A Multimodal Sensing Method and System for Safety Control of Industrial Process Anomalies

Abstract

This invention relates to a multimodal sensing-based industrial process anomaly safety control method, belonging to the field of intelligent industrial manufacturing. This method collects data on equipment, environment, materials, and personnel behavior through a multimodal heterogeneous sensor array, and generates a standardized safety dataset through spatiotemporal alignment and fusion. It employs a differentiable industrial field coupling propagation model to estimate production environment field data, achieving dynamic compensation and enhanced perception. A digital twin evolutionary network embedded with production cycle time, energy efficiency, and safety coupling constraints is constructed to output equipment health trends and residual safety capability assessments. Based on an interpretable causal reasoning engine and aligned with a safety procedure knowledge graph, root cause diagnosis and compensation calculation for process deviations are achieved. Temporal dependency constraints are introduced into the attention mechanism to achieve hierarchical early warning and proactive process reconfiguration. This invention improves the safety level and production continuity assurance capabilities of industrial production.

Description

Technical Field

[0001] This invention relates to the technical field of intelligent industrial manufacturing, and in particular to a multimodal sensing method and system for safety control of industrial process anomalies. Background Technology

[0002] With the deepening of industrialization and intelligent manufacturing, process industries such as chemical, metallurgical, and energy face multiple challenges, including increased production intensity, greater equipment complexity, and stricter safety standards. Industrial process safety control has become a core element in ensuring personnel safety, equipment asset integrity, and production continuity. However, at the environmental perception level, existing industrial safety monitoring technologies rely on isolated monitoring using single-modal sensors (such as vibration probes and temperature instruments). In complex industrial environments with strong electromagnetic interference, high-temperature drift, and mechanical vibration coupling, sensor data is susceptible to noise contamination and distortion. Furthermore, the lack of online estimation and dynamic compensation capabilities for interference fields leads to frequent misjudgments of equipment status and a serious deficiency in the reliability of basic safety monitoring data, thus reducing the safety of industrial production. Therefore, improvements are needed. Summary of the Invention

[0003] To improve the safety of industrial production, this application provides a multimodal sensing method and system for industrial process anomaly safety control.

[0004] Firstly, the above-mentioned inventive objective of this application is achieved through the following technical solution:

[0005] A multimodal sensing method for safety control of industrial process anomalies, the method comprising the following steps:

[0006] By collecting equipment operating parameters, environmental state variables, material flow attribute data, and personnel operation behavior video streams in industrial processes through multimodal heterogeneous sensor arrays, and generating standardized monitoring semantic industrial safety datasets after spatiotemporal alignment and multi-scale feature fusion.

[0007] The industrial safety dataset is input into the dynamic industrial environment interference suppression layer. This layer estimates the field data of the work area in real time through a differentiable industrial field coupling propagation model, and generates an enhanced sensing data stream after interference suppression. The field data includes temperature drift field, electromagnetic noise field and mechanical vibration velocity field.

[0008] The enhanced sensing data stream is fed into the industrial process digital twin evolution network. The industrial process digital twin evolution network constructs a device mechanism coupled dynamics model based on a physical information driven architecture and embeds multi-objective coupled constraints of production cycle time, energy efficiency and safety, and outputs the device health evolution trend characteristics and remaining safe operation capability assessment value.

[0009] The health evolution characteristics are aligned with the industrial safety procedure knowledge graph. An interpretable causal reasoning engine is used to evaluate the current work process deviation and safety procedure compliance in real time, and outputs risk compensation and adaptive adjustment instructions for process parameters.

[0010] The risk compensation instruction and the remaining safe operation capability assessment value are input into the safety control decision module. This module introduces time-dependent constraints into the attention mechanism, generates hierarchical safety early warning information, and triggers proactive process refactoring and dynamic response control instructions.

[0011] By adopting the above technical solutions, the core problems in industrial process safety monitoring, such as insufficient interference suppression capabilities, lagging equipment status prediction, ambiguous risk root cause location, and passive early warning response, are systematically solved. Compared with existing technologies, the dynamic interference suppression layer realizes online estimation and adaptive compensation of the industrial environment, significantly improving the reliability of perception; the digital twin evolutionary network deeply integrates equipment mechanisms and production constraints, realizing accurate quantitative prediction of health status and remaining capacity, filling the technical gap in predictive maintenance; the causal reasoning engine provides transparent risk diagnosis capabilities, enabling safety management to shift from experience-driven to data-physical dual-driven; the proactive process reconfiguration mechanism realizes a paradigm shift from "post-event alarm shutdown" to "pre-event early warning and avoidance", greatly reducing the risk of production interruption and safety accident losses. The overall solution provides a smart safety control system with autonomous perception, proactive early warning, and dynamic reconfiguration capabilities for high-risk industrial processes, significantly improving the safety, continuity, and economy of industrial production.

[0012] In a preferred example, this application can be further configured as follows: the step of generating an industrial safety dataset with standardized monitoring semantics after collecting equipment operating parameters, environmental state variables, material flow attribute data, and personnel operation behavior video streams in an industrial process through a multimodal heterogeneous sensor array, and performing spatiotemporal alignment and multi-scale feature fusion includes the following steps:

[0013] Vibration acceleration sensors, temperature thermocouples, pressure transmitters and current sensors are deployed on the body of industrial equipment, and visual inspection cameras and laser rangefinders are integrated on the material conveyor belt.

[0014] An industrial IoT communication channel detection submodule is constructed, which estimates the channel delay spread and packet loss rate parameters by sending orthogonal frequency division multiplexing pilot signals and receiving echoes.

[0015] The collected equipment operating parameters, environmental variables, material properties, and behavioral video streams are time-stamped and aligned. A dynamic time warping algorithm is used to eliminate timing mismatch caused by differences in the sampling frequencies of multimodal sensors.

[0016] The aligned heterogeneous data are fused into multi-scale feature volumes under a unified monitoring semantic framework to generate a standardized industrial safety dataset containing equipment operating condition information, material status information, and personnel behavior information.

[0017] In a preferred embodiment, this application can be further configured such that the step of inputting the industrial safety dataset into a dynamic industrial environment interference suppression layer, which estimates the field data of the work area in real time through a differentiable industrial field coupling propagation model, and generating an enhanced sensing data stream after interference suppression, includes the following steps:

[0018] A differentiable industrial field coupling propagation sub-model is constructed, and the point cloud data of equipment layout is used as spatial constraints to inversely deduce the eddy current attenuation of electromagnetic noise in metal structures and the convective diffusion intensity distribution of temperature field in air medium.

[0019] Based on the inferred intensity distribution, the temperature drift field, electromagnetic noise field, and mechanical vibration velocity field are estimated. The temperature field is modeled as a learnable three-dimensional implicit representation, and a smoothing constraint is applied through the heat conduction equation.

[0020] Based on the estimated interference field distribution, the temperature drift deviation of the sensor signal and the vibration noise floor are dynamically compensated to generate an enhanced sensing data stream after interference suppression.

[0021] A rendering consistency self-supervised mechanism is introduced, in which the compensated data stream is re-input into the propagation model to generate simulated sensor readings, and the residual loss is calculated with the original readings to update the model parameters, thereby achieving unsupervised online learning.

[0022] In a preferred embodiment, this application can be further configured as follows: The step of sending the enhanced sensing data stream into an industrial process digital twin evolution network, wherein the industrial process digital twin evolution network constructs a device mechanism coupled dynamics model based on a physical information-driven architecture and embeds multi-objective coupled constraints of production cycle time, energy efficiency, and safety, and outputs device health evolution trend characteristics and remaining safe operating capability assessment values, includes the following steps:

[0023] A coupled dynamic sub-model of equipment mechanism is constructed, and the rotating machinery rotor system and the transmission gearbox are respectively modeled as a rigid multibody dynamic element and a nonlinear meshing damping system. The equation of motion is established through the principle of virtual work.

[0024] The dynamic model incorporates multi-objective coupled constraints of production cycle time, energy efficiency, and safety. Based on the production process control theory, the dynamic balance coefficient between equipment load rate and safety threshold is calculated and fed back to the dynamic equation to correct the boundary conditions.

[0025] Using the enhanced sensing data stream as the observation input, the unscented Kalman filter algorithm is used to recursively estimate the hidden state of the dynamic model, and outputs the evolution trend of bearing torque, gear contact stress and wear state of key components.

