Dangerous waste storage environment real-time monitoring and risk early warning system based on digital twinning

Abstract

The invention relates to the technical field of environmental monitoring of artificial intelligence, and particularly discloses a dangerous waste storage environment real-time monitoring and risk early warning system based on digital twinning, which collects multi-modal environmental data in real time through an intelligent sensor network deployed in a dangerous waste storage facility, and constructs a digital twinning model synchronized with a physical entity; a multi-modal data fusion and physical embedding technology is adopted, and virtual risk parameters of an area which is not actually measured are calculated based on an environment evolution rule; an abnormal association mode between the measured data and the virtual risk parameters is analyzed and identified through the dynamic association network, and risk level judgment and early warning signal triggering are achieved; a sensor monitoring strategy is dynamically adjusted according to an early warning result to form closed-loop management and control; according to the invention, the sensor limitation of the traditional monitoring system is broken through, the advanced accurate early warning of the hidden risk is realized, and the monitoring efficiency is obviously improved through adaptive resource configuration.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring technology using artificial intelligence, specifically to a real-time monitoring and risk warning system for hazardous waste storage environments based on digital twins. Background Technology

[0002] Currently, environmental monitoring of hazardous waste storage mainly relies on physical sensor networks deployed within the storage facilities. Risk warnings are achieved through real-time acquisition and threshold determination of environmental parameters such as temperature, humidity, and specific gas concentrations. However, its monitoring capabilities are strictly limited by the location and types of physical sensors being monitored. It is ineffective for areas not covered by sensors or for hazardous substances that existing sensor technology cannot directly detect (such as unknown gases associated with certain chemical reactions), creating "sensor blind spots." Existing systems are mostly static or semi-static, with sensor monitoring strategies and operating parameters (such as sampling frequency and range) typically fixed. This prevents dynamic adjustments based on actual risk conditions, leading to resource waste during non-risk periods and potential missed alarms during high-risk periods due to insufficient monitoring density or accuracy. Consequently, efficient and accurate allocation of monitoring resources is difficult to achieve.

[0003] One technical problem solved by this invention is:

[0004] Traditional monitoring systems rely entirely on readings from a limited number of physical sensors, resulting in inherent blind spots in their "field of view." When hazardous waste undergoes slow, complex chemical reactions (such as generating asphyxiating gases without triggering specific gas sensors), or when multiple sensor readings are collectively misjudged due to environmental anomalies, the system fails due to a lack of insight into the physicochemical processes within the blind spots. The core of this invention lies in achieving "virtual sensing" capabilities by constructing a digital twin and deeply integrating physical mechanisms. This means that based on limited measured data and physicochemical laws, risk parameters within the sensor blind spots can be calculated, transforming risk perception from passively "monitoring existing anomalies" to proactively "predicting future risks," thus solving the problems of missed and delayed early warnings caused by insufficient sensing capabilities. Summary of the Invention

[0005] The purpose of this invention is to provide a real-time monitoring and risk warning system for hazardous waste storage environment based on digital twins, so as to solve the problems mentioned above.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A real-time monitoring and risk warning system for hazardous waste storage environments based on digital twins includes:

[0008] The hazardous waste environmental data acquisition module collects environmental status data in real time through intelligent sensors deployed in hazardous waste storage facilities and records it as measured data;

[0009] The multimodal data fusion processing module is used to extract and fuse cross-modal features from measured data, and based on preset environmental evolution rules, to deduce virtual risk parameters of unmeasured data from the measured data, generating a comprehensive environmental situation characterization.

[0010] The virtual sensing and risk simulation module is used to perform dynamic analysis of the comprehensive environmental situation based on time-series changes. By identifying abnormal correlation patterns between measured data and virtual risk parameters, it determines the risk level of the hazardous waste storage environment and triggers corresponding early warning signals.

[0011] The dynamic risk identification and early warning module dynamically adjusts the monitoring strategy and working parameter combination of the smart sensor based on the triggered early warning signal, and feeds back the optimized configuration instructions to the hazardous waste environmental data acquisition module.

[0012] As a further aspect of the present invention: the comprehensive environmental situation characterization generated by deriving virtual risk parameters from measured data to obtain unmeasured data specifically includes:

[0013] A multi-level spatiotemporal causal graph structure for hazardous waste storage environment is constructed. Measured data of temperature, gas concentration, and humidity are used as observation nodes, while chemical reaction processes and fluid transport processes are used as hidden nodes. Multi-scale interaction relationships between different nodes are established through directed edges in the graph structure.

[0014] Based on the spatiotemporal causal graph structure, the reverse inference calculation of physical constraints is performed. When a data combination pattern of a specific observation node is detected, bidirectional probability propagation is carried out along the directed edges of the graph to solve the hidden node state distribution that explains the current measured data pattern.

[0015] Based on the hidden node state distribution obtained from the solution, a virtual risk parameter spatial field is generated for the corresponding unmeasured area, including the chemical substance concentration gradient field, reaction rate distribution field, and heat release intensity field. These virtual risk parameters are then spatiotemporally aligned and fused with the measured data to form a comprehensive environmental situation characterization.

[0016] As a further aspect of the present invention: the bidirectional probability propagation along the directed edges of the graph to solve for the hidden node state distribution that interprets the current measured data pattern specifically includes:

[0017] Construct a dynamic causal strength matrix and dynamically adjust the connection strength between nodes based on the time series characteristics of the measured data;

[0018] The system performs iterative calculations of forward causal deduction and reverse evidence propagation. In the forward phase, the state confidence is passed along the directed edge starting from the observation node. In the reverse phase, the evidence weight is passed back from the abnormal pattern node. The uncertainty of the node state is gradually eliminated through bidirectional information interaction.

[0019] The bidirectional propagation results within continuous time segments are dynamically weighted and integrated to generate a hidden node state probability distribution with spatiotemporal continuity.

[0020] Based on the converged probability distribution of the hidden node states, the maximum posterior probability states of key hidden nodes and their spatial correlation patterns are extracted to form a complete mathematical description of the risk evolution process in unmonitored areas.

