Big data-driven hydrogen fluoride purification risk entropy assessment and multi-parameter monitoring system

Abstract

The invention discloses a big data-driven hydrogen fluoride purification risk entropy assessment and multi-parameter monitoring system, which relates to the technical field of chemical process monitoring, comprises a multi-mode heterogeneous sensor network, and realizes high-precision detection and three-dimensional monitoring by using a nanowire FET (Field Effect Transistor) and the like. The space-time tensor decomposition module extracts features and identifies a causal relationship; the risk entropy manifold learning module evaluates the risk; the self-adaptive twin network monitors the abnormity and simulates and disposes the abnormity; and a knowledge graph-reinforcement learning system decision making subsystem and a plurality of subsystems such as an optical monitoring subsystem and a multi-scale modeling subsystem are also provided, so that whole-process risk management and control are realized. The hydrogen fluoride purification monitoring level is greatly improved, the detection sensitivity and the anomaly detection accuracy are remarkably improved, risk early warning is more timely, decision response is accelerated, the production efficiency is improved, energy consumption is reduced, the system has self-repairing and self-power-supply capabilities, data are safe and traceable, reliable operation of the system is guaranteed, and economic losses and potential safety hazards are reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of chemical process monitoring, and in particular to a big data-driven hydrogen fluoride purification risk entropy assessment and multi-parameter monitoring system. Background Art

[0002] In the field of hydrogen fluoride purification, traditional monitoring systems rely on a single parameter threshold to judge risks, making it difficult to cope with potential threats under complex working conditions. Existing sensors mostly use electrochemical or semiconductor technology to detect hydrogen fluoride concentrations, with response times of up to several minutes and a detection limit of only ppm, making it impossible to capture trace leaks in a timely manner. Monitoring equipment for key parameters such as temperature and pressure is discretely distributed, with low data acquisition frequency, making it impossible to form spatiotemporally continuous monitoring data, making it difficult to quickly locate the root cause when abnormal events occur. In addition, traditional systems lack the ability to predict the health status of equipment, and often cause unplanned shutdowns due to equipment aging or hidden failures, resulting in significant economic losses and safety hazards.

[0003] With the development of industrial big data and artificial intelligence technologies, some companies have attempted to apply machine learning to chemical process monitoring, but existing methods have significant limitations. Most systems only perform simple statistical analysis on historical data and are unable to dynamically adapt to real-time changes in process parameters, resulting in poor model generalization capabilities. In terms of risk assessment, traditional entropy calculation methods are based on static probability distributions and do not consider the causal relationship between parameters and the spatiotemporal evolution characteristics, resulting in delayed risk warnings. At the same time, existing decision support systems rely on manual experience to formulate disposal plans, making it difficult to quickly generate optimal strategies under complex working conditions and unable to meet the strict real-time and accuracy requirements of the hydrogen fluoride purification process.

[0004] The monitoring system for the hydrogen fluoride purification process also faces challenges in the areas of edge computing and data security. Traditional edge nodes have limited computing power and are unable to process massive amounts of sensor data, often leading to data transmission delays and packet loss. The lack of effective encryption and traceability mechanisms during data storage and transmission makes it difficult to quickly locate the responsible party in the event of data tampering or leakage. Furthermore, the existing system lacks equipment self-repair and energy self-sufficiency. Sensor failures require manual replacement and rely on external power supplies. This can lead to monitoring blind spots in extreme environments, seriously impacting system reliability. Summary of the Invention

[0005] The present invention proposes a big data-driven hydrogen fluoride purification risk entropy assessment and multi-parameter monitoring system to solve the problems mentioned in the above-mentioned prior art.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions: A big data-driven hydrogen fluoride purification risk entropy assessment and multi-parameter monitoring system, including the following modules: Multimodal heterogeneous sensor networks: deploying surface-functionalized nanowire field-effect transistor sensor arrays to specifically recognize HF molecules; integrating fiber Bragg grating temperature sensors to monitor temperature using wavelength division multiplexing; and employing plasmon resonance sensors to detect trace metal impurities. Spatiotemporal tensor decomposition feature extraction module: This module constructs multi-source data into a four-dimensional tensor and uses a non-negative matrix three-factor decomposition algorithm to extract spatiotemporal features. It also introduces causal structure learning, constructs a causal graph between parameters using the PC algorithm, identifies key influencing paths, and designs a sliding window entropy rate estimation method to calculate the entropy change rate of the parameter sequence in real time. in, is the rate of change of entropy at time t; n is the number of parameters; is the probability distribution; Risk entropy manifold learning and evaluation module: construct Riemann manifold risk space, calculate risk evolution path through geodesic distance, total risk entropy , where M is a Riemannian manifold space; is the position in the manifold space The risk density function at ; For manifold volume elements, a graph neural network is used to dynamically update the manifold structure and automatically assign parameter weights through the attention mechanism; Adaptive twin network monitoring module: This module builds a digital twin of the physical system, uses neural radiation field technology to achieve 3D reconstruction of the equipment, and designs a twin network synchronization algorithm. When parameter anomalies are detected, it conducts virtual intervention experiments in the twin, evaluates the effectiveness of different treatment options using a Monte Carlo tree search algorithm, and generates a control strategy. Knowledge graph-reinforcement learning decision system: Build a knowledge graph in the field of hydrogen fluoride purification, design a knowledge-enhanced deep reinforcement learning decision model, embed the knowledge graph as prior knowledge input into the DRL agent, use the policy gradient algorithm to optimize the control strategy, and balance safety and economy through a reward shaping mechanism.

