Intelligent supervision system for dangerous goods family

Abstract

The application discloses a kind of dangerous house wisdom supervision systems, it is related to dangerous waste intelligent supervision technical field, including, acquisition text data in the process of dangerous waste supervision, time series monitoring data and image data, generate standardization multimodal feature matrix, and pass through entity identification and relationship extraction, construct dynamic knowledge graph;Based on dynamic knowledge graph, extract the dynamic node features of each entity node, and identify high-risk entities and colonies, while calculating dynamic risk weight, generating risk quantification list;The dynamic risk weight in risk quantification list is used as risk cost, constructs multi-objective logistics path optimization function that fuses economic cost, time cost and risk cost, and generates global optimal transport plan.The application realizes the upgrade from discrete risk point monitoring to global risk network situation awareness, and can actively warn potential high-risk colonies.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent supervision of hazardous waste, and particularly relates to a hazardous waste intelligent supervision system. BACKGROUND

[0002] With the increasingly stringent environmental protection regulations and the popularity of Internet of Things technology, hazardous waste supervision technology has gradually developed from the traditional mode of relying on manual recording and regular patrol to the information management method based on electronic joint forms and basic databases. In recent years, related technologies have further integrated sensor networks, global positioning systems and rule engines to realize static monitoring of hazardous waste flow nodes and preliminary cost optimization of transportation paths. At the data analysis level, existing solutions mostly use independent machine learning models to analyze text reports, sensor time series data or video images in isolation, or use graph databases to construct fixed association networks between enterprises to identify explicit risks.

[0003] The deficiencies of current hazardous waste supervision technology mainly exist in two key dimensions: first, at the data integration and knowledge construction level, multi-modal information such as text reports, time series monitoring data and image data lack a unified spatio-temporal alignment framework and semantic-level fusion mechanism, resulting in that the entity relationship of the constructed knowledge graph is updated with a lag, and the risk dynamic transmission effect of hazardous waste in the storage, transportation and disposal process cannot be accurately described; second, at the decision optimization level, although the existing path planning methods can consider economic cost and time efficiency, they generally simplify the risk cost as a fixed constraint term, and fail to establish a dynamic correlation mechanism between risk weight and real-time multi-modal monitoring data, so that the optimized scheme is difficult to respond to sudden environmental abnormalities, equipment failures or risk situation changes during transportation. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a hazardous waste intelligent supervision system to solve the problem of supervision lag caused by insufficient multi-modal data fusion and static risk weight.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] This invention provides a hazardous waste management intelligent supervision system, comprising: a data acquisition module for collecting text data, time-series monitoring data, and image data during the hazardous waste supervision process, generating a standardized multimodal feature matrix, performing entity recognition and relationship extraction, and constructing a dynamic knowledge graph; a risk analysis module for extracting dynamic node features of each entity node from the dynamic knowledge graph, identifying high-risk entities and clusters, calculating dynamic risk weights, and generating a risk quantification list; a path optimization module for using the dynamic risk weights in the risk quantification list as risk costs, constructing a multi-objective logistics path optimization function that integrates economic costs, time costs, and risk costs, and generating a globally optimal transportation plan; a dynamic scheduling module for executing the globally optimal transportation plan, collecting multimodal transportation monitoring data, combining the dynamic knowledge graph for anomaly detection and dynamic rescheduling, and generating full-cycle task execution data; and a closed-loop update module for collecting disposal process data and generating hazardous waste disposal feedback data after the hazardous waste arrives at the hazardous waste collection center, updating entity relationships and dynamic risk weights in the dynamic knowledge graph based on the full-cycle task execution data, and synchronously updating the multi-objective logistics path optimization function to form a closed-loop optimization.

[0008] As a preferred embodiment of the hazardous waste management system of the present invention, the text data includes enterprise declaration records, hazardous waste transfer manifests, business qualification documents and administrative penalty documents;

[0009] The time-series monitoring data includes GPS trajectory, vehicle speed, load, temperature, humidity, and pressure;

[0010] The image data includes hazardous waste storage monitoring video frames, photos of the exterior of transport vehicles, operation videos of disposal facilities, and close-ups of hazardous waste packaging labels.

[0011] As a preferred embodiment of the hazardous waste management intelligent supervision system of the present invention, the specific steps for collecting text data, time-series monitoring data, and image data during the hazardous waste supervision process to generate a standardized multimodal feature matrix are as follows.

[0012] Text data, time-series monitoring data, and image data are collected during the hazardous waste supervision process and time-stamped and aligned to form multimodal hazardous waste data.

[0013] Multimodal joint feature extraction and standardization processing are performed on hazardous waste multimodal data to generate text semantic feature vectors, standardized temporal feature vectors, and image visual feature vectors;

[0014] Text feature vectors, standardized temporal feature vectors, and image visual feature vectors are aligned and concatenated according to their feature dimensions to generate a standardized multimodal feature matrix.

[0015] As a preferred embodiment of the hazardous affairs management intelligent supervision system of the present invention, the specific steps for performing entity recognition and relationship extraction to construct a dynamic knowledge graph are as follows:

[0016] The standardized multimodal feature matrix is ​​input into the multimodal fusion algorithm. The dynamic interaction weights between text semantic feature vectors, temporal monitoring quantization feature vectors and image visual feature vectors are obtained through cross-modal attention mechanism. The weighted linear combination and dimensionality reduction are then performed to generate a fusion feature vector in a unified semantic space.

[0017] Based on the fused feature vectors, a graph attention network is used to identify hazardous waste enterprise entities, transport vehicle entities, and disposal facility entities, and the affiliation, transportation, and disposal relationships between entities are extracted to generate a set of entity relationship triples.

[0018] Based on the set of entity relation triples, an initial knowledge graph is constructed in the graph database, and the entity attributes and relation weights are dynamically updated through a temporal convolutional network to form a dynamic knowledge graph.

