Dynamic calculation system for risk of major hazard source based on AI large model enabling

Abstract

The invention discloses a major hazard source risk dynamic calculation system based on AI large model enabling, and relates to the technical field of risk calculation, and the system comprises a multi-modal data collection and preprocessing module which is used for achieving the synchronous collection of original data through the butt joint of an industrial protocol with a sensor; the heterogeneous data space-time fusion engine module is used for constructing a space-time diagram model and analyzing a nonlinear coupling relationship of multi-source data; the online incremental learning and model fine tuning module is used for triggering dynamic parameter adjustment based on the real-time data flow; a risk conduction probability calculation module; a multi-modal knowledge self-evolution module; and a self-adaptive threshold management and alarm module. According to the method, sliding window dynamic confidence interval calculation is combined with a time-varying confidence coefficient adjustment mechanism, self-adaptive fitting of a threshold value to real environment disturbance is achieved by embedding a periodic correction term, the bidirectional contradiction between detection sensitivity and false alarm suppression is effectively cracked, and a complete technical closed loop from dynamic sensing and cross-domain verification to rapid linkage is formed.

Description

Technical Field

[0001] The present invention relates to the technical field of risk calculation, and more particularly to a dynamic risk calculation system for major hazard sources empowered by an AI large model. Background Art

[0002] The risk assessment technology for major hazard sources has gradually shifted from traditional manual experience judgment to intelligent methods. Typical technologies include: 1. Risk analysis models: The industry widely uses static models such as the Analytic Hierarchy Process (AHP), Fault Tree Analysis (FTA), and Event Tree Analysis (ETA). The risk weights are determined by expert scoring, and risk assessment is carried out in combination with historical accident statistics. Such methods rely on a priori rule bases and have weak modeling capabilities for the dynamic evolution process of complex systems. 2. Sensor monitoring and threshold warning: By deploying single-point sensors such as temperature, pressure, and gas concentration, and combining with the PLC / DCS system to give real-time alarms for over-limit parameters. Typical examples include the pressure safety interlock of storage tanks and the ventilation system triggered by the LEL concentration threshold of toxic gases. Such technologies only focus on the instantaneous anomalies of single parameters and do not associate with the potential coupling risks of multi-source data. 3. Application of early prediction algorithms: Some improved solutions introduce machine learning models such as time series analysis (such as ARIMA prediction), Support Vector Machine (SVM), and random forest to predict the probability of equipment failure based on historical training data. However, limited by the low dimensionality of model input features and insufficient generalization ability, it is difficult to meet the requirements of multi-modal data fusion and rapid migration in industrial scenarios. 4. Industry standard frameworks: Internationally common standards and 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 dynamically correct the results in real-time by associating with actual operation data.

[0003] The existing technologies generally show characteristics such as static rule dominance, local single-point analysis, and lack of cross-domain risk conduction mechanisms, resulting in significant bottlenecks in the timeliness and accuracy of dynamic risk assessment for complex hazard sources. Therefore, there are technical problems in the existing risk assessment methods for major hazard sources, including the following aspects:

[0004] 1. The lag of traditional static risk assessment is serious: The existing technologies mostly rely on manual regular inspections and fixed threshold judgments. The data update cycle is long (usually in weeks / months), and it is impossible to perceive the dynamic changes of hazard sources in real-time (such as instantaneous fluctuations in storage tank pressure, cumulative trends of gas concentration, and accelerated corrosion rate of equipment), resulting in delayed risk warnings and difficulty in promptly responding to safety hazards that rapidly evolve in a short period.

[0005] 2. Insufficient analysis capabilities for coupling of multi-source risk factors: Traditional methods usually use linear models or independent weight superposition methods, which make it difficult to effectively quantify the nonlinear interactive relationships among multi-dimensional heterogeneous data such as environmental parameters (temperature, humidity, wind speed), equipment status (vibration, temperature), and human operation records. In particular, they lack modeling capabilities for complex coupling risks such as the superposition of equipment aging and extreme weather, process anomalies, and chain reactions of 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 mostly customized for specific scenarios (such as chemical parks, oil and gas storage), relying on expert experience to set the rule base. When faced with process changes, new types of hazardous sources, or regional environmental differences, a lot of manpower is required to readjust parameters, making it difficult to quickly adapt to new scenarios. In addition, traditional models are sensitive to the density and quality of historical data, and the prediction reliability drops sharply when data is sparse (such as new equipment) and labels are missing (such as unrecorded small accidents).

[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 leakage leading to chain explosions, and the probabilistic transmission chain of equipment failures triggering safety system malfunctions). This makes it difficult to predict the cascading effects of major risks, affecting the precise scheduling of prevention and control resources.

[0008] 5. Rigid alarm rules and high false alarm rate: Traditional methods trigger alarms based on fixed thresholds or simple statistical models (such as the 3σ principle). They cannot dynamically adapt to operating condition fluctuations (such as normal equipment deformation deviations caused by day and night temperature differences) and data quality disturbances (such as temporary sensor drift). They frequently generate false alarms (such as normal maintenance operations being misjudged 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 model empowerment to address the above technical issues.

[0010] The present invention provides a major hazard source risk dynamic calculation system based on AI big 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 a spatiotemporal graph model and analyze the nonlinear coupling relationship of 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] A risk conduction probability calculation module, which is used to generate risk diffusion paths by combining Bayesian networks and CFD simulations;

[0015] A multi-modal knowledge self-evolution module, which is used to continuously update the risk knowledge base using large model technology;

[0016] An adaptive threshold management and alarm module, which is used to dynamically optimize alarm rules and output emergency decisions.

[0017] Furthermore, the multi-modal data acquisition and preprocessing module includes:

[0018] A sensor data acquisition module, which is used to acquire analog and digital signals of the physical device body, judge the circuit according to the signal type to distinguish the input type and shunt and forward;

[0019] A sliding window segmentation module, which is used to cut data blocks based on the circular buffer of the edge computing terminal according to timestamps;

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

[0021] A time alignment verification module, which is used to correct the clocks of each sensor. If the clock deviation exceeds the tolerance, interpolation compensation is triggered;

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

[0023] A modal normalization module, which is used to perform normalization processing on the interpolated time series data.

[0024] Furthermore, the heterogeneous data spatio-temporal fusion engine module includes:

[0025] An input module, which is used to receive the preprocessed time series data and extract multi-modal feature vectors;

[0026] A spatio-temporal graph construction module, which is used to construct a spatio-temporal graph composed of nodes and edges, and use a temporal convolutional network to extract the temporal dependencies of each node;

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

[0028] A risk level prediction module, which is used to output risk level prediction results, and trigger adaptive adjustment of the local neuron learning rate according to the latest accumulated data.