[0021] As a further aspect of the present invention: the step of generating a virtual risk parameter space field corresponding to the unmeasured region based on the solved hidden node state distribution specifically includes:

[0022] Establish the field strength coupling relationship between the hidden node and its spatial location, and transform the state probability value of the hidden node into the intensity distribution on the spatial basis function;

[0023] A multi-scale potential energy field is constructed based on spatial basis functions, and the propagation path and diffusion intensity of virtual parameters are determined by calculating the potential energy field gradient.

[0024] Perform virtual parameter diffusion driven by the potential energy field gradient, and propagate iteratively along the potential energy channel in multiple rounds;

[0025] By adaptively fusing the results of multiple diffusion rounds with the boundary conditions of measured data, and through boundary consistency verification and spatial smoothing, a continuous virtual risk parameter space field with physical rationality is generated.

[0026] As a further aspect of the present invention: the step of determining the risk level of the hazardous waste storage environment and triggering corresponding early warning signals by identifying abnormal correlation patterns between measured data and virtual risk parameters specifically includes:

[0027] Real-time measured data and virtual risk parameters are used as correlation nodes. Dynamic weighted connections are established by calculating the time-varying correlation strength between nodes. Abnormal correlation patterns are characterized by abrupt changes in the connection weights between nodes.

[0028] Starting from the initial abnormal node, trace the abnormal propagation trajectory along the strong connection path in the dynamic interconnection network, and record the key nodes passed through during the abnormal propagation process and their impact intensity.

[0029] The warning level threshold is dynamically adjusted based on the coverage of the abnormal propagation trajectory and the impact intensity of key nodes. When the trajectory coverage exceeds the spatial threshold and the impact intensity exceeds the dynamic threshold, the corresponding warning signal is triggered.

[0030] The key node sequences and their impact intensity information in the abnormal propagation trajectory are encoded into metadata of the early warning signal, forming complete early warning information containing the risk tracing path.

[0031] As a further aspect of the present invention: the step of tracing the anomaly propagation trajectory along the strong connection path in the dynamic association network from the initial anomaly node, and recording the key nodes passed through during the anomaly propagation process and their influence intensity, specifically includes:

[0032] Starting from the initial abnormal node, based on the real-time data of connection weights in the dynamic association network, connections with weights exceeding the dynamic threshold are selected as candidate propagation paths.

[0033] Deep exploration is conducted along candidate propagation paths. The path weights are recalculated and their persistence is verified at each level of propagation. Only paths that continuously satisfy the weight conditions are retained as valid propagation branches.

[0034] Construct a dynamic graph of the propagation trajectory, integrate the validated effective propagation branches in time series, and record the key nodes and their cumulative impact strength on each branch;

[0035] Generate anomaly propagation trajectories with spatiotemporal characteristics, sort key nodes in the dynamic graph according to propagation time sequence, and label the spatial location information and influence intensity parameters of each node to form a complete anomaly propagation trajectory description.

[0036] As a further aspect of the present invention: the construction of the dynamic spectrum of the propagation trajectory integrates the verified effective propagation branches according to the time series, and records the key nodes and their cumulative influence intensity on each branch, specifically including:

[0037] The initial impact intensity data of key nodes are normalized to convert the impact intensity values ​​of different dimensions into a standard range of 0 to 1, forming a standardized initial impact intensity.

[0038] Dynamic weights are determined based on the correlation strength and propagation delay of adjacent nodes in the propagation path;

[0039] The degree of attenuation of the influence intensity is calculated layer by layer according to the time sequence of the abnormal propagation;

[0040] Starting from the initial node, the cumulative influence intensity is accumulated step by step along the propagation path. The cumulative influence intensity of each node is equal to its standardized initial influence intensity plus the sum of the influence intensity of all predecessor nodes after attenuation and weight adjustment.

[0041] As a further aspect of the present invention: the dynamic adjustment of the monitoring strategy and operating parameter combination of the intelligent sensor specifically includes:

[0042] Analyze the spatial location information and risk type characteristics contained in the early warning signals, and determine the scope and boundaries of key monitoring areas based on spatial impact factors;

[0043] Based on the characteristics of risk type, configure differentiated monitoring mode combinations. For gas diffusion risk, activate high-frequency sampling and continuous monitoring mode, and for temperature anomaly risk, activate multi-sensor collaborative verification mode.

[0044] Based on the coverage of key monitoring areas, the smart sensors in non-core areas adopt an energy-saving operation strategy, and achieve dynamic balance of monitoring resources by adjusting the sampling interval and transmission frequency.

[0045] The optimized monitoring strategy is transformed into the operating parameters of a single smart sensor, including sampling frequency, measurement range, trigger threshold, and communication protocol parameters, and the rationality of the configuration is verified through a digital twin system.

[0046] As a further aspect of the present invention: the intelligent sensor covering non-core areas based on key monitoring areas adopts an energy-saving operation strategy, which achieves dynamic balance of monitoring resources by adjusting the sampling interval and transmission frequency, specifically including:

[0047] The boundaries of non-core areas are determined based on the spatial distribution characteristics of key monitoring areas, and non-core areas are divided into monitoring sub-areas of different priorities based on spatial correlation analysis.

[0048] Initial energy-saving operation parameters are calculated based on the spatial distance and risk propagation path relationship between each monitoring sub-area and the key monitoring area;

[0049] The sampling interval of non-core areas is dynamically adjusted based on real-time risk assessment results. When the risk assessment value is lower than the set threshold, the sampling interval is gradually extended and the data transmission frequency is reduced according to a preset ratio.

[0050] Monitor the quality of the adjusted data collection to ensure that the fluctuation range of key parameters is within the allowable range, thus forming a closed-loop resource optimization allocation.