[0019] As a preferred embodiment of the hazardous waste management intelligent supervision system of the present invention, the hazardous waste enterprise entity refers to the enterprise entity that generates hazardous waste;

[0020] The transport vehicle entity refers to the vehicle entity used for transporting hazardous waste;

[0021] The term "disposal facility entity" refers to the facility entity for the treatment, disposal, and utilization of hazardous waste.

[0022] As a preferred embodiment of the intelligent risk management system described in this invention, the steps of extracting dynamic node features of each entity node from a dynamic knowledge graph, identifying high-risk entities and clusters, calculating dynamic risk weights, and generating a risk quantification list are as follows.

[0023] Based on dynamic knowledge graphs, temporal graph convolutional networks are used to extract dynamic node features of each entity node;

[0024] By inputting dynamic node features into a risk fluid dynamics model, the transmission process of risk in a dynamic knowledge graph relationship network is simulated, and a risk concentration distribution is generated.

[0025] High-risk clusters in the risk concentration distribution are identified using a risk-aware graph clustering algorithm.

[0026] The risk concentration distribution and high-risk clusters are weighted and aggregated, and risk dispersion analysis is performed to calculate dynamic risk weights and generate a risk quantification list.

[0027] As a preferred embodiment of the hazardous materials management intelligent supervision system described in this invention, the following steps are taken: The dynamic risk weights in the risk quantification list are used as risk costs to construct a multi-objective logistics path optimization function that integrates economic costs, time costs, and risk costs, and to generate a globally optimal transportation plan.

[0028] Map the dynamic risk weights in the risk quantification list to the risk cost parameters of each route segment in the transportation network;

[0029] The economic cost parameters and time cost parameters of each route segment are extracted from the basic database of the transportation network and fused with the risk cost parameters to construct a multi-objective logistics route optimization function;

[0030] The multi-objective logistics path optimization function is transformed into a single-objective optimization problem by linear weighted combination.

[0031] A distributed parallel optimization algorithm is used to solve the single-objective optimization problem and obtain the globally optimal transportation path.

[0032] By segmenting the global optimal transportation route into tasks and associating risk information through path serialization and risk mapping, a global optimal transportation plan is generated.

[0033] As a preferred embodiment of the hazardous materials management system described in this invention, the globally optimal transportation path includes a path sequence, cost details, and risk warning.

[0034] As a preferred embodiment of the hazardous materials management intelligent supervision system described in this invention, the steps of executing the globally optimal transportation plan, collecting multimodal transportation monitoring data, combining dynamic knowledge graphs for anomaly detection and dynamic rescheduling, and generating full-cycle task execution data are as follows.

[0035] The digital twin synchronization engine executes the globally optimal transportation plan and collects multimodal transportation monitoring data in real time.

[0036] Multimodal transportation monitoring data is spatiotemporally aligned and fused with dynamic knowledge graphs to generate standardized transportation status sequences;

[0037] Anomaly detection and analysis are performed on standardized transportation status sequences to generate a set of abnormal events;

[0038] Perform rescheduling decisions on the set of abnormal events to generate optimized transportation routes;

[0039] Based on the optimized transportation route, a complete decision record of the entire task execution cycle data is generated through a blockchain-based evidence storage architecture.

[0040] As a preferred embodiment of the hazardous waste management intelligent supervision system described in this invention, the following steps are taken: after hazardous waste arrives at the hazardous waste collection center, data on the disposal process is collected and hazardous waste disposal feedback data is generated. Combined with full-cycle task execution data, entity relationships and dynamic risk weights in the dynamic knowledge graph are updated, and the multi-objective logistics path optimization function is updated synchronously to form a closed-loop optimization. The specific steps are as follows:

[0041] After the hazardous waste arrives at the hazardous waste collection center, data on the disposal process is collected and hazardous waste disposal feedback data is generated. This data is then integrated with the data from the entire task execution cycle to generate a comprehensive updated dataset.

[0042] Extract multi-dimensional task features and anomaly event correlation matrices from the comprehensive updated dataset;

[0043] Based on multi-dimensional task characteristics, dynamic risk weights in dynamic knowledge graphs are updated through an adversarial meta-learning framework.

[0044] Based on the abnormal event association matrix, entity association relationships in a dynamic knowledge graph are evolved through a multi-scale graph neural network.

[0045] Based on the updated dynamic risk weights and evolving entity relationships, the parameters of the multi-objective logistics path optimization function are synchronously updated through an online meta-tuning algorithm, forming a closed-loop optimization.

[0046] The beneficial effects of this invention are as follows: by constructing a technical chain of dynamic knowledge graph generation, risk quantification, path optimization, and closed-loop update, it realizes the leap from static control to dynamic intelligent decision-making in hazardous waste supervision, and solves the cognitive limitations caused by data silos; through temporal graph convolutional networks and risk fluid dynamics models, it realizes the upgrade from monitoring discrete risk points to global risk network situational awareness, and can proactively warn of potential high-risk clusters. Attached Figure Description

[0047] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Fig. 1 This is a schematic diagram of the Danger Management Smart Supervision System.

[0049] Fig. 2 A flowchart for generating a standardized multimodal feature matrix.

[0050] Fig. 3 A flowchart for constructing a dynamic knowledge graph.

[0051] Fig. 4 The flowchart illustrates the anomaly detection and dynamic rescheduling process for executing the globally optimal transportation plan. Detailed Implementation

[0052] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0053] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0054] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0055] Reference Figs. 1-4 As one embodiment of the present invention, this embodiment provides a smart monitoring system for hazardous waste management, comprising the following steps:

[0056] The data acquisition module is used to collect text data, time-series monitoring data, and image data during the hazardous waste supervision process, generate a standardized multimodal feature matrix, perform entity recognition and relationship extraction, and construct a dynamic knowledge graph.

[0057] Text data, time-series monitoring data, and image data are collected during the hazardous waste supervision process and synchronized with timestamps to form multimodal hazardous waste data.

