variance of the equipment life cycle (the initial value comes from the factory setting); X i Represents the original data.

[0098] The heterogeneous data spatiotemporal fusion engine module is used to build a spatiotemporal graph model and analyze the nonlinear coupling relationship of multi-source data.

[0099] Among them, the hardware of the heterogeneous data spatiotemporal fusion engine module relies on GPU servers (such as NVIDIA A100, 40GB video memory).

[0100] In the description of the present invention, the heterogeneous data spatiotemporal fusion engine module includes: an input module (not shown in the figure), a spatiotemporal graph construction module (not shown in the figure), an attention network module (not shown in the figure) and a risk level prediction module (not shown in the figure).

[0101] The input module (input layer) is used to receive preprocessed time series data and extract multimodal feature vectors (12-dimensional features such as temperature gradient, vibration spectrum peak, and gas concentration change rate).

[0102] The spatiotemporal graph construction module (spatiotemporal graph construction layer) is used to construct a spatiotemporal graph consisting of nodes and edges, and use a temporal convolutional network to extract the temporal dependencies of each node.

[0103] In the description of the present invention, constructing a spatiotemporal graph consisting of nodes and edges and using a temporal convolutional network to extract the temporal dependencies of each node includes:

[0104] Step S11: Set each sensor as a node and attach a device attribute label (such as material: 304 stainless steel).

[0105] Step S12: Calculate the edge weights of two nodes based on the statistical risk factor coupling strength of the historical accident database. For example, the initial weights of node i (temperature) and node j (vibration) are: ; where P casual is the joint accident probability, P imdependence is the independent probability.

[0106] Step S13: construct a temporal convolutional network based on causal dilated convolution, set network structure parameters, and use the temporal convolutional network to extract the temporal dependency of each node, where the network structure parameters include input dimension, convolution kernel diameter, dilation factor, number of output channels, and activation function.

[0107] Specifically, the temporal convolutional network (temporal convolution module) uses the TCN network (temporal convolutional network, also known as the temporal convolution module) to extract the temporal dependency of each node (convolution kernel diameter D=5).

[0108] Regarding the core structure and mathematical formula of TCN, the temporal convolution module adopts a causal dilated convolution design to ensure that the model does not leak future information while improving the ability to capture long-term temporal dependencies. The module structure includes the following core parts:

[0109] 1. Based on the characteristics of industrial time series data, the TCN parameter configuration is shown in Table 1:

[0110] Table 1: TCN parameter configuration table

[0111]

[0112] 2. Temporal dependency extraction process (forward propagation step), taking the vibration signal time series as an example (60-second window, 1 sampling point per second), its input tensor is X∈R 12x60 (12-dimensional features, such as spectral amplitude, variance, etc.) The convolution kernel includes the first layer residual block, the second layer residual block, the third layer residual block, and the fourth layer residual block.

[0113] The first residual block uses a dilated convolution with D=5, d=1, and an output dimension of 60×64. The feature extraction is to learn short-term fluctuations related to vibration intensity (e.g., 3 abnormal peaks per minute).

[0114] Second-layer residual block: dilated convolution with D=5, d=2, and output dimension 60×128; periodic features identify equipment start-stop cycles (e.g., load changes every 15 minutes);

[0115] The third residual block has a dilated convolution of D=5, d=4, and an output dimension of 60×256. The long-term dependencies are used to capture slow trend drifts caused by corrosion (e.g., a 0.1% drop in the pressure benchmark per week).

[0116] The fourth layer of residual block: the dilated convolution is D=5, d=8, and the output dimension is 60×64; the feature compression content is to retain the most discriminative temporal pattern and output to the graph attention network (GAT).

[0117] The output result is the hidden state of each time step, and HϵR12x60 describes the evolution of device states at different time scales.

[0118] In addition, the selection basis of the core diameter D=5 is based on the physical meaning that the local mode of industrial equipment failure usually appears within 5 seconds (such as the window of pressure surge and vibration mutation).

[0119] 3. Regarding TCN's suitability for industrial scenarios, its advantages over traditional RNNs / LSTMs are reflected in three key areas: parallel computing, long-term memory, and stable training. Parallel computing refers to TCN's support for a fully convolutional structure, enabling parallel processing of the entire time window, resulting in 3-5 times higher computational efficiency than LSTM. Long-term memory refers to its ability to cover a longer history with a fixed number of parameters through dilated convolutions (e.g., RF with 61 steps vs. LSTM, which struggles to capture dependencies beyond 50 steps due to vanishing gradients). Stable training eliminates the exploding / vanishing gradient issues common in LSTM, resulting in a 40% faster convergence of the loss function.

[0120] The present invention is verified by experiments. Specifically, the effects of different D values ​​on the fault detection F1-score are tested in 100 sets of sensor data, as shown in Table 2:

[0121] Table 2: Impact of different D values ​​on fault detection F1-score

[0122]

[0123] Conclusion: D=5 achieves the best balance between preserving details and suppressing noise.

[0124] The attention network module (GAT layer) is used to dynamically adjust the weights between nodes based on the attention adaptation mechanism.

