Solid waste traceability analysis method based on knowledge graph

Abstract

The invention relates to a solid waste traceability analysis method based on a knowledge graph, and the method comprises the steps: completing the time-space collection, denoising standardization, unified alignment and geocoding through collection and fusion of multi-modal data, such as a structured circulation document, a GPS track, an RFID signal, video monitoring and an environment sensor, and constructing a node-level confidence probability model; and the abnormal node state is cooperatively corrected by using a multi-modal signal, the solid waste flow direction knowledge graph is dynamically updated by a corrected result, global traceability and service decision feedback of an abnormal path are realized, and the judgment accuracy of the abnormal traceability of the solid waste flow direction is effectively improved in combination with continuous model iteration and knowledge closed-loop optimization.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of "solid waste flow intelligent supervision", and in particular to a solid waste traceability analysis method based on a knowledge graph. BACKGROUND

[0002] In the current field of solid waste flow supervision and traceability analysis, with the increasing requirements for ecological environment supervision, solid waste full-process digital traceability and abnormality identification have become an important technical direction for smart environmental protection, city management and hazardous waste whole-process supervision. The existing mainstream technologies generally rely on information management of transfer documents, Internet of Things (GPS, RFID), video monitoring and sensor integration, and other multi-modal data collection methods, and cooperate with rule engines or basic abnormality detection algorithms to visually track and check the compliance of solid waste flow links. In recent years, new generation cognitive intelligent technologies such as knowledge graph, causal reasoning and multi-source information fusion have been gradually introduced into the solid waste supervision scene, promoting the transformation of solid waste flow management from traditional post-tracing to real-time early warning and intelligent traceability. Specifically, the commonly seen solid waste flow monitoring system in the industry mainly supports the centralized collection of transfer node data, basic data cleaning and automatic reconciliation, and assists in realizing path tracking by using conventional trajectory fitting, RFID time sequence abnormality detection and other methods, but the adaptive fusion of multi-source heterogeneous data, the refinement of abnormality identification and the level of uncertainty correction are insufficient.

[0003] The current solid waste flow traceability abnormality identification field faces the following key technical bottlenecks:

[0004] (1) Spatial drift and time error generated by multi-modal spatio-temporal data collection are difficult to be effectively and adaptively corrected through simple data alignment and rule discrimination, abnormal nodes frequently misjudge, which seriously affects the accuracy and business operability of abnormal traceability.

[0005] (2) The existing knowledge graph structure cannot dynamically optimize node and path weights according to high confidence data, resulting in that abnormal link reasoning does not have a quasi-real-time closed loop, and abnormal suspicious scores lack multi-source confidence fusion.

[0006] (3) The utilization of heterogeneous data supplements (such as video, RFID supplementary recording and sensing information) in abnormality identification is limited, and a complete multi-modal perception signal automatic collaborative correction mechanism has not been established, and the abnormal traceability path cannot dynamically eliminate data uncertainty.

[0007] (4) The traceability abnormality judgment lacks fine probability quantization and causal link priority sorting, and the abnormality judgment process is difficult to transparently interact, which hinders the decision-making closed loop and model adaptive evolution.

[0008] (5) The closed loop mechanism of data-model-knowledge is weak, and it is difficult to form an intelligent abnormality identification platform with iterative self-learning and continuous optimization and strong generalization ability. SUMMARY

[0009] The application provides a solid waste traceability analysis method based on a knowledge graph, aiming to solve one of the technical problems in the prior art.

[0010] The application provides a solid waste traceability analysis method based on a knowledge graph, specifically comprising:

[0011] S1: Collecting multi-modal data of solid waste flow key nodes, the multi-modal data including structured flow documents, GPS trajectory data, radio frequency identification signals, video monitoring images and environmental sensor spatio-temporal information, to obtain solid waste flow multi-modal original data sets across different geographical regions and time periods.

[0012] S2: Spatio-temporal data cleaning and unified preprocessing of the multi-modal original data set, including denoising, format standardization and timestamp alignment processing of various data, to eliminate synchronization errors caused by heterogeneous collection devices and unify geographic coding.

[0013] S3: For different node geographical regions and time period labels, input the preprocessed multi-modal data into a Bayesian network inference model to dynamically calculate the confidence probability distribution interval of the spatio-temporal information of each node, and realize the probability quantization of the solid waste flow data quality.

[0014] S4: Based on the node confidence probability distribution interval output by the Bayesian network, use an abnormal deviation discrimination algorithm to identify the spatio-temporal sequence anomaly in the solid waste flow path, and judge the spatial and temporal consistency of the actual flow path and the declared path under the specified probability threshold.

[0015] S5: If abnormal deviation is found in the path discrimination of a specific node, call the heterogeneous perception signals associated with the abnormal node, including supplementary video monitoring, RFID supplementary information and residual sensor data, to reposition and probability correct the abnormal node spatio-temporal state in multi-modal collaboration.

[0016] S6: Taking the key node confidence probability distribution corrected by multi-modal collaboration as input, dynamically updating the solid waste flow node and path relationship in the knowledge graph, obtaining the solid waste flow information corrected and injected into the knowledge graph, and realizing high confidence injection of solid waste flow knowledge in different geographical regions and different time periods.

[0017] S7: Based on the solid waste flow information corrected and injected into the knowledge graph, using a weighted causal reasoning algorithm to calculate the global abnormal suspicious score of each flow path, and adjusting the priority of the abnormal judgment chain according to the spatio-temporal weight.

[0018] S8: The suspiciously traced path generated by the abnormal judgment chain is classified and displayed according to high, medium and low suspicious levels, and the confidence interval and key abnormal node information of the corresponding path are output to the result visualization interface to support intuitive interpretation and disposal feedback of business decision makers.

[0019] The application provides a solid waste tracing analysis method based on a knowledge graph, and through dynamic fusion of multi-source heterogeneous sensing data and Bayesian confidence inference, high-precision identification of abnormal tracing paths in solid waste flow business and improvement of decision confidence are realized.

[0020] (1) The industry problem of "time and space errors in solid waste flow data collection and synchronization in complex scenarios leading to abnormal reasoning errors" is solved. Through multi-modal information collaborative collection (structured documents, GPS, RFID, video, environmental sensing, etc.), supplemented by node physical and spatial unified coding, data synchronization alignment and multi-source feedback repositioning mechanism, high consistency fusion of cross-device, cross-platform and cross-scene full-link time and space data is realized, and the problem of high false positives or false negatives caused by heterogeneous collection devices, data packet loss and clock errors is fundamentally solved.

[0021] (2) The Bayesian network probability inference is innovatively introduced to realize the quantification of solid waste flow node time and space information quality and adaptive correction of uncertainty. Through spatial-time division, window sliding, structured Bayesian probability modeling and MCMC sampling of node-level multi-modal data, the method can generate a dynamic confidence probability distribution interval for each key node, and clearly measure the reliability and abnormal sensitivity of the node data. Compared with the traditional flow comparison method based on single rule or static threshold, the model can reduce the abnormal detection error by more than 30% in uncertain data scenarios, effectively improving the robustness and logical adaptive ability of identification, and being suitable for extremely complex and multi-node switching actual business environment.

[0022] (3) A multi-modal collaborative correction process is realized. For the suspected abnormal nodes screened out by the abnormal judgment algorithm, the application automatically triggers the accurate backtracking mechanism of heterogeneous data (such as video supplement, RFID supplement, residual sensing, etc.), dynamically corrects the node state through information fusion and probability repositioning, and realizes adaptive optimization of the confidence interval of the abnormal node. Compared with the previous dependence on manual inspection or static supplement, the overall correction efficiency is improved by more than 100%, the traceability and recoverability of abnormal judgment are significantly enhanced, and the cost of manual review is effectively reduced.

[0023] (4) Introduce a confidence-based knowledge graph dynamic injection and causal reasoning mechanism. The node and path information corrected in coordination is updated to the knowledge graph structure in real time, combined with a weighted causal reasoning algorithm, which can not only quantitatively sort the abnormal suspicious scores of solid waste flow path, but also dynamically adjust the link priority considering the time and space weights. This mechanism breaks through the technical bottleneck of "fragmentation and static" in the existing solid waste supervision platform, realizes high-priority warning and risk sorting of abnormal links for the whole process, and significantly improves the business response speed.