[0007] Furthermore, it also includes an adaptive optical monitoring subsystem: deploying a distributed fiber optic sensing network, using stimulated Brillouin scattering technology to achieve simultaneous measurement of temperature and strain, developing a differential absorption lidar system, measuring HF concentration distribution through dual-wavelength laser, designing a photoacoustic spectroscopy detection unit, using the characteristic absorption spectrum of HF molecules to achieve trace detection, and constructing a three-dimensional dynamic monitoring field through the fusion of multiple optical technologies.

[0008] Furthermore, it also includes a multi-scale physics-data fusion modeling subsystem: establishing a multi-scale model architecture, using molecular dynamics to simulate the transmission and reaction process of HF molecules at the micro level, using computational fluid dynamics to simulate fluid flow and heat transfer at the meso level, building a system-level state space model at the macro level, designing a multi-scale coupling algorithm, realizing the connection of models of different scales through coarse-graining technology, developing a data assimilation module, and using ensemble Kalman filtering to integrate real-time monitoring data into the physical model.

[0009] Furthermore, it also includes a cognitive computing early warning subsystem: building a three-level cognitive computing architecture, using convolutional neural networks at the bottom layer to identify abnormal parameter patterns, long and short-term memory networks in the middle layer to predict abnormal evolution trends, and the top layer to generate disposal suggestions through the Transformer architecture. It designs a situational awareness module, generates a dynamic adjustment strategy for early warning thresholds through analogical reasoning, deploys an adversarial training mechanism, and simulates extreme working conditions through generative adversarial networks.

[0010] Furthermore, it also includes a digital watermark traceability subsystem: implanting physical unclonable functional chips in equipment components to generate digital identity identification, designing a quantum random number generator to assign a unique watermark to each data point, realizing data traceability through a chaotic encryption algorithm, developing a blockchain evidence storage module, packaging device data on the chain, and using a proof-of-stake mechanism to ensure that data cannot be tampered with. When an accident occurs, the problem component can be located through the digital watermark.

[0011] Furthermore, it also includes a dynamic causal graph reasoning subsystem: building a time-varying causal graph model and evaluating the impact of parameter perturbations on the system through counterfactual reasoning: in, The effect change of the outcome variable Y caused by the change of parameter X; is the expected function; is the causal intervention operation; Y is the outcome variable; X is the intervention variable; x is the value of variable X; is the change in variable X. When it is detected that the key parameters deviate from the normal range, the root cause is located through causal diagram reasoning.

[0012] Furthermore, it also includes an edge intelligent collaborative computing architecture: designing a three-level edge computing node, device-level nodes completing data preprocessing, regional-level nodes implementing local risk assessment, factory-level nodes making global optimization decisions, developing an adaptive task offloading algorithm, dynamically allocating computing tasks based on network latency, node load, and data importance, and deploying a federated learning framework.

[0013] Furthermore, it also includes a self-repairing sensor network subsystem: designing a redundant deployment strategy for sensor nodes, using a cellular topology to achieve ≥3-fold coverage, developing a sensor health status assessment algorithm, identifying faulty nodes through multi-sensor data consistency verification, and initiating a self-repair mechanism when a sensor failure is detected: a micro-robot driven by a micro-electromechanical system carries a spare sensor module to the fault point.

[0014] Furthermore, it also includes energy collection and self-powered electronic systems: deploying thermoelectric power generation modules, using the temperature gradient of the purification process to generate electricity, integrating vibration energy harvesters, using piezoelectric materials to convert equipment vibration energy into electrical energy, developing radio frequency energy collection units, obtaining energy from surrounding wireless signals, supplementing system energy consumption, designing intelligent energy management systems, and optimizing energy collection efficiency through maximum power point tracking technology.

[0015] Furthermore, it also includes a metaverse decision support subsystem: building a metaverse decision-making platform based on virtual reality / augmented reality, developing an interactive interface, designing a risk deduction engine, simulating system responses under extreme working conditions in the metaverse, automatically generating emergency plans through genetic algorithms, deploying group decision-making modules, and supporting remote experts to make collaborative decisions through virtual avatars.