[0125] In the description of the present invention, based on the attention adaptation mechanism, dynamically adjusting the weights between nodes includes:

[0126] Based on the node feature vector, learnable parameter matrix and attention vector, the attention coefficient is calculated and updated in real time through back propagation of the learnable parameter matrix and attention vector, so that the connection weights between sensor nodes can be adaptively adjusted as the distribution of sensor data changes.

[0127] In the description of the present invention, the attention coefficient calculation formula is:

[0128] ;

[0129] Where, α ij represents the connection weight between node i and node j; T represents the transposed operation matrix;

[0130] a represents the attention vector: the dimension is 2d', which is used to calculate the correlation strength between nodes;

[0131] W represents the learnable matrix: the dimension is d'xd (d'=32), which is used for feature space mapping;

[0132] h i and h j Represents the node feature vector: the feature vector of each sensor node i, with a dimension of d (the value is 64 dimensions).

[0133] In addition, the attention network module has dynamic characteristics, which is represented by the real-time update of W and a parameters through back propagation, so that the attention weights are adaptively adjusted as the distribution of sensor data changes. For example, when the correlation between temperature and vibration signals increases under high temperature conditions, α ij The value will automatically increase.

[0134] At the same time, the feature aggregation formula is:

[0135] ;

[0136] Where h i ′ represents the aggregated feature; σ represents the activation function, which is used to enhance the nonlinear expression capability; N(i) represents the neighbor set of node i, which includes sensors with spatiotemporal correlation (spatial distance < 50m and time series isotropy > 0.7).

[0137] It should be noted that feature aggregation refers to the process of fusing and compressing multi-dimensional industrial perception data at the spatiotemporal level. It includes temporal dynamic features (output of TCN (temporal convolutional network)), topological correlation features (attention coefficient α of graph attention network (GAT)), and so on. ij ), process knowledge features (knowledge graph embedding).

[0138] In addition, an example of the association between dynamic weight adjustment and risk transmission is provided. The background is a local gas leak in a chemical plant area, and the leakage diffusion path needs to be dynamically tracked. In this scenario, the dynamic weight adjustment process includes the following steps:

[0139] 1. Time T0: Gas sensor C detects that the concentration exceeds the standard, and module 2 starts attention calculation;

[0140] 2. T0+30s: Based on the data from wind speed sensor D (south wind level 3), the wind speed increases from 0.2 to 0.65, and C is strengthened in association with downwind equipment;

[0141] 3. Trigger the CFD simulation of the risk transmission probability calculation module to predict the diffusion direction of the concentration cloud map;

[0142] 4. T0+5min: The concentration of sensor E in the downwind direction increases and simultaneously rises to 0.53;

[0143] 5. The system issues an early warning: "It is expected that the concentration in area E will reach the explosion limit in 8 minutes. It is recommended to close valve V205."

[0144] Through the above design, the attention network module (GAT layer) can dynamically adjust the connection weights between sensor nodes according to real-time working conditions, effectively capture the nonlinear time-varying correlations between hazard source parameters, and solve the defect of traditional methods that rely on fixed topology.

[0145] The risk level prediction module (output layer) is used to output the risk level prediction results and trigger the adaptive adjustment of the local neuron learning rate based on the latest accumulated data.

[0146] Specifically, the output layer uses a fully connected network (number of neurons: 256→64→3), corresponding to low, medium, and high risk levels. Every time 100 new data are accumulated (approximately 10 minutes), the learning rate of local neurons (the last two layers) is adaptively adjusted (Adam optimizer) to achieve online fine-tuning of the network.

[0147] The output of the risk level prediction module can be used for a variety of purposes, including the following:

[0148] 1. Risk level as Bayesian network input: When the heterogeneous data spatiotemporal fusion engine module 2 outputs a risk level ≥ medium, the conditional probability table (CPT) of the conduction probability calculation module is dynamically updated.

[0149] For example, the prior probability of the original node "tank leak" is P(leak) = 0.05; when the risk level increases to "medium," the prior probability is revised to P(leak) = 0.15 x Phigh. Regarding the transmission path search range constraint, when the risk level is low, only the independent failure probability of the current device is evaluated (a single-node calculation); when the risk level is medium or higher, a plant-wide transmission path search is activated, and the Bayesian network traverses all relevant device nodes (such as upstream and downstream pipelines and adjacent tanks).

[0150] 2. CFD simulation parameter coupling: If the risk level is "high" and a leak scenario is involved, the CFD module will shorten the simulation time interval (from 5 seconds to 1 second) and increase the grid resolution (from 1m³ to 0.5m³) to capture the rapidly diffusing concentration gradient.

[0151] 3. Example linkage process: Assuming a reactor temperature anomaly (risk level determined to be "medium"), the system will trigger a Bayesian network update, increasing the failure probability of the temperature anomaly node and the adjacent cooling system by 50%. At the same time, CFD simulation is initiated to predict the heat diffusion area. If the simulation shows a probability of reaching the adjacent storage tank within 10 minutes greater than 40%, the attention weight for the temperature node in module 2 is increased, and the risk level is forcibly upgraded to "high", triggering an emergency stop command.