[0026] Based on the current state estimate, the performance degradation trajectory of the equipment in the remaining production cycle is projected forward, and the remaining safe operation capability assessment value and health index are output.

[0027] In a preferred example, this application can be further configured as follows: the step of aligning the health evolution characteristics with the industrial safety procedure knowledge graph, evaluating the current work process deviation and safety procedure compliance in real time through an interpretable causal reasoning engine, and outputting risk compensation and adaptive adjustment instructions for process parameters includes the following steps:

[0028] The equipment health evolution characteristics are mapped to the production process cycle point sequence, and subgraph isomorphic matching is performed with the pre-stored industrial safety procedure knowledge graph to construct a process-safety alignment graph structure.

[0029] Run an interpretable causal reasoning engine on the alignment graph to identify key state variables that cause process deviations through causal discovery algorithms and quantify the contribution of each variable to safety deviations.

[0030] Based on the contribution analysis results, the risk compensation amount of production cycle time and process parameters is calculated online, and the compensation amount is decomposed into equipment load fine-tuning component and material flow rate replanning component.

[0031] When causal reasoning discovers that multivariate coupling leads to systemic safety risks, it triggers an adaptive adjustment command for process parameters, dynamically modifying equipment speed, feed rate, and personnel work intervals to ensure that the overall safety compliance is maintained above the qualified threshold.

[0032] In a preferred embodiment, this application can be further configured as follows: The step of inputting the risk compensation instruction and the remaining safe operation capability assessment value into the safety control decision module, which introduces time-dependent constraints in the attention mechanism, generates hierarchical safety early warning information, and triggers proactive process refactoring and dynamic response control instructions, includes the following steps:

[0033] The remaining safe operation capability assessment value, the historical sequence of risk compensation instructions, and the evolution trend of wear status of key components are encoded into temporal feature vectors and input into the attention mechanism network;

[0034] By introducing temporal dependency constraints into the attention mechanism, the attention weights are forced to follow an autoregressive process, which suppresses overfitting to random disturbances and enhances attention to progressive failure modes.

[0035] The graded safety warning generator outputs three levels of warning information based on the intensity of the degradation mode of attention focus: normal operation monitoring level, preventive maintenance level, and emergency shutdown level;

[0036] When the warning level reaches the preventive maintenance level, the proactive process refactoring instruction is automatically triggered. The digital twin simulates strategies for slowing down production, switching equipment, or bypassing processes, and selects the optimal refactoring solution for execution.

[0037] In a preferred example, this application can be further configured to include an industrial IoT edge-cloud collaborative security architecture, comprising the following steps:

[0038] A lightweight interference suppression layer and a causal inference engine are deployed on the industrial equipment side. The forward calculation of the industrial field propagation model is realized by using a field-programmable gate array to meet the millisecond-level real-time requirements.

[0039] A digital twin evolutionary network and a safety early warning module are run on the edge computing node in the workshop. The dynamic model is deployed on the graphics processor cluster through a model parallel partitioning strategy.

[0040] The enterprise cloud platform maintains a global knowledge graph of industrial safety procedures and regularly collects rare fault data from the edge to perform incremental training. The updated model parameters are then sent to the edge via an encrypted channel.

[0041] Establish a three-level data security flow control mechanism at the cloud, edge, and terminal levels. The terminal only uploads a summary of abnormal operating conditions, the edge terminal uploads a daily health report, and the cloud pulls all historical data on demand, reducing communication bandwidth consumption and ensuring data security.

[0042] Secondly, the above-mentioned inventive objective of this application is achieved through the following technical solutions:

[0043] A multimodal sensing industrial process anomaly safety control device, the device comprising: an industrial safety dataset generation unit, used to collect equipment operating parameters, environmental state variables, material flow attribute data and personnel operation behavior video streams in the industrial process through a multimodal heterogeneous sensor array, and generate an industrial safety dataset with standardized monitoring semantics after spatiotemporal alignment and multi-scale feature fusion.

[0044] An enhanced sensing data stream generation unit is used to input the industrial safety dataset into a dynamic industrial environment interference suppression layer. This layer estimates the field data of the work area in real time through a differentiable industrial field coupling propagation model and generates an enhanced sensing data stream after interference suppression. The field data includes temperature drift field, electromagnetic noise field and mechanical vibration velocity field.

[0045] The equipment health and operational capability assessment unit is used to send the enhanced sensing data stream into the industrial process digital twin evolution network. The industrial process digital twin evolution network constructs a coupled dynamic model of equipment mechanism based on a physical information driven architecture and embeds multi-objective coupled constraints of production cycle time, energy efficiency and safety, and outputs the equipment health evolution trend characteristics and the remaining safe operational capability assessment value.

[0046] An adaptive adjustment instruction output unit is used to align the health evolution characteristics with the industrial safety procedure knowledge graph, and to evaluate the current work process deviation and safety procedure compliance in real time through an interpretable causal reasoning engine, and output risk compensation amount and adaptive adjustment instructions for process parameters.

[0047] The dynamic response control instruction generation unit is used to input the risk compensation instruction and the remaining safe operation capability assessment value into the safety control decision module. This module introduces time-dependent constraints in the attention mechanism, generates hierarchical safety early warning information, and triggers proactive process refactoring and dynamic response control instructions.

[0048] Thirdly, the above-mentioned objectives of this application are achieved through the following technical solutions:

[0049] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described multimodal sensing industrial process anomaly safety control method.

[0050] Fourthly, the above-mentioned objectives of this application are achieved through the following technical solutions:

[0051] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described multimodal sensing industrial process anomaly safety control method. Attached Figure Description

[0052] Figure 1 This is a flowchart of a multimodal sensing industrial process anomaly safety control method according to an embodiment of this application;

[0053] Figure 2 This is a schematic diagram of a multimodal sensing industrial process anomaly safety control device according to an embodiment of this application;

[0054] Figure 3 This is a schematic diagram of an electronic device according to an embodiment of this application.

[0055] Icon labels:

[0056] 1. Industrial safety dataset generation unit; 2. Enhanced perception data stream generation unit; 3. Equipment health and operational capability assessment unit; 4. Adaptive adjustment instruction output unit; 5. Dynamic response control instruction generation unit. Detailed Implementation

[0057] The present application will be further described in detail below with reference to the accompanying drawings.

[0058] In one embodiment, such as Figure 1 As shown, this application discloses a multimodal sensing method for safety control of industrial process anomalies, which specifically includes the following steps:

[0059] S10: Collect equipment operating parameters, environmental state variables, material flow attribute data and personnel operation behavior video streams in industrial processes through multimodal heterogeneous sensor arrays, and generate standardized monitoring semantic industrial safety datasets after spatiotemporal alignment and multi-scale feature fusion.

[0060] In this embodiment, taking the reactor production process as an example, the multimodal heterogeneous sensor array is deployed as follows: a triaxial vibration accelerometer and a temperature thermocouple are installed on the reactor body to collect the reactor vibration spectrum (0-10kHz) and reaction temperature (0-150℃) in real time; a pressure transmitter and an electromagnetic flowmeter are integrated into the material feed pipeline to monitor the feed pressure (0-2MPa) and flow rate (0-500L / h); an explosion-proof industrial camera is installed at the top window of the reactor to collect video streams of the color change of the reactant material and the height of the foam layer; an ambient temperature sensor and a combustible gas detector are deployed on the workshop wall. The sampling frequencies of each sensor differ significantly (vibration 1kHz, temperature 1Hz, video 25fps). The system uses a dynamic time warping algorithm to perform timestamp alignment on all data, downsampling high-frequency vibration data and linearly interpolating low-frequency temperature data to unify and align them to a millisecond-level time reference. Under the unified monitoring semantic framework, equipment operating conditions, material properties, ambient gas concentrations, and video behavior features are fused at the channel dimension to generate a standardized industrial safety dataset that includes time-series vibration spectrum, temperature curves, pressure-flow phase diagrams, and personnel operation behavior labels, providing structured input for subsequent interference suppression.

[0061] S20: Input the industrial safety dataset into the dynamic industrial environment interference suppression layer. This layer estimates the field data of the work area in real time through a differentiable industrial field coupling propagation model and generates an enhanced sensing data stream after interference suppression.