[0016] Compared with the existing technology, the beneficial effects of the present invention are: The system uses nanowire field-effect transistor sensors and optical fiber sensing technology to reduce the detection limit of hydrogen fluoride concentration to 10 -12 mol / L, with a response time reduced to less than 5ms. This represents a nine-order-of-magnitude increase in sensitivity compared to traditional sensors, enabling detection of potential leaks up to 72 hours in advance. Combining a distributed sensor network with spatiotemporal tensor analysis, the system enables three-dimensional monitoring and dynamic feature extraction of process parameters, improving anomaly detection accuracy and effectively avoiding missed and false alarms.

[0017] In terms of risk assessment and decision support, the risk entropy model based on Riemannian manifold learning can quantify and analyze complex relationships between parameters. It dynamically updates risk status through graph neural networks, improving the accuracy of risk warnings. The combination of knowledge graphs and reinforcement learning enables full automation from risk identification to response plan generation, reducing decision response time from four hours to three minutes, improving production efficiency, and reducing energy consumption. The virtual simulation environment created by digital twin technology supports risk deduction and strategy optimization without interrupting production, significantly enhancing the system's risk mitigation capabilities.

[0018] The system's edge intelligent collaborative computing architecture and self-healing mechanism further enhance reliability. Three-level edge nodes enable hierarchical data processing, and a task offloading algorithm reduces system latency and energy consumption. The sensor network boasts redundant coverage and self-healing capabilities, shortening fault repair time to less than 10 minutes and ensuring continuous monitoring. The energy-harvesting self-powered electronic system utilizes temperature differences, vibration, and radio frequency energy to maintain core functions for 72 hours in the event of a power outage. Furthermore, digital watermark traceability and blockchain evidence storage technologies enable full data lifecycle management, ensuring data security and traceability, and comprehensively enhancing the intelligence level of the hydrogen fluoride purification process. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a schematic block diagram of the hydrogen fluoride purification risk entropy assessment and multi-parameter monitoring system proposed in the present invention; Figure 2 This is a comparison chart of the risk assessment accuracy of the GNN model and the traditional model; Figure 3 This is a combined diagram of the spatial distribution of horizontal and vertical positioning errors corresponding to HF leakage concentration. DETAILED DESCRIPTION

[0020] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0021] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise" and the like to indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0022] In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the said features. In the description of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined. In addition, the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be a connection between the two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances. The present invention will be further described in detail below with reference to the accompanying drawings.

[0023] Reference Figures 1 to 3 :A big data driven hydrogen fluoride purification risk entropy assessment and multi-parameter monitoring system, including the following modules: Multimodal heterogeneous sensor network: This system deploys a nanowire field-effect transistor (FET) sensor array, using zinc oxide nanowires as the sensitive material. Trimethylsilyl (TMS) groups are surface-modified to enhance the specific adsorption of HF molecules. The sensor array is integrated onto a flexible polyimide substrate, forming a wrappable structure. Groups of four sensors each for temperature, pressure, concentration, and flow are deployed every 20 cm along the pipeline's inner wall, forming a three-dimensional monitoring grid.

[0024] The fiber Bragg grating (FBG) temperature sensor utilizes chirped grating technology, combining wavelength-division multiplexing (WDM) and time-division multiplexing (TDM) to achieve simultaneous measurement at 1,000 sensing points. Each grating is inscribed on an 80μm diameter single-mode optical fiber and encapsulated in a stainless steel capillary for enhanced corrosion resistance. The system uses a demodulator to detect the wavelength shift of the reflected light. The temperature measurement range is -50°C to +200°C, with an accuracy of ±0.05°C.

[0025] The plasmon resonance sensor is based on a silicon substrate modified with gold nanoparticles. When HF molecules adsorb onto the surface, they cause a localized change in the refractive index, resulting in a shift in the plasmon resonance peak. Using surface-enhanced Raman scattering (SERS) technology, p-mercaptobenzoic acid (4-MBA) molecules are modified on the gold nanoparticles to serve as Raman probes, enabling highly sensitive detection of metal impurities such as Al, Fe, and Ni, with a detection limit as low as 0.1 ppb.

[0026] All sensor data is transmitted to edge computing nodes via an industrial-grade wireless mesh network, using the IEEE802.15.4e standard protocol. This allows for a transmission distance of up to 500 meters and a transmission rate of 1 Mbps, ensuring real-time data delivery. The network employs a hierarchical routing strategy, prioritizing the transmission of critical data over less important data.

[0027] Spatiotemporal tensor decomposition feature extraction module: The collected multi-source data is constructed into a four-dimensional tensor, whose four dimensions are time, space, parameters, and modality. The time dimension resolution is set to 100ms, which means that the changes in data over time can be captured more precisely; the spatial dimension corresponds to the physical location of the sensor, making it easier to determine the source of the data in space; the parameter dimension covers 20 process parameters such as temperature, pressure, and concentration, comprehensively reflecting the key indicators in the process; the modal dimension is used to distinguish data from different types of sensors. The tensor is initialized using the stochastic gradient descent method. After 500 iterations, the learning rate is set to 0.001. By continuously adjusting the parameters, the tensor gradually reaches the appropriate initial state.