[0058] The specific process includes: during the supervision of hazardous waste, each batch of hazardous waste is uniquely identified by electronic tags with QR codes deployed on hazardous waste packaging, transport vehicles, and disposal facilities; the government data interface is used to automatically obtain enterprise declaration records, hazardous waste transfer manifests, business qualification documents, and administrative penalty documents as text data; vehicle-mounted terminals and IoT sensors upload real-time sensor data such as GPS trajectory, vehicle speed, load, temperature, humidity, and pressure as time-series monitoring data; video surveillance equipment captures monitoring video frames of hazardous waste storage, photos of the exterior of transport vehicles, operation videos of disposal facilities, and close-ups of hazardous waste packaging labels as image data; and the electronic tag QR code is used as the entity identifier. The text data, time-series monitoring data, and image data are synchronized and aligned with the timestamps according to the collection time to form multimodal hazardous waste data.

[0059] Multimodal joint feature extraction and standardization processing are performed on hazardous waste multimodal data to generate text semantic feature vectors, standardized temporal feature vectors, and image visual feature vectors.

[0060] The specific process includes multimodal joint feature extraction and standardization of hazardous waste multimodal data. First, the text data is segmented and embedded, and the words are converted into numerical vectors and then aggregated to generate text semantic feature vectors. For time-series monitoring data, the min-max normalization method is used to linearly scale the data according to the value range of each sensor signal to generate standardized time-series feature vectors. Then, spatial hierarchical features are extracted layer by layer from the image data through a convolutional neural network to generate image visual feature vectors. Finally, text semantic feature vectors, standardized time-series feature vectors, and image visual feature vectors are formed.

[0061] Text feature vectors, standardized temporal feature vectors, and image visual feature vectors are aligned and concatenated according to their feature dimensions to generate a standardized multimodal feature matrix.

[0062] The specific process includes adjusting the text feature vector, standardized temporal feature vector, and image visual feature vector to the same dimension through zero padding or truncation to complete feature dimension alignment. Then, the text feature vector, standardized temporal feature vector, and image visual feature vector after feature dimension alignment are connected into a whole vector in a preset order to generate a standardized multimodal feature matrix.

[0063] It should be noted that the preset order is determined in advance based on the semantic relevance and feature importance of multimodal data in hazardous waste supervision tasks. For example, text semantic feature vectors reflect hazardous waste declaration and compliance information, standardized temporal feature vectors reflect real-time status changes during transportation, and image visual feature vectors provide visual evidence of on-site operations and packaging markings. The three are arranged in the order of text, temporal, and image to prioritize semantic integrity, process traceability, and visual verifiability.

[0064] The standardized multimodal feature matrix is ​​input into the multimodal fusion algorithm. The dynamic interaction weights between the text semantic feature vector, the temporal monitoring quantization feature vector, and the image visual feature vector are obtained through the cross-modal attention mechanism. The weighted linear combination and dimensionality reduction are then performed to generate a fusion feature vector in a unified semantic space.

[0065] The specific process includes inputting a standardized multimodal feature matrix into a multimodal fusion algorithm, using a cross-modal attention mechanism to obtain the correlation scores between each pair of text semantic feature vectors, temporal monitoring quantitative feature vectors, and image visual feature vectors, generating dynamic interaction weights based on the correlation scores, then using the dynamic interaction weights to perform a weighted linear combination of the text semantic feature vectors, temporal monitoring quantitative feature vectors, and image visual feature vectors, and performing dimensionality reduction through linear projection or principal component analysis to generate a fusion feature vector in a unified semantic space.

[0066] It should be noted that the multimodal fusion algorithm is an algorithm that dynamically models the interaction relationship between text semantic feature vectors, temporal monitoring quantization feature vectors, and image visual feature vectors through cross-modal attention mechanisms, and performs weighted fusion and dimensional reduction to generate a unified semantic space representation.

[0067] Cross-modal attention mechanisms are a type of mechanism that analyzes the correlation between different modal features (including text semantic feature vectors, time-series monitoring quantization feature vectors, and image visual feature vectors) and dynamically assigns attention weights to enhance semantic alignment and information interaction between modalities.

[0068] Based on the fused feature vectors, a graph attention network is used to identify hazardous waste enterprise entities, transport vehicle entities, and disposal facility entities, and the affiliation, transportation, and disposal relationships between entities are extracted to generate a set of entity relationship triples.

[0069] The specific process includes learning node representations based on the fused feature vectors and using a graph attention network to learn the semantic information contained in the fused feature vectors. This process identifies hazardous waste enterprise entities, transport vehicle entities, and disposal facility entities. The process also determines the membership, transportation, and disposal relationships between entities based on the attention weights between nodes. Each relationship is represented as a structured expression consisting of a head entity, a relationship type, and a tail entity, generating a set of entity relationship triples.

[0070] It should be noted that semantic information includes the declaration description of hazardous waste enterprises, the operating status of transport vehicles, temporal behavior patterns, and the operational image features of disposal facilities.

[0071] Based on the set of entity relation triples, an initial knowledge graph is constructed in the graph database, and the entity attributes and relation weights are dynamically updated through a temporal convolutional network to form a dynamic knowledge graph.

[0072] The specific process includes: based on the entity relationship triple set, using hazardous waste enterprises, transport vehicles, and disposal facilities as graph nodes in the graph database, and using membership, transportation, and disposal relationships as connecting edges, to establish a structured initial knowledge graph. Historical and real-time time-series monitoring data are then input into a temporal convolutional network. By sliding the convolution kernel along the time dimension, temporal patterns are extracted to obtain the time-varying characteristics of entity attributes. The node attribute values ​​and the relation weights on the edges are then adjusted so that the knowledge graph can reflect the current regulatory status, forming a dynamic knowledge graph.

