An intelligent identification method for identifying risks of abnormal stay and illegal dumping of solid waste transportation

Abstract

The application discloses an intelligent identification method for identifying abnormal stay and illegal dumping risks of solid waste transportation, and belongs to the technical field of artificial intelligence. The method comprises the following steps: constructing an initial heterogeneous graph taking vehicles as nodes and business logic association as edges; real-time monitoring of vehicle cargo state change events as a graph structure evolution trigger signal; in response to the monitored events, dynamically adjusting the topological structure of the current heterogeneous graph by selectively disconnecting, activating or attenuating business edges, so that the graph structure evolves in real time with the business state, and an evolved heterogeneous graph is generated; inputting the evolved heterogeneous graph into a graph neural network model to extract structural features and identify abnormal stay and illegal dumping risks. Through the event-driven graph evolution mechanism, the graph structure and the transportation business state are synchronized in real time, which helps to alleviate the information lag problem and improve the accuracy and timeliness of risk identification.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an intelligent identification method for the risks of abnormal stops and illegal dumping during the transportation of solid and hazardous waste. Background Technology

[0002] In the field of solid and hazardous waste transportation supervision, existing technologies mainly identify abnormal behavior through GPS trajectory monitoring and preset rule judgment. More advanced methods use graph neural networks to construct vehicle interaction graphs and detect anomalies by analyzing the spatiotemporal correlations between vehicles.

[0003] However, existing graph construction methods have the following technical drawbacks:

[0004] The vehicle interaction graphs constructed by existing technologies are essentially static snapshots or time-slice sequences—whether they are static graphs sampled based on fixed time windows or graph sequences arranged in chronological order, the updates to their graph structures depend on preset time intervals. This time-driven update mechanism has fundamental flaws: On the one hand, the graph's topology lags far behind the real-time changes in the transportation business status. For example, after a vehicle finishes unloading, its business association with the destination should be immediately disconnected, but the time-driven mechanism must wait for the next time window to update. During this period, the risk identification model still makes judgments based on outdated associations, which is prone to misjudgment or omission. On the other hand, the frequent changes in cargo status (loading, unloading, waybill changes) in solid and hazardous waste transportation are only recorded as node attributes, rather than as event signals that trigger dynamic adjustments to the graph topology, causing the graph structure to be unable to perceive the real-time evolution of business intentions. In addition, the existing mechanism lacks the ability to selectively evolve business edges, and cannot distinguish between edges that need to be disconnected immediately (such as the association with the original destination after being empty), edges that need to be activated (such as the potential association between an empty vehicle and nearby cargo sources), and edges that need to decay over time (such as long-term unfulfilled business intentions), thus failing to accurately depict the changing process of dynamic business intentions. Summary of the Invention

[0005] The technical problem this invention aims to solve is to provide an intelligent identification method for the risks of abnormal stops and illegal dumping in the transportation of solid and hazardous waste, addressing the issue that the updating of the vehicle interaction diagram structure lags behind changes in business status and cannot reflect the dynamic evolution of business intentions in real time.

[0006] To address the aforementioned technical problems, this invention provides an intelligent identification method for the risks of abnormal stops and illegal dumping during the transportation of solid and hazardous waste. The method includes: constructing an initial heterogeneous graph P0 with vehicles as nodes and business logic relationships between vehicles as edges, wherein the business logic relationships are determined based on the vehicle's electronic waybill data; for any given moment, a current transportation heterogeneous graph Pn, where n is an integer greater than or equal to 0, and the current transportation heterogeneous graph Pn is the initial heterogeneous graph P0 or an intermediate heterogeneous graph obtained by at least one evolution of the initial heterogeneous graph P0; and real-time monitoring of cargo status change events for each vehicle, wherein the cargo status change events include at least the elements present in the current transportation heterogeneous graph P0. The n-structure evolution triggers loading / unloading events and waybill status update events; in response to the monitored cargo status change events, the topology and edge weights of the business edges in the current heterogeneous transportation graph Pn are dynamically adjusted so that the topology of the current heterogeneous transportation graph Pn evolves in real time with the business status, generating an evolved heterogeneous graph Pn+1; wherein, the dynamic adjustment includes selectively disconnecting, activating or attenuating business edges according to the type and content of cargo status changes; the evolved heterogeneous graph Pn+1 is input into a graph neural network model, and heterogeneous graph structural features are extracted from the heterogeneous graph Pn+1 to identify abnormal stays and illegal dumping risks in the transportation of solid and hazardous waste.

[0007] Optionally, determining the business logic association based on the vehicle's electronic waybill data includes: acquiring the vehicle's electronic waybill data, which at least includes unstructured cargo description text; semantically encoding the cargo description text using a pre-trained language model to generate a waybill semantic embedding vector; calculating the cosine similarity between the waybill semantic embedding vectors of any two vehicles, and establishing a business edge between the two vehicles when the cosine similarity exceeds a preset threshold.

[0008] Optionally, the semantic encoding using a pre-trained language model includes: concatenating the cargo description text, cargo owner information, destination information, and transportation timeliness information into a standardized waybill text; inputting the waybill text into the pre-trained language model, and taking the hidden state of the [CLS] token output by the language model as the waybill semantic embedding vector.