[0028] The non-negative matrix tri-factor decomposition (NMTF) algorithm is used to extract spatiotemporal features. When implementing the algorithm, the decomposition rank is set to 15, and the iterative solution is performed by the alternating least squares method with a convergence threshold of 10. -6 After decomposition, three factor matrices are obtained, corresponding to temporal features, spatial features, and parameter features respectively. The decomposition quality is evaluated by reconstruction error to ensure the accuracy and reliability of feature extraction.

[0029] Introducing causal structure learning, using PC algorithm to construct causal graph between parameters. Setting significance level = 0.01, and the maximum condition set size k = 5. Causal relationships between parameters are determined by calculating conditional independence. In a causal graph, each node represents a parameter, and directed edges represent causal relationships. For example, analysis revealed that temperature change is the root cause of both pressure and reaction rate, with a causal strength value of 0.85 (ranging from 0 to 1, with values closer to 1 indicating stronger causal relationships), thus identifying key influencing paths.

[0030] A sliding window entropy rate estimation method is designed to calculate the entropy change rate of the parameter sequence in real time. The window size is set to 100 sampling points and the step size is 10 points. The kernel density estimation (KDE) is used to calculate the entropy rate, using the Gaussian kernel as the kernel function and the bandwidth parameter according to the Silverman rule. ,in It represents the rate of change of entropy at time t, reflecting how fast the uncertainty of parameter sequence changes over time; n is the number of parameters, reflecting the scale of parameters involved in the calculation; is the probability distribution of parameter i at time t, which is used to calculate the entropy value of parameter i at that time and further participates in the calculation of the entropy change rate. When the threshold value exceeds 0.2 bit / s, the system triggers a level 1 warning.

[0031] Risk Entropy Manifold Learning and Assessment Module: This module constructs a Riemannian manifold risk space and employs a hyperbolic geometry model to map a 20-dimensional high-dimensional parameter space, including temperature, pressure, and concentration, onto a 3D manifold surface. The manifold embedding algorithm uses the Laplace Eigenmap (LE), with a neighborhood parameter k=12 and an eigenvector dimension of 3, to effectively map high-dimensional parameters onto the low-dimensional manifold surface. The risk evolution path is quantitatively analyzed by calculating the geodesic distance on the manifold, which reflects the trajectory and degree of risk change within the manifold space.

[0032] Total risk entropy The calculation formula is ,in represents the total risk entropy, which is used to measure the uncertainty of the overall risk; M is the constructed Riemannian manifold space, which is the basic space for risk analysis; is the position in the manifold space The risk density function at is estimated by the Gaussian mixture model (GMM), with the number of mixture components set to 5, and the expectation maximization (EM) algorithm is used to iteratively optimize the parameters. It describes the density of risk distribution at different locations in the manifold space. It is a manifold volume element. It uses discretization approximation to divide the manifold surface into grid cells with a side length of 0.01. The volume of each cell is calculated based on the local curvature and is used to measure small changes in the space volume in the integral operation.

[0033] A graph neural network (GNN) is used to dynamically update the manifold structure. Specifically, the Graph Attention Network (GAT) architecture is selected, consisting of three hidden layers, eight attention heads per layer, and a dropout rate of 0.1. The network inputs are parameter feature vectors and the manifold topology, and the output is the risk entropy value. The attention mechanism automatically assigns parameter weights, effectively improving the accuracy of risk assessment.

[0034] The system categorizes risk entropy into five levels: safe (0-0.2), low risk (0.2-0.4), medium risk (0.4-0.6), high risk (0.6-0.8), and extremely high risk (0.8-1). Different monitoring strategies and response mechanisms are matched to different risk levels to achieve hierarchical risk management. For example, high risk levels may trigger more frequent monitoring and more urgent disposal processes, thereby ensuring effective risk management.

[0035] Adaptive twin network monitoring module: The digital twin is constructed using Neural Radiant Field (NeRF) technology. Twenty depth cameras and ten industrial-grade laser scanners deployed throughout the purification workshop collect equipment geometry data. A multi-view stereo (MVS) algorithm is used to reconstruct the 3D model, achieving a vertex accuracy of 0.5mm. Texture mapping employs a feature point matching method to ensure visual consistency between the model and the physical equipment.

[0036] The twin network synchronization algorithm is based on a spatiotemporal attention mechanism. Temporal attention weights are calculated using a gated recurrent unit (GRU), while spatial attention weights are extracted using a convolutional neural network (CNN). The synchronization error is defined as the root mean square error (RMSE) between the digital twin model's predictions and the measured values of the physical system. The system dynamically adjusts the prediction parameters using a Kalman filter.

[0037] When a parameter anomaly is detected, the system initiates a Monte Carlo Tree Search (MCTS) algorithm within the digital twin. The number of simulations is set to 10,000, with each simulation consisting of 10 time steps, each corresponding to one minute of actual system operation. The value of each action is calculated using the Upper Confidence Bound applied to Trees (UCT) formula, and the action sequence with the highest value is selected as the optimal control strategy.