[0029] Furthermore, constructing a spatio-temporal graph composed of nodes and edges and using a temporal convolutional network to extract the temporal dependencies of each node includes:

[0030] Each sensor is set as a node and is attached with device attribute labels;

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

[0032] Based on causal dilated convolution, a temporal convolutional network is constructed, the network structure parameters are set, and the temporal dependencies of each node are extracted by using the temporal convolutional network. Among them, the network structure parameters include the input dimension, the convolution kernel diameter, the dilation factor, the number of output channels, and the activation function.

[0033] Furthermore, based on the attention adaptive mechanism, dynamically adjusting the weights between nodes includes:

[0034] Based on the node feature vectors, the learnable parameter matrix, and the attention vectors, the attention coefficients are calculated and are updated in real time through the backpropagation of the learnable parameter matrix and the attention vectors, so that the connection weights between sensor nodes are adaptively adjusted following the changes in the sensor data distribution.

[0035] Furthermore, the formula for calculating the attention coefficient 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 represent the node feature vectors; T represents the transpose operation matrix.

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

[0039] A Bayesian network inference engine module, which is used to construct a Bayesian network by using the physical devices and the environmental states, and to regularly correct the joint probability distribution in the conditional probability table by using the acquired real-time sensor data, and calculate the risk probabilities corresponding to each physical device node;

[0040] A CFD physical simulation engine module, which 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 spatio-temporal distribution of the combustible gas diffusion concentration cloud map;

[0041] A risk conduction path integration module, which is used to mark the risk paths of combustible gas leakage based on the concentration field data output by the CFD simulation engine and the device topology diagram.

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

[0043] A multi-modal input parsing module, which is used to parse structured data and unstructured data respectively;

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

[0045] A knowledge graph fusion module, which is used to detect the conflict between new triples and old knowledge. If there is a conflict, it will trigger expert review. After the review is passed, the node attributes of the knowledge graph will be updated.

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

[0047] ;

[0048] In the formula, Confidence represents the confidence; W i represents the weight factor of data source i; C i represents the content consistency score of the triples in data source i; η consistency represents the consistency factor among multiple sources.

[0049] Detecting the conflict between new triples and old knowledge includes: attribute value conflict, statistical significance conflict, logical reasoning conflict, and spatio - temporal context conflict.

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

[0051] A benchmark curve generation module, which is used to fit the parameter change trend according to the equipment service life.

[0052] A time - fluctuation correction module, which is used to calculate the instantaneous variance through exponentially weighted moving average.

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

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

[0055] 1. Aiming at the false alarm problem caused by the traditional static threshold being vulnerable to periodic interference (such as the natural pressure fluctuation caused by the day - night temperature difference in the storage tank) in the dynamic industrial scenario, the present invention adopts a sliding - window dynamic confidence interval calculation combined with a time - varying confidence coefficient adjustment mechanism, and realizes the adaptive fitting of the threshold to the real - environment disturbance by embedding a periodic correction term (such as in the pressure working condition), effectively cracking 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 that it is difficult to achieve both accuracy and real - time performance in the field of major hazard source risk monitoring for a long time.

[0056] 2. Based on the insufficient reliability of the knowledge graph of the original monitoring system, the present invention constructs a cognitive closed-loop and introduces a physical verification mechanism based on CFD simulation to conduct secondary verification of knowledge conflicts. When semantic reasoning determines the leakage risk of the storage tank flange, the simulation engine reconstructs the three-dimensional fluid field distribution in real time to verify the physical consistency of the pressure gradient and the leakage diffusion mode, and corrects the confidence level of the mislabeled nodes of "suspected leakage" in the knowledge base from level 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-depth 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 system end-to-end response time is compressed from 5 to 8 minutes of 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 illustrative 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 is the system principle block diagram of a major hazard source risk dynamic calculation system empowered by an AI large model according to an embodiment of the present invention;

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

[0061] Figure 3 is the structure diagram of a spatio-temporal graph neural network (STGNN) according to an embodiment of the present invention;

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

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

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

[0065] Figure 7 is the system working timing diagram according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0066] To make the objectives, 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 merely used to explain the present invention and are not intended to limit the present invention.

[0067] Please refer to Figure 1 , the present invention provides a dynamic risk calculation system for major hazard sources empowered by an AI large model, including: a multi-modal data acquisition and preprocessing module, a heterogeneous data spatio-temporal fusion engine module, an online incremental learning and model fine-tuning module, a risk conduction probability calculation module, a multi-modal knowledge self-evolution module, and an adaptive threshold management and alarm module.

[0068] As Figure 1 shown, the output ports of the multi-modal data acquisition and preprocessing module include: Port A: Transmit the cleaned time series data (including 9 types of parameters such as temperature and pressure) to the heterogeneous data spatio-temporal fusion engine module; Port B: Provide labeled samples (including abnormal event labels and sensor raw values) to the online incremental learning and model fine-tuning module.

[0069] The connection methods of the heterogeneous data spatio-temporal fusion engine module include: Bidirectional connection: Share weight parameters with the online incremental learning and model fine-tuning module (arrow marked "Real-time synchronization of model parameters"); Lower-level port: Transmit the fused feature vectors to the risk conduction probability calculation module.

[0070] The output branches of the risk conduction probability calculation module include: Branch 1: Input the risk diffusion path data (including probability values) into the multi-modal knowledge self-evolution module; Branch 2: Trigger the real-time threshold adjustment signal of the adaptive threshold management and alarm module.

[0071] The feedback path of the multi-modal knowledge self-evolution module: Push the updated entity relationship weights to the heterogeneous data spatio-temporal fusion engine module (arrow marked "Reverse update of knowledge graph").

[0072] The input dependence of the adaptive threshold management and alarm module: Receive the real-time evaluation results of the risk conduction probability calculation module and the latest threshold rules of the multi-modal knowledge self-evolution module.

[0073] The multi-modal data acquisition and preprocessing module is used to connect to sensors through industrial protocols to achieve synchronous acquisition of raw data. The process of data preprocessing implemented by this multi-modal data acquisition and preprocessing module is as Figure 2 shown.

[0074] In the description of the present invention, the multi-modal 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).

[0075] The sensor data acquisition module is used to acquire the analog signals and digital signals of the physical device body, and distinguish the input types according to the signal types and shunt and forward them.

[0076] Specifically, the input of the sensor data acquisition module is the analog signal (current 4 - 20mA) and digital signal (RS485 message) from the device body; it distinguishes the input types through a signal type judgment circuit (integrated circuit model: TI ADS1115) and shunts and forwards them.

[0077] The sliding window segmentation is used to cut data blocks based on the circular buffer of the edge computing terminal according to the time stamp.