[0062] Specifically, the aforementioned chemical workshop experiences strong electromagnetic interference (25kHz high-frequency noise generated by the frequency converter) and a temperature gradient field (a 90℃ temperature difference between the reactor surface (120℃) and the workshop environment (30℃), leading to convection). A dynamic industrial environment interference suppression layer initiates a differentiable industrial field coupling propagation sub-model. Using the three-dimensional layout point cloud of the reactor and pipelines as spatial constraints, it inversely extrapolates the eddy current attenuation distribution of electromagnetic noise in the metal structure (attenuation coefficient 0.85 / m) and the convective diffusion intensity of the temperature field in the air (diffusion coefficient 0.023m² / s). The model estimates that the temperature drift field on the reactor surface exhibits a gradient distribution with a high center and low edges, and calculates in real-time the thermocouple measurement deviation caused by thermal expansion (approximately +3℃). Based on the estimation results, the system dynamically compensates for temperature sensor readings, subtracting the influence of thermal drift; simultaneously, it adaptively filters the vibration signal to eliminate the electromagnetic noise floor (25kHz and its harmonic components). After compensation, the temperature measurement accuracy improved from ±5℃ to ±1℃, the vibration signal-to-noise ratio improved from 12dB to 28dB, and the generated enhanced sensing data stream significantly improved the reliability of equipment condition assessment.

[0063] S30: The enhanced sensing data stream is sent into the industrial process digital twin evolution network. The industrial process digital twin evolution network constructs a device mechanism coupled dynamics model based on a physical information driven architecture and embeds multi-objective coupled constraints of production cycle time, energy efficiency and safety, and outputs the device health evolution trend characteristics and remaining safe operation capability assessment value.

[0064] Specifically, after enhancing the input of the perception data stream into the industrial process digital twin evolution network, the network constructs a coupled dynamic sub-model of equipment mechanism: the reactor agitator rotor system is modeled as a rigid multibody dynamic unit, considering the viscous resistance of the material; the transmission gearbox is modeled as a nonlinear meshing damping system, considering the decrease in meshing stiffness caused by tooth surface wear. Multi-objective coupled constraints of production cycle time, energy efficiency, and safety are embedded in the dynamic model: the optimal reaction temperature curve (120℃±2℃ maintained for 3 hours) is calculated based on chemical reaction kinetics, and the maximum allowable vibration amplitude of the agitator (peak-to-peak value ≤ 0.5mm) is calculated as a safety boundary based on the equipment safety threshold. Unscented Kalman filtering is used to recursively estimate the hidden states of the dynamic model, revealing an abnormally high amplitude of the bearing outer ring characteristic frequency (BPF=120Hz) in the current vibration spectrum, diagnosed as bearing wear; simultaneously, it is estimated that the gear meshing stress shows a fatigue accumulation trend due to long-term load fluctuations. Based on state estimation, the network forward extrapolates the performance degradation trajectory over the next 72 hours, predicting that the remaining bearing life is approximately 50 hours. The output remaining safe operating capability assessment value is "can continue to operate until the end of this round of production (approximately 48 hours)," and the health index is "early warning level," indicating that a spare bearing needs to be prepared and a planned shutdown maintenance should be arranged.

[0065] S40: Align the health evolution characteristics with the industrial safety procedure knowledge graph, and evaluate the current work process deviation and safety procedure compliance in real time through an interpretable causal reasoning engine, and output risk compensation amount and adaptive adjustment instructions for process parameters;

[0066] Specifically, health evolution characteristics revealed excessive reactor vibration, temperature control deviation (actual 118℃ compared to standard 120℃), and personnel operation videos showing workers not wearing protective masks as required. The system maps these characteristics to a production process cycle point sequence and performs subgraph isomorphic matching with the industrial safety regulations knowledge graph. Knowledge graph nodes include safety constraints such as "reaction temperature 120℃±2℃," "vibration amplitude ≤0.5mm," and "personnel protective equipment complete." The causal reasoning engine identifies key causal variables: bearing wear (65% contribution to vibration), PID controller integral saturation (25% contribution to temperature deviation), and insufficient worker safety awareness (10% contribution to lack of protection). Based on contribution analysis, the system calculates risk compensation online: equipment load fine-tuning component (reducing stirring speed by 10% to decrease bearing load) and material flow rate replanning component (extending reaction time by 0.5 hours to compensate for insufficient temperature). When a systemic safety risk is detected due to multivariate coupling (bearing wear + controller saturation), an adaptive adjustment command for process parameters is automatically triggered: switching to the backup temperature control loop and forcibly locking the upper limit of the stirring speed to 85% of the rated value, ensuring that the overall safety compliance recovers to the acceptable threshold.

[0067] S50: Input the risk compensation instruction and the remaining safe operation capability assessment value into the safety control decision module. This module introduces time-dependent constraints into the attention mechanism, generates hierarchical safety early warning information, and triggers proactive process refactoring and dynamic response control instructions.

[0068] The remaining safe operating capability assessment value is "operable for 48 hours," but the bearing wear evolution trend shows signs of accelerated fatigue crack propagation rate. After this temporal characteristic is input into the attention mechanism network, the temporal dependency constraint forces the attention weight to focus on the long-term trend of crack propagation rather than instantaneous vibration fluctuations. The graded safety early warning generator outputs a "preventive maintenance level" warning based on the attention intensity. The system's proactive process reconfiguration command is automatically triggered, and three strategies are pre-simulated in the digital twin: Strategy 1 "Slowdown operation," reducing the stirring speed to 60% of the rated value, extending the remaining life to 80 hours but reducing the capacity by 20%; Strategy 2 "Switch to standby reactor," starting the standby reactor and transferring materials to achieve uninterrupted production; Strategy 3 "Process bypass," temporarily using the intermediate storage tank to store materials and suspending the feeding of this reactor. The preliminary assessment showed that Strategy 3 could be switched over within 2 hours without affecting the overall production cycle. The system then selected this option and issued control commands. The robot automatically closed the feed valve and bypassed the material to the storage tank. After the workers completed the mask-wearing operation according to the AR navigation prompts, they performed the cooling and depressurization operation inside the reactor, successfully avoiding unplanned shutdowns and safety accidents.

[0069] In summary, through multimodal perception, physical information-driven modeling, interpretable causal reasoning, and an end-to-end closed-loop design for safety control decisions, this system systematically solves core challenges in industrial process safety monitoring, such as insufficient interference suppression capabilities, lagging equipment status prediction, ambiguous risk root cause localization, and passive early warning response. Compared with existing technologies, the dynamic interference suppression layer enables online estimation and adaptive compensation of the industrial environment, significantly improving perception reliability; the digital twin evolutionary network deeply integrates equipment mechanisms with production constraints, achieving accurate quantitative prediction of health status and remaining capacity, filling the technological gap in predictive maintenance; the causal reasoning engine provides transparent risk diagnosis capabilities, enabling safety management to shift from experience-driven to data-physical dual-driven; and the proactive process reconfiguration mechanism achieves a paradigm shift from "post-event alarm shutdown" to "pre-event early warning and avoidance," significantly reducing the risk of production interruption and safety accident losses. The overall solution provides a smart safety control system for high-risk industrial processes with autonomous perception, proactive early warning, and dynamic reconfiguration capabilities, significantly improving the safety, continuity, and economy of industrial production.

[0070] In step S10: Collecting equipment operating parameters, environmental state variables, material flow attribute data, and personnel work behavior video streams from industrial processes using a multimodal heterogeneous sensor array, and generating a standardized industrial safety dataset with monitoring semantics after spatiotemporal alignment and multi-scale feature fusion, the following steps are included:

[0071] S11: Deploy vibration acceleration sensors, temperature thermocouples, pressure transmitters and current sensors on the body of industrial equipment, and integrate visual inspection cameras and laser rangefinders on the material conveyor belt.

[0072] S12: Construct an industrial IoT communication channel detection submodule to estimate channel delay spread and packet loss rate parameters by sending orthogonal frequency division multiplexing pilot signals and receiving echoes.

[0073] S13. Perform timestamp alignment on the collected equipment operating parameters, environmental variables, material properties and behavioral video streams, and use a dynamic time warping algorithm to eliminate timing mismatch caused by the sampling frequency difference of multimodal sensors;

[0074] S14. The aligned heterogeneous data are fused with multi-scale feature volumes under a unified monitoring semantic framework to generate a standardized industrial safety dataset containing equipment operating condition information, material status information, and personnel behavior information.