[0152] Online incremental learning and model fine-tuning module, used to trigger dynamic parameter adjustments based on real-time data streams.

[0153] In the description of the present invention, the online incremental learning and model fine-tuning module includes: an attention module insertion and channel evaluation module (not shown in the figure), a channel differentiation and sharing decision module (not shown in the figure), a parameter fine-tuning and update module (not shown in the figure), a task-specific filter extension module (not shown in the figure), a catastrophic forgetting suppression module (not shown in the figure) and a dynamic weight synchronization and conflict detection module (not shown in the figure).

[0154] The attention module insertion and channel evaluation module is used to insert the attention module into the neural network, generate statistics through channel importance evaluation, and distinguish shared channels from auxiliary channels to solve the problem of task feature conflict in incremental learning and prevent new task training from interfering with the key channels of old tasks.

[0155] Specifically, the attention module insertion and channel evaluation module need to insert the channel attention module after the temporal convolutional network (TCN) layer. The structure is as follows:

[0156] 1. Input feature map: the output from the TCN layer (dimension is [Batch, Channels, Time]).

[0157] 2. Global average pooling: compress along the time dimension into a channel description vector (dimension [Batch, Channels]).

[0158] 3. Fully connected layer: The first layer compresses the number of channels to 1 / 16 (e.g. 64 channels → 4 nodes), and the second layer restores the number of channels to the original number. The activation function is Sigmoid.

[0159] 4. Output layer: Generates a channel weight vector (dimension [Batch, Channels]) to indicate the importance of each channel.

[0160] To generate channel importance statistics, the network with the attention module inserted is trained using new task samples (data labeled with port B from the multimodal data acquisition module) with a cross-entropy loss function, implementing bare-net training. The mean channel weights of all training samples are then calculated and sorted from high to low, with the top 30% of channels selected as candidate shared channels. Finally, the channel importance ranking results are stored in an in-memory database (such as Redis) for use by the channel differentiation module.

[0161] The channel differentiation and sharing decision module is used to dynamically divide shared channels (shared by multiple tasks) and auxiliary channels (task-specific) based on the intersection of the importance of new and old task channels, achieving a balance between parameter reuse and decoupling of task-specific features.

[0162] The input is the candidate shared channels for the new task (e.g., the channel index list {1, 3, 5, 7}) and the historical shared channels for the old task (obtained through the reverse update interface of the multimodal knowledge self-evolution module, e.g., {3, 5, 9, 11}). The intersection of the two ({3, 5}) is calculated to define the global shared channels; the non-intersecting portions ({1, 7} and {9, 11}) are classified as auxiliary channels for the new and old tasks, respectively.

[0163] All tasks share the same set of parameters, allowing for bidirectional synchronous updates (real-time synchronization with the weights of the spatiotemporal fusion engine module). Auxiliary channels are established, including those for old tasks and new tasks. The old task auxiliary channel is used to freeze parameters (forward propagation only, backward updates are prohibited); the new task auxiliary channel is used to allocate an independent parameter space, initialized to a random normal distribution.

[0164] By maintaining a global channel mapping table with fields including channel index, role (shared / auxiliary), task ID, and last update time, it receives reverse weight update signals from the multimodal knowledge self-evolution module and dynamically adjusts channel allocation (for example, recalculating intersections when entity relationship weights change beyond a threshold).

[0165] The parameter fine-tuning and updating module is used to update the parameters in a fine-tuning manner on the shared channel, while freezing the old task parameters of the auxiliary channel to ensure the learning efficiency of the new task and avoid performance degradation of the old task.

[0166] The parameter fine-tuning and update module updates parameters on shared channels using low-rank fine-tuning to ensure performance for new tasks while maintaining compatibility with existing tasks. Through low-rank adaptation, the convolutional layer weight matrices of the shared channels are decomposed into low-rank components. During training, only matrices A and B are optimized, while the original weights W are frozen. This reduces the number of trainable parameters to 0.1%-1% of the full parameter fine-tuning requirement.

[0167] By establishing elastic weight constraints and based on the Fisher information matrix of old task data, the importance of shared channel parameters to historical tasks is quantified, and a regularization term is added to the loss function to limit the update amplitude of important parameters (for example, the update amplitude of parameters with Fisher value > 0.5 is limited to ±0.01).

[0168] After each incremental training is completed, the updated A and B matrices are synchronized to the TCN layer of the spatiotemporal fusion engine module through bidirectional connections to ensure that the risk prediction model is updated in real time.

[0169] The task-specific filter extension module is used to add parallel filters for specific tasks on auxiliary channels to improve the ability to extract task-specific features and solve the problem of insufficient feature expression capabilities of shared channels.

[0170] Specifically, the task-specific filter extension module adds dedicated filters for newly added tasks to the auxiliary channel, improving task-specific feature extraction. By adding parallel 1x1 convolution branches after the TCN layer in the auxiliary channel, each new task is assigned an independent filter bank (e.g., four filter banks). The filter weights are initialized using the He normal distribution, and the bias term is initialized to 0.