[0078] Specifically, the parameter settings of the sliding window segmentation include: window length = 60 seconds, step size = 10 seconds, and data reuse in the overlapping area; during the implementation process, the sliding window segmentation module needs to cut data blocks based on the circular buffer (capacity 1MB) of the edge computing terminal according to the time stamp.

[0079] The denoising processing module is used to perform wavelet packet decomposition on the data signal to achieve high-frequency noise filtering.

[0080] Specifically, the algorithm flow of the denoising processing module includes:

[0081] 1. Perform 5-layer wavelet packet decomposition on the vibration signal (mother wavelet: Daubechies 4);

[0082] 2. Calculate the energy proportion of each sub-band. If the high-frequency band energy < 5% threshold, then set it to zero;

[0083] 3. Reconstruct the signal and output it to the downstream.

[0084] The time alignment verification module is used to correct the clocks of each sensor. If the clock deviation exceeds the tolerance, then trigger interpolation compensation.

[0085] Specifically, the synchronization mechanism of the time alignment verification module is to use the NTP protocol (Network Time Protocol) to correct the clocks of each sensor, with an error tolerance of ±5ms; if the clock deviation exceeds the tolerance, then it is identified as an abnormality and trigger interpolation compensation (linear interpolation method).

[0086] Missing value LSTM interpolation module, which is used to obtain the time series data collected in the past N sampling periods, and based on the long short-term memory network, predict the current missing value and replace the placeholder.

[0087] Specifically, the network structure of the long short-term memory network is a single-layer LSTM (long short-term memory network), and the number of neurons in the hidden layer = 64; the input format is the data sequence in the past 10 minutes (sampling interval of 1 second); the output of this long short-term memory network is corrected to predict the current missing value and replace the placeholder.

[0088] Modal normalization module, which is used to perform normalization processing on the interpolated time series data.

[0089] Specifically, the calculation formula for modal normalization is:

[0090] ;

[0091] In the formula, μ windows represents the mean value of this window; σ base represents the baseline variance of the device life cycle (the initial value comes from the factory setting); X i represents the original data.

[0092] Heterogeneous data spatio-temporal fusion engine module, which is used to construct a spatio-temporal graph model and analyze the non-linear coupling relationship of multi-source data.

[0093] Among them, the hardware of the heterogeneous data spatio-temporal fusion engine module relies on a GPU server (such as NVIDIA A100, 40GB video memory).

[0094] In the description of the present invention, the heterogeneous data spatio-temporal fusion engine module includes: an input module (not shown in the figure), a spatio-temporal 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).

[0095] Input module (input layer), which is used to receive the preprocessed time series data and extract multi-modal feature vectors (12-dimensional features such as temperature gradient, vibration spectrum peak value, gas concentration change rate, etc.).

[0096] Spatio-temporal graph construction module (spatio-temporal graph construction layer), which is used to construct a spatio-temporal graph composed of nodes and edges, and use a temporal convolutional network to extract the temporal dependencies of each node.

[0097] In the description of the present invention, constructing a spatio-temporal graph composed of nodes and edges and using a temporal convolutional network to extract the temporal dependencies of each node includes:

[0098] Step S11, set each sensor as a node and attach device attribute labels (such as material: 304 stainless steel).

[0099] Step S12: Statistically analyze the coupling strength of risk factors based on the historical accident database, and calculate the edge weights of two nodes. For example, the initial weights of node i (temperature) and node j (vibration) are as follows: ; where P casual is the joint accident probability, and P imdependence is the independent probability.

[0100] Step S13: Construct a temporal convolutional network based on causal dilated convolution, set the network structure parameters, and use the temporal convolutional network to extract the temporal dependencies of each node. Among them, the network structure parameters include the input dimension, the diameter of the convolutional kernel, the dilation factor, the number of output channels, and the activation function.

[0101] Specifically, for the temporal convolutional network (temporal convolution module): The TCN network (temporal convolutional network, also known as the temporal convolution module) is used to extract the temporal dependencies of each node (the diameter of the convolutional kernel D = 5).

[0102] Regarding the core structure and mathematical formula of TCN, the temporal convolution module adopts causal dilated convolution design to ensure that the model does not leak future information and at the same time improves the ability to capture long-term temporal dependencies. The module structure includes the following core parts: 1. For the characteristics of industrial temporal data, the TCN parameter configuration is shown in Table 1:

[0103] Table 1: TCN Parameter Configuration Table

[0104] 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 convolutional kernel includes the first residual block, the second residual block, the third residual block, and the fourth residual block.

[0105] Among them, for the first residual block: the dilated convolution is D = 5, d = 1, and the output dimension is 60×64; the content of feature extraction is to learn the short-term fluctuations related to the vibration intensity (such as 3 abnormal peaks per minute); For the second residual block: the dilated convolution is D = 5, d = 2, and the output dimension is 60×128; the content of the periodic feature is to identify the equipment start-stop cycle (such as a load change every 15 minutes); For the third residual block: the dilated convolution is D = 5, d = 4, and the output dimension is 60×256; the content of the long-term dependency is to capture the slow trend drift caused by corrosion (such as a 0.1% decrease in the pressure baseline per week); Fourth-layer residual block: dilation convolution with D = 5, d = 8, output dimension 60×64; the content of feature compression is to retain the most discriminative temporal patterns and output them to the graph attention network (GAT).

[0106] The content of the output result is the hidden state at each time step, HϵR12x60, which depicts the evolution of the device state at different time scales.

[0107] In addition, regarding the selection basis of the kernel diameter D = 5, its physical meaning is that the local patterns of industrial equipment failures usually appear within 5 seconds (such as the windows of sudden pressure increase and vibration mutation).

[0108] 3. Regarding the adaptability of TCN to industrial scenarios and its advantages compared with traditional RNN / LSTM, it is reflected in three aspects: parallel computing, long-term memory, and stable training. Among them, parallel computing means that TCN supports a fully convolutional structure and can process the entire time window in parallel, with the computing efficiency being 3-5 times higher than that of LSTM; long-term memory means that through dilation convolution, it can cover a longer history with a fixed number of parameters (such as RF = 61 steps vs. LSTM having difficulty capturing dependencies beyond 50 steps due to gradient vanishing); stable training means that there is no problem of gradient explosion / vanishing in LSTM, and the convergence speed of the loss function is increased by 40%.

[0109] The present invention is verified by experiments. Specifically, the influence of different D values on the fault detection F1-score is tested in 100 groups of sensor data as shown in Table 2:

[0110] Table 2: Influence table of different D values on the fault detection F1-score

[0111] Conclusion: D = 5 achieves the optimal balance between retaining details and suppressing noise.

[0112] Attention network module (GAT layer), used to dynamically adjust the weights between nodes based on the attention adaptive mechanism.