[0024] (5) Support visual hierarchical display and business feedback loop of abnormal traceability results. Through high, medium and low hierarchical marking of path suspicious levels, as well as node confidence interval and abnormal point highlight output, the intuitive nature and decision efficiency of terminal business personnel are greatly improved. At the same time, business feedback and new data automatically flow back to iterate the training of Bayesian model and knowledge graph, realizing the full-link closed-loop optimization of data-model-knowledge, and continuously enhancing the generalization ability and self-learning ability of the system. The accuracy of long-term running abnormality discrimination remains at a high level. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 is a flowchart of a solid waste traceability analysis method based on a knowledge graph according to the present application. DETAILED DESCRIPTION

[0026] Embodiments of the present application are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be understood as limiting the present application.

[0027] The following disclosure provides many different embodiments or examples for implementing different structures of the present application. In order to simplify the disclosure of the present application, the components and settings of specific examples are described below. Of course, they are only examples and the purpose is not to limit the present application. In addition, the present application can repeatedly refer to numerals and / or letters in different examples, and such repetition is for the purpose of simplification and clarity, which itself does not indicate the relationship between the various embodiments and / or settings discussed. In addition, the present application provides examples of various specific processes and materials, but those skilled in the art can realize the application of other processes and / or the use of other materials.

[0028] The present application provides a solid waste traceability analysis method based on a knowledge graph, specifically including the following steps:

[0029] S1: Collect multi-modal data of solid waste flow at key nodes, including structured transfer documents, GPS trajectory data, RFID signals, video monitoring images, and environmental sensor spatio-temporal information, to obtain multi-modal raw data sets of solid waste flow across different geographical regions and time periods.

[0030] S2: Perform spatio-temporal data cleaning and unified preprocessing on the multi-modal raw data sets, including denoising, format normalization, and timestamp alignment of various data, to eliminate synchronization errors caused by heterogeneous collection devices and unify geographic coding.

[0031] S3: For different node geographical regions and time period labels, input the preprocessed multi-modal data into a Bayesian network inference model to dynamically calculate the confidence probability distribution interval of each node spatio-temporal information, and realize the probability quantification of solid waste flow data quality.

[0032] S4: Based on the node confidence probability distribution interval output by the Bayesian network, use an abnormal deviation discrimination algorithm to identify spatio-temporal sequence abnormalities in the solid waste transfer path, and judge the spatial and temporal consistency of the actual transfer path and the declared path under a specified probability threshold.

[0033] S5: If an abnormal deviation is found in the path discrimination at a specific node, call the heterogeneous perception signals associated with the abnormal node, including supplementary video monitoring, RFID supplementary information, and residual sensor data, to perform multi-modal collaborative repositioning and probability correction of the abnormal node spatio-temporal state.

[0034] S6: Take the multi-modal collaborative corrected key node confidence probability distribution as input, dynamically update the solid waste flow node and path relationship in the knowledge graph, obtain the corrected and injected knowledge graph solid waste flow information, and realize high-confidence injection of solid waste flow knowledge in different geographical regions and different time periods.

[0035] S7: Based on the corrected and injected knowledge graph solid waste flow information, use a weighted causal reasoning algorithm to calculate the global abnormal suspicious score of each flow path, and adjust the priority of the abnormal decision chain according to the spatio-temporal weight.

[0036] S8: The suspicious traceability path generated by the abnormal decision chain is displayed in high, medium, and low suspicious levels, and the confidence interval and key abnormal node information of the corresponding path are output to the result visualization interface to support intuitive interpretation and disposal feedback of business decision makers.

[0037] S1: Collect multimodal data on the flow of solid waste to key nodes. This multimodal data includes structured transfer documents, GPS trajectory data, RFID signals, video surveillance images, and spatiotemporal information from environmental sensors to obtain a raw multimodal dataset of solid waste flow spanning different geographical regions and time periods. Specifically, it includes:

[0038] S1.1: Define the physical environment and geolocation of key nodes in the flow of solid waste, and generate standardized geographic labels based on the map information system (GIS) module to support the consistency fusion of subsequent spatiotemporal data and the accuracy of node positioning.

[0039] In the solid waste flow data collection process, the physical attributes and spatial geographic information of key nodes in the solid waste flow are used as input conditions to perform node environment definition and geographic location coding processes, ensuring that standardized spatial labels are provided for subsequent multimodal data normalization and fusion and accurate anomaly tracing and reasoning.

[0040] Using physical environment identification algorithms, the attributes of each key business node in the solid waste circulation chain (such as waste-generating enterprises, collection centers, transfer stations, transit warehouses, and final disposal points) are qualitatively characterized, including elements such as the building use, process function, and capacity of the node, and preliminary node environmental description data is organized.

[0041] The geographic information acquisition module, combined with a high-precision GPS coordinate acquisition instrument and a mapping data interface, acquires the geographic coordinates (latitude, longitude, altitude, etc.) corresponding to the physical location of the node, as well as basic spatial data such as its administrative division and terrain features.

[0042] By utilizing the map information system (GIS) module and employing spatial vector coding algorithms to standardize the encoding of node geographic information, the actual collected spatial data is converted into a unified geographic code in the GIS database and assigned a unique spatial label ID to support cross-location of multi-source spatiotemporal data.

[0043] Furthermore, through spatial topology modeling algorithms, the spatial adjacency of each node in the solid waste flow path, its geographical unit (such as park, street, urban area, administrative division), and the distance, orientation, and spatial nesting relationship between nodes within the flow path are analyzed to generate spatial topology description indicators at the node level and path level.

[0044] Through the spatial reference conversion mechanism of the GIS module, the original geographic information of nodes collected from different types of equipment (such as RFID, circulation documents, video surveillance, environmental sensors, etc.) is uniformly converted to a coordinate system (such as WGS-84 to GCJ-02 or BD-09, etc.), ensuring that all subsequent heterogeneous spatiotemporal data have a unified spatial reference, forming a spatial standard support system for data collection, analysis, fusion and intelligent reasoning.

[0045] Through multi-level algorithm processing of node physical environment definition and geographic location coding, the physical and spatial heterogeneous information of solid waste flow to key nodes in the entire chain is transformed into standardized and structured geographic tag data. This provides a unified coordinate basis for the subsequent synchronous collection, feature analysis and high-precision anomaly inference modules of multimodal spatiotemporal information of solid waste flow, and significantly improves the spatial consistency and positioning accuracy of nodes in the solid waste traceability chain.

[0046] For example, taking the deployment of a provincial solid waste management platform in a typical industrial park as an example, for 80 solid waste flow business nodes, a high-precision resolution GIS (better than 5m) is used to collect the physical address attributes of each node, and a third-party high-precision differential GPS data acquisition device is linked to obtain the actual coordinates ((x,y,z)) of the node. Hierarchical spatial coding is used in the GIS platform to assign a unique spatial identifier to each node, such as "CN-GD-4403-PL001". Through spatial topology modeling, the spatial distance, relative orientation, and administrative region affiliation of the three nodes—Enterprise A (CN-GD-4403-PL001), Transfer Station B (CN-GD-4403-PL023), and Disposal Site C (CN-GD-4403-PL078)—are linked one-to-one, forming a structured spatial link. The original coordinates (WGS-84) collected from the GPS positioning device are batch-converted to the local reference system (GCJ-02) using GIS tools, with a coordinate conversion error better than 1.5 meters. The output spatial coding table is structured, allowing multimodal data (such as RFID scans, environmental sensor curves, and video surveillance frames) to be efficiently associated with spatial tags like "CN-GD-4403-PL001" in subsequent tasks involving solid waste flow data collection and source tracing anomaly inference. This significantly improves data fusion efficiency and spatial accuracy of anomaly tracing results. Actual business scenario testing shows a spatial tag matching success rate of 98.3% and an average geographical location error of 1.1 meters, effectively supporting accurate identification of solid waste path tracing in complex geographical scenarios.