[0171] A binary mask matrix is ​​generated based on the task identifier to limit the scope of gradient propagation. For example, the loss of task A is backpropagated only to its dedicated filter parameters, and the gradient of the filter parameters of task B is set to zero. Finally, the shared channel features are concatenated with the dedicated filter outputs of the auxiliary channels along the channel dimension and input into the subsequent fully connected layer for joint decision making.

[0172] The catastrophic forgetting suppression module is used to prevent the performance degradation of old tasks caused by the update of shared channel parameters, so as to maintain the recognition stability of the model for historical tasks.

[0173] The catastrophic forgetting suppression module obtains historically labeled samples (such as leak event data) from port B of the multimodal data acquisition module. K-Means clustering is used to select representative samples (cluster centers) and store them in a buffer. During each incremental training, new and old samples are mixed proportionally (25% old samples + 75% new samples) to form a training batch.

[0174] The dynamic weight synchronization and conflict detection module is used to manage the synchronous update of shared channel parameters among multiple tasks, detect parameter conflicts and trigger repairs to ensure the compatibility of multi-task parameter updates.

[0175] Specifically, each task independently calculates the gradient of the shared channel, assigns weights according to the proportion of task loss (for example, if the loss of task A accounts for 60%, the gradient weight is 0.6), and updates the global parameters after weighted averaging. The similarity of the gradient directions of the shared channel between the new and old tasks is monitored. If the similarity is <0.5, it is determined to be a direction conflict, triggering the following operations: roll back to the previous stable parameter version (loaded from the version snapshot); dynamically reduce the learning rate (for example, from 1e-4 to 5e-5).

[0176] After every 100 training iterations, a snapshot of the model parameters and optimizer state (momentum, learning rate) is saved to the non-volatile memory of the edge computing node to support fast recovery.

[0177] The risk transmission probability calculation module is used to combine Bayesian networks and CFD simulation to generate risk diffusion paths.

[0178] In the description of the present invention, the risk transmission probability calculation module includes: a Bayesian network reasoning engine module (not shown in the figure), a CFD physical simulation engine module (not shown in the figure) and a risk transmission path integration module (not shown in the figure).

[0179] The Bayesian network inference engine module is used to build a Bayesian network using physical devices and environmental states. It uses the acquired real-time sensor data to regularly revise the joint probability distribution in the conditional probability table and calculate the risk probability corresponding to each physical device node.

[0180] Specifically, the Bayesian network structure diagram (including CFD module linkage) is as follows Figure 4 As shown in Figure 2, the structure and content of the Bayesian network inference engine include the following aspects:

[0181] 1. Equipment node definition: Physical entities such as storage tank V101, pump P203, pipeline L305, etc. constitute network nodes.

[0182] 2. Conditional Probability Table (CPT): Initialized based on historical data statistics and expert experience. For example, when the tank pressure exceeds the limit (V101_Fault=1) and the temperature is greater than 60°C (V101_Temp=High), the explosion probability of the adjacent pipeline L305 is 0.78.

[0183] 3. Dynamic update mechanism: The joint probability distribution in CPT is revised every hour based on real-time data.

[0184] For initial risk identification in the Bayesian network, the input is real-time sensor data (pressure, temperature, gas concentration, etc.); the Bayesian network structure includes: Bayesian network nodes represent physical equipment or environmental status (such as storage tank V101, pump P203, wind speed direction, etc.).

[0185] It should be noted that the key operations of the Bayesian network inference engine include the following:

[0186] 1. Dynamic update of the conditional probability table (CPT), where the update formula of the conditional probability table is:

[0187] ;

[0188] Where, represents the updated conditional probability; represents the prior conditional probability; α represents the historical experience weight (default 0.7, adjustable); I(·) represents the indicator function (1 if the condition is met, 0 otherwise); T represents the statistical window length (such as the data for the last seven days); X i Represents the current device node; t represents the observation point index of the time window; Pa (X i ) indicates X i The parent node collection.

[0189] 2. Risk propagation path search. This algorithm uses depth-first search (DFS) to traverse the sub-nodes of the Bayesian network to generate potential impact chains; its path sequence is S=〔V101→P203→L305〕.

[0190] The CFD physical simulation engine module is used to start the CFD simulation engine when the risk probability of any physical device node exceeds a preset threshold, and simulate and output the spatiotemporal distribution of the combustible gas diffusion concentration cloud map.

[0191] Specifically, the input variables of the CFD physical simulation engine include: leak location, wind speed, and terrain data (GIS elevation model); the output result is the spatiotemporal distribution of the combustible gas diffusion concentration cloud map (updated at 5-second intervals); the simulation process requires the use of the finite volume method (FVM) to solve the NS equations and calculate the turbulent diffusion model (k-ε model).

[0192] The CFD physics simulation engine is triggered when the risk probability of any equipment node exceeds a threshold. The simulation process includes 3D modeling, meshing, and mesh generation. An unstructured tetrahedral mesh (average edge length 0.5m) is used, with local meshing (0.1m) near the leak point. The input requires factory GIS information, equipment 3D models (STEP format), and environmental data (wind speed, temperature gradient).