[0113] In the description of the present invention, dynamically adjusting the weights between nodes based on the attention adaptive mechanism includes:

[0114] Based on the node feature vector, learnable parameter matrix, and attention vector, calculate the attention coefficient, which is updated in real time through the backpropagation of the learnable parameter matrix and attention vector, so that the connection weights between sensor nodes can be adaptively adjusted following the change of the sensor data distribution.

[0115] In the description of the present invention, the calculation formula of the attention coefficient is:

[0116] ;

[0117] In the formula, αij represents the connection weight between node i and node j; T represents the transpose operation matrix;

[0118] a represents the attention vector: with a dimension of 2d’, used to calculate the correlation strength between nodes;

[0119] W represents the learnable matrix: with a dimension of d’xd (d’ = 32), used for feature space mapping;

[0120] h i and h j represent the node feature vectors: the feature vectors of each sensor node i, with a dimension of d (taking a value of 64 dimensions).

[0121] In addition, the attention network module has dynamic characteristics, which are expressed as the parameters of W and a being updated in real-time through backpropagation, enabling the attention weights to be adaptively adjusted according to the changes in the sensor data distribution. For example, when the correlation between the temperature and vibration signals strengthens under high-temperature working conditions, the value of α ij will automatically increase.

[0122] Meanwhile, the feature aggregation formula is:

[0123] ;

[0124] In the formula, h i ′ represents the aggregated feature; σ represents the activation function, used to enhance the non-linear expression ability; N(i) represents the neighbor set of node i, including sensors with spatio-temporal correlations (spatial distance < 50m and time series isotropy > 0.7).

[0125] It should be noted that feature aggregation refers to the process of fusing and compressing multi-dimensional industrial perception data in the spatio-temporal dimension. It includes temporal dynamic features (the output of TCN (temporal convolutional network)), topological correlation features (the attention coefficient α ij of the graph attention network (GAT)), and process knowledge features (knowledge graph embedding).

[0126] In addition, an example of the association between dynamic weight adjustment and risk conduction is provided. The background is that a local gas leak occurs in a chemical industrial park, and it is necessary to dynamically track the leakage diffusion path. In this scenario, the process of dynamic weight adjustment includes the following steps:

[0127] 1. At time T0: Gas sensor C detects an excessive concentration, and module 2 starts attention calculation;

[0128] 2. At T0 + 30s: According to the data of wind speed sensor D (3-level south wind), it increases from 0.2 to 0.65, enhancing the association between C and the downwind equipment;

[0129] 3. Trigger the CFD simulation of the risk conduction probability calculation module to predict the diffusion direction of the concentration contour map;

[0130] 4. T0 + 5 min: The concentration of downwind sensor E rises and synchronously increases to 0.53;

[0131] 5. The system issues a warning: "It is predicted that the concentration in area E will reach the explosion limit in 8 minutes. It is recommended to close valve V205."

[0132] Through the above design, the attention network module (GAT layer) can dynamically adjust the connection weights between sensor nodes according to the real-time working conditions, effectively capture the non-linear time-varying correlations between hazard source parameters, and solve the defect of traditional methods relying on fixed topologies.

[0133] The risk level prediction module (output layer) is used to output the risk level prediction result and trigger the adaptive adjustment of the local neuron learning rate according to the latest accumulated data.

[0134] Specifically, the output layer adopts a fully connected network (number of neurons: 256 → 64 → 3), corresponding to low, medium, and high risk levels; every time 100 new data (about 10 minutes) are accumulated, it triggers the adaptive adjustment of the local neuron (the last two layers) learning rate (Adam optimizer) to achieve online fine-tuning of the network.

[0135] Among them, the output results of the risk level prediction module can be used for multiple purposes, including the following aspects: 1. The risk level is used as the input of the Bayesian network: When the risk level output by the heterogeneous data spatio-temporal fusion engine module 2 is ≥ medium, it triggers the dynamic update of the conditional probability table (CPT) of the conduction probability calculation module.

[0136] For example, the prior probability of the original node "tank leakage" is P(leakage) = 0.05; when the risk level rises to "medium", the prior probability is corrected to P(leakage) = 0.15xP high. For the conduction path search range constraint, when the risk level is low, only the independent failure probability of the current device (single node calculation) is evaluated; when the risk level ≥ medium, it is necessary to activate the plant-wide propagation path search, and the Bayesian network traverses all relevant device nodes (such as upstream and downstream pipelines, adjacent tanks).

[0137] 2. CFD simulation parameter coupling: If the risk level is "high" and a leakage scenario is involved, the CFD module will shorten the simulation time interval (from 5 s → 1 s); and increase the grid resolution (from 1 m³ → 0.5 m³) to capture the rapidly diffusing concentration gradient.

[0138] 3. Example linkage process: Assume that the temperature of a certain reactor is abnormal (the risk level is determined as "medium"). The system will trigger the update of the Bayesian network, that is, the failure probability of the temperature abnormal node and the adjacent cooling system will increase by 50%; at the same time, start the CFD simulation to predict the heat diffusion area. If the simulation shows that the probability of reaching the adjacent storage tank within 10 minutes > 40%, then increase the attention weight of the temperature node in Module 2, and force the risk level to be upgraded to "high" and trigger an emergency stop instruction.

[0139] Online incremental learning and model fine-tuning module, used to trigger dynamic parameter adjustment based on real-time data streams.

[0140] 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 discrimination 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 expansion 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).

[0141] The attention module insertion and channel evaluation module is used to insert an attention module in the neural network, generate statistics through channel importance evaluation, distinguish shared channels and auxiliary channels, so as to solve the problem of task feature conflicts in incremental learning and avoid key channels of new task training interfering with old tasks.

[0142] Specifically, the attention module insertion and channel evaluation module needs to insert a channel attention module after the time convolutional network (TCN) layer, and the structure is as follows: 1. Input feature map: The output from the TCN layer (dimension [Batch, Channels, Time]).

[0143] 2. Global average pooling: Compress along the time dimension into a channel description vector (dimension [Batch, Channels]).

[0144] 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 to the original number of channels, and the activation function is Sigmoid.

[0145] 4. Output layer: Generate a channel weight vector (dimension [Batch, Channels]), indicating the importance of each channel.

[0146] Generate channel importance statistics. Train the network with the inserted attention module using new task samples (the labeled data from port B of the multimodal data acquisition module). The loss function is cross-entropy to achieve the training of the empty network. Then, calculate the mean of the channel weights of all training samples, sort them from high to low, and select the top 30% channels as candidate shared channels. Finally, store the channel importance ranking results in an in-memory database (such as Redis) for the channel discrimination module to call.

[0147] Channel discrimination and sharing decision module, which is used to dynamically divide shared channels (shared by multiple tasks) and auxiliary channels (task-specific) according to the intersection of channel importance between new and old tasks, so as to achieve a balance between parameter reuse and decoupling of task-specific features.