[0051] The beneficial effects of this invention are:

[0052] (1) This invention uses a multimodal data fusion processing module to construct a spatiotemporal causal graph structure and perform bidirectional probabilistic inference by taking the physical laws of hazardous waste chemical reactions as the core constraint. Based on limited measured data (such as weak temperature anomalies and humidity changes), it can "infer" the chemical processes occurring in the sensor blind zone and calculate virtual risk parameters (such as asphyxiating gas concentration field and reaction rate distribution). This effect elevates risk perception from passively "monitoring what has happened" to actively "predicting what will happen", realizing early warning of hidden risks and solving the problem of early warning blind zones caused by limited sensor deployment or incomplete monitoring types.

[0053] (2) After the dynamic risk identification and early warning module determines the risk level and type, the intelligent sensing configuration adaptive module dynamically adjusts the strategy of the entire sensor network accordingly. Specifically, the system analyzes the spatial positioning information of the early warning signal, initiates high-frequency sampling and multi-sensor collaborative verification mode in the core risk area, and adopts energy-saving operation strategy (such as extending the sampling interval) in non-core areas. This resource allocation ensures that the monitoring force is always concentrated in the most critical parts, while reducing the overall energy consumption of the system during safe periods, forming an intelligent closed loop from risk identification to resource scheduling, thereby improving the operating efficiency and long-term endurance of the entire system while ensuring monitoring accuracy. Attached Figure Description

[0054] The invention will now be further described with reference to the accompanying drawings.

[0055] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Please see Figure 1 As shown, this invention is a real-time monitoring and risk early warning system for hazardous waste storage environments based on digital twins, comprising:

[0058] The hazardous waste environmental data acquisition module collects environmental status data in real time through intelligent sensors deployed in hazardous waste storage facilities and records it as measured data;

[0059] The multimodal data fusion processing module is used to extract and fuse cross-modal features from measured data, and based on preset environmental evolution rules, to deduce virtual risk parameters of unmeasured data from the measured data, generating a comprehensive environmental situation characterization.

[0060] The virtual sensing and risk simulation module is used to perform dynamic analysis of the comprehensive environmental situation based on time-series changes. By identifying abnormal correlation patterns between measured data and virtual risk parameters, it determines the risk level of the hazardous waste storage environment and triggers corresponding early warning signals.

[0061] The dynamic risk identification and early warning module dynamically adjusts the monitoring strategy and working parameter combination of the smart sensor based on the triggered early warning signal, and feeds back the optimized configuration instructions to the hazardous waste environmental data acquisition module.

[0062] In the hazardous waste environmental data acquisition module, multiple types of intelligent sensors are deployed within the hazardous waste storage facility to monitor and collect data on the storage environment in real time. These sensors include, but are not limited to, temperature sensors, humidity sensors, gas concentration sensors, liquid level sensors, pressure sensors, and high-definition cameras. These sensors are deployed in key locations within the storage facility according to a spatial three-dimensional distribution principle, including inside hazardous waste containers, storage areas, ventilation ducts, and drainage ditches, forming a complete spatial monitoring network. The data acquisition process is synchronously triggered through the sensor network, ensuring that data from each monitoring point has a unified time reference and spatial coordinates.

[0063] The real-time collected environmental data constitutes the system's measured data, specifically including the following categories: temperature data reflects the thermal state of the storage environment; humidity data characterizes the air humidity level; gas concentration data monitors the volume concentration of specific hazardous gases; liquid level data records the height of hazardous waste materials inside the container; pressure data reflects the internal pressure state of the sealed container; and video data provides real-time visual information of the storage area. All measured data are accompanied by precise timestamps and spatial location identifiers to ensure the spatiotemporal traceability of the data.

[0064] In the multimodal data fusion processing module, when extracting cross-modal features from measured data, the monitoring data from different sources are first converted into time-series data in a unified format. Temperature data, gas concentration data, and humidity data are aligned according to their acquisition time points. Video data is processed using image processing techniques to extract visual feature sequences such as brightness changes and fog concentration. Feature extraction for different modal data uses the same sliding time window, with a window length set to 60 consecutive sampling points. During feature extraction, the gradient and fluctuation characteristics of temperature data are calculated, the distribution and abrupt change characteristics of gas concentration data are extracted, the spatial difference characteristics of humidity data are analyzed, and the texture and motion characteristics of video feature data are calculated. All feature values ​​are normalized to ensure they fall within the same numerical range.

[0065] The pre-defined environmental evolution rules are implemented by establishing a physicochemical knowledge base for hazardous waste storage environments. This knowledge base contains basic physicochemical property data of hazardous waste substances, possible chemical reaction types and reaction conditions between different substances, and the movement patterns of fluids within the storage environment. Specifically, it includes the thermal decomposition temperature range of substances, the concentration conditions required for oxidation reactions, the characteristics of acid-base neutralization reactions, and the movement patterns of gas diffusion. These rules are stored in the form of conditional statements; when monitoring data meets a specific combination of conditions, the corresponding environmental evolution rule will be activated. The condition part of the rule contains a logical combination of multiple monitoring parameters, and the conclusion part of the rule specifies the possible environmental change processes and their characteristic manifestations.

[0066] The estimation of virtual risk parameters is based on the correlation between environmental evolution rules and measured data characteristics. When a continuous upward trend in temperature data is detected, coupled with fluctuations in the concentration of a specific gas, the generation rate and distribution range of potentially unmonitored substances are estimated based on the corresponding chemical reaction relationships in the environmental evolution rules. The estimation process employs a weighted inference method, weighting and fusing the output results of different rules according to the reliability of the rules. For each virtual risk parameter, its possible value range and probability distribution are calculated simultaneously. The estimation results include the distribution of chemical substance concentrations, reaction rates, and heat generation intensity in the unmonitored area. These distribution data are represented in the form of a two-dimensional grid, with each grid cell containing an estimated risk parameter value for that location.