[0075] In the nitration reactor production process of this application embodiment, the multimodal heterogeneous sensor array is deployed and implemented as follows:

[0076] In step S11, triaxial vibration acceleration sensors are deployed at key locations on the reactor body to monitor its mechanical operation; thermocouples are installed to sense the reaction temperature in real time; pressure transmitters are integrated to acquire pressure changes within the reactor; and current sensors are configured to track the electrical parameters of the stirring motor. Above the material feed conveyor belt, a high-resolution industrial vision inspection camera is integrated to identify material impurities and flow status; and a laser rangefinder is installed to accurately measure the material layer height. These sensors constitute a multi-dimensional sensing network covering equipment operating conditions, material properties, and environmental conditions, achieving physical-layer acquisition of status information throughout the entire process.

[0077] In step S12, due to strong electromagnetic interference in the workshop and the wide distribution of equipment, an industrial IoT communication channel detection submodule is constructed. This module periodically sends orthogonal frequency division multiplexing pilot signals, which are transmitted via a wireless channel and then the echo signals are received. By analyzing the degree of signal distortion, the channel delay spread parameters and packet loss rate are estimated. When an increase in the packet loss rate is detected in a certain area, the system adaptively reduces the data transmission rate and enables forward error correction coding to ensure the reliable transmission of critical security data. At the same time, the channel quality parameters are included as environmental state variables in subsequent fusion analysis.

[0078] In step S13, the sampling frequencies of the various sensors differ significantly (kilohertz for vibration sensors, hertz for temperature sensors, and tens of frames for video streams). The system performs timestamp alignment on the acquired data. When using a dynamic time warping algorithm for alignment, the high-frequency time axis of the vibration data is used as a reference. Temperature data is linearly interpolated, and video frames are timestamped to eliminate timing mismatches caused by asynchronous sampling. For example, when capturing a sudden change in vibration, synchronized temperature and pressure data can accurately reflect the causal timing of the change in operating conditions at that moment, avoiding misjudgments of fault propagation paths due to time misalignment.

[0079] In step S14, the aligned heterogeneous data are fused within a unified monitoring semantic framework: vibration spectrum features, temperature curve features, and pressure-flow phase diagram features are cascaded along the channel dimension to construct an equipment operating condition feature body; material impurity features identified by the visual camera are fused with material layer morphology features measured by laser ranging to construct a material state feature body; and personnel operation video streams are encoded into operation specification tags after behavior recognition, forming a personnel behavior feature body. The final standardized industrial safety dataset, indexed by a unified timestamp, contains equipment operating condition information, material state information, and personnel behavior information, providing a structured and semantically consistent input foundation for subsequent interference suppression and state assessment.

[0080] For steps S11-S14, this multimodal heterogeneous sensor array construction and data fusion process, compared to traditional distributed sensor deployment and isolated data analysis, achieves a leap from "single-point monitoring" to "global perception" in industrial process safety monitoring. Industrial IoT channel detection ensures communication reliability in complex electromagnetic environments and solves the problem of blind spots in status monitoring caused by security data packet loss. The dynamic time warping algorithm effectively eliminates timing mismatch caused by heterogeneous sampling frequencies, ensuring accurate reconstruction of causal relationships from multiple data sources. Feature fusion under a unified monitoring semantic framework organically integrates three key elements: equipment, materials, and personnel, enabling the safety monitoring system to possess global situational awareness capabilities. This significantly improves the accuracy of industrial process anomaly detection and the comprehensiveness of fault root cause localization, laying a high-quality data foundation for subsequent intelligent safety control.

[0081] In step S20: Inputting the industrial safety dataset into the dynamic industrial environment interference suppression layer, which estimates the field data of the work area in real time through a differentiable industrial field coupling propagation model and generates an enhanced sensing data stream after interference suppression, the steps include the following:

[0082] S21: Construct a differentiable industrial field coupling propagation sub-model, using equipment layout point cloud data as spatial constraints, and inversely deduce the eddy current attenuation of electromagnetic noise in metal structures and the convective diffusion intensity distribution of temperature field in air medium.

[0083] S22: Based on the inferred intensity distribution, the temperature drift field, electromagnetic noise field, and mechanical vibration velocity field are estimated. The temperature field is modeled as a learnable three-dimensional implicit representation, and a smoothing constraint is applied through the heat conduction equation.

[0084] S23: Based on the estimated interference field distribution, dynamically compensate for the temperature drift deviation and vibration noise floor of the sensor signal to generate an enhanced sensing data stream after interference suppression;

[0085] S24: Introducing a rendering consistency self-supervised mechanism, the compensated data stream is re-input into the propagation model to generate simulated sensor readings, and the residual loss is calculated with the original readings to update the model parameters, thus achieving unsupervised online learning.

[0086] Specifically, in step S21, the system constructs a differentiable industrial field coupling propagation sub-model. Using the 3D laser point cloud layout data of the reactor, pipelines, and power distribution cabinets within the workshop as spatial constraints, the model inversely extrapolates the eddy current attenuation distribution of electromagnetic noise in the metal structure: when the frequency converter generates 25kHz high-frequency interference, the model calculates that the eddy current penetration depth of this noise along the reactor wall is approximately 3mm, and the noise intensity is enhanced due to impedance abrupt changes at the pipeline flange connection. Simultaneously, the model extrapolates the convective diffusion intensity distribution of the temperature field in the air medium, identifying the natural convection path formed between the 120℃ high-temperature zone on the reactor surface and the 30℃ low-temperature zone in the workshop environment, and quantifying the thermal diffusion intensity distribution driven by the temperature gradient. This inverse extrapolation transforms the intangible interference field into a computable spatial field distribution, providing a physically interpretable basis for subsequent estimations.

[0087] In step S22, based on the inferred intensity distribution, the system estimates the temperature drift field, electromagnetic noise field, and mechanical vibration velocity field of the work area. The temperature drift field is modeled as a learnable three-dimensional implicit representation, and the temperature deviation values ​​of each point in the workshop space are parameterized through a multilayer perceptron. To ensure the spatial continuity of the estimate, the model applies a smoothing constraint to the heat conduction equation, forcing the temperature gradient of adjacent spatial points to conform to Fourier's law of thermal conductivity. For example, if the estimated temperature drift value in a certain local area shows isolated hot spots, violating the continuity of heat conduction, the smoothing constraint term will drive the model to adjust its parameters so that its temperature distribution conforms to the physical law of natural heat diffusion from high-temperature areas to low-temperature areas. The field estimate output by this step effectively reflects the real interference distribution of the industrial environment and avoids misjudgments caused by local measurement noise.

[0088] In step S23, the system dynamically compensates for sensor signals based on the estimated interference field distribution. For thermocouples mounted on the reactor surface, the temperature drift field estimation shows a +3°C thermal conduction deviation in the high-temperature region, which the system subtracts in real time. For signals acquired by vibration sensors, the electromagnetic noise field estimation indicates a -15dB noise floor in the 25kHz frequency band, and the system activates an adaptive notch filter to eliminate interference in this band. For example, when the reactor stirring motor starts, electromagnetic interference noise is mixed into the vibration signal. After compensation, the noise amplitude drops from 0.8m / s² to below 0.1m / s², highlighting the true mechanical vibration characteristics. This dynamic compensation mechanism significantly improves the accuracy of sensor data, and the generated enhanced sensing data stream provides a high signal-to-noise ratio input for subsequent digital twin modeling.

[0089] In step S24, a self-supervised rendering consistency mechanism is introduced. The compensated enhanced sensing data stream (temperature, vibration, and pressure signals) is re-input into the industrial field propagation model. The model forward renders and generates simulated sensor readings based on the currently estimated interference field parameters. If the residual between the simulated readings and the original real readings is large (e.g., the temperature residual exceeds 2°C), it indicates that the current interference field estimation is inaccurate. The residual loss is updated through backpropagation to update the model parameters, making the temperature drift field estimate closer to the real workshop thermal environment. This process requires no manual annotation, achieving unsupervised online learning, enabling the model to continuously and adaptively optimize with changes in production load (e.g., the difference between winter and summer ambient temperatures), maintaining stable and reliable interference suppression performance over the long term.

[0090] The dynamic industrial environment interference suppression layer achieves online estimation and adaptive compensation of industrial field interference through the inverse derivation and forward rendering closed loop of the differentiable propagation model, overcoming the limitation of traditional fixed filters that cannot adapt to changes in operating conditions. Implicit representation of the temperature field and constraints from the heat conduction equation ensure the physical rationality and spatial continuity of the estimation results, avoiding misjudgment of isolated noise points. The dynamic compensation mechanism accurately eliminates temperature drift and electromagnetic noise in sensor signals, significantly improving the data signal-to-noise ratio and laying a reliable foundation for equipment condition assessment. The self-supervised learning mechanism enables the model to self-evolve, continuously optimizing as the production environment changes, reducing the maintenance costs of manual recalibration. Overall, this layer enhances the perception robustness and measurement accuracy of the industrial safety monitoring system under strong interference and variable operating conditions, providing a high-quality enhanced perception data stream for subsequent predictive maintenance and precise safety control.