[0047] S1.2: Collect structured transfer document data from the enterprise / regulatory unit, and use electronic document recognition and parsing algorithms to extract the declaration information, operator identity, operation time and destination attributes of solid waste flow nodes, and generate a structured transfer document feature set.

[0048] Using the geographic tags and node types of solid waste flow to key nodes as initial input conditions, the original electronic transfer document sets corresponding to the actual business of the nodes are imported for the solid waste management platform that is simultaneously connected to by enterprises and regulatory units.

[0049] An electronic document recognition and parsing algorithm (based on layout structure analysis and content semantic annotation strategy, with parameters including template type library, character set adaptive threshold, and key field dictionary) is adopted to achieve automatic recognition and format parsing of various types of circulation documents such as batch-imported PDFs, scanned documents, or structured tables.

[0050] Furthermore, through field deep extraction and semantic annotation algorithms (supporting regular expression extraction, text semantic discrimination based on the Named Entity Recognition (NER) engine, core business fields such as solid waste category, quantity, transfer node (sending, receiving), operator signature, declaration time, and destination code in the original transfer documents are parsed item by item and intelligently mapped to a standard business field set to achieve standardized data extraction.

[0051] Furthermore, an automatic identity and compliance verification algorithm (including online compliance verification and fingerprint comparison of fields such as the applicant's ID number, operator's job code, and corporate legal person code) is applied to batch verify the identity of the operators and corporate information involved in the transfer documents, eliminate non-compliant or forged identity records, and generate document structure data groups that have passed identity authentication.

[0052] Furthermore, a time-domain aggregation and destination association algorithm (parameters being event timestamp tolerance and node geographic label matching accuracy) is adopted. For each document data item, the unified archiving of operation time and destination spatial label binding are performed. The business time point and target destination attribute of the delivery behavior in the document data are mapped one by one with the node spatial code of the previous stage, and the document node feature vector set with accurate spatiotemporal alignment is output.

[0053] The structured transfer document feature set generation module transforms all data items that have undergone batch identification, field parsing, identity verification, and spatiotemporal aggregation into a standardized solid waste flow declaration dataset. This enables the output of complete, accurate, and highly spatiotemporally consistent features of solid waste flow node-level declaration data, significantly improving the quality of the basic data for subsequent source tracing path anomaly identification.

[0054] For example, the provincial-level hazardous waste comprehensive supervision cloud platform receives approximately 27,000 original PDFs and spreadsheets of "Hazardous Waste Transfer Manifests" uploaded annually by 86 waste-generating enterprises. The platform is configured with 36 recognition model templates based on layout parsing and NER (Non-Executive Analyzer), achieving 100% template library coverage. The character recognition threshold is set to 0.96, and the platform automatically synchronizes with the public security identity database and tax legal entity database for identity compliance verification. Extracted fields include transfer order number, waste code, spatial codes of the transfer and disposal enterprises, the handler's ID number and mobile phone number, generation date, outbound time, and signed-in time. After batch recognition, the average document recognition rate reaches 99.2%, the field extraction accuracy rate is 98.7%, and the abnormal identity rejection rate is 5.1%. By mapping time-series data (time deviation threshold ±3 minutes) and binding it with spatial tags, all document data was successfully aggregated to the corresponding geographical nodes and event timelines, generating structured feature items. Each document generated 22 standardized fields, forming a complete solid waste flow document feature data table. This fully prepares for subsequent integration with GPS trajectories, RFID, environmental, and video multimodal signals. This step effectively ensured the efficient coupling of business data and physical flow paths, improving the data traceability accuracy by more than 12% in cross-regional declarations, intermodal transport at distribution nodes, and link anomaly investigation tasks.

[0055] S1.3: Collect data from the positioning equipment on the transfer vehicles or containers, using high-precision positioning equipment for collaborative data collection.

[0056] GPS trajectory data is used to optimize trajectory sampling accuracy through differential positioning calibration algorithms, forming a standardized GPS trajectory feature vector sequence.

[0057] S1.4: Deploy radio frequency identification (RFID) sensing devices at solid waste transfer nodes to collect tag identification signals during the solid waste entering / leaving the warehouse / transfer process, use a time synchronization mechanism to calibrate RFID collection events, and generate RFID signal event records with timestamps.

[0058] S1.5: Deploy video surveillance systems at key nodes and in the solid waste transfer chain, using intelligent video stream acquisition and key frame extraction algorithms to obtain spatiotemporal image sequences and abnormal behavior characteristics of key nodes, and generate video surveillance image feature data blocks.

[0059] S1.6: Deploy multiple types of environmental sensors to collect environmental data related to solid waste circulation, such as temperature, humidity, gas concentration, and infrared detection. Utilize a heterogeneous signal synchronization alignment algorithm to accurately match the environmental sensor data with the corresponding node spatiotemporal labels, generating sensor spatiotemporal information feature vectors.

[0060] S1.7: By integrating the above-mentioned structured flow document feature set, GPS trajectory feature vector sequence, RFID signal event record, video surveillance image feature data block and sensor spatiotemporal information feature vector, a complete multidimensional raw dataset of solid waste flow is generated based on the multimodal feature integration framework, providing high-value raw input for subsequent data preprocessing and spatiotemporal anomaly inference modules.

[0061] S2: Perform spatiotemporal data cleaning and unified preprocessing on the multimodal raw dataset, including denoising, format normalization, and timestamp alignment of various types of data, to eliminate synchronization errors caused by heterogeneous acquisition devices and unify geocoding, specifically including:

[0062] S2.1: The structured transfer documents, GPS trajectory data, RFID signals, video surveillance images and environmental sensor spatiotemporal information in the multimodal raw dataset are uniformly collected to aggregate the original solid waste flow multimodal raw dataset across different geographical regions and time periods, forming a normalized multimodal raw data input.

[0063] S2.2: Anomaly detection and low-noise filtering algorithms are used to perform multidimensional denoising on the numerical data (such as GPS trajectory data, RFID signals, and spatiotemporal information of environmental sensors) in the normalized multimodal raw data input. Low-confidence values ​​caused by random errors, sensor drift and packet loss are removed to realize a noise-suppressed dataset of spatiotemporal numerical information of solid waste flow.

[0064] S2.3: For the noise suppression dataset, normalized coding and structure standardization algorithms are used to perform field mapping and unified format conversion on heterogeneous data types such as structured flow documents, radio frequency identification signals, and spatiotemporal information from environmental sensors, so as to achieve a consistent and standardized dataset for the multimodal data structure of solid waste flow.

[0065] S2.4: Using a multi-timezone timestamp alignment algorithm, various types of data in the standardized dataset are aligned according to the event occurrence time stamp, the physical time of the acquisition device, and the system synchronization time signal to achieve a unified time reference for aligning the dataset of solid waste flow multimodal data.

[0066] S2.5: Based on Geographic Information System (GIS) geocoding mapping, the spatial coordinates corresponding to GPS trajectory data, radio frequency identification signals and environmental sensor spatiotemporal information in the time-aligned dataset are spatially reduced and labeled with standard geographic tags, outputting a solid waste flow direction multimodal geocoding dataset with a unified spatial coordinate system and accurate geographic identification.

[0067] S3: For the geographical region and time period labels of different nodes, the preprocessed multimodal data is input into the Bayesian network inference model to dynamically calculate the confidence probability distribution interval of the spatiotemporal information of each node, thereby realizing the probability quantification of the solid waste flow data quality. Specifically, this includes:

[0068] S3.1: For the multimodal dataset generated by the unified spatiotemporal data preprocessing module, the node sample space is divided based on the geographic region label of each key node in order to provide spatial prior distribution input for the Bayesian network inference model and achieve refined spatial representation dimension at the node level.