[0067] The construction of the multi-level spatiotemporal causal graph structure begins with the identification of observation nodes. Each sensor's monitoring location is considered an observation node, with node attributes including measured data such as temperature, gas concentration, and humidity at that location. Hidden nodes are set according to the chemical and physical processes defined in the environmental evolution rules; each hidden node represents a possible reaction or mass transport process. Directed edges between nodes are established based on the causal relationships between processes, with the edge direction indicating the direction of influence transmission. The edge weights are determined based on the probability of the process occurring and its degree of influence, quantified by analyzing the sequence and correlation of changes in different parameters in historical data. The graph structure comprises three levels: a data layer containing all observation nodes, a process layer containing chemical and physical processes directly related to the observation data, and an effect layer containing derivative effects triggered by the basic processes.

[0068] Backpropagation computation based on a spatiotemporal causal graph structure begins by identifying anomalous patterns in the observed nodes. An anomalous pattern refers to a combination of changes in the data from multiple observed nodes within a specific time window that conforms to a certain environmental evolution rule. When such a pattern is detected, bidirectional probability propagation computation is initiated. Forward computation starts from the observed nodes and propagates the confidence level of the observed data to the hidden nodes along directed edges. The confidence level depends on the degree of matching between the observed data and the expected pattern. Backward computation starts from the known anomalous pattern nodes and propagates evidence weights backward along directed edges to the relevant hidden nodes. The weight values ​​reflect the hidden node's explanatory power for the current anomalous pattern. Through multiple iterative computations, the state probabilities of each hidden node are gradually adjusted until the probability distribution tends to stabilize. In each iteration, the update of the node state probability considers the information propagated by all neighboring nodes and uses a weighted average method for fusion.

[0069] The distribution of hidden node states is determined by analyzing the convergence results of bidirectional probability propagation. The computation is considered to have reached convergence when the change in the state probabilities of all hidden nodes is less than a set threshold in three consecutive iterations. After convergence, the state probability of each hidden node represents the likelihood of the corresponding process occurring, and the probability values ​​are converted into a relative probability distribution through normalization. The identification of key hidden nodes is based on the magnitude of the node state probabilities and the node's centrality in the graph structure; nodes with probability values ​​higher than the threshold and high centrality are selected as key nodes. The states of key nodes are analyzed to extract their probability distribution characteristics, including probability peaks, distribution ranges, and trends. Based on the spatial distribution relationships of key nodes, a risk evolution path description for unmonitored areas is established, which includes information such as the risk origin location, propagation direction, and impact range.

[0070] The coupling of field strength between hidden nodes and their spatial locations is achieved through spatial basis functions. Each hidden node is associated with a set of spatial basis functions, the distribution of which is determined by the node type: Gaussian basis functions are used for chemical reaction processes, exponentially decaying basis functions for mass transport processes, and polynomial basis functions for thermodynamic processes. The center of each basis function corresponds to the spatial coordinates of the node, and the broadening of the basis functions is determined by the influence range of the process. The state probability values ​​of the hidden nodes are mapped to the corresponding spatial basis functions; the higher the probability value, the larger the amplitude of the basis function. The spatial basis functions of multiple nodes are superimposed on a two-dimensional plane to form a continuous spatial intensity distribution. The calculation of the intensity distribution considers the interaction between the basis functions, and the intensity values ​​in overlapping regions are processed using the principle of taking the larger value.

[0071] The construction of the multi-scale potential energy field is based on spatial intensity distribution data. At the coarse-grained scale, the monitoring area is divided into larger grid cells, and the average intensity value of each cell is calculated as the potential energy baseline value for that cell. At the fine-grained scale, each coarse grid is further subdivided into finer grids, and the potential energy values ​​of the finer grids are adjusted based on measured data from neighboring observation points. The gradient of the potential energy field is calculated using the central difference method. At each grid point, the potential energy difference between itself and its neighboring grid points is calculated. The magnitude of the difference represents the propagation driving force of the virtual parameter in that direction, and the direction of the difference determines the propagation path. By analyzing the distribution characteristics of the potential energy field gradient, the main channels and barrier regions for virtual parameter propagation are identified. The update frequency of the potential energy field is consistent with the acquisition frequency of the monitoring data to ensure timely reflection of changes in the environmental state.

[0072] The virtual parameter diffusion driven by the potential energy field gradient employs an iterative calculation method. In each iteration, the virtual parameter propagates from the high potential energy region to the low potential energy region, with the propagation amount proportional to the potential energy gradient. The virtual parameter value of each grid point is calculated based on the parameter values ​​of its neighboring grid points and the potential energy gradient, considering the attenuation effect of the parameter during propagation. The attenuation coefficient is determined based on the distance between grid points and the medium properties; the greater the distance, the more significant the attenuation. Iterative calculations continue until the following convergence condition is met: the change in virtual parameters at all grid points between two consecutive iterations is less than a set threshold, or the maximum number of iterations is reached. The converged virtual parameter field displays the parameter estimates at each spatial location. The spatial distribution of parameter values ​​should conform to the physical propagation laws, i.e., gradually decreasing from the source region outwards.

[0073] The fusion of multi-round diffusion results with measured data employs an adaptive weighting method. At grid points with measured data, virtual parameter values ​​are corrected according to the measured data, with the correction weight depending on the reliability and representativeness of the measured data. At grid points without measured data, the virtual parameter values ​​obtained from the diffusion calculation are retained. Boundary consistency verification ensures that at the boundaries of the monitoring area, the virtual parameter values ​​satisfy physical boundary conditions, such as zero normal gradient at closed boundaries and parameter values ​​consistent with background values ​​at open boundaries. Spatial smoothing employs an anisotropic filtering method, with smoothing along the potential energy gradient direction being less than in the vertical direction, preserving the main characteristics of the parameter distribution while eliminating abnormal fluctuations. The final generated virtual risk parameter spatial field is stored in the form of a two-dimensional matrix, where the rows and columns of the matrix correspond to spatial grid coordinates, and the matrix element values ​​are the estimated virtual parameter values ​​at that location.