[0091] In step S30: sending the enhanced sensing data stream into the industrial process digital twin evolution network, which constructs a coupled dynamic model of equipment mechanism based on a physical information-driven architecture and embeds multi-objective coupled constraints of production cycle time, energy efficiency, and safety, and outputs the equipment health evolution trend characteristics and remaining safe operating capability assessment value, the steps include the following:

[0092] S31: Construct a coupled dynamic sub-model of equipment mechanism, model the rotating machinery rotor system and the transmission gearbox as a rigid multibody dynamics unit and a nonlinear meshing damping system, respectively, and establish the equation of motion through the principle of virtual work;

[0093] S32: Embed the multi-objective coupled constraints of production cycle time, energy efficiency and safety into the dynamic model, calculate the dynamic balance coefficient between equipment load rate and safety threshold based on production process control theory, and feed it back to the dynamic equation to correct the boundary conditions;

[0094] S33: Using the enhanced sensing data stream as the observation input, the unscented Kalman filter algorithm is used to recursively estimate the hidden state of the dynamic model, and outputs the evolution trend of bearing torque, gear contact stress and wear state of key components.

[0095] S34: Based on the current state estimate, forward extrapolate the performance degradation trajectory of the equipment during the remaining production cycle, and output the remaining safe operation capability assessment value and health index.

[0096] In the embodiments of this application, the following example is given:

[0097] In step S31, the system constructs a coupled dynamic sub-model of the equipment mechanism. The stirring motor-reducer-stirring blade rotor system at the top of the reactor is modeled as a rigid multibody dynamic element, where the stirring blade's moment of inertia is calculated to be 15 kg·m², and the torsional stiffness of the stirring shaft is 2.5 × 10⁻⁶. 6 The gearbox of the reducer is modeled as a nonlinear meshing damping system with a gear module of 6 mm and a gear ratio of 25:73. The meshing stiffness varies periodically along the meshing line, and a nonlinear term for the increase in meshing clearance due to tooth surface wear is introduced. When establishing the overall motion equation using the principle of virtual work, the virtual work of inertial force of rigid elements and the elastic deformation work of flexible elements participate in the balance, accurately characterizing the dynamic behavior of the coupling of gear meshing impact and shaft torsional vibration when the agitator rotates in viscous material, providing a model basis that conforms to the physical mechanism for subsequent state estimation.

[0098] In step S32, a multi-objective coupled constraint of production cycle time, energy efficiency, and safety is embedded in the kinetic model. Based on the nitration reaction process, the production cycle time constraint requires the stirring speed to be maintained at 120 rpm for 3 hours to complete the reaction cycle; the energy efficiency constraint requires the motor load rate to be maintained at 75%-85% to optimize energy consumption; and the safety constraint requires the peak-to-peak value of the stirring shaft vibration amplitude to be ≤0.5 mm and the bearing temperature to be ≤75℃. The system calculates a dynamic balance coefficient: when it detects that the material viscosity increases due to a decrease in temperature, it automatically increases the upper limit of the load rate to 90% to maintain the speed, while tightening the vibration safety threshold to 0.4 mm to compensate for additional mechanical stress. This coefficient feeds back to correct the boundary conditions of the kinetic equations, enabling the model to automatically match the safe operating envelope under the current production conditions during state estimation, avoiding misjudgments or omissions caused by fixed thresholds.

[0099] In step S33, the enhanced sensing data stream (compensated vibration spectrum, temperature, and pressure signals) output from S20 is used as the observation input, and an unscented Kalman filter algorithm is employed to recursively estimate the hidden states of the dynamic model. Based on the abnormally high amplitude of the 120Hz characteristic frequency in the vibration spectrum, the filter estimates a 0.15mm wear gap in the bearing outer ring; based on the accumulated energy of the gearbox acoustic emission signal, it estimates that the contact stress on the large gear tooth surface has decreased from 800MPa to 650MPa due to wear, but the fatigue accumulation has reached 78% of the critical value; based on the fluctuation of the stirring current, it estimates a 5% increase in the unbalanced torque at the end of the agitator due to material fouling. These accurate estimates of hidden states provide reliable internal state information for subsequent life prediction, overcoming the bottleneck of difficult direct measurement in highly interfering industrial environments.

[0100] In step S34, based on the currently estimated bearing wear clearance of 0.15mm and fatigue accumulation of 78%, the digital twin evolutionary network forward models the performance degradation trajectory of the equipment over the remaining production cycle. Using Paris's crack propagation law, it predicts that the outer ring crack of the bearing will propagate to a critical size after 50 hours under the current load; if the gear contact stress continues to decrease, pitting failure will occur after 120 hours. After comprehensive evaluation, the remaining safe operating capability assessment value is "can continue operating until the end of this production cycle (48 hours)," the health index is "warning level (yellow)," and it is suggested that "a planned shutdown is required at hour 48 to replace the bearing and simultaneously check the gear meshing status." This predictive assessment enables the company to accurately arrange maintenance plans before an accident occurs, avoiding the risk of entire batch of material scrap and safety accidents caused by sudden shutdowns.

[0101] For steps S31-S34, the following beneficial effects are achieved:

[0102] The industrial process digital twin evolution network, by constructing a coupled dynamic model of equipment mechanisms, overcomes the limitations of traditional condition monitoring that only reflects surface signals. It achieves accurate estimation of the hidden states within equipment (bearing clearance, gear stress, wear accumulation), solving the problem of opaque condition perception under strong interference environments. The embedding of multi-objective coupled constraints—production cycle time, energy efficiency, and safety—enables the model to dynamically match the safe operating envelope under actual working conditions, avoiding the rigidity and conservatism of fixed threshold strategies. The unscented Kalman filter algorithm effectively integrates multimodal sensing data, improving the accuracy and noise resistance of condition estimation. Forward extrapolation of performance degradation trajectories achieves a leap from "current condition monitoring" to "remaining life prediction," providing a quantitative basis for predictive maintenance and significantly reducing the risk of unplanned downtime. Overall, this network improves the transparency, accuracy, and quantification of industrial safety monitoring, providing a solid model foundation and decision support for the long-term safe operation of high-risk industrial processes.

[0103] In step S40: Aligning the health evolution characteristics with the industrial safety procedure knowledge graph, evaluating the current work process deviation and safety procedure compliance in real time through an interpretable causal reasoning engine, and outputting risk compensation and adaptive adjustment instructions for process parameters, the following steps are included:

[0104] S41: Map the equipment health evolution characteristics to a sequence of production process cycle points, perform subgraph isomorphic matching with the pre-stored industrial safety procedure knowledge graph, and construct a process-safety alignment graph structure;

[0105] S42: Run an interpretable causal reasoning engine on the alignment graph to identify key state variables that cause process deviations through causal discovery algorithms and quantify the contribution of each variable to safety deviations.

[0106] S43: Based on the contribution analysis results, calculate the risk compensation amount of production cycle time and process parameters online, and decompose the compensation amount into equipment load fine-tuning component and material flow rate replanning component;

[0107] S44: When causal reasoning discovers that multivariate coupling leads to systemic safety risks, an adaptive adjustment command for process parameters is triggered to dynamically modify equipment speed, feed rate, and personnel work intervals to ensure that the overall safety compliance is maintained above the qualified threshold.

[0108] For steps S41-S44, in step S41, the equipment health evolution characteristics show a bearing wear of 0.15mm, an actual reaction temperature of 118℃ (standard 120℃), and personnel video analysis indicating they are not wearing protective masks. The system maps these characteristics to a production process cycle point sequence: feeding stage (0-30min), reaction stage (30-210min), and discharging stage (210-240min). The cycle sequence is matched with a pre-stored industrial safety procedure knowledge graph, with nodes containing safety constraints such as "reaction temperature 120±2℃", "vibration amplitude ≤0.5mm", and "personnel protective equipment complete". The matching reveals a temperature node deviation of -2℃, a vibration node exceeding the standard (actual 0.65mm), and a personnel behavior violation (not wearing a mask) in the reaction stage cycle. A process-safety alignment diagram is constructed, clearly identifying the stage and type of deviation.