[0073] The training process of the temporal convolutional network is as follows: Historical and real-time temporal monitoring data are aligned by entity and organized into multivariate time series. Each time step includes monitoring indicators such as the amount of waste generated by hazardous waste enterprises, the location, speed, and load of transport vehicles, and the treatment load of disposal facilities. These time series are input into the temporal convolutional network, and local and global temporal patterns are extracted layer by layer in the time dimension through multi-layer causal convolution and residual connections. The time-varying feature representation of each entity at the current time step is output. During training, known node attribute changes or relation strength labels (such as attribute adjustment records based on abnormal events or expert annotations) are used as supervision signals. The error between the output time-varying features and the true attribute values ​​is calculated, and the convolution kernel weights are optimized using the backpropagation algorithm. In the iterative process, the temporal convolutional network gradually learns the normal evolution law and abnormal offset pattern of various entity attributes. Ultimately, the output time-varying features can accurately drive the dynamic update of node attribute values ​​and edge relation weights in the graph database, supporting the continuous evolution of dynamic knowledge graphs.

[0074] It should be noted that a graph database is a database that stores data in a graph structure, using nodes to represent entities and edges to represent relationships between entities, and supports efficient relationship traversal and graph query operations.

[0075] The time-varying characteristics of physical attributes include monitoring indicators that change over time, such as the amount of waste generated by hazardous waste enterprises, the location, speed, load, temperature, humidity of transport vehicles, and the treatment load and operating status of disposal facilities.

[0076] The risk analysis module is used to extract the dynamic node features of each entity node from the dynamic knowledge graph, identify high-risk entities and clusters, calculate dynamic risk weights, and generate a risk quantification list.

[0077] Based on dynamic knowledge graphs, temporal graph convolutional networks are used to extract dynamic node features of each entity node.

[0078] The specific process involves using a temporal graph convolutional network based on a dynamic knowledge graph to jointly model the attributes and connections of each entity node and its neighboring nodes at multiple time steps. By superimposing convolutional operations along the time dimension on the dynamic knowledge graph, neighborhood information is aggregated layer by layer, and temporal evolution patterns are captured (e.g., the changing trends of the position and speed of transport vehicles in consecutive time steps reflect the continuity and abnormal deviation of their driving state; the periodic fluctuations in the amount of waste generated by hazardous waste enterprises reflect the regularity of their production activities; and the accumulation and release process of the treatment load of disposal facilities reveals the dynamic changes in their treatment capacity), thereby extracting the dynamic node features of each entity node.

[0079] It should be noted that neighborhood information refers to the local subgraph structure formed by other entity nodes directly connected to the current entity node and their connection relationships in the dynamic knowledge graph. This includes the dynamic node characteristics, node attributes, and temporal evolution status of adjacent hazardous waste enterprises, transport vehicles, or disposal facilities, as well as the type and weight of their affiliation, transportation, or disposal relationships with the current entity node.

[0080] The training process of the temporal graph convolutional network (TMB) involves taking a sequence of snapshots of a dynamic knowledge graph across multiple consecutive time steps as input. Each time step contains the attribute features of entity nodes and their connectivity relationships. In each layer, the TMB performs both spatial and temporal convolutions. Spatial convolutions aggregate the features of each entity node and its neighboring nodes at the current time step, while temporal convolutions fuse the representations of the same node across adjacent time steps along the temporal dimension. This jointly models structural dependencies and temporal evolution, and by stacking multiple layers, it captures dynamic information from multi-hop neighborhoods and long time spans, generating dynamic node features for each entity node at the final time step. During training, a set of abnormal events or labeled risk states are used as supervision signals. The loss between the output dynamic node features and the true labels is compared, and the network weights are optimized end-to-end using backpropagation. During iteration, the network gradually learns the interaction patterns between typical temporal patterns and graph structures, such as changes in the position and speed of transport vehicles, fluctuations in waste generation by hazardous waste enterprises, and dynamic loads of disposal facilities, forming a temporal graph convolutional network that accurately represents the dynamic behavior of entities.

[0081] By inputting dynamic node features into the risk fluid dynamics model, the transmission process of risk in the dynamic knowledge graph relationship network is simulated, and the risk concentration distribution is generated.

[0082] The specific process includes inputting dynamic node characteristics into a risk fluid dynamics model. The risk fluid dynamics model regards each entity node in the dynamic knowledge graph as a state carrier in the risk propagation process, and regards the affiliation, transportation, and disposal relationships between entity nodes as the paths of risk transmission. Based on the temporal attributes and association strength reflected by the dynamic node characteristics, the transmission intensity and cumulative effect of risk on each relationship path are obtained. Referring to the physical process of fluid flow, diffusion, convergence, and retention in a pipeline network, the propagation, convergence, and retention behavior of risk in the dynamic knowledge graph relationship network are simulated, and the transmission process of risk from high-risk entity nodes to associated entity nodes is dynamically deduced, ultimately generating a risk concentration distribution.

[0083] Furthermore, the construction process of the risk fluid dynamics model is based on the entity nodes and relation edges in the dynamic knowledge graph. Hazardous waste enterprises, transport vehicles, and disposal facilities are taken as risk carrier nodes, and affiliation, transportation, and disposal relationships are taken as risk transmission channels. The risk state of each node is defined as being determined by its dynamic node characteristics. The flow of risk in the relation channels is set to follow the evolution rules analogous to the diffusion, transmission, and retention of fluid in a connected pipe. By setting the initial risk source, transmission rate, and attenuation coefficient based on the dynamic node characteristics, a dynamic propagation mechanism of risk evolution over time in the dynamic knowledge graph is established, thus forming the risk fluid dynamics model.

[0084] High-risk clusters in the risk concentration distribution are identified using a risk-aware graph clustering algorithm.

[0085] The specific process includes analyzing the risk concentration distribution of each entity node in the dynamic knowledge graph through a risk-aware graph clustering algorithm, and classifying entity nodes with high risk concentration and close correlation into the same group based on the risk concentration similarity between nodes and the graph topology connection relationship, thereby identifying high-risk clusters in the risk concentration distribution.

[0086] It should be noted that graph clustering algorithm is an algorithm that divides nodes into several subgroups with internally tightly connected and externally sparsely connected based on the connection relationships and attribute similarities between nodes in a graph structure.

[0087] The risk concentration distribution is weighted and aggregated with high-risk clusters, and risk dispersion analysis is performed to calculate dynamic risk weights and generate a risk quantification list. The expression is as follows:

[0088] ;

[0089] in, Indicates the first Dynamic risk weights for high-risk clusters. Indicates the index number of the high-risk cluster. This represents the risk dispersion adjustment coefficient.