[0009] Optionally, the initial weight of the business edge is dynamically calculated based on the business association mode, which includes at least one or more of the following: Same cargo owner / same destination mode: When the cargo owner information or destination is the same in the electronic waybill data of two vehicles, the business edge weight increases by a first preset value; Upstream and downstream connection mode: When the destination of the waybill of the first vehicle is the same as the origin of the waybill of the second vehicle, and the cargo status of the first vehicle is fully loaded and awaiting delivery, and the cargo status of the second vehicle is empty and returning, the business edge weight increases by a second preset value, and the second preset value is greater than the first preset value; Value collaboration mode: When it is decoded from the waybill semantic embedding vector that the cargo of both vehicles is high-value cargo, and the real-time location of both vehicles is located in a preset high-risk area, the business edge weight increases by a third preset value.

[0010] Optionally, selectively disconnecting, activating, or attenuating service edges includes: when the cargo status of the first vehicle changes from fully loaded to empty, disconnecting the service edge related to the destination for the first vehicle and activating the potential service association between the first vehicle and nearby empty vehicles; when the cargo status of the second vehicle changes from empty to fully loaded and its current location is close to the loading point, activating the service edge between the second vehicle and empty vehicles in the destination area; setting an attenuation coefficient for each service edge that decays over time, and when two vehicles do not physically approach each other within a preset time window after the service edge is established, the weight of the service edge gradually decreases according to the attenuation coefficient until it disappears.

[0011] Optionally, the dynamic adjustment further includes an attention adjustment mechanism that incorporates business status awareness: during the neighbor aggregation process of the graph neural network, an attention weight is calculated for each vehicle node to its neighbor nodes; the calculation formula for the attention weight incorporates a business status adjustment factor, which is dynamically generated based on the current cargo status, travel progress, and timeliness of the vehicle node; when the travel progress of the first vehicle exceeds a preset threshold and the cargo status is fully loaded and ready for delivery, the attention weight of the first vehicle to empty vehicle nodes in the destination area is increased.

[0012] Optionally, the graph neural network model is a heterogeneous graph neural network, which includes: a heterogeneous graph attention layer, used to learn the contribution of edges of different business association types to node feature updates; a meta-path learning layer, used to capture high-order composite relationships between vehicles through predefined meta-paths, wherein the meta-paths include at least a composite path of "vehicle-business edge-vehicle"; and an anomaly detection layer, used to output the anomaly score of each vehicle node or the classification result of whether each vehicle subgraph is an anomalous subgraph.

[0013] Optionally, the method for identifying abnormal stops and illegal dumping risks in the transportation of solid and hazardous waste includes identifying abnormal vehicle subgraph structures, specifically: identifying vehicle subgraphs connected by business edges and recently physically close, which are determined to be illegal connection patterns; identifying vehicle subgraphs in the same high-risk area that are densely connected by business edges to form a cluster structure, which are determined to be abnormal clustering patterns; and identifying vehicle pairs connected by business edges and whose trajectories have a fixed time delay similarity, which are determined to be abnormal following patterns.

[0014] Optionally, it also includes a historical association edge construction step: when two vehicles have been physically close during the existence of a business edge, the business edge is transformed into a permanent historical association edge and stored in the historical association database; in the subsequent risk identification process, when a newly added vehicle has a business association with a vehicle in the historical association database, the initial risk score of the newly added vehicle is increased.

[0015] Optionally, it also includes an early warning step: based on the vehicle's current trajectory and business semantic embedding vector, predict vehicle pairs that may be physically close in the future through a link prediction task; when it is predicted that two vehicles with high business semantic similarity and current physical distance less than a preset threshold will be physically close within a preset time window, generate early warning information.

[0016] Compared with existing technologies, the present invention has the following advantages: The present invention abandons the traditional graph structure's reliance on time window updates. By using loading and unloading and waybill updates as real-time trigger signals for the graph topology, and establishing a "disconnect-activate-decay" control rule for business edges, the graph structure can evolve synchronously with changes in business status. This ensures that the risk identification model always makes judgments based on the latest business intentions, which helps to solve the problem of untimely and inaccurate risk identification caused by graph structure lag, and significantly improves the accuracy and timeliness of risk identification in solid and hazardous waste transportation. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the system architecture of an embodiment of the present invention.

[0019] Figure 2 This is a flowchart illustrating a risk identification method according to an embodiment of the present invention.

[0020] Figure 3 This is a schematic diagram of the structure of a graph neural network model according to an embodiment of the present invention.

[0021] Figure 4 This is a schematic diagram of the dynamic evolution process of heterogeneous graphs according to an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0023] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” used in the embodiments of this invention and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise. “Multiple” generally includes at least two, but does not exclude the inclusion of at least one.

[0024] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0025] Depending on the context, the words “if” or “suppose” as used here can be interpreted as “when” or “in response to determination” or “in response to detection.” Similarly, depending on the context, the phrases “if determination” or “if detection (of the stated condition or event)” can be interpreted as “when determination” or “in response to determination” or “when detection (of the stated condition or event)” or “in response to detection (of the stated condition or event).”

[0026] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a product or system comprising a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a product or system. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the product or system that includes said element.

[0027] Furthermore, the timing of the steps in the following method embodiments is merely an example and not a strict limitation.

[0028] The intelligent identification method for identifying abnormal stops and illegal dumping risks during the transportation of solid and hazardous waste provided in this embodiment of the invention can be executed by an electronic device, such as a PC, laptop, or smartphone, or a server. The server can be a physical server containing an independent host, a virtual server, a cloud server, or a server cluster.