[0109] In step S42, an interpretable causal reasoning engine is run on the alignment graph. The causal discovery algorithm identifies three key causal variables: bearing wear (contributing 65% to excessive vibration), PID controller integral saturation (contributing 25% to temperature deviation), and insufficient personnel safety awareness (contributing 10% to lack of protection). The engine generates a visual causal chain diagram, showing the transmission path from "bearing wear 0.15mm" to "increased vibration noise" to "vibration amplitude exceeding the standard by 0.65mm," as well as the logical chain from "controller saturation" to "insufficient heating power" to "temperature deviation of 118℃," transforming the ambiguous safety deviation phenomenon into an interpretable parametric causal chain.

[0110] In step S43, based on the contribution analysis results, the risk compensation amount is calculated online. For excessive vibration, a fine-tuning instruction for the equipment load is generated: reducing the stirring speed from 120 rpm to 108 rpm (a 10% reduction), thus decreasing the bearing load and predicting that the vibration amplitude can be reduced to 0.48 mm, meeting the standard. For temperature deviation, a replanning instruction for the material flow rate is generated: extending the reaction time from 180 min to 210 min (an increase of 30 min), compensating for the decrease in reaction rate caused by insufficient temperature. This decomposition of compensation amounts ensures that the adjustment strategy is both precise and executable, avoiding secondary disturbances caused by a single large adjustment.

[0111] In step S44, causal reasoning detected a strong coupling between two variables: bearing wear and controller saturation, both of which jointly lead to systemic safety risks (if the bearing suddenly seizes up and the controller malfunctions, it may cause the reaction to runaway). The system automatically triggered an adaptive adjustment command for process parameters: immediately switching to the backup temperature control loop and forcibly locking the upper limit of the stirring speed to 85% of the rated value (102 rpm), while extending the feed interval from 30 min to 45 min to ensure that the reactor's thermal load is within a safe envelope under abnormal operating conditions. After the adjustment, the overall safety compliance rate rose from 78% to 92%, remaining above the acceptable threshold, successfully avoiding the risk of cascading failures.

[0112] Steps S41-S44, the process of safety alignment and causal reasoning mechanism, represents a leap from "surface deviation alarms" to "deep root cause diagnosis" compared to traditional safety monitoring based on fixed threshold alarms. Knowledge graph alignment transforms discrete safety procedures into computable structured constraints, enabling deviation assessment to possess semantic-level understanding capabilities. The causal reasoning engine provides contribution quantification and visualization of transmission paths, allowing operators to accurately locate fault sources rather than blindly troubleshooting, significantly shortening fault handling time. The hierarchical compensation strategy (equipment fine-tuning + flow rate replanning) balances adjustment accuracy with production continuity, avoiding material scrap losses caused by emergency shutdowns. The multivariate coupled detection and adaptive adjustment mechanism enables proactive avoidance of systemic risks, elevating safety control from passive response to proactive prevention. Overall, this step enables industrial process safety monitoring to possess transparent diagnostic, refined compensation, and forward-looking adjustment capabilities, significantly improving the inherent safety level and production continuity assurance capabilities under complex operating conditions.

[0113] In step S50: The risk compensation instruction and the remaining safe operation capability assessment value are input into the safety control decision module. This module introduces time-dependent constraints in the attention mechanism, generates hierarchical safety early warning information, and triggers proactive process refactoring and dynamic response control instructions. This step includes the following steps:

[0114] The remaining safe operation capability assessment value, the historical sequence of risk compensation instructions, and the evolution trend of wear status of key components are encoded into temporal feature vectors and input into the attention mechanism network;

[0115] By introducing temporal dependency constraints into the attention mechanism, the attention weights are forced to follow an autoregressive process, which suppresses overfitting to random disturbances and enhances attention to progressive failure modes.

[0116] The graded safety warning generator outputs three levels of warning information based on the intensity of the degradation mode of attention focus: normal operation monitoring level, preventive maintenance level, and emergency shutdown level;

[0117] When the warning level reaches the preventive maintenance level, the proactive process refactoring instruction is automatically triggered. The digital twin simulates strategies for slowing down production, switching equipment, or bypassing processes, and selects the optimal refactoring solution for execution.

[0118] It should be noted that this application also includes an industrial IoT edge-cloud collaborative security architecture, including the following steps:

[0119] S701: Deploys a lightweight interference suppression layer and a causal inference engine on the industrial equipment side, and uses a field-programmable gate array to realize forward calculation of the industrial field propagation model, meeting the millisecond-level real-time requirements;

[0120] S702: Runs a digital twin evolutionary network and a safety early warning module on the edge computing node of the workshop, and deploys the dynamic model on the graphics processor cluster through a model parallel partitioning strategy;

[0121] S703: Maintains a global library of industrial safety procedure knowledge graphs on the enterprise cloud platform, and regularly collects rare fault data from the edge to perform incremental training. The updated model parameters are then sent to the edge via an encrypted channel.

[0122] S704: Establish a three-level data security flow control mechanism at the cloud, edge, and terminal levels. The terminal only uploads a summary of abnormal operating conditions, the edge terminal uploads a daily health report, and the cloud pulls all historical data on demand, reducing communication bandwidth consumption and ensuring data security.

[0123] In this embodiment, a lightweight interference suppression layer and a causal inference engine are integrated into the end-side controller of each reactor. Specifically, an FPGA acceleration card is embedded in the main control PLC of the reactor to solidify the forward calculation logic (temperature drift field estimation, electromagnetic noise compensation algorithm) of the industrial field propagation model into hardware circuitry. When the vibration sensor generates raw data at a frequency of 1kHz, the FPGA completes noise floor filtering and temperature drift compensation within 1 millisecond, directly outputting an enhanced signal to the local safety relay to meet the hard real-time requirements of emergency shutdown (response time <10ms). At the same time, the lightweight causal inference engine retains only the local causal link calculation unit of bearing-vibration-temperature, directly determining at the end side that "if the bearing temperature > 75℃ and the vibration amplitude > 0.5mm, then trigger an early warning," achieving microsecond-level fast response without waiting for cloud commands.

[0124] The workshop edge computing nodes are configured with a GPU cluster (4 NVIDIA A100 GPUs). A model-parallel partitioning strategy is used to split the digital twin evolution network of the reactor into six sub-models: reaction kinetics model, stirring and transmission system model, sealing state model, temperature control loop model, safety threshold model, and lifespan prediction model. Each GPU is responsible for the parallel computation of two sub-models. When three reactors simultaneously upload augmented sensing data, the cluster completes state estimation and remaining lifespan prediction for all devices within 0.5 seconds, and runs an attention mechanism network to identify common degradation patterns across devices (such as early wear in a batch of bearings). The edge nodes cache early warning information and aggregate it to generate a workshop-level safety situation heatmap, which is then uploaded to the enterprise cloud platform to achieve collaborative monitoring and rapid decision-making at the workshop level.

[0125] In this embodiment, the enterprise cloud platform maintains a global knowledge graph of industrial safety regulations containing 5,000 safety regulations (such as GB 30871 Hazardous Chemicals Operation Standard) and 2,000 historical accident cases. At the beginning of each month, the cloud platform automatically collects rare fault data reported by edge nodes in each workshop (such as "bearing seizure rate increases by 15% in winter"). It then updates the weight parameters of the causal inference engine through incremental training, strengthening the confidence of the causal link "low temperature - lubrication failure - bearing failure". The updated model parameters are encrypted using the national cryptographic standard SM4, packaged, and distributed to edge nodes in each workshop via a dedicated VPN channel. After verifying the integrity of the parameter signature, the edge nodes perform hot-loading updates during nighttime production breaks, improving the ability to identify seasonal faults without downtime and continuously empowering edge-local capabilities with global knowledge from the cloud.

[0126] In step S704, a three-tiered data security flow control mechanism is established: The reactor at the edge only uploads an abnormal operating condition summary (data size <200 bytes) containing timestamps, fault codes, and location coordinates via the industrial firewall when a Level 3 warning (emergency shutdown level) is detected; edge nodes upload daily health reports (approximately 5KB) every morning, including health scores, warning frequencies, and maintenance recommendations for each piece of equipment in the workshop; the cloud only issues commands to retrieve the reactor's complete vibration spectrum and temperature curves for the past 30 days (approximately 50MB) during accident investigations or model training. This mechanism reduces daily communication traffic to less than 5% of the traditional continuous upload mode, saving over 300,000 RMB in communication costs annually. Simultaneously, encryption and access control ensure the security of production data during transmission and storage, avoiding the risk of sensitive process information leakage.