[0090] in, Indicates the first The weighted average risk concentration value of a high-risk cluster is expressed as:

[0091] ;

[0092] It should be noted that, The index number representing a single entity within a high-risk cluster. Indicates the first A collection of entities in a high-risk cluster. The entity is The weighting coefficient within the high-risk cluster. Representing entities The risk concentration value.

[0093] in, Indicates the first The standard deviation of the risk concentration within a high-risk cluster is expressed as:

[0094] ;

[0095] It should be noted that the risk dispersion adjustment coefficient is a correction factor obtained by linear mapping of the normalized risk dispersion. It is used to reflect the concentration of risk distribution within high-risk clusters, and its value ranges from greater than 0 to less than or equal to 1.

[0096] The risk concentration distribution is weighted and aggregated with high-risk clusters, and risk dispersion analysis is performed. The specific process includes: calculating the dynamic risk weight of each high-risk cluster, where the dynamic risk weight is determined by the weighted average risk concentration value and the risk dispersion adjustment coefficient. The weighted average risk concentration value is obtained by multiplying the risk concentration value of each entity in the high-risk cluster by its weight coefficient in the high-risk cluster, summing the results, and normalizing the result. The risk dispersion is obtained by calculating the root mean square of the sum of squared deviations of the risk concentration values ​​of each entity in the high-risk cluster from the weighted average value. Finally, all high-risk clusters and their corresponding dynamic risk weights are compiled into a risk quantification list.

[0097] The route optimization module is used to construct a multi-objective logistics route optimization function that integrates economic costs, time costs, and risk costs by using the dynamic risk weights in the risk quantification list as risk costs, and to generate a globally optimal transportation plan.

[0098] The dynamic risk weights in the risk quantification list are mapped to the risk costs of each route segment in the transportation network.

[0099] The specific process involves associating the dynamic risk weights in the risk quantification list with the corresponding relationships between hazardous waste enterprises, transport vehicles, and disposal facilities and the path segments in the transportation network (for example, if a transport vehicle departs from hazardous waste enterprise A and travels to disposal facility B via road segment L, then road segment L is associated with the transport vehicle, hazardous waste enterprise A, and disposal facility B). The path segments involved in each high-risk cluster are assigned corresponding dynamic risk weights as the risk cost of that path segment, thereby completing the mapping of dynamic risk weights to the risk costs of each path segment in the transportation network.

[0100] The economic and time costs of each route segment are extracted from the basic database of the transportation network and integrated with the risk cost to construct a multi-objective logistics route optimization function.

[0101] The specific process includes extracting the economic and time costs of each route segment from the basic database of the transportation network. The economic costs include cost indicators such as fuel consumption, toll fees, labor costs, and vehicle depreciation corresponding to the route segment, while the time costs include duration indicators such as travel time, waiting time, and scheduling delays of the route segment. The economic and time costs are then linearly weighted and integrated with the risk cost parameters mapped to each route segment according to preset weights, so that the comprehensive cost of each route segment simultaneously reflects the three factors of economy, time, and risk, thereby constructing a multi-objective logistics route optimization function.

[0102] It should be noted that the transportation network basic database is a structured data set that stores the economic and time costs of each route segment in the transportation network. It is obtained by integrating data from the transportation geographic information system, historical transportation records, logistics billing standards, and real-time traffic information.

[0103] The preset weights are set in advance based on the priority requirements of economic costs, time costs, and risk costs in hazardous waste supervision tasks, and in accordance with policy regulations and safety standards.

[0104] The multi-objective logistics path optimization function is transformed into a single-objective optimization problem by using linear weighted combination.

[0105] The specific process includes multiplying the three cost items—economic cost, time cost, and risk cost—into their respective preset weights in the multi-objective logistics path optimization function. These preset weights reflect the relative importance of the three types of costs in the hazardous waste supervision task. The weighted costs are then summed over each path segment to obtain the comprehensive cost of that path segment. Furthermore, the comprehensive costs of all path segments along the entire transportation path are accumulated to form a single comprehensive objective function. This comprehensive objective function integrates economic, time, and risk factors, transforming the optimization problem that originally required balancing multiple objectives into a single-objective optimization problem that only needs to minimize the comprehensive objective function.

[0106] A distributed parallel optimization algorithm is used to solve the single-objective optimization problem and obtain the globally optimal transportation path.

[0107] The specific process includes using a distributed parallel optimization algorithm to solve the single-objective optimization problem, dividing the transportation network into multiple sub-regions, assigning each sub-region to different computing nodes to search for local optimal paths in parallel, and coordinating the optimization direction of each sub-region by periodically exchanging boundary path information and the current optimal solution between nodes, gradually converging to the globally optimal transportation path covering the entire network.

[0108] It should be noted that the distributed parallel optimization algorithm divides the entire optimization problem into several sub-problems according to the topology or geographical distribution of the transportation network. Each sub-problem is assigned to a computing node for independent solution. Each computing node maintains partial solutions locally and periodically exchanges boundary states and solution information with other nodes. The algorithm gradually approaches the global optimal solution through cooperative iteration.

[0109] By segmenting the global optimal transportation route into tasks and associating risk information through path serialization and risk mapping, a global optimal transportation plan is generated.

[0110] The specific process includes dividing the globally optimal transportation route into nodes according to the order of hazardous waste enterprises, transport vehicles, and disposal facilities along the route, forming a series of continuous route segments. Each route segment corresponds to a specific transportation task. Through risk mapping, the dynamic risk weights related to each entity in the risk quantification list are matched to the involved route segments, so that each route segment not only includes the starting point, ending point, estimated time, and economic cost, but also the corresponding risk cost. Finally, all route segments and associated multi-dimensional cost information are integrated into a structured task sequence to generate a globally optimal transportation plan.