[0074] The generation of a comprehensive environmental situational characterization is accomplished by integrating measured data and virtual risk parameters. First, a unified spatial reference system is established, mapping various data types onto the same two-dimensional grid system. Each grid cell contains multiple data layers: the measured data layer stores directly monitored values ​​such as temperature and gas concentration at that location; the virtual parameter layer stores calculated chemical concentration gradients, reaction rates, and heat release intensities; and the metadata layer stores the data source identifier and timestamp. Relationships are established between different data layers, so when a user queries a specific location, both the measured data and virtual parameters for that location are displayed simultaneously. The situational characterization is updated synchronously with monitoring data acquisition to ensure it reflects the latest environmental status. The characterization data is displayed using a layered visualization technique, with different colors and transparency settings for different data layers, facilitating user understanding of the overall environmental condition.

[0075] In the virtual sensing and risk simulation module, the construction of the dynamic correlation network begins with node definition. The measured data of each monitoring point and the corresponding virtual risk parameters are defined as network nodes. The connections between nodes are established by calculating the time-varying correlation strength, which is obtained by calculating the Pearson correlation coefficient between the data sequences of two nodes, using a sliding time window of 30 sampling points. The update frequency of connection weights is consistent with the data acquisition frequency. When a change in connection weight is detected to exceed three standard deviations of the historical average, the connection is marked as an anomalous connection. The identification of anomalous correlation patterns is based on the abrupt change characteristics of connection weights. When the connection weights of a node and three or more adjacent nodes change significantly simultaneously, that node is determined to be an initial anomalous node. Each node stores the historical connection weight records for the most recent 100 time points for analyzing weight change trends.

[0076] Tracking anomaly propagation trajectories begins with identifying the initial anomalous node. Determining the initial anomalous node requires meeting two conditions: the change in connection weights between this node and its neighboring nodes exceeds a dynamic threshold, and this change persists for at least three consecutive sampling periods. The dynamic threshold is calculated based on the weight fluctuations of this node over the last 50 time points, taking 2.5 times the historical weight standard deviation. Candidate propagation paths are selected based on real-time connection weight data, retaining only those with weight values ​​exceeding the dynamic threshold and a clear direction. During the path verification phase, each candidate path undergoes a three-level depth exploration, with each level verifying the persistence and directional consistency of the path weights. Establishing an effective propagation branch requires that the path weights remain stable over two consecutive time points, and that the propagation direction conforms to the physical laws of environmental evolution.

[0077] The construction of the dynamic propagation trajectory graph employs a temporal integration method. Validated effective propagation branches are arranged according to a time series, with each branch's time signature accurate to the second. Key nodes are identified based on their centrality during propagation, selecting nodes with a degree centrality higher than 0.7 as key nodes. The calculation of influence intensity considers three dimensions: the node's own anomaly degree, the number of connections the node has in the network, and the node's positional weight in the propagation path. The initial influence intensity value of each key node is calculated using a range normalization method, mapping the original influence intensity value to a range of 0 to 1. The normalization calculation uses a min-max scaling method to ensure comparability of data across different dimensions. The dynamic graph is updated synchronously with the anomaly propagation process; the graph structure is updated immediately upon detecting a new effective propagation branch.

[0078] The cumulative impact strength is calculated using a recursive accumulation algorithm. First, the initial impact strength of key nodes is normalized using the formula (current value - minimum value) / (maximum value - minimum value), where the minimum and maximum values ​​are taken from the node's historical impact strength data over the past 24 hours. Dynamic weights are determined based on the association strength and propagation delay of adjacent nodes. The association strength weight is the average of the connection weights of two nodes, and the propagation delay weight is calculated using an exponential decay function, decreasing by 0.8 times for each additional sampling period. Time decay calculation is performed layer by layer according to the time sequence of anomaly propagation: the decay coefficient is 1 for the first layer, 0.7 for the second layer, and 0.5 for the third layer and beyond. The recursive accumulation process starts from the initial node. The cumulative impact strength of each node equals its standardized initial impact strength plus the sum of the decayed and weighted impact strengths of all predecessor nodes, where the impact strength of predecessor nodes needs to be multiplied by the association strength weight and the time delay weight.

[0079] The warning level is determined based on the coverage area of ​​the abnormal propagation trajectory and the influence intensity of key nodes. The trajectory coverage area is obtained by calculating the number of grid cells traversed by the abnormal propagation trajectory. The spatial threshold is set as a proportion of the total monitored area: Level 1 warning corresponds to 5% coverage, Level 2 warning to 10%, and Level 3 warning to 20%. The dynamic threshold is adjusted considering the influence intensity distribution of key nodes; when the coefficient of variation of influence intensity exceeds 0.5, the threshold level is increased accordingly. Triggering a warning signal requires two conditions to be met simultaneously: the trajectory coverage area exceeds the currently set spatial threshold, and the average influence intensity of all key nodes exceeds the dynamic threshold. Threshold adjustments follow a gradual principle, with each adjustment not exceeding 20% ​​of the previous value to avoid frequent changes in the warning level due to short-term fluctuations.

[0080] The generation of early warning information includes a metadata encoding process. The key node sequence is arranged chronologically according to the anomaly propagation, with each node containing spatial coordinates, an impact intensity value, and a node type identifier. Spatial location information is represented using a three-dimensional coordinate system, accurate to 0.1 meters. The impact intensity parameter records the baseline value before normalization and the standard value after normalization. The description of the risk tracing path includes three elements: propagation direction, propagation speed, and impact range. The propagation direction is calculated using the spatial coordinates of the key node sequence, the propagation speed is calculated as the ratio of the distance between nodes to the time difference, and the impact range is determined by calculating the convex hull area of ​​the anomaly propagation trajectory. Complete early warning information is stored in a structured data format, containing four parts: a message header, a tracing path data block, an impact intensity data block, and a spatial distribution data block, ensuring the integrity and parseability of the information.