[0029] Unless otherwise defined, all technical and scientific terms used in the embodiments of this application have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in the embodiments of this application is for the purpose of describing the embodiments of this application only and is not intended to limit this application.

[0030] Before providing a further detailed description of the embodiments of this application, some of the nouns and terms involved in the embodiments of this application will be explained. The nouns and terms involved in the embodiments of this application are subject to the following interpretations.

[0031] (1) Initial heterogeneous graph: refers to a graph structure data in which the vehicles involved in the transportation of solid and hazardous waste are the nodes of the graph, and the business logic relationship between the vehicles based on their electronic waybill data is the edge of the graph, used to represent a snapshot of the business status of the entire transportation network at a specific moment. This graph is "heterogeneous" because the edges can represent different types of business relationships, such as "related to the same cargo owner" and "related to upstream and downstream connections".

[0032] (2) Cargo status change events: These are specific business signals that can trigger real-time evolution of the graph structure. These events signify that the transportation business has entered a new stage, thus requiring a reassessment and adjustment of the business relationships between vehicles. In this application, such events include at least vehicle loading and unloading events (e.g., a significant change in cargo weight detected by a weighbridge system or on-board sensors) and electronic waybill status update events (e.g., a change in waybill status from "in transit" to "received").

[0033] (3) Waybill Semantic Embedding Vector: This refers to a high-dimensional mathematical vector that uses deep learning models (such as pre-trained language models) to transform unstructured waybill text information (such as cargo description, consignor, destination, etc.) into a digital representation that computers can understand and compute. This vector can capture the deeper "business meaning" behind the waybill, making waybills with similar business purposes closer to each other in the vector space.

[0034] (4) Business edge: This refers to the edge connecting two vehicle nodes in a heterogeneous graph, representing a logical business-level relationship between the two vehicles. This relationship is not based on physical spatiotemporal co-occurrence, but on the inherent connection of their transportation tasks, such as transporting similar goods, serving the same cargo owner, or having a potential cargo handover relationship. The establishment of business edges is the basis for risk identification.

[0035] (5) Business Status Adjustment Factor: This refers to a dynamic weighting factor introduced during the calculation of neighbor information aggregation in the graph neural network. This factor is dynamically generated based on the real-time business status of the vehicle node (such as whether the current cargo is fully loaded or empty, the percentage of the trip completed, and whether the time for transportation is urgent). It is used to adjust the attention weight of the node to different neighbor nodes, so that the model can "intelligently" focus on the most important related information in the current business scenario.

[0036] (6) Meta-path: refers to a predefined composite path pattern in a heterogeneous graph, consisting of alternating nodes and edges of different types. It is used to capture higher-order and more complex semantic relationships between nodes that go beyond direct connections. For example, a meta-path such as “Vehicle A - (business edge) - Vehicle B - (physical proximity) - Vehicle C” can reveal that Vehicle A had an indirect cargo transfer with Vehicle C through Vehicle B, which is crucial for discovering complex gang crime patterns.

[0037] (7) Historical association edge: This refers to a situation where two vehicles actually approach or interact physically during the existence of a temporary "business edge" (e.g., GPS trajectories highly overlap within a specific spatiotemporal window). The system will solidify this "verified" business edge into a permanent historical association and store it in a dedicated database. It constitutes the vehicle's "social" risk profile, used for rapid risk assessment of new vehicles associated with it in the future.

[0038] Reference Figure 1 and Figure 2 This invention provides an intelligent identification method for the risks of abnormal stops and illegal dumping during the transportation of solid and hazardous waste. This method aims to address the fundamental problem that existing risk identification models rely on static vehicle relationship graphs that lag behind business changes. This invention abandons the traditional "time-driven" model based on fixed-time-window graph structure updates and innovatively adopts an "event-driven" model, enabling the graph's topology to evolve synchronously with real-time changes in the transportation business status, thereby significantly improving the accuracy and timeliness of risk identification.

[0039] In a basic implementation, such as Figure 2As shown, the method includes the following core steps. First, in step 201, an initial heterogeneous graph P0 is constructed, with vehicles as nodes and business logic relationships between vehicles as edges; here, the nodes represent all en route solid and hazardous waste transport vehicles, and the edges represent the inherent business-based connections between them. For example, P0 = (V, E0). Wherein, the node set V represents all currently en route solid and hazardous waste transport vehicles. The edge set E0 represents the business logic relationships between vehicles, which are determined based on the vehicle's electronic waybill data.

[0040] Next, in step 202, the system monitors cargo status change events for each vehicle in real time. It continuously monitors the cargo status change events for each transport vehicle. These events are key indicators of business process progress; for example, when a vehicle completes unloading at a processing plant, its cargo status changes from "fully loaded" to "empty." Or, when a vehicle loads new cargo at a transfer station, its waybill status is updated to "picked up." These events are the core signal sources driving the evolution of the graph structure. Cargo status change events include at least loading / unloading events and waybill status update events, which serve as trigger signals for the evolution of the current transport heterogeneous graph Pn structure. Figure 1 As shown, these events can come from multiple data sources 110, such as sensor data reported by the vehicle terminal 111, status updates from the electronic waybill management system 112, and weighing records from the weighbridge system 113 of the treatment plant or waste-generating enterprise. These heterogeneous event streams are aggregated and transmitted through a unified message queue 120 (e.g., using Kafka Topics 121 for topic isolation).