[0148] Among them, it is necessary to input the candidate shared channels of the new task (such as the channel index list {1, 3, 5, 7}), and the historical shared channels of the old task (obtained through the reverse update interface of the multimodal knowledge self-evolution module, such as {3, 5, 9, 11}). By calculating the intersection of the two ({3, 5}), it is defined as the global shared channel; the non-intersection parts ({1, 7} and {9, 11}) are respectively classified into the auxiliary channels of the new and old tasks.

[0149] All tasks share the same set of parameters and allow bidirectional synchronous updates (real-time weight synchronization with the spatio-temporal fusion engine module). And establish auxiliary channels, including the auxiliary channels of the old task and the auxiliary channels of the new task. Among them, the auxiliary channels of the old task are used for parameter freezing (only forward propagation, reverse update is prohibited); the auxiliary channels of the new task are used to allocate independent parameter spaces, which are initialized with a random normal distribution.

[0150] By maintaining a global channel mapping table, the fields include: channel index, role (shared / auxiliary), task identification to which it belongs, and the last update time. Receive the reverse weight update signal of the multimodal knowledge self-evolution module and dynamically adjust the channel allocation (such as recalculating the intersection when the entity relationship weight changes exceed the threshold).

[0151] Parameter fine-tuning and update module, which is used to update parameters in a fine-tuning manner on the shared channels, and at the same time freeze the old task parameters of the auxiliary channels to ensure the learning efficiency of the new task and avoid the performance degradation of the old task.

[0152] Among them, the parameter fine-tuning and update module updates parameters in a low-rank fine-tuning manner on the shared channels to ensure the performance of the new task and be compatible with the old task. Through low-rank adaptation, perform low-rank decomposition on the convolutional layer weight matrix of the shared channels. During the training process, only optimize matrices A and B, freeze the original weight W, and the number of trainable parameters is reduced to 0.1% - 1% of the full parameter fine-tuning.

[0153] By establishing elastic weight constraints, based on the Fisher information matrix of the old task data, quantify the importance of shared channel parameters for historical tasks, and add a regularization term 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).

[0154] After each incremental training is completed, synchronize the updated A and B matrices to the TCN layer of the spatio-temporal fusion engine module through a bidirectional connection to ensure that the risk prediction model is updated in real time.

[0155] The task-specific filter extension module is used to add parallel filters for specific tasks on the auxiliary channel to enhance the task-specific feature extraction ability and solve the problem of insufficient feature expression ability of the shared channel.

[0156] Specifically, the task-specific filter extension module adds exclusive filters for new tasks on the auxiliary channel to enhance the task-specific feature extraction ability. By adding a parallel 1x1 convolution branch after the TCN layer of the auxiliary channel, an independent filter bank (such as 4 groups of filters) is assigned to each new task; the filter weights are initialized using the He normal distribution, and the bias term is initialized to 0.

[0157] Generate a binary mask matrix according to the task identifier to limit the range of gradient propagation. For example, the loss of task A only backpropagates to its exclusive filter parameters, and the gradient of the filter parameters of task B is set to zero. Finally, concatenate the shared channel features with the output of the exclusive filters on the auxiliary channel along the channel dimension and input them into the subsequent fully connected layer for joint decision-making.

[0158] The catastrophic forgetting suppression module is used to prevent the performance decay of old tasks caused by the update of shared channel parameters to maintain the recognition stability of the model for historical tasks.

[0159] In the catastrophic forgetting suppression module, obtain historical labeled samples (such as leakage event data) from port B of the multi-modal data acquisition module, select representative samples (cluster centers) through K-Means clustering and store them in the buffer. During each incremental training, mix the old and new samples in proportion (25% old samples + 75% new samples) to form a training batch.

[0160] 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.

[0161] Specifically, each task independently calculates the gradients of the shared channels, allocates weights according to the task loss ratio (for example, if the loss of task A accounts for 60%, the gradient weight is 0.6), updates the global parameters after weighted averaging, and monitors the similarity of the gradient directions of the shared channels between the old and new tasks. If the similarity < 0.5, it is determined that there is a direction conflict, and the following operations are triggered: roll back to the previous stable parameter version (loaded from the version snapshot); dynamically reduce the learning rate (for example, adjust from 1e-4 to 5e-5).

[0162] After every 100 training iterations, save the model parameter snapshots and optimizer states (momentum, learning rate) to the non-volatile memory of the edge computing node to support fast recovery.

[0163] The risk conduction probability calculation module is used to generate the risk diffusion path by combining the Bayesian network and CFD simulation.

[0164] In the description of the present invention, the risk conduction probability calculation module includes: a Bayesian network inference engine module (not shown in the figure), a CFD physical simulation engine module (not shown in the figure), and a risk conduction path integration module (not shown in the figure).

[0165] The Bayesian network inference engine module is used to construct a Bayesian network using the physical device and environmental states, regularly correct the joint probability distribution in the conditional probability table using the acquired real-time sensor data, and calculate the risk probabilities corresponding to each physical device node.

[0166] Specifically, the schematic diagram of the Bayesian network structure (including the linkage with the CFD module) is as Figure 4 shown. The structure and content of the Bayesian network inference engine include the following aspects: 1. Device node definition: Physical entities such as storage tank V101, pump P203, pipeline L305, etc. constitute the network nodes.

[0167] 2. Conditional probability table (CPT): Initialized based on historical data statistics and expert experience. For example, when the storage tank pressure exceeds the limit (V101_Fault = 1) and the temperature > 60°C (V101_Temp = High), the explosion probability of the adjacent pipeline L305 is 0.78.

[0168] 3. Dynamic update mechanism: Correct the joint probability distribution in the CPT every hour according to the real-time data.

[0169] For the initial risk identification in the Bayesian network, its input is real-time sensor data (pressure, temperature, gas concentration, etc.); the Bayesian network structure includes: Bayesian network nodes representing physical devices or environmental states (such as storage tank V101, pump P203, wind direction, etc.).

[0170] It should be noted that the key operations of the Bayesian network inference engine include the following: 1. Dynamically update the conditional probability table (CPT). The update formula for the conditional probability table is:

[0171] ;

[0172] In the formula, 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, otherwise 0); T represents the statistical window duration (such as data for the most recent seven days); X i represents the current device node; t represents the observation point index of the time window; Pa(X i ) represents the set of parent nodes of X i .

[0173] 2. Search for the risk propagation path. This algorithm traverses the child nodes of the Bayesian network using depth-first search (DFS) to generate potential influence chains; its path sequence is S = 〔V101 → P203 → L305〕.

[0174] The CFD physical simulation engine module is used to start the CFD simulation engine and simulate and output the spatio-temporal distribution of the combustible gas diffusion concentration cloud map when the risk probability of any physical device node exceeds the preset threshold.