[0193] During the simulation process, the control equations are solved, such as the diffusion model, including the coupled mass conservation and momentum equations, taking into account the turbulence effect (k-ε model), and the formula is:

[0194] ;

[0195] Where, Indicates gas concentration (kg / m 3 ); represents the turbulent diffusion coefficient; represents the leakage source term (with the leakage rate Q leak related).

[0196] It should be noted that the CFD physics simulation engine uses the finite volume method (FVM) to discretize equations and the SIMPLEC algorithm to solve the velocity-pressure coupled field. Its output for the time-varying concentration field is the time step, such as Δt = 0.1 seconds (satisfying the CFL condition). The output data also includes a record of the gas concentration distribution across the plant every 5 seconds. .

[0197] The risk transmission path integration module is used to mark the risk path of combustible gas leakage based on the concentration field data and equipment topology diagram output by the CFD simulation engine.

[0198] Specifically, the input data of the risk transmission path integration module includes CFD concentration field data and equipment topology diagram (adjacency matrix A n×n The risk transmission path integration module needs to realize the transmission probability calculation and identification through the integration algorithm, including the following contents:

[0199] 1. For explosion hazard area marking, the criteria for determining the area marking are: (LEL is the lower explosion limit, such as methane LEL = 5% volume concentration);

[0200] Dangerous area set .

[0201] 2. Cascade path probability calculation. During the calculation process, the topology impact weight includes the physical connection weight between devices (such as pipeline length L ij , circulationQ ij ), the calculation formula is:

[0202] ;

[0203] Explosion conduction probability: If device i is in hazardous area D risk , then the probability that the device is affected The calculation formula is:

[0204] ;

[0205] in, is the failure probability of upstream device i.

[0206] 3. In the process of time window fusion, it is necessary to establish cascade time constraints and set the safety response time of the device T j (For example, pump P203 must be shut down within 2 minutes after gas coverage). If CFD predicts the gas arrival time T arrive < T j , the path risk level increases by one level.

[0207] Final path probability The calculation formula is:

[0208] ;

[0209] in, P source Initial leakage probability; I (·) represents the temporal compliance indicator function (1 if compliant, 0 otherwise).

[0210] When the Bayesian network detects a tank leak probability > 0.5, it triggers the CFD module to generate a gas diffusion prediction (e.g., area coverage after 5 minutes). If the concentration exceeds 60% of the lower explosion limit (LEL), it is marked as a high-risk path.

[0211] The multimodal knowledge self-evolution module is used to continuously update the risk knowledge base using large model technology.

[0212] Among them, the multimodal knowledge self-evolution module runs on a distributed cluster (Hadoop+Spark architecture).

[0213] In the description of the present invention, the multimodal knowledge self-evolution module includes: a multimodal input parsing module (not shown in the figure), a knowledge triple generation module (not shown in the figure) and a knowledge graph fusion module (not shown in the figure).

[0214] Multimodal input parsing module, used to parse structured data and unstructured data separately.

[0215] Specifically, structured data: sensor log (CSV format) → entity extraction (device ID, alarm code);

[0216] Unstructured data: Maintenance report (PDF) → BERT model extracts "faulty component: pump bearing"; maintenance video → YOLOv5 identifies "flange rust area ≥ 30%."

[0217] The knowledge triple generation module is used to establish an entity relationship grammar template, generate knowledge triples consisting of equipment-failure mode-root cause, and calculate the confidence of the newly generated knowledge triples.

[0218] Specifically, the entity relationship syntax template is: <equipment, failure mode, root cause> (such as <reactor R201, seal failure, chloride ion stress corrosion>);

[0219] It should be noted that the confidence calculation for knowledge triples is based on the source reliability weight (0.9 for sensor data vs. 0.7 for manual reports). The weighting factors and confidence levels include the following:

[0220] 1. Definition and quantification of weighting factors. A portion of this data comes from sensor data (weight = 0.9). First, the sensor data's error rate must be verified. This involves analyzing and labeling 10,000 hours of historical sensor data. The standard deviation (σsensor) of the temperature / pressure sensor and the average deviation from the manually calibrated value is ±0.5%, and the false alarm rate (FP Rate) is approximately 2%. The drift ratio for 30 days of continuous operation is ≤0.1% / day (tested according to GB / T 25919-2022). This ensures the stability of the sensor data.

[0221] The weight calculation formula for sensor data is:

[0222] ;

[0223] Another portion of the data comes from manual reports (weight = 0.7). We sampled 500 historical maintenance reports and compared them with the post-fault tracing results. The accuracy of key fields (such as the faulty component and cause) was 68%, which was used to calculate the sampling accuracy. The median report delay time (Tdelay) was 48 hours. Based on the logistic function attenuation factor, the calculation formula is:

[0224] ;

[0225] Finally, the comprehensive weight is calculated. , normalized to the interval [0, 1], and then revised upward to 0.7 (compensated by expert experience). In addition, due to the ambiguity of descriptions in manual reports (such as "suspected leakage" is quantified), its basic weight needs to be compensated for the scenario through expert experience.