[0127] The edge-cloud collaborative security architecture achieves hard real-time response for critical tasks through lightweight edge-side processing, while FPGA hardware acceleration ensures millisecond-level interference suppression and causal inference, meeting the stringent timeliness requirements of emergency security control. The parallel computing capabilities of the edge-side GPU cluster support the real-time operation of complex models such as digital twin evolutionary networks, balancing the computational load through model partitioning and avoiding single-point computing power bottlenecks. Centralized maintenance of the knowledge graph and execution of incremental training in the cloud enable continuous accumulation of global security experience and iterative improvement of model generalization capabilities, ensuring the long-term evolution of the monitoring system. A three-level data flow control mechanism specifically optimizes communication efficiency and security, prioritizing limited bandwidth for anomaly warnings and critical commands, significantly reducing operating costs and the risk of information leakage. The overall architecture achieves efficient and secure collaboration of "real-time edge processing, edge collaborative analysis, and global cloud optimization," significantly improving the system's deployment flexibility, operational economy, and security reliability in industrial environments with limited communication and heterogeneous computing resources, providing a scalable technical paradigm for large-scale industrial IoT security monitoring.

[0128] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0129] In one embodiment, a multimodal sensing industrial process anomaly safety control device is provided, which corresponds one-to-one with the multimodal sensing industrial process anomaly safety control method described in the above embodiments. For example... Figure 2 As shown, the multimodal perception industrial process anomaly safety control device includes an industrial safety dataset generation unit 1, which is used to collect equipment operating parameters, environmental state variables, material flow attribute data and personnel operation behavior video streams in the industrial process through a multimodal heterogeneous sensor array. After spatiotemporal alignment and multi-scale feature fusion, it generates an industrial safety dataset with standardized monitoring semantics.

[0130] The enhanced sensing data stream generation unit 2 is used to input the industrial safety dataset into the dynamic industrial environment interference suppression layer. This layer estimates the field data of the work area in real time through a differentiable industrial field coupling propagation model and generates an enhanced sensing data stream after interference suppression. The field data includes temperature drift field, electromagnetic noise field and mechanical vibration velocity field.

[0131] Equipment health and operational capability assessment unit 3 is used to send the enhanced sensing data stream into the industrial process digital twin evolution network. The industrial process digital twin evolution network constructs a coupled dynamic model of equipment mechanism based on a physical information driven architecture and embeds multi-objective coupled constraints of production cycle time, energy efficiency and safety, and outputs the equipment health evolution trend characteristics and remaining safe operational capability assessment value.

[0132] The adaptive adjustment instruction output unit 4 is used to align the health evolution characteristics with the industrial safety procedure knowledge graph, and to evaluate the current work process deviation and safety procedure compliance in real time through an interpretable causal reasoning engine, and output risk compensation amount and adaptive adjustment instructions for process parameters.

[0133] The dynamic response control instruction generation unit 5 is used to input the risk compensation instruction and the remaining safe operation capability assessment value into the safety control decision module. This module introduces time-dependent constraints in the attention mechanism, generates hierarchical safety early warning information, and triggers proactive process refactoring and dynamic response control instructions.

[0134] Specific limitations regarding the multimodal sensing industrial process anomaly safety control device can be found in the limitations of the multimodal sensing industrial process anomaly safety control method described above, and will not be repeated here. Each module in the aforementioned multimodal sensing industrial process anomaly safety control device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in the electronic device, or stored in the memory of the electronic device as software, so that the processor can call and execute the corresponding operations of each module.

[0135] In one embodiment, an electronic device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 3 As shown, this electronic device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores the database. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a multimodal sensing method for industrial process anomaly safety control.

[0136] In one embodiment, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps:

[0137] By collecting equipment operating parameters, environmental state variables, material flow attribute data, and personnel operation behavior video streams in industrial processes through multimodal heterogeneous sensor arrays, and generating standardized monitoring semantic industrial safety datasets after spatiotemporal alignment and multi-scale feature fusion.

[0138] The industrial safety dataset is input into the dynamic industrial environment interference suppression layer. This layer estimates the field data of the work area in real time through a differentiable industrial field coupling propagation model, and generates an enhanced sensing data stream after interference suppression. The field data includes temperature drift field, electromagnetic noise field and mechanical vibration velocity field.

[0139] The enhanced sensing data stream is fed into the industrial process digital twin evolution network. The industrial process digital twin evolution network constructs a device mechanism coupled dynamics model based on a physical information driven architecture and embeds multi-objective coupled constraints of production cycle time, energy efficiency and safety, and outputs the device health evolution trend characteristics and remaining safe operation capability assessment value.

[0140] The health evolution characteristics are aligned with the industrial safety procedure knowledge graph. An interpretable causal reasoning engine is used to evaluate the current work process deviation and safety procedure compliance in real time, and outputs risk compensation and adaptive adjustment instructions for process parameters.

[0141] The risk compensation instruction and the remaining safe operation capability assessment value are input into the safety control decision module. This module introduces time-dependent constraints into the attention mechanism, generates hierarchical safety early warning information, and triggers proactive process refactoring and dynamic response control instructions.

[0142] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0143] By collecting equipment operating parameters, environmental state variables, material flow attribute data, and personnel operation behavior video streams in industrial processes through multimodal heterogeneous sensor arrays, and generating standardized monitoring semantic industrial safety datasets after spatiotemporal alignment and multi-scale feature fusion.

[0144] The industrial safety dataset is input into the dynamic industrial environment interference suppression layer. This layer estimates the field data of the work area in real time through a differentiable industrial field coupling propagation model, and generates an enhanced sensing data stream after interference suppression. The field data includes temperature drift field, electromagnetic noise field and mechanical vibration velocity field.

[0145] The enhanced sensing data stream is fed into the industrial process digital twin evolution network. The industrial process digital twin evolution network constructs a device mechanism coupled dynamics model based on a physical information driven architecture and embeds multi-objective coupled constraints of production cycle time, energy efficiency and safety, and outputs the device health evolution trend characteristics and remaining safe operation capability assessment value.

[0146] The health evolution characteristics are aligned with the industrial safety procedure knowledge graph. An interpretable causal reasoning engine is used to evaluate the current work process deviation and safety procedure compliance in real time, and outputs risk compensation and adaptive adjustment instructions for process parameters.

[0147] The risk compensation instruction and the remaining safe operation capability assessment value are input into the safety control decision module. This module introduces time-dependent constraints into the attention mechanism, generates hierarchical safety early warning information, and triggers proactive process refactoring and dynamic response control instructions.

[0148] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0149] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0150] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A multimodal sensing method for industrial process anomaly safety control, characterized in that, The method includes the following steps: collecting equipment operating parameters, environmental state variables, material flow attribute data and personnel operation behavior video streams in industrial processes through a multimodal heterogeneous sensor array, and generating an industrial safety dataset with standardized monitoring semantics after spatiotemporal alignment and multi-scale feature fusion. The industrial safety dataset is input into the dynamic industrial environment interference suppression layer. This layer estimates the field data of the work area in real time through a differentiable industrial field coupling propagation model, and generates an enhanced sensing data stream after interference suppression. The field data includes temperature drift field, electromagnetic noise field and mechanical vibration velocity field. The enhanced sensing data stream is fed into the industrial process digital twin evolution network. The industrial process digital twin evolution network constructs a device mechanism coupled dynamics model based on a physical information driven architecture and embeds multi-objective coupled constraints of production cycle time, energy efficiency and safety, and outputs the device health evolution trend characteristics and remaining safe operation capability assessment value. The health evolution characteristics are aligned with the industrial safety procedure knowledge graph. An interpretable causal reasoning engine is used to evaluate the current work process deviation and safety procedure compliance in real time, and outputs risk compensation and adaptive adjustment instructions for process parameters. The risk compensation instruction and the remaining safe operation capability assessment value are input into the safety control decision module. This module introduces time-dependent constraints into the attention mechanism, generates hierarchical safety early warning information, and triggers proactive process refactoring and dynamic response control instructions.