[0175] Specifically, the input variables of the CFD physical simulation engine include: leakage location, wind speed, terrain data (GIS elevation model); the output result is the spatio-temporal distribution of the combustible gas diffusion concentration cloud map (updated at 5-second intervals); the simulation process needs to solve the N-S equations using the finite volume method (FVM) and calculate the turbulence diffusion model (k-ε model).

[0176] The trigger condition for the CFD physical simulation engine is to start the CFD engine when the risk probability of any device node exceeds the threshold. The simulation process of the CFD physical simulation engine includes: three-dimensional modeling and mesh generation, and mesh generation. Unstructured tetrahedral meshes (average side length 0.5m) are used, and local refinement (0.1m) is performed near the leakage point. Factory GIS geographic information, device 3D models (STEP format), and environmental data (wind speed, temperature gradient) need to be input.

[0177] During the simulation, solve the control equations, such as the diffusion model, including coupling the mass conservation and momentum equations and considering the turbulence effect (k-ε model). The formula is:

[0178] ;

[0179] In the formula, represents the gas concentration (kg / m 3 ); represents the turbulent diffusion coefficient; represents the leakage source term (related to the leakage rate Q leak ).

[0180] It should be noted that the CFD physical simulation engine discretizes the equations using the finite volume method (FVM) and solves the velocity-pressure coupling field using the SIMPLEC algorithm. Its output content for the time-varying concentration field is the time step, such as Δt = 0.1 s (meeting the CFL condition); the output data also includes the gas concentration distribution of the entire plant area recorded every 5 seconds .

[0181] The risk conduction path integration module is used to mark the risk paths of combustible gas leakage based on the concentration field data output by the CFD simulation engine and the equipment topology relationship diagram.

[0182] Specifically, the input data of the risk conduction path integration module includes the CFD concentration field data and the equipment topology relationship diagram (adjacency matrix A n×n ); this risk conduction path integration module needs to implement the conduction probability calculation and identification through an integration algorithm, including the following content: 1. For the marking of explosion hazard areas, the judgment condition for this area marking is (LEL is the lower explosion limit, such as for methane, LEL = 5% volume concentration);

[0183] The set of hazard areas is .

[0184] 2. Cascade path probability calculation. During the calculation, the topological influence weight includes the physical connection weight between devices (such as pipeline length L ij , flow rate Q ij ), and its calculation formula is:

[0185] ;

[0186] Explosion conduction probability: If device i is in the hazard area D risk , then the probability that this device is affected is calculated by the formula:

[0187] ;

[0188] Among them, is the failure probability of the upstream device i.

[0189] 3. During the process of time window fusion, it is necessary to establish cascade time constraints. Let the safety response time of the deviceT j (e.g., pump P203 must be closed within 2 minutes after gas coverage), if the CFD predicts the gas arrival time T arrive < T j , then the path risk level is increased by one level.

[0190] Final path probability The calculation formula is:

[0191] ; where P source Initial leakage probability; I (·) represents the time compliance indicator function (compliance is 1, otherwise 0).

[0192] When the Bayesian network detects that the probability of tank leakage > 0.5, it triggers the CFD module to generate a gas diffusion prediction (such as the coverage area after 5 minutes). If the concentration exceeds 60% of the lower explosive limit (LEL), it is marked as a high-risk path.

[0193] The multi-modal knowledge self-evolution module is used to continuously update the risk knowledge base by using large model technology.

[0194] Among them, the multi-modal knowledge self-evolution module runs on a distributed cluster (Hadoop + Spark architecture).

[0195] In the description of the present invention, the multi-modal knowledge self-evolution module includes: a multi-modal 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).

[0196] The multi-modal input parsing module is used to parse structured data and unstructured data respectively.

[0197] Specifically, structured data: sensor log (CSV format) → entity extraction (device ID, alarm code);

[0198] Unstructured data: maintenance report (PDF) → BERT model extracts "faulty component: pump bearing"; maintenance video → YOLOv5 identifies "the proportion of rusty flange area ≥ 30%".

[0199] The knowledge triple generation module is used to establish an entity relationship grammar template, generate knowledge triples composed of device - fault mode - root cause, and calculate the confidence of the newly generated knowledge triples.

[0200] Specifically, the entity-relationship grammar template: <equipment, failure mode, root cause> (such as <Reactor R201, seal failure, chloride stress corrosion>);

[0201] It should be noted that for the confidence calculation of knowledge triples, based on the source reliability weights (sensor data 0.9 vs manual report 0.7), the weight factors and confidence include the following: 1. Definition and quantification basis of weight factors. Some of the data comes from sensor data (weight = 0.9). First, the error rate verification of sensor data needs to be carried out. That is, through the annotation and analysis of 10,000 hours of sensor historical data, the standard deviation (σsensor) of temperature / pressure sensors and the average deviation from the manual calibration value are ±0.5%, and the false alarm rate (FP Rate) is about 2%. The drift rate (Drift Ratio) ≤ 0.1% / day for 30 consecutive days of operation (tested according to the GB / T 25919-2022 standard) to achieve the quantification of sensor data stability.

[0202] The weight calculation formula for sensor data is:

[0203] ;

[0204] Another part of the data comes from manual reports (weight = 0.7). 500 historical maintenance reports are extracted and compared with the results of post-fault traceability. The accuracy rate of key fields (such as faulty components, reasons) is 68%, which is used to calculate the sampling accuracy rate; for the median of the report delay time (Tdelay) is 48 hours, according to the decay factor of the Logistic function, the calculation formula is:

[0205] ;

[0206] Finally, the comprehensive weight is calculated, , after normalizing to the [0, 1] interval, it is corrected upward to 0.7 (expert experience compensation). In addition, due to the ambiguity of the description in manual reports (such as "suspected leakage" for quantification), the basic weight needs to be compensated by expert experience for the scenario.

[0207] 2. Confidence synthesis formula. The calculation formula for the confidence of the newly generated knowledge triples is:

[0208] ;

[0209] In the formula, Confidence represents the 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 consistencyRepresents the consistency factor among multiple sources (if both the sensor and the manual report mention the same cause, then η = 1.2; otherwise, η = 0.8).

[0210] The knowledge graph fusion module is used to detect conflicts between new triples and old knowledge. If a conflict exists, it triggers expert review, and after the review is passed, it updates the attributes of the knowledge graph nodes.

[0211] Among them, detecting conflicts between new triples and old knowledge includes attribute value conflicts, statistical significance conflicts, logical reasoning conflicts, and spatio-temporal context conflicts. The specific content is as follows:

[0212] 1. Attribute value conflict: The numerical attribute exceeds the allowable error range or the categorical attribute does not match;

[0213] 2. Statistical significance conflict: Judge the difference in the distribution of new and old data based on chi-square test or t-test;

[0214] 3. Logical reasoning conflict: Reason based on ontology (OWL) axioms to detect logical contradictions after adding triples;

[0215] 4. Spatio-temporal context conflict: The marked spatio-temporal conditions (such as temperature and pressure ranges) are inconsistent, and the conclusions are contradictory under the intersection working conditions.