[0041] Then, in the judgment phase of step 205, it is checked whether the aforementioned cargo status change event has been detected. If no event is detected, the system will continue to maintain the monitoring state in step 202. Once a detected cargo status change event is detected, the process proceeds to step 203. In this step, the topology and edge weights of the business edges in the current transportation heterogeneous graph Pn are dynamically adjusted, so that the topology of the current transportation heterogeneous graph Pn evolves in real time with the business status, generating the evolved heterogeneous graph Pn+1; here, Pn represents the graph at the current moment, where n is an integer greater than or equal to 0, and P0 is the initial heterogeneous graph. For example, an unloading event may cause a business edge to be disconnected, while a new loading event may activate a completely new business edge. In this way, the current transportation heterogeneous graph Pn evolves into a new heterogeneous graph Pn+1 that better reflects the latest business reality. This evolution process is carried out by... Figure 1 The graph evolution engine 131 within the stream processing module 130 is responsible for execution. This engine triggers the logic of graph evolution 133 based on the received events and may rely on state storage 132 to maintain the latest state of the vehicle.

[0042] Finally, in step 204, the evolved heterogeneous graph Pn+1 is input into the graph neural network model to extract heterogeneous graph structural features from Pn+1, which are used to identify abnormal stops and illegal dumping risks during the transportation of solid and hazardous waste. The graph neural network model can be used to learn and extract deep-level pattern features from complex graph structure data. By analyzing the node connection patterns, subgraph structure, and feature distribution in the evolved heterogeneous graph Pn+1, the model can effectively identify potential abnormal stops and illegal dumping risks during the transportation of solid and hazardous waste. For example, the model may find that several vehicles that are not related in business form a tight cluster in the graph, which may indicate an abnormal aggregation. After identifying the risk, the system can output the result and return to step 202 to continue listening for new events, forming a closed-loop process of continuous monitoring, dynamic evolution, and real-time identification.

[0043] This basic implementation establishes a risk identification framework that synchronizes business operations. Each update to the graph structure is triggered by a cargo status change event, rather than relying on a fixed time interval, thereby ensuring that the business relationship graph upon which the risk identification model is based remains synchronized with the actual transportation status, which helps reduce misjudgments and omissions caused by information lag.

[0044] In a preferred embodiment, the process of constructing the initial heterogeneous graph P0 in step 201 above is specified. To accurately capture the subtle, potential business relationships between vehicles, the determination of the initial business logic relationships includes deep semantic analysis based on natural language processing. Specifically, the system first acquires the vehicle's electronic waybill data. This data, in addition to structured information (such as license plate number and cargo owner ID), more importantly contains unstructured cargo description text, such as "waste circuit boards containing precious metals HW49" or "HW08 mixture of waste mineral oil and oily wastewater".

[0045] After acquiring these texts, the system uses a pre-trained language model to semantically encode the cargo description text, generating a waybill semantic embedding vector for each waybill. This process ensures that even if the text descriptions differ, as long as the business essence is similar (e.g., both transporting used batteries), the generated vectors will be very close in mathematical space. Subsequently, the system quantifies the similarity of the business content between any two vehicles by calculating the cosine similarity between their waybill semantic embedding vectors. For example, the system calculates the cosine similarity between any two vehicles... i and j Waybill semantic embedding vector and Cosine similarity between = When the calculated cosine similarity exceeds a preset threshold (e.g., 0.9), the system considers the two vehicles to be highly correlated in terms of business, and thus establishes a business edge between their corresponding nodes. ∈ E 0, as part of the initial heterogeneous graph P0. This semantic-based edge construction method can uncover "hidden" relationships that are difficult to find using traditional rule-based methods, enriching the information content of the graph and providing a more solid foundation for subsequent risk identification.

[0046] Furthermore, in the semantic encoding process using a pre-trained language model, to ensure the generated waybill semantic embedding vector contains more comprehensive business information, the system performs a text concatenation preprocessing step. Specifically, it doesn't just use the cargo description text, but concatenates multiple dimensions of information, including the cargo description text, shipper information, destination information, and transportation timeliness requirements, into a longer waybill text sequence according to a standardized template. For example, a complete sequence might be: "[CLS] Cargo: Discarded mercury-containing lamp tubes HW29; Shipper: XX Lighting Company; Destination: YY Hazardous Waste Disposal Center; Timeliness: Within 24 hours [SEP]". This standardized, more information-rich waybill text is then input into a pre-trained language model (e.g., the BERT model). After processing, the model extracts the hidden state vector corresponding to the special token "[CLS]" from its output layer. This vector is generally considered to aggregate the global semantic information of the entire input sequence; therefore, it is used as the waybill semantic embedding vector representing the comprehensive business meaning of the waybill. In this way, the generated vectors not only contain the characteristics of the goods themselves, but also incorporate business context information, which helps to improve the accuracy and reliability of similarity calculation.

[0047] In another preferred embodiment, the service edge in the initial heterogeneous graph P0 The initial weights are not fixed values, but are dynamically calculated based on business association patterns to more precisely characterize the strength and risk level of the association. This dynamic calculation considers at least the following patterns: The first is the "same cargo owner / same destination pattern." When the system parses the structured data of the electronic waybill and finds that the cargo owner information of two vehicles is exactly the same, or that their transport destinations point to the same geographical location, the system determines that they have a clear business collaboration relationship. The weight of their corresponding business edge will be increased by an additional first preset value, such as 0.2, on top of the basic semantic similarity. This enhances the ability to identify vehicle clusters serving the same business entity.