[0038] Knowledge Graph - Reinforcement Learning Decision System: The knowledge graph is constructed using a bottom-up approach, extracting entities and relationships from text such as process manuals, equipment specifications, and operating procedures. Named Entity Recognition (NER) uses the BERT-BiLSTM-CRF model. After fine-tuning on hydrogen fluoride purification data, entity recognition accuracy has improved. Relationship extraction uses a graph convolutional network (GAT-GCN) with an attention mechanism.

[0039] The deep reinforcement learning (DRL) decision model uses the proximal policy optimization (PPO) algorithm. The state space includes 50 dimensions such as the current risk entropy value, equipment status, and process parameters, and the action space includes 15 discrete actions such as valve opening adjustment and temperature setting value adjustment. The reward function is designed as: in, is the risk entropy reduction value, For energy saving, is the process stability index, weight coefficient =0.6, =0.2, =0.2.

[0040] The meta-learning framework uses the Model-Agnostic Meta-Learning (MAML) algorithm and is pre-trained under five typical operating conditions, with each task containing 100 episodes. When the system encounters a new operating condition, it can quickly adapt after only five episodes, achieving a policy convergence speed five times faster than traditional DRL methods.

[0041] The present invention also includes an adaptive optics monitoring subsystem: a distributed fiber-optic sensing network uses single-mode communication fiber as the sensing medium, detecting Rayleigh backscattered light via an optical time-domain reflectometer (OTDR). Stimulated Brillouin scattering (SBS) temperature measurement is based on the linear relationship between the Brillouin frequency shift and temperature, with a temperature sensitivity of 1.1 MHz / °C and a measurement range of -40°C to +200°C. Strain measurement is based on changes in the power of Brillouin scattered light, with a strain sensitivity of 0.002 με and a measurement range of 0 to 10,000 με.

[0042] The Differential Absorption LiDAR (DIAL) system uses an optical parametric oscillator (OPO) pumped by a Nd:YAG laser as its light source, emitting dual-wavelength lasers at 2.4μm (HF absorption peak) and 2.5μm (reference wavelength). The system has a scanning frequency of 10Hz, a vertical resolution of 0.5m, a horizontal coverage range of 100m, and a measurement accuracy of ±0.5ppm.

[0043] The photoacoustic spectroscopy detection unit utilizes quartz-enhanced photoacoustic spectroscopy (QEPAS) technology, with a 1.39μm distributed feedback laser (DFB) as the light source and a quartz tuning fork resonating at 32.768kHz. The system employs wavelength modulation spectroscopy (WMS) technology, and the second harmonic detection signal intensity exhibits a linear relationship with HF concentration, with a detection limit as low as 0.1ppm and a response time of less than 5s.

[0044] The present invention also includes a multi-scale physics-data fusion modeling subsystem: Molecular dynamics (MD) simulation uses LAMMPS software, and the simulation system includes 10 6 HF molecules and 10 4 The force field selected was ReaxFFreactiveforcefield, with a time step of 0.1 fs and a simulation duration of 10 ns. Microscopic parameters such as molecular diffusion coefficients and reaction rate constants were calculated to provide boundary conditions for the mesoscopic model.

[0045] Computational fluid dynamics (CFD) simulations were performed using ANSYS Fluent software to establish a three-dimensional flow model. Unstructured tetrahedral meshing was used, with a mesh size of 1 mm in critical areas and a global mesh count of 5 million. The k-ω SST model was used as the turbulence model, and the energy equation accounted for the thermal effects of chemical reactions. Reaction rate constants obtained from MD simulations were imported into the CFD model using the UDF interface.

[0046] The system-level state-space model was constructed using MATLAB / Simulink and includes 12 state variables, including temperature, pressure, and concentration, as well as eight control inputs and five output variables. Model parameters were determined using system identification methods and updated online using the recursive least squares (RLS) method. The data assimilation module employed an ensemble Kalman filter (EnKF) with 100 members, generating the ensemble by perturbing observations and model parameters.

[0047] The present invention also includes a cognitive computing early warning subsystem: a convolutional neural network (CNN) employing the ResNet-18 architecture takes as input a two-dimensional time-frequency plot of a parameter time series, generated via a short-time Fourier transform (STFT). Based on ImageNet pre-training, the network was fine-tuned on a hydrogen fluoride process dataset with a learning rate of 0.0001, a batch size of 32, and a training cycle of 50 epochs. It can identify eight abnormal patterns, including temperature fluctuations and sudden pressure changes.