[0081] The early warning signal triggering process employs a multi-level verification mechanism. When the initial conditions are met, the verification procedure is initiated, with a verification cycle of three consecutive sampling points. During the verification period, the coverage area of ​​the abnormal propagation trajectory and the impact intensity of key nodes are continuously monitored. When both indicators maintain an upward trend during the verification period, the early warning signal is confirmed as triggered. The final determination of the early warning level is based on the average value during the verification period, the coverage area is the arithmetic mean of the three sampling points, and the impact intensity is the maximum value of the three sampling points. The early warning signal includes four basic elements: timestamp, early warning level, affected area, and expected duration. The timestamp is accurate to the second, the early warning level is represented by numbers 1-3, the affected area is described using a grid coordinate set, and the expected duration is estimated based on the speed and range of abnormal propagation, calculated by dividing the radius of the affected area by the propagation speed.

[0082] The metadata encoding adopts a hierarchical structure. The first layer contains basic information about the warning signal, including the signal identifier, generation time, and validity period. The second layer contains source tracing data, recording the spatial coordinates and timestamps of the key node sequence. The third layer stores impact intensity parameters, including the initial impact intensity, standardized impact intensity, and cumulative impact intensity of each key node. The fourth layer saves a snapshot of the environmental state, recording the measured data and virtual risk parameters of each monitoring point when the warning is triggered. The encoding process uses data compression technology, performing dictionary encoding on recurring patterns and differential encoding on numerical data. The final generated warning information data packet size is controlled within 1KB, ensuring transmission efficiency while retaining all necessary information.

[0083] The transmission of early warning information employs a reliable transmission protocol. Each early warning signal requires an acknowledgment from the receiver after transmission. If no acknowledgment is received within a specified time, a retransmission mechanism is initiated. The maximum number of retransmissions is three, with increasing intervals between each retransmission: 2 seconds for the first, 4 seconds for the second, and 8 seconds for the third. During transmission, data packets are checked and calculated, using a cyclic redundancy check algorithm to ensure data integrity. Early warning information has three priority levels, corresponding to different transmission quality requirements. Level 1 warnings use the highest priority, requiring a transmission delay of less than 100 milliseconds; Level 2 warnings require a transmission delay of less than 500 milliseconds; and Level 3 warnings require a transmission delay of less than 1 second. The transmission protocol supports resuming interrupted transmissions; when network interruptions are resolved, unsuccessfully transmitted early warning information can be automatically retransmitted.

[0084] In the dynamic risk identification and early warning module, the analysis of early warning signals begins with the extraction of spatial positioning information. This spatial positioning information is contained in the metadata of the early warning signal, represented in grid coordinates, with each grid cell measuring 1 meter × 1 meter. The analysis process first identifies all affected grid coordinates marked in the early warning signal, then calculates the distribution center point and dispersion of these coordinate points. The spatial impact factor is calculated based on two parameters: the reciprocal of the distance between the grid point and the risk source point, and the risk propagation channel weight at the grid point's location. The reciprocal distance weight is calculated according to an exponential decay law; for every additional meter of distance, the weight decreases by 0.7 times. The risk propagation channel weight is determined based on the propagation paths of risk factors in historical data: the weight coefficient for primary ventilation channels is 1.2, for secondary channels it is 0.8, and for obstruction areas it is 0.3. The scope of the key monitoring area is determined by connecting all grid points with a spatial impact factor greater than 0.6 to form a convex polygon, with boundary accuracy controlled within 0.5 meters.

[0085] The monitoring mode configuration employs differentiated strategies based on the characteristics of the risk type. For gas diffusion risks, a high-frequency sampling mode is activated, with the sampling interval adjusted from the usual 300 seconds to 30 seconds, and the data upload interval adjusted from 600 seconds to 60 seconds. The continuous monitoring mode requires the sensor to automatically switch to real-time transmission mode when it detects a change in concentration. For temperature anomaly risks, a multi-sensor collaborative verification mode is enabled. This mode requires the temperature sensor to establish a linkage with adjacent gas and humidity sensors. When the temperature sensor detects an anomaly, it automatically triggers all associated sensors within a 5-meter radius to enter enhanced monitoring mode, with the sampling frequency uniformly increased to three times the normal frequency. The monitoring mode configuration is issued in the form of command codes, which include three parts: mode identifier, parameter set, and effective time.

[0086] The implementation of energy-saving operation strategies begins with the delineation of non-core areas. The boundaries of non-core areas are determined by calculating spatial correlation, and buffer zones are formed by expanding outward from the boundaries of key monitoring areas. The width of the buffer zones is dynamically adjusted according to the risk type: a 10-meter buffer zone is set for gas diffusion risks, and a 5-meter buffer zone is set for temperature anomaly risks. Non-core areas are divided into three priority monitoring sub-areas: the first-level sub-area is closest to the key area and maintains the regular monitoring frequency; the sampling interval for the second-level sub-area is extended to 600 seconds; and the sampling interval for the third-level sub-area is extended to 1800 seconds. The initial energy-saving operation parameters are calculated based on the relationship between spatial distance and risk propagation paths. For every 1 meter increase in distance from the key area, the base value of the sampling interval increases by 50 seconds, while also considering the weighting adjustment of risk propagation paths.

[0087] The dynamic adjustment mechanism adjusts operating parameters based on real-time risk assessment results. The risk assessment value is updated every 300 seconds, taking into account the stability, trend, and fluctuation of monitoring data over the past 30 minutes. When the risk assessment value falls below 0.3, the energy-saving adjustment program is activated. The sampling interval is adjusted gradually, with each adjustment not exceeding 50% of the current value. The initial adjustment extends the sampling interval to 1.5 times its original value, and the data transmission frequency is correspondingly reduced to 2 / 3 of its original value. After each adjustment, three sampling cycles are observed to confirm that the monitoring data remains stable before proceeding with the next adjustment. During the adjustment process, protection thresholds are set to ensure that the sampling interval does not exceed 3600 seconds and the data transmission interval does not exceed 7200 seconds.