[0048] The second type is the "upstream and downstream connection mode," which is a typical high-risk behavior in the illegal transfer of solid and hazardous waste. The system continuously monitors vehicle status and location. When it finds that the destination of the first vehicle's waybill and the origin of the second vehicle's waybill are geographically highly overlapping (e.g., less than 1 kilometer apart), and within a similar time window, the first vehicle's cargo status is "fully loaded and awaiting delivery," while the second vehicle's cargo status is "empty return trip" or "empty and awaiting loading," the system determines that this is highly likely an illegal "vehicle-to-vehicle" cargo transfer. In this case, the weight of the business side is increased by a second preset value significantly higher than the former, such as 0.8. By assigning this mode a very high weight, the model can be highly sensitive to potential illegal transfer networks.

[0049] The third type is the "value collaboration model." The system combines semantic and geographic information for a comprehensive judgment. When decoding and analyzing the semantic embedding vectors of waybills (e.g., through a classifier), and it is found that both vehicles are transporting high-value or high-environmentally-risk solid and hazardous waste (such as catalysts containing precious metals or highly toxic chemical waste), and their real-time GPS locations simultaneously appear in a "high-risk area" pre-marked as a high-incidence zone for illegal dumping, the system considers this "aggregation of high-value goods in a dangerous location" itself a risk signal. Therefore, the business edge weights between them are increased by a third preset value, such as 0.5. The introduction of this model allows the model to combine the inherent risks of the goods with the risks of the external environment for a more comprehensive risk assessment.

[0050] Based on the above embodiments, the dynamic adjustment process of the diagram, i.e., step 203, will be described in more detail. For example... Figure 4 As shown, this process includes selectively disconnecting, activating, or attenuating business edges. For example, in a graph before evolution Pn400, vehicle A (fully loaded) 401 has a business edge with destination 402 (or another entity representing that destination) because its waybill destination is 402. At this time, there may be an empty vehicle C 403 nearby, but there is no effective connection between the two. When the system detects an event: vehicle A unloads 420, causing its state to change from "fully loaded" to "empty", the graph evolution engine will perform an "edge disconnection" operation. In the graph after evolution Pn+1 450, the business edge that originally connected vehicle A (empty) 451 and destination 452 is logically disconnected or its weight is reduced to a very low value because this transportation business has been completed. At the same time, the system will perform an "edge activation" operation to actively detect whether there are other potential business opportunities near vehicle A (empty) 451. For example, it will activate the calculation of potential business associations with nearby empty vehicle C 453. If their waybill semantics (such as return freight information) are similar, a new business edge may be established.

[0051] Similarly, when the system detects that the cargo status of a second vehicle changes from "empty" to "fully loaded" and its current location is very close to the loading point specified on its waybill, it will assume that the vehicle has started a new transportation task. At this time, the system will proactively activate the business edge between this vehicle and other empty vehicles in the destination area. This simulates the real business scenario in the solid waste transportation network where vehicles usually look for return cargo after unloading, and proactively monitors potential return transfer behaviors.

[0052] Furthermore, to ensure the graph reflects the timeliness of business intent, the system assigns a decay coefficient to each business edge that decreases over time. The calculation formula is as follows: .in, It is the decay coefficient at time t, where t0 is the timestamp of when the business edge was established or most recently activated. It is an adjustable decay rate. If two vehicles do not physically approach each other within a preset time window (e.g., 24 hours) after the business edge is established (e.g., the GPS track distance is always greater than 1 kilometer), then the weight of this edge will gradually decrease according to this decay coefficient until it eventually falls below a certain threshold and is removed from the graph. This "use it or lose it" mechanism helps to retain only those currently valid and meaningful business connections in the graph, and helps to avoid the interference of outdated information on risk assessment. Furthermore, in order to enable the graph neural network model to more intelligently understand and utilize the dynamically evolving graph structure, a business state-aware attention adjustment mechanism is also introduced in the dynamic adjustment process. When the graph neural network aggregates neighbor node information, the attention weight of a vehicle node to its neighbor nodes depends not only on the similarity of their features, but also on a "business state adjustment factor" dynamically generated according to the business state of the vehicle node itself. The formula for calculating this attention weight can be expressed as:

[0053] in, Its meaning is the attention weight of the central node i to the neighbor node j, which represents the proportion of information from the neighbor node j when updating the features of node i. and These are the feature vectors of node i and node j, such as waybill semantic embedding, location, trajectory features, etc. It is a learnable feature transformation matrix; It is a learnable attention parameter vector; This represents a vector concatenation operation; LeakyReLU is the activation function. It is the set of neighboring nodes of node i; and This is a business status adjustment factor. This is dynamically generated based on the current cargo status (e.g., fully loaded, empty), journey progress (e.g., 95% complete), and time urgency (e.g., less than 1 hour until the scheduled delivery time) of the central node i. For example, when the first vehicle's journey progress exceeds a preset threshold (e.g., 90%) and its cargo status is "fully loaded and awaiting delivery," the system determines that it is about to unload, which presents a high risk of illegally connecting with a return vehicle. Therefore, the system increases the business status adjustment factor for the first vehicle on neighboring vehicle nodes located within its destination area that are "empty." The value of this will directly result in these "suspicious" unloaded vehicles receiving higher attention weights in the attention calculation. This allows the model to incorporate more information from these high-risk neighbors when updating the node representation of the first vehicle, enhancing the model's ability to predict impending connection risks.