2. The multimodal sensing method for industrial process anomaly safety control according to claim 1, characterized in that, The step of generating a standardized industrial safety dataset with monitoring semantics by collecting equipment operating parameters, environmental state variables, material flow attribute data, and personnel operation behavior video streams from industrial processes through multimodal heterogeneous sensor arrays, followed by spatiotemporal alignment and multi-scale feature fusion, includes the following steps: Vibration acceleration sensors, temperature thermocouples, pressure transmitters and current sensors are deployed on the body of industrial equipment, and visual inspection cameras and laser rangefinders are integrated on the material conveyor belt. An industrial IoT communication channel detection submodule is constructed, which estimates the channel delay spread and packet loss rate parameters by sending orthogonal frequency division multiplexing pilot signals and receiving echoes. The collected equipment operating parameters, environmental variables, material properties, and behavioral video streams are time-stamped and aligned. A dynamic time warping algorithm is used to eliminate timing mismatch caused by differences in the sampling frequencies of multimodal sensors. The aligned heterogeneous data are fused into multi-scale feature volumes under a unified monitoring semantic framework to generate a standardized industrial safety dataset containing equipment operating condition information, material status information, and personnel behavior information.

3. The multimodal sensing method for industrial process anomaly safety control according to claim 2, characterized in that, The step of inputting the industrial safety dataset into the dynamic industrial environment interference suppression layer, which estimates the field data of the work area in real time through a differentiable industrial field coupling propagation model and generates an enhanced sensing data stream after interference suppression, includes the following steps: A differentiable industrial field coupling propagation sub-model is constructed, and the point cloud data of equipment layout is used as spatial constraints to inversely deduce the eddy current attenuation of electromagnetic noise in metal structures and the convective diffusion intensity distribution of temperature field in air medium. Based on the inferred intensity distribution, the temperature drift field, electromagnetic noise field, and mechanical vibration velocity field are estimated. The temperature field is modeled as a learnable three-dimensional implicit representation, and a smoothing constraint is applied through the heat conduction equation. Based on the estimated interference field distribution, the temperature drift deviation of the sensor signal and the vibration noise floor are dynamically compensated to generate an enhanced sensing data stream after interference suppression. A rendering consistency self-supervised mechanism is introduced, in which the compensated data stream is re-input into the propagation model to generate simulated sensor readings, and the residual loss is calculated with the original readings to update the model parameters, thereby achieving unsupervised online learning.

4. The multimodal sensing method for industrial process anomaly safety control according to claim 3, characterized in that, The step of feeding the enhanced sensing data stream into the industrial process digital twin evolution network, which constructs a coupled dynamic model of equipment mechanism based on a physical information-driven architecture and embeds multi-objective coupled constraints of production cycle time, energy efficiency, and safety, and outputs the equipment health evolution trend characteristics and remaining safe operating capability assessment value, includes the following steps: A coupled dynamic sub-model of equipment mechanism is constructed, and the rotating machinery rotor system and the transmission gearbox are respectively modeled as a rigid multibody dynamic element and a nonlinear meshing damping system. The equation of motion is established through the principle of virtual work. The dynamic model incorporates multi-objective coupled constraints of production cycle time, energy efficiency, and safety. Based on the production process control theory, the dynamic balance coefficient between equipment load rate and safety threshold is calculated and fed back to the dynamic equation to correct the boundary conditions. Using the enhanced sensing data stream as the observation input, the unscented Kalman filter algorithm is used to recursively estimate the hidden state of the dynamic model, and outputs the evolution trend of bearing torque, gear contact stress and wear state of key components. Based on the current state estimate, the performance degradation trajectory of the equipment in the remaining production cycle is projected forward, and the remaining safe operation capability assessment value and health index are output.

5. The multimodal sensing method for industrial process anomaly safety control according to claim 4, characterized in that, The step of aligning the health evolution characteristics with the industrial safety procedure knowledge graph, evaluating the current work process deviation and safety procedure compliance in real time through an interpretable causal reasoning engine, and outputting risk compensation and adaptive adjustment instructions for process parameters includes the following steps: The equipment health evolution characteristics are mapped to the production process cycle point sequence, and subgraph isomorphic matching is performed with the pre-stored industrial safety procedure knowledge graph to construct a process-safety alignment graph structure. Run an interpretable causal reasoning engine on the alignment graph to identify key state variables that cause process deviations through causal discovery algorithms and quantify the contribution of each variable to safety deviations. Based on the contribution analysis results, the risk compensation amount of production cycle time and process parameters is calculated online, and the compensation amount is decomposed into equipment load fine-tuning component and material flow rate replanning component. When causal reasoning discovers that multivariate coupling leads to systemic safety risks, it triggers an adaptive adjustment command for process parameters, dynamically modifying equipment speed, feed rate, and personnel work intervals to ensure that the overall safety compliance is maintained above the qualified threshold.

6. The multimodal sensing method for industrial process anomaly safety control according to claim 1, characterized in that, The process of inputting the risk compensation instruction and the remaining safe operation capability assessment value into the safety control decision module, which introduces time-dependent constraints into the attention mechanism to generate hierarchical safety early warning information and trigger proactive process refactoring and dynamic response control instructions, includes the following steps: The remaining safe operation capability assessment value, the historical sequence of risk compensation instructions, and the evolution trend of wear status of key components are encoded into temporal feature vectors and input into the attention mechanism network; By introducing temporal dependency constraints into the attention mechanism, the attention weights are forced to follow an autoregressive process, which suppresses overfitting to random disturbances and enhances attention to progressive failure modes. The graded safety warning generator outputs three levels of warning information based on the intensity of the degradation mode of attention focus: normal operation monitoring level, preventive maintenance level, and emergency shutdown level; When the warning level reaches the preventive maintenance level, the proactive process refactoring instruction is automatically triggered. The digital twin simulates strategies for slowing down production, switching equipment, or bypassing processes, and selects the optimal refactoring solution for execution.

7. The multimodal sensing method for industrial process anomaly safety control according to claim 1, characterized in that, An industrial IoT edge-cloud collaborative security architecture is set up, including the following steps: A lightweight interference suppression layer and a causal inference engine are deployed on the industrial equipment side. The forward calculation of the industrial field propagation model is realized by using a field-programmable gate array to meet the millisecond-level real-time requirements. A digital twin evolutionary network and a safety early warning module are run on the edge computing node in the workshop. The dynamic model is deployed on the graphics processor cluster through a model parallel partitioning strategy. The enterprise cloud platform maintains a global knowledge graph of industrial safety procedures and regularly collects rare fault data from the edge to perform incremental training. The updated model parameters are then sent to the edge via an encrypted channel. Establish a three-level data security flow control mechanism at the cloud, edge, and terminal levels. The terminal only uploads a summary of abnormal operating conditions, the edge terminal uploads a daily health report, and the cloud pulls all historical data on demand, reducing communication bandwidth consumption and ensuring data security.

8. A multimodal sensing industrial process anomaly safety control device, characterized in that, The device includes an industrial safety dataset generation unit (1), which is used to collect equipment operating parameters, environmental state variables, material flow attribute data and personnel operation behavior video streams in the industrial process through a multimodal heterogeneous sensor array, and generate an industrial safety dataset with standardized monitoring semantics after spatiotemporal alignment and multi-scale feature fusion. The enhanced sensing data stream generation unit (2) is used to input the industrial safety dataset into the dynamic industrial environment interference suppression layer. This layer estimates the field data of the work area in real time through a differentiable industrial field coupling propagation model and generates an enhanced sensing data stream after interference suppression. The field data includes temperature drift field, electromagnetic noise field and mechanical vibration velocity field. Equipment health and operational capability assessment unit (3) is used to send the enhanced sensing data stream into the industrial process digital twin evolution network. The industrial process digital twin evolution network constructs a device mechanism coupled dynamics model based on a physical information driven architecture and embeds production cycle time-energy efficiency-safety multi-objective coupled constraints, and outputs the device health evolution trend characteristics and remaining safe operational capability assessment value. The adaptive adjustment instruction output unit (4) is used to align the health evolution characteristics with the industrial safety procedure knowledge graph, evaluate the current work process deviation and safety procedure compliance in real time through an interpretable causal reasoning engine, and output the risk compensation amount and process parameter adaptive adjustment instructions. The dynamic response control instruction generation unit (5) is used to input the risk compensation instruction and the remaining safe operation capability assessment value into the safety control decision module. This module introduces time-dependent constraints in the attention mechanism, generates hierarchical safety early warning information, and triggers proactive process reconstruction and dynamic response control instructions.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of a multimodal sensing industrial process anomaly safety method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of a multimodal sensing industrial process anomaly safety method as described in any one of claims 1 to 7.

Images