[0111] The dynamic scheduling module is used to execute the globally optimal transportation plan, collect multimodal transportation monitoring data, and combine dynamic knowledge graphs for anomaly detection and dynamic rescheduling to generate full-cycle data for task execution.

[0112] The digital twin synchronization engine executes the globally optimal transportation plan and collects multimodal transportation monitoring data in real time.

[0113] The specific process includes executing the globally optimal transportation plan through a digital twin synchronization engine, synchronizing the task instructions and risk information of each path segment in the globally optimal transportation plan to the digital twin of the corresponding physical entity, driving the transport vehicles to run along the planned path, and collecting multimodal transportation monitoring data composed of text data, time-series monitoring data and image data in real time during the execution of the globally optimal transportation plan, so as to achieve dynamic consistency between the physical transportation process and the digital space.

[0114] It should be noted that multimodal transportation monitoring data includes text data such as enterprise declaration records and hazardous waste transfer manifests during transportation, time-series monitoring data such as GPS trajectory, vehicle speed, load, temperature, humidity, and pressure, as well as image data such as exterior photos of transport vehicles, monitoring video frames of hazardous waste storage, operation videos of disposal facilities, and close-ups of hazardous waste packaging labels.

[0115] A digital twin synchronization engine is a technical mechanism that enables state synchronization and command execution between a physical transportation entity and its digital twin. It is an existing technical component based on dynamic knowledge graphs, standardized transportation state sequences, and multimodal transportation monitoring data.

[0116] Multimodal transportation monitoring data is spatiotemporally aligned and fused with dynamic knowledge graphs to generate standardized transportation status sequences.

[0117] The specific process includes: spatiotemporally aligning and fusing multimodal transportation monitoring data with a dynamic knowledge graph; assigning a unified timestamp to text data, time-series monitoring data, and image data based on the collection time; and spatially matching the GPS trajectory of the transport vehicle with the geographical location information of entity nodes in the dynamic knowledge graph. This ensures that each record in the multimodal transportation monitoring data is precisely associated with the corresponding hazardous waste enterprise, transport vehicle, and disposal facility entity in the dynamic knowledge graph in time and space. Finally, the aligned multi-source information is integrated in chronological order to generate a standardized transportation status sequence.

[0118] Anomaly detection and analysis are performed on standardized transportation status sequences to generate a set of abnormal events.

[0119] The specific process includes performing anomaly detection and analysis on the standardized transportation state sequence, constructing the spatial dependency relationship between entity nodes in the dynamic knowledge graph and the temporal evolution law in the standardized transportation state sequence, extracting the state characteristics of each entity node step by step and identifying abnormal behaviors that deviate from the normal pattern, extracting event records that meet the preset anomaly judgment conditions, and generating an abnormal event set.

[0120] It should be noted that the pre-set anomaly judgment conditions are based on the regulatory standards for hazardous waste transportation, the statistical characteristics of historical anomaly cases, and the normal behavior patterns of multimodal transportation monitoring data, and are pre-set by risk control requirements.

[0121] The system makes rescheduling decisions on the set of abnormal events and generates optimized transportation routes.

[0122] The specific process includes making rescheduling decisions on the set of abnormal events through a metacognitive dynamic optimizer. The metacognitive dynamic optimizer re-evaluates feasible paths and adjusts transportation plans based on the type, location, risk level, and impact of each abnormal event in the set on the transportation task, combined with the current transportation network status and multi-objective logistics path optimization function, and generates optimized transportation paths.

[0123] It should be noted that the metacognitive dynamic optimizer refers to a mechanism that performs self-evaluation and reflection on existing transportation scheduling decisions based on a set of abnormal events. It is an existing intelligent optimization method consisting of an anomaly perception module, a decision evaluation module, and a route replanning module. It can dynamically adjust the optimization objectives and constraints according to the type, location, and degree of impact of abnormal events to generate optimized transportation routes.

[0124] Based on the optimized transportation route, a complete decision record of the entire task execution cycle data is generated through a blockchain-based evidence storage architecture.

[0125] The specific process includes, based on the optimized transportation route, storing key data of the transportation task at each stage, such as initial planning, risk assessment, route optimization, anomaly handling and final execution, on the blockchain through a blockchain evidence storage architecture. This includes the global optimal transportation plan, abnormal event set, rescheduling decision results and multimodal transportation monitoring data, which are organized in chronological order and written into the immutable ledger of the blockchain to generate complete decision records of the entire lifecycle of task execution data.

[0126] It should be noted that the blockchain evidence storage architecture is a technical architecture that utilizes the immutability, timestamp, and distributed consensus characteristics of blockchain to reliably record and store key data.

[0127] The closed-loop update module is used to collect disposal process data and generate hazardous waste disposal feedback data after hazardous waste arrives at the hazardous waste collection center. Combined with the full-cycle data of task execution, it updates the entity relationship and dynamic risk weight in the dynamic knowledge graph and updates the multi-objective logistics path optimization function simultaneously to form a closed-loop optimization.

[0128] After the hazardous waste arrives at the hazardous waste collection center, data on the disposal process is collected and hazardous waste disposal feedback data is generated. This data is then integrated with the full-cycle data of the task execution to generate a comprehensive updated dataset.

[0129] The specific process includes collecting temperature, pressure, pollutant concentration, and operation record data during the disposal process through a deployed sensor network after the hazardous waste arrives at the hazardous waste collection center, generating hazardous waste disposal feedback data. The hazardous waste disposal feedback data, along with the full-cycle data of the task execution that records information from the start to the end of the task, are input into the data fusion engine and integrated using a correlation alignment method based on timestamps and entity identifiers to generate a comprehensive updated dataset that includes the disposal results and context information of the entire process.

[0130] Extract multi-dimensional task features and anomalous event correlation matrices from the comprehensive updated dataset.

[0131] The specific process includes using a data parsing engine to analyze the globally optimal transportation plan, optimized transportation routes, multimodal transportation monitoring data, and abnormal event set contained in the comprehensive update dataset. From this, multi-dimensional task characteristics such as economic cost, time consumption, risk level, route structure, and entity type of the transportation task are extracted. Based on the location, timestamp, and associated entities of the abnormal events on the transportation route, the correspondence between the abnormal events and each task characteristic is constructed to form an abnormal event correlation matrix.