[0054] In one specific implementation, the graph neural network model used in step 204 is described in detail. For example... Figure 3 As shown, the graph neural network model 300 is designed as a heterogeneous graph neural network, specifically for processing complex graph data containing various node and edge types. When the evolved heterogeneous graph Pn+1 301 is input into the model, it undergoes processing through the following layers. First is the heterogeneous graph attention layer 302, which distinguishes different types of business edges in the graph (such as same-owner edges, upstream / downstream connection edges, value collaboration edges, etc.) and learns independent transformation matrices and attention weights for each type of edge. This allows the model to assign higher weights to 'connection' relationships compared to 'same-owner' relationships, and correspondingly, to give them a greater influence in information transmission.

[0055] Next, the data flows through the meta-path learning layer 303. To capture higher-order, complex relationships between vehicles beyond direct connections (e.g., an illicit transport chain linked by an intermediary vehicle), this layer learns through predefined meta-paths. These meta-paths could be, for example, "vehicle-business edge-vehicle" to capture direct business partnerships, or "vehicle-business edge-vehicle-business edge-vehicle" to discover second-degree business connections. The model "walks" along these meta-paths on the graph, aggregating information from all nodes along the path to generate a feature representation for each vehicle node that incorporates higher-order neighborhood structures. This helps identify organized, multi-stage gang-related patterns of crime.

[0056] Finally, the node features, incorporating multi-layered information, are fed into the anomaly detection layer 304. This layer can be a simple multilayer perceptron or a more complex classifier. Its task is to output an anomaly score for each vehicle node based on the final node feature representation, or to directly classify whether the vehicle subgraphs formed by business edges are anomaly subgraphs. Ultimately, the model outputs anomaly score / classification result 305, providing regulators with a clear risk warning.

[0057] Furthermore, the process described above for identifying abnormal stops and illegal dumping risks during the transportation of solid and hazardous waste is specifically manifested in identifying specific abnormal vehicle subgraph structures within the graph. These structures are a visual representation of complex risk behaviors on the graph topology. For example, the model will focus on identifying "illegal connection patterns," that is, looking for vehicle pairs or small-scale subgraphs connected by high-weight "upstream and downstream connection" business edges in the graph, and whose GPS trajectories show physical proximity (distance less than 50 meters) in recent times (e.g., within the past hour). This pattern is a high-risk indicator indicating possible illegal transfer or resale.

[0058] The model also identifies "abnormal clustering patterns." When it detects multiple vehicles forming a clique on the graph through dense business edges in a high-risk area (such as a remote industrial park or a suburban area), and these vehicles also exhibit clustering behavior in physical space, the system determines that this may be an illegal solid and hazardous waste collection or processing point. Finally, the model also identifies "abnormal following patterns," which involves searching for vehicle pairs on the graph connected by business edges whose historical GPS trajectories show a fixed time lag and are highly similar in space. For example, if vehicle A's trajectory always leads vehicle B's trajectory by about 30 minutes, this may suggest a collaborative operation between a "scout vehicle" and a "transport vehicle."

[0059] To enable the system's risk identification capabilities to continuously learn and evolve, this invention also includes a step of constructing historical association edges. When the system detects that two vehicles (e.g., vehicle A and vehicle B) have indeed had one or more physical proximity interactions during the existence of their business edge, the system transforms this "fact-verified" business edge, along with information such as the time, location, and business context of the interaction, into a permanent "historical association edge" and stores it in a dedicated historical association graph database. In future risk identification processes, if a newly added vehicle C establishes a new business edge with a "habitual offender" vehicle A in the historical database, the system will automatically retrieve vehicle A's historical association network when calculating vehicle C's initial risk score. It will discover that vehicle A had an illegal interaction record with vehicle B, thereby increasing vehicle C's initial risk score and monitoring level based on this historical association. This facilitates the rapid inheritance and dissemination of risk profiles, improving the efficiency of discovering emerging risk groups.

[0060] Furthermore, to achieve the leap from "post-event detection" to "pre-event warning," this invention also includes an early warning step. The system utilizes a link prediction model based on a graph attention network. This model takes the currently dynamically evolving heterogeneous graph and the real-time features of vehicles as input. Its task is to predict which pairs of vehicles in the graph, which have not yet physically approached each other, have the highest probability of physically approaching each other within a future time window (e.g., the next 2 hours). When the model predicts that two vehicles have extremely high similarity in their waybill semantic embedding vectors (meaning highly consistent business purposes), and their current physical distance is less than a certain warning threshold (e.g., 5 kilometers), and the model predicts that their probability of physically approaching each other within the next 2 hours exceeds 90%, the system immediately generates a high-priority early warning message and pushes it to the monitoring center. This warning message clearly identifies the suspicious vehicle pairs, their possible meeting points, and times, allowing regulatory personnel to intervene and deploy controls in advance, helping to prevent illegal activities.