[0088] Data quality monitoring focuses on the fluctuation range of key parameters. Key parameters include temperature, specific gas concentrations, and humidity, with allowable fluctuation ranges determined based on historical data statistics. The allowable fluctuation range for temperature is ±15% of the average value of the most recent 24 hours; for gas concentration, it is ±25%; and for humidity, it is ±20%. During monitoring, data quality indicators for each sampling point are recorded, including data integrity, numerical reasonableness, and continuity of change. Data integrity requires a successful data acquisition rate of no less than 95%. Numerical reasonableness checks exclude outliers that significantly exceed physical possibilities. Continuity of change requires that the numerical changes at adjacent sampling points conform to the gradual change pattern of environmental parameters.

[0089] The conversion of monitoring strategies into operating parameters is achieved through a parameter mapping table. Sampling frequency is converted into specific sampling time intervals, the measurement range is adjusted based on the distribution of recent monitoring data, and trigger thresholds are set with different sensitivities according to the warning level. Communication protocol parameters include data transmission compression algorithms, encryption methods, and verification mechanisms. Digital twin verification is completed through simulation testing. A 24-hour simulation test is run in a virtual environment with the new parameter settings, monitoring the integrity of data transmission, response timeliness, and resource consumption in the virtual environment. During verification, the coordination of parameter settings is carefully checked to ensure that there are no conflicts in parameter configurations between different sensors, and that the total energy consumption of all sensors does not exceed the power supply capacity of the power system. Verified parameter configurations take effect immediately, and a detailed log of this adjustment is recorded.

[0090] The working principle of this invention is as follows: A closed-loop management process is constructed, from physical perception to virtual simulation and then to decision feedback. First, the system uses various intelligent sensors deployed at key locations in hazardous waste storage facilities to collect real-time environmental status data such as temperature, humidity, gas concentration, liquid level, pressure, and video images, forming a standardized multi-source monitoring dataset. Then, using multimodal data fusion and physical embedding technology, cross-modal features are extracted from the measured data. Based on preset physical laws of hazardous waste chemical reactions and environmental evolution rules, virtual risk parameters for unmeasured areas are calculated by constructing a spatiotemporal causal graph, performing bidirectional probability simulation, and potential field-driven diffusion calculations, generating a comprehensive environmental situation characterization that integrates measured and virtual data. Furthermore, abnormal correlation patterns between measured data and virtual risk parameters are analyzed through dynamic correlation network analysis to track abnormal propagation trajectories. Based on the trajectory coverage and the influence intensity of key nodes, the risk level is dynamically determined, and graded early warning signals are triggered. Finally, based on the risk characteristics analyzed from the early warning signals, the system dynamically adjusts the monitoring strategies and operating parameter combinations of the intelligent sensors to achieve optimized resource allocation for enhanced monitoring in key areas and energy-saving operation in non-core areas. The system also feeds back optimization instructions to the data acquisition terminal, forming a continuously self-optimizing intelligent closed loop of "perception-deduction-early warning-control", which significantly improves the accuracy and timeliness of environmental risk early warning for hazardous waste storage.

[0091] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A real-time monitoring and risk early warning system for hazardous waste storage environment based on digital twins, characterized in that, include: The hazardous waste environmental data acquisition module collects environmental status data in real time through intelligent sensors deployed in hazardous waste storage facilities and records it as measured data; The multimodal data fusion processing module is used to extract and fuse cross-modal features from measured data, and based on preset environmental evolution rules, to deduce virtual risk parameters of unmeasured data from the measured data and generate a comprehensive environmental situation characterization. The virtual sensing and risk simulation module is used to perform dynamic analysis of the comprehensive environmental situation based on time-series changes. By identifying abnormal correlation patterns between measured data and virtual risk parameters, it determines the risk level of the hazardous waste storage environment and triggers corresponding early warning signals. The dynamic risk identification and early warning module dynamically adjusts the monitoring strategy and working parameter combination of the smart sensor based on the triggered early warning signal, and feeds back the optimized configuration instructions to the hazardous waste environmental data acquisition module.

2. The real-time monitoring and risk early warning system for hazardous waste storage environment based on digital twin as described in claim 1, characterized in that, The process of deriving virtual risk parameters from measured data to generate a comprehensive environmental situation characterization specifically includes: A multi-level spatiotemporal causal graph structure for hazardous waste storage environment is constructed. Measured data of temperature, gas concentration, and humidity are used as observation nodes, while chemical reaction processes and fluid transport processes are used as hidden nodes. Multi-scale interaction relationships between different nodes are established through directed edges in the graph structure. Based on the spatiotemporal causal graph structure, the reverse inference calculation of physical constraints is performed. When a data combination pattern of a specific observation node is detected, bidirectional probability propagation is carried out along the directed edges of the graph to solve the hidden node state distribution that explains the current measured data pattern. Based on the hidden node state distribution obtained from the solution, a virtual risk parameter spatial field is generated for the corresponding unmeasured area, including the chemical substance concentration gradient field, reaction rate distribution field, and heat release intensity field. These virtual risk parameters are then spatiotemporally aligned and fused with the measured data to form a comprehensive environmental situation characterization.

3. The real-time monitoring and risk early warning system for hazardous waste storage environment based on digital twins as described in claim 2, characterized in that, The bidirectional probability propagation along the directed edges of the graph is used to solve for the hidden node state distribution that interprets the current measured data pattern, specifically including: Construct a dynamic causal strength matrix and dynamically adjust the connection strength between nodes based on the time series characteristics of the measured data; The system performs iterative calculations of forward causal deduction and reverse evidence propagation. In the forward phase, the state confidence is passed along the directed edge starting from the observation node. In the reverse phase, the evidence weight is passed back from the abnormal pattern node. The uncertainty of the node state is gradually eliminated through bidirectional information interaction. The bidirectional propagation results within continuous time segments are dynamically weighted and integrated to generate a hidden node state probability distribution with spatiotemporal continuity. Based on the converged probability distribution of the hidden node states, the maximum posterior probability states of key hidden nodes and their spatial correlation patterns are extracted to form a complete mathematical description of the risk evolution process in unmonitored areas.