[0132] It should be noted that the data parsing engine refers to the technical component used to perform structured parsing tasks on multi-source heterogeneous information in the entire lifecycle of the data. It is based on a natural language processing module, a time-series data parsing module, an image semantic understanding module, and a graph structure alignment module.

[0133] Based on multi-dimensional task characteristics, dynamic risk weights in a dynamic knowledge graph are updated using an adversarial meta-learning framework.

[0134] The specific process includes updating the dynamic risk weights in the dynamic knowledge graph based on multi-dimensional task features using an adversarial meta-learning framework. The adversarial meta-learning framework uses multi-dimensional task features to construct support sets and query sets, simulates the changes in risk distribution under different transportation scenarios during the meta-learning process, and optimizes the generalization ability of risk weights through adversarial training mechanisms, so that dynamic risk weights can adapt to newly emerging abnormal patterns and task features, thereby achieving continuous calibration and updating of dynamic risk weights in the dynamic knowledge graph.

[0135] Based on the abnormal event association matrix, entity association relationships in a dynamic knowledge graph are evolved through a multi-scale graph neural network.

[0136] The specific process includes: based on the abnormal event association matrix, evolving the entity association relationships in the dynamic knowledge graph through a multi-scale graph neural network; the multi-scale graph neural network aggregates the abnormal patterns and task features revealed by the abnormal event association matrix within multiple neighborhood hop count ranges, respectively capturing local tight interactions and global sparse dependencies, and updating the semantic strength and structural connections of the affiliation, transportation, and disposal relationships between hazardous waste enterprises, transport vehicles, and disposal facilities, thereby realizing the dynamic evolution of entity association relationships in the dynamic knowledge graph.

[0137] The training process of a multi-scale graph neural network is as follows: Entity nodes in the dynamic knowledge graph are initialized as embedding vectors containing static attributes and dynamic node features, with the abnormal event association matrix used as a supervision signal. Subsequently, in each layer of the multi-scale graph neural network, message passing is performed within the 1-hop, 2-hop, and even multi-hop neighborhood ranges, aggregating neighbor node features at different scales to form node representations at multiple scales. These multi-scale representations are then fused to generate a comprehensive embedding for each entity node. By comparing the comprehensive embedding with the abnormal association patterns revealed in the abnormal event association matrix, a loss function is obtained, and the parameters of the multi-scale graph neural network are updated using the backpropagation algorithm. In multiple rounds of iterative training, the multi-scale graph neural network gradually learns how to adjust the relational weights between nodes according to abnormal events, ultimately enabling the multi-scale graph neural network to accurately reflect the semantic strength and structural changes of the affiliation, transportation, and disposal relationships between hazardous waste enterprises, transport vehicles, and disposal facilities, realizing the dynamic evolution of entity association relationships in the dynamic knowledge graph.

[0138] Based on the updated dynamic risk weights and evolving entity relationships, the parameters of the multi-objective logistics path optimization function are synchronously updated through an online meta-tuning algorithm, forming a closed-loop optimization.

[0139] The specific process includes updating the parameters of the multi-objective logistics path optimization function synchronously through an online meta-tuning algorithm based on the updated dynamic risk weights and evolving entity relationships. The online meta-tuning algorithm uses the latest dynamic risk weights to adjust the contribution ratio of risk costs and corrects the connection effectiveness and cost structure of path segments in the transportation network based on the evolving entity relationships. This ensures that the weights and calculation methods of the three components of the multi-objective logistics path optimization function—economic cost, time cost, and risk cost—continuously adapt to the current regulatory status, forming a closed-loop optimization.

[0140] It should be noted that online meta-tuning algorithms refer to optimization methods that continuously receive new feedback data (such as updated dynamic risk weights and evolving entity relationships) during task execution and adjust the parameters of multi-objective logistics path optimization functions in real time through meta-learning mechanisms. The core of these methods is to learn parameter adjustment strategies based on the experience of multiple historical tasks, so that when faced with the new state of the current task, it can quickly adapt to and optimize the weight configuration of the three components of economic cost, time cost, and risk cost in the objective function, thereby maintaining the consistency between the path optimization results and the latest regulatory situation.

[0141] In summary, this invention achieves a leap from static control to dynamic intelligent decision-making in hazardous waste supervision by constructing a technical chain of dynamic knowledge graph generation, risk quantification, path optimization, and closed-loop update, thus solving the cognitive limitations caused by data silos. Through temporal graph convolutional networks and risk fluid dynamics models, it upgrades from monitoring discrete risk points to global risk network situational awareness, enabling proactive early warning of potential high-risk clusters.

[0142] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A smart monitoring system for hazardous materials management, characterized in that: include, The data acquisition module is used to collect text data, time-series monitoring data and image data in the process of hazardous waste supervision, generate a standardized multimodal feature matrix, and perform entity recognition and relationship extraction to construct a dynamic knowledge graph. The risk analysis module is used to extract dynamic node features of each entity node from the dynamic knowledge graph, identify high-risk entities and clusters, calculate dynamic risk weights, and generate a risk quantification list. The specific steps are as follows. Based on dynamic knowledge graphs, temporal graph convolutional networks are used to extract dynamic node features of each entity node; By inputting dynamic node features into a risk fluid dynamics model, the transmission process of risk in a dynamic knowledge graph relationship network is simulated, and a risk concentration distribution is generated. High-risk clusters in the risk concentration distribution are identified using a risk-aware graph clustering algorithm. The risk concentration distribution and high-risk clusters are weighted and aggregated, and risk dispersion analysis is performed to calculate dynamic risk weights and generate a risk quantification list. The route optimization module is used to construct a multi-objective logistics route optimization function that integrates economic cost, time cost and risk cost by using the dynamic risk weight in the risk quantification list as risk cost, and generate a globally optimal transportation plan. The dynamic scheduling module is used to execute the globally optimal transportation plan, collect multimodal transportation monitoring data, and combine dynamic knowledge graphs for anomaly detection and dynamic rescheduling to generate full-cycle data of task execution. The closed-loop update module is used to collect disposal process data and generate hazardous waste disposal feedback data after hazardous waste arrives at the hazardous waste collection center. Combined with the full-cycle data of task execution, it updates the entity relationship and dynamic risk weight in the dynamic knowledge graph and updates the multi-objective logistics path optimization function simultaneously to form a closed-loop optimization.