[0069] S3.2: Based on the spatially segmented multimodal data received in S3.1, the time period labels associated with nodes are used to generate temporal subsets of the data in the geographic region node sample space according to the time dimension using a sliding window to obtain spatial-temporal joint subset data, providing continuous spatiotemporal input for the Bayesian network inference model, and realizing the subdivision and extraction of temporal features of different historical and real-time scenarios.

[0070] Using the multimodal preprocessed data within the geographic region node sample space after S3.1 spatial division as the input object, the structured flow document features, GPS trajectory features, radio frequency identification signals, video surveillance image features, and environmental sensor spatiotemporal feature sets of each key node are loaded into the continuous time series processing unit.

[0071] A method is adopted to associate node time period labels and generate time series subsets using a sliding window (parameters: window length L, step size Δt). The node data is segmented and processed in batches according to the time axis to ensure coverage of all historical and real-time time period intervals. Specifically, by setting the window length L and the window step size Δt, sliding window partitioning of multimodal data is implemented within the sample space of each geographic region node to form overlapping or non-overlapping temporal data subsets.

[0072] Furthermore, by aggregating the timestamps of data within each window, the multi-source temporal features of transfer documents, GPS trajectories, RFID signals, video features, and sensor data are synchronously encapsulated to construct a unified standard input structure at the window level, ensuring the temporal continuity data requirements of the Bayesian network.

[0073] A temporal completion algorithm is used to fill in missing data or perform interpolation (such as linear interpolation, forward padding, etc.) when there are missing packets or incomplete data in the window, so as to prevent data breakpoints from causing estimation errors in subsequent probability modeling.

[0074] By using the time-period weight correction method, based on the metadata such as the important business periods of the node and the historically frequent abnormal periods, the temporal labels of the window time-series subset are mapped to weight indicators, introducing time-series weight priors for subsequent joint probability modeling of the Bayesian network, and realizing adaptive enhancement of the model's sensitivity to anomalies in key periods.

[0075] Through the above chain processing method, the multimodal raw data within the spatially partitioned nodes is further finely segmented and feature synchronized according to the windowed time dimension, and transformed into a continuous spatiotemporal input subset that can be directly input into the Bayesian network inference model, thereby realizing the refined extraction and expression of node temporal features.

[0076] For example, at the disposal plant node of solid waste transportation company A, the spatial code is "X123", and the time window length is set to L = 2 hours and the time step is Δt = 30 minutes. For the normalized multimodal spatiotemporal data from June 1st to June 7th, 2024, a sliding window processing method is used to generate time-series subsets that slide every 2 hours and every 30 minutes, resulting in a total of 312 window units. Within each window, GPS trajectories are resampled to 5-minute intervals, RFID signals are clustered according to event occurrence time, transfer documents are aggregated according to declaration timestamps, 10 keyframes are collected from video feature points, and environmental detection is processed using a 10-minute average. To address the 5% packet loss in RFID within the window, forward padding interpolation is used to compensate. Historical data analysis shows that 8:00-10:00 and 18:00-20:00 are the peak periods for anomalies at this plant node each day. These time-period subsets are assigned a weight w = 1.5 to enhance time-period sensitivity. The final output, with each windowed subset containing a standardized multimodal feature structure, complete time-series labels, and time-series weight parameters, provides high-quality, continuous spatiotemporal input for subsequent Bayesian network probabilistic modeling. Validation results show that after sliding windowing and completion processing, the window subset coverage is 100%, the feature time series is complete, the data synchronization error is controlled within 1 minute, and the recall rate of the downstream anomaly detection model is improved by 6.2%.

[0077] S3.3: Based on the spatial-temporal joint subset data generated in S3.2, a spatiotemporal confidence probability model for each node is constructed using structured Bayesian network modeling technology. The Bayesian network structure is obtained by coupling and modeling the features of multi-source data in a joint probability graph manner, providing a probabilistic structural basis for subsequent dynamic inference and anomaly detection.

[0078] The spatial-temporal joint subset data output from step S3.2 is used as the input object. The input data includes the normalized multimodal feature structure of each solid waste flow key node within the sliding time window, covering the features of the structured flow documents after time-series aggregation, GPS trajectory features, radio frequency identification signals, video surveillance image features, and spatiotemporal information features of environmental sensors, and is accompanied by spatial geocoding, time labels, and time period weight parameters.

[0079] Using structured Bayesian network modeling techniques, a spatial-temporal windowed subset of multimodal data is used as input to automatically generate a node-centric joint probabilistic graph structure. Specifically, through the setting of the structured Bayesian network topology, each node serves as a core variable node in the probabilistic graph, and various multi-source modal features (such as declaration fields, trajectory points, label events, video frame features, and environmental indicators) serve as conditional parent-child nodes or collaborative observation variables, respectively. The network structure is established based on the physical flow and causal logic of the nodes.

[0080] Furthermore, by applying data-driven structural learning algorithms (such as the expectation-maximization (EM) algorithm, greedy search, AIC / BIC minimization, and other model selection criteria), the dependency structure between nodes is automatically filtered and optimized within the window, and the directional, directed, or undirected edge relationships between feature variables and master nodes are corrected, thereby achieving dynamic tuning of the Bayesian network structure for data-adaptive purposes.

[0081] Using parameter learning algorithms (such as maximum likelihood estimation / maximum a posteriori estimation), under a fixed network structure, the probability parameters in each conditional probability table (CPT) are jointly solved for the windowed spatiotemporal data of each node. During parameter solving, the frequency of occurrence of multi-source features such as circulation documents, trajectories, tags, videos, and sensors is statistically analyzed under different node values ​​and parent node observation conditions, generating a non-unimodal probability distribution table that comprehensively reflects the linkage of multimodal features.

[0082] By employing a joint probability distribution expression, the above conditional probabilities are globally multiplied to achieve full probability modeling of the spatiotemporal feature vectors of nodes. The specific formula is as follows:

[0083]

[0084] Where X i For multimodal feature variables, Pa(X) i X in a Bayesian network i The parent node feature set.

[0085] Furthermore, to quantify the impact of multimodal observation variables on confidence outputs at nodes within a specific space-time window, a Markov chain Monte Carlo (MCMC) sampling method is employed to generate the sampled value distribution of the joint probability distribution within the window sample, thereby obtaining key confidence interval indicators.

[0086] Through the above structured Bayesian network modeling process, the multimodal spatiotemporal features of each spatial-time node are transformed into a windowed probability structure with joint probability distribution expression capabilities, providing an accurate probability modeling foundation for subsequent spatiotemporal dynamic reasoning, anomaly detection, and confidence interval estimation.

[0087] For example, in a municipal solid waste collection scenario, for the spatial node "F001" at the disposal plant, with a time window T = [8:00-10:00], the input data includes 12 points of GPS trajectory sampling over 5 minutes, 3 RFID entry signal events, 2 transfer documents, 20 abnormal feature frames from video surveillance, and 24 sets of environmental gas concentrations, along with the geographic tag "X135" and a time period weight w = 1.6. A Bayesian network subgraph centered on "F001@T" is automatically generated, and the node dependency structure is minimized using the AIC criterion: GPS trajectory → entry signal → document → abnormal video → environmental parameters. Maximum likelihood parameter estimation is applied to statistically analyze the GPS trajectory density within the legal range, calculate the conditional probability of RFID signal occurrence frequency within the specified time window, and collaboratively infer the probability of "normal / abnormal" attributes based on all observed features. A joint probability distribution is used to express the confidence level of abnormal events at the node, and a confidence interval [0.31, 0.89] is obtained through 10,000 MCMC samplings. Finally, the multimodal spatiotemporal confidence probability model under the time window of the disposal node is output, providing sample input for the downstream anomaly inference module and improving the sensitivity and practicality of anomaly identification.

[0088] S3.4: Using the Bayesian network structure constructed in S3.3, conditional probability inference between nodes is performed on the multimodal data of each key node. Based on the feature distribution of the node input and the parameter estimation of the global Bayesian network, the confidence probability distribution interval of the spatiotemporal features of the node is obtained.