[0216] Specifically, the knowledge self-evolution process is as Figure 5 shown. If there is a conflict between the new triple and the old knowledge (such as new data indicating that "the corrosion rate of 304 stainless steel increases by 2 times when Cl⁻ > 200 ppm"), it triggers expert review (push the entry to be confirmed on the Web interface); after the review is passed, it updates the attributes of the graph nodes and feeds back to the graph model weight calculator of the heterogeneous data spatio-temporal fusion engine module.

[0217] The adaptive threshold management and alarm module is used to dynamically optimize the alarm rules and output emergency decisions.

[0218] 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).

[0219] The reference curve generation module is used to fit the parameter change trend according to the service life of the device.

[0220] The time fluctuation correction module is used to calculate the instantaneous variance through exponentially weighted moving average.

[0221] 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 periods and the output risk level is greater than or equal to medium level.

[0222] Specifically, the adaptive threshold management and alarm module is deployed at the edge (implemented by a low-power FPGA), as Figure 6 shown. The key steps of the adaptive threshold management and alarm module include: 1. Benchmark curve generation, fitting the parameter change trend (cubic polynomial regression) according to the equipment service life (in months); 2. Time-varying fluctuation correction, calculating the instantaneous variance through the exponentially weighted moving average (EWMA). The calculation formula is:

[0223] ;

[0224] where λ = 0.85, μ t is the sliding mean window (24 hours); λ represents the decay factor, with a value range of 0 ≤ λ < 1, which is used to control the weight decay rate of historical data. The larger the value, the more persistent the influence of historical data and the lower the sensitivity (industrial default value: λ = 0.9); o t represents the exponentially weighted variance estimate at time t (Variance Estimate), which 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"; o t-1 represents the exponentially weighted variance estimate at time t - 1, which is used for the historical variance estimate of the previous moment and is used for smooth iterative calculation. represents the observed data value at time t (Observation Value), which is used for the physical quantities collected by the sensor in real time (such as temperature, vibration frequency, chloride ion concentration); The mean estimate at time t is usually updated by EWMA: , which is used to dynamically characterize the system steady-state benchmark (such as "average corrosion rate" or "normal operating condition pressure") and support anomaly detection.

[0225] 3. Dynamic threshold calculation. The definition of the safety boundary is: \text{Threshold}_{\text{upper}}=\mu_t + 3\times\sigma_t\times\gamma_{\text{age_factor}}; The definition of the aging factor is: \gamma_{\text{age_factor}}=1 + 0.05\times\text{number of years in operation}; The alarm logic is that when the hazard source parameters exceed Threshold upper for 3 consecutive sampling periods, and the risk level output by module two is ≥ medium, the linkage control is triggered (such as closing the upstream valve and starting the exhaust).

[0226] It should be noted that the timing diagram of the working process of this system is as follows Figure 7 as shown.

[0227] It should be noted that the timing diagram of the working process of this system is as follows Figure 7 as shown.

[0228] Taking the storage tank leakage event as an example, the cooperation process of each module is as follows:

[0229] 1. At time T0: The pressure sensor detects a sudden change in the pressure value (rise rate > 10 kPa / s), and the preprocessing module marks it as abnormal;

[0230] 2. At T0 + 2s: The heterogeneous data spatio-temporal fusion engine module 2 determines that the risk level has risen to "medium", and module four starts the Bayesian network to query related devices;

[0231] 3. At T0 + 5s: It is found that the historical failure rate of the adjacent pressure relief valve is 20%, triggering the CFD simulation to predict the diffusion path;

[0232] 4. At T0 + 30s: The multi-modal knowledge self-evolution module 5 updates the confidence of the triple "storage tank overpressure causes flange leakage" in the knowledge graph;

[0233] 5. At T0 + 60s: The adaptive threshold management and alarm module 6 generates an adjustment instruction according to the new threshold, and the platform pops up a window to prompt "start regional evacuation within 5 minutes".

[0234] The following is a supplementary description in combination with the technical architecture and core content adopted by this system.

[0235] I. Regarding the technical architecture of this system, the overall architecture diagram - the hierarchical division is the 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 level marked event; event trigger → knowledge graph leakage risk reasoning → CFD / PINN physical field verification → operation instruction issuance.

[0236] 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 )}; the output is {lower limit of the current window threshold L ( t ), upper limit H ( t ), confidence level (high / medium / low)}.

[0237] II. Core technical details of dynamic threshold generation. For the calculation of the sliding window parameter, the formula expression of this sliding window parameter:

[0238] ; In the formula, k 0 represents the basic confidence coefficient (default 2.5, adjustable range 1.8 - 3.2); T represents the period corresponding to the periodic environmental disturbance (for example, the diurnal temperature difference period is set to 24 hours); σ windows represents the standard deviation of features within the window; α represents the risk confidence coefficient.

[0239] In addition, regarding the details of the machine learning alternative, it is specifically reflected in the following aspects: 1. LSTM model structure, input layer → 3 - layer LSTM (hidden module 128 → 64 → 32) → fully - connected layer (outputs the predicted pressure change rate); 2. Training data features, time - series features are the average pressure, variance, and autocorrelation coefficient in the past 1 hour; environmental features are temperature (EMA smoothing), equipment operation duration, and pipeline flow velocity; 3. Dynamic threshold generation logic, using the 90% confidence interval (quantile regression method) of the predicted value as the upper and lower limits of the threshold.

[0240] In summary, by means of the above - mentioned technical solution of the present invention, aiming at the false - alarm problem caused by the susceptibility of traditional static thresholds to periodic interference (such as the natural pressure fluctuation caused by the diurnal temperature difference in a storage tank) in a dynamic industrial scenario, the present invention adopts a sliding - window dynamic confidence - interval calculation combined with a time - varying confidence - coefficient adjustment mechanism, and realizes the adaptive fitting of the threshold to the real - environment disturbance by embedding a periodic correction term (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 that it is difficult to achieve both accuracy and real - time performance in the field of major - hazard source risk monitoring for a long time. Based on the insufficient reliability of the knowledge graph of the original monitoring system, the present invention constructs a cognitive closed - loop and introduces a physical verification mechanism based on CFD simulation to perform secondary verification of knowledge conflicts - when semantic reasoning determines the risk of leakage of a storage - tank flange, the simulation engine reconstructs the three - dimensional fluid - field distribution in real time to verify the physical consistency of the pressure gradient and the leakage diffusion mode, and corrects the confidence level of the mislabeled node of "suspected leakage" in the knowledge base from level 0.72 to below 0.18, significantly improving the reliability of cross - domain knowledge coupling. Based on the response delay of the original monitoring system, the present invention further combines the resource - collaboration strategy of edge - side signal pre - processing and cloud - side deep analysis, deploys a lightweight AI model to achieve millisecond - level abnormal perception, and at the same time only uploads high - confidence events to the cloud to trigger the dynamic evolution of the knowledge graph, reducing the overall data transmission volume by 73%, and shortening the system end - to - end response time from 5 - 8 minutes of the traditional architecture to within 45 seconds.