[0226] 2. Confidence synthesis formula. The confidence calculation formula of the newly generated knowledge triple is:

[0227] ;

[0228] Where, Confidence represents confidence; W i represents the weight factor of data source i; C i represents the content consistency score of the triple in data source i; η consistency represents the consistency factor among multiple sources (if both the sensor and the manual report mention the same cause, η = 1.2; otherwise η = 0.8).

[0229] The knowledge graph fusion module is used to detect conflicts between new triples and old knowledge. If there is a conflict, it triggers expert review. After the review is passed, the knowledge graph node attributes are updated.

[0230] The conflicts between new triples and old knowledge include attribute value conflicts, statistical significance conflicts, logical reasoning conflicts, and spatiotemporal context conflicts. The specific contents are as follows:

[0231] 1. Attribute value conflict: The numerical attribute exceeds the allowable error range or the categorical attribute does not match;

[0232] 2. Statistical significance conflict: Use chi-square test or t-test to determine the difference in the distribution of new and old data;

[0233] 3. Logical reasoning conflicts: Reasoning based on ontology (OWL) axioms to detect logical contradictions after triples are added;

[0234] 4. Conflict of time and space context: The time and space conditions (such as temperature and pressure ranges) are not consistent, and the conclusions under the intersection conditions are contradictory.

[0235] Specifically, the knowledge self-evolution process is as follows Figure 5 As shown in the figure, if the new triplet conflicts with the old knowledge (for example, the new data shows that "the corrosion rate of 304 stainless steel increases by 2 times when Cl⁻>200ppm"), an expert review is triggered (the web interface pushes the pending confirmation item); after the review is passed, the graph node attributes are updated and fed back to the graph model weight calculator of the heterogeneous data spatiotemporal fusion engine module.

[0236] Adaptive threshold management and alarm module, used to dynamically optimize alarm rules and output emergency decisions.

[0237] In the description of the present invention, the adaptive threshold management and alarm module includes: a reference curve generation module (not shown in the figure), a time fluctuation correction module (not shown in the figure) and a dynamic threshold calculation module (not shown in the figure).

[0238] The benchmark curve generation module is used to fit the parameter change trend according to the service life of the equipment.

[0239] The time fluctuation correction module is used to calculate the instantaneous variance through exponentially weighted moving average.

[0240] The dynamic threshold calculation module is used to trigger the linkage control mechanism when the hazard source parameters exceed the preset threshold for three consecutive sampling cycles and the output risk level is greater than or equal to medium.

[0241] Specifically, the adaptive threshold management and alarm module is deployed at the edge (implemented by low-power FPGA), such as Figure 6 As shown in the figure, the key steps of the adaptive threshold management and alarm module include:

[0242] 1. Baseline curve generation, fitting parameter change trends based on equipment service time (months) (cubic polynomial regression);

[0243] 2. Time fluctuation correction: calculate the instantaneous variance through the exponentially weighted moving average (EWMA). The calculation formula is:

[0244] ;

[0245] Where, λ=0.85, μ t is a sliding mean window (24 hours); λ represents the attenuation factor, with a value range of 0≤λ<1. It is used to control the weight attenuation rate of historical data. The larger the value, the more lasting the impact of historical data and the lower the sensitivity (industrial default value: λ=0.9); t The exponentially weighted variance estimate at time t is used to quantify the volatility of the current system state, such as the variance of the corrosion rate or the fluctuation amplitude of the gas concentration. t-1 Represents the exponentially weighted variance estimate at time t-1, which is used as the historical estimate of the variance at the previous time, for smoothing iterative calculations. Represents the observation value at time t, which is used for physical quantities collected by the sensor in real time (such as temperature, vibration frequency, and chloride ion concentration); The mean estimate at time t is usually updated by EWMA: , used to dynamically characterize the system steady-state benchmark (such as "average corrosion rate" or "normal operating pressure") and support anomaly detection.

[0246] 3. Dynamic threshold calculation, the definition of the safety boundary is:

[0247] \text{Threshold}_{\text{upper}}=\mu_t+3\times\sigma_t\times\gamma_{\text{age_factor}};

[0248] The aging factor is defined as:

[0249] \gamma_{\text{age_factor}}=1+0.05\times\text{number of years in operation};

[0250] The alarm logic is that when the hazardous source parameter exceeds 3 consecutive sampling periods, Threshold upper , and when the output risk level of module 2 is ≥ medium, the linkage control is triggered (such as closing the upstream valve and starting exhaust).

[0251] It should be noted that the workflow sequence diagram of this system is as follows Figure 7 shown.

[0252] It should be noted that the workflow sequence diagram of this system is as follows Figure 7 shown.

[0253] Taking a tank leakage incident as an example, the collaboration process of each module is as follows:

[0254] 1. Time T0: The pressure sensor detects a sudden change in pressure (increase rate > 10 kPa / s), and the pre-processing module marks the abnormality;

[0255] 2. T0+2s: Heterogeneous data spatiotemporal fusion engine module 2 determines that the risk level has risen to "medium", and module 4 initiates the Bayesian network to query related devices;

[0256] 3. T0+5s: The historical failure rate of adjacent pressure relief valves is found to be 20%, triggering CFD simulation to predict the diffusion path;

[0257] 4. T0+30s: Multimodal knowledge self-evolution module 5 updates the confidence of the triple “tank overpressure causes flange leakage” in the knowledge graph;

[0258] 5. T0+60s: The adaptive threshold management and alarm module 6 generates an adjustment instruction based on the new threshold, and the platform pops up a window prompting "Initiate regional evacuation within 5 minutes."