[0089] Based on the structured Bayesian network model established in S3.3, the input is a normalized multimodal dataset under the same node's spatial-temporal window, including structured circulation document features, GPS trajectory features, radio frequency identification signals, video surveillance image features, and spatiotemporal features of environmental sensors, as well as the geographic label, time label, and time segment weight parameters of the spatial node.

[0090] Using a conditional probability inference algorithm, under a Bayesian network topology, the observation state of the multimodal feature input variables of each node is first encoded. Through probability mapping, the actual observation data is mapped to the value state set of each variable in the network, thereby realizing the probability prior update of the independent variable to the target node.

[0091] Furthermore, utilizing the observed features of a node and the state of its parent node, and through the chained conditional probability formula dependency, conditional probability chain updates are performed sequentially on the node and its parent variables, following the formula below:

[0092]

[0093] Where X represents the target node variable, and Pa(X) is the feature combination of its parent nodes. In practice, for the observation state within each window, the posterior probability of the node under the current feature observation is queried and calculated based on the Bayesian network parameter table.

[0094] Furthermore, a forward inference mechanism is adopted to traverse each observation sample within the space-time window in batches, and to perform sequential conditional inference under a multi-node joint probability structure, thereby realizing the dynamic probability output of nodes under different observation conditions.

[0095] A confidence interval fitting method (such as Bootstrap resampling and Markov chain Monte Carlo sampling) is used to sample a large number of nodes in the conditional probability inference output. The mean, variance and confidence interval of the distribution of nodes under different observation states are calculated to obtain the confidence probability interval index of the multimodal input of the nodes.

[0096] Furthermore, through probability normalization and confidence adjustment algorithms, the conditional probability outputs of all observed samples under the node window are fitted to the overall confidence probability distribution interval of the node-level spatiotemporal features, and the output is as shown in [p]. low ,p high [], which serves as a quantitative indicator of the quality of the current node within this time window.

[0097] By using the chain-like conditional probability inference and confidence interval evaluation method described above, the multimodal feature input based on the previous structured Bayesian network modeling is dynamically transformed into the confidence probability distribution interval of each node, thereby achieving the probabilistic quantification effect of spatiotemporal information on the flow of solid waste to key nodes.

[0098] For example, at a municipal solid waste transfer station, the spatial node "Y210" and the time window T = [15:00-17:00] are input with a GPS trajectory density feature of 0.92, RFID signal detection of 3 entry / exit events, video surveillance identification of 2 vehicle entry / exit events, and 8 sets of data reported by temperature and humidity sensors, along with the geographic label "Y210" and a window weight of 1.3. Using a trained Bayesian network, each modal input variable is first encoded as an observation state. After querying the parameter table, the following are obtained: GPS trajectory belongs to high-density coverage, with a conditional probability of 0.82; the node state probability under high-frequency RFID exit conditions is 0.79; and the conditional probability that the number of video surveillance entry / exit events matches the document record is 0.85. Based on the network topology and parent node states, the conditional probability formula is applied layer by layer:

[0099] P(node|GPS,RFID,Video,Sensor)

[0100] =P(GPS|Pa(GPS))·P(RFID|GPS)·P(Video|RFID)·P(Sensor|Video)

[0101] Substituting the above probabilities, the mean confidence probability distribution of node "Y210@T" was obtained through 1000 MCMC samplings, with a confidence interval of 0.81 and a confidence range of [0.62, 0.93]. The actual state of this node within this time window showed high consistency with the reported information, and the above interval can serve as a key data indicator for subsequent anomaly detection and quality assessment.

[0102] S3.5: The spatiotemporal confidence quality index of each node is formed by combining the geographic region label and time period label of each node in the solid waste flow network with the confidence probability distribution interval of each node obtained in S3.4. These indicators are used as inputs for solid waste anomaly detection and subsequent causal inference to achieve high-confidence data quality quantification.

[0103] S4: Based on the node confidence probability distribution interval output by the Bayesian network, an anomaly deviation discrimination algorithm is used to identify spatiotemporal sequence anomalies in the solid waste transfer path, and to determine the spatial and temporal consistency between the actual transfer path and the declared path under a specified probability threshold. Specifically, this includes:

[0104] S4.1: Using the spatiotemporal confidence probability distribution interval of solid waste flow nodes in the Bayesian network model as input conditions, based on the specified spatial distance threshold and time window parameters, extract the confidence probability distribution feature set of all nodes in the solid waste flow path, and generate the basic data packet for anomaly detection.

[0105] Using the spatiotemporal confidence probability distribution interval of the Bayesian network of solid waste flow to multiple nodes as input conditions, node-level confidence probability feature extraction is performed based on a pre-set spatial distance threshold d_thr and time window parameter t_win. A spatiotemporal window feature ensemble algorithm (parameters: node index N, spatial threshold d_thr, time window t_win) is employed to traverse all nodes in the solid waste flow path point by point, extracting the confidence probability distribution interval of each node. Synchronously aggregate with corresponding spatial geocoding, timestamp labels, declaration path identifiers, and window weights to achieve feature distribution normalization. Furthermore, using a confidence probability spatial-temporal pruning algorithm (parameters: geocoding mapping set, time period label group), the effective confidence probability distributions of all nodes are filtered and organized to conform to the set consistency criteria between spatial d_thr and temporal t_win, thus constructing a structured confidence probability distribution feature set.

[0106] Furthermore, a structured data regularization and standardization processing method is adopted to unify the confidence probability features from different modalities according to the spatiotemporal arrangement of nodes, generating a data packet with a multidimensional feature format, including: unique node identifier, spatial coordinates, timestamp, declaration status, window weight, measured confidence probability distribution interval, etc., to ensure the structural integrity of the data packet and the compatibility with subsequent algorithms.

[0107] Furthermore, through feature verification and missing data completion mechanisms, the consistency and completeness of the confidence probability distribution intervals of each node in the data packet and its spatial and temporal labels are detected. Missing or abnormal nodes are filled by probability interpolation or neighborhood mean, thereby improving the overall reliability of the data packet.

[0108] By employing aggregation and indexing optimization algorithms, the feature set is merged and indexed according to the order of the solid waste circulation network path and the actual and declared path labels, outputting a feature set of confidence probability distributions for all nodes in the solid waste circulation path. This is the basic data packet used in the anomaly detection algorithm.

[0109] Through the above processing method, the node confidence probability information output by the Bayesian network is standardized, serialized and formed into a structured feature set, realizing efficient data preparation for subsequent spatiotemporal consistency judgment of solid waste circulation paths.

[0110] For example, in the inter-regional transportation of municipal solid waste, for the five key transfer nodes numbered "P101-P105", the Bayesian network inference module outputs the following confidence probability distribution intervals for each node within a 2-hour time window: P101: [0.85, 0.92] (geographic code A12, time 09:00-11:00), P102: [0.71, 0.89] (A15, 11:00-13:00), P103: [0.52, 0.67] (B04, 13:00-15:00), P104: [0.80, 0.94] (B08, 15:00-17:00), P105: [0.66, 0.81] (C01, 17:00-19:00). A spatial distance threshold d_thr = 20 kilometers and a time window t_win = 2 hours were set. The spatial and temporal tags, window weights, and actual and reported path identifiers for each node were integrated into a structured feature format data packet. For a node (e.g., P103) exhibiting partial confidence interval loss (caused by RFID packet loss), the probability interval was corrected to [0.60, 0.78] using the average probability of its immediate neighbors (0.69). The aggregated data forms a complete... The dataset provides standardized input for downstream consistency judgment and spatial-temporal anomaly detection. Experimental results show that after feature extraction, data normalization, and completion, the feature set coverage reaches 100%, and the effectiveness of node confidence intervals is improved by 8.7% for multimodal input scenarios, providing a solid data foundation for the accuracy of subsequent anomaly tracing analysis.

[0111] S4.2: Apply the spatiotemporal consistency verification algorithm to the confidence probability distribution feature set to analyze the temporal arrangement and spatial adjacency of each solid waste flow node. By calculating the overlap of confidence probabilities between nodes and the interaction relationship with the probability distribution of adjacent points, spatial and temporal consistency analysis indicators are generated.