[0241] It should be understood that although the steps in the flowchart of the accompanying drawings are shown sequentially according to the indication of the arrows, these steps are not necessarily executed sequentially in the order indicated by the arrows. Unless otherwise clearly stated in this document, there is no strict order restriction for the execution of these steps, and they can be executed in other orders. Moreover, at least a part of the steps in the flowchart of the accompanying drawings may include multiple sub-steps or multiple stages. 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 alternately or in turn with at least a part of other steps or sub-steps or stages of other steps.

Claims

1. A major hazard source risk dynamic calculation system empowered by an AI large model, characterized in that, Including: A multi-modal data acquisition and preprocessing module, which is used to dock sensors through industrial protocols to achieve synchronous acquisition of raw data; A heterogeneous data spatio-temporal fusion engine module, which is used to construct a spatio-temporal graph model and analyze the non-linear coupling relationship of multi-source data; An online incremental learning and model fine-tuning module, which is used to trigger dynamic parameter adjustment based on real-time data streams; A risk conduction probability calculation module, which is used to generate a risk diffusion path by combining a Bayesian network and CFD simulation; A multi-modal knowledge self-evolution module, which is used to continuously update the risk knowledge base using large model technology; An adaptive threshold management and alarm module, which is used to dynamically optimize alarm rules and output emergency decisions.

2. The dynamic risk calculation system for major hazard sources empowered by an AI large model according to claim 1, wherein, The multi-modal data acquisition and preprocessing module includes: A sensor data acquisition module, which is used to acquire analog and digital signals of the physical device body, judge the circuit according to the signal type to distinguish the input type, and split and forward; A sliding window segmentation module, which is used to cut data blocks according to timestamps based on the circular buffer of the edge computing terminal; A denoising processing module, which is used to perform wavelet packet decomposition on data signals to achieve high-frequency noise filtering; A time alignment verification module, which is used to correct the clocks of each sensor. If the clock deviation exceeds the tolerance, interpolation compensation is triggered; A missing value LSTM interpolation module, which is used to obtain time series data collected in the past N sampling periods, and based on the long short-term memory network, predict the current missing value and replace the placeholder; A modal normalization module, which is used to perform normalization processing on the interpolated time series data.

3. The dynamic risk calculation system for major hazard sources empowered by an AI large model according to claim 1, wherein, The heterogeneous data spatio-temporal fusion engine module includes: An input module, which is used to receive the preprocessed time series data and extract multi-modal feature vectors; A spatio-temporal graph construction module, which is used to construct a spatio-temporal graph composed of nodes and edges, and use a temporal convolutional network to extract the temporal dependencies of each node; An attention network module, which is used to dynamically adjust the weights between nodes based on an attention adaptive mechanism; A risk level prediction module, which is used to output the risk level prediction result, and trigger adaptive adjustment of the local neuron learning rate according to the latest accumulated data.

4. The dynamic risk calculation system for major hazard sources empowered by an AI large model according to claim 3, characterized in that, The construction of the spatio-temporal graph composed of nodes and edges, and the use of a temporal convolutional network to extract the temporal dependencies of each node include: Setting each sensor as a node and attaching device attribute labels; Based on the historical accident database, statistically calculate the coupling strength of risk factors and calculate the edge weights of two nodes; Construct a temporal convolutional network based on causal dilated convolution, set the network structure parameters, and use the temporal convolutional network to extract the temporal dependencies of each node, where the network structure parameters include input dimension, convolution kernel diameter, dilation factor, output channels, and activation function.

5. The dynamic risk calculation system for major hazard sources empowered by the AI large model according to claim 3, wherein, The dynamic adjustment of the weights between nodes based on the attention adaptive mechanism includes: Calculating the attention coefficient based on the node feature vector, learnable parameter matrix, and attention vector, and updating in real time through the backpropagation of the learnable parameter matrix and attention vector, so that the connection weights between sensor nodes can be adaptively adjusted following the change of the sensor data distribution.

6. The dynamic risk calculation system for major hazard sources empowered by the AI large model according to claim 5, characterized in that The calculation formula of the attention coefficient is: ； 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 represent the node feature vectors; T represents the transpose operation matrix.

7. The dynamic risk calculation system for major hazard sources empowered by the AI large model according to claim 1, wherein, The risk conduction probability calculation module includes: The Bayesian network inference engine module is used to construct a Bayesian network using physical devices and environmental states, regularly correct the joint probability distribution in the conditional probability table using the acquired real-time sensor data, and calculate the risk probabilities 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 a preset threshold, and simulate and output the spatio-temporal distribution of the combustible gas diffusion concentration cloud map; The risk conduction path integration module is used to mark the risk paths of combustible gas leakage based on the concentration field data output by the CFD simulation engine and the device topology diagram.

8. The dynamic risk calculation system for major hazard sources empowered by an AI large model according to claim 1, wherein The multi-modal knowledge self-evolution module includes: The multi-modal input parsing module is used to parse structured data and unstructured data respectively; The knowledge triple generation module is used to establish an entity relationship grammar template, generate knowledge triples consisting of device - failure mode - root cause, and calculate the confidence of the newly generated knowledge triples; The knowledge graph fusion module is used to detect conflicts between new triples and old knowledge. If there are conflicts, it triggers expert review, and after the review passes, it updates the node attributes of the knowledge graph.

9. The dynamic risk calculation system for major hazard sources empowered by an AI large model according to claim 8, characterized in that, The calculation formula for the confidence of the newly generated knowledge triples is: ； In the formula, Confidence represents the confidence; W i Denotes the weight factor of data source i; C i Indicates the content consistency score of the triples in data source i; η consistency represents the consistency factor among multiple sources; The detection of conflicts between new triples and old knowledge includes: attribute value conflicts, statistical significance conflicts, logical reasoning conflicts, and spatio-temporal context conflicts.

10. The dynamic risk calculation system for major hazard sources empowered by an AI large model according to claim 1, characterized in that, The adaptive threshold management and alarm module includes: The reference curve generation module is used to fit the parameter change trend according to the service life of the device; The time fluctuation correction module is used to calculate the 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 periods and the output risk level is greater than or equal to medium level.

Images