[0048] The Long Short-Term Memory (LSTM) network consists of three hidden layers, each with 128 neurons and a dropout rate of 0.2. The input is the CNN output feature vector and the current parameter value, and the output is the parameter prediction for the next 10 minutes. The Transformer architecture uses 12 encoder layers and 6 decoder layers, each with 8 attention heads. The model was pre-trained on large-scale chemical process text data and then fine-tuned on hydrogen fluoride data. Through the attention mechanism, the model can capture long-range dependencies between parameters and generate more reasonable disposal recommendations.

[0049] The context awareness module maintains a knowledge base of 500 historical cases, each of which includes anomaly characteristics, a resolution plan, and an evaluation of its effectiveness. When a new anomaly is detected, the most similar case is calculated using cosine similarity, with a similarity threshold set to 0.8.

[0050] This invention also includes a digital watermark traceability subsystem: the physically unclonable function (PUF) chip uses SRAMPUF technology to generate a unique ID based on the randomness of the SRAM cell state when the chip is powered on. Each PUF chip contains 1024 SRAM cells and uses BCH error correction code for error correction, with a Hamming distance threshold set to 0.1. The chip is encapsulated in epoxy resin and can withstand temperatures ranging from -40°C to +120°C and HF gas corrosion environments.

[0051] The quantum random number generator (QRNG) is based on the principle of vacuum state fluctuations and uses InGaAs single-photon detectors to detect photon noise. The generated random numbers are verified by the NIST SP800-22 test suite, passing all 15 randomness tests. The random number generation rate is 10 Mbps, and the random numbers are encrypted using the AES-256 algorithm to ensure the security of the watermark.

[0052] The blockchain evidence storage module utilizes the Hyperledger Fabric platform, with five organization nodes and two sorting nodes. The block size is set at 2MB, with a block generation time of 10 seconds. Smart contracts are written in Go, enabling data on-chain, permissions management, and data query capabilities. Using the PBFT consensus algorithm, transaction confirmation times are less than 3 seconds, with a throughput of 2000 TPS.

[0053] The present invention also includes a dynamic causal graph inference subsystem: the time-varying causal graph (TVCG) model uses a sliding window mechanism with a window size of 1000 sampling points, and the causal graph is updated every time 100 new points are added. The Granger causality test uses a VAR model, and the lag order is determined by the AIC criterion, with a maximum order of 10. The significance level α is set to 0.01, and a Bonferroni correction is used to address multiple hypothesis testing.

[0054] The counterfactual reasoning algorithm employs a causal effect estimation framework, calculating intervention effects through backdoor adjustment formulas. In practice, the system pre-calculates and stores causal effect tables for common intervention scenarios. When reasoning is needed, the system directly looks up the table to obtain the results, accelerating reasoning speeds to milliseconds.

[0055] The system constructed a causal graph containing 30 parameters, in which temperature was identified as the root cause node affecting the reaction rate with a causal strength value of 0.85; the causal strength between pressure and conversion rate was 0.72.

[0056] This invention also includes an edge intelligent collaborative computing architecture: edge computing nodes adopt a three-tiered design. Device-level nodes are based on STM32H7 series microcontrollers, equipped with 2MB of SRAM and 16MB of Flash, and run the FreeRTOS operating system. The nodes implement data acquisition, preprocessing, and feature extraction, with data processing latency less than 1ms and power consumption less than 500mW.

[0057] Regional nodes are based on an NXP MX8MPlus processor, equipped with 4GB of LPDDR4 memory and 64GB of eMMC storage, and run Ubuntu Linux. They implement local risk assessment and data fusion, supporting simultaneous processing of 100 sensor data channels with a data processing throughput of 100MB / s.

[0058] The factory-level nodes are based on Intel Xeon D-2141I processors, equipped with 32GB of DDR4 memory and a 1TB NVMe SSD, and run the Docker containerization platform. They implement global optimization decisions and model training, support deep learning frameworks such as TensorFlow and PyTorch, and achieve a floating-point computing capability of 120 TOPS.

[0059] The adaptive task offloading algorithm is based on a Markov decision process (MDP) model. The state space includes network latency, node load, and data importance, and the action space is the task allocation strategy. The optimal strategy is trained using the Q-learning algorithm, reducing overall system latency and energy consumption.

[0060] The present invention also includes a self-healing sensor network subsystem: Sensor redundancy is deployed in a honeycomb topology, with each hexagonal grid measuring 1 meter on a side and containing seven sensor nodes (one at the center and six around it). Nodes use a distributed consensus algorithm (DCA) for data fusion. A micro-robot powered by a microelectromechanical system (MEMS) employs a wheeled mechanism and is equipped with infrared and ultrasonic sensors for autonomous navigation. The robot measures 5 cm × 5 cm × 3 cm, has a payload capacity of 100 g, and a maximum speed of 10 cm / s. The backup sensor module carries a standardized interface and supports hot-swappable replacement.

[0061] The laser welding system uses a 1064nm Nd:YAG pulsed laser with a maximum output power of 50W and a pulse width of 1-20ms. The welding process is monitored in real time by a machine vision system, and welding parameters are adjusted using image processing algorithms. The welding time is less than 30 seconds.