[0112] The set of confidence probability distribution features of solid waste transfer path nodes output from step S4.1 is used as the input object. The input data includes the confidence probability distribution interval, spatial geocode, timestamp, declaration path identifier and window weight parameter of each node.

[0113] A spatiotemporal consistency verification algorithm (parameters: spatial adjacency matrix A, temporal arrangement index T) is used to analyze the temporal arrangement and spatial adjacency characteristics of each node in the solid waste flow path.

[0114] Furthermore, by using an overlap calculation method (parameters: node pair (i,j), confidence probability interval [p_{low}^i,p_{high}^i],[p_{low}^j,p_{high}^j]), the intersection interval of the confidence probability distributions between adjacent nodes is calculated, and the overlap of the probability distributions is quantified. The specific calculation formula is as follows:

[0115]

[0116] Among them, Overlap i,j represents the overlap of the confidence intervals between nodes i and j, with a value range of [0,1].

[0117] Furthermore, an adjacency probability distribution interaction relationship analysis algorithm is adopted (parameters: the set of preceding and following adjacency points of node k in the path N_k, probability distribution set [p_{low}^k, p_{high}^k]), which generates a node-level spatiotemporal consistency quantification index by calculating the mean and variance of the probability distribution interaction between a node and all its direct adjacency nodes.

[0118] Furthermore, a spatial consistency analysis method (parameters: geocoding mapping table, spatial buffer threshold d_thr) is used to compare the physical distance of spatial codes between nodes, determine the rationality of the actual spatial layout of the connecting nodes, and generate a spatial consistency score by combining the overlap probability interval.

[0119] Furthermore, a time consistency discrimination method (parameters: time series index, time window t_win) is applied to compare the differences between node timestamps and their probability interval sequences, calculate the time series difference measure of consecutive nodes, and generate a time consistency index accordingly.

[0120] By integrating the above multiple algorithms, a comprehensive analysis index of spatial and temporal consistency is output, including a node probability overlap matrix, spatial consistency score, and temporal consistency score, providing an accurate data foundation for downstream abnormal deviation probability judgment.

[0121] By using a spatiotemporal consistency verification algorithm and a probability distribution interaction analysis method, the set of confidence probability distribution features is transformed into spatial and temporal consistency analysis indicators, providing basic data support for the detection of anomalies at nodes in the solid waste circulation path.

[0122] For example, in a municipal solid waste inter-regional transportation route containing five nodes (P101-P105), the confidence probability distribution intervals of nodes P102 and P103 in the input feature set are [0.71, 0.89] and [0.60, 0.78], respectively. Using the formula for calculating overlap, we obtain:

[0123]

[0124] This indicates that the confidence intervals between P102 and P103 have low overlap, suggesting an initial indication of abnormal data continuity.

[0125] Furthermore, in the spatial adjacency analysis, the physical distance between P103 geocode B04 and P104 code B08 is 18 kilometers, which is lower than the spatial threshold d_thr = 20 kilometers, and is therefore judged to be spatially continuous.

[0126] In the time consistency judgment, the time windows of nodes P103 and P104 are 13:00-15:00 and 15:00-17:00 respectively, with an interval of exactly 2 hours, which meets the set time window t_win=2 hours, and the time consistency score is 1.

[0127] Comprehensive analysis shows that the probability distribution overlap of segment P102-P103 in this path is low, which is a potential anomaly area. Spatial and temporal consistency indicators show that the path is physically and temporally continuous, but some probability characteristics are abnormal.

[0128] The final output node pair overlap matrix, spatial consistency score, and temporal consistency score serve as key input data for subsequent anomaly detection, enabling refined quantitative risk analysis of spatiotemporal anomalies in the solid waste flow path.

[0129] S4.3: Based on spatial and temporal consistency analysis indicators, the probability deviation discrimination algorithm is called to perform spatial distance and time window consistency tests on each node pair of the actual circulation path and the corresponding declaration path, and outputs the consistency probability value of the actual path and the declaration path of each node in the form of probability.

[0130] The spatial and temporal consistency analysis indicators delivered in step S4.2 are used as inputs, including parameters such as the probability interval overlap of node pairs, spatial consistency score, and temporal consistency score.

[0131] A probabilistic deviation discrimination algorithm (parameters: spatial distance threshold d_thr, time window t_win, overlap lower limit θ_o, consistency score threshold θ_s) is used to perform consistency tests on each corresponding node pair of the actual circulation path and the declared path.

[0132] Furthermore, through the spatial distance consistency discrimination submodule, based on the spatial geocoding information of the node pair and combined with the spatial consistency score, the physical distance d_{i,i′} between the path node i and the declared node i′ is compared with the threshold d_thr. If d_{i,i′}>d_thr, the spatial consistency of the node pair is judged as abnormal; otherwise, it is recorded as spatially consistent.

[0133] Furthermore, the time window consistency judgment submodule calculates the time difference Δt_{i,i′} between node i and the reporting node i′ based on the timestamp label of the node pair and the time window parameter t_win. The time window consistency test is performed using the following formula:

[0134]

[0135] Among them, t i and t i , representing the time stamps of the actual node and the reporting node, respectively.

[0136] Furthermore, using the probability interval overlap measurement operator, the overlap calculation formula is applied to each actual-reporting node pair:

[0137]

[0138] in, These represent the upper and lower confidence intervals of the actual nodes. The upper and lower confidence intervals of the reporting nodes, Overlap i,i′ Characterizes the consistency of the probability distributions of two nodes.

[0139] Furthermore, a consistency probability output module is employed to output a consistency probability value for each actual-reporting node pair based on spatial consistency, temporal consistency, and probability interval overlap. A weighted probability combination formula is used:

[0140] Pr consist (i,i′)=w s Consist space (i,i′)+w t Consist time (i,i′)+w o Overlap i,i′

[0141] Among them, w s ,w t ,w o Consist is a weighting coefficient for spatial and temporal consistency and overlap. space With Consist time These are the spatial and temporal consistency discriminant values, respectively. The above parameters are weighted and normalized, Pr consist (i,i′) takes values ​​in [0,1].

[0142] Furthermore, traversing all actual and declared node pairs along the solid waste circulation path, the above-mentioned discrimination and probability output process is executed sequentially, and the Probability Output for each node pair is calculated. consist (i,i′) is written into the node consistency probability matrix, serving as a core technical indicator for accurate identification of path-level anomaly locations.

[0143] By using a probability deviation discrimination algorithm and node consistency probability output, the three indicators of spatial and temporal consistency and joint probability interval overlap are transformed into the consistency probability value of the actual path and the declared path of each node, thereby realizing the quantitative detection of the consistency probability of the multi-node traceability path of solid waste flow in complex spatiotemporal anomaly scenarios.

[0144] For example, in the inter-regional transportation route of municipal solid waste (Route ID: C315), the actual route node N2 (geographic code B04, time 13:15-15:15, confidence interval [0.62, 0.80]) and the declared route node N2′ (geographic code B05, time 13:30-15:30, confidence interval [0.74, 0.90]) form a node pair. The physical distance d_{N2, N2′} = 8 kilometers, which is lower than the spatial threshold d_thr = 20 kilometers, so Consist_{space}(N2, N2′) = 1. The node time difference is |13:15-13:30| = 15 minutes, which is less than t_win = 2 hours, so Consist_{time}(N2, N2′) = 1. The overlap is:

[0145]

[0146] Set weight w s =0.25,w t =0.25,w o =0.5, then the probability of agreement is:

[0147] Pr consist (N2, N2′)=0.25×1+0.25×1+0.5×0.21=0.25+0.25+0.105=0.605

[0148] The output here shows that the consistency probability of node N2 is 0.605, which is close to the anomaly detection threshold of 0.6. The entire process is executed on all node pairs, generating a node-level consistency probability output matrix. Practical applications show that this judgment can accurately identify abnormal nodes in scenarios involving spatiotemporal synchronization errors, data packet loss, or business path tampering.