4. The real-time monitoring and risk early warning system for hazardous waste storage environment based on digital twins according to claim 2, characterized in that, The step of generating a virtual risk parameter space field for the corresponding unmeasured region based on the solved hidden node state distribution specifically includes: Establish the field strength coupling relationship between the hidden node and its spatial location, and transform the state probability value of the hidden node into the intensity distribution on the spatial basis function; A multi-scale potential energy field is constructed based on spatial basis functions, and the propagation path and diffusion intensity of virtual parameters are determined by calculating the potential energy field gradient. Perform virtual parameter diffusion driven by the potential energy field gradient, and propagate iteratively along the potential energy channel in multiple rounds; By adaptively fusing the results of multiple diffusion rounds with the boundary conditions of measured data, and through boundary consistency verification and spatial smoothing, a continuous virtual risk parameter space field with physical rationality is generated.

5. The real-time monitoring and risk early warning system for hazardous waste storage environment based on digital twin as described in claim 1, characterized in that, The method of identifying abnormal correlation patterns between measured data and virtual risk parameters to determine the risk level of the hazardous waste storage environment and trigger corresponding early warning signals specifically includes: Real-time measured data and virtual risk parameters are used as correlation nodes. Dynamic weighted connections are established by calculating the time-varying correlation strength between nodes. Abnormal correlation patterns are characterized by abrupt changes in the connection weights between nodes. Starting from the initial abnormal node, trace the abnormal propagation trajectory along the strong connection path in the dynamic interconnection network, and record the key nodes passed through during the abnormal propagation process and their impact intensity. Based on the coverage of the abnormal propagation trajectory and the impact intensity of key nodes, the warning level determination threshold is dynamically adjusted. When the trajectory coverage exceeds the spatial threshold and the impact intensity exceeds the dynamic threshold, the corresponding level of warning signal is triggered. The key node sequences and their impact intensity information in the abnormal propagation trajectory are encoded into metadata of the early warning signal, forming complete early warning information containing the risk tracing path.

6. The real-time monitoring and risk early warning system for hazardous waste storage environment based on digital twin as described in claim 5, characterized in that, Starting from the initial anomalous node, the process of tracing the anomaly propagation trajectory along the strong connection paths in the dynamic interconnected network, and recording the key nodes passed through during the anomaly propagation process and their impact intensity, specifically includes: Starting from the initial abnormal node, based on the real-time data of connection weights in the dynamic association network, connections with weights exceeding the dynamic threshold are selected as candidate propagation paths. Deep exploration is conducted along candidate propagation paths. The path weights are recalculated and their persistence is verified at each level of propagation. Only paths that continuously satisfy the weight conditions are retained as valid propagation branches. Construct a dynamic graph of the propagation trajectory, integrate the validated effective propagation branches in time series, and record the key nodes and their cumulative impact strength on each branch; Generate anomaly propagation trajectories with spatiotemporal characteristics, sort key nodes in the dynamic graph according to propagation time sequence, and label the spatial location information and influence intensity parameters of each node to form a complete description of the anomaly propagation trajectory.

7. The real-time monitoring and risk early warning system for hazardous waste storage environment based on digital twins according to claim 6, characterized in that, The constructed dynamic graph of the propagation trajectory integrates the validated effective propagation branches according to time series, recording the key nodes and their cumulative influence intensity on each branch, specifically including: The initial impact intensity data of key nodes are normalized to convert the impact intensity values ​​of different dimensions into a standard range of 0 to 1, forming a standardized initial impact intensity. Dynamic weights are determined based on the correlation strength and propagation delay of adjacent nodes in the propagation path; The degree of attenuation of the influence intensity is calculated layer by layer according to the time sequence of the abnormal propagation; Starting from the initial node, the cumulative influence intensity is accumulated step by step along the propagation path. The cumulative influence intensity of each node is equal to its standardized initial influence intensity plus the sum of the influence intensity of all predecessor nodes after attenuation and weight adjustment.

8. The real-time monitoring and risk early warning system for hazardous waste storage environment based on digital twin as described in claim 1, characterized in that, The dynamic adjustment of the monitoring strategy and operating parameter combination of the intelligent sensor specifically includes: Analyze the spatial location information and risk type characteristics contained in the early warning signals, and determine the scope and boundaries of key monitoring areas based on spatial impact factors; Based on the characteristics of risk type, configure differentiated monitoring mode combinations. For gas diffusion risk, activate high-frequency sampling and continuous monitoring mode, and for temperature anomaly risk, activate multi-sensor collaborative verification mode. Based on the coverage of key monitoring areas, the smart sensors in non-core areas adopt an energy-saving operation strategy, and achieve dynamic balance of monitoring resources by adjusting the sampling interval and transmission frequency. The optimized monitoring strategy is transformed into the operating parameters of a single smart sensor, including sampling frequency, measurement range, trigger threshold, and communication protocol parameters, and the rationality of the configuration is verified through a digital twin system.

9. The real-time monitoring and risk early warning system for hazardous waste storage environment based on digital twin as described in claim 8, characterized in that, The intelligent sensors covering key monitoring areas employ an energy-saving operation strategy for non-core areas. This strategy achieves dynamic balancing of monitoring resources by adjusting sampling intervals and transmission frequencies. Specifically, this includes: The boundaries of non-core areas are determined based on the spatial distribution characteristics of key monitoring areas, and non-core areas are divided into monitoring sub-areas of different priorities based on spatial correlation analysis. Initial energy-saving operation parameters are calculated based on the spatial distance and risk propagation path relationship between each monitoring sub-area and the key monitoring area; The sampling interval of non-core areas is dynamically adjusted based on real-time risk assessment results. When the risk assessment value is lower than the set threshold, the sampling interval is gradually extended and the data transmission frequency is reduced according to a preset ratio. Monitor the quality of the adjusted data collection to ensure that the fluctuation range of key parameters is within the allowable range, thus forming a closed-loop resource optimization allocation.

Images