[0062] The present invention also includes an energy-harvesting self-powered subsystem: the thermoelectric power generation module utilizes Bi2Te3-based thermoelectric materials, with an aluminum heat sink at the hot end and a water-cooled heat sink at the cold end. When the temperature difference is 50°C, a single module outputs 2.5W. With 20 modules connected in series, the total output reaches 50W, which can meet the energy consumption requirements of a regional edge node.

[0063] The vibration energy harvester uses PZT-5H piezoelectric ceramic material and is designed as a cantilever beam structure with a resonant frequency of 120Hz, matching the vibration frequency of industrial equipment. At a vibration acceleration of 1g, a single harvester outputs 15mW of power. The system deploys 10 harvesters in key locations on the equipment, achieving a total output power of 150mW, sufficient to continuously power sensor nodes.

[0064] The RF energy harvesting unit utilizes a 2.4GHz microstrip antenna and rectifier circuit to receive Wi-Fi signals within a 5m radius. When the signal strength is -20dBm, the output power is 5mW. The system utilizes a maximum power point tracking (MPPT) circuit and adaptive impedance matching technology to improve energy harvesting efficiency.

[0065] The intelligent energy management system is based on an STM32L4 microcontroller and uses dynamic voltage scaling (DVS) technology to adjust the supply voltage according to load demand. Energy storage uses a supercapacitor with a capacity of 100F and an operating voltage range of 2.7-5.5V.

[0066] This invention also includes a Metaverse decision support subsystem: The Metaverse platform is developed based on the Unity3D engine and uses WebGL technology for cross-platform access. The 3D reconstruction of the physical system achieves millimeter-level accuracy, and through texture mapping and lighting rendering, achieves visual consistency with the real scene. The platform supports up to 4K resolution and 60 FPS frame rate, with latency less than 20ms in VR mode.

[0067] The multi-dimensional interactive interface integrates gesture recognition, voice control, and brain-computer interface capabilities. Gesture recognition utilizes a Leap Motion controller, while voice control utilizes the iFlytek speech recognition engine, supporting command recognition within a 10-meter range. The brain-computer interface utilizes the EmotivInsight headband, which detects alpha and beta waves for simple command input.

[0068] The risk deduction engine is based on Monte Carlo simulations, generating 10,000 random scenarios per simulation, covering various risks such as equipment failure, operational errors, and external interference. The genetic algorithm uses a population size of 200, 50 iterations, a crossover probability of 0.8, and a mutation probability of 0.05. Through these simulations, the system can predict potential risks within 72 hours.

[0069] The group decision-making module supports simultaneous online collaboration among 10 experts, each of whom is rendered as a highly human avatar using motion capture and facial expression recognition technology. The system records all interactions during the decision-making process and uses sentiment analysis algorithms to assess decision quality.

[0070] The above are only preferred specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solutions and inventive concepts of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A big data driven hydrogen fluoride purification risk entropy assessment and multi-parameter monitoring system, characterized in that: Includes the following modules: Multimodal heterogeneous sensor networks: deploying surface-functionalized nanowire field-effect transistor sensor arrays to specifically recognize HF molecules; integrating fiber Bragg grating temperature sensors to monitor temperature using wavelength division multiplexing; and employing plasmon resonance sensors to detect trace metal impurities. Spatiotemporal tensor decomposition feature extraction module: This module constructs multi-source data into a four-dimensional tensor and uses a non-negative matrix three-factor decomposition algorithm to extract spatiotemporal features. It also introduces causal structure learning, constructs a causal graph between parameters using the PC algorithm, identifies key influencing paths, and designs a sliding window entropy rate estimation method to calculate the entropy change rate of the parameter sequence in real time. in, is the rate of change of entropy at time t; n is the number of parameters; is the probability distribution; Risk entropy manifold learning and evaluation module: construct Riemann manifold risk space, calculate risk evolution path through geodesic distance, total risk entropy , where M is a Riemannian manifold space; is the position in manifold space The risk density function at ; For manifold volume elements, a graph neural network is used to dynamically update the manifold structure and automatically assign parameter weights through the attention mechanism; Adaptive twin network monitoring module: This module builds a digital twin of the physical system, uses neural radiation field technology to achieve 3D reconstruction of the equipment, and designs a twin network synchronization algorithm. When parameter anomalies are detected, it conducts virtual intervention experiments in the twin, evaluates the effectiveness of different treatment options using a Monte Carlo tree search algorithm, and generates a control strategy. Knowledge graph-reinforcement learning decision system: Build a knowledge graph in the field of hydrogen fluoride purification, design a knowledge-enhanced deep reinforcement learning decision model, embed the knowledge graph as prior knowledge input into the DRL agent, use the policy gradient algorithm to optimize the control strategy, and balance safety and economy through a reward shaping mechanism.