[0149] S4.4: Set an anomaly identification probability threshold, compare the node consistency probability values ​​output by the probability deviation discrimination algorithm with the threshold, mark solid waste flow nodes and associated paths with consistency probabilities lower than the preset probability threshold, and initially screen out spatiotemporal nodes suspected of having anomalies.

[0150] S4.5: Aggregate the selected abnormal spatiotemporal nodes and paths to generate an abnormal deviation marker sequence of the complete source tracing path, providing a probability-quantified abnormal detection input basis for subsequent multimodal collaborative correction and knowledge graph dynamic injection modules.

[0151] S5: If an abnormal deviation is detected at a specific node during path discrimination, the heterogeneous sensing signals associated with the abnormal node are invoked, including supplementary video surveillance, RFID data, and residual sensor data, to relocate and probabilistically correct the spatiotemporal state of the abnormal node through multimodal collaboration. Specifically, this includes:

[0152] S5.1: Based on the node anomaly probability distribution output by the Bayesian network model, target nodes with high anomaly suspicion scores are extracted and used as input conditions for the precise invocation of multimodal sensing signals in this round, so as to lock the abnormal position in the solid waste flow link.

[0153] S5.2: Based on the spatial geocoding and timestamp window corresponding to the abnormal node, automatically match and obtain supplementary video surveillance signals to realize the association between the spatial coordinates of the abnormal node and the physical operation event, which is used to verify the spatiotemporal consistency between the actual operation trajectory of the node and the abnormal judgment of the target node extracted in the previous step.

[0154] S5.3: Utilize the RFID supplementary information associated with abnormal nodes to parse the electronic tag data that identifies solid waste entry and exit operations. Through a time sequence consistency discrimination algorithm, identify the flow events in the RFID signal that match the spatial code and time range of the node to verify the actual physical flow status of the abnormal node.

[0155] S5.4: The remaining sensor data is called and spatiotemporally filtered. The physical inventory change sequence of the abnormal node is fused with the aforementioned RFID and video surveillance signal results in a multimodal manner. A new comprehensive probability distribution of the node is generated using the confidence joint distribution mechanism to realize spatiotemporal state relocation under multimodal signals.

[0156] S5.5: Based on the corrected node comprehensive probability distribution output by the multimodal signal relocation algorithm, the original node anomaly probability is dynamically corrected using a probability correction method to form a corrected node confidence probability distribution as the final output result, which is used for subsequent high-confidence injection of knowledge graph and global anomaly causal inference.

[0157] S6: Using the confidence probability distribution of key nodes after multimodal collaborative correction as input, the solid waste flow direction nodes and path relationships in the knowledge graph are dynamically updated to obtain the corrected solid waste flow direction information injected into the knowledge graph, achieving high-confidence injection of solid waste flow direction knowledge for different geographical regions and time periods, specifically including:

[0158] S6.1: Extract entity features from the confidence probability distribution data of key nodes after multimodal collaborative correction, identify and structure key information nodes (such as the location of solid waste flow, flow time, node type and confidence interval) based on the entity extraction algorithm, so as to form a solid waste flow node attribute set as the basic input for knowledge fusion.

[0159] S6.2: Using an entity alignment algorithm, the entity attribute set of the confidence probability distribution of key nodes is accurately compared with the existing solid waste flow nodes in the knowledge graph to realize the consistency verification of node identity and output the entity alignment list that meets the confidence threshold, providing accurate node mapping for subsequent path relationship fusion.

[0160] S6.3: Based on the entity alignment results, a dynamic relationship mapping algorithm is applied to assign probability weights to the upstream and downstream transfer relationships between nodes in the solid waste flow path, construct a confidence probability weighted path relationship set, and provide input basis for the optimization of relationship edges in the knowledge graph through data structured output.

[0161] S6.4: Based on the confidence probability weighted path relationship set, execute the knowledge injection engine algorithm to dynamically inject the solid waste flow node attribute set and path relationship set into the knowledge graph structure with high confidence labels, realize the dynamic high confidence fusion of multi-dimensional knowledge of solid waste flow in different geographical areas and time periods, and produce the latest knowledge graph structure data.

[0162] S6.5: For the updated knowledge graph structure data, a confidence verification algorithm is used to perform consistency and closed-loop verification on the relationship between solid waste flow nodes and paths. Potential spatiotemporal information conflicts or redundant relationships are automatically identified and labeled, realizing global consistency optimization of solid waste flow knowledge in the knowledge graph, and providing a high-quality knowledge foundation for subsequent causal reasoning.

[0163] S7: Based on the corrected and injected solid waste flow information into the knowledge graph, a weighted causal reasoning algorithm is used to calculate the global anomaly and suspicion score for each flow path, and the priority of the anomaly judgment chain is adjusted according to the spatiotemporal weights, specifically including:

[0164] S7.1: The node and path relationship information of the solid waste flow direction knowledge graph after multimodal collaborative correction is extracted in a hierarchical manner to construct a global spatiotemporal link subgraph of each solid waste flow direction path on the knowledge graph, so as to provide an input dataset with structured upstream and downstream semantic relationships for subsequent causal reasoning algorithms.

[0165] S7.2: Based on the global spatiotemporal link subgraph of solid waste flow path, a weighted causal reasoning algorithm is applied to integrate the confidence probability distribution intervals of node and path associations, model the causal impact of each solid waste flow path, and generate an initial anomaly suspicion score to realize the quantitative characterization of the abnormal impact of solid waste flow path.

[0166] S7.3: For the abnormal and suspicious scores generated by the causal reasoning algorithm, the spatiotemporal weight information of the solid waste flow path nodes (including the geographical span of the path, the temporal span, the local confidence probability, etc.) is called, and the spatiotemporal weighting operation is performed to dynamically correct the original abnormal and suspicious scores, so as to output an accurate abnormal judgment score that combines global reasoning and local feature perception.

[0167] S7.4: Based on the multi-path anomaly judgment score output by the weighted causal reasoning algorithm, the priority of the anomaly judgment chain is adjusted. According to the comprehensive anomaly score, spatiotemporal weight distribution and business priority strategy, all solid waste flow anomaly links are prioritized and sorted to form a sorting list, which provides decision support data for subsequent anomaly reporting, automatic alarm and business handling systems.

[0168] S7.5: The priority ranking results of the anomaly judgment chain and its associated spatiotemporal weight distribution information are structured and archived to form an anomaly tracing and discrimination structured result package that can be used for business visualization, subsequent closed-loop feedback and continuous model self-learning optimization, so as to realize the data accumulation and scalable call of the discrimination results.

[0169] S8: The suspicious tracing paths generated by the abnormal judgment chain are displayed in a hierarchical manner according to high, medium, and low suspicion levels, and the confidence interval and key abnormal node information of the corresponding paths are output to the result visualization interface to support the intuitive interpretation and handling feedback of business decision-makers. Specifically, this includes:

[0170] S8.1: Perform hierarchical calculation of the confidence probability distribution interval of the solid waste flow to the suspicious tracing path in the global anomaly judgment chain, so as to determine the high, medium and low suspicious confidence level of each tracing path under the Bayesian network, and realize the generation of confidence level labels for the tracing path.

[0171] S8.2: Based on high, medium and low doubt confidence level labels, the doubtful tracing paths of solid waste flow involved in the abnormal judgment chain are grouped and organized in a multi-level manner, and the key abnormal node information and confidence probability distribution range involved in each path are associated to form a hierarchical structured dataset of doubtful tracing paths.

[0172] S8.3: Using data visualization rendering algorithms, the hierarchical structured dataset of the above-mentioned suspicious tracing paths is mapped to the result visualization interface, high-confidence abnormal nodes are highlighted, and an interactive confidence interval numerical display module is designed for each tracing path, enabling business decision-makers to intuitively interpret the tracing judgment results.

[0173] S8.4: For suspicious tracing paths of each confidence level, generate a visual summary report containing path ID, confidence level range, list of key abnormal nodes, and spatiotemporal attribute metadata of the path occurrence, so that business decision-makers can conduct online judgment, event assignment, and emergency response.