[0259] The following is a supplementary explanation based on the technical architecture and core content of this system.

[0260] 1. Regarding the technical architecture of this system, the overall architecture diagram is divided into edge perception layer → local computing layer → cloud knowledge layer → physical verification layer; the key data flow is marked as pressure sensor → sliding window extraction rate parameter → dynamic threshold comparison → confidence annotation event; event triggering → knowledge graph leakage risk reasoning → CFD / PINN physical field verification → operation instruction issuance.

[0261] In addition, regarding the module interface protocol, the input of the dynamic threshold generation module of the present invention is {sensor ID, timestamp, pressure value, temperature compensation coefficient, confidence adjustment factor k ( t )}; Output is {current window threshold lower limit L ( t ), upper limit H ( t ), confidence level (high / medium / low)}.

[0262] 2. Core technical details of dynamic threshold generation. For the calculation of sliding window parameters, the formula expression of the sliding window parameters is:

[0263] ;

[0264] Where, k 0 represents the basic confidence coefficient (default is 2.5, adjustment range is 1.8~3.2); T Indicates the period corresponding to periodic environmental disturbances (e.g. the period of day and night temperature difference is set to 24 hours); σ windows represents the characteristic standard deviation within the window; α represents the risk confidence coefficient.

[0265] In addition, the details of the machine learning alternative are specifically reflected in the following aspects: 1. LSTM model structure, input layer → 3-layer LSTM (hidden modules 128 → 64 → 32) → fully connected layer (output predicted pressure change rate); 2. Training data characteristics, time series characteristics are the mean, variance, and autocorrelation coefficient of pressure in the past hour; environmental characteristics are temperature (EMA smoothing), equipment operating time, and pipeline flow rate; 3. Dynamic threshold generation logic, with the 90% confidence interval of the predicted value (quantile regression method) as the upper and lower limits of the threshold.

[0266] To sum up, with the help of the above technical scheme of the present invention, in order to solve the problem of false alarm caused by periodic interference (such as natural pressure fluctuation caused by day and night temperature difference of storage tanks) of traditional static thresholds in dynamic industrial scenarios, the present invention adopts sliding window dynamic confidence interval calculation combined with time-varying confidence coefficient adjustment mechanism, and realizes adaptive fitting of thresholds to real environmental disturbances by embedding periodic correction terms (such as under pressure conditions), effectively solving the two-way contradiction between detection sensitivity and false alarm suppression, forming a complete technical closed loop from dynamic perception, cross-domain verification to rapid linkage, and systematically overcoming the key problem of long-standing accuracy and real-time performance in the field of major hazardous source risk monitoring. In response to the fact that the reliability of the knowledge graph of the original monitoring system is not strong enough, the present invention constructs a cognitive closed loop and introduces a physical verification mechanism based on CFD simulation to conduct a secondary verification of knowledge conflicts. When semantic reasoning determines the risk of leakage of the tank flange, the simulation engine reconstructs the three-dimensional fluid field distribution in real time, verifies the physical consistency of the pressure gradient and the leakage diffusion pattern, and corrects the confidence level of the mislabeled "suspected leakage" nodes in the knowledge base from 0.72 to below 0.18, significantly improving the reliability of cross-domain knowledge coupling. In response to the response delay of the original monitoring system, the present invention further combines the resource collaboration strategy of edge-side signal preprocessing and cloud-side deep analysis, deploys a lightweight AI model to achieve millisecond-level anomaly perception, and only uploads high-confidence events to the cloud to trigger the dynamic evolution of the knowledge graph. The overall data transmission volume is reduced by 73%, and the end-to-end response time of the system is compressed from 5 to 8 minutes in the traditional architecture to within 45 seconds.

[0267] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown in sequence as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a portion of the sub-steps or stages of other steps.

Claims

1. A major hazard source risk dynamic calculation system based on AI large model empowerment, characterized by: include: Multimodal data acquisition and preprocessing module, used to connect sensors through industrial protocols to achieve synchronous acquisition of raw data; Heterogeneous data spatiotemporal fusion engine module, used to build spatiotemporal graph models and analyze nonlinear coupling relationships between multi-source data; Online incremental learning and model fine-tuning module, used to trigger dynamic parameter adjustments based on real-time data streams; Risk transmission probability calculation module, used to combine Bayesian networks and CFD simulation to generate risk diffusion paths; Multimodal knowledge self-evolution module, used to continuously update the risk knowledge base using large model technology; Adaptive threshold management and alarm module, used to dynamically optimize alarm rules and output emergency decisions; The heterogeneous data spatiotemporal fusion engine module includes: The spatiotemporal graph construction module is used to construct a spatiotemporal graph consisting of nodes and edges, and use a temporal convolutional network to extract the temporal dependencies of each node; Attention network module, which is used to dynamically adjust the weights between nodes based on the attention adaptation mechanism; The risk level prediction module is used to output the risk level prediction results and trigger the adaptive adjustment of the local neuron learning rate based on the latest accumulated data; The dynamic adjustment of inter-node weights based on the attention adaptive mechanism includes: Based on the node feature vector, learnable parameter matrix and attention vector, the attention coefficient is calculated and updated in real time through back propagation of the learnable parameter matrix and attention vector, so that the connection weights between sensor nodes can be adaptively adjusted as the distribution of sensor data changes.