2. The hazardous waste management intelligent supervision system as described in claim 1, characterized in that: The text data includes enterprise declaration records, hazardous waste transfer manifests, business qualification documents, and administrative penalty documents; The time-series monitoring data includes GPS trajectory, vehicle speed, load, temperature, humidity, and pressure; The image data includes hazardous waste storage monitoring video frames, photos of the exterior of transport vehicles, operation videos of disposal facilities, and close-ups of hazardous waste packaging labels.

3. The hazardous materials management intelligent supervision system as described in claim 2, characterized in that: The process of collecting text data, time-series monitoring data, and image data during hazardous waste management, and generating a standardized multimodal feature matrix, involves the following specific steps. Text data, time-series monitoring data, and image data are collected during the hazardous waste supervision process and time-stamped and aligned to form multimodal hazardous waste data. Multimodal joint feature extraction and standardization processing are performed on hazardous waste multimodal data to generate text semantic feature vectors, standardized temporal feature vectors, and image visual feature vectors; Text feature vectors, standardized temporal feature vectors, and image visual feature vectors are aligned and concatenated according to their feature dimensions to generate a standardized multimodal feature matrix.

4. The hazardous materials management intelligent supervision system as described in claim 3, characterized in that: The specific steps for entity recognition and relation extraction to construct a dynamic knowledge graph are as follows. The standardized multimodal feature matrix is ​​input into the multimodal fusion algorithm. The dynamic interaction weights between text semantic feature vectors, temporal monitoring quantization feature vectors and image visual feature vectors are obtained through cross-modal attention mechanism. The weighted linear combination and dimensionality reduction are then performed to generate a fusion feature vector in a unified semantic space. Based on the fused feature vectors, a graph attention network is used to identify hazardous waste enterprise entities, transport vehicle entities, and disposal facility entities, and the affiliation, transportation, and disposal relationships between entities are extracted to generate a set of entity relationship triples. Based on the set of entity relation triples, an initial knowledge graph is constructed in the graph database, and the entity attributes and relation weights are dynamically updated through a temporal convolutional network to form a dynamic knowledge graph.

5. The hazardous materials management intelligent supervision system as described in claim 4, characterized in that: The hazardous waste enterprise entity refers to the enterprise entity that generates hazardous waste; The transport vehicle entity refers to the vehicle entity used for transporting hazardous waste; The term "disposal facility entity" refers to the facility entity for the treatment, disposal, and utilization of hazardous waste.

6. The hazardous materials management intelligent supervision system as described in claim 5, characterized in that: The process involves using the dynamic risk weights in the risk quantification list as risk costs, constructing a multi-objective logistics path optimization function that integrates economic costs, time costs, and risk costs, and generating a globally optimal transportation plan. The specific steps are as follows: Map the dynamic risk weights in the risk quantification list to the risk cost parameters of each route segment in the transportation network; The economic cost parameters and time cost parameters of each route segment are extracted from the basic database of the transportation network and fused with the risk cost parameters to construct a multi-objective logistics route optimization function; The multi-objective logistics path optimization function is transformed into a single-objective optimization problem by linear weighted combination. A distributed parallel optimization algorithm is used to solve the single-objective optimization problem and obtain the globally optimal transportation path. By segmenting the global optimal transportation route into tasks and associating risk information through path serialization and risk mapping, a global optimal transportation plan is generated.

7. The hazardous materials management intelligent supervision system as described in claim 6, characterized in that: The globally optimal transportation route includes a route sequence, cost details, and risk warnings.

8. The hazardous materials management intelligent supervision system as described in claim 7, characterized in that: The process involves executing the globally optimal transportation plan, collecting multimodal transportation monitoring data, combining dynamic knowledge graphs for anomaly detection and dynamic rescheduling, and generating full-cycle task execution data. The specific steps are as follows: The digital twin synchronization engine executes the globally optimal transportation plan and collects multimodal transportation monitoring data in real time. Multimodal transportation monitoring data is spatiotemporally aligned and fused with dynamic knowledge graphs to generate standardized transportation status sequences; Anomaly detection and analysis are performed on standardized transportation status sequences to generate a set of abnormal events; Perform rescheduling decisions on the set of abnormal events to generate optimized transportation routes; Based on the optimized transportation route, a complete decision record of the entire task execution cycle data is generated through a blockchain-based evidence storage architecture.

9. The hazardous materials management intelligent supervision system as described in claim 8, characterized in that: After the hazardous waste arrives at the hazardous waste collection center, data on the disposal process is collected and hazardous waste disposal feedback data is generated. Combined with data from the entire task execution cycle, the entity relationships and dynamic risk weights in the dynamic knowledge graph are updated, and the multi-objective logistics path optimization function is updated simultaneously to form a closed-loop optimization. The specific steps are as follows: After the hazardous waste arrives at the hazardous waste collection center, data on the disposal process is collected and hazardous waste disposal feedback data is generated. This data is then integrated with the data from the entire task execution cycle to generate a comprehensive updated dataset. Extract multi-dimensional task features and anomaly event correlation matrices from the comprehensive updated dataset; Based on multi-dimensional task characteristics, dynamic risk weights in dynamic knowledge graphs are updated through an adversarial meta-learning framework. Based on the abnormal event association matrix, entity association relationships in a dynamic knowledge graph are evolved through a multi-scale graph neural network. Based on the updated dynamic risk weights and evolving entity relationships, the parameters of the multi-objective logistics path optimization function are synchronously updated through an online meta-tuning algorithm, forming a closed-loop optimization.

Images