[0061] This invention can be materialized as a hardware and software integrated intelligent monitoring system. (See reference...) Figure 1The system comprises a data acquisition layer, an event processing layer, and an analytics application layer. The data acquisition layer consists of data sources 110 deployed in various locations, including vehicle-mounted terminals 111 on thousands of transport vehicles, a business platform recording electronic waybills 112, and weighbridge systems 113 distributed throughout waste generation and disposal units. Data generated from these sources is fed into the system in real time via message queues 120. The event processing layer is the core of the system, consisting of a graph evolution engine 131 built on a stream processing framework 130 (such as Apache Flink or Spark Streaming) and a state store 132 (such as Redis) for caching real-time vehicle status. The graph evolution engine 131 consumes events from the message queue and triggers graph evolution 133 according to preset rules. The analytics application layer deploys the aforementioned heterogeneous graph neural network model. It receives the latest graph structure generated by the graph evolution engine, performs risk calculations, and presents the results to regulators through a visualization interface or triggers automatic early warning processes through an API interface.

[0062] Now, we will describe an implementation of the present invention that integrates multiple technical features through a complete end-to-end scenario. Assume there are vehicles A and B. Vehicle A's waybill is described as "transporting HW08 waste mineral oil to processing plant A", and its status is "fully loaded"; vehicle B's waybill is described as "returning empty from near processing plant A", and its status is "empty".

[0063] First, the system acquires the electronic waybills for both vehicles and semantically encodes their waybill text using a pre-trained language model, generating waybill semantic embedding vectors. Since both involve "Processing Plant A" and similar business scenarios, their vector cosine similarity is very high (e.g., 0.92), exceeding a threshold, thus establishing an initial business edge. Next, the system identifies this as conforming to an "upstream and downstream connection pattern," therefore setting the initial weight of this business edge to a high value (e.g., 0.8). In this way, an initial heterogeneous graph P0 is constructed, in which a high-risk strong connection exists between A and B.

[0064] During transport, the system monitors the status of both vehicles in real time. When vehicle A's GPS shows that its journey progress has reached 95%, the business status awareness attention adjustment mechanism takes effect, and the system dynamically generates a business status adjustment factor for vehicle A. This factor increases the attention weight of vehicle A towards its neighbor vehicle B in the graph neural network calculation. Meanwhile, the link prediction model, based on the high-weight business edge between A and B, the continuously decreasing physical distance, and similar semantic vectors, predicts that they are highly likely to physically approach each other in an industrial park near processing plant A within the next hour. The system then generates and issues an early warning message.

[0065] When the cargo status change event (full load → empty load) of vehicle A is detected, the graph evolution engine performs the following operations: (1) disconnect the business edge between vehicle A and the corresponding node of processing plant A; (2) activate the potential business association calculation between vehicle A and nearby empty vehicles, and further enhance the business edge weight between A and B. If the GPS trajectories of vehicle A and vehicle B overlap within a preset spatiotemporal window (distance ≤ 50 meters, time window ≤ 30 minutes), the anomaly detection layer outputs that the anomaly score of the subgraph exceeds the preset threshold, and the system generates an illegal connection risk warning.

[0066] After the event concludes and is manually verified, the system transforms the verified business edge between A and B into a permanent historical association edge and stores it in the historical association database. Several months later, when a new vehicle C establishes a business edge with vehicle A for transporting the same type of goods, the system automatically raises C's initial risk level based on A's historical risk records, placing it under close monitoring. Through this process, the system provides an intelligent regulatory method integrating semantic understanding, dynamic evolution, early warning, assisted identification, and self-learning.

[0067] The technical solution provided by this invention can be widely applied to the supervision of solid and hazardous waste transportation by environmental protection departments and transportation management departments at all levels. It can serve as a core intelligent engine, integrated into existing national or provincial solid and hazardous waste transportation monitoring platforms. By linking data with existing systems such as electronic fences, video surveillance, and historical case databases, this invention can elevate isolated, rule-based monitoring methods into a three-dimensional, intelligent regulatory system capable of analyzing the overall business network and perceiving dynamic risk evolution. For example, when the system issues an early warning, the monitoring platform can automatically dispatch drones or cameras near the warning location for focused monitoring and evidence collection, helping to improve the efficiency and deterrent effect of regulatory enforcement. Furthermore, through macroscopic analysis of long-accumulated graph-structured data, it can also reveal regional and industry-specific networks and patterns of illegal solid and hazardous waste transfer, providing data-driven decision support for policy formulation and source control, thereby contributing to environmental safety at a higher level.

[0068] In addition, embodiments of the present invention provide a non-transitory machine-readable storage medium storing executable code. When the executable code is executed by a processor of an electronic device, the processor can at least implement the intelligent identification method for abnormal stay and illegal dumping risks in solid and hazardous waste transportation as provided in the foregoing embodiments.