[0174] S8.5: Collect decision-makers' interpretations, actions, feedback, and newly added annotations on the visualization interface, and automatically feed them back to the source tracing path hierarchical structured dataset and knowledge graph correction link, providing an input loop for the dynamic optimization and adaptive learning of subsequent Bayesian inference models and visualization hierarchical algorithms.

[0175] S9: Based on feedback from business decision-makers and newly collected multimodal data, dynamically iteratively train and correct the Bayesian network model and associated knowledge graph to adaptively improve the accuracy of identifying abnormal sources of solid waste flow and the model's generalization ability, achieving closed-loop optimization of data-model-knowledge, specifically including:

[0176] Long-term effectiveness and flexible response capabilities are key technological aspects for improving the accuracy and sustainable evolution of intelligent traceability platforms.

[0177] S9.1: Based on the identification results of abnormal solid waste flow paths, receive feedback information from business decision-makers (such as the accuracy of anomaly judgment, the evaluation of suspicious level, and the results of on-site disposal), and form a structured feedback dataset for key node information involving abnormal paths to guide subsequent model parameter correction and knowledge structure optimization.

[0178] S9.2: Perform spatiotemporal correlation feature extraction processing on newly collected multimodal data (including structured transfer documents, GPS trajectory data, RFID signals, video surveillance images and environmental sensor spatiotemporal information) with the key feedback information to form multimodal feature-feedback label pairs, providing standardized input for supervised retraining of the Bayesian network model.

[0179] S9.3: Based on multimodal feature-feedback label pairs, an incremental Bayesian network parameter learning algorithm is used to iteratively train and correct the structure of existing Bayesian network models, thereby achieving dynamic updates to spatiotemporal information uncertainties and improving the adaptive discrimination ability of node confidence probability distribution.

[0180] S9.4: Based on the node confidence probability distribution after iterative training of the Bayesian network model, dynamically adjust and optimize the entity attributes, edge weights, and causal inference weights of the solid waste flow node and path relationship in the knowledge graph, so as to achieve synchronous evolution of the knowledge graph structure and the model confidence association.

[0181] S9.5: The optimized knowledge graph nodes and causal reasoning chains are used to update the anomaly tracing reasoning judgment process, recalculate the global anomaly suspicion score chain and anomaly priority weight, and form a closed-loop solid waste flow anomaly tracing model - knowledge synchronous evolution system, providing highly adaptive support for the next round of judgment and business decision-making.

[0182] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0183] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0184] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A solid waste source tracing analysis method based on knowledge graphs, specifically including: S1: Collect multimodal data on solid waste flow to key nodes to obtain a raw multimodal dataset of solid waste flow across different geographical regions and time periods; S2: Perform spatiotemporal data cleaning and unified preprocessing on the multimodal raw dataset, including denoising, format normalization and timestamp alignment of various types of data; S3: For the geographical region and time period labels of different nodes, the preprocessed multimodal data is input into the Bayesian network inference model to dynamically calculate the confidence probability distribution interval of the spatiotemporal information of each node; S4: Based on the confidence probability distribution range of each node, use the anomaly deviation discrimination algorithm to identify spatiotemporal sequence anomalies in the solid waste transfer path, and determine the spatial and temporal consistency between the actual transfer path and the declared path under the specified probability threshold. S5: If an abnormal deviation is found in a specific node during path discrimination, the heterogeneous sensing signal associated with the abnormal node is invoked to relocate and probabilistically correct the spatiotemporal state of the abnormal node in a multimodal collaborative manner. S6: Using the corrected confidence probability distribution of key nodes as input, dynamically update the solid waste flow direction nodes and path relationships in the knowledge graph to obtain the corrected solid waste flow direction information injected into the knowledge graph; S7: Based on the solid waste flow information that has been corrected and injected into the knowledge graph, calculate the global anomaly suspicion score for each flow path, and adjust the priority of the anomaly judgment chain according to the spatiotemporal weights. S8: The suspicious tracing paths generated by the abnormal judgment chain are displayed in a hierarchical manner according to high, medium and low suspicion levels, and the confidence interval and key abnormal node information of the corresponding path are output to the result visualization interface to support the intuitive interpretation and handling feedback of business decision-makers.

2. The solid waste source tracing analysis method based on knowledge graphs according to claim 1, characterized in that: The multimodal data includes structured transfer documents, GPS trajectory data, radio frequency identification signals, video surveillance images, and spatiotemporal information from environmental sensors.

3. The solid waste source tracing analysis method based on knowledge graphs according to claim 2, characterized in that: Structured transfer document data is collected from enterprises / regulatory units, and electronic document recognition and parsing algorithms are used to extract declaration information, operator identity, operation time and destination attributes of solid waste flow nodes, and generate a structured transfer document feature set.

4. The solid waste source tracing analysis method based on knowledge graphs according to claim 2, characterized in that: The physical environment and geographic location coding of key nodes in the flow of solid waste are defined, and standardized geographic labels are generated based on the map information system module to support the consistency fusion of subsequent spatiotemporal data and the accuracy of node positioning.

5. The solid waste source tracing analysis method based on knowledge graphs according to claim 2, characterized in that: Data is collected from positioning equipment on vehicles or containers, and GPS trajectory data is collected in collaboration with high-precision positioning equipment. The accuracy of trajectory sampling is optimized through differential positioning calibration algorithm to form a standardized GPS trajectory feature vector sequence.

6. The solid waste source tracing analysis method based on knowledge graphs according to claim 1, characterized in that: The heterogeneous sensing signals in step S5 include: supplementary video surveillance, RFID supplementary recording information, and surplus sensor data.

7. The solid waste source tracing analysis method based on knowledge graphs according to claim 1, characterized in that: Step S3 specifically includes: For the multimodal dataset generated by the unified spatiotemporal data preprocessing module, the node sample space is divided based on the geographic region label of each key node; Based on the received spatially segmented multimodal data, the time period labels associated with nodes are used to generate time-series subsets of the data within the geographic region node sample space according to the time dimension using a sliding window, thus obtaining spatial-temporal joint subset data. Based on the joint spatial-temporal subset data, a spatiotemporal confidence probability model for each node is constructed, and the features of multi-source data are coupled and modeled in a joint probability graph manner to obtain a Bayesian network structure. Using the constructed Bayesian network structure, conditional probability inference between nodes is performed on the multimodal data of each key node. Based on the feature distribution of the node input and the parameter estimation of the global Bayesian network, the confidence probability distribution interval of the spatiotemporal features of the node is obtained. The confidence probability distribution intervals of the spatiotemporal characteristics of each node are combined with the geographical region label and time period label of the node in the solid waste flow network to form the spatiotemporal confidence quality index of the node output by the algorithm, and these indicators are used as inputs for solid waste anomaly discrimination and subsequent causal inference.

8. The solid waste source tracing analysis method based on knowledge graphs according to claim 1, characterized in that: Step S4 specifically includes: Using the confidence probability distribution interval of the spatiotemporal features of each node as input conditions, and based on the specified spatial distance threshold and time window parameters, the confidence probability distribution feature set of all nodes in the solid waste transfer path is extracted to generate a basic data package for anomaly detection. A spatiotemporal consistency verification algorithm is applied to the confidence probability distribution feature set to analyze the temporal arrangement and spatial adjacency of each solid waste flow node. By calculating the overlap of confidence probabilities between nodes and the interaction relationship with the probability distribution of adjacent points, spatial and temporal consistency analysis indicators are generated. Based on spatial and temporal consistency analysis indicators, a probability deviation discrimination algorithm is invoked to perform spatial distance and temporal window consistency tests on each node pair of the actual circulation path and the corresponding declared path, and outputs the consistency probability value of the actual path and the declared path of each node in probabilistic form; an anomaly identification probability threshold is set, and the consistency probability values ​​are compared with the threshold to mark the solid waste flow nodes and associated paths with consistency probabilities lower than the preset probability threshold, and preliminarily screen out spatiotemporal nodes suspected of having anomalies; the screened abnormal spatiotemporal nodes and paths are aggregated to generate an anomaly deviation mark sequence of the complete traceability path.

Images