2. The big data driven hydrogen fluoride purification risk entropy assessment and multi-parameter monitoring system according to claim 1 is characterized in that: It also includes an adaptive optical monitoring subsystem: deploying a distributed fiber optic sensing network, using stimulated Brillouin scattering technology to achieve simultaneous measurement of temperature and strain, developing a differential absorption lidar system, measuring HF concentration distribution through dual-wavelength laser, designing a photoacoustic spectroscopy detection unit, using the characteristic absorption lines of HF molecules to achieve trace detection, and constructing a three-dimensional dynamic monitoring field through the fusion of multiple optical technologies.

3. The big data driven hydrogen fluoride purification risk entropy assessment and multi-parameter monitoring system according to claim 1 is characterized in that: It also includes a multi-scale physics-data fusion modeling subsystem: establishing a multi-scale model architecture, using molecular dynamics to simulate the transmission and reaction process of HF molecules at the micro level, using computational fluid dynamics to simulate fluid flow and heat transfer at the meso level, building a system-level state space model at the macro level, designing a multi-scale coupling algorithm, and realizing the connection between models of different scales through coarse-graining technology, developing a data assimilation module, and using ensemble Kalman filtering to integrate real-time monitoring data into the physical model.

4. The big data driven hydrogen fluoride purification risk entropy assessment and multi-parameter monitoring system according to claim 1, characterized in that: It also includes a cognitive computing early warning subsystem: building a three-level cognitive computing architecture, using convolutional neural networks at the bottom layer to identify abnormal parameter patterns, long and short-term memory networks in the middle layer to predict abnormal evolution trends, and the top layer to generate disposal suggestions through the Transformer architecture. It designs a situational awareness module, generates a dynamic adjustment strategy for early warning thresholds through analogical reasoning, deploys an adversarial training mechanism, and simulates extreme working conditions through generative adversarial networks.

5. The big data driven hydrogen fluoride purification risk entropy assessment and multi-parameter monitoring system according to claim 1, characterized in that: It also includes a digital watermark traceability subsystem: implanting physical unclonable functional chips in device components to generate digital identity identification, designing a quantum random number generator to assign a unique watermark to each data point, achieving data traceability through a chaotic encryption algorithm, developing a blockchain evidence storage module, packaging device data on the chain, and using a proof-of-stake mechanism to ensure that data cannot be tampered with. When an accident occurs, the problem component can be located through the digital watermark.

6. The big data driven hydrogen fluoride purification risk entropy assessment and multi-parameter monitoring system according to claim 1, characterized in that: It also includes a dynamic causal graph reasoning subsystem: constructing a time-varying causal graph model and evaluating the impact of parameter perturbations on the system through counterfactual reasoning: in, The effect change of the outcome variable Y caused by the change of parameter X; is the expected function; is the causal intervention operation; Y is the outcome variable; X is the intervention variable; x is the value of variable X; is the change in variable X. When it is detected that the key parameters deviate from the normal range, the root cause is located through causal diagram reasoning.

7. The big data driven hydrogen fluoride purification risk entropy assessment and multi-parameter monitoring system according to claim 1, characterized in that: It also includes an edge intelligent collaborative computing architecture: designing a three-level edge computing node, with device-level nodes completing data preprocessing, regional-level nodes implementing local risk assessment, and factory-level nodes making global optimization decisions. It also develops an adaptive task offloading algorithm, dynamically allocates computing tasks based on network latency, node load, and data importance, and deploys a federated learning framework.

8. The big data driven hydrogen fluoride purification risk entropy assessment and multi-parameter monitoring system according to claim 1, characterized in that: It also includes a self-repairing sensor network subsystem: designing a redundant deployment strategy for sensor nodes, using a cellular topology to achieve ≥3-fold coverage, developing a sensor health status assessment algorithm, identifying faulty nodes through multi-sensor data consistency verification, and initiating a self-repair mechanism when a sensor failure is detected: a micro-robot driven by a micro-electromechanical system carries a spare sensor module to the fault point.

9. The big data driven hydrogen fluoride purification risk entropy assessment and multi-parameter monitoring system according to claim 1, characterized in that: It also includes energy harvesting self-powered electronic systems: deploying thermoelectric power generation modules to generate electricity using the temperature gradient of the purification process, integrating vibration energy harvesters to convert equipment vibration energy into electrical energy using piezoelectric materials, developing radio frequency energy harvesting units to obtain energy from surrounding wireless signals to supplement system energy consumption, designing intelligent energy management systems, and optimizing energy harvesting efficiency through maximum power point tracking technology.

10. The big data driven hydrogen fluoride purification risk entropy assessment and multi-parameter monitoring system according to claim 1, characterized in that: It also includes a metaverse decision support subsystem: building a metaverse decision-making platform based on virtual reality / augmented reality, developing an interactive interface, designing a risk deduction engine, simulating system responses under extreme working conditions in the metaverse, automatically generating emergency plans through genetic algorithms, deploying group decision-making modules, and supporting remote experts to make collaborative decisions through virtual avatars.

Images