[0069] The device embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separate. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0070] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of a necessary general-purpose hardware platform, or by a combination of hardware and software. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a computer product. The present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent identification method for the risks of abnormal stoppage and illegal dumping during the transportation of solid and hazardous waste, characterized in that, include: Construct an initial heterogeneous graph P0 with vehicles as nodes and business logic relationships between vehicles as edges, wherein the business logic relationships are determined based on the vehicle's electronic waybill data; For any given moment, the current heterogeneous transportation graph Pn is a heterogeneous graph P0 or an intermediate heterogeneous graph obtained by at least one evolution of the initial heterogeneous graph P0. The cargo status change events of each vehicle are monitored in real time. The cargo status change events include at least loading and unloading events and waybill status update events, which serve as trigger signals for the structural evolution of the current heterogeneous transportation graph Pn. In response to the detected cargo status change event, the topology and edge weights of the business edges in the current heterogeneous transportation graph Pn are dynamically adjusted so that the topology of the current heterogeneous transportation graph Pn evolves in real time with the business status, generating the evolved heterogeneous graph Pn+1; wherein, the dynamic adjustment includes selectively disconnecting, activating or attenuating business edges according to the type and content of the cargo status change. The evolved heterogeneous graph Pn+1 is input into a graph neural network model, and heterogeneous graph structural features are extracted from the heterogeneous graph Pn+1 to identify the risks of abnormal stay and illegal dumping during the transportation of solid and hazardous waste.

2. The method according to claim 1, characterized in that, Determining the business logic association based on the vehicle's electronic waybill data includes: Obtain electronic waybill data for the vehicle, wherein the electronic waybill data includes at least unstructured cargo description text; The cargo description text is semantically encoded using a pre-trained language model to generate a waybill semantic embedding vector. Calculate the cosine similarity between the semantic embedding vectors of any two vehicles' waybills. When the cosine similarity exceeds a preset threshold, establish a business edge between the two vehicles.

3. The method according to claim 2, characterized in that, The semantic encoding using a pre-trained language model includes: The cargo description text, cargo owner information, destination information, and transit time information are combined into a standardized waybill text; The waybill text is input into a pre-trained language model, and the hidden state of the [CLS] token output by the language model is taken as the semantic embedding vector of the waybill.

4. The method according to claim 2, characterized in that, The initial weight of the business edge is dynamically calculated based on the business association pattern, which includes at least one or more of the following: Same shipper / same destination mode: When the shipper information or the destination is the same in the electronic waybill data of two vehicles, the business edge weight is increased by the first preset value; Upstream and downstream connection mode: When the destination of the first vehicle's waybill is the same as the origin of the second vehicle's waybill, and the cargo status of the first vehicle is fully loaded and awaiting delivery, while the cargo status of the second vehicle is empty and returning, the business edge weight is increased by a second preset value, which is greater than the first preset value. Value Collaboration Mode: When the goods of both vehicles are high-value goods and the real-time locations of both vehicles are located within a preset high-risk area, the business edge weight is increased by a third preset value.

5. The method according to claim 1, characterized in that, The selective disconnection, activation, or attenuation of service edges includes: When the cargo status of the first vehicle changes from fully loaded to empty, disconnect the business edge related to the destination of the first vehicle and activate the potential business association between the first vehicle and nearby empty vehicles. When the cargo status of the second vehicle changes from empty to fully loaded and its current location is close to the loading point, activate the business edge between the second vehicle and the empty vehicles in the destination area. A decay coefficient is set for each service edge, which decreases over time. When two vehicles do not physically approach each other within a preset time window after the service edge is established, the weight of the service edge gradually decreases according to the decay coefficient until it disappears.

6. The method according to claim 5, characterized in that, The dynamic adjustment also includes an attention regulation mechanism that introduces business state awareness: In the neighbor aggregation process of the graph neural network, attention weights for each vehicle node to its neighbor nodes are calculated; the calculation formula of the attention weights incorporates a business state adjustment factor, which is dynamically generated based on the current cargo status, travel progress and time urgency of the vehicle node. When the first vehicle's journey progress exceeds a preset threshold and the cargo status is fully loaded and ready for delivery, the attention weight of the first vehicle to empty vehicle nodes in the destination area is increased.

7. The method according to claim 1, characterized in that, The graph neural network model is a heterogeneous graph neural network, which includes: A heterogeneous graph attention layer is used to learn the contribution of edge pairs to node feature updates for different business association types. The meta-path learning layer is used to capture high-order composite relationships between vehicles through predefined meta-paths, wherein the meta-paths include at least a composite path of "vehicle-business edge-vehicle". The anomaly detection layer is used to output the anomaly score for each vehicle node or the classification result of whether each vehicle subgraph is an anomaly subgraph.

8. The method according to claim 1, characterized in that, The method for identifying abnormal stops and illegal dumping risks during the transportation of solid and hazardous waste includes identifying abnormal vehicle sub-graph structures, specifically: Identify vehicle subgraphs connected by business edges that have recently experienced physical proximity and determine them as illegal connection patterns; Identify vehicle subgraphs within the same high-risk area that form clusters due to dense connections of business edges, and determine them as abnormal clustering patterns. Identify vehicle pairs connected by business edges and whose trajectories have a fixed time delay similarity, and determine them as abnormal following patterns.

9. The method according to claim 1, characterized in that, It also includes a historical association edge construction step: when two vehicles have physically approached each other during the existence of a business edge, the business edge is transformed into a permanent historical association edge and stored in the historical association database; In the subsequent risk identification process, when a newly registered vehicle has a business relationship with a vehicle in the historical association database, the initial risk score of the newly registered vehicle will be increased.

10. The method according to claim 1, characterized in that, It also includes early warning steps: Based on the vehicle's current trajectory and business semantic embedding vector, a link prediction task is used to predict vehicle pairs that may be physically close in the future. When it is predicted that two vehicles with high business semantic similarity and a current physical distance of less than a preset threshold will physically approach each other within a preset time window, an early warning message is generated.

Images