2. According to claim 1, a dynamic calculation system for major hazard source risks based on AI large model empowerment is characterized by: The multimodal data acquisition and preprocessing module includes: The sensor data acquisition module is used to collect analog and digital signals from the physical device itself, distinguish the input type based on the signal type judgment circuit, and then divert and forward the signals; Sliding window segmentation module, which is used to cut data blocks by timestamp based on the ring buffer of edge computing terminals; Denoising processing module, used to perform wavelet packet decomposition on data signals to achieve high-frequency noise filtering; The time alignment module is used to calibrate the clocks of each sensor. If the clock deviation exceeds the tolerance, interpolation compensation is triggered. The missing value LSTM interpolation module is used to obtain time series data collected in the past N sampling periods, predict the current missing values ​​and replace the placeholders based on the long short-term memory network; The modal normalization module is used to normalize the interpolated time series data.

3. The system for dynamic calculation of major hazard source risks based on AI large model empowerment according to claim 1 is characterized in that: The heterogeneous data spatiotemporal fusion engine module also includes: The input module is used to receive preprocessed time series data and extract multimodal feature vectors.

4. The system for dynamic calculation of major hazard source risks based on AI large model empowerment according to claim 3 is characterized in that: The construction of a spatiotemporal graph consisting of nodes and edges and the use of a temporal convolutional network to extract the temporal dependencies of each node include: Set each sensor as a node with device attribute labels; Based on the historical accident database, the coupling strength of risk factors is statistically analyzed and the edge weights of the two nodes are calculated; A temporal convolutional network is constructed based on causal dilated convolution, and the network structure parameters are set. The temporal convolutional network is used to extract the temporal dependency of each node, wherein the network structure parameters include input dimension, convolution kernel diameter, dilation factor, number of output channels and activation function.

5. The system for dynamic calculation of major hazard source risks based on AI large model empowerment according to claim 4 is characterized in that: The attention coefficient calculation formula is: α ij =softmax(LeakyReLU(a T [Wh i ||Wh j ])); Where, α ij represents the connection weight between node i and node j; a represents the attention vector; W represents the learnable matrix; h i and h j represents the node feature vector; T represents the transposed operation matrix.

6. The system for dynamic calculation of major hazard source risks based on AI large model empowerment according to claim 1 is characterized in that: The risk transmission probability calculation module includes: The Bayesian network inference engine module is used to build a Bayesian network using physical devices and environmental states. It uses the acquired real-time sensor data to regularly revise the joint probability distribution in the conditional probability table and calculate the risk probability corresponding to each physical device node. The CFD physical simulation engine module is used to start the CFD simulation engine when the risk probability of any physical device node exceeds the preset threshold, and simulate and output the spatiotemporal distribution of the combustible gas diffusion concentration cloud map; The risk transmission path integration module is used to mark the risk path of combustible gas leakage based on the concentration field data and equipment topology diagram output by the CFD simulation engine.

7. The system for dynamic calculation of major hazard source risks based on AI large model empowerment according to claim 1 is characterized in that: The multimodal knowledge self-evolution module includes: Multimodal input parsing module, used to parse structured data and unstructured data separately; A knowledge triple generation module is used to establish an entity relationship grammar template, generate a knowledge triple consisting of equipment-failure mode-root cause, and calculate the confidence of the newly generated knowledge triple; The knowledge graph fusion module is used to detect conflicts between new triples and old knowledge. If there is a conflict, it triggers expert review. After the review is passed, the knowledge graph node attributes are updated.

8. The system for dynamic calculation of major hazard source risks based on AI large model empowerment according to claim 7 is characterized in that: The calculation formula for the confidence of the newly generated knowledge triple is: Where, Confidence represents confidence; W i represents the weight factor of data source i; C i represents the content consistency score of the triple in data source i; η consistency Represents the consistency factor among multiple sources; The conflicts detected between the new triples and the old knowledge include: attribute value conflicts, statistical significance conflicts, logical reasoning conflicts and spatiotemporal context conflicts.

9. The system for dynamic calculation of major hazard source risks based on AI large model empowerment according to claim 1 is characterized in that: The adaptive threshold management and alarm module includes: A benchmark curve generation module is used to fit parameter change trends based on the equipment's service life; Time fluctuation correction module, used to calculate instantaneous variance through exponentially weighted moving average; The dynamic threshold calculation module is used to trigger the linkage control mechanism when the hazard source parameters exceed the preset threshold for three consecutive sampling cycles and the output risk level is greater than or equal